This course develops the analytic theory of several complex variables through the geometry of real hypersurfaces and CR structures, with a focus on how boundary geometry governs holomorphic behavior. It begins by setting up real hypersurfaces in $\mathbb{C}^n$ and the intrinsic notion of a CR manifold, then uses model spaces such as the Heisenberg group and Siegel domain to illustrate the basic mechanisms behind [boundary regularity](/theorems/99), extension, and invariance. Along the way, the course introduces the tangential [Cauchy-Riemann equations](/page/Cauchy-Riemann%20Equations) as the natural differential system on CR manifolds and shows how they control the local and global behavior of CR functions. The middle chapters build from local analysis to deeper geometric and PDE phenomena. The Lewy [extension theorem](/theorems/59) and pseudoconvexity at the boundary lead into the Levi problem, explaining when boundary geometry forces holomorphic extendability and domain of holomorphy properties. Chern-Moser theory then provides a normal form and a set of CR invariants for hypersurfaces, while the $\bar{\partial}$-Neumann problem connects these geometric ideas to boundary regularity for the $\bar{\partial}$-operator. The final chapters broaden the perspective by studying CR embeddings, the Boutet de Monvel-Guillemin framework, propagation of CR singularities via the FBI transform, and selected global directions that connect analysis, geometry, and microlocal methods. # Introduction This opening chapter fixes the scope of the course and explains why boundary geometry in several complex variables is not a peripheral topic but one of the main places where holomorphic phenomena become visible. In one complex variable, boundaries are often controlled by conformal mapping, harmonic measure, and contour methods; in several variables, real hypersurfaces carry their own complex tangential geometry. The central objects of the course are CR structures, Levi forms, pseudoconvexity, tangential Cauchy-Riemann operators, and the extension or failure of extension of CR functions across boundaries. The course sits after a first encounter with holomorphic functions on domains in $\mathbb C^n$ and before the more microlocal theory of the $\bar{\partial}$-Neumann problem. We shall repeatedly pass between three languages: geometry of real submanifolds, analysis of differential operators acting only in complex tangential directions, and boundary values of holomorphic functions. The purpose of this introduction is to identify the questions that organise the lectures and to set the notation used throughout. ## The Boundary as a Complex Object The first problem is to understand what complex analysis can see on a real hypersurface. A real hypersurface $M \subset \mathbb C^n$ has real dimension $2n-1$, so it cannot itself be a complex submanifold. Nevertheless, at each point there are tangent directions that are stable under multiplication by $i$, and those directions form the complex tangent bundle. [definition: Real Hypersurface] Let $U \subset \mathbb C^n$ be open. A real hypersurface in $U$ is a smooth embedded submanifold $M \subset U$ of real codimension $1$. [/definition] A hypersurface is therefore too large to be complex and too small to be open. To make the surviving complex directions usable in analysis, we need a pointwise definition of the tangent directions that remain stable under the ambient complex structure. [definition: Complex Tangent Space] Let $M \subset \mathbb C^n$ be a real hypersurface and let $p \in M$. The complex tangent space at $p$ is \begin{align*} H_pM = T_pM \cap iT_pM. \end{align*} The holomorphic tangent space is the $(1,0)$ part \begin{align*} T_p^{1,0}M = \{L \in T_p^{1,0}\mathbb C^n : L\rho = 0\}, \end{align*} where $\rho$ is any local defining function for $M$ near $p$. [/definition] The definition refers to a local defining function, but the resulting space does not depend on the chosen defining function. The lectures begin by checking this invariance and then using it to define geometric quantities intrinsic to the hypersurface. [example: The Unit Sphere] Let $M=\{z\in\mathbb C^n:|z|^2=1\}$ and choose the defining function \begin{align*} \rho(z)=|z|^2-1=\sum_{k=1}^n z_k\bar z_k-1. \end{align*} Fix $p\in M$, and write an arbitrary $(1,0)$ vector at $p$ as \begin{align*} L=\sum_{j=1}^n a_j\partial_{z_j}. \end{align*} For each pair $j,k$, the complex coordinates $z_k$ and $\bar z_k$ are treated as independent when differentiating with respect to $z_j$, so \begin{align*} \partial_{z_j}(z_k\bar z_k)=(\partial_{z_j}z_k)\bar z_k+z_k(\partial_{z_j}\bar z_k)=\delta_{jk}\bar z_k+z_k\cdot 0=\delta_{jk}\bar z_k. \end{align*} Therefore \begin{align*} \partial_{z_j}\rho(z)=\partial_{z_j}\left(\sum_{k=1}^n z_k\bar z_k-1\right)=\sum_{k=1}^n\delta_{jk}\bar z_k-0=\bar z_j. \end{align*} Evaluating at $p$ gives \begin{align*} \partial_{z_j}\rho(p)=\bar p_j. \end{align*} Applying $L$ to $\rho$ at $p$ gives \begin{align*} L\rho(p)=\left(\sum_{j=1}^n a_j\partial_{z_j}\right)\rho(p)=\sum_{j=1}^n a_j\partial_{z_j}\rho(p). \end{align*} Substituting $\partial_{z_j}\rho(p)=\bar p_j$ yields \begin{align*} L\rho(p)=\sum_{j=1}^n a_j\bar p_j. \end{align*} By the definition of the holomorphic tangent space, $L\in T_p^{1,0}M$ exactly when $L\rho(p)=0$, and hence \begin{align*} T_p^{1,0}M=\left\{\sum_{j=1}^n a_j\partial_{z_j}:\sum_{j=1}^n a_j\bar p_j=0\right\}. \end{align*} The equation $\sum_{j=1}^n a_j\bar p_j=0$ is nontrivial. Indeed, since $p\in M$, \begin{align*} \sum_{j=1}^n |p_j|^2=|p|^2=1. \end{align*} If every coefficient $\bar p_j$ were $0$, then every $p_j$ would be $0$, so the left side would be $0$, contradicting $\sum_{j=1}^n |p_j|^2=1$. Thus at least one $\bar p_j$ is nonzero. The map \begin{align*} \sum_{j=1}^n a_j\partial_{z_j}\longmapsto \sum_{j=1}^n a_j\bar p_j \end{align*} is therefore a nonzero complex linear functional on $T_p^{1,0}\mathbb C^n$, and $T_p^{1,0}M$ is its kernel, a complex hyperplane. Next compute the Levi form. Since $\partial_{z_j}\rho=\bar z_j$, differentiating with respect to $\bar z_k$ gives \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\bar z_k}(p)=\frac{\partial}{\partial\bar z_k}(\bar z_j)(p)=\delta_{jk}. \end{align*} For $L=\sum_{j=1}^n a_j\partial_{z_j}\in T_p^{1,0}M$, the Levi form is \begin{align*} \mathcal L_{\rho,p}(L,\overline L)=\sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial\bar z_k}(p)a_j\overline{a_k}=\sum_{j,k=1}^n\delta_{jk}a_j\overline{a_k}. \end{align*} When $j\ne k$, $\delta_{jk}=0$, and when $j=k$, $\delta_{jj}=1$. Hence \begin{align*} \sum_{j,k=1}^n\delta_{jk}a_j\overline{a_k}=\sum_{j=1}^n a_j\overline{a_j}=\sum_{j=1}^n |a_j|^2. \end{align*} Thus \begin{align*} \mathcal L_{\rho,p}(L,\overline L)=\sum_{j=1}^n |a_j|^2. \end{align*} If $L\ne 0$, then some coefficient $a_j$ is nonzero, so the corresponding summand $|a_j|^2$ is positive and every other summand is nonnegative. Therefore $\mathcal L_{\rho,p}(L,\overline L)>0$ for every nonzero $L\in T_p^{1,0}M$, so the unit sphere is the basic model of a strongly pseudoconvex boundary. [/example] This example already shows the pattern of the course. The ambient holomorphic coordinates give concrete formulas, but the important information is invariant under biholomorphic changes of coordinates. ## Levi Geometry and Pseudoconvexity The next problem is to measure how the complex tangent directions bend inside the hypersurface. Ordinary curvature is not the right invariant for holomorphic extension questions. The relevant object is the Levi form, a Hermitian form on complex tangent vectors obtained by differentiating a defining function twice in complex directions. [definition: Levi Form of a Hypersurface] Let $M = \{\rho = 0\} \subset \mathbb C^n$ be a smooth real hypersurface with $d\rho \ne 0$ on $M$. For $p \in M$, the Levi form of $M$ with respect to $\rho$ is the Hermitian form on $T_p^{1,0}M$ given by \begin{align*} \mathcal L_{\rho,p}(L,\overline{K}) = \sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial \bar z_k}(p)L_j\overline{K_k}. \end{align*} [/definition] Changing the defining function rescales the Levi form by a positive factor when the orientation is fixed. Thus the sign of the Levi form is the invariant datum, while its numerical normalization is not. The boundary condition needed for extension theory is precisely the absence of negative complex-tangential curvature on the side from which holomorphic functions are to extend. [definition: Pseudoconvex Hypersurface] Let $M \subset \mathbb C^n$ be an oriented real hypersurface with a local defining function $\rho$ chosen so that the pseudoconvex side is $\{\rho < 0\}$. The hypersurface is pseudoconvex at $p \in M$ if $\mathcal L_{\rho,p}$ is positive semidefinite on $T_p^{1,0}M$. [/definition] Pseudoconvexity is the boundary version of holomorphic convexity. In later chapters it will control both local extension results and estimates for the tangential Cauchy-Riemann complex. [example: Sphere and Hyperquadric] For the sphere $M_S=\{|z|^2=1\}\subset\mathbb C^n$, take \begin{align*} \rho_S(z)=|z|^2-1=\sum_{m=1}^n z_m\bar z_m-1. \end{align*} We compute its Levi form by differentiating $\rho_S$ in the holomorphic and antiholomorphic coordinate directions. Treating $z_m$ and $\bar z_m$ as independent coordinates, for each pair $j,m$, \begin{align*} \partial_{z_j}(z_m\bar z_m)=(\partial_{z_j}z_m)\bar z_m+z_m(\partial_{z_j}\bar z_m)=\delta_{jm}\bar z_m+z_m\cdot 0=\delta_{jm}\bar z_m. \end{align*} Therefore \begin{align*} \frac{\partial\rho_S}{\partial z_j}=\partial_{z_j}\left(\sum_{m=1}^n z_m\bar z_m-1\right)=\sum_{m=1}^n\delta_{jm}\bar z_m-0=\bar z_j. \end{align*} Differentiating with respect to $\bar z_k$ gives \begin{align*} \frac{\partial^2\rho_S}{\partial z_j\partial\bar z_k}=\frac{\partial}{\partial\bar z_k}(\bar z_j)=\delta_{jk}. \end{align*} Let $p\in M_S$ and let \begin{align*} L=\sum_{j=1}^n a_j\partial_{z_j}\in T_p^{1,0}M_S. \end{align*} By the definition of the Levi form, \begin{align*} \mathcal L_{\rho_S,p}(L,\overline L)=\sum_{j,k=1}^n \frac{\partial^2\rho_S}{\partial z_j\partial\bar z_k}(p)a_j\overline{a_k}. \end{align*} Substituting $\frac{\partial^2\rho_S}{\partial z_j\partial\bar z_k}(p)=\delta_{jk}$ yields \begin{align*} \mathcal L_{\rho_S,p}(L,\overline L)=\sum_{j,k=1}^n \delta_{jk}a_j\overline{a_k}. \end{align*} The summand is $0$ when $j\ne k$, and when $j=k$ it is $a_j\overline{a_j}=|a_j|^2$, so \begin{align*} \mathcal L_{\rho_S,p}(L,\overline L)=\sum_{j=1}^n |a_j|^2. \end{align*} If $L\ne 0$, then some coefficient $a_j$ is nonzero, so the corresponding term $|a_j|^2$ is positive and all other terms are nonnegative. Hence the Levi form of the sphere is positive definite on every complex tangent space. For the hyperquadric \begin{align*} M_Q=\{(z_1,z_2,w)\in\mathbb C^3:\operatorname{Im}w=|z_1|^2-|z_2|^2\}, \end{align*} use the defining function \begin{align*} \rho_Q(z_1,z_2,w)=|z_1|^2-|z_2|^2-\operatorname{Im}w=z_1\bar z_1-z_2\bar z_2-\frac{w-\bar w}{2i}. \end{align*} The holomorphic first derivatives are \begin{align*} \frac{\partial\rho_Q}{\partial z_1}=\partial_{z_1}(z_1\bar z_1)-\partial_{z_1}(z_2\bar z_2)-\partial_{z_1}\left(\frac{w-\bar w}{2i}\right)=\bar z_1-0-0=\bar z_1. \end{align*} \begin{align*} \frac{\partial\rho_Q}{\partial z_2}=\partial_{z_2}(z_1\bar z_1)-\partial_{z_2}(z_2\bar z_2)-\partial_{z_2}\left(\frac{w-\bar w}{2i}\right)=0-\bar z_2-0=-\bar z_2. \end{align*} \begin{align*} \frac{\partial\rho_Q}{\partial w}=\partial_w(z_1\bar z_1)-\partial_w(z_2\bar z_2)-\partial_w\left(\frac{w-\bar w}{2i}\right)=0-0-\frac{1}{2i}=\frac{i}{2}. \end{align*} Differentiating these with respect to the antiholomorphic variables gives \begin{align*} \frac{\partial^2\rho_Q}{\partial z_1\partial\bar z_1}=\frac{\partial}{\partial\bar z_1}(\bar z_1)=1. \end{align*} \begin{align*} \frac{\partial^2\rho_Q}{\partial z_2\partial\bar z_2}=\frac{\partial}{\partial\bar z_2}(-\bar z_2)=-1. \end{align*} \begin{align*} \frac{\partial^2\rho_Q}{\partial w\partial\bar w}=\frac{\partial}{\partial\bar w}\left(\frac{i}{2}\right)=0. \end{align*} The remaining mixed second derivatives with distinct variables are $0$: $\bar z_1$ depends only on $\bar z_1$, $-\bar z_2$ depends only on $\bar z_2$, and $i/2$ is constant. Thus for \begin{align*} L=a_1\partial_{z_1}+a_2\partial_{z_2}+b\partial_w\in T_p^{1,0}M_Q, \end{align*} the Levi form is \begin{align*} \mathcal L_{\rho_Q,p}(L,\overline L)=1\cdot a_1\overline{a_1}+(-1)\cdot a_2\overline{a_2}+0\cdot b\overline b=|a_1|^2-|a_2|^2. \end{align*} The positive and negative values occur on actual complex tangent vectors. At $p=(p_1,p_2,p_w)\in M_Q$, the first derivatives above give \begin{align*} \frac{\partial\rho_Q}{\partial z_1}(p)=\bar p_1. \end{align*} \begin{align*} \frac{\partial\rho_Q}{\partial z_2}(p)=-\bar p_2. \end{align*} \begin{align*} \frac{\partial\rho_Q}{\partial w}(p)=\frac{i}{2}. \end{align*} Set \begin{align*} L_+=\partial_{z_1}+2i\bar p_1\partial_w. \end{align*} Then \begin{align*} L_+\rho_Q(p)=1\cdot\bar p_1+(2i\bar p_1)\cdot\frac{i}{2}. \end{align*} Since $(2i)(i/2)=i^2=-1$, \begin{align*} L_+\rho_Q(p)=\bar p_1-\bar p_1=0. \end{align*} Hence $L_+\in T_p^{1,0}M_Q$. In the formula $\mathcal L_{\rho_Q,p}(L,\overline L)=|a_1|^2-|a_2|^2$, this vector has $a_1=1$ and $a_2=0$, so \begin{align*} \mathcal L_{\rho_Q,p}(L_+,\overline{L_+})=|1|^2-|0|^2=1. \end{align*} Similarly, set \begin{align*} L_-=\partial_{z_2}-2i\bar p_2\partial_w. \end{align*} Then \begin{align*} L_-\rho_Q(p)=1\cdot(-\bar p_2)+(-2i\bar p_2)\cdot\frac{i}{2}. \end{align*} Since $(-2i)(i/2)=-i^2=1$, \begin{align*} L_-\rho_Q(p)=-\bar p_2+\bar p_2=0. \end{align*} Hence $L_-\in T_p^{1,0}M_Q$. In the same Levi-form formula this vector has $a_1=0$ and $a_2=1$, so \begin{align*} \mathcal L_{\rho_Q,p}(L_-,\overline{L_-})=|0|^2-|1|^2=-1. \end{align*} The sphere is the model with strictly positive Levi geometry, while this hyperquadric is the basic model where the Levi form has both positive and negative complex tangential directions. [/example] The course first treats hypersurfaces because the Levi form has a simple scalar-valued form there. Later, the same idea is reformulated for higher-codimension CR manifolds, where the Levi form takes values in a normal bundle. ## CR Manifolds and Tangential Cauchy-Riemann Equations Once the complex tangent distribution has been identified, the next problem is to express holomorphicity using only data intrinsic to the boundary. A [holomorphic function](/page/Holomorphic%20Function) restricts to a boundary function annihilated by all $(0,1)$ tangent vector fields. This leads to CR functions and to the operator $\bar{\partial}_b$. [definition: CR Function on a Hypersurface] Let $M \subset \mathbb C^n$ be a smooth real hypersurface. A smooth function $f:M\to\mathbb C$ is a CR function if \begin{align*} \overline{L}f = 0 \end{align*} for every smooth local section $L$ of $T^{1,0}M$. [/definition] A CR function is a boundary object satisfying the tangential part of the Cauchy-Riemann equations. The missing normal equation is precisely why extension is a serious question rather than part of the definition. Here $\Lambda_b^{0,q}$ denotes the bundle of tangential $(0,q)$-forms: at a point of $M$, its elements are alternating $q$-linear forms on the antiholomorphic tangent space $T^{0,1}_pM$. When an ambient antiholomorphic form is restricted to $M$, the tangential projection keeps only its values on vectors in $T^{0,1}M$ and discards conormal components. Thus projecting the ambient $\bar{\partial}$ operator means differentiating first and then retaining only this tangential antiholomorphic component. [definition: Tangential Cauchy-Riemann Operator] Let $M \subset \mathbb C^n$ be a smooth real hypersurface. For each $q \ge 0$, the tangential Cauchy-Riemann operator is the differential operator \begin{align*} \bar{\partial}_b:C^\infty(M;\Lambda_b^{0,q}) \longrightarrow C^\infty(M;\Lambda_b^{0,q+1}) \end{align*} obtained by projecting the ambient $\bar{\partial}$ operator onto tangential $(0,q+1)$-forms. In degree $0$, this gives \begin{align*} \bar{\partial}_b:C^\infty(M) \longrightarrow C^\infty(M;\Lambda_b^{0,1}). \end{align*} [/definition] The operator $\bar{\partial}_b$ is not elliptic in the usual sense, because it differentiates only in tangential complex directions. Its analytic behaviour depends on the Levi geometry, and this dependence is one of the recurring themes of the course. [example: Boundary Values of Holomorphic Functions] Let $\Omega\subset\mathbb C^n$ have smooth boundary $M$, let $F\in C^1(\overline\Omega)$ be holomorphic on $\Omega$, and set $f=F|_M$. We show that $f$ is a CR function on $M$. Fix $q\in M$, and let \begin{align*} L=\sum_{j=1}^n a_j\partial_{z_j} \end{align*} be a local section of $T^{1,0}M$ near $q$. Its conjugate tangential vector field is \begin{align*} \overline L=\sum_{j=1}^n \overline{a_j}\partial_{\bar z_j}. \end{align*} Because $\overline L$ is tangent to $M$, differentiating the restricted function $f=F|_M$ in the direction $\overline L$ agrees with differentiating the $C^1$ extension $F$ in that same tangential direction and then restricting to $M$. Hence \begin{align*} (\overline L f)(q)=\left(\sum_{j=1}^n \overline{a_j(q)}\partial_{\bar z_j}\right)F(q). \end{align*} By linearity of differentiation, \begin{align*} (\overline L f)(q)=\sum_{j=1}^n \overline{a_j(q)}\frac{\partial F}{\partial\bar z_j}(q). \end{align*} Since $F$ is holomorphic on $\Omega$, the ambient Cauchy-Riemann equations give, for every $x\in\Omega$ and every $1\le j\le n$, \begin{align*} \frac{\partial F}{\partial\bar z_j}(x)=0. \end{align*} Because $F\in C^1(\overline\Omega)$, each derivative $\partial F/\partial\bar z_j$ is continuous up to $M$. Therefore, for the boundary point $q\in M$, \begin{align*} \frac{\partial F}{\partial\bar z_j}(q)=\lim_{\Omega\ni x\to q}\frac{\partial F}{\partial\bar z_j}(x). \end{align*} The expression inside the limit is identically $0$ on $\Omega$, so \begin{align*} \frac{\partial F}{\partial\bar z_j}(q)=\lim_{\Omega\ni x\to q}0=0. \end{align*} Substituting this into the formula for the tangential derivative gives \begin{align*} (\overline L f)(q)=\sum_{j=1}^n \overline{a_j(q)}\cdot 0. \end{align*} Since each summand is $0$, \begin{align*} (\overline L f)(q)=0. \end{align*} This holds for every $q\in M$ and every local section $L$ of $T^{1,0}M$, so $f$ is a CR function on $M$. Boundary values of holomorphic functions therefore always satisfy the tangential Cauchy-Riemann equations; the extension problem asks when a CR function arises in this way. [/example] The intrinsic viewpoint is needed because many CR manifolds are not initially presented as boundaries. The course develops abstract CR structures after the embedded hypersurface case has supplied the motivating examples. ## Extension, Non-Extension, and the Lewy Phenomenon The central analytic problem is whether a CR function on a real hypersurface is the boundary value of a holomorphic function on one side, on both sides, or on neither side. In one complex variable, local boundary values are strongly constrained by the availability of holomorphic coordinates. In several variables, the Levi form creates asymmetric extension phenomena. [quotetheorem:9190] [citeproof:9190] The theorem is a prototype rather than the final extension result. It gives a one-sided conclusion, not a two-sided extension theorem, and it does not say that every smooth CR function extends when the Levi hypothesis is absent. For example, on Levi-flat models such as $\operatorname{Im} w=0$, the CR equations only impose holomorphicity along the complex leaves, so smooth dependence in the remaining real direction can obstruct holomorphic extension to a prescribed side. Thus the positive Levi eigenvalue is not a decorative assumption: it supplies the analytic-disc direction along which holomorphic information propagates. [example: Why the Side Matters] For the sphere $M_S=\{|z|^2=1\}$, take \begin{align*} \rho_S(z)=|z|^2-1=\sum_{m=1}^n z_m\bar z_m-1. \end{align*} The bounded side is $\{\rho_S<0\}$, because \begin{align*} \rho_S(z)<0 \Longleftrightarrow |z|^2-1<0 \Longleftrightarrow |z|^2<1. \end{align*} For each $j$ and $m$, treating $z_m$ and $\bar z_m$ as independent coordinates gives \begin{align*} \partial_{z_j}(z_m\bar z_m)=(\partial_{z_j}z_m)\bar z_m+z_m(\partial_{z_j}\bar z_m)=\delta_{jm}\bar z_m+z_m\cdot 0=\delta_{jm}\bar z_m. \end{align*} Therefore \begin{align*} \partial_{z_j}\rho_S=\partial_{z_j}\left(\sum_{m=1}^n z_m\bar z_m-1\right)=\sum_{m=1}^n\delta_{jm}\bar z_m-0=\bar z_j. \end{align*} Differentiating with respect to $\bar z_k$ gives \begin{align*} \frac{\partial^2\rho_S}{\partial z_j\partial\bar z_k}=\frac{\partial\bar z_j}{\partial\bar z_k}=\delta_{jk}. \end{align*} Let $p\in M_S$ and let \begin{align*} L=\sum_{j=1}^n a_j\partial_{z_j}\in T_p^{1,0}M_S. \end{align*} By the definition of the Levi form, \begin{align*} \mathcal L_{\rho_S,p}(L,\overline L)=\sum_{j,k=1}^n \frac{\partial^2\rho_S}{\partial z_j\partial\bar z_k}(p)a_j\overline{a_k}. \end{align*} Substituting $\frac{\partial^2\rho_S}{\partial z_j\partial\bar z_k}(p)=\delta_{jk}$ gives \begin{align*} \mathcal L_{\rho_S,p}(L,\overline L)=\sum_{j,k=1}^n\delta_{jk}a_j\overline{a_k}. \end{align*} When $j\ne k$, the factor $\delta_{jk}$ is $0$; when $j=k$, the factor $\delta_{jj}$ is $1$. Hence \begin{align*} \mathcal L_{\rho_S,p}(L,\overline L)=\sum_{j=1}^n a_j\overline{a_j}=\sum_{j=1}^n |a_j|^2. \end{align*} If $L\ne 0$, then at least one coefficient $a_j$ is nonzero, so the corresponding summand $|a_j|^2$ is positive and every other summand is nonnegative. Thus \begin{align*} \mathcal L_{\rho_S,p}(L,\overline L)>0. \end{align*} By the *[Lewy Extension Principle](/theorems/9190)*, smooth CR data near such a point extend to the side selected by this positive Levi geometry, here the bounded side $\{|z|^2<1\}$. The sign cannot be read from the real hypersurface alone. For \begin{align*} M_Q=\{(z_1,z_2,w)\in\mathbb C^3:\operatorname{Im}w=|z_1|^2-|z_2|^2\}, \end{align*} use the defining function \begin{align*} \rho_Q=z_1\bar z_1-z_2\bar z_2-\operatorname{Im}w=z_1\bar z_1-z_2\bar z_2-\frac{w-\bar w}{2i}. \end{align*} The holomorphic first derivatives are computed term by term: \begin{align*} \frac{\partial\rho_Q}{\partial z_1}=\partial_{z_1}(z_1\bar z_1)-\partial_{z_1}(z_2\bar z_2)-\partial_{z_1}\left(\frac{w-\bar w}{2i}\right)=\bar z_1-0-0=\bar z_1. \end{align*} \begin{align*} \frac{\partial\rho_Q}{\partial z_2}=\partial_{z_2}(z_1\bar z_1)-\partial_{z_2}(z_2\bar z_2)-\partial_{z_2}\left(\frac{w-\bar w}{2i}\right)=0-\bar z_2-0=-\bar z_2. \end{align*} \begin{align*} \frac{\partial\rho_Q}{\partial w}=\partial_w(z_1\bar z_1)-\partial_w(z_2\bar z_2)-\partial_w\left(\frac{w-\bar w}{2i}\right)=0-0-\frac{1}{2i}=\frac{i}{2}. \end{align*} Differentiating these first derivatives with respect to antiholomorphic variables gives \begin{align*} \frac{\partial^2\rho_Q}{\partial z_1\partial\bar z_1}=\frac{\partial}{\partial\bar z_1}(\bar z_1)=1. \end{align*} \begin{align*} \frac{\partial^2\rho_Q}{\partial z_2\partial\bar z_2}=\frac{\partial}{\partial\bar z_2}(-\bar z_2)=-1. \end{align*} \begin{align*} \frac{\partial^2\rho_Q}{\partial w\partial\bar w}=\frac{\partial}{\partial\bar w}\left(\frac{i}{2}\right)=0. \end{align*} The mixed second derivatives with distinct variables vanish: $\partial_{\bar z_2}(\bar z_1)=\partial_{\bar w}(\bar z_1)=0$, $\partial_{\bar z_1}(-\bar z_2)=\partial_{\bar w}(-\bar z_2)=0$, and every antiholomorphic derivative of the constant $i/2$ is $0$. Therefore, for \begin{align*} L=a_1\partial_{z_1}+a_2\partial_{z_2}+b\partial_w\in T_p^{1,0}M_Q, \end{align*} the Levi form is \begin{align*} \mathcal L_{\rho_Q,p}(L,\overline L)=1\cdot a_1\overline{a_1}+(-1)\cdot a_2\overline{a_2}+0\cdot b\overline b=|a_1|^2-|a_2|^2. \end{align*} At $p=(p_1,p_2,p_w)\in M_Q$, set \begin{align*} L_+=\partial_{z_1}+2i\bar p_1\partial_w. \end{align*} Using the first derivatives of $\rho_Q$ at $p$, \begin{align*} L_+\rho_Q(p)=1\cdot\bar p_1+(2i\bar p_1)\cdot\frac{i}{2}. \end{align*} Since $(2i)(i/2)=i^2=-1$, \begin{align*} L_+\rho_Q(p)=\bar p_1-\bar p_1=0. \end{align*} Thus $L_+\in T_p^{1,0}M_Q$. In the formula $\mathcal L_{\rho_Q,p}(L,\overline L)=|a_1|^2-|a_2|^2$, this vector has $a_1=1$ and $a_2=0$, so \begin{align*} \mathcal L_{\rho_Q,p}(L_+,\overline{L_+})=|1|^2-|0|^2=1. \end{align*} Similarly, set \begin{align*} L_-=\partial_{z_2}-2i\bar p_2\partial_w. \end{align*} Then \begin{align*} L_-\rho_Q(p)=1\cdot(-\bar p_2)+(-2i\bar p_2)\cdot\frac{i}{2}. \end{align*} Since $(-2i)(i/2)=-i^2=1$, \begin{align*} L_-\rho_Q(p)=-\bar p_2+\bar p_2=0. \end{align*} Thus $L_-\in T_p^{1,0}M_Q$. In the same Levi-form formula, this vector has $a_1=0$ and $a_2=1$, so \begin{align*} \mathcal L_{\rho_Q,p}(L_-,\overline{L_-})=|0|^2-|1|^2=-1. \end{align*} If the defining function is replaced by $-\rho_Q$, then \begin{align*} L(-\rho_Q)(p)=-L\rho_Q(p). \end{align*} Hence $L\rho_Q(p)=0$ if and only if $L(-\rho_Q)(p)=0$, so the same two vectors remain complex tangent. Also, \begin{align*} \frac{\partial^2(-\rho_Q)}{\partial z_j\partial\bar z_k}(p)=-\frac{\partial^2\rho_Q}{\partial z_j\partial\bar z_k}(p). \end{align*} Therefore \begin{align*} \mathcal L_{-\rho_Q,p}(L,\overline L)=-\mathcal L_{\rho_Q,p}(L,\overline L) \end{align*} for every $L\in T_p^{1,0}M_Q$. Applying this to the two tangent vectors above gives \begin{align*} \mathcal L_{-\rho_Q,p}(L_+,\overline{L_+})=-\mathcal L_{\rho_Q,p}(L_+,\overline{L_+})=-1. \end{align*} \begin{align*} \mathcal L_{-\rho_Q,p}(L_-,\overline{L_-})=-\mathcal L_{\rho_Q,p}(L_-,\overline{L_-})=1. \end{align*} Thus the side chosen by the defining function changes which complex directions have positive Levi geometry, so extension statements must record the orientation, the defining function, and the Levi signature rather than only the underlying real hypersurface. [/example] Non-extension results are just as important as extension theorems. Later lectures study examples where formal CR solvability fails, where weak pseudoconvexity produces loss of estimates, and where the tangential complex has nontrivial cohomology. ## The Analytical Complex on the Boundary The final organising problem is to study boundary differential forms, not only functions. The operator $\bar{\partial}_b$ extends to a complex of tangential $(0,q)$-forms, and its cohomology measures obstructions to solving boundary Cauchy-Riemann equations. [definition: Tangential Cauchy-Riemann Complex] Let $M \subset \mathbb C^n$ be a smooth CR hypersurface. The tangential Cauchy-Riemann complex is the sequence \begin{align*} 0 \longrightarrow C^\infty(M) \xrightarrow{\bar{\partial}_b} C^\infty(M;\Lambda_b^{0,1}) \xrightarrow{\bar{\partial}_b} C^\infty(M;\Lambda_b^{0,2}) \xrightarrow{\bar{\partial}_b} \cdots, \end{align*} where $\Lambda_b^{0,q}$ denotes the bundle of tangential $(0,q)$-forms. [/definition] This complex is the boundary analogue of the Dolbeault complex. The equation $\bar{\partial}_b u=f$ has the obvious compatibility condition $\bar{\partial}_b f=0$, but on a boundary this condition need not be sufficient. The obstruction is measured by closed tangential forms modulo those that already come from a previous potential. [definition: CR Cohomology] For a smooth CR hypersurface $M$, the $q$-th CR cohomology group is \begin{align*} H_b^{0,q}(M)=\frac{\ker(\bar{\partial}_b:C^\infty(M;\Lambda_b^{0,q})\to C^\infty(M;\Lambda_b^{0,q+1}))}{\operatorname{im}(\bar{\partial}_b:C^\infty(M;\Lambda_b^{0,q-1})\to C^\infty(M;\Lambda_b^{0,q}))}. \end{align*} [/definition] The cohomology groups record the failure of solving $\bar{\partial}_b u = f$. In favourable geometric settings they vanish or are finite-dimensional; in weakly pseudoconvex settings they can be large and sensitive to microlocal behaviour. [example: The First Cohomological Obstruction] Let $f\in C^\infty(M;\Lambda_b^{0,1})$ and assume that $\bar{\partial}_b f=0$. This means that applying the next differential in the tangential Cauchy-Riemann complex to $f$ gives the zero tangential $(0,2)$-form, so \begin{align*} f\in \ker(\bar{\partial}_b:C^\infty(M;\Lambda_b^{0,1})\to C^\infty(M;\Lambda_b^{0,2})). \end{align*} The first CR cohomology group is defined by quotienting this kernel by the image of the previous differential: \begin{align*} H_b^{0,1}(M)=\frac{\ker(\bar{\partial}_b:C^\infty(M;\Lambda_b^{0,1})\to C^\infty(M;\Lambda_b^{0,2}))}{\operatorname{im}(\bar{\partial}_b:C^\infty(M)\to C^\infty(M;\Lambda_b^{0,1}))}. \end{align*} Since $f$ lies in the numerator, it determines the coset \begin{align*} [f]=f+\operatorname{im}(\bar{\partial}_b:C^\infty(M)\to C^\infty(M;\Lambda_b^{0,1})). \end{align*} Thus \begin{align*} [f]\in H_b^{0,1}(M). \end{align*} Now suppose that the boundary equation $\bar{\partial}_b u=f$ has a smooth solution $u\in C^\infty(M)$. The map in degree $0$ is \begin{align*} \bar{\partial}_b:C^\infty(M)\to C^\infty(M;\Lambda_b^{0,1}). \end{align*} Since $\bar{\partial}_b u=f$, the form $f$ is an element of the image of this map: \begin{align*} f\in \operatorname{im}(\bar{\partial}_b:C^\infty(M)\to C^\infty(M;\Lambda_b^{0,1})). \end{align*} Let \begin{align*} I=\operatorname{im}(\bar{\partial}_b:C^\infty(M)\to C^\infty(M;\Lambda_b^{0,1})). \end{align*} Then $f\in I$, and the coset represented by $f$ is \begin{align*} f+I=I. \end{align*} The coset $I$ is the zero element of the quotient [vector space](/page/Vector%20Space), so \begin{align*} [f]=0\in H_b^{0,1}(M). \end{align*} Therefore solvability of $\bar{\partial}_b u=f$ forces the cohomology class of $f$ to vanish. Equivalently, if $[f]\ne 0$ in $H_b^{0,1}(M)$, then no smooth function $u$ can satisfy $\bar{\partial}_b u=f$; this is the first cohomological obstruction to solving the boundary Cauchy-Riemann equation. [/example] ## Prerequisites and Course Trajectory The course assumes one complex variable, the basic theory of smooth manifolds, vector bundles and differential forms, elementary functional analysis, and the first definitions from several complex variables. The most frequently used background facts are the Cauchy-Riemann equations in $\mathbb C^n$, the ambient $\bar{\partial}$ operator, plurisubharmonic functions, pseudoconvex domains, and the decomposition of complexified tangent and cotangent bundles into $(1,0)$ and $(0,1)$ parts. [remark: Background From Earlier Several Complex Variables] The course uses holomorphic functions on domains in $\mathbb C^n$, the ambient $\bar{\partial}$ operator, plurisubharmonic functions, pseudoconvex domains, and basic examples such as balls, polydiscs, and strongly pseudoconvex domains. The new ingredient is that the boundary itself becomes the main geometric and analytic object. [/remark] The lectures progress from embedded hypersurfaces to intrinsic CR manifolds, then to extension theorems and finally to the analysis of $\bar{\partial}_b$. This order is deliberate: the concrete Levi form supplies intuition, CR structures make the theory invariant, extension theorems explain the analytic stakes, and boundary complexes provide the PDE framework. [explanation: How to Read These Notes] The notes should be read with the boundary examples in mind. When a definition is given abstractly, it is useful to test it on the unit sphere, a hyperquadric, and the boundary of a pseudoconvex domain. When a theorem mentions a sign condition on the Levi form, the reader should ask which side of the hypersurface is being selected and which differential operator gains an estimate from that sign. Proofs in these notes are included when they are part of the course's mathematical development. Some outside results from elliptic regularity, functional analysis, or microlocal analysis will be used as background tools. Their role here is to support the boundary analysis rather than to replace the CR-geometric argument. [/explanation] With these tools in hand, the boundary is no longer just a place where holomorphic functions stop; it becomes a geometric object whose local structure controls analytic behavior. The next chapter makes that shift explicit by studying real hypersurfaces in $\mathbb C^n$ as the first natural setting where CR geometry appears. # 1. Real Hypersurfaces in $\mathbb C^n$ Real hypersurfaces are the first place where several complex variables sees the boundary as a geometric object rather than as a passive edge of a domain. In one complex variable every smooth real curve has the same local complex tangential dimension, but in $\mathbb C^n$ with $n \ge 2$ a real hypersurface carries a distinguished complex distribution and a second-order invariant, the Levi form. The prerequisites for this chapter are multivariable calculus over $\mathbb R^{2n}$, the decomposition of complex differentials into $\partial$ and $\bar\partial$ parts, and the [inverse function theorem](/theorems/51) for regular level sets. This chapter builds the basic language: defining functions, real and complex tangent spaces, the Levi form, and the model classes of Levi-flat, nondegenerate, and strictly pseudoconvex hypersurfaces. ## Defining Functions and Complex Tangencies The first question is how a real hypersurface sitting in $\mathbb C^n$ can remember the ambient complex structure. A smooth defining equation gives an ordinary real tangent hyperplane, but the complex geometry comes from asking which tangent vectors remain tangent after multiplication by $i$. [definition: Smooth Real Hypersurface] Let $M \subset \mathbb C^n$ be a subset and let $k \ge 1$. The set $M$ is a $C^k$ real hypersurface if for every $p \in M$ there are an open neighbourhood $U \subset \mathbb C^n$ of $p$ and a function $\rho \in C^k(U;\mathbb R)$ such that \begin{align*} M \cap U = \{z \in U : \rho(z)=0\}. \end{align*} The differential satisfies $d\rho_p \ne 0$. The function $\rho$ is a local defining function for $M$ near $p$. [/definition] The defining function packages the hypersurface as a regular level set, so its first derivative gives the tangent hyperplane. This leads to the real tangent space, which is the ambient real tangent space constrained by the first-order equation defining $M$. [definition: Real Tangent Space] Let $M \subset \mathbb C^n$ be a $C^1$ real hypersurface, let $p \in M$, and let $\rho$ be a local defining function near $p$. The real tangent space of $M$ at $p$ is \begin{align*} T_pM := \{v \in T_p\mathbb C^n : d\rho_p(v)=0\}. \end{align*} [/definition] This definition is independent of the chosen defining function because two defining functions for the same hypersurface differ by a nowhere-zero real factor along $M$. The next refinement asks which real tangent directions are compatible with the complex structure, since holomorphic curves can only move in directions stable under multiplication by $i$. [definition: Holomorphic Tangent Space] Let $M \subset \mathbb C^n$ be a $C^1$ real hypersurface and let $p \in M$. Let $J:T_p\mathbb C^n \to T_p\mathbb C^n$ denote the ambient complex structure, namely multiplication by $i$ on tangent vectors. The holomorphic tangent space at $p$ is the real vector space \begin{align*} H_p(M) := T_pM \cap J(T_pM). \end{align*} Its $(1,0)$ part is \begin{align*} T_{1,0,p}M := T_{1,0,p}\mathbb C^n \cap (\mathbb C \otimes_{\mathbb R} T_pM), \end{align*} and its $(0,1)$ part is $T_{0,1,p}M := \overline{T_{1,0,p}M}$. [/definition] The real space $H_p(M)$ has real dimension $2n-2$, while $T_{1,0,p}M$ has complex dimension $n-1$. For an [open set](/page/Open%20Set) $U \subset M$, we write $H(M)|_U=\bigsqcup_{p\in U}H_p(M)$ for the resulting real distribution. Thus a hypersurface has one missing real direction, the normal direction, but only one missing complex direction in the complexified tangent space. The sphere is the standard test case because its defining function has a simple gradient and its complex tangencies can be written explicitly. [example: Sphere Tangencies] For the unit sphere $S^{2n-1}=\{z\in\mathbb C^n:|z|^2=1\}$, use the defining function $\rho(z)=|z|^2-1=\sum_{j=1}^n z_j\overline z_j-1$. Fix $p\in S^{2n-1}$ and identify $T_p\mathbb C^n$ with $\mathbb C^n$. For a real tangent vector $v=(v_1,\dots,v_n)$, the derivative of the term $z_j\overline z_j$ in the direction $v_j$ is $v_j\overline{p_j}+p_j\overline{v_j}$, so \begin{align*} d\rho_p(v)=\sum_{j=1}^n v_j\overline{p_j}+\sum_{j=1}^n p_j\overline{v_j}. \end{align*} Since $\sum_{j=1}^n p_j\overline{v_j}$ is the complex conjugate of $\sum_{j=1}^n \overline{p_j}v_j$, this becomes \begin{align*} d\rho_p(v)=2\operatorname{Re}\sum_{j=1}^n \overline{p_j}v_j. \end{align*} Therefore \begin{align*} v\in T_pS^{2n-1} \end{align*} exactly when \begin{align*} \operatorname{Re}\sum_{j=1}^n \overline{p_j}v_j=0. \end{align*} To be holomorphically tangent, the real vector $v$ must satisfy both $v\in T_pS^{2n-1}$ and $iv\in T_pS^{2n-1}$. Put \begin{align*} \alpha=\sum_{j=1}^n \overline{p_j}v_j. \end{align*} The first condition gives \begin{align*} 0=d\rho_p(v)=2\operatorname{Re}\alpha. \end{align*} For the second condition, substitute $iv=(iv_1,\dots,iv_n)$ into the same formula: \begin{align*} d\rho_p(iv)=2\operatorname{Re}\sum_{j=1}^n \overline{p_j}(iv_j). \end{align*} Since $\sum_{j=1}^n \overline{p_j}(iv_j)=i\alpha$, we get \begin{align*} 0=d\rho_p(iv)=2\operatorname{Re}(i\alpha). \end{align*} Writing $\alpha=a+ib$ with $a,b\in\mathbb R$, one has $i\alpha=ia-b=-b+ia$, hence $\operatorname{Re}(i\alpha)=-b=-\operatorname{Im}\alpha$. Thus the two conditions are \begin{align*} \operatorname{Re}\alpha=0 \end{align*} and \begin{align*} \operatorname{Im}\alpha=0. \end{align*} So $\alpha=0$, meaning the holomorphic tangent directions are precisely those orthogonal to $p$ for the standard Hermitian product: \begin{align*} T_{1,0,p}S^{2n-1}=\left\{v\in\mathbb C^n:\sum_{j=1}^n \overline{p_j}v_j=0\right\}. \end{align*} The missing complex direction is the complex normal line $\mathbb Cp$, because \begin{align*} \sum_{j=1}^n \overline{p_j}p_j=\sum_{j=1}^n |p_j|^2=|p|^2=1. \end{align*} Thus $p$ itself is not complex tangent, and the complex tangent space consists exactly of the Hermitian orthogonal complement of the normal direction $p$. [/example] The sphere example is the reference model for curved boundaries. To measure curvature in complex tangential directions, we need a [second derivative](/page/Second%20Derivative) of the defining function that ignores directions transverse to $H_p(M)$. ## The Levi Form and Its Invariance The next question is which part of the second-order behaviour of $M$ is intrinsic to the CR geometry. A defining function changes under rescaling, and coordinates change under biholomorphism, so the useful second-order object must transform by a controlled non-zero factor. [definition: Levi Form] Let $M \subset \mathbb C^n$ be a $C^2$ real hypersurface, let $p \in M$, and let $\rho$ be a local defining function near $p$. The Levi form of $M$ at $p$ with respect to $\rho$ is the Hermitian form \begin{align*} \mathcal L_{\rho,p}: T_{1,0,p}M \times T_{1,0,p}M \to \mathbb C \end{align*} defined by \begin{align*} \mathcal L_{\rho,p}(Z,W) := \sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial \overline z_k}(p) Z_j \overline{W_k}, \end{align*} for $Z,W \in T_{1,0,p}M$. [/definition] The harmless convention-dependent factor is less important than the Hermitian matrix restricted to complex tangential directions. Since the definition used a particular defining function, the first invariance question is whether changing that function alters the geometric information carried by the Levi form. [quotetheorem:9191] [citeproof:9191] This lemma means that the rank and signature of the Levi form are intrinsic after the side of the hypersurface is fixed. The hypothesis that $\rho$ and $\widetilde\rho$ define the same hypersurface germ is essential: for $M=\{\operatorname{Im} w=0\}\subset\mathbb C^2$, replacing $\rho=\operatorname{Im}w$ by $\rho+|z|^2$ changes the zero set to the curved hypersurface $\{\operatorname{Im}w=-|z|^2\}$ and creates a nonzero Levi form. The sign of $a(p)$ records the choice of side; for the sphere, $\rho=|z|^2-1$ has positive Levi form on the ball side convention used here, while $-\rho$ reverses the sign. The restriction to $T_{1,0,p}M$ is also essential: for the plane $\{\operatorname{Im}w=0\}$, the defining functions $\rho=\operatorname{Im}w$ and $e^{\operatorname{Re}w}\rho$ have the same Levi form on complex tangential vectors, but their full mixed Hessians in the transverse $w$ direction differ at points with $\rho=0$. The next invariance question is stronger: a CR-geometric notion should also survive holomorphic coordinate changes, not only rescalings of the same defining equation. [quotetheorem:9192] [citeproof:9192] The theorem explains why the Levi form is the first CR invariant encountered in boundary theory. Holomorphicity is essential here: the real diffeomorphism $F(z,w)=(z,w+i|z|^2)$ sends the flat hypersurface $\{\operatorname{Im}w=0\}\subset\mathbb C^2$ to $\{\operatorname{Im}w=|z|^2\}$, changing the Levi form from zero to positive even though it is a smooth real coordinate change. Local invertibility is also essential; for instance, $F(z,w)=(z^2,w)$ is holomorphic but has rank-deficient differential at $(0,0)$, so it cannot identify the one-dimensional CR tangent spaces there. As in the defining-function lemma, the identity only concerns complex tangential directions; transverse second derivatives depend on the chosen defining function and coordinates. The Heisenberg hypersurface gives a computable non-compact model where this invariant has constant positive signature. [example: Heisenberg Hypersurface] In $\mathbb C^{n+1}$ write points as $(z',w)=(z_1,\dots,z_n,w)$, and consider \begin{align*} M=\{(z',w):\operatorname{Im}(w)=|z'|^2\}. \end{align*} Use the defining function \begin{align*} \rho(z',w)=|z'|^2-\operatorname{Im}(w)=\sum_{j=1}^n z_j\overline z_j-\frac{w-\overline w}{2i}. \end{align*} For $p=(p',p_w)\in M$ and a $(1,0)$ vector $Z=(Z_1,\dots,Z_n,Z_w)$, the complex tangency condition is $\partial\rho_p(Z)=0$. For $1\le j\le n$, \begin{align*} \frac{\partial}{\partial z_j}\left(\sum_{\ell=1}^n z_\ell\overline z_\ell-\frac{w-\overline w}{2i}\right)=\overline z_j. \end{align*} Also, \begin{align*} \frac{\partial}{\partial w}\left(\sum_{\ell=1}^n z_\ell\overline z_\ell-\frac{w-\overline w}{2i}\right)=-\frac{1}{2i}=\frac{i}{2}. \end{align*} Therefore \begin{align*} \partial\rho_p(Z)=\sum_{j=1}^n \overline{p_j}Z_j+\frac{i}{2}Z_w. \end{align*} Thus \begin{align*} T_{1,0,p}M=\left\{Z\in\mathbb C^{n+1}:\sum_{j=1}^n \overline{p_j}Z_j+\frac{i}{2}Z_w=0\right\}. \end{align*} The coefficient of $Z_w$ is $i/2\ne 0$, so this is the kernel of a nonzero complex linear functional on $\mathbb C^{n+1}$ and hence has complex dimension $n$. Now compute the mixed second derivatives. For $1\le j,k\le n$, \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}=\frac{\partial}{\partial\overline z_k}(\overline z_j)=\delta_{jk}. \end{align*} Since the $w$-part $-(w-\overline w)/(2i)$ is linear in $w$ and $\overline w$, differentiating it once leaves a constant and differentiating it twice gives zero. Hence \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\overline w}=0,\quad \frac{\partial^2\rho}{\partial w\partial\overline z_k}=0,\quad \frac{\partial^2\rho}{\partial w\partial\overline w}=0. \end{align*} For $Z,W\in T_{1,0,p}M$, the Levi form is therefore \begin{align*} \mathcal L_{\rho,p}(Z,W)=\sum_{j,k=1}^n \delta_{jk}Z_j\overline{W_k}=\sum_{j=1}^n Z_j\overline{W_j}. \end{align*} In particular, \begin{align*} \mathcal L_{\rho,p}(Z,Z)=\sum_{j=1}^n |Z_j|^2. \end{align*} Each term $|Z_j|^2$ is a nonnegative real number, so if $\mathcal L_{\rho,p}(Z,Z)=0$, then $Z_1=\cdots=Z_n=0$. Substituting this into the tangency equation gives \begin{align*} \frac{i}{2}Z_w=0. \end{align*} Multiplying by $2/i$ gives $Z_w=0$, so $Z=0$. Thus the Levi form is positive definite on the $n$-dimensional complex tangent space. The Heisenberg hypersurface is therefore the basic non-compact model with constant positive Levi signature, matching the sphere in stereographic-type CR coordinates. [/example] The Heisenberg hypersurface is the simplest model where the Levi form has constant positive signature. More general quadratic models allow both positive and negative Levi eigenvalues. ## Levi Classes and Quadratic Models Once the Levi form is defined, the natural classification problem is to distinguish hypersurfaces by the rank and signature of this Hermitian form. These classes govern the local behaviour of holomorphic functions, the possibility of extension across the boundary, and the PDE type of the tangential Cauchy-Riemann equations. [definition: Levi Flat Hypersurface] A $C^2$ real hypersurface $M \subset \mathbb C^n$ is Levi-flat if for every $p \in M$ the Levi form vanishes identically on $T_{1,0,p}M$. [/definition] Levi-flat hypersurfaces have no complex normal curvature. After the zero-curvature case, the next threshold is whether the Levi form has a kernel; this separates hypersurfaces with degenerate complex curvature from those with a genuine Hermitian signature. [definition: Levi Nondegenerate Hypersurface] A $C^2$ real hypersurface $M \subset \mathbb C^n$ is Levi-nondegenerate at $p \in M$ if the Levi form at $p$ is a nondegenerate Hermitian form on $T_{1,0,p}M$. [/definition] Nondegeneracy allows mixed signature as well as definite signature. The course now isolates the positive definite nondegenerate case because it is the condition under which boundaries behave like complex-convex barriers for holomorphic functions. [definition: Strictly Pseudoconvex Hypersurface] Let $M \subset \mathbb C^n$ be a $C^2$ real hypersurface and let $p \in M$. The hypersurface is strictly pseudoconvex at $p$ if, for a defining function whose negative side is the chosen domain side, the Levi form is positive definite on $T_{1,0,p}M$. [/definition] Strict pseudoconvexity is the boundary analogue of strict convexity, but only complex tangential directions are tested. This distinction is essential: real curvature in the missing normal direction is irrelevant for holomorphic boundary behaviour. Quadratic hypersurfaces display every possible Levi signature in a form that can be computed from the coefficients. [example: Real Quadrics] For integers $a,b \ge 0$ with $a+b=n$, define \begin{align*} Q(a,b)=\left\{(z',w)\in\mathbb C^{n+1}: \operatorname{Im}(w)=\sum_{j=1}^a |z_j|^2-\sum_{j=a+1}^{n}|z_j|^2\right\}. \end{align*} Use the defining function \begin{align*} \rho(z',w)=\sum_{j=1}^a z_j\overline z_j-\sum_{j=a+1}^{n}z_j\overline z_j-\frac{w-\overline w}{2i}. \end{align*} For $p=(p',p_w)\in Q(a,b)$ and $Z=(Z_1,\dots,Z_n,Z_w)\in T_{1,0,p}\mathbb C^{n+1}$, the complex tangency condition is $\partial\rho_p(Z)=0$. For $1\le j\le a$, \begin{align*} \frac{\partial}{\partial z_j}(z_j\overline z_j)=\overline z_j. \end{align*} For $a+1\le j\le n$, \begin{align*} \frac{\partial}{\partial z_j}(-z_j\overline z_j)=-\overline z_j. \end{align*} Also, \begin{align*} \frac{\partial}{\partial w}\left(-\frac{w-\overline w}{2i}\right)=-\frac{1}{2i}=\frac{i}{2}. \end{align*} Thus \begin{align*} \partial\rho_p(Z)=\sum_{j=1}^a \overline{p_j}Z_j-\sum_{j=a+1}^{n}\overline{p_j}Z_j+\frac{i}{2}Z_w. \end{align*} Therefore \begin{align*} T_{1,0,p}Q(a,b)=\left\{Z\in\mathbb C^{n+1}:\sum_{j=1}^a \overline{p_j}Z_j-\sum_{j=a+1}^{n}\overline{p_j}Z_j+\frac{i}{2}Z_w=0\right\}. \end{align*} Now compute the mixed second derivatives. If $1\le j,k\le a$, then \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}=\frac{\partial}{\partial\overline z_k}(\overline z_j)=\delta_{jk}. \end{align*} If $a+1\le j,k\le n$, then \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}=\frac{\partial}{\partial\overline z_k}(-\overline z_j)=-\delta_{jk}. \end{align*} If $j\le a<k$ or $k\le a<j$, then $\partial\rho/\partial z_j$ contains only $\overline z_j$, not $\overline z_k$, so \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}=0. \end{align*} The $w$-part $-(w-\overline w)/(2i)$ is linear in $w$ and $\overline w$, so every mixed second derivative involving $w$ or $\overline w$ is zero. Hence, for $Z,W\in T_{1,0,p}Q(a,b)$, \begin{align*} \mathcal L_{\rho,p}(Z,W)=\sum_{j=1}^a Z_j\overline{W_j}-\sum_{j=a+1}^{n}Z_j\overline{W_j}. \end{align*} The projection $Z\mapsto (Z_1,\dots,Z_n)$ identifies $T_{1,0,p}Q(a,b)$ with $\mathbb C^n$. Indeed, once $Z_1,\dots,Z_n$ are chosen, the tangency equation gives \begin{align*} \frac{i}{2}Z_w=-\sum_{j=1}^a \overline{p_j}Z_j+\sum_{j=a+1}^{n}\overline{p_j}Z_j. \end{align*} Multiplying by $2/i=-2i$ gives \begin{align*} Z_w=2i\left(\sum_{j=1}^a \overline{p_j}Z_j-\sum_{j=a+1}^{n}\overline{p_j}Z_j\right). \end{align*} Conversely, if $Z_1=\cdots=Z_n=0$, then the tangency equation becomes \begin{align*} \frac{i}{2}Z_w=0. \end{align*} Multiplying by $2/i$ gives $Z_w=0$, so the projection has zero kernel. Under this identification, writing $\xi_j=Z_j$, the Levi form on diagonal entries is \begin{align*} \mathcal L_{\rho,p}(Z,Z)=\sum_{j=1}^a |\xi_j|^2-\sum_{j=a+1}^{n}|\xi_j|^2. \end{align*} The coordinate vectors $e_1,\dots,e_a$ give positive values $1$, and the coordinate vectors $e_{a+1},\dots,e_n$ give negative values $-1$. Hence the Levi signature is $(a,b)$: it has $a$ positive and $b$ negative directions. Thus $Q(n,0)$ is strictly pseudoconvex, while $Q(a,b)$ with $a,b>0$ is Levi-nondegenerate of mixed signature. [/example] Quadrics are local normal models for Levi-nondegenerate hypersurfaces. When the Levi form vanishes identically, this model computation no longer gives positive or negative curvature directions, so it is not clear whether flatness is merely a second-order degeneracy or reflects genuine complex geometry inside the boundary. The missing test is local existence of complex leaves, not another signature computation. A precise integrability result is needed to turn vanishing of the Levi form into actual complex hypersurfaces inside the boundary, and to distinguish Levi-flatness from accidental pointwise degeneracy. [quotetheorem:9193] [citeproof:9193] This result is the first link between the Levi form and actual complex submanifolds inside the boundary. The smoothness assumption is not a cosmetic detail: the proof uses an integrability theorem for the CR distribution, and low regularity versions require additional hypotheses. The conclusion is local, so it does not assert that the whole hypersurface is globally a product or that the complex leaves fit together without monodromy; a Levi-flat hypersurface with compact transverse direction may have leaves returning to the same coordinate patch in a non-product way. Nor does it assert holomorphic flattening of $M$ to $\{\operatorname{Im}z_n=0\}$ in the smooth category; such a holomorphic first integral is an additional feature, available under stronger hypotheses such as real analyticity. The Levi-flat hypothesis itself cannot be dropped: the unit sphere has positive Levi form and contains no local complex hypersurface through a boundary point tangent to $H_p(M)$. Levi-flatness also does not mean the ambient domain is flat in any real Euclidean sense; it only says that the complex tangential second-order curvature vanishes. Product domains illustrate how Levi-flat pieces can occur even on important pseudoconvex examples. [example: Boundary of the Bidisc] The bidisc $\mathbb D^2 \subset \mathbb C^2$ is \begin{align*} \mathbb D^2=\{(z,w): |z|<1,\ |w|<1\}. \end{align*} A point belongs to $\partial\mathbb D^2$ exactly when at least one coordinate has modulus $1$ and neither coordinate has modulus greater than $1$. Hence \begin{align*} \partial\mathbb D^2 = (\partial\mathbb D \times \overline{\mathbb D}) \cup (\overline{\mathbb D}\times \partial\mathbb D). \end{align*} Consider a point $(p,q)\in \partial\mathbb D\times \mathbb D$, so $|p|=1$ and $|q|<1$. In a neighbourhood of $(p,q)$, the second coordinate remains strictly inside $\mathbb D$, so the relevant smooth boundary face is defined by \begin{align*} \rho(z,w)=|z|^2-1=z\overline z-1. \end{align*} For a $(1,0)$ vector $Z=(Z_z,Z_w)$, the complex tangency condition is $\partial\rho_{(p,q)}(Z)=0$. Since \begin{align*} \frac{\partial\rho}{\partial z}=\overline z \end{align*} and \begin{align*} \frac{\partial\rho}{\partial w}=0, \end{align*} evaluation at $(p,q)$ gives \begin{align*} \partial\rho_{(p,q)}(Z)=\overline p Z_z+0\cdot Z_w=\overline p Z_z. \end{align*} Thus $Z$ is complex tangent exactly when \begin{align*} \overline p Z_z=0. \end{align*} Because $|p|=1$, we have $p\ne 0$ and therefore $\overline p\ne 0$, so the last equation is equivalent to $Z_z=0$. Hence \begin{align*} T_{1,0,(p,q)}(\partial\mathbb D\times \mathbb D)=\{(0,Z_w):Z_w\in\mathbb C\}. \end{align*} The mixed second derivatives of $\rho$ are obtained from $\partial\rho/\partial z=\overline z$ and $\partial\rho/\partial w=0$: \begin{align*} \frac{\partial^2\rho}{\partial z\partial\overline z}=1,\quad \frac{\partial^2\rho}{\partial z\partial\overline w}=0,\quad \frac{\partial^2\rho}{\partial w\partial\overline z}=0,\quad \frac{\partial^2\rho}{\partial w\partial\overline w}=0. \end{align*} For tangent vectors $Z=(0,Z_w)$ and $W=(0,W_w)$, the Levi form is \begin{align*} \mathcal L_{\rho,(p,q)}(Z,W)=1\cdot Z_z\overline{W_z}+0\cdot Z_z\overline{W_w}+0\cdot Z_w\overline{W_z}+0\cdot Z_w\overline{W_w}. \end{align*} Substituting $Z_z=0$ and $W_z=0$ gives \begin{align*} \mathcal L_{\rho,(p,q)}(Z,W)=1\cdot 0\cdot \overline{0}+0\cdot 0\cdot\overline{W_w}+0\cdot Z_w\cdot\overline{0}+0\cdot Z_w\overline{W_w}=0. \end{align*} So the face $\partial\mathbb D\times\mathbb D$ has identically vanishing Levi form. The other smooth face is computed with the coordinates reversed. If $(p,q)\in \mathbb D\times\partial\mathbb D$, then $|p|<1$ and $|q|=1$, and a local defining function is \begin{align*} \sigma(z,w)=|w|^2-1=w\overline w-1. \end{align*} For $Z=(Z_z,Z_w)$, \begin{align*} \frac{\partial\sigma}{\partial z}=0 \end{align*} and \begin{align*} \frac{\partial\sigma}{\partial w}=\overline w, \end{align*} so \begin{align*} \partial\sigma_{(p,q)}(Z)=0\cdot Z_z+\overline q Z_w=\overline q Z_w. \end{align*} Since $|q|=1$, the equation $\overline q Z_w=0$ is equivalent to $Z_w=0$, and hence \begin{align*} T_{1,0,(p,q)}(\mathbb D\times\partial\mathbb D)=\{(Z_z,0):Z_z\in\mathbb C\}. \end{align*} The only nonzero mixed second derivative of $\sigma$ is \begin{align*} \frac{\partial^2\sigma}{\partial w\partial\overline w}=1. \end{align*} Therefore, for tangent vectors $Z=(Z_z,0)$ and $W=(W_z,0)$, \begin{align*} \mathcal L_{\sigma,(p,q)}(Z,W)=0\cdot Z_z\overline{W_z}+0\cdot Z_z\overline{W_w}+0\cdot Z_w\overline{W_z}+1\cdot Z_w\overline{W_w}=0. \end{align*} Thus each smooth boundary face is Levi-flat. The corner torus $\partial\mathbb D\times\partial\mathbb D$ is not a smooth real hypersurface, because both independent equations $|z|^2-1=0$ and $|w|^2-1=0$ hold there. This is why product domains can have Levi-flat boundary pieces even though their full boundary may have corners. [/example] The bidisc also warns that pseudoconvex domains need not have strictly pseudoconvex smooth boundary everywhere. Boundary regularity and corner behaviour will return later when the $\bar\partial$-Neumann problem is discussed. ## Strict Pseudoconvexity and Local Convexity The final question of the chapter is how the Levi form relates to the geometric shape of a domain. In real convexity, positive second fundamental form means that the boundary bends away from supporting hyperplanes; in complex analysis, strict pseudoconvexity means the boundary bends away from complex tangent hyperplanes. [quotetheorem:9194] The course uses this result as an organizing principle rather than proving it here. The $C^2$ boundary hypothesis is needed because the statement reads off the quadratic Taylor part of a defining function; for a merely $C^1$ boundary such as $r(z,w)=\operatorname{Re}w+|z|^{3/2}$ near the origin in $\mathbb C^2$, the second-order Levi expression is not defined at the boundary point. The phrase "second-order" is a limitation: the theorem gives a local normal form at one boundary point, not a global convex embedding of the whole domain. Strict positivity is also necessary, since the weakly pseudoconvex model $r(z,w)=2\operatorname{Re}w+|z|^4$ has zero Levi form at the origin and cannot be converted into a positive Hermitian quadratic model by a biholomorphism. This is the precise invariant content behind the slogan that strict pseudoconvexity is the biholomorphic substitute for strict convexity; ordinary real convexity itself is not biholomorphically invariant. The unit sphere is the model where the real and complex convexity pictures coincide. [example: Sphere as the Convex Model] For the unit ball $\mathbb B^n=\{z\in\mathbb C^n: |z|<1\}$, the boundary is $S^{2n-1}=\{z:|z|^2=1\}$. Use the defining function \begin{align*} \rho(z)=|z|^2-1=\sum_{j=1}^n z_j\overline z_j-1. \end{align*} Fix $p\in S^{2n-1}$. For a $(1,0)$ vector $Z=(Z_1,\dots,Z_n)$, the complex tangency condition is $\partial\rho_p(Z)=0$. Since \begin{align*} \frac{\partial}{\partial z_j}\left(\sum_{\ell=1}^n z_\ell\overline z_\ell-1\right)=\overline z_j, \end{align*} evaluation at $p$ gives \begin{align*} \partial\rho_p(Z)=\sum_{j=1}^n \overline{p_j}Z_j. \end{align*} Thus \begin{align*} T_{1,0,p}S^{2n-1}=\left\{Z\in\mathbb C^n:\sum_{j=1}^n \overline{p_j}Z_j=0\right\}. \end{align*} Now compute the mixed second derivatives. From $\partial\rho/\partial z_j=\overline z_j$, differentiating with respect to $\overline z_k$ gives \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}=\frac{\partial\overline z_j}{\partial\overline z_k}=\delta_{jk}. \end{align*} Therefore, for $Z,W\in T_{1,0,p}S^{2n-1}$, the Levi form is \begin{align*} \mathcal L_{\rho,p}(Z,W)=\sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}(p)Z_j\overline{W_k}. \end{align*} Substituting $\partial^2\rho/(\partial z_j\partial\overline z_k)=\delta_{jk}$ gives \begin{align*} \mathcal L_{\rho,p}(Z,W)=\sum_{j,k=1}^n \delta_{jk}Z_j\overline{W_k}. \end{align*} The terms with $j\ne k$ vanish because $\delta_{jk}=0$, and the terms with $j=k$ have $\delta_{jj}=1$, so \begin{align*} \mathcal L_{\rho,p}(Z,W)=\sum_{j=1}^n Z_j\overline{W_j}. \end{align*} In particular, \begin{align*} \mathcal L_{\rho,p}(Z,Z)=\sum_{j=1}^n Z_j\overline{Z_j}. \end{align*} Since $Z_j\overline{Z_j}=|Z_j|^2$, this becomes \begin{align*} \mathcal L_{\rho,p}(Z,Z)=\sum_{j=1}^n |Z_j|^2. \end{align*} Each term $|Z_j|^2$ is nonnegative, and the sum is zero only when $Z_1=\cdots=Z_n=0$. Hence $\mathcal L_{\rho,p}(Z,Z)>0$ for every nonzero complex tangent vector $Z$, so the sphere is strictly pseudoconvex with respect to the ball side. The same defining function also shows ordinary Euclidean strict convexity. Identify $\mathbb C^n$ with $\mathbb R^{2n}$ by writing $z_j=x_j+iy_j$ and listing the real coordinates as $x=(x_1,y_1,\dots,x_n,y_n)$. Then \begin{align*} |z|^2=\sum_{j=1}^n |z_j|^2=\sum_{j=1}^n (x_j^2+y_j^2). \end{align*} Equivalently, after relabelling the real coordinates as $x_1,\dots,x_{2n}$, \begin{align*} \rho(x)=x_1^2+\cdots+x_{2n}^2-1. \end{align*} For $1\le r,s\le 2n$, the first derivatives are \begin{align*} \frac{\partial\rho}{\partial x_r}=2x_r. \end{align*} Differentiating once more gives \begin{align*} \frac{\partial^2\rho}{\partial x_r\partial x_s}=2\delta_{rs}. \end{align*} Thus for a real vector $v=(v_1,\dots,v_{2n})$, \begin{align*} D^2\rho_p(v,v)=\sum_{r,s=1}^{2n}2\delta_{rs}v_rv_s. \end{align*} Again the off-diagonal terms vanish and the diagonal terms remain, so \begin{align*} D^2\rho_p(v,v)=2\sum_{r=1}^{2n}v_r^2=2|v|^2. \end{align*} For every nonzero real tangent vector $v\in T_pS^{2n-1}$, one has $|v|^2>0$, hence $D^2\rho_p(v,v)>0$. Thus the unit sphere is the model case where real strict convexity and strict pseudoconvexity coincide. [/example] The main point is that strict pseudoconvexity is invariant under biholomorphism, while ordinary real convexity is not. Later chapters exploit this invariant convexity through CR functions, extension theorems, and estimates for the tangential $\bar\partial_b$ operator. Having seen how strict pseudoconvexity is detected on a hypersurface, we now strip away the ambient coordinates and ask what part of that structure is intrinsic. CR manifolds provide the abstract language for carrying the same tangential complex geometry without relying on an embedding. # 2. CR Manifolds — Intrinsic Theory This chapter turns from hypersurfaces already sitting inside complex Euclidean space to CR manifolds as intrinsic objects. The guiding question is how much of the boundary geometry from Chapter 1 can be recognized without first choosing an ambient complex manifold. We isolate the complex tangent directions, formulate their integrability, and then compare smooth and real-analytic CR structures through the local embeddability problem. ## Abstract CR Structures The first problem is to describe the tangential Cauchy-Riemann directions using only the smooth manifold $M$, rather than a defining function in some $\mathbb C^N$. On an embedded hypersurface these directions are the $(1,0)$ tangent vectors lying in the complexified tangent bundle of $M$; intrinsically we take this subbundle as part of the structure and ask for the bracket closure that makes CR functions behave like boundary values of holomorphic functions. [definition: Abstract CR Manifold] Let $M$ be a smooth real manifold of dimension $2n+k$. An abstract CR structure of CR dimension $n$ and CR codimension $k$ on $M$ is a smooth complex vector subbundle \begin{align*} T^{1,0}M \subset T_{\mathbb C}M := TM \otimes_{\mathbb R} \mathbb C \end{align*} of complex rank $n$ such that, with $T^{0,1}M := \overline{T^{1,0}M}$, \begin{align*} T^{1,0}M \cap T^{0,1}M = \{0\}, \qquad [\Gamma(T^{1,0}M),\Gamma(T^{1,0}M)] \subset \Gamma(T^{1,0}M). \end{align*} [/definition] The first condition says that the chosen complex directions do not contain real tangent vectors after complexification. The second condition is formal integrability: it is the Frobenius-type condition needed for the tangential CR equations to form a differential system closed under commutators. To connect this complex subbundle back to real geometry, we now extract the real distribution on which the intrinsic complex structure acts. [definition: Holomorphic Tangent Bundle] For an abstract CR manifold $(M,T^{1,0}M)$, the holomorphic tangent bundle is the real subbundle \begin{align*} H(M) := \operatorname{Re}\bigl(T^{1,0}M \oplus T^{0,1}M\bigr) \subset TM. \end{align*} [/definition] The bundle $H(M)$ is the distribution of directions in which the CR geometry lives. It has real rank $2n$, and multiplication by $i$ on $T^{1,0}M$ induces a bundle map \begin{align*} J:H(M)\to H(M) \end{align*} with $J^2=-\operatorname{id}_{H(M)}$. The remaining $k$ real directions are transverse to the complex tangent distribution and account for the CR codimension. [example: Sphere CR Structure] Let $S^{2n-1}=\{z\in\mathbb C^n:|z|^2=1\}$ and set \begin{align*} \rho(z)=|z|^2-1=\sum_{\ell=1}^n z_\ell\bar z_\ell-1. \end{align*} A $(1,0)$ vector at $z$ has the form $L_v=\sum_{j=1}^n v_j\partial/\partial z_j$. It is tangent to the level set $\rho=0$ precisely when $L_v\rho=0$. Using the Wirtinger rule $\partial\bar z_\ell/\partial z_j=0$, we have \begin{align*} \frac{\partial}{\partial z_j}\left(\sum_{\ell=1}^n z_\ell\bar z_\ell-1\right)=\sum_{\ell=1}^n \delta_{j\ell}\bar z_\ell=\bar z_j. \end{align*} Therefore \begin{align*} L_v\rho=\sum_{j=1}^n v_j\frac{\partial \rho}{\partial z_j}=\sum_{j=1}^n v_j\bar z_j. \end{align*} Thus the induced CR bundle is \begin{align*} T^{1,0}_zS^{2n-1}=\left\{v\in\mathbb C^n:\sum_{j=1}^n \bar z_jv_j=0\right\}. \end{align*} For each $z\in S^{2n-1}$, the linear functional $v\mapsto \sum_{j=1}^n\bar z_jv_j$ is nonzero, since at $v=z$ it has value \begin{align*} \sum_{j=1}^n\bar z_jz_j=\sum_{j=1}^n |z_j|^2=|z|^2=1. \end{align*} A nonzero complex linear functional on $\mathbb C^n$ has kernel of complex dimension $n-1$, so $\operatorname{rank}_{\mathbb C}T^{1,0}S^{2n-1}=n-1$. Since \begin{align*} \dim_{\mathbb R}S^{2n-1}=2n-1=2(n-1)+1, \end{align*} the sphere has CR dimension $n-1$ and CR codimension $1$. It remains to verify formal integrability. Let \begin{align*} L=\sum_{j=1}^n a_j\frac{\partial}{\partial z_j} \end{align*} and \begin{align*} K=\sum_{j=1}^n b_j\frac{\partial}{\partial z_j} \end{align*} be local sections of $T^{1,0}S^{2n-1}$, extended as ambient $(1,0)$ vector fields near the sphere. Applying $LK$ and $KL$ to a [test function](/page/Test%20Function) $f$ gives \begin{align*} LKf=\sum_{j=1}^n\sum_{m=1}^n a_j\frac{\partial b_m}{\partial z_j}\frac{\partial f}{\partial z_m}+\sum_{j=1}^n\sum_{m=1}^n a_jb_m\frac{\partial^2 f}{\partial z_j\partial z_m}. \end{align*} Similarly, \begin{align*} KLf=\sum_{j=1}^n\sum_{m=1}^n b_j\frac{\partial a_m}{\partial z_j}\frac{\partial f}{\partial z_m}+\sum_{j=1}^n\sum_{m=1}^n b_ja_m\frac{\partial^2 f}{\partial z_j\partial z_m}. \end{align*} The second-derivative terms cancel because mixed Wirtinger derivatives commute and $a_jb_m=b_ma_j$, so \begin{align*} [L,K]=\sum_{m=1}^n\left(\sum_{j=1}^n a_j\frac{\partial b_m}{\partial z_j}-\sum_{j=1}^n b_j\frac{\partial a_m}{\partial z_j}\right)\frac{\partial}{\partial z_m}. \end{align*} Thus $[L,K]$ is still of type $(1,0)$. Because $L$ and $K$ are sections of $T^{1,0}S^{2n-1}$, their tangency condition says \begin{align*} L\rho=0 \end{align*} on $S^{2n-1}$ and \begin{align*} K\rho=0 \end{align*} on $S^{2n-1}$. If a smooth function vanishes on $S^{2n-1}$, every complexified tangent vector to $S^{2n-1}$ annihilates its restriction there; hence, at each point of the sphere, \begin{align*} L(K\rho)=0 \end{align*} and \begin{align*} K(L\rho)=0. \end{align*} Therefore \begin{align*} [L,K]\rho=L(K\rho)-K(L\rho)=0-0=0 \end{align*} on $S^{2n-1}$. The bracket is both of type $(1,0)$ and tangent to the sphere, so it is again a section of $T^{1,0}S^{2n-1}$. The displayed bundle therefore defines the standard hypersurface-type CR structure on the sphere. [/example] The sphere illustrates the model case: the intrinsic bundle is not an arbitrary distribution, but the remnant of ambient holomorphic directions. For a general embedded real submanifold, however, the intersection with ambient holomorphic tangent directions could change rank or fail to be closed under brackets. The next issue is to identify a clean geometric hypothesis that makes the inherited tangential Cauchy-Riemann equations constant-rank and involutive. Genericity supplies exactly that missing transversality, so the embedded construction produces an intrinsic CR structure rather than a pointwise collection of complex tangent vectors. [quotetheorem:9195] [citeproof:9195] Genericity is the hypothesis that prevents the intersection from changing size. For instance, a complex curve in $\mathbb C^2$ is not a generic real submanifold: its real tangent space is already invariant under multiplication by $i$, so $T_pM+iT_pM$ is only the tangent space of the curve, not all of $T_p\mathbb C^2$. The theorem therefore does not say that every embedded real submanifold has the CR structure expected of a boundary; it says that the submanifolds with enough real transverse directions inherit a constant-rank tangential Cauchy-Riemann system. Conversely, it also does not imply that an arbitrary abstract CR manifold has an ambient realization. That question is precisely the local embeddability problem addressed at the end of the chapter, where the same bracket-closed PDE system may or may not have enough CR functions to provide coordinates. ## CR Dimension, Codimension, and Hypersurface Type Once the CR bundle has been specified, the next question is how to measure its size and how close the structure is to a real hypersurface. The two integers $n$ and $k$ control the number of complex tangential directions and the number of missing real directions, and they determine which phenomena resemble hypersurface theory. [definition: CR Dimension and CR Codimension] Let $(M,T^{1,0}M)$ be an abstract CR manifold with $\operatorname{rank}_{\mathbb C}T^{1,0}M=n$ and $\dim_{\mathbb R}M=2n+k$. The integer $n$ is the CR dimension of $M$, and the integer $k$ is the CR codimension of $M$. [/definition] For an embedded generic submanifold $M\subset \mathbb C^N$, the CR codimension agrees with the real codimension of $M$ in $\mathbb C^N$. This comparison motivates a separate definition for the case with one transverse real direction, because that is the intrinsic replacement for a real hypersurface. [definition: Hypersurface-Type CR Manifold] A CR manifold $(M,T^{1,0}M)$ is of hypersurface type if its CR codimension is $1$, so \begin{align*} \dim_{\mathbb R}M = 2n+1. \end{align*} [/definition] In hypersurface type there is a single transverse real direction. This is the setting where the Levi form, pseudoconvexity, and extension phenomena most closely parallel the boundary theory from Chapter 1. [example: Graph Hypersurfaces] Let $\varphi \in C^\infty(\mathbb C^n \times \mathbb R;\mathbb R)$, write $z_{n+1}=s+it$, and set \begin{align*} \rho(z',z_{n+1})=t-\varphi(z',s). \end{align*} Then \begin{align*} M=\rho^{-1}(0)=\{(z',s+it):t=\varphi(z',s)\}. \end{align*} Equivalently, $M$ is parametrized by \begin{align*} \Phi(z',s)=(z',s+i\varphi(z',s)). \end{align*} The parameters $(z',s)\in \mathbb C^n\times \mathbb R$ have real dimension $2n+1$, and the first $n$ complex coordinates together with the real part $s=\operatorname{Re} z_{n+1}$ recover the parameters from $\Phi(z',s)$. Thus $\Phi$ is an injective immersion onto $M$, so $M$ is a smooth real hypersurface in $\mathbb C^{n+1}$ of real dimension $2n+1$. The same conclusion is visible from the defining function. Since \begin{align*} d\rho=dt-d\varphi, \end{align*} the coefficient of $dt$ in $d\rho$ is $1$, so $d\rho$ is nonzero at every point. Therefore $M$ has one real transverse direction, hence it is of hypersurface type. An ambient $(1,0)$ vector has the form \begin{align*} L=\sum_{j=1}^n a_j\frac{\partial}{\partial z_j}+a_{n+1}\frac{\partial}{\partial z_{n+1}}. \end{align*} For $1\le j\le n$, the variables $s$ and $t$ do not depend on $z_j$, so \begin{align*} \frac{\partial \rho}{\partial z_j}=\frac{\partial t}{\partial z_j}-\frac{\partial}{\partial z_j}\varphi(z',s)=0-\varphi_{z_j}=-\varphi_{z_j}. \end{align*} For the last coordinate, use \begin{align*} s=\frac{z_{n+1}+\bar z_{n+1}}{2}. \end{align*} Hence \begin{align*} \frac{\partial s}{\partial z_{n+1}}=\frac{1}{2}. \end{align*} Also \begin{align*} t=\frac{z_{n+1}-\bar z_{n+1}}{2i}. \end{align*} Therefore \begin{align*} \frac{\partial t}{\partial z_{n+1}}=\frac{1}{2i}. \end{align*} Applying the chain rule to $\rho=t-\varphi(z',s)$ gives \begin{align*} \frac{\partial \rho}{\partial z_{n+1}}=\frac{\partial t}{\partial z_{n+1}}-\varphi_s\frac{\partial s}{\partial z_{n+1}}=\frac{1}{2i}-\frac{1}{2}\varphi_s. \end{align*} Thus \begin{align*} L\rho=\sum_{j=1}^n a_j\frac{\partial \rho}{\partial z_j}+a_{n+1}\frac{\partial \rho}{\partial z_{n+1}}. \end{align*} Substituting the derivatives just computed yields \begin{align*} L\rho=-\sum_{j=1}^n a_j\varphi_{z_j}+a_{n+1}\left(\frac{1}{2i}-\frac{1}{2}\varphi_s\right). \end{align*} A $(1,0)$ vector is tangent to the level set $\rho=0$ exactly when $L\rho=0$ along $M$, so the induced CR bundle is \begin{align*} T^{1,0}M=\left\{L\in T^{1,0}\mathbb C^{n+1}: -\sum_{j=1}^n a_j\varphi_{z_j}+a_{n+1}\left(\frac{1}{2i}-\frac{1}{2}\varphi_s\right)=0\right\}. \end{align*} Finally, \begin{align*} \frac{1}{2i}-\frac{1}{2}\varphi_s=-\frac{i}{2}-\frac{\varphi_s}{2}=-\frac{i+\varphi_s}{2}. \end{align*} Because $\varphi_s$ is real, the complex number $i+\varphi_s$ has imaginary part $1$, so it is never zero. Hence the tangency equation uniquely determines $a_{n+1}$ from the freely chosen coefficients $a_1,\dots,a_n$, and $T^{1,0}M$ has complex rank $n$. The graph presentation therefore gives a concrete hypersurface-type CR manifold whose complex tangential directions are exactly the ambient $(1,0)$ directions annihilating the defining function. [/example] Graphs are useful because they let local CR geometry be written in coordinates while retaining the intrinsic interpretation. A second model, the Heisenberg group, packages the same hypersurface-type structure with a noncommutative group law. [example: Heisenberg Group CR Structure] Identify $H^n$ with $\mathbb C^n\times \mathbb R$ and coordinates $(z,t)=(z_1,\dots,z_n,t)$. Define \begin{align*} L_j=\frac{\partial}{\partial z_j}+i\bar z_j\frac{\partial}{\partial t} \end{align*} for $1\le j\le n$, and let $T^{1,0}H^n$ be their complex span. Since complex conjugation sends $\partial/\partial z_j$ to $\partial/\partial \bar z_j$ and sends $i\bar z_j\partial_t$ to $-iz_j\partial_t$, the conjugate bundle $T^{0,1}H^n$ is spanned by \begin{align*} \bar L_j=\frac{\partial}{\partial \bar z_j}-iz_j\frac{\partial}{\partial t}. \end{align*} We first check that the two bundles have zero intersection. Suppose \begin{align*} \sum_{j=1}^n a_jL_j=\sum_{j=1}^n b_j\bar L_j. \end{align*} Substituting the definitions gives \begin{align*} \sum_{j=1}^n a_j\frac{\partial}{\partial z_j}+\left(\sum_{j=1}^n ia_j\bar z_j\right)\frac{\partial}{\partial t}=\sum_{j=1}^n b_j\frac{\partial}{\partial \bar z_j}-\left(\sum_{j=1}^n ib_jz_j\right)\frac{\partial}{\partial t}. \end{align*} The coordinate vector fields $\partial/\partial z_j$, $\partial/\partial \bar z_j$, and $\partial/\partial t$ are linearly independent over $\mathbb C$, so comparison of the $\partial/\partial z_j$ coefficients gives $a_j=0$ for every $j$, and comparison of the $\partial/\partial \bar z_j$ coefficients gives $b_j=0$ for every $j$. Hence \begin{align*} T^{1,0}H^n\cap T^{0,1}H^n=\{0\}. \end{align*} The fields $L_1,\dots,L_n$ are also pointwise linearly independent, because a relation $\sum_j a_jL_j=0$ forces all $\partial/\partial z_j$ coefficients $a_j$ to vanish. Thus $T^{1,0}H^n$ has complex rank $n$. Since $\mathbb C^n\times\mathbb R$ has real dimension $2n+1=2n+1$, this is a hypersurface-type CR structure of CR dimension $n$. To verify formal integrability, compute the brackets of the spanning fields. For a test function $f$, \begin{align*} L_jL_kf=\frac{\partial}{\partial z_j}\left(\frac{\partial f}{\partial z_k}+i\bar z_k\frac{\partial f}{\partial t}\right)+i\bar z_j\frac{\partial}{\partial t}\left(\frac{\partial f}{\partial z_k}+i\bar z_k\frac{\partial f}{\partial t}\right). \end{align*} Using $\partial \bar z_k/\partial z_j=0$ and $\partial \bar z_k/\partial t=0$, this expands to \begin{align*} L_jL_kf=\frac{\partial^2 f}{\partial z_j\partial z_k}+i\bar z_k\frac{\partial^2 f}{\partial z_j\partial t}+i\bar z_j\frac{\partial^2 f}{\partial t\partial z_k}-\bar z_j\bar z_k\frac{\partial^2 f}{\partial t^2}. \end{align*} Interchanging $j$ and $k$ gives \begin{align*} L_kL_jf=\frac{\partial^2 f}{\partial z_k\partial z_j}+i\bar z_j\frac{\partial^2 f}{\partial z_k\partial t}+i\bar z_k\frac{\partial^2 f}{\partial t\partial z_j}-\bar z_k\bar z_j\frac{\partial^2 f}{\partial t^2}. \end{align*} The second derivatives commute in these smooth coordinates, and the scalar coefficients commute. Therefore each term in $L_jL_kf$ cancels with the corresponding term in $L_kL_jf$, so \begin{align*} [L_j,L_k]f=L_jL_kf-L_kL_jf=0. \end{align*} Since this holds for every test function $f$, \begin{align*} [L_j,L_k]=0. \end{align*} Now let \begin{align*} L=\sum_{j=1}^n a_jL_j \end{align*} and \begin{align*} K=\sum_{k=1}^n b_kL_k \end{align*} be arbitrary smooth local sections of $T^{1,0}H^n$. The bracket identity $[aX,bY]=ab[X,Y]+aX(b)Y-bY(a)X$ gives \begin{align*} [L,K]=\sum_{j=1}^n\sum_{k=1}^n a_jb_k[L_j,L_k]+\sum_{j=1}^n\sum_{k=1}^n a_jL_j(b_k)L_k-\sum_{j=1}^n\sum_{k=1}^n b_kL_k(a_j)L_j. \end{align*} Since $[L_j,L_k]=0$, the first double sum is zero, and the remaining terms are smooth linear combinations of the fields $L_1,\dots,L_n$. Hence $[L,K]$ is again a section of $T^{1,0}H^n$, so the formal integrability condition holds. The mixed brackets show where the missing transverse direction appears. For a test function $f$, \begin{align*} L_j\bar L_kf=\frac{\partial}{\partial z_j}\left(\frac{\partial f}{\partial \bar z_k}-iz_k\frac{\partial f}{\partial t}\right)+i\bar z_j\frac{\partial}{\partial t}\left(\frac{\partial f}{\partial \bar z_k}-iz_k\frac{\partial f}{\partial t}\right). \end{align*} Using $\partial z_k/\partial z_j=\delta_{jk}$ and $\partial z_k/\partial t=0$, this becomes \begin{align*} L_j\bar L_kf=\frac{\partial^2 f}{\partial z_j\partial \bar z_k}-i\delta_{jk}\frac{\partial f}{\partial t}-iz_k\frac{\partial^2 f}{\partial z_j\partial t}+i\bar z_j\frac{\partial^2 f}{\partial t\partial \bar z_k}+\bar z_jz_k\frac{\partial^2 f}{\partial t^2}. \end{align*} Similarly, \begin{align*} \bar L_kL_jf=\frac{\partial}{\partial \bar z_k}\left(\frac{\partial f}{\partial z_j}+i\bar z_j\frac{\partial f}{\partial t}\right)-iz_k\frac{\partial}{\partial t}\left(\frac{\partial f}{\partial z_j}+i\bar z_j\frac{\partial f}{\partial t}\right). \end{align*} Using $\partial \bar z_j/\partial \bar z_k=\delta_{kj}$ and $\partial \bar z_j/\partial t=0$, this becomes \begin{align*} \bar L_kL_jf=\frac{\partial^2 f}{\partial \bar z_k\partial z_j}+i\delta_{kj}\frac{\partial f}{\partial t}+i\bar z_j\frac{\partial^2 f}{\partial \bar z_k\partial t}-iz_k\frac{\partial^2 f}{\partial t\partial z_j}+\bar z_jz_k\frac{\partial^2 f}{\partial t^2}. \end{align*} Subtracting, the second-derivative terms cancel by commutation of coordinate derivatives, while the first-order $\partial_t f$ terms give \begin{align*} [L_j,\bar L_k]f=-i\delta_{jk}\frac{\partial f}{\partial t}-i\delta_{kj}\frac{\partial f}{\partial t}. \end{align*} Since $\delta_{kj}=\delta_{jk}$, this is \begin{align*} [L_j,\bar L_k]f=-2i\delta_{jk}\frac{\partial f}{\partial t}. \end{align*} Therefore \begin{align*} [L_j,\bar L_k]=-2i\delta_{jk}\frac{\partial}{\partial t}. \end{align*} Thus the holomorphic directions close among themselves, while their commutators with conjugate directions recover the real transverse direction $\partial_t$. [/example] The Heisenberg group is the local nilpotent model for strongly pseudoconvex hypersurfaces. It will reappear when boundary behavior is rescaled near a point. ## CR Functions and CR Maps After the geometric structure is in place, the analytic question is which functions and maps preserve the tangential Cauchy-Riemann equations. This replaces ambient holomorphicity by differentiation only along the $T^{0,1}$ directions. [definition: CR Function] Let $(M,T^{1,0}M)$ be a CR manifold. A function $f\in C^\infty(M;\mathbb C)$ is a CR function if \begin{align*} \bar L f = 0 \end{align*} for every smooth section $\bar L\in \Gamma(T^{0,1}M)$. [/definition] CR functions are the intrinsic candidates for boundary values of holomorphic functions. To compare two CR manifolds, however, we need maps that send complex tangent directions on the source to complex tangent directions on the target. [definition: CR Map] Let $(M,T^{1,0}M)$ and $(N,T^{1,0}N)$ be CR manifolds. A smooth map $F:M\to N$ is a CR map if its complexified differential satisfies \begin{align*} dF_p(T^{1,0}_pM) \subset T^{1,0}_{F(p)}N \end{align*} for every $p\in M$. [/definition] An ordinary smooth map can destroy the tangential Cauchy-Riemann system by sending a complex tangent direction into a transverse or anti-holomorphic direction. The CR condition is exactly what prevents this failure: it makes differentiation of pullbacks compatible with the $T^{0,1}$ equations. This definition is coordinate-free and is stable under composition. In local CR coordinates, it says that the components of $F$ satisfy the tangential Cauchy-Riemann equations relative to the target structure. [definition: CR Diffeomorphism] A CR diffeomorphism is a diffeomorphism $F:M\to N$ such that both $F$ and $F^{-1}$ are CR maps. [/definition] CR diffeomorphisms are the isomorphisms in the category of CR manifolds. For a single CR manifold, the natural symmetry problem is to separate diffeomorphisms that preserve the CR equations from those that merely preserve the underlying smooth manifold. The self-isomorphisms form the intrinsic symmetry group of the CR structure. [definition: CR Automorphism Group] For a CR manifold $M$, the CR automorphism group is \begin{align*} \operatorname{Aut}_{CR}(M) := \{F:M\to M : F \text{ is a CR diffeomorphism}\}. \end{align*} [/definition] The automorphism group measures the intrinsic symmetry of the CR structure. For homogeneous examples, this group can be large; for a generic smooth CR manifold it may be very small. [example: Automorphisms of the Sphere] For $n\ge 2$, the *boundary extension theorem for the ball* says that every CR automorphism of the standard sphere $S^{2n-1}=\partial B^n$ is the boundary value of a biholomorphic automorphism of $B^n$. The algebraic model for these automorphisms comes from the Hermitian form on $\mathbb C^{n+1}$ given by \begin{align*} \langle (z,w),(\zeta,\eta)\rangle_{n,1}=\sum_{j=1}^n z_j\overline{\zeta_j}-w\overline{\eta}. \end{align*} In the affine chart $w=1$, its null cone cuts out the sphere, since \begin{align*} \langle (z,1),(z,1)\rangle_{n,1}=\sum_{j=1}^n z_j\overline{z_j}-1=\sum_{j=1}^n |z_j|^2-1=|z|^2-1. \end{align*} Thus $\langle (z,1),(z,1)\rangle_{n,1}=0$ exactly when $|z|=1$. Let $A\in U(n,1)$, so \begin{align*} \langle A\xi,A\eta\rangle_{n,1}=\langle \xi,\eta\rangle_{n,1}. \end{align*} Write the block action on the affine vector $(z,1)$ as \begin{align*} A(z,1)=(Pz+q,rz+d), \end{align*} where $P$ is an $n\times n$ matrix, $q\in\mathbb C^n$, $r$ is a complex row vector, and $d\in\mathbb C$. Since $A$ preserves the Hermitian form, \begin{align*} \langle A(z,1),A(z,1)\rangle_{n,1}=\langle (z,1),(z,1)\rangle_{n,1}. \end{align*} Substituting the two sides gives \begin{align*} |Pz+q|^2-|rz+d|^2=|z|^2-1. \end{align*} On the chart where $rz+d\ne 0$, the induced projective map is \begin{align*} F_A(z)=\frac{Pz+q}{rz+d}. \end{align*} If $|z|<1$ and $rz+d=0$, then the identity above gives \begin{align*} |Pz+q|^2=|z|^2-1<0, \end{align*} which is impossible because $|Pz+q|^2\ge 0$. Hence the denominator is nonzero on $B^n$. For such $z$, \begin{align*} |F_A(z)|^2=\left|\frac{Pz+q}{rz+d}\right|^2=\frac{|Pz+q|^2}{|rz+d|^2}. \end{align*} The identity $|Pz+q|^2-|rz+d|^2=|z|^2-1<0$ gives \begin{align*} |Pz+q|^2<|rz+d|^2. \end{align*} Dividing by the positive number $|rz+d|^2$ gives \begin{align*} |F_A(z)|^2<1. \end{align*} Thus $F_A$ maps $B^n$ into $B^n$. The same construction applied to $A^{-1}\in U(n,1)$ gives a holomorphic map $F_{A^{-1}}:B^n\to B^n$. Since projective actions compose according to matrix multiplication, $F_{A^{-1}}\circ F_A=F_I$ and $F_A\circ F_{A^{-1}}=F_I$, where \begin{align*} F_I(z)=z. \end{align*} Therefore $F_A$ is a biholomorphic automorphism of the unit ball. Now take $|z|=1$. If $rz+d=0$, then the form identity gives \begin{align*} |Pz+q|^2=0, \end{align*} so $Pz+q=0$. Hence $A(z,1)=(0,0)$, contradicting invertibility of $A$. Thus the denominator is also nonzero on the sphere. For $|z|=1$, the same identity gives \begin{align*} |Pz+q|^2-|rz+d|^2=0. \end{align*} Hence \begin{align*} |Pz+q|^2=|rz+d|^2. \end{align*} Dividing by $|rz+d|^2$ gives \begin{align*} |F_A(z)|^2=\frac{|Pz+q|^2}{|rz+d|^2}=1. \end{align*} So the boundary value of $F_A$ maps $S^{2n-1}$ to itself, and because $F_A$ is biholomorphic in the ball, its boundary restriction is a CR automorphism of the sphere. Multiplying $A$ by a scalar $\lambda$ with $|\lambda|=1$ does not change the projective map, since \begin{align*} F_{\lambda A}(z)=\frac{\lambda(Pz+q)}{\lambda(rz+d)}=\frac{Pz+q}{rz+d}=F_A(z). \end{align*} The scalar matrices with $|\lambda|=1$ are exactly the scalar matrices in $U(n,1)$, so the effective projective group is \begin{align*} PU(n,1)=U(n,1)/\{\lambda I:|\lambda|=1\}. \end{align*} Together with the boundary extension theorem and the standard classification of ball automorphisms by these projective unitary transformations, this identifies $\operatorname{Aut}_{CR}(S^{2n-1})$ with $PU(n,1)$. The intrinsic CR structure on the sphere therefore recovers exactly the projective holomorphic symmetry group of the ball. [/example] The sphere example is a first instance of a boundary rigidity principle: an intrinsic CR equivalence can force ambient holomorphic equivalence. Later chapters develop this phenomenon through extension theorems. ## Real-Analytic and Smooth CR Structures The final issue in this chapter is whether an abstract CR manifold can be realized locally inside some $\mathbb C^N$. This is the local embeddability problem, and it separates real-analytic CR geometry from smooth CR geometry in a decisive way. [definition: Local CR Embedding] A CR manifold $(M,T^{1,0}M)$ is locally CR-embeddable at $p\in M$ if there is a neighbourhood $U\subset M$ of $p$ and a smooth CR map \begin{align*} F:U\to \mathbb C^N \end{align*} for some $N$, such that $F$ is an embedding and the CR structure on $U$ is the one induced by $F(U)\subset \mathbb C^N$. [/definition] Local embeddability asks whether the abstract tangential system has enough independent CR functions to serve as coordinates. Real analyticity supplies a strong existence mechanism for such functions. [quotetheorem:9196] [citeproof:9196] The theorem is a major structural reason that real-analytic CR manifolds behave like embedded hypersurfaces: analytic complexification converts the tangential Cauchy-Riemann system into a holomorphic Frobenius problem. Its conclusion is deliberately local. It does not assert a global embedding of $M$, does not prescribe a minimal target dimension, and does not remove possible global topological or monodromy obstructions. The contrast with smooth CR geometry is sharp: Nirenberg's example below shows that formal integrability alone can leave the overdetermined PDE $\bar Lf=0$ with too few local solutions to form coordinates. [quotetheorem:9197] [citeproof:9197] The hypotheses matter in two directions. The example is smooth and formally integrable, so it rules out any theorem saying that the abstract CR axioms alone force local realization in complex Euclidean space. At the same time, it is a local non-existence result at a chosen point; it does not say that all smooth CR manifolds are non-embeddable, nor does it contradict the many smooth CR structures that arise as actual hypersurfaces. Its role in the course is to mark where several complex variables becomes an analysis problem: CR functions are solutions of an overdetermined first-order PDE system, and later extension and rigidity theorems must either assume embeddedness or prove enough solvability to recover it. [remark: Real-Analytic Versus Smooth] Real-analytic CR structures can be complexified, so integrability becomes a holomorphic Frobenius problem. Smooth CR structures lack this complexification mechanism, and the local solvability of the tangential Cauchy-Riemann equations becomes a genuine PDE issue. [/remark] The intrinsic CR viewpoint clarifies what the Heisenberg group captures: the flat model of a nontrivial tangential complex structure. From there, the Siegel domain serves as the standard strictly pseudoconvex template on which later operators and estimates can be tested explicitly. # 3. The Heisenberg Group and Siegel Domain This chapter replaces a general strictly pseudoconvex boundary by its flat local model. The prerequisites are the Levi form, CR tangent spaces, holomorphic coordinate changes, and the elementary language of Lie groups and vector fields from the preceding chapters. The model is the Heisenberg group, which appears as the boundary of the Siegel upper half-space after separating complex tangential directions from the real normal direction. The main questions are how the group law encodes the CR structure, how the Siegel domain is equivalent to the unit ball, and why the resulting anisotropic geometry is the correct scale for boundary analysis. ## The Heisenberg Group as a Lie Group The boundary models from the preceding chapter still look local and coordinate-dependent. We want a single homogeneous space whose translations preserve the CR directions and the contact form, so that local computations can be made in a flat setting before being transferred back to curved hypersurfaces. [definition: Heisenberg Group] The Heisenberg group $\mathbb H^n$ is the smooth manifold $\mathbb C^n \times \mathbb R$ equipped with the multiplication map \begin{align*} m:\mathbb H^n\times\mathbb H^n\to\mathbb H^n \end{align*} defined by \begin{align*} m((z,t),(w,s))=(z,t)(w,s) = \left(z+w, t+s+2\operatorname{Im}(z\cdot \bar{w})\right), \end{align*} where $z,w\in \mathbb C^n$, $t,s\in \mathbb R$, and $z\cdot \bar{w}=\sum_{j=1}^n z_j\bar{w}_j$. The identity element is $(0,0)$, and the inversion map \begin{align*} \iota:\mathbb H^n\to\mathbb H^n,\qquad \iota(z,t)=(-z,-t) \end{align*} is smooth. [/definition] These formulas look close to addition on $\mathbb C^n\times\mathbb R$, but the central correction changes how paths compose. The extra term is the first signal that this model is not Euclidean: horizontal directions do not commute, and their commutator produces the missing real direction. The case $n=1$ already displays this phenomenon, so it is worth computing the product in real coordinates before introducing the general [Lie algebra](/page/Lie%20Algebra) notation. [example: The First Heisenberg Product] For $n=1$, write $z=x+iy$ and $w=u+iv$, with $x,y,u,v\in\mathbb R$. Then $\bar w=u-iv$, so the complex product is \begin{align*} z\bar w=(x+iy)(u-iv) \end{align*} and expanding the four terms gives \begin{align*} (x+iy)(u-iv)=xu-ixv+iyu-i^2yv. \end{align*} Since $i^2=-1$, this becomes \begin{align*} xu-ixv+iyu-i^2yv=xu+yv+i(yu-xv). \end{align*} Therefore \begin{align*} \operatorname{Im}(z\bar w)=yu-xv. \end{align*} Substituting this into the Heisenberg product formula $(z,t)(w,s)=(z+w,t+s+2\operatorname{Im}(z\bar w))$ gives \begin{align*} (x,y,t)(u,v,s)=(x+u,y+v,t+s+2(yu-xv)). \end{align*} Now multiply the two basic horizontal steps in both orders. First, \begin{align*} (1,0,0)(0,1,0)=(1+0,0+1,0+0+2(0\cdot 0-1\cdot 1)). \end{align*} The central coordinate is \begin{align*} 0+0+2(0\cdot 0-1\cdot 1)=2(0-1)=-2, \end{align*} so \begin{align*} (1,0,0)(0,1,0)=(1,1,-2). \end{align*} In the reverse order, \begin{align*} (0,1,0)(1,0,0)=(0+1,1+0,0+0+2(1\cdot 1-0\cdot 0)). \end{align*} The central coordinate is \begin{align*} 0+0+2(1\cdot 1-0\cdot 0)=2(1-0)=2, \end{align*} so \begin{align*} (0,1,0)(1,0,0)=(1,1,2). \end{align*} Both products have the same horizontal coordinate $(1,1)$, but their central coordinates are $-2$ and $2$, so the order of horizontal motion is recorded in the center of $\mathbb H^1$. [/example] The example isolates the central correction term in the product. To compute systematically with that correction, we need the Lie algebra formula that truncates the Baker-Campbell-Hausdorff series for this two-step nilpotent group. [quotetheorem:9198] [citeproof:9198] The nilpotence hypothesis is essential here: for a general Lie group the Baker-Campbell-Hausdorff series contains higher commutators and does not collapse to a single bracket term. A concrete warning comes from $\mathfrak{sl}_2(\mathbb C)$ with basis $H,E,F$ satisfying $[H,E]=2E$, $[H,F]=-2F$, and $[E,F]=H$. For $X=aH$ and $Y=bE$, the commutator $[X,Y]=2abE$ does not commute with $X$, since $[X,[X,Y]]=4a^2bE$ may be nonzero, so terms beyond $\frac12[X,Y]$ remain in the Baker-Campbell-Hausdorff expansion. Thus this theorem does not say that all CR boundaries have this exact product; it says that the flat nilpotent approximation does. To turn this algebraic information into differential operators on the boundary model, we next define the left-invariant fields whose values at the identity are the standard horizontal and central directions. [definition: Standard Left Invariant Vector Fields] Write $z_j=x_j+iy_j$ on $\mathbb C^n$. The standard left-invariant real vector fields are differential operators \begin{align*} X_j,Y_j,T:C^\infty(\mathbb H^n)\to C^\infty(\mathbb H^n) \end{align*} defined by \begin{align*} X_j = \partial_{x_j}+2y_j\partial_t. \end{align*} \begin{align*} Y_j = \partial_{y_j}-2x_j\partial_t. \end{align*} \begin{align*} T = \partial_t. \end{align*} The corresponding complex vector fields are \begin{align*} Z_j,\bar Z_j:C^\infty(\mathbb H^n;\mathbb C)\to C^\infty(\mathbb H^n;\mathbb C) \end{align*} defined by \begin{align*} Z_j = \frac{1}{2}(X_j-iY_j). \end{align*} \begin{align*} \bar Z_j = \frac{1}{2}(X_j+iY_j). \end{align*} [/definition] These fields span the CR directions and their conjugates, but the main structural point is not only their span. We need their commutators to see whether the horizontal directions generate the missing central direction, since that bracket generation is what later drives subelliptic estimates. [quotetheorem:9199] [citeproof:9199] The coefficient signs matter because they determine the orientation of the contact form and the sign convention for the Levi form. If the central terms in $X_j$ and $Y_j$ were omitted, all horizontal brackets would vanish and the model would reduce to Euclidean tangential geometry, losing the normal direction generated by the CR structure. The theorem therefore supplies both the horizontal directions and the transverse direction generated by their brackets. It does not say that every vector field on the boundary is left-invariant, nor that all commutators on a curved CR hypersurface have constant coefficients; those features belong to the flat model. This motivates the following definition of the contact form: it is the one-form whose kernel is exactly the horizontal bundle spanned by the $X_j$ and $Y_j$. [definition: Standard Contact Form] The standard contact form on $\mathbb H^n$ is the smooth one-form $\theta\in\Omega^1(\mathbb H^n)$ given by \begin{align*} \theta = dt +2\sum_{j=1}^n (x_jdy_j-y_jdx_j). \end{align*} For each $p\in\mathbb H^n$, its value is the [linear map](/page/Linear%20Map) \begin{align*} \theta_p:T_p\mathbb H^n\to\mathbb R. \end{align*} The horizontal bundle is the subbundle \begin{align*} H\mathbb H^n = \ker \theta. \end{align*} [/definition] With this convention, $X_j$ and $Y_j$ lie in $H\mathbb H^n$, while $T$ is transverse to it. Thus the model carries a contact structure together with the complex splitting generated by $Z_j$ and $\bar Z_j$. [example: Horizontal Curves And The Central Coordinate] Let $\gamma:[a,b]\to\mathbb H^n$ be a smooth curve with \begin{align*} \gamma(r)=(z(r),t(r)) \end{align*} and \begin{align*} z_j(r)=x_j(r)+iy_j(r). \end{align*} We compute the condition that $\gamma$ be horizontal, meaning $\dot\gamma(r)\in H_{\gamma(r)}\mathbb H^n=\ker\theta_{\gamma(r)}$ for every $r$. The standard contact form is \begin{align*} \theta=dt+2\sum_{j=1}^n(x_jdy_j-y_jdx_j), \end{align*} and the velocity vector is \begin{align*} \dot\gamma(r)=\sum_{j=1}^n\dot x_j(r)\partial_{x_j}+\sum_{j=1}^n\dot y_j(r)\partial_{y_j}+\dot t(r)\partial_t. \end{align*} The coordinate one-forms evaluate on this vector by \begin{align*} dt(\dot\gamma(r))=\dot t(r). \end{align*} For each $j$, \begin{align*} dy_j(\dot\gamma(r))=\dot y_j(r). \end{align*} For each $j$, \begin{align*} dx_j(\dot\gamma(r))=\dot x_j(r). \end{align*} Therefore \begin{align*} (x_jdy_j)_{\gamma(r)}(\dot\gamma(r))=x_j(r)\dot y_j(r). \end{align*} Similarly, \begin{align*} (y_jdx_j)_{\gamma(r)}(\dot\gamma(r))=y_j(r)\dot x_j(r). \end{align*} Substituting these evaluations into $\theta$ gives \begin{align*} \theta_{\gamma(r)}(\dot\gamma(r))=\dot t(r)+2\sum_{j=1}^n\left(x_j(r)\dot y_j(r)-y_j(r)\dot x_j(r)\right). \end{align*} Thus $\dot\gamma(r)\in\ker\theta_{\gamma(r)}$ exactly when $\theta_{\gamma(r)}(\dot\gamma(r))=0$, which is equivalent to \begin{align*} \dot t(r)+2\sum_{j=1}^n\left(x_j(r)\dot y_j(r)-y_j(r)\dot x_j(r)\right)=0. \end{align*} Solving for $\dot t(r)$ gives \begin{align*} \dot t(r)=-2\sum_{j=1}^n\left(x_j(r)\dot y_j(r)-y_j(r)\dot x_j(r)\right). \end{align*} Integrating both sides from $a$ to $b$ yields \begin{align*} \int_a^b\dot t(r)\,dr=-2\int_a^b\sum_{j=1}^n\left(x_j(r)\dot y_j(r)-y_j(r)\dot x_j(r)\right)\,dr. \end{align*} By the [fundamental theorem of calculus](/theorems/632), \begin{align*} \int_a^b\dot t(r)\,dr=t(b)-t(a). \end{align*} Hence the central displacement of a horizontal curve is \begin{align*} t(b)-t(a)=-2\int_a^b\sum_{j=1}^n\left(x_j(r)\dot y_j(r)-y_j(r)\dot x_j(r)\right)\,dr. \end{align*} If the projected curve $z(r)$ is closed and its signed symplectic area is denoted by \begin{align*} A_{\mathrm{sym}}(z)=\int_a^b\sum_{j=1}^n\left(x_j(r)\dot y_j(r)-y_j(r)\dot x_j(r)\right)\,dr, \end{align*} then substitution gives \begin{align*} t(b)-t(a)=-2A_{\mathrm{sym}}(z). \end{align*} So a horizontal curve can return to its initial $z$-coordinate while its central coordinate changes; that change is exactly minus twice the signed symplectic area swept out by the horizontal projection. [/example] ## The Siegel Upper Half-Space And The Cayley Transform The Heisenberg group becomes relevant to complex analysis because it is naturally the boundary of a strictly pseudoconvex domain. The next question is how this boundary model compares with the most familiar strictly pseudoconvex domain, the unit ball. [definition: Siegel Upper Half Space] The Siegel upper half-space is \begin{align*} \mathcal U^n=\{(z',z_{n+1})\in \mathbb C^n\times \mathbb C: \operatorname{Im}(z_{n+1})>|z'|^2\}. \end{align*} Its boundary is \begin{align*} \partial\mathcal U^n=\{(z',z_{n+1})\in \mathbb C^n\times \mathbb C: \operatorname{Im}(z_{n+1})=|z'|^2\}. \end{align*} [/definition] The boundary can be parameterised by $(z,t)\in\mathbb C^n\times\mathbb R$ through \begin{align*} (z,t)\longmapsto (z,t+i|z|^2). \end{align*} This identifies the Siegel boundary with the Heisenberg manifold. However, the parameterisation alone does not explain how the model relates to the compact sphere, where many global formulas are easiest to derive. A plain affine change of variables cannot send a bounded ball to an unbounded half-space, so the comparison requires a fractional holomorphic map with one boundary point sent to infinity. To compare it with the ball and transfer familiar spherical formulas, we need an explicit biholomorphism between the two domains. [quotetheorem:9200] [citeproof:9200] The exclusion of the point with last coordinate $-1$ is forced by the denominator, and geometrically it is the point sent to infinity in the Siegel model. The theorem is not merely a set-theoretic correspondence: holomorphicity and the explicit inverse are what allow analytic objects on one domain to be transported to the other. Without the ball condition $|\zeta|<1$, the displayed positivity identity would not place the image in $\mathcal U^n$: for instance, $\zeta=(0,\dots,0,2)$ gives image last coordinate $-i/3$, whose imaginary part is negative while $|z'|^2=0$. The transform turns spherical CR geometry into Heisenberg CR geometry, and it is the same bridge used later to compare boundary kernels on the ball with their anisotropic counterparts on $\mathbb H^n$. The next boundary statement is needed because CR analysis is performed on the hypersurface, not only in the interior domain. [quotetheorem:9201] [citeproof:9201] The CR conclusion would fail for an arbitrary smooth boundary diffeomorphism, because such a map need not preserve complex tangent directions. The puncture is also not cosmetic: at the omitted point the formula has a pole, and the Heisenberg model records that point as infinity rather than as an ordinary boundary point. Thus the theorem explains why estimates on the ball can be rewritten as anisotropic estimates on $\mathbb H^n$ without losing the CR structure. The next example records the boundary form of the transform explicitly. [example: Explicit Boundary Cayley Transform] Take $\zeta=(\zeta',\eta)\in\partial B^{n+1}$ with $\eta=\zeta_{n+1}\ne -1$. Then $1+\eta\ne0$, and the boundary condition is \begin{align*} |\zeta'|^2+|\eta|^2=1. \end{align*} Under the Cayley transform, \begin{align*} z=\frac{\zeta'}{1+\eta}. \end{align*} The last coordinate is \begin{align*} w=i\frac{1-\eta}{1+\eta}. \end{align*} To separate the real and imaginary parts of $w$, multiply numerator and denominator by $1+\bar\eta$: \begin{align*} w=i\frac{(1-\eta)(1+\bar\eta)}{(1+\eta)(1+\bar\eta)}. \end{align*} Since $(1+\eta)(1+\bar\eta)=|1+\eta|^2$, this becomes \begin{align*} w=i\frac{(1-\eta)(1+\bar\eta)}{|1+\eta|^2}. \end{align*} Expanding the numerator gives \begin{align*} (1-\eta)(1+\bar\eta)=1+\bar\eta-\eta-\eta\bar\eta. \end{align*} Because $\eta\bar\eta=|\eta|^2$, this is \begin{align*} 1+\bar\eta-\eta-\eta\bar\eta=1-|\eta|^2+\bar\eta-\eta. \end{align*} Writing $\eta=a+ib$ with $a,b\in\mathbb R$, we have $\bar\eta=a-ib$, so \begin{align*} \bar\eta-\eta=(a-ib)-(a+ib)=-2ib=-2i\operatorname{Im}\eta. \end{align*} Therefore \begin{align*} (1-\eta)(1+\bar\eta)=(1-|\eta|^2)-2i\operatorname{Im}\eta. \end{align*} Substituting this into the formula for $w$ gives \begin{align*} w=i\frac{(1-|\eta|^2)-2i\operatorname{Im}\eta}{|1+\eta|^2}. \end{align*} Distributing the factor $i$ in the numerator, \begin{align*} i\left((1-|\eta|^2)-2i\operatorname{Im}\eta\right)=i(1-|\eta|^2)-2i^2\operatorname{Im}\eta. \end{align*} Since $i^2=-1$, this becomes \begin{align*} i(1-|\eta|^2)-2i^2\operatorname{Im}\eta=i(1-|\eta|^2)+2\operatorname{Im}\eta. \end{align*} Hence \begin{align*} w=\frac{2\operatorname{Im}\eta+i(1-|\eta|^2)}{|1+\eta|^2}. \end{align*} Since $|1+\eta|^2$ is a positive real number, the real and imaginary parts are \begin{align*} \operatorname{Re}w=\frac{2\operatorname{Im}\eta}{|1+\eta|^2}. \end{align*} \begin{align*} \operatorname{Im}w=\frac{1-|\eta|^2}{|1+\eta|^2}. \end{align*} Now compute the boundary height determined by $z$. Since $1+\eta\ne0$, \begin{align*} |z|^2=\left|\frac{\zeta'}{1+\eta}\right|^2. \end{align*} Using $|a/b|^2=|a|^2/|b|^2$ for $b\ne0$, \begin{align*} |z|^2=\frac{|\zeta'|^2}{|1+\eta|^2}. \end{align*} The sphere condition $|\zeta'|^2+|\eta|^2=1$ gives $|\zeta'|^2=1-|\eta|^2$, so \begin{align*} |z|^2=\frac{1-|\eta|^2}{|1+\eta|^2}. \end{align*} Comparing this with the formula for $\operatorname{Im}w$ yields \begin{align*} \operatorname{Im}w=|z|^2. \end{align*} Thus $C(\zeta)=(z,w)$ lies on $\partial\mathcal U^n$, and since every boundary point has the form $(z,t+i|z|^2)$ with $t=\operatorname{Re}w$, we get \begin{align*} C(\zeta)=(z,t+i|z|^2). \end{align*} Here \begin{align*} t=\operatorname{Re}\left(i\frac{1-\zeta_{n+1}}{1+\zeta_{n+1}}\right)=\frac{2\operatorname{Im}\zeta_{n+1}}{|1+\zeta_{n+1}|^2}. \end{align*} The number $t$ is the Heisenberg central coordinate of the boundary point, while the excluded point $\zeta_{n+1}=-1$ is exactly where the denominator of the Cayley transform vanishes. [/example] ## Boundary Translations, Koranyi Geometry, And Dilations Euclidean balls do not respect the vector fields $X_j,Y_j,T$: the central direction has twice the scaling weight of a horizontal direction. The problem is to find the metric-scale objects preserved by the group structure and compatible with the CR boundary. [definition: Boundary Action Of The Heisenberg Group] For $(a,\tau)\in\mathbb H^n$, the abstract Heisenberg coordinate action is the left translation map \begin{align*} L_{(a,\tau)}:\mathbb H^n\to\mathbb H^n, \qquad L_{(a,\tau)}(z,t)=(a,\tau)(z,t). \end{align*} The corresponding Siegel boundary action is the map \begin{align*} \Lambda_{(a,\tau)}:\partial\mathcal U^n\to\partial\mathcal U^n \end{align*} defined by \begin{align*} \Lambda_{(a,\tau)}(z,t+i|z|^2)=\left(a+z,\tau+t+2\operatorname{Im}(a\cdot\bar z)+i|a+z|^2\right). \end{align*} [/definition] The formula for $\Lambda_{(a,\tau)}$ is obtained from the embedding $(z,t)\mapsto (z,t+i|z|^2)$ of $\mathbb H^n$ into $\partial\mathcal U^n$. These translations act transitively on the boundary coordinates. To use them as changes of base point in CR analysis, we must check that they preserve the contact distribution and its complex structure. [quotetheorem:9202] [citeproof:9202] The word "left" is important: the vector fields used above were built to be invariant under left translations, so the proof would not apply unchanged to an arbitrary diffeomorphism of the manifold. For instance, the smooth map $F(z,t)=(z,t+\operatorname{Re}z_1)$ on $\mathbb H^n$ has \begin{align*} F_*X_1=X_1+\partial_t, \end{align*} after comparing at corresponding points, and this changes the horizontal distribution because $\partial_t=T$ is transverse to $\ker\theta$. Thus a smooth boundary coordinate change can fail to be CR even when it is a diffeomorphism of the underlying manifold. The theorem does not claim that every boundary automorphism of the Siegel domain is a Heisenberg translation; it isolates the subgroup that moves base points while preserving the flat CR frame. The translation theorem gives homogeneous base points, but it does not yet give the correct notion of scale. This motivates the following definition of anisotropic dilations, with weight one on horizontal variables and weight two on the central variable because $T$ arises from horizontal brackets. [definition: Anisotropic Dilations] The Heisenberg dilation family is the map \begin{align*} \delta:(0,\infty)\times\mathbb H^n\to\mathbb H^n \end{align*} defined by \begin{align*} \delta(r,(z,t))=\delta_r(z,t)=(rz,r^2t). \end{align*} [/definition] The definition assigns weight one to the complex tangential variables and weight two to the central variable. This weighting is not only a convenient convention; it is forced by the group law and by the commutator relation that produces $T$ from two horizontal derivatives. The next theorem is needed to turn that scaling rule into an actual symmetry of the boundary model: if $\delta_r$ failed to respect the product or the vector fields, then scaled neighbourhoods would not preserve the CR translation structure used in estimates. [quotetheorem:9203] [citeproof:9203] Ordinary Euclidean distance treats $t$ as having the same size as the components of $z$, so it is incompatible with these pushforward relations and with the bracket relation that produces $T$. A concrete failure occurs on the central axis: the Euclidean distance from $(0,0)$ to $(0,t)$ scales by $r^2$ under $\delta_r$, whereas horizontal distances scale by $r$; no Euclidean ball can be homogeneous for both behaviours at once. To measure distance in a way that is both left-invariant and homogeneous under these dilations, we now introduce the Koranyi norm. [definition: Koranyi Norm] The Koranyi norm is the map \begin{align*} \|\cdot\|_K:\mathbb H^n\to[0,\infty), \qquad \|(z,t)\|_K = \left(|z|^4+t^2\right)^{1/4}. \end{align*} The associated left-invariant Koranyi distance is the map \begin{align*} d_K:\mathbb H^n\times\mathbb H^n\to[0,\infty), \qquad d_K(p,q)=\|q^{-1}p\|_K. \end{align*} [/definition] The fourth root is chosen so that the norm is homogeneous of degree one under $\delta_r$. This construction does not make central and horizontal displacements interchangeable; instead it records that a central displacement of size $r^2$ belongs to the same scale as a horizontal displacement of size $r$. The level sets are the boundary balls naturally seen by the tangential Cauchy-Riemann equations, and the homogeneous dimension of $\mathbb H^n$ is $Q=2n+2$. [example: Koranyi Ball As A CR Analogue Of A Euclidean Ball] For $r>0$, the Koranyi ball centered at the identity is the set of points satisfying $\|(z,t)\|_K<r$. By the definition of the Koranyi norm, \begin{align*} \|(z,t)\|_K=\left(|z|^4+t^2\right)^{1/4}. \end{align*} Therefore the ball condition is \begin{align*} \left(|z|^4+t^2\right)^{1/4}<r. \end{align*} Both sides are nonnegative, and the function $a\mapsto a^4$ is strictly increasing on $[0,\infty)$, so this inequality is equivalent to \begin{align*} \left(\left(|z|^4+t^2\right)^{1/4}\right)^4<r^4. \end{align*} Since $\left(A^{1/4}\right)^4=A$ for $A\ge0$, this becomes \begin{align*} |z|^4+t^2<r^4. \end{align*} Thus \begin{align*} B_K(0,r)=\{(z,t)\in\mathbb H^n: |z|^4+t^2<r^4\}. \end{align*} This inequality displays the two scales. If $t=0$, then the condition becomes $|z|^4<r^4$. Since $|z|\ge0$ and $r>0$, taking fourth roots gives \begin{align*} |z|<r. \end{align*} If $z=0$, then the condition becomes $t^2<r^4$. Taking square roots gives \begin{align*} |t|<r^2. \end{align*} For an arbitrary point in $B_K(0,r)$, the inequalities $|z|^4\ge0$ and $t^2\ge0$ give \begin{align*} |z|^4\le |z|^4+t^2<r^4. \end{align*} Hence $|z|<r$. Similarly, \begin{align*} t^2\le |z|^4+t^2<r^4, \end{align*} so $|t|<r^2$. Conversely, consider the box \begin{align*} \{(z,t): |z|<r/2,\ |t|<r^2/2\}. \end{align*} For a point in this box, $|z|<r/2$ implies \begin{align*} |z|^4<\left(\frac r2\right)^4=\frac{r^4}{16}. \end{align*} Also $|t|<r^2/2$ implies \begin{align*} t^2<\left(\frac{r^2}{2}\right)^2=\frac{r^4}{4}. \end{align*} Adding the two strict inequalities gives \begin{align*} |z|^4+t^2<\frac{r^4}{16}+\frac{r^4}{4}. \end{align*} Since $\frac{r^4}{4}=\frac{4r^4}{16}$, the right-hand side is \begin{align*} \frac{r^4}{16}+\frac{4r^4}{16}=\frac{5r^4}{16}. \end{align*} Because $5/16<1$, we have \begin{align*} \frac{5r^4}{16}<r^4. \end{align*} Therefore $|z|^4+t^2<r^4$, so the box lies inside $B_K(0,r)$. A Koranyi ball has horizontal radius of order $r$ and central radius of order $r^2$; near a strictly pseudoconvex boundary point, this is the CR analogue of a Euclidean ball because the normal direction is thinner than the complex tangential directions under Heisenberg scaling. [/example] Koranyi balls describe the correct neighbourhoods. For boundary integral operators, we also need the correct translation-invariant product of functions and kernels. Euclidean convolution would use $p-q$, but subtraction is not the operation that preserves the left-invariant CR vector fields; it would ignore the central correction term in the group law. Thus Euclidean convolution must be replaced by group convolution. [definition: Convolution On The Heisenberg Group] The Heisenberg group convolution is the bilinear map \begin{align*} *:C_c^\infty(\mathbb H^n)\times C_c^\infty(\mathbb H^n)\to C^\infty(\mathbb H^n) \end{align*} defined by \begin{align*} (f*g)(p)=\int_{\mathbb H^n} f(q)g(q^{-1}p)\,dq, \end{align*} where $dq$ denotes [Lebesgue measure](/page/Lebesgue%20Measure) on $\mathbb C^n\times\mathbb R$. [/definition] The compact support assumption ensures that the displayed integral is finite and gives a smooth output; broader $L^1$ and distributional versions require additional analytic hypotheses. The measure $dq$ is both left and right Haar measure for this nilpotent group, which is why no modular correction appears in the formula. Kernels homogeneous with respect to $\delta_r$ interact with this convolution in the same way that homogeneous kernels interact with Euclidean convolution. [example: Convolution With A Radial Koranyi Kernel] Let $q=(\zeta,\tau)$ and $p=(z,t)$. Since $q^{-1}=(-\zeta,-\tau)$, the group product gives \begin{align*} q^{-1}p=(-\zeta,-\tau)(z,t)=(z-\zeta,-\tau+t+2\operatorname{Im}((-\zeta)\cdot\bar z)). \end{align*} Because $(-\zeta)\cdot\bar z=-(\zeta\cdot\bar z)$ and $\operatorname{Im}(-A)=-\operatorname{Im}(A)$ for $A\in\mathbb C$, this becomes \begin{align*} q^{-1}p=(z-\zeta,t-\tau-2\operatorname{Im}(\zeta\cdot\bar z)). \end{align*} For $K(q)=\|q\|_K^{-\alpha}$ away from the identity, the convolution formula therefore gives \begin{align*} (K*f)(p)=\int_{\mathbb H^n}\frac{f(z-\zeta,t-\tau-2\operatorname{Im}(\zeta\cdot\bar z))}{(|\zeta|^4+\tau^2)^{\alpha/4}}\,d\zeta\,d\tau. \end{align*} The denominator vanishes exactly when $|\zeta|^4+\tau^2=0$, which is equivalent to $|\zeta|=0$ and $\tau=0$, so the only local singularity is at $q=(0,0)$. Under the dilation $\delta_r(\zeta,\tau)=(r\zeta,r^2\tau)$, \begin{align*} \|\delta_r(\zeta,\tau)\|_K=\left(|r\zeta|^4+(r^2\tau)^2\right)^{1/4}. \end{align*} Since $|r\zeta|=r|\zeta|$ for $r>0$, \begin{align*} |r\zeta|^4+(r^2\tau)^2=r^4|\zeta|^4+r^4\tau^2=r^4(|\zeta|^4+\tau^2). \end{align*} Taking fourth roots gives \begin{align*} \|\delta_r(\zeta,\tau)\|_K=r\left(|\zeta|^4+\tau^2\right)^{1/4}=r\|(\zeta,\tau)\|_K. \end{align*} The variable $\zeta\in\mathbb C^n$ has $2n$ real coordinates, each multiplied by $r$, while $\tau$ is multiplied by $r^2$. Hence the Jacobian determinant of $\delta_r$ is \begin{align*} r^{2n}r^2=r^{2n+2}. \end{align*} Writing $Q=2n+2$, this implies \begin{align*} |B_K(0,r)|=r^Q|B_K(0,1)|. \end{align*} Fix $\varepsilon>0$ and define \begin{align*} A_k=\{q:2^{-(k+1)}\varepsilon\le \|q\|_K<2^{-k}\varepsilon\},\qquad k=0,1,2,\dots. \end{align*} These annuli cover $B_K(0,\varepsilon)\setminus\{0\}$. On $A_k$ we have $\|q\|_K\ge 2^{-(k+1)}\varepsilon$, and since $a\mapsto a^{-\alpha}$ is decreasing for $\alpha>0$, \begin{align*} \|q\|_K^{-\alpha}\le (2^{-(k+1)}\varepsilon)^{-\alpha}=2^{(k+1)\alpha}\varepsilon^{-\alpha}. \end{align*} Also $A_k\subset B_K(0,2^{-k}\varepsilon)$, so \begin{align*} |A_k|\le |B_K(0,2^{-k}\varepsilon)|=(2^{-k}\varepsilon)^Q|B_K(0,1)|=2^{-kQ}\varepsilon^Q|B_K(0,1)|. \end{align*} Therefore \begin{align*} \int_{A_k}\|q\|_K^{-\alpha}\,dq\le 2^{(k+1)\alpha}\varepsilon^{-\alpha}2^{-kQ}\varepsilon^Q|B_K(0,1)|. \end{align*} Combining the powers gives \begin{align*} \int_{A_k}\|q\|_K^{-\alpha}\,dq\le 2^\alpha \varepsilon^{Q-\alpha}|B_K(0,1)|2^{-k(Q-\alpha)}. \end{align*} Summing over $k$ yields \begin{align*} \int_{B_K(0,\varepsilon)}\|q\|_K^{-\alpha}\,dq\le 2^\alpha\varepsilon^{Q-\alpha}|B_K(0,1)|\sum_{k=0}^{\infty}2^{-k(Q-\alpha)}. \end{align*} The series is geometric with ratio $2^{-(Q-\alpha)}$, and it converges exactly when $2^{-(Q-\alpha)}<1$, which is equivalent to $Q-\alpha>0$. Thus, for $0<\alpha<Q$, the kernel $\|q\|_K^{-\alpha}$ is locally integrable at the identity, and the convolution averages $f$ over Koranyi shells whose volume scales like $r^Q$. [/example] ## The Flat Model For Strictly Pseudoconvex CR Geometry The remaining issue is why this model describes an arbitrary strictly pseudoconvex hypersurface rather than only the Siegel boundary. The answer is that after choosing coordinates adapted to a boundary point, the lowest-order nondegenerate part of the defining function is the Heisenberg quadric. [definition: Heisenberg Quadric] The Heisenberg quadric in $\mathbb C^{n+1}$ is the hypersurface \begin{align*} M_0=\{(z',z_{n+1})\in\mathbb C^n\times\mathbb C: \operatorname{Im}(z_{n+1})=|z'|^2\}. \end{align*} [/definition] This is exactly $\partial\mathcal U^n$. Its Levi form is positive definite, and its CR vector fields are the flat normal forms of the tangential directions on any strictly pseudoconvex hypersurface. The precise local statement is the following normal form. [quotetheorem:9204] [citeproof:9204] Strict pseudoconvexity is the hypothesis that makes the Hermitian quadratic part positive definite; without it, the leading model need not be the Heisenberg quadric. For example, the hypersurface \begin{align*} \operatorname{Im}w=|z_1|^2-|z_2|^2 \end{align*} in $\mathbb C^3$ has mixed Levi signature, while \begin{align*} \operatorname{Im}w=|z_1|^4 \end{align*} has degenerate Levi form at the origin. Neither has the positive definite quadratic model $v=|z'|^2$ at the chosen point. The theorem is local and coordinate-dependent, so it does not identify the whole hypersurface globally with $\mathbb H^n$. It justifies treating $\mathbb H^n$ as the tangent object to strictly pseudoconvex CR geometry. It is not the tangent vector space; it is the nilpotent approximation that remembers the first noncommuting behaviour of CR vector fields. [remark: Weighted Tangent Geometry] Ordinary tangent planes miss the Levi bracket because they linearise all directions with the same weight. The Heisenberg scaling keeps horizontal directions at weight one and the Reeb direction at weight two, so the Levi form remains visible in the limiting model. This is why subelliptic estimates and boundary kernels are naturally written in Heisenberg coordinates. [/remark] [Normal coordinates](/theorems/2713) also explain the role of the Cayley transform. Locally, a strictly pseudoconvex boundary resembles the Siegel boundary, while globally the ball provides a compact reference model with one point sent to infinity. [example: Reading A Boundary Point In Normal Coordinates] Let $w=u+iv$, and suppose that in local holomorphic coordinates the hypersurface $M$ is written as \begin{align*} v=|z'|^2+R(z',u). \end{align*} Assume that the remainder has weighted order at least three for the scaling $(z',u)\mapsto (rz',r^2u)$, so that, in a fixed coordinate neighbourhood, \begin{align*} |R(z',u)|\le C\left(|z'|^3+|u||z'|+u^2\right). \end{align*} Fix a scale $r>0$ and introduce rescaled variables by \begin{align*} z'=r\xi,\qquad u=r^2\tau. \end{align*} Substituting these into the defining equation gives \begin{align*} v=|r\xi|^2+R(r\xi,r^2\tau). \end{align*} Since $r>0$ is real, $|r\xi|=r|\xi|$, hence \begin{align*} |r\xi|^2=(r|\xi|)^2=r^2|\xi|^2. \end{align*} Therefore \begin{align*} v=r^2|\xi|^2+R(r\xi,r^2\tau). \end{align*} Dividing both sides by $r^2$ gives \begin{align*} \frac{v}{r^2}=|\xi|^2+\frac{R(r\xi,r^2\tau)}{r^2}. \end{align*} Now apply the assumed remainder estimate at the point $(r\xi,r^2\tau)$: \begin{align*} |R(r\xi,r^2\tau)|\le C\left(|r\xi|^3+|r^2\tau||r\xi|+(r^2\tau)^2\right). \end{align*} The three terms on the right are \begin{align*} |r\xi|^3=(r|\xi|)^3=r^3|\xi|^3. \end{align*} Also, \begin{align*} |r^2\tau||r\xi|=r^2|\tau|\,r|\xi|=r^3|\tau||\xi|. \end{align*} Finally, \begin{align*} (r^2\tau)^2=r^4\tau^2. \end{align*} Substituting these identities into the estimate gives \begin{align*} |R(r\xi,r^2\tau)|\le C\left(r^3|\xi|^3+r^3|\tau||\xi|+r^4\tau^2\right). \end{align*} Dividing by $r^2$ yields \begin{align*} \left|\frac{R(r\xi,r^2\tau)}{r^2}\right|\le C\left(r|\xi|^3+r|\tau||\xi|+r^2\tau^2\right). \end{align*} On a [bounded set](/page/Bounded%20Set) where $|\xi|\le M$ and $|\tau|\le M$, the right-hand side is bounded by \begin{align*} C\left(rM^3+rM^2+r^2M^2\right). \end{align*} If $0<r\le1$, then $r^2M^2\le rM^2$, so \begin{align*} C\left(rM^3+rM^2+r^2M^2\right)\le Cr\left(M^3+2M^2\right). \end{align*} Thus \begin{align*} \frac{R(r\xi,r^2\tau)}{r^2}=O(r) \end{align*} uniformly on bounded sets of $(\xi,\tau)$ as $r\to0$. Since $v=\operatorname{Im}w$, the rescaled boundary equation becomes \begin{align*} \frac{\operatorname{Im}w}{r^2}=|\xi|^2+O(r). \end{align*} Hence the rescaled hypersurfaces converge locally to the model equation $\operatorname{Im}W=|\xi|^2$, the Heisenberg quadric. The scaling therefore treats $z'$ as size $r$ and both $\operatorname{Re}w$ and $\operatorname{Im}w$ as size $r^2$, which is exactly the anisotropic scale used in Koranyi geometry. [/example] The chapter has built the flat CR model in three equivalent ways: as a nilpotent Lie group, as the boundary of the Siegel upper half-space, and as the normal form of a strictly pseudoconvex hypersurface. In later analysis, the tangential Cauchy-Riemann operator $\bar\partial_b$ and the Kohn Laplacian will be compared with their Heisenberg counterparts before curvature and lower-order terms are restored. Once the model geometry is fixed, the next step is to turn the tangential structure into equations. The chapter on $\bar\partial_b$ builds the differential complex that measures whether CR data can be locally and globally solved. # 4. The Tangential Cauchy-Riemann Equations This chapter turns the geometric CR structure from the preceding chapters into a differential complex. It assumes the earlier definitions of CR manifolds, formally integrable CR bundles, the Levi form, pseudoconvexity, and the Dolbeault operator $\bar\partial$ on complex manifolds. The central question is how much of the ordinary Dolbeault operator survives after restricting attention to directions tangent to a CR manifold. The answer is the tangential Cauchy-Riemann operator $\bar\partial_b$, whose kernel gives CR functions and whose cohomology measures boundary obstructions to solving tangential Cauchy-Riemann equations. ## Tangential Forms and the Operator $\bar\partial_b$ The ordinary $\bar\partial$ operator differentiates in all antiholomorphic directions of an ambient complex manifold. On a CR manifold, only the antiholomorphic tangent directions are intrinsic, so the first problem is to define the part of $\bar\partial$ that depends only on the CR structure and not on arbitrary extensions off the manifold. Let $M$ be a smooth CR manifold of CR dimension $n-1$ and real dimension $2n-1$, with CR bundle $T^{1,0}M \subset \mathbb C TM$ and conjugate bundle $T^{0,1}M=\overline{T^{1,0}M}$. The complexified tangent bundle has a distinguished quotient of covectors that detect antiholomorphic tangential directions. [definition: Tangential Type $(0,q)$ Form] A smooth tangential form of type $(0,q)$ on $M$ is a smooth section of \begin{align*} \Lambda^{0,q}_b M := \Lambda^q (T^{0,1}M)^* . \end{align*} The space of smooth tangential $(0,q)$-forms is denoted $C^\infty(M,\Lambda^{0,q}_bM)$. [/definition] This definition records only the components of a form on $T^{0,1}M$. If $M$ is realised as a hypersurface in a complex manifold, an ambient $(0,q)$-form restricts to such a tangential form by evaluating it on $(0,1)$ tangent vector fields. [example: Tangential One-Forms on a Hypersurface] Let $M \subset \mathbb C^n$ be a smooth real hypersurface with defining function $\rho$, so $M=\{\rho=0\}$ and $d\rho\ne0$ on $M$. Choose a local frame $\bar L_1,\dots,\bar L_{n-1}$ for $T^{0,1}M$ and let $\bar\omega_1,\dots,\bar\omega_{n-1}$ be the dual coframe, meaning $\bar\omega_j(\bar L_k)=\delta_{jk}$. For a tangential $(0,1)$-form $\alpha$, set \begin{align*} a_j=\alpha(\bar L_j). \end{align*} We show that these functions are exactly the local coefficients of $\alpha$. Define \begin{align*} \beta=\sum_{j=1}^{n-1}a_j\bar\omega_j. \end{align*} For each basis vector $\bar L_k$, evaluation gives \begin{align*} \beta(\bar L_k)=\left(\sum_{j=1}^{n-1}a_j\bar\omega_j\right)(\bar L_k). \end{align*} By linearity of evaluation, \begin{align*} \beta(\bar L_k)=\sum_{j=1}^{n-1}a_j\bar\omega_j(\bar L_k). \end{align*} Using the duality relation $\bar\omega_j(\bar L_k)=\delta_{jk}$, \begin{align*} \beta(\bar L_k)=\sum_{j=1}^{n-1}a_j\delta_{jk}. \end{align*} The factor $\delta_{jk}$ is $0$ when $j\ne k$ and $1$ when $j=k$, so the sum has only one nonzero term: \begin{align*} \beta(\bar L_k)=a_k. \end{align*} Since $a_k=\alpha(\bar L_k)$, the forms $\alpha$ and $\beta$ agree on every vector in the frame of $T^{0,1}M$. Hence \begin{align*} \alpha=\sum_{j=1}^{n-1}a_j\bar\omega_j. \end{align*} The ambient normal component disappears because $\bar\partial\rho$ vanishes on every antiholomorphic tangential vector. If $\bar L\in T^{0,1}M$, then $\bar L$ is tangent to the level set $\rho=0$, so \begin{align*} d\rho(\bar L)=\bar L\rho=0 \end{align*} on $M$. Since $\bar L$ has type $(0,1)$), every $(1,0)$-form vanishes on $\bar L$, and therefore \begin{align*} (\partial\rho)(\bar L)=0. \end{align*} Because $d\rho=\partial\rho+\bar\partial\rho$, \begin{align*} (\bar\partial\rho)(\bar L)=d\rho(\bar L)-(\partial\rho)(\bar L)=0-0=0. \end{align*} Thus, if two ambient $(0,1)$-forms differ by $g\,\bar\partial\rho$, then for every $\bar L\in T^{0,1}M$ their difference evaluates to \begin{align*} (g\,\bar\partial\rho)(\bar L)=g(\bar\partial\rho)(\bar L)=g\cdot0=0. \end{align*} They induce the same tangential form. Thus a tangential $(0,1)$-form records precisely the coefficients along $\bar\omega_1,\dots,\bar\omega_{n-1}$ and forgets the ambient conormal component generated by $\bar\partial\rho$. [/example] The example identifies the data that a tangential form contains, but it also raises the next problem: differentiating coefficients may produce normal terms. This motivates defining an operator by differentiating first and then projecting back to the intrinsic antiholomorphic tangential component. [definition: Tangential Cauchy-Riemann Operator] For $0\le q\le n-2$, the tangential Cauchy-Riemann operator is the first-order differential operator \begin{align*} \bar\partial_b:C^\infty(M,\Lambda^{0,q}_bM)\longrightarrow C^\infty(M,\Lambda^{0,q+1}_bM) \end{align*} defined as follows. Let \begin{align*} \pi_b:\Lambda^{q+1}\mathbb C T^*M\longrightarrow \Lambda_b^{0,q+1}M \end{align*} be the bundle projection obtained by restricting covectors to $T^{0,1}M$. If $\alpha\in C^\infty(M,\Lambda_b^{0,q}M)$ and $\widetilde\alpha$ is any smooth exterior form on $M$ whose restriction to $T^{0,1}M$ equals $\alpha$, then \begin{align*} \bar\partial_b\alpha:=\pi_b(d\widetilde\alpha). \end{align*} [/definition] Formal integrability ensures that this definition is independent of the chosen representative $\widetilde\alpha$. At the top degree $q=n-1$ there is no bundle $\Lambda_b^{0,n}M$ in the CR complex; when a formula needs an outgoing top-degree map, it is interpreted as the zero map from $C^\infty(M,\Lambda_b^{0,n-1}M)$ to the zero space. In local frames, this definition says that $\bar\partial_b$ differentiates coefficients only in the $T^{0,1}M$ directions and adds the correction terms coming from the variation of the coframe. For functions the formula has no coframe correction. [example: Local Formula on Functions] Let $\bar L_1,\dots,\bar L_{n-1}$ be a local frame for $T^{0,1}M$, and let $\bar\omega_1,\dots,\bar\omega_{n-1}$ be the dual coframe, so $\bar\omega_j(\bar L_k)=\delta_{jk}$. For $u\in C^\infty(M)$, the definition of the [exterior derivative](/theorems/1525) on functions gives \begin{align*} du(\bar L_k)=\bar L_k(u). \end{align*} Because $\bar\partial_bu$ is obtained by restricting $du$ to antiholomorphic tangential directions, we have \begin{align*} (\bar\partial_bu)(\bar L_k)=du(\bar L_k)=\bar L_k(u). \end{align*} Now define the tangential one-form \begin{align*} \beta=\sum_{j=1}^{n-1}(\bar L_j u)\bar\omega_j. \end{align*} Evaluating $\beta$ on $\bar L_k$ and using linearity of one-forms gives \begin{align*} \beta(\bar L_k)=\left(\sum_{j=1}^{n-1}(\bar L_j u)\bar\omega_j\right)(\bar L_k)=\sum_{j=1}^{n-1}(\bar L_j u)\bar\omega_j(\bar L_k). \end{align*} By the duality relation $\bar\omega_j(\bar L_k)=\delta_{jk}$, \begin{align*} \beta(\bar L_k)=\sum_{j=1}^{n-1}(\bar L_j u)\delta_{jk}. \end{align*} Since $\delta_{jk}=0$ for $j\ne k$ and $\delta_{kk}=1$, the sum reduces to \begin{align*} \beta(\bar L_k)=\bar L_k u. \end{align*} Thus $\beta(\bar L_k)=(\bar\partial_bu)(\bar L_k)$ for every $k$. A tangential one-form is determined by its values on the local frame $\bar L_1,\dots,\bar L_{n-1}$, so \begin{align*} \bar\partial_b u=\sum_{j=1}^{n-1}(\bar L_j u)\bar\omega_j. \end{align*} Therefore $\bar\partial_bu=0$ holds exactly when every coefficient in this coframe expansion vanishes: \begin{align*} \bar\partial_bu=0 \quad\Longleftrightarrow\quad \bar L_j u=0\text{ for every }1\le j\le n-1. \end{align*} Thus the CR equation on functions is precisely the system of antiholomorphic tangential equations $\bar L_j u=0$. [/example] The local formula makes the operator concrete, but a cohomology theory requires more than a first-order system. If commutators of the antiholomorphic tangent fields produced transverse components, then applying $\bar\partial_b$ twice would create a genuine second-order obstruction instead of zero. The quotient defining CR cohomology is meaningful only when formal integrability removes this obstruction. [quotetheorem:9205] [citeproof:9205] This proposition is the algebraic foundation for the rest of the chapter. Formal integrability is exactly the hypothesis that prevents commutators of antiholomorphic tangent vector fields from producing components outside $T^{0,1}M$. The hypothesis cannot be omitted. On $\mathbb R^5$ with coordinates $(z_1,z_2,t)\in\mathbb C^2\times\mathbb R$, set \begin{align*} \bar L_1=\frac{\partial}{\partial \bar z_1},\qquad \bar L_2=\frac{\partial}{\partial \bar z_2}+\bar z_1\frac{\partial}{\partial t}, \end{align*} and let the proposed antiholomorphic bundle be spanned by $\bar L_1,\bar L_2$. Then $[\bar L_1,\bar L_2]=\partial/\partial t$, which is not in the span of $\bar L_1,\bar L_2$. For $u(z_1,z_2,t)=t$, the coefficient of the would-be square on $(\bar L_1,\bar L_2)$ is \begin{align*} \bar L_1(\bar L_2u)-\bar L_2(\bar L_1u)=1. \end{align*} Thus the proposed tangential operator does not define a complex. Nilpotence is therefore the analytic expression of the CR involutivity condition, not a formal consequence of projecting $d^2=0$. The result also does not solve any equation by itself: it only makes the quotient defining cohomology meaningful, while the vanishing, finite-dimensionality, or regularity of those cohomology groups requires additional geometry and estimates. ## CR Functions and Boundary Values The first meaningful equation in the complex is the homogeneous equation on functions. It asks which smooth functions are annihilated by every antiholomorphic tangential derivative, so it is the boundary analogue of holomorphicity. [definition: CR Function] Let $M$ be a CR manifold. A smooth function $u\in C^\infty(M)$ is a CR function if \begin{align*} \bar\partial_bu=0. \end{align*} The vector space of smooth CR functions on $M$ is denoted $\mathcal O_b(M)$. [/definition] The definition is intrinsic, but its importance comes from the embedded case. If $M$ is the boundary of a domain in a complex manifold, holomorphic functions on the domain automatically satisfy the tangential equations on the boundary whenever the boundary values are smooth. [quotetheorem:9206] [citeproof:9206] The theorem is one-way and uses both hypotheses in an essential way. The neighbourhood hypothesis cannot be replaced by holomorphicity only in the interior without adding boundary regularity: on the unit disc, $F(z)=1/(1-z)$ is holomorphic in the disc but has no smooth boundary value at $z=1$, so the expression $\bar\partial_b(F|_{\partial\mathbb D})$ is not even defined as a smooth boundary equation. The holomorphicity hypothesis is also necessary. On $S^{2n-1}\subset\mathbb C^n$ with $n\ge2$, the smooth boundary function $u(z)=\bar z_1$ is the restriction of a smooth ambient function but is not CR. Indeed, \begin{align*} \bar L=z_2\frac{\partial}{\partial \bar z_1}-z_1\frac{\partial}{\partial \bar z_2} \end{align*} is tangent to the sphere because $\bar L(|z|^2-1)=z_2z_1-z_1z_2=0$, and $\bar L$ is of type $(0,1)$. However $\bar L u=z_2$, which is nonzero on an open subset of the sphere. Conversely, a smooth CR function on an abstract or weakly embedded CR manifold need not be the boundary value of a holomorphic function on a prescribed side. For instance, on nonminimal CR hypersurfaces local CR functions can be constant along CR orbits in ways that obstruct one-sided holomorphic extension. Earlier chapters studied geometric conditions such as Levi convexity and minimality that control this extension problem; here those conditions reappear through analysis of $\bar\partial_b$. [example: CR Functions on the Unit Sphere] Let $n\ge 2$ and let $S^{2n-1}=\partial B^n=\{z\in\mathbb C^n:|z|^2=1\}$. Put \begin{align*} \rho(z)=|z|^2-1=\sum_{\ell=1}^n z_\ell\bar z_\ell-1. \end{align*} For $1\le j<k\le n$, define \begin{align*} \bar L_{jk}=z_j\frac{\partial}{\partial\bar z_k}-z_k\frac{\partial}{\partial\bar z_j}. \end{align*} Since \begin{align*} \frac{\partial\rho}{\partial\bar z_\ell}=z_\ell, \end{align*} we have \begin{align*} \bar L_{jk}\rho=z_j\frac{\partial\rho}{\partial\bar z_k}-z_k\frac{\partial\rho}{\partial\bar z_j}=z_jz_k-z_kz_j=0. \end{align*} Thus $\bar L_{jk}$ is tangent to the level set $\rho=0$, hence tangent to $S^{2n-1}$, and it has type $(0,1)$ because it is a linear combination of the coordinate fields $\partial/\partial\bar z_\ell$. These fields generate all antiholomorphic tangential directions. Indeed, a $(0,1)$ vector \begin{align*} V=\sum_{\ell=1}^n a_\ell\frac{\partial}{\partial\bar z_\ell} \end{align*} is tangent to the sphere exactly when \begin{align*} V\rho=\sum_{\ell=1}^n a_\ell z_\ell=0. \end{align*} On a patch where $z_r\ne0$, this condition gives \begin{align*} a_r=-\frac{1}{z_r}\sum_{\ell\ne r}a_\ell z_\ell. \end{align*} Therefore \begin{align*} \sum_{\ell\ne r}\frac{a_\ell}{z_r}\bar L_{r\ell}=\sum_{\ell\ne r}a_\ell\frac{\partial}{\partial\bar z_\ell}-\frac{1}{z_r}\sum_{\ell\ne r}a_\ell z_\ell\frac{\partial}{\partial\bar z_r}. \end{align*} Using the displayed formula for $a_r$, this becomes \begin{align*} \sum_{\ell\ne r}\frac{a_\ell}{z_r}\bar L_{r\ell}=\sum_{\ell\ne r}a_\ell\frac{\partial}{\partial\bar z_\ell}+a_r\frac{\partial}{\partial\bar z_r}=V. \end{align*} If $F$ is holomorphic in a neighbourhood of $\overline{B^n}$ and $u=F|_{S^{2n-1}}$, then \begin{align*} \frac{\partial F}{\partial\bar z_\ell}=0 \end{align*} for every $\ell$. Hence, on $S^{2n-1}$, \begin{align*} \bar L_{jk}u=\left(z_j\frac{\partial F}{\partial\bar z_k}-z_k\frac{\partial F}{\partial\bar z_j}\right)\bigg|_{S^{2n-1}}=z_j\cdot0-z_k\cdot0=0. \end{align*} Since the vector fields $\bar L_{jk}$ generate $T^{0,1}S^{2n-1}$, this is equivalent to $\bar\partial_bu=0$. Thus every holomorphic function smooth up to the sphere has a CR boundary value. Conversely, by the *Bochner extension theorem for the sphere*, every smooth function $u$ on $S^{2n-1}$ satisfying \begin{align*} \bar L_{jk}u=0 \end{align*} for all $j<k$ is the boundary value of a holomorphic function $U$ on $B^n$ that is smooth up to $S^{2n-1}$. Equivalently, its spherical harmonic expansion contains only terms of antiholomorphic degree $0$: \begin{align*} u\sim\sum_{p\ge0}u_{p,0}, \end{align*} where $u_{p,0}$ is the restriction to the sphere of a holomorphic homogeneous polynomial of degree $p$. Thus the CR equations remove the antiholomorphic spherical modes and leave exactly the boundary traces of holomorphic [power series](/page/Power%20Series) in $z_1,\dots,z_n$. [/example] To see the operator in coordinates, the simplest non-flat model is the Heisenberg hypersurface. It is locally equivalent to the boundary of the Siegel upper half-space and models strongly pseudoconvex boundaries after a first normalisation. [example: Explicit Computation on $H^1$] Let \begin{align*} H^1=\{(z,w)\in\mathbb C^2:\operatorname{Im}(w)=|z|^2\}. \end{align*} Write $z=x+iy$ and parametrize $H^1$ by coordinates $(z,t)\in\mathbb C\times\mathbb R$ through \begin{align*} w=t+i|z|^2=t+iz\bar z. \end{align*} In these coordinates, consider \begin{align*} \bar L=\frac{\partial}{\partial \bar z}-iz\frac{\partial}{\partial t}. \end{align*} Along $H^1$, the restricted coordinate functions are $z$, $\bar z$, and $w=t+iz\bar z$. Since $t$ and $z$ are independent of $\bar z$, and since $\partial\bar z/\partial\bar z=1$, we have \begin{align*} \frac{\partial}{\partial\bar z}(t+iz\bar z)=0+i\left(\frac{\partial z}{\partial\bar z}\bar z+z\frac{\partial\bar z}{\partial\bar z}\right)=0+i(0\cdot\bar z+z\cdot1)=iz. \end{align*} Also, \begin{align*} \frac{\partial}{\partial t}(t+iz\bar z)=1+i\left(\frac{\partial z}{\partial t}\bar z+z\frac{\partial\bar z}{\partial t}\right)=1+i(0\cdot\bar z+z\cdot0)=1. \end{align*} Therefore \begin{align*} \bar L(w)=iz-iz\cdot1=0. \end{align*} Similarly, \begin{align*} \bar L(z)=\frac{\partial z}{\partial\bar z}-iz\frac{\partial z}{\partial t}=0-iz\cdot0=0. \end{align*} Thus $\bar L$ annihilates the restrictions of the ambient holomorphic coordinates $z$ and $w$, so it is an antiholomorphic tangential vector field on $H^1$. Since $H^1$ has CR dimension $1$, its antiholomorphic tangent bundle is locally generated by $\bar L$. Hence a smooth function $u\in C^\infty(H^1)$ satisfies $\bar\partial_bu=0$ exactly when its derivative in this generator vanishes: \begin{align*} \bar\partial_bu=0 \quad\Longleftrightarrow\quad \bar L u=0. \end{align*} Substituting the formula for $\bar L$ gives \begin{align*} \bar L u=\frac{\partial u}{\partial\bar z}-iz\frac{\partial u}{\partial t}. \end{align*} Therefore the CR equation on $H^1$ is \begin{align*} \bar\partial_bu=0 \quad\Longleftrightarrow\quad \frac{\partial u}{\partial\bar z}-iz\frac{\partial u}{\partial t}=0. \end{align*} For $u(z,t)=z$, the coordinate derivatives are \begin{align*} \frac{\partial z}{\partial\bar z}=0 \end{align*} and \begin{align*} \frac{\partial z}{\partial t}=0. \end{align*} Thus \begin{align*} \bar L(z)=0-iz\cdot0=0. \end{align*} So $z$ is a CR function on $H^1$. For $u(z,t)=\bar z$, the coordinate derivatives are \begin{align*} \frac{\partial\bar z}{\partial\bar z}=1 \end{align*} and \begin{align*} \frac{\partial\bar z}{\partial t}=0. \end{align*} Thus \begin{align*} \bar L(\bar z)=1-iz\cdot0=1. \end{align*} Therefore $\bar z$ is not CR. The Heisenberg CR equation is not merely the flat equation $\partial u/\partial\bar z=0$; the additional $t$-derivative term records the non-flat tangential geometry of the hypersurface. [/example] This computation shows why the equation is not merely the flat equation $\partial u/\partial\bar z=0$. The extra $t$-derivative encodes the Levi geometry of the hypersurface and is responsible for the hypoelliptic rather than elliptic behaviour of the boundary system. ## Kohn-Rossi Cohomology Once $\bar\partial_b^2=0$ is available, the next question is what obstructs solving $\bar\partial_bu=f$. The obstruction is cohomological: a right-hand side must be closed, and two solutions differing by an exact term represent the same cohomological information. [definition: Kohn-Rossi Cohomology] For a CR manifold $M$, the Kohn-Rossi cohomology groups are \begin{align*} H_b^{0,q}(M):=\frac{\ker\{\bar\partial_b:C^\infty(M,\Lambda^{0,q}_bM)\to C^\infty(M,\Lambda^{0,q+1}_bM)\}}{\operatorname{im}\{\bar\partial_b:C^\infty(M,\Lambda^{0,q-1}_bM)\to C^\infty(M,\Lambda^{0,q}_bM)\}}. \end{align*} For $q=0$, this is interpreted as $H_b^{0,0}(M)=\ker(\bar\partial_b:C^\infty(M)\to C^\infty(M,\Lambda_b^{0,1}M))$. For $q=n-1$, the numerator uses the zero outgoing map from top-degree tangential forms. [/definition] Thus $H_b^{0,0}(M)$ is the space of CR functions. For $q\ge1$, the group $H_b^{0,q}(M)$ measures the failure of local or global solvability of the equation $\bar\partial_bu=f$ for closed tangential forms $f$. [example: Meaning of $H_b^{0,1}$] Let $f\in C^\infty(M,\Lambda_b^{0,1}M)$ satisfy $\bar\partial_b f=0$. In degree $1$, the Kohn-Rossi group is the quotient \begin{align*} H_b^{0,1}(M)=\frac{\ker\{\bar\partial_b:C^\infty(M,\Lambda_b^{0,1}M)\to C^\infty(M,\Lambda_b^{0,2}M)\}}{\operatorname{im}\{\bar\partial_b:C^\infty(M)\to C^\infty(M,\Lambda_b^{0,1}M)\}}. \end{align*} Because $\bar\partial_b f=0$, the form $f$ belongs to the kernel in the numerator: \begin{align*} f\in \ker\{\bar\partial_b:C^\infty(M,\Lambda_b^{0,1}M)\to C^\infty(M,\Lambda_b^{0,2}M)\}. \end{align*} Thus $f$ determines the cohomology class \begin{align*} [f]=f+\operatorname{im}\{\bar\partial_b:C^\infty(M)\to C^\infty(M,\Lambda_b^{0,1}M)\}. \end{align*} The zero element of the quotient is the coset of the zero form: \begin{align*} [0]=0+\operatorname{im}\{\bar\partial_b:C^\infty(M)\to C^\infty(M,\Lambda_b^{0,1}M)\}. \end{align*} Therefore $[f]=0$ means that the two cosets are equal: \begin{align*} f+\operatorname{im}\bar\partial_b=0+\operatorname{im}\bar\partial_b. \end{align*} For any quotient vector space $V/W$, two cosets $v+W$ and $v'+W$ are equal exactly when $v-v'\in W$. Applying this with $v=f$, $v'=0$, and $W=\operatorname{im}\bar\partial_b$, we get \begin{align*} [f]=0 \quad\Longleftrightarrow\quad f-0\in \operatorname{im}\bar\partial_b. \end{align*} Since $f-0=f$, this is \begin{align*} [f]=0 \quad\Longleftrightarrow\quad f\in \operatorname{im}\{\bar\partial_b:C^\infty(M)\to C^\infty(M,\Lambda_b^{0,1}M)\}. \end{align*} Unpacking the definition of the image gives \begin{align*} f\in \operatorname{im}\bar\partial_b \quad\Longleftrightarrow\quad \text{there exists }u\in C^\infty(M)\text{ such that }\bar\partial_bu=f. \end{align*} Thus a closed tangential $(0,1)$-form represents the zero class exactly when it is $\bar\partial_b$ of a smooth function, while a nonzero class is exactly the obstruction to solving the equation $\bar\partial_bu=f$. [/example] The example explains cohomology as a solvability obstruction on the boundary; the next question is how this obstruction relates to ordinary Dolbeault obstructions on the two sides of the boundary. This motivates the Andreotti-Hill exact sequence, whose role is to place $H_b^{0,q}(M)$ inside a larger ambient cohomological comparison. [quotetheorem:9207] The hypotheses in the statement are not cosmetic: the two sides of $M$ are placed in the same smooth up-to-boundary category, so restriction and connecting maps are defined on the cohomology groups appearing in the sequence. A concrete failure occurs already for $X=\mathbb C$, $\Omega=\{z:|z|<1\}$, and $M=S^1$ if the exterior term is taken to mean unrestricted holomorphic functions on $\mathbb C\setminus\overline{\Omega}$. The function \begin{align*} f(z)=\frac{1}{z-1} \end{align*} is holomorphic on the exterior region, but it has a pole at the boundary point $1\in S^1$. Its boundary jump is not a smooth tangential form on $M$, so the boundary map is not defined on this element. The theorem above avoids this mismatch by fixing the smooth up-to-boundary cohomology groups before forming the sequence. In this course the sequence is used as a comparison device showing that boundary cohomology is part of the same exact algebra relating the inside and outside complex geometry. [remark: Exactness as Solvability] Exactness says that a boundary cohomology class vanishes after passing through one map precisely when it comes from the preceding ambient cohomology group. In analytic terms, this translates extension and jump problems across $M$ into statements about solving $\bar\partial$ on the adjacent regions. [/remark] The long exact sequence also explains why strongly pseudoconvex boundaries often have finite-dimensional and rigid boundary cohomology. Under positivity assumptions, interior Dolbeault cohomology is controlled by vanishing and finiteness theorems, and the boundary groups inherit this structure through exactness. ## The Kohn Laplacian and Subelliptic Estimates Cohomology defines the obstruction spaces, but analysis asks whether solutions have estimates. The boundary operator is not elliptic, because it differentiates only in complex tangential directions; the missing real transverse direction must be recovered through commutators encoded by the Levi form. Fix a compact orientable CR manifold $M$ with a smooth density and a Hermitian metric on tangential forms. These choices define Hilbert spaces $L^2(M,\Lambda_b^{0,q}M)$ for each degree. For $0\le q\le n-2$, the operator \begin{align*} \bar\partial_b:L^2(M,\Lambda_b^{0,q}M)\supset \operatorname{Dom}(\bar\partial_b)\longrightarrow L^2(M,\Lambda_b^{0,q+1}M) \end{align*} is first defined on smooth forms and then closed in $L^2$. Its Hilbert-space adjoint at degree $q$ is \begin{align*} \bar\partial_b^*:L^2(M,\Lambda_b^{0,q+1}M)\supset \operatorname{Dom}(\bar\partial_b^*)\longrightarrow L^2(M,\Lambda_b^{0,q}M), \end{align*} characterised by $(\bar\partial_b u,v)_{L^2}=(u,\bar\partial_b^*v)_{L^2}$ for smooth compactly supported local representatives, and then by closure. Equivalently, on $(0,q)$-forms in the Kohn Laplacian, $\bar\partial_b^*$ denotes the adjoint of the preceding map $\bar\partial_b:L^2(M,\Lambda_b^{0,q-1}M)\to L^2(M,\Lambda_b^{0,q}M)$, with no preceding term when $q=0$. The second-order operator controlling the complex is the Kohn Laplacian. [definition: Kohn Laplacian] On smooth tangential $(0,q)$-forms, the Kohn Laplacian is the differential operator \begin{align*} \Box_b:C^\infty(M,\Lambda_b^{0,q}M)\longrightarrow C^\infty(M,\Lambda_b^{0,q}M),\qquad \Box_b = \bar\partial_b\bar\partial_b^*+\bar\partial_b^*\bar\partial_b. \end{align*} As an unbounded operator on $L^2(M,\Lambda_b^{0,q}M)$, its natural domain is the set of forms $u$ for which $u$, $\bar\partial_bu$, $\bar\partial_b^*u$, and $\Box_bu$ lie in the corresponding $L^2$ spaces in the weak sense. Its quadratic form is the map \begin{align*} Q_b:\operatorname{Dom}(\bar\partial_b)\cap\operatorname{Dom}(\bar\partial_b^*)\longrightarrow [0,\infty),\qquad Q_b[u,u]=\|\bar\partial_bu\|_{L^2}^2+\|\bar\partial_b^*u\|_{L^2}^2. \end{align*} [/definition] The Kohn Laplacian packages the two first-order equations into one nonnegative operator, but its quadratic form does not by itself say how many derivatives are controlled. This motivates the following estimate, which records the fractional regularity gain that replaces ellipticity on a CR boundary. [definition: Subelliptic Estimate for $\Box_b$] Let $H^\varepsilon(M,\Lambda_b^{0,q}M)$ denote the Sobolev space of tangential $(0,q)$-forms of order $\varepsilon$, defined using any finite smooth atlas, [partition of unity](/page/Partition%20of%20Unity), and local tangential frame; on compact $M$ the resulting norm is independent of these choices up to equivalence. A subelliptic estimate of order $\varepsilon>0$ at degree $q$ is an inequality of the form \begin{align*} \|u\|_{H^\varepsilon(M,\Lambda_b^{0,q}M)}^2 \le C\bigl(Q_b[u,u]+\|u\|_{L^2(M,\Lambda_b^{0,q}M)}^2\bigr) \end{align*} for all smooth tangential $(0,q)$-forms $u$, with a constant $C>0$ independent of $u$. [/definition] This estimate means that the tangential Cauchy-Riemann equations control more derivatives than their visible first-order directions suggest. The natural question is which Levi geometries force such control; Kohn's condition $Y(q)$ gives the degree-dependent sign hypothesis used in the standard theorem. [definition: Kohn Condition $Y(q)$] Let $M$ be a CR hypersurface of CR dimension $m$, with a chosen coorientation so that the Levi form has signed eigenvalues. For $0\le q\le m$, the condition $Y(q)$ holds if, at every point of $M$, the Levi form has either at least $m-q+1$ positive eigenvalues or at least $q+1$ negative eigenvalues. [/definition] The two alternatives are sign-specific; the condition is not a demand for $\max(q+1,m-q+1)$ eigenvalues of an arbitrary common sign. In particular, a strictly pseudoconvex hypersurface, whose Levi form is positive definite after choosing the pseudoconvex coorientation, satisfies $Y(q)$ for $1\le q\le m$ but not for $q=0$. The condition isolates the degrees where the Levi form supplies enough commutator positivity to compensate for the missing real direction. This motivates the analytic theorem that turns the geometric sign condition into a Sobolev estimate. [quotetheorem:9208] The analytic mechanism is microlocal decomposition together with positivity of the Levi form in the cones where the principal symbol of $\Box_b$ is degenerate. Compactness and the global smooth structure are needed to pass from local microlocal estimates to a uniform global constant, while the hypersurface and $Y(q)$ assumptions identify the Levi directions that recover the missing characteristic derivative. When $Y(q)$ fails, the estimate can fail even on smooth compact examples: on a Levi-flat model such as $M=\mathbb C_z\times\mathbb R_t$ with CR vector field $\bar L=\partial/\partial\bar z$, choose a nonzero cutoff $\chi\in C_c^\infty(\mathbb C_z\times\mathbb R_t)$ and set $u_k(z,t)=\chi(z,t)e^{ikt}$. These test functions have large $H^\varepsilon$ norm in the missing $t$ direction, while $\bar L u_k$ does not see the $t$ oscillation except through the cutoff. This model explains why the commutators of CR vector fields must generate transverse control. The theorem is therefore a geometric sufficient condition for subellipticity, not a general property of every CR boundary. [remark: Why the Estimate Is Subelliptic] An elliptic second-order operator would control one full derivative in the $H^1$ norm from its quadratic form. The Kohn Laplacian loses ellipticity in the characteristic direction, so the theorem gives only $H^\varepsilon$ control for some $0<\varepsilon\le1$. The gain is still strong enough to imply hypoellipticity in many geometric settings. [/remark] The remark explains the strength and limitation of the estimate. Once a subelliptic bound is available, the analytic question is whether the formal quotient spaces $H_b^{0,q}$ behave like finite-dimensional obstruction spaces rather than uncontrolled spaces of weak solutions. Closed range, Hodge representatives, and finite-dimensional harmonic spaces are the functional-analytic consequences that make this possible. [quotetheorem:9209] [citeproof:9209] These consequences close the circle between geometry, cohomology, and analysis, but they depend on the estimate rather than on the formal identity $\bar\partial_b^2=0$ alone. A compact Levi-flat model shows what can fail without subellipticity. Let \begin{align*} M=(\mathbb C/\Lambda)\times S^1 \end{align*} with CR bundle generated along the complex torus factor by $\partial/\partial\bar z$. Then $\bar\partial_b$ differentiates only in the elliptic-curve direction and does not see the $S^1$ variable. Every smooth function of the $S^1$ variable alone is $\Box_b$-harmonic in degree $0$, so $\mathcal H_b^{0,0}(M)$ is infinite-dimensional. This contradicts the finite-dimensional conclusion of the theorem and also shows why no positive-order estimate of the displayed kind can hold in degree $0$: high-frequency functions $e^{ikt}$ have zero graph norm but growing $H^\varepsilon$ norm. For nonclosed range, Rossi's nonembeddable perturbations of the standard CR structure on $S^3$ give a classical compact example in which $H_b^{0,1}$ is non-Hausdorff, equivalently the relevant $\bar\partial_b$ range is not closed. The theorem also does not assert vanishing of $H_b^{0,q}(M)$; it gives a finite-dimensional Hodge representation of whatever obstruction space remains. The formal boundary equations set up the extension problem, but they do not yet explain when a CR function is truly the boundary value of a holomorphic one. The Lewy extension theorem answers that question in the simplest nontrivial cases and shows how geometry can force [analytic continuation](/page/Analytic%20Continuation) across the boundary. # 5. The Lewy Extension Theorem These notes follow the course progression through several complex variables, CR geometry, and boundary behaviour of holomorphic functions. The guiding prerequisites are the Cauchy-Riemann equations, basic distribution theory, the Levi form from the preceding chapter, and the $\bar\partial$ complex on domains in $\mathbb C^n$. The aim of this chapter is to connect those ingredients to Lewy's extension theorem: strict pseudoconvexity turns the tangential CR equations on a boundary into genuine holomorphic extension on the pseudoconvex side. The same discussion also explains why this is a theorem with real content, since Lewy's and Henkin's examples show that smooth CR equations may be locally non-solvable or non-extendable when the geometric hypotheses are weakened. ## Lewy's Operator and the First Obstruction The guiding problem is to understand why the CR equations on a hypersurface are not just harmless restrictions of the ambient Cauchy-Riemann equations. Lewy's 1956 example isolates this difficulty in the smallest model: a hypersurface in $\mathbb C^2$ whose CR vector field has a commutator pointing in the missing real direction. This model explains both the possibility of extension under a sign condition and the possibility of non-solvability for a related inhomogeneous equation. [definition: Heisenberg Hypersurface] The Heisenberg hypersurface $H^1 \subset \mathbb C^2$ is \begin{align*} H^1=\{(z,w)\in \mathbb C^2: \operatorname{Im}(w)=|z|^2\}. \end{align*} [/definition] Writing $w=t+i|z|^2$ with $t\in \mathbb R$, the standard CR operator of type $(0,1)$ on a coordinate neighbourhood $U\subset H^1$ is the linear differential operator \begin{align*} L_H:C^\infty(U;\mathbb C)\to C^\infty(U;\mathbb C),\qquad L_Hu=\frac{\partial u}{\partial \bar z}-i z\frac{\partial u}{\partial t}. \end{align*} The field $L_H$ is tangential because differentiating the parametrisation $(z,t)\mapsto (z,t+i|z|^2)$ in the anti-holomorphic direction gives $\partial_{\bar z}(t+i|z|^2)=iz$, which is cancelled by the term $-iz\partial_t$. A smooth function $u$ on $H^1$ is CR exactly when $L_Hu=0$, so the local boundary problem becomes a first-order PDE with complex coefficients. Lewy's original non-solvable operator below is the sign-reversed model; it is obtained from this one by reversing the $t$-orientation, so the analytic obstruction is the same while the displayed Heisenberg parametrisation keeps the correct sign. [example: Boundary Values of Holomorphic Functions on the Heisenberg Hypersurface] Let $F$ be holomorphic near a point of $H^1\subset\mathbb C^2$, and parametrize the boundary by $(z,t)\mapsto (z,w(z,t))$, where \begin{align*} w(z,t)=t+i|z|^2=t+iz\bar z. \end{align*} Set \begin{align*} u(z,t)=F(z,w(z,t)). \end{align*} We show that this boundary value satisfies $L_Hu=0$. On the real coordinate chart, $z$ and $\bar z$ are treated as independent Wirtinger variables. Since \begin{align*} \frac{\partial z}{\partial\bar z}=0,\quad \frac{\partial\bar z}{\partial\bar z}=1,\quad \frac{\partial w}{\partial\bar z}=\frac{\partial}{\partial\bar z}(t+iz\bar z)=iz,\quad \frac{\partial\bar w}{\partial\bar z}=\frac{\partial}{\partial\bar z}(t-iz\bar z)=-iz, \end{align*} the chain rule gives \begin{align*} \frac{\partial u}{\partial\bar z}=F_z(z,w)\cdot 0+F_w(z,w)\cdot iz+F_{\bar z}(z,w)\cdot 1+F_{\bar w}(z,w)\cdot(-iz). \end{align*} Because $F$ is holomorphic in the ambient variables $(z,w)$, its antiholomorphic derivatives vanish: \begin{align*} F_{\bar z}(z,w)=0,\quad F_{\bar w}(z,w)=0. \end{align*} Thus \begin{align*} \frac{\partial u}{\partial\bar z}=0+izF_w(z,w)+0+0=izF_w(z,w)=izF_w(z,t+i|z|^2). \end{align*} For the $t$-derivative, the coordinate derivatives are \begin{align*} \frac{\partial z}{\partial t}=0,\quad \frac{\partial\bar z}{\partial t}=0,\quad \frac{\partial w}{\partial t}=\frac{\partial}{\partial t}(t+iz\bar z)=1,\quad \frac{\partial\bar w}{\partial t}=\frac{\partial}{\partial t}(t-iz\bar z)=1. \end{align*} Applying the chain rule again, \begin{align*} \frac{\partial u}{\partial t}=F_z(z,w)\cdot 0+F_w(z,w)\cdot 1+F_{\bar z}(z,w)\cdot 0+F_{\bar w}(z,w)\cdot 1. \end{align*} Using $F_{\bar z}=F_{\bar w}=0$ gives \begin{align*} \frac{\partial u}{\partial t}=0+F_w(z,w)+0+0=F_w(z,w)=F_w(z,t+i|z|^2). \end{align*} Substituting these two identities into the Heisenberg CR operator, \begin{align*} L_Hu=\frac{\partial u}{\partial\bar z}-iz\frac{\partial u}{\partial t}=izF_w(z,t+i|z|^2)-izF_w(z,t+i|z|^2)=0. \end{align*} Therefore every ambient holomorphic function restricts to a function on $H^1$ annihilated by $L_H$, so $L_H$ is exactly the tangential Cauchy-Riemann operator for this parametrized Heisenberg boundary. [/example] The example shows that $L_H$ is the correct homogeneous CR operator for this orientation. The next test is whether the corresponding sign-equivalent inhomogeneous equation has local solutions for all smooth data, since such solvability would allow arbitrary defects in an almost-CR function to be corrected. Lewy's non-solvable equation gives the negative answer and exposes the analytic obstruction carried by the commutator of the complex field and its conjugate. [quotetheorem:9210] [citeproof:9210] This theorem is a PDE obstruction, not yet an extension theorem. The smoothness of $f$ matters because the failure is not caused by a rough right-hand side; even $C^\infty$ data may fail to be locally in the range of $L$. The distributional formulation is essential: if non-solvability were asserted only among smooth functions, the obstruction could be dismissed as a regularity defect, while Lewy's statement rules out even weak local solutions. Locality is essential in a different direction: the theorem is stronger than any global boundary obstruction, since shrinking $V$ never repairs the equation. The hypothesis that the coefficient of $\partial_t$ is tied to the complex variable is also essential; for the constant-coefficient operator $\partial_{\bar z}$, the [Cauchy transform](/page/Cauchy%20Transform) solves $\partial_{\bar z}u=f$ locally for smooth $f$, so the bracket-generated missing direction is the new phenomenon. The theorem does not say that the sign-equivalent homogeneous CR equation has no solutions; restrictions of holomorphic functions in the Heisenberg model still solve the homogeneous equation. Instead, it says that the inhomogeneous correction problem can fail, which is exactly the kind of correction one would try to use when turning an almost-CR boundary function into a genuine boundary value. We next record the boundary non-extension phenomenon in the language of CR functions. [quotetheorem:9211] [citeproof:9211] Lewy's example is deliberately local and pathological from the extension point of view. Smoothness of $M$ and $u$ is not enough: the Levi-flat model $M=\{\operatorname{Im} z_n=0\}$ already has smooth CR functions $a(\operatorname{Re} z_n)h(z')$ that fail to extend when $a$ is not real-analytic, and Lewy's construction gives a sharper two-sided failure. The absence of a strict pseudoconvex side is essential; on the unit sphere, the Lewy extension theorem below gives extension for every smooth CR function into the ball. The CR hypothesis is also essential in the positive direction: a general smooth boundary function on the sphere with antiholomorphic Fourier modes is not the trace of a holomorphic function. The theorem does not claim that every CR function on such a hypersurface fails to extend; restrictions of ambient holomorphic functions still give many CR functions that extend. Rather, it produces one smooth CR function whose boundary behaviour is incompatible with either side, showing that the CR equations alone are not enough. The next section imposes a strong sign condition on the Levi form; the sign supplies exactly the missing analytic control. ## Strict Pseudoconvexity and One-Sided Extension The central question is now positive: if $M$ is the boundary of a strictly pseudoconvex domain, does every smooth CR function on $M$ arise as the boundary value of a holomorphic function inside? The theorem says yes. It is one-sided because the Levi sign selects the pseudoconvex side and gives no symmetric information about the exterior. [definition: Smooth CR Function on a Hypersurface] Let $M\subset \mathbb C^n$ be a smooth real hypersurface. A function $u\in C^\infty(M;\mathbb C)$ is a smooth CR function if \begin{align*} \bar Z u=0 \end{align*} for every smooth local section $\bar Z$ of $T^{0,1}M$, where each such section acts as a first-order linear differential operator \begin{align*} \bar Z:C^\infty(U;\mathbb C)\longrightarrow C^\infty(U;\mathbb C) \end{align*} on every coordinate neighbourhood $U\subset M$ on which $\bar Z$ is defined. [/definition] This definition packages the tangential Cauchy-Riemann equations intrinsically. On a boundary $M=\partial \Omega$, the question is whether such a tangential object is the trace of an ambient holomorphic function on $\Omega$. [quotetheorem:9212] [citeproof:9212] This is the positive counterpart to Lewy's obstruction. The strict pseudoconvex hypothesis supplies a preferred side and enough positivity in the Levi form to solve the correction problem that fails in the non-solvable model. The conclusion is one-sided: it produces holomorphic extension into the pseudoconvex domain, not across both sides of the hypersurface. Smoothness matters because the CR equations and boundary correction are differentiated repeatedly; with finite boundary regularity one expects correspondingly finite extension regularity. The theorem is therefore not merely a statement about the sphere, but the local mechanism that lets CR boundary data become holomorphic interior data throughout the later regularity theory. [example: Extension from the Sphere to the Ball] Let $M=S^{2n-1}=\partial B(0,1)\subset \mathbb C^n$ with $n\ge 2$, and let $u\in C^\infty(S^{2n-1})$ be CR. We verify that the sphere is strictly pseudoconvex and then apply the extension theorem. For the defining function \begin{align*} \rho(z)=|z|^2-1=\sum_{m=1}^n z_m\bar z_m-1, \end{align*} the Wirtinger variables $z_m$ and $\bar z_m$ are treated independently, so \begin{align*} \frac{\partial}{\partial z_j}(z_m\bar z_m)=\frac{\partial z_m}{\partial z_j}\bar z_m+z_m\frac{\partial\bar z_m}{\partial z_j}. \end{align*} Since $\partial z_m/\partial z_j=\delta_{jm}$ and $\partial\bar z_m/\partial z_j=0$, this becomes \begin{align*} \frac{\partial}{\partial z_j}(z_m\bar z_m)=\delta_{jm}\bar z_m+z_m\cdot 0=\delta_{jm}\bar z_m. \end{align*} Therefore \begin{align*} \frac{\partial \rho}{\partial z_j}=\sum_{m=1}^n \delta_{jm}\bar z_m=\bar z_j. \end{align*} Differentiating this identity with respect to $\bar z_k$ gives \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\bar z_k}=\frac{\partial}{\partial\bar z_k}(\bar z_j)=\delta_{jk}. \end{align*} A $(1,0)$ vector at $z$ has the form $\xi=\sum_{j=1}^n \xi_j\partial/\partial z_j$. It is complex tangent to $S^{2n-1}$ exactly when $\xi\rho=0$. Using the first derivatives of $\rho$, \begin{align*} \xi\rho=\sum_{j=1}^n \xi_j\frac{\partial\rho}{\partial z_j}=\sum_{j=1}^n \xi_j\bar z_j=\sum_{j=1}^n \bar z_j\xi_j. \end{align*} Thus $\xi\in T_z^{1,0}S^{2n-1}$ exactly when \begin{align*} \sum_{j=1}^n \bar z_j\xi_j=0. \end{align*} For such a tangent vector, the Levi form is \begin{align*} \mathcal L_\rho(z;\xi)=\sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial\bar z_k}\xi_j\overline{\xi_k}. \end{align*} Substituting $\partial^2\rho/\partial z_j\partial\bar z_k=\delta_{jk}$ gives \begin{align*} \mathcal L_\rho(z;\xi)=\sum_{j,k=1}^n \delta_{jk}\xi_j\overline{\xi_k}. \end{align*} The Kronecker delta leaves only the terms with $j=k$, so \begin{align*} \mathcal L_\rho(z;\xi)=\sum_{j=1}^n \xi_j\overline{\xi_j}=\sum_{j=1}^n |\xi_j|^2. \end{align*} If $\xi\ne 0$, then some $\xi_j\ne 0$, hence $|\xi_j|^2>0$ and $\sum_{m=1}^n|\xi_m|^2>0$. Therefore the unit ball is strictly pseudoconvex. By the *Lewy Extension Theorem*, there is a function \begin{align*} F\in\mathcal O(B(0,1))\cap C^\infty(\overline{B(0,1)}) \end{align*} such that \begin{align*} F|_{S^{2n-1}}=u. \end{align*} The extension is unique. If $F_1$ and $F_2$ are two such extensions, set $G=F_1-F_2$. Then $G$ is holomorphic on $B(0,1)$ and continuous on $\overline{B(0,1)}$. On the boundary, \begin{align*} G|_{S^{2n-1}}=F_1|_{S^{2n-1}}-F_2|_{S^{2n-1}}=u-u=0. \end{align*} The maximum principle applied to $G$ gives \begin{align*} \sup_{B(0,1)}|G|\le \sup_{S^{2n-1}}|G|. \end{align*} Since $G|_{S^{2n-1}}=0$, the right-hand side is $0$, so \begin{align*} \sup_{B(0,1)}|G|\le 0. \end{align*} Because $|G|\ge 0$ pointwise, this implies $|G|=0$ on $B(0,1)$, hence $G=0$ and $F_1=F_2$. The tangential CR fields also show explicitly why anti-holomorphic boundary modes are excluded. For $1\le j,k\le n$, define \begin{align*} \bar L_{jk}=z_j\frac{\partial}{\partial\bar z_k}-z_k\frac{\partial}{\partial\bar z_j}. \end{align*} Since \begin{align*} \frac{\partial\rho}{\partial\bar z_k}=z_k, \end{align*} we get \begin{align*} \bar L_{jk}\rho=z_j\frac{\partial\rho}{\partial\bar z_k}-z_k\frac{\partial\rho}{\partial\bar z_j}=z_jz_k-z_kz_j=0. \end{align*} Thus each $\bar L_{jk}$ is tangent to the sphere. For a holomorphic monomial \begin{align*} z^\alpha=z_1^{\alpha_1}\cdots z_n^{\alpha_n}, \end{align*} there is no dependence on any $\bar z_\ell$. Hence \begin{align*} \frac{\partial z^\alpha}{\partial\bar z_k}=0. \end{align*} Similarly, \begin{align*} \frac{\partial z^\alpha}{\partial\bar z_j}=0. \end{align*} Substituting these two derivatives into $\bar L_{jk}$ gives \begin{align*} \bar L_{jk}(z^\alpha)=z_j\cdot 0-z_k\cdot 0=0. \end{align*} So holomorphic monomials satisfy the tangential CR equations. By contrast, take the anti-holomorphic coordinate function $\bar z_k$ with $j\ne k$. The independent-variable rules give \begin{align*} \frac{\partial\bar z_k}{\partial\bar z_k}=1. \end{align*} Since $j\ne k$, they also give \begin{align*} \frac{\partial\bar z_k}{\partial\bar z_j}=0. \end{align*} Therefore \begin{align*} \bar L_{jk}(\bar z_k)=z_j\frac{\partial\bar z_k}{\partial\bar z_k}-z_k\frac{\partial\bar z_k}{\partial\bar z_j}=z_j\cdot 1-z_k\cdot 0=z_j. \end{align*} This is not identically zero on $S^{2n-1}$, because at the point with $z_j=1$ and all other coordinates equal to $0$, the value is \begin{align*} \bar L_{jk}(\bar z_k)=1. \end{align*} Thus holomorphic monomials satisfy the tangential CR equations, while basic anti-holomorphic modes do not. The extension theorem identifies the remaining smooth CR boundary functions on the sphere exactly as boundary traces of holomorphic functions in the ball. [/example] This example is the higher-dimensional analogue of extending positive Fourier modes from the unit circle, but there is an important difference. In one complex variable a smooth boundary function on the circle must satisfy a global moment condition to be the boundary value of a holomorphic function. In several variables, strict pseudoconvexity and the CR equations enforce the compatible boundary structure locally and globally. ## The Bochner-Hartogs Reduction to a Compactly Supported Equation The proof just sketched rests on an analytic engine: solve a $\bar\partial$ equation whose data is supported away from the boundary on the chosen side. The natural question is why this reduction is possible and where pseudoconvexity enters. The Bochner-Hartogs viewpoint answers by converting boundary extension into a correction problem for an arbitrary smooth extension. [definition: Compactly Supported One-Sided Data] Let $\Omega\subset \mathbb C^n$ be a smooth domain and let $f$ be a smooth $(0,1)$-form on $\Omega$. The form $f$ is compactly supported on the interior side if there is a compact set $K\subset\Omega$ such that $f=0$ on $\Omega\setminus K$. [/definition] Compact support away from $M$ is valuable because the boundary values are then untouched by any correction term that vanishes near $M$. The Bochner-Hartogs reduction is needed to manufacture exactly this situation: it starts with arbitrary CR boundary data and produces an interior $\bar\partial$-closed error whose support is away from the boundary, so the extension problem becomes a solvable correction problem. [quotetheorem:9213] [citeproof:9213] This reduction explains the role of dimension $n\ge2$. Compactly supported $\bar\partial$ solvability is a Hartogs-type phenomenon and has no direct analogue for arbitrary planar domains: on an annulus in $\mathbb C$, the equation $\partial_{\bar z}v=f$ with compactly supported $v$ imposes contour moment conditions, so compact support cannot be expected for arbitrary closed data. The CR hypothesis is essential because without it the boundary error of an arbitrary extension has a tangential component on $M$ that no interior correction vanishing near $M$ can change; a smooth non-CR function on the sphere gives a concrete failure. Smooth boundary is used in the collar and Taylor construction; a domain with corners need not have a smooth CR bundle or a smooth normal Taylor expansion. Strict pseudoconvexity is also not used in the formal Taylor step alone, as the Levi-flat product boundary has the same local collar smoothness but admits non-extendable CR data. The theorem is only a reduction: it does not itself prove that the compactly supported equation can be solved. The sign of the Levi form is what turns this Hartogs phenomenon into a boundary extension theorem. [quotetheorem:9214] [citeproof:9214] The extension theorem is therefore not a purely formal consequence of CR geometry. It is a boundary regularity and support theorem for the $\bar\partial$ complex, and strict pseudoconvexity is the geometric hypothesis that makes the analytic estimate work. The strict sign cannot be replaced by arbitrary smoothness of the boundary: the Levi-flat boundary $\{\operatorname{Im} z_n=0\}$ has smooth non-real-analytic transverse CR data, and Henkin's degenerate mixed-sign models have support-controlled tangential cohomology obstructions. The assumption $n\ge2$ is also essential for the support statement; in one variable compactly supported solutions to $\partial_{\bar z}v=f$ are obstructed by contour integrals and residues. The compact-support hypothesis on $f$ separates this theorem from the boundary problem: it supplies interior data for a PDE, while the Bochner-Hartogs reduction supplies the special $f$ attached to a CR boundary value. The theorem does not by itself extend boundary functions; it gives the invisible correction term that the reduction needs. [example: Correcting an Almost Holomorphic Extension] Suppose $u\in C^\infty(\partial\Omega)$ is CR on a strictly pseudoconvex boundary, and let $\tilde u\in C^\infty(\overline\Omega)$ satisfy \begin{align*} \tilde u|_{\partial\Omega}=u. \end{align*} Assume its antiholomorphic defect \begin{align*} f=\bar\partial\tilde u \end{align*} is supported in a compact set $K\subset\Omega$. By *[Compactly Supported Dolbeault Solvability on Strictly Pseudoconvex Domains](/theorems/9214)*, choose $v\in C_c^\infty(\Omega)$ with \begin{align*} \bar\partial v=f. \end{align*} Set \begin{align*} F=\tilde u-v. \end{align*} The operator $\bar\partial$ is linear on smooth functions, so \begin{align*} \bar\partial F=\bar\partial(\tilde u-v). \end{align*} Applying linearity to the difference gives \begin{align*} \bar\partial(\tilde u-v)=\bar\partial\tilde u-\bar\partial v. \end{align*} Using $f=\bar\partial\tilde u$, this becomes \begin{align*} \bar\partial\tilde u-\bar\partial v=f-\bar\partial v. \end{align*} Using $\bar\partial v=f$, we get \begin{align*} f-\bar\partial v=f-f. \end{align*} Since subtraction of equal $(0,1)$-forms gives the zero form, \begin{align*} f-f=0. \end{align*} Therefore \begin{align*} \bar\partial F=0. \end{align*} Thus $F$ is holomorphic on $\Omega$. It remains to check that the correction did not change the boundary value. Since $v\in C_c^\infty(\Omega)$, its support is a compact subset of the open set $\Omega$. Hence there is an open collar neighbourhood $U$ of $\partial\Omega$ in $\overline\Omega$ such that \begin{align*} v|_U=0. \end{align*} In particular, \begin{align*} v|_{\partial\Omega}=0. \end{align*} Restricting $F=\tilde u-v$ to the boundary gives \begin{align*} F|_{\partial\Omega}=(\tilde u-v)|_{\partial\Omega}. \end{align*} Restriction commutes with subtraction, so \begin{align*} (\tilde u-v)|_{\partial\Omega}=\tilde u|_{\partial\Omega}-v|_{\partial\Omega}. \end{align*} Substituting $\tilde u|_{\partial\Omega}=u$ and $v|_{\partial\Omega}=0$ gives \begin{align*} \tilde u|_{\partial\Omega}-v|_{\partial\Omega}=u-0. \end{align*} Finally, \begin{align*} u-0=u. \end{align*} Thus \begin{align*} F|_{\partial\Omega}=u. \end{align*} The compactly supported correction $v$ cancels exactly the interior defect $\bar\partial\tilde u$ and is zero near the boundary, so the subtraction produces a holomorphic function with the original prescribed trace. [/example] ## Levi-Flat and Degenerate Mixed Levi Geometry The final question is what happens when the Levi form is not positive. The answer is not merely that the proof fails; extension itself can fail. Levi-flat hypersurfaces have too many complex leaves, while Henkin's degenerate mixed-sign examples combine sign change with local cohomological obstruction. Both mechanisms allow CR functions whose local behaviour is incompatible with holomorphic extension to either side. [definition: Levi-Flat Hypersurface] A smooth real hypersurface $M\subset\mathbb C^n$ is Levi-flat if its Levi form vanishes identically on $T^{1,0}M$. [/definition] After this definition, $M$ is foliated locally by complex hypersurfaces under the standard integrability hypotheses. CR functions may then vary holomorphically along leaves without having the transverse control required for ambient holomorphic extension. [example: Leafwise Holomorphic Data on a Levi-Flat Model] In $\mathbb C^n$ write $z_n=t+is$, where $t,s\in\mathbb R$, and set $z'=(z_1,\dots,z_{n-1})$. The Levi-flat model hypersurface is \begin{align*} M=\{s=0\}=\{(z',z_n):z_n=t\in\mathbb R\}. \end{align*} A defining function for $M$ is \begin{align*} \rho(z)=\operatorname{Im}(z_n)=s. \end{align*} Since $z_n=t+is$ and $\bar z_n=t-is$, we have \begin{align*} s=\frac{z_n-\bar z_n}{2i}. \end{align*} Thus \begin{align*} \rho(z)=\frac{z_n-\bar z_n}{2i}. \end{align*} For $1\le j\le n-1$, the Wirtinger variables $z_j,\bar z_j$ are independent of $z_n,\bar z_n$, so \begin{align*} \frac{\partial z_n}{\partial\bar z_j}=0. \end{align*} Also, \begin{align*} \frac{\partial\bar z_n}{\partial\bar z_j}=0. \end{align*} Therefore \begin{align*} \frac{\partial\rho}{\partial\bar z_j}=\frac{\partial}{\partial\bar z_j}\left(\frac{z_n-\bar z_n}{2i}\right)=\frac{1}{2i}\left(0-0\right)=0. \end{align*} It follows that each vector field $\partial/\partial\bar z_j$, $1\le j\le n-1$, is tangent to $M$. These are exactly the leafwise $(0,1)$ directions on the product model, where the leaves are obtained by fixing the real parameter $t$. Let \begin{align*} u(z',t)=a(t)h(z') \end{align*} with $a\in C^\infty(\mathbb R)$ and $h$ holomorphic in $z'$. For $1\le j\le n-1$, the variable $t$ is independent of $\bar z_j$, so \begin{align*} \frac{\partial a(t)}{\partial\bar z_j}=0. \end{align*} By the product rule, \begin{align*} \frac{\partial u}{\partial\bar z_j}(z',t)=\frac{\partial a(t)}{\partial\bar z_j}h(z')+a(t)\frac{\partial h}{\partial\bar z_j}(z'). \end{align*} Since $h$ is holomorphic in $z'$, its antiholomorphic derivatives vanish: \begin{align*} \frac{\partial h}{\partial\bar z_j}(z')=0. \end{align*} Substituting the two vanishing derivatives gives \begin{align*} \frac{\partial u}{\partial\bar z_j}(z',t)=0\cdot h(z')+a(t)\cdot 0=0. \end{align*} Thus every tangential $(0,1)$ derivative of $u$ vanishes, so $u$ is CR on $M$. Now fix a point $(z'_0,t_0)\in M$ with $h(z'_0)\ne 0$. Suppose that a holomorphic function $F$ is defined near $(z'_0,t_0)\in\mathbb C^n$ and has boundary value $u$ on $M$. Holding $z'=z'_0$ fixed gives a one-variable holomorphic function \begin{align*} g(\zeta)=F(z'_0,\zeta) \end{align*} near $\zeta=t_0$. For real $t$ near $t_0$, the point $(z'_0,t)$ lies in $M$, because its last coordinate has imaginary part $0$. Hence \begin{align*} g(t)=F(z'_0,t). \end{align*} Using the boundary condition $F|_M=u$ gives \begin{align*} F(z'_0,t)=u(z'_0,t). \end{align*} Using the definition of $u$ gives \begin{align*} u(z'_0,t)=a(t)h(z'_0). \end{align*} Combining these identities yields \begin{align*} g(t)=a(t)h(z'_0) \end{align*} for all real $t$ sufficiently close to $t_0$. Because $g$ is holomorphic near $t_0$, it has a convergent power series expansion \begin{align*} g(\zeta)=\sum_{m=0}^{\infty} c_m(\zeta-t_0)^m \end{align*} for $\zeta$ in some complex neighbourhood of $t_0$. Restricting this identity to real $t$ in that neighbourhood gives \begin{align*} g(t)=\sum_{m=0}^{\infty} c_m(t-t_0)^m. \end{align*} Since $g(t)=a(t)h(z'_0)$ and $h(z'_0)\ne 0$, division by the nonzero constant $h(z'_0)$ gives \begin{align*} a(t)=\frac{g(t)}{h(z'_0)}. \end{align*} Substituting the power series for $g(t)$ gives \begin{align*} a(t)=\frac{1}{h(z'_0)}\sum_{m=0}^{\infty} c_m(t-t_0)^m. \end{align*} Equivalently, \begin{align*} a(t)=\sum_{m=0}^{\infty} \frac{c_m}{h(z'_0)}(t-t_0)^m. \end{align*} This is a convergent power series in the real variable $t-t_0$, so $a$ would be real-analytic at $t_0$. Therefore, if $a$ is smooth but not real-analytic at $t_0$, no holomorphic extension can have boundary value $u$ near any point $(z'_0,t_0)$ with $h(z'_0)\ne 0$. [/example] This model shows the mechanism behind non-extension: the CR equations do not differentiate in the transverse real direction when the Levi form is flat. Henkin's examples show that comparable non-extension can persist beyond this elementary product model. [quotetheorem:9215] [citeproof:9215] Henkin's theorem marks the boundary of the Lewy extension theorem. Strict pseudoconvexity is not a decorative hypothesis; it is the condition that prevents the transverse freedom seen in Levi-flat geometry and the degenerate mixed-sign obstructions present in Henkin's models. The Levi-flat alternative is necessary in the statement because the product model $M=\{\operatorname{Im}z_n=0\}$ already gives non-extension from a non-real-analytic transverse factor. The mixed Levi alternative is more delicate: non-degenerate mixed-sign hypersurfaces have their own propagation theorems, so the counterexample must be read with the degeneracy and the Henkin construction included in the hypotheses. Smoothness alone is therefore not a substitute for a Levi sign, while the CR condition alone is not a substitute for ambient holomorphicity. The theorem is existential, so it does not say that all CR functions on Levi-flat or degenerate mixed hypersurfaces fail to extend; restrictions of ambient holomorphic functions still extend by definition. Its force is that weakened Levi hypotheses allow at least one smooth CR datum whose local behaviour violates necessary holomorphic continuation. The preceding Levi-flat model gives the concrete limitation: smooth dependence on a real transverse parameter is much weaker than holomorphic dependence on the corresponding complex variable. [remark: One-Sided Means Levi-Side] For a strictly pseudoconvex boundary, the extension side is the pseudoconvex side determined by the sign convention for the defining function. Reversing the orientation changes the sign of the Levi form and therefore changes which side is supported by the $\bar\partial$ estimates. The theorem does not assert simultaneous extension to both sides, and Lewy-Henkin examples show that two-sided extension is a much stronger property. [/remark] The chapter leaves us with a precise moral. CR functions are boundary holomorphic only when the boundary geometry supplies enough analytic control to solve the missing normal equation. Strict pseudoconvexity supplies that control; Levi-flatness and Henkin's degenerate mixed Levi geometry leave enough freedom for non-extension. After extension from the boundary is understood locally, the course turns to the global condition that governs when domains behave like genuine holomorphic regions. Pseudoconvexity at the boundary connects the Levi form to the Levi problem and explains why boundary convexity is the correct analytic substitute for interior holomorphic convexity. # 6. Pseudoconvexity at the Boundary and the Levi Problem This chapter returns to the Levi problem from the boundary side. Earlier parts of the course treated pseudoconvexity as an interior condition, expressed through plurisubharmonic functions, holomorphic convexity, and domains of holomorphy. We now ask how much of that global analytic condition can be read directly from a smooth real hypersurface bounding the domain. The answer is that, for smooth boundaries, the sign of the Levi form on complex tangent directions is the local boundary signature of pseudoconvexity. ## Reading Pseudoconvexity from a Defining Function Suppose $\Omega \subset \mathbb C^n$ has smooth boundary. The local problem is to decide whether holomorphic functions inside $\Omega$ can develop boundary barriers or holomorphic extension phenomena by looking only at derivatives of a defining function near $\partial \Omega$. Since complex tangent directions are the directions in which a holomorphic curve can touch the boundary to first order, the second-order behaviour of the defining function along those directions is the relevant quantity. [definition: Smooth Defining Function] Let $\Omega \subset \mathbb C^n$ be a domain with smooth boundary. A smooth defining function for $\Omega$ near $p \in \partial \Omega$ is a function $\rho \in C^\infty(U;\mathbb R)$ on a neighbourhood $U$ of $p$ such that \begin{align*} \Omega \cap U = \{z \in U : \rho(z) < 0\}. \end{align*} It also satisfies \begin{align*} \partial\Omega \cap U = \{z \in U : \rho(z)=0\}. \end{align*} Finally, \begin{align*} d\rho_p \neq 0. \end{align*} [/definition] The choice of sign fixes the convention for the Levi form below. Multiplying $\rho$ by a positive smooth factor changes the Levi form on complex tangent vectors by the same positive factor at the boundary point, so the condition of non-negativity is intrinsic. To isolate the directions where the boundary can be tested by holomorphic curves, we first record the complex tangent space. [definition: Complex Tangent Space] Let $M=\partial\Omega$ and let $p \in M$. The holomorphic tangent space of $M$ at $p$ is \begin{align*} H_p(M)=T_p^{1,0}(M)=\left\{\xi=\sum_{j=1}^n \xi_j\frac{\partial}{\partial z_j}: \sum_{j=1}^n \frac{\partial \rho}{\partial z_j}(p)\xi_j=0\right\}. \end{align*} [/definition] This is the complex part of the real tangent hyperplane. It is independent of the defining function, because the equation only records which complex directions annihilate the complex normal covector. Once these tangential directions have been separated from the normal direction, the next invariant is the second-order change of $\rho$ along them. [definition: Levi Form of a Defining Function] Let $\rho$ be a smooth defining function for $\Omega$ near $p \in \partial\Omega$. The Levi form of $\rho$ at $p$ is the real-valued quadratic form \begin{align*} \mathcal L_\rho(p;\cdot):H_p(\partial\Omega) \to \mathbb R \end{align*} defined by \begin{align*} \mathcal L_\rho(p;\xi)=\sum_{j,k=1}^n \frac{\partial^2 \rho}{\partial z_j\partial \bar z_k}(p)\xi_j\overline{\xi_k}, \qquad \xi \in H_p(\partial\Omega). \end{align*} Its associated Hermitian sesquilinear form is the map $H_p(\partial\Omega)\times H_p(\partial\Omega)\to\mathbb C$ obtained by replacing $\xi_j\overline{\xi_k}$ with $\xi_j\overline{\eta_k}$. [/definition] The restriction to $H_p(\partial\Omega)$ is essential. In the complex normal direction the value depends on how the defining function is extended off the boundary, while tangent values measure the curvature of the hypersurface itself. The central boundary question is therefore whether non-negative curvature in precisely these directions is equivalent to pseudoconvexity. [quotetheorem:3392] [citeproof:3392] The theorem translates pseudoconvexity into a computation, but it does so only because the boundary has enough differentiability for the complex Hessian of $\rho$ to exist and vary continuously. Without a $C^2$ defining function there may be no pointwise Levi form, even though pseudoconvexity can still be formulated by plurisubharmonic exhaustions. The result also does not say that every non-negative Levi form is strictly positive, or that weakly pseudoconvex boundary points enjoy the same regularity as strictly pseudoconvex ones. Its role is to make the first bridge from global pseudoconvexity to CR boundary geometry; the next distinction is between weak non-negativity and strict positivity. [example: Unit Ball Levi Form] Let $\Omega=B(0,1)\subset \mathbb C^n$, and take the defining function \begin{align*} \rho(z)=|z|^2-1=\sum_{m=1}^n z_m\overline{z_m}-1. \end{align*} For $p\in \partial B(0,1)$, the complex tangent space is found from the first derivatives of $\rho$. Treating $z_j$ and $\overline z_j$ as independent variables in the Wirtinger calculus, \begin{align*} \frac{\partial}{\partial z_j}(z_m\overline z_m)=\delta_{jm}\overline z_m. \end{align*} Therefore \begin{align*} \frac{\partial \rho}{\partial z_j}(z)=\sum_{m=1}^n \delta_{jm}\overline z_m=\overline z_j. \end{align*} Thus $\xi=(\xi_1,\ldots,\xi_n)\in H_p(\partial B(0,1))$ exactly when \begin{align*} \sum_{j=1}^n \frac{\partial \rho}{\partial z_j}(p)\xi_j=\sum_{j=1}^n \overline{p_j}\xi_j=0. \end{align*} The second derivatives are equally explicit: \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial \overline z_k}(z)=\frac{\partial}{\partial \overline z_k}\overline z_j=\delta_{jk}. \end{align*} Hence for any $\xi\in H_p(\partial B(0,1))$, \begin{align*} \mathcal L_\rho(p;\xi)=\sum_{j,k=1}^n \delta_{jk}\xi_j\overline{\xi_k}. \end{align*} Since $\delta_{jk}=0$ when $j\ne k$ and $\delta_{jj}=1$, this sum reduces to \begin{align*} \mathcal L_\rho(p;\xi)=\sum_{j=1}^n \xi_j\overline{\xi_j}. \end{align*} For each $j$, $\xi_j\overline{\xi_j}=|\xi_j|^2$, so \begin{align*} \mathcal L_\rho(p;\xi)=\sum_{j=1}^n |\xi_j|^2=|\xi|^2. \end{align*} If $\xi\ne 0$, then at least one $|\xi_j|^2$ is positive and all the other terms are non-negative, so $\mathcal L_\rho(p;\xi)>0$. Thus the sphere has positive Levi curvature in every non-zero complex tangent direction, making the unit ball the basic strictly pseudoconvex model. [/example] This computation also explains why the ball is the local reference object. A hypersurface whose Levi form has the same positive-definite behaviour resembles the sphere to second order along complex tangential directions. ## Strict Pseudoconvexity and the Spherical Model The next question is what changes when the Levi form is not only non-negative but positive in every non-zero complex tangent direction. This stronger condition is the boundary analogue of strict plurisubharmonicity, and it is the regime where the local CR geometry is most rigid. [definition: Strictly Pseudoconvex Boundary Point] Let $\Omega \subset \mathbb C^n$ have $C^2$ boundary and let $p \in \partial\Omega$. The boundary is strictly pseudoconvex at $p$ if, for some smooth defining function $\rho$ with $\Omega=\{\rho<0\}$ near $p$, \begin{align*} \mathcal L_\rho(p;\xi)>0 \end{align*} for every non-zero $\xi \in H_p(\partial\Omega)$. [/definition] Because a positive change of defining function rescales the restricted Levi form by a positive number at $p$, this is a property of the oriented hypersurface rather than of the chosen formula for $\rho$. For global boundary analysis we need the same positivity at all boundary points, since estimates and normal forms are applied uniformly across the hypersurface. This leads from a pointwise condition to a domain-level condition. [definition: Strictly Pseudoconvex Domain] A domain $\Omega \subset \mathbb C^n$ with $C^2$ boundary is strictly pseudoconvex if $\partial\Omega$ is strictly pseudoconvex at every $p \in \partial\Omega$. [/definition] Strict positivity excludes flat complex-tangential boundary directions. It is stronger than boundary pseudoconvexity and gives a robust class of domains for $\bar\partial$ estimates, boundary regularity, and CR embedding results. [remark: Strong Versus Strict Pseudoconvexity] In many analytic texts, strong pseudoconvexity means the existence of a defining function whose complex Hessian is positive definite in all complex directions near the boundary. Strict pseudoconvexity refers to positivity of the Levi form on the complex tangent space at the boundary. For smooth bounded domains these formulations are closely related after modifying the defining function by a convex increasing function, but the distinction is useful because the boundary condition is intrinsic while the ambient Hessian condition uses an extension away from $\partial\Omega$. [/remark] The unit ball provides the prototype, but strict pseudoconvexity does not mean that every boundary is biholomorphically spherical. The natural classification problem asks how much of a strictly pseudoconvex hypersurface can be simplified by holomorphic coordinates and which terms remain as CR invariants. Chern-Moser normal form is the organizing answer: after a local holomorphic change of coordinates near a strictly pseudoconvex boundary point, the hypersurface can be put into a standard form whose leading model is the sphere and whose higher-order terms carry CR invariants. In this course we use this normal-form picture as geometric guidance rather than as a technical tool. Its hypotheses are doing real work: strict pseudoconvexity is what makes the positive quadratic term $|z'|^2$ available, while weakly pseudoconvex hypersurfaces can have flat complex-tangential directions and require different normal forms. The statement is local near a chosen boundary point, not a global classification of the whole boundary. It also does not say that the quadratic term determines the biholomorphic type; the higher-order normal form terms contain the CR invariants that can obstruct local equivalence to the sphere. These invariants later reappear in boundary questions for the tangential Cauchy-Riemann operator and in regularity problems for the $\bar\partial$-Neumann problem. [example: Perturbing the Sphere] Write $z'=(z_1,\ldots,z_{n-1})$ and consider the hypersurface near $0$ defined by \begin{align*} \rho(z',w)=|z'|^2+\varepsilon |z_1|^4-\operatorname{Im} w=0. \end{align*} Since \begin{align*} \operatorname{Im} w=\frac{w-\overline w}{2i}, \end{align*} the term $-\operatorname{Im} w$ is linear in $w$ and $\overline w$, so every second derivative involving this term is $0$. For the quadratic part, \begin{align*} |z'|^2=\sum_{m=1}^{n-1}z_m\overline z_m. \end{align*} Thus \begin{align*} \frac{\partial |z'|^2}{\partial z_j}=\sum_{m=1}^{n-1}\delta_{jm}\overline z_m=\overline z_j. \end{align*} Taking the $\overline z_k$ derivative gives \begin{align*} \frac{\partial^2 |z'|^2}{\partial z_j\partial\overline z_k}=\frac{\partial \overline z_j}{\partial\overline z_k}=\delta_{jk}. \end{align*} For the quartic perturbation, \begin{align*} |z_1|^4=(z_1\overline z_1)^2=z_1^2\overline z_1^2. \end{align*} Hence \begin{align*} \frac{\partial |z_1|^4}{\partial z_1}=2z_1\overline z_1^2. \end{align*} Taking the $\overline z_1$ derivative gives \begin{align*} \frac{\partial^2 |z_1|^4}{\partial z_1\partial\overline z_1}=2z_1\cdot 2\overline z_1=4|z_1|^2. \end{align*} If $j\ne 1$, then $\partial |z_1|^4/\partial z_j=0$; if $k\ne 1$, then $\partial(2z_1\overline z_1^2)/\partial\overline z_k=0$. Therefore the $z'$-block of the complex Hessian is \begin{align*} \left(\frac{\partial^2\rho}{\partial z_j\partial\overline z_k}\right)_{1\le j,k\le n-1}=\operatorname{diag}(1+4\varepsilon |z_1|^2,1,\ldots,1). \end{align*} Let $p=(a,w_0)$, with $a=(a_1,\ldots,a_{n-1})$, and let $\xi=(\xi',\xi_w)\in H_p(\partial\Omega)$. Since all second derivatives involving $w$ vanish, the Levi form only sees the $z'$ components: \begin{align*} \mathcal L_\rho(p;\xi)=(1+4\varepsilon |a_1|^2)|\xi_1|^2+\sum_{j=2}^{n-1}|\xi_j|^2. \end{align*} At $p=0$, this becomes \begin{align*} \mathcal L_\rho(0;\xi)=\sum_{j=1}^{n-1}|\xi_j|^2. \end{align*} At the origin the first derivatives in the $z'$ variables vanish, while \begin{align*} \frac{\partial\rho}{\partial w}(0)=\frac{i}{2}. \end{align*} Thus the complex tangent condition at $0$ is \begin{align*} \frac{i}{2}\xi_w=0, \end{align*} so $\xi_w=0$. Therefore a non-zero complex tangent vector at $0$ has $\xi'\ne 0$, and then \begin{align*} \mathcal L_\rho(0;\xi)=\sum_{j=1}^{n-1}|\xi_j|^2>0. \end{align*} The same positivity persists after shrinking the neighbourhood. If $\varepsilon\ge 0$, then $1+4\varepsilon |a_1|^2\ge 1$. If $\varepsilon<0$, choose the neighbourhood so that $4|\varepsilon||a_1|^2<1$; then \begin{align*} 1+4\varepsilon |a_1|^2=1-4|\varepsilon||a_1|^2>0. \end{align*} For any tangent vector with $\xi'\ne 0$, the displayed Levi form is then positive. If $\xi'=0$, the tangent equation forces $\xi_w=0$ because $\partial\rho/\partial w=i/2$, so there is no non-zero tangent vector with only a $w$ component. Hence the hypersurface is strictly pseudoconvex near $0$. The quartic term leaves the positive quadratic Levi model at the origin unchanged, while changing higher-order CR data beyond that model. [/example] ## Boundary Distance and Plurisubharmonic Exhaustions The boundary criterion should match the older definition of pseudoconvexity from SCV I: a domain is pseudoconvex when it admits a continuous or smooth plurisubharmonic exhaustion. The bridge is the boundary distance function, or more precisely functions built from the defining function that blow up at the boundary while retaining plurisubharmonicity. [definition: Plurisubharmonic Exhaustion] Let $\Omega \subset \mathbb C^n$ be a domain. A plurisubharmonic exhaustion of $\Omega$ is a plurisubharmonic function $\psi:\Omega \to \mathbb R$ such that for every $c \in \mathbb R$ the sublevel set \begin{align*} \{z \in \Omega : \psi(z)<c\} \end{align*} is relatively compact in $\Omega$. [/definition] Exhaustions convert boundary behaviour into interior convexity. Near a smooth boundary, functions such as $-\log(-\rho)$ or powers of the boundary distance become candidates for $\psi$, but their complex Hessians contain both normal terms and the tangential Levi form. This raises the comparison problem: does the boundary Levi condition produce exactly the same class of domains as the exhaustion definition? [quotetheorem:9216] [example: Model Exhaustion for the Ball] For $B(0,1)\subset \mathbb C^n$, define \begin{align*} \psi(z)=-\log(1-|z|^2), \end{align*} where \begin{align*} |z|^2=\sum_{m=1}^n z_m\overline z_m. \end{align*} Since $|z|<1$ on $B(0,1)$, we have $1-|z|^2>0$, so $\psi$ is smooth and real-valued on the ball. We compute its complex Hessian and then check that its sublevel sets stay compactly inside the ball. For $1\le j\le n$, Wirtinger differentiation gives \begin{align*} \frac{\partial}{\partial z_j}(1-|z|^2)=-\frac{\partial}{\partial z_j}\sum_{m=1}^n z_m\overline z_m=-\sum_{m=1}^n\delta_{jm}\overline z_m=-\overline z_j. \end{align*} Using $\frac{d}{du}(-\log u)=-u^{-1}$ and the chain rule, \begin{align*} \frac{\partial \psi}{\partial z_j}(z)=-\frac{1}{1-|z|^2}\cdot(-\overline z_j)=\frac{\overline z_j}{1-|z|^2}. \end{align*} Now differentiate this expression with respect to $\overline z_k$. The product rule gives \begin{align*} \frac{\partial}{\partial\overline z_k}\left(\frac{\overline z_j}{1-|z|^2}\right)=\frac{\delta_{jk}}{1-|z|^2}+\overline z_j\frac{\partial}{\partial\overline z_k}(1-|z|^2)^{-1}. \end{align*} Since \begin{align*} \frac{\partial}{\partial\overline z_k}(1-|z|^2)=-z_k, \end{align*} another use of the chain rule gives \begin{align*} \frac{\partial}{\partial\overline z_k}(1-|z|^2)^{-1}=-(1-|z|^2)^{-2}(-z_k)=\frac{z_k}{(1-|z|^2)^2}. \end{align*} Therefore \begin{align*} \frac{\partial^2\psi}{\partial z_j\partial\overline z_k}(z)=\frac{\delta_{jk}}{1-|z|^2}+\frac{\overline z_jz_k}{(1-|z|^2)^2}. \end{align*} For $\xi=(\xi_1,\ldots,\xi_n)\in\mathbb C^n$, the associated Hermitian quadratic form is \begin{align*} \sum_{j,k=1}^n\frac{\partial^2\psi}{\partial z_j\partial\overline z_k}(z)\xi_j\overline{\xi_k}=\sum_{j,k=1}^n\frac{\delta_{jk}\xi_j\overline{\xi_k}}{1-|z|^2}+\sum_{j,k=1}^n\frac{\overline z_jz_k\xi_j\overline{\xi_k}}{(1-|z|^2)^2}. \end{align*} Because $\delta_{jk}=0$ for $j\ne k$ and $\delta_{jj}=1$, \begin{align*} \sum_{j,k=1}^n\delta_{jk}\xi_j\overline{\xi_k}=\sum_{j=1}^n\xi_j\overline{\xi_j}=\sum_{j=1}^n|\xi_j|^2. \end{align*} For the second sum, factor the $j$- and $k$-dependent terms: \begin{align*} \sum_{j,k=1}^n\overline z_jz_k\xi_j\overline{\xi_k}=\left(\sum_{j=1}^n\overline z_j\xi_j\right)\left(\sum_{k=1}^n z_k\overline{\xi_k}\right). \end{align*} The second factor is the complex conjugate of the first, so \begin{align*} \left(\sum_{j=1}^n\overline z_j\xi_j\right)\left(\sum_{k=1}^n z_k\overline{\xi_k}\right)=\left|\sum_{j=1}^n\overline z_j\xi_j\right|^2. \end{align*} Hence \begin{align*} \sum_{j,k=1}^n\frac{\partial^2\psi}{\partial z_j\partial\overline z_k}(z)\xi_j\overline{\xi_k}=\frac{\sum_{j=1}^n|\xi_j|^2}{1-|z|^2}+\frac{\left|\sum_{j=1}^n\overline z_j\xi_j\right|^2}{(1-|z|^2)^2}. \end{align*} On $B(0,1)$ the denominators $1-|z|^2$ and $(1-|z|^2)^2$ are positive, while $\sum_j|\xi_j|^2$ and $\left|\sum_j\overline z_j\xi_j\right|^2$ are non-negative. Thus the complex Hessian is positive semidefinite, so $\psi$ is plurisubharmonic on $B(0,1)$. It remains to verify the exhaustion property. Since $0<1-|z|^2\le 1$ on $B(0,1)$, we have $\log(1-|z|^2)\le 0$, and therefore $\psi(z)\ge 0$. If $c\le 0$, then $\{z\in B(0,1):\psi(z)<c\}=\varnothing$. If $c>0$, then \begin{align*} \psi(z)<c \Longleftrightarrow -\log(1-|z|^2)<c. \end{align*} Multiplying by $-1$ reverses the inequality: \begin{align*} -\log(1-|z|^2)<c \Longleftrightarrow \log(1-|z|^2)>-c. \end{align*} Since the exponential function is increasing, \begin{align*} \log(1-|z|^2)>-c \Longleftrightarrow 1-|z|^2>e^{-c}. \end{align*} Rearranging gives \begin{align*} 1-|z|^2>e^{-c} \Longleftrightarrow |z|^2<1-e^{-c}. \end{align*} Thus for $c>0$ the sublevel set is \begin{align*} \{z\in B(0,1):\psi(z)<c\}=\{z\in\mathbb C^n:|z|<\sqrt{1-e^{-c}}\}. \end{align*} Because $c>0$ implies $0<e^{-c}<1$, the radius $\sqrt{1-e^{-c}}$ is strictly less than $1$, so the closure of this sublevel set lies inside $B(0,1)$. Therefore $\psi$ is a plurisubharmonic exhaustion of the ball, and its logarithmic blow-up as $|z|\to 1$ is exactly what keeps all finite sublevel sets away from the boundary. [/example] This example shows how strict boundary curvature creates an interior exhaustion. In the weakly pseudoconvex case, the construction needs more care because tangential directions may have zero Levi curvature. [example: Product Domain Boundary] Let $p=(p_1,p_2)\in \partial\Omega_1\times \Omega_2$, where $\Omega_1\subset\mathbb C^{n_1}$ and $\Omega_2\subset\mathbb C^{n_2}$, and let $\rho_1$ be a smooth defining function for $\Omega_1$ near $p_1$. Near $p$, the product face $\partial\Omega_1\times\Omega_2$ is defined by \begin{align*} \rho(z_1,z_2)=\rho_1(z_1). \end{align*} For $\xi=(\xi^{(1)},\xi^{(2)})\in\mathbb C^{n_1}\times\mathbb C^{n_2}$, the dependence of $\rho$ only on $z_1$ gives, for $1\le j\le n_1$, \begin{align*} \frac{\partial \rho}{\partial z_{1,j}}(p)=\frac{\partial \rho_1}{\partial z_{1,j}}(p_1). \end{align*} For $1\le \ell\le n_2$, differentiating a function independent of $z_{2,\ell}$ gives \begin{align*} \frac{\partial \rho}{\partial z_{2,\ell}}(p)=0. \end{align*} Therefore the complex tangent condition at $p$ is \begin{align*} \sum_{j=1}^{n_1}\frac{\partial \rho_1}{\partial z_{1,j}}(p_1)\xi^{(1)}_j+\sum_{\ell=1}^{n_2}0\cdot \xi^{(2)}_\ell=0. \end{align*} Since the second sum is $0$, this is equivalent to \begin{align*} \sum_{j=1}^{n_1}\frac{\partial \rho_1}{\partial z_{1,j}}(p_1)\xi^{(1)}_j=0. \end{align*} Thus $\xi\in H_p(\partial\Omega)$ exactly when $\xi^{(1)}\in H_{p_1}(\partial\Omega_1)$, and the component $\xi^{(2)}$ is unrestricted. The complex Hessian has the same block structure. For $1\le j,k\le n_1$, \begin{align*} \frac{\partial^2\rho}{\partial z_{1,j}\partial\overline z_{1,k}}(p)=\frac{\partial^2\rho_1}{\partial z_{1,j}\partial\overline z_{1,k}}(p_1). \end{align*} Because $\rho$ is independent of all $z_2$ and $\overline z_2$ variables, for $1\le j\le n_1$ and $1\le \ell,m\le n_2$, \begin{align*} \frac{\partial^2\rho}{\partial z_{1,j}\partial\overline z_{2,\ell}}(p)=0. \end{align*} Similarly, \begin{align*} \frac{\partial^2\rho}{\partial z_{2,\ell}\partial\overline z_{1,j}}(p)=0. \end{align*} And also, \begin{align*} \frac{\partial^2\rho}{\partial z_{2,\ell}\partial\overline z_{2,m}}(p)=0. \end{align*} Substituting these blocks into the Levi form gives \begin{align*} \mathcal L_\rho(p;\xi)=\sum_{j,k=1}^{n_1}\frac{\partial^2\rho_1}{\partial z_{1,j}\partial\overline z_{1,k}}(p_1)\xi^{(1)}_j\overline{\xi^{(1)}_k}. \end{align*} By the definition of the Levi form for $\rho_1$, this is \begin{align*} \mathcal L_\rho(p;\xi)=\mathcal L_{\rho_1}(p_1;\xi^{(1)}). \end{align*} Hence, if $\Omega_1$ is pseudoconvex at $p_1$, then $\mathcal L_{\rho_1}(p_1;\xi^{(1)})\ge 0$ for every $\xi^{(1)}\in H_{p_1}(\partial\Omega_1)$, so \begin{align*} \mathcal L_\rho(p;\xi)\ge 0. \end{align*} This same formula shows why the product face is not strictly pseudoconvex when $n_2\ge 1$. Choose a non-zero vector $\eta\in\mathbb C^{n_2}$ and set $\xi=(0,\eta)$. The tangent equation becomes \begin{align*} \sum_{j=1}^{n_1}\frac{\partial \rho_1}{\partial z_{1,j}}(p_1)\cdot 0+\sum_{\ell=1}^{n_2}0\cdot \eta_\ell=0, \end{align*} so $\xi\in H_p(\partial\Omega)$. Its Levi value is \begin{align*} \mathcal L_\rho(p;(0,\eta))=\mathcal L_{\rho_1}(p_1;0). \end{align*} Expanding the right-hand side, \begin{align*} \mathcal L_{\rho_1}(p_1;0)=\sum_{j,k=1}^{n_1}\frac{\partial^2\rho_1}{\partial z_{1,j}\partial\overline z_{1,k}}(p_1)\cdot 0\cdot \overline{0}=0. \end{align*} Thus the product has non-zero complex tangent directions parallel to the interior factor on which the Levi form vanishes. The smooth product face is pseudoconvex under the pseudoconvexity hypothesis on $\Omega_1$, but it is not strictly pseudoconvex whenever the second factor has positive complex dimension. [/example] The product example is a useful warning: pseudoconvexity permits flat complex directions. Boundary regularity and extension results often become sharper only after imposing strict positivity or finite-type hypotheses. ## Domains of Holomorphy and the Levi Problem at the Boundary The final question is how the boundary computations complete the Levi problem narrative. In SCV I and II the Levi problem states that pseudoconvex domains are exactly domains of holomorphy. The boundary form of the theory says that, for smooth domains, this analytic property is visible in the CR geometry of $\partial\Omega$. [definition: Domain of Holomorphy] A domain $\Omega \subset \mathbb C^n$ is a domain of holomorphy if there exists a holomorphic function $f:\Omega\to\mathbb C$ that cannot be holomorphically extended to any strictly larger domain across any non-empty boundary neighbourhood. [/definition] This definition records maximality for holomorphic functions. A domain that fails it has hidden analytic continuation through some boundary piece, so its boundary is not a genuine obstruction for all holomorphic functions. [quotetheorem:3416] [citeproof:3416] The theorem is the conceptual endpoint of the chapter, but its smooth-boundary wording should not be mistaken for the most general Levi problem. On a nonsmooth pseudoconvex domain, such as a polydisc at a corner point, the domain may still be a domain of holomorphy while no single boundary hypersurface has a well-defined Levi form at the corner. Conversely, the theorem does not claim strict pseudoconvexity, boundary regularity for the $\bar\partial$-Neumann problem, or spherical CR geometry; it only identifies holomorphic maximality with non-negative Levi curvature where the second-order boundary invariant exists. Boundary CR geometry, plurisubharmonic convexity, and holomorphic maximality are three views of the same phenomenon when the boundary is sufficiently smooth. The hypotheses are therefore deliberately balanced. Smoothness is needed to read pseudoconvexity from the Levi form, while plurisubharmonic exhaustion is the more flexible analytic language behind the result. The theorem is used below as a baseline: once non-negative Levi curvature is known to capture holomorphic maximality, examples such as worm domains can be read as warnings that pseudoconvexity alone still leaves serious regularity questions unresolved. [example: Worm Domains and Smoothness Warnings] Diederich-Fornaess worm domains give smooth bounded pseudoconvex domains whose boundary behaviour is weaker than strict pseudoconvexity. At a smooth boundary point $p\in\partial\Omega$, choose a smooth defining function $\rho$ with $\Omega=\{\rho<0\}$ near $p$. A vector $\xi=(\xi_1,\ldots,\xi_n)\in\mathbb C^n$ is complex tangent exactly when it annihilates the complex normal covector, namely \begin{align*} \sum_{j=1}^n \frac{\partial\rho}{\partial z_j}(p)\xi_j=0. \end{align*} For such a tangent vector, the Levi form is obtained by inserting $\xi$ into the Hermitian matrix of mixed second derivatives: \begin{align*} \mathcal L_\rho(p;\xi)=\sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}(p)\xi_j\overline{\xi_k}. \end{align*} By *[The Levi Condition Is a Boundary Test](/theorems/3392)*, pseudoconvexity at a $C^2$ boundary point requires \begin{align*} \mathcal L_\rho(p;\xi)\ge 0 \end{align*} for every $\xi\in H_p(\partial\Omega)$. Strict pseudoconvexity would require the stronger condition \begin{align*} \mathcal L_\rho(p;\xi)>0 \end{align*} for every non-zero $\xi\in H_p(\partial\Omega)$. Thus the distinction is visible in a single tangent direction. If $q\in\partial\Omega$ is a weakly pseudoconvex point and $\eta\in H_q(\partial\Omega)$ is non-zero with \begin{align*} \mathcal L_\rho(q;\eta)=0, \end{align*} then the pseudoconvex inequality still holds in that direction because \begin{align*} 0\ge 0. \end{align*} But strict pseudoconvexity fails in the same direction because the required inequality would be \begin{align*} 0>0, \end{align*} which is false. The lesson is that smooth pseudoconvexity permits non-zero complex tangent directions with zero Levi curvature, and those weak directions are precisely where stronger boundary regularity arguments can fail. For a non-smooth pseudoconvex domain the obstruction is more basic. The tangent equation already uses the first derivatives $\partial\rho/\partial z_j(p)$, and the Levi form uses the second derivatives \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\overline z_k}(p). \end{align*} If no $C^2$ defining function is available at $p$, then these second derivative entries need not exist, so the displayed Hermitian sum has no well-defined pointwise value. The exhaustion formulation avoids this boundary Hessian: it asks instead for a plurisubharmonic exhaustion on the interior of $\Omega$. This is why the boundary Levi criterion is a smooth-boundary theorem, while the general Levi problem is stated in terms of pseudoconvexity or plurisubharmonic exhaustion rather than a pointwise defining-function Hessian. [/example] The Levi problem therefore has two complementary faces. In the interior theory, pseudoconvexity is controlled by plurisubharmonic functions and holomorphic convexity; at a smooth boundary, the same condition is encoded by the Levi form on $H(\partial\Omega)$. The next stage of the course studies what the tangential Cauchy-Riemann operator detects beyond this sign condition, especially extension, obstruction, and regularity phenomena on CR manifolds. The Levi condition captures the basic sign information, but it does not exhaust the geometry of strictly pseudoconvex boundaries. The next chapter refines this picture through Chern-Moser theory, where local equivalence, normal forms, and CR invariants detect the deeper structure hidden beyond the Levi form. # 7. Chern-Moser Theory and CR Invariants The preceding chapters treated CR maps and boundary extension as analytic objects: solutions of tangential Cauchy-Riemann equations, boundary values of holomorphic maps, and transformations preserving the Levi form. This chapter asks a more rigid geometric question: given two strictly pseudoconvex real hypersurfaces, how can we decide whether a local CR diffeomorphism carries one to the other? Chern-Moser theory answers this by putting hypersurfaces into canonical coordinates and reading off the coefficients that survive all allowable coordinate changes. The guiding model is the unit sphere, whose large automorphism group makes it the flat object in this geometry. Normal forms isolate the part of a defining equation that cannot be removed by biholomorphic changes of variables, while pseudo-Hermitian geometry packages the same information in terms of a contact form, connection, torsion, and curvature. ## The CR Equivalence Problem in Dimension Three The first problem is to turn local CR equivalence into a calculation. For hypersurfaces in $\mathbb C^2$, a point has a one-dimensional complex tangent direction, so the Levi form gives positivity but not a complete invariant. Cartan's 1932 analysis showed that higher-order data, organised through an equivalence method, are needed to distinguish hypersurfaces that have the same Levi form. [definition: Local CR Equivalence] Let $M,N \subset \mathbb C^2$ be smooth real hypersurfaces, and let $p \in M$, $q \in N$. A local CR equivalence from $(M,p)$ to $(N,q)$ is a diffeomorphism $F: U \to V$ between neighbourhoods $U \subset M$ of $p$ and $V \subset N$ of $q$ such that $F(p)=q$ and \begin{align*} dF_x(T_x^{1,0}M)=T_{F(x)}^{1,0}N \end{align*} for every $x \in U$. [/definition] A CR equivalence preserves the complex tangent bundle, hence also preserves the Levi form up to the natural scaling coming from the choice of defining function. For strictly pseudoconvex hypersurfaces this first invariant only says that the geometry resembles the sphere to second order; the equivalence problem begins at higher order. [example: Sphere And Heisenberg Hypersurface] Let $(\zeta,\eta)$ be coordinates on the source copy of $\mathbb C^2$, and consider the unit sphere \begin{align*}\lvert \zeta\rvert^2+\lvert \eta\rvert^2=1\end{align*} near $(0,1)$. On the neighbourhood where $1+\eta\ne 0$, define \begin{align*}z=\frac{\zeta}{1+\eta},\qquad w=i\,\frac{1-\eta}{1+\eta}\end{align*} with $w=u+iv$. The coordinate functions are holomorphic in $(\zeta,\eta)$ on this neighbourhood. To solve for $\eta$, multiply the equation for $w$ by $1+\eta$: \begin{align*}w(1+\eta)=i(1-\eta).\end{align*} Expanding both sides gives \begin{align*}w+w\eta=i-i\eta.\end{align*} Moving the $\eta$-terms to the left and the remaining terms to the right gives \begin{align*}\eta(w+i)=i-w.\end{align*} Thus, wherever $w+i\ne 0$, \begin{align*}\eta=\frac{i-w}{w+i}.\end{align*} Since \begin{align*}1+\eta=1+\frac{i-w}{w+i}=\frac{w+i+i-w}{w+i}=\frac{2i}{w+i},\end{align*} the equation $z=\zeta/(1+\eta)$ gives \begin{align*}\zeta=z(1+\eta)=\frac{2iz}{w+i}.\end{align*} Hence the map is biholomorphic between the indicated neighbourhoods. We compute the image of the sphere. Multiplying the numerator and denominator in the formula for $w$ by $1+\bar\eta$ gives \begin{align*}w=i\,\frac{(1-\eta)(1+\bar\eta)}{|1+\eta|^2}.\end{align*} The numerator expands as \begin{align*}(1-\eta)(1+\bar\eta)=1+\bar\eta-\eta-|\eta|^2.\end{align*} Therefore \begin{align*}w=\frac{i(1-|\eta|^2)+i(\bar\eta-\eta)}{|1+\eta|^2}.\end{align*} If $\eta=a+ib$, then \begin{align*}i(\bar\eta-\eta)=i(a-ib-a-ib)=i(-2ib)=2b,\end{align*} so $i(\bar\eta-\eta)$ is real. Since $1-|\eta|^2$ and $|1+\eta|^2$ are real, the imaginary part of $w$ is \begin{align*}\operatorname{Im}w=\frac{1-|\eta|^2}{|1+\eta|^2}.\end{align*} On the sphere, $|\zeta|^2+|\eta|^2=1$, so $1-|\eta|^2=|\zeta|^2$. Hence \begin{align*}\operatorname{Im}w=\frac{|\zeta|^2}{|1+\eta|^2}.\end{align*} Using $z=\zeta/(1+\eta)$, this becomes \begin{align*}\operatorname{Im}w=\left|\frac{\zeta}{1+\eta}\right|^2=|z|^2.\end{align*} Conversely, suppose $w=u+iv$ and $\operatorname{Im}w=v=|z|^2$. For the inverse formulas, \begin{align*}|\eta|^2=\left|\frac{i-w}{w+i}\right|^2=\frac{u^2+(1-v)^2}{u^2+(v+1)^2}\end{align*} and \begin{align*}|\zeta|^2=\left|\frac{2iz}{w+i}\right|^2=\frac{4|z|^2}{u^2+(v+1)^2}=\frac{4v}{u^2+(v+1)^2}.\end{align*} Adding the two expressions gives \begin{align*}|\zeta|^2+|\eta|^2=\frac{u^2+(1-v)^2+4v}{u^2+(v+1)^2}.\end{align*} The numerator is \begin{align*}u^2+(1-v)^2+4v=u^2+1-2v+v^2+4v=u^2+1+2v+v^2=u^2+(v+1)^2,\end{align*} so \begin{align*}|\zeta|^2+|\eta|^2=1.\end{align*} Thus the Cayley transform carries the sphere near $(0,1)$ exactly to the Heisenberg hypersurface $\operatorname{Im}w=|z|^2$ near $(0,0)$. In these coordinates the defining equation is exactly the Levi quadric and contains no higher-weight remainder terms. The model also has visible symmetries: $z\mapsto e^{i\theta}z$ preserves $|z|^2$ because $|e^{i\theta}z|^2=|e^{i\theta}|^2|z|^2=|z|^2$, and for real $\lambda\ne 0$ the dilation $(z,w)\mapsto(\lambda z,\lambda^2w)$ sends the equation $\operatorname{Im}w=|z|^2$ to \begin{align*}\operatorname{Im}(\lambda^2w)=|\lambda z|^2.\end{align*} Since $\operatorname{Im}(\lambda^2w)=\lambda^2\operatorname{Im}w$ and $|\lambda z|^2=\lambda^2|z|^2$, the equation is preserved. This is why the quadric is the flat model for the local CR geometry of the sphere. [/example] The example gives the expected normal target, but it does not yet provide a test for deciding whether a general hypersurface reaches that target. This motivates Cartan's equivalence principle, which replaces arbitrary coordinate changes by a canonical coframe whose curvature functions can be compared. [quotetheorem:9217] [citeproof:9217] The real-analytic hypothesis is used at the final integration step: Cartan's method first produces equivalence of complete jets and an analytic absolute parallelism, and analyticity is what promotes this formal agreement to an actual local CR equivalence. Strict pseudoconvexity is also essential, because the Levi form gives the nondegenerate Hermitian structure used to reduce the coframe bundle; for Levi-degenerate hypersurfaces the structure group does not reduce in the same way and additional degeneracies appear. The theorem is therefore not a finite-order test and does not say that matching a small list of curvature quantities is enough. This principle is complete but not yet convenient for calculation, so Chern-Moser normal form converts the same invariant content into coefficients of a defining equation. ## Chern-Moser Normal Coordinates The normal-form problem begins with a choice: if $M \subset \mathbb C^{n+1}$ is strictly pseudoconvex and $p \in M$, choose holomorphic coordinates $(z,w) \in \mathbb C^n \times \mathbb C$ centred at $p$ with $w=u+iv$ and complex tangent space $\{w=0\}$. The Levi form can then be diagonalised, so the defining equation starts with the model quadric $v=|z|^2$. [definition: Normal Coordinates For A Strictly Pseudoconvex Hypersurface] Let $M \subset \mathbb C^{n+1}$ be a smooth strictly pseudoconvex hypersurface and $p \in M$. Holomorphic coordinates $(z,w)$ centred at $p$, with $w=u+iv$, are normal coordinates at $p$ if $T_pM=\{v=0\}$, $T_p^{1,0}M=\{w=0\}$, and a local defining equation for $M$ has the form \begin{align*} v = |z|^2 + F(z,\bar z,u), \end{align*} where $F$ has weighted order at least $3$ for the weights $\operatorname{wt}(z)=\operatorname{wt}(\bar z)=1$ and $\operatorname{wt}(u)=2$. [/definition] Normal coordinates set up the calculation, but many higher-order terms in $F$ still change when the coordinates are changed without disturbing the Levi-normalised part. The next definition records the additional trace conditions that single out the coefficients which the normalising process is not allowed to remove. [definition: Chern-Moser Normal Form] For a strictly pseudoconvex hypersurface in $\mathbb C^{n+1}$, a formal defining equation \begin{align*} v = |z|^2 + \sum_{k,l \ge 2} F_{kl}(z,\bar z,u) \end{align*} is in Chern-Moser normal form if each $F_{kl}$ is bihomogeneous of degree $k$ in $z$ and degree $l$ in $\bar z$, and the Chern-Moser trace normalisations \begin{align*} \operatorname{tr} F_{22}=0, \qquad \operatorname{tr}^2 F_{23}=0, \qquad \operatorname{tr}^3 F_{33}=0 \end{align*} hold, with the trace taken using the Levi form. [/definition] For CR dimension $1$, the trace conditions have a degenerate form, and the surviving invariant is more naturally identified with Cartan's scalar curvature. The definition nevertheless raises the key existence question: can every strictly pseudoconvex hypersurface be brought to this form, and how much coordinate freedom remains after doing so? [quotetheorem:9218] [proofunderconstruction:9218] Real analyticity is needed for the convergence part of the theorem: in the smooth category the same weight-by-weight procedure gives formal normal forms, but formal coordinate changes need not converge to holomorphic coordinates. Strict pseudoconvexity is needed because the trace operator and its complementary normal-form decomposition use a positive definite Levi form; for indefinite or degenerate Levi forms the normalisation has a different structure and may have extra moduli. The uniqueness statement is also relative to the chosen normalisation data and the residual isotropy of the quadric, so it is not uniqueness among all possible holomorphic coordinates. The theorem reduces local equivalence to comparing normal-form coefficients after accounting for this residual group. The next example shows the flat case, where the algorithm has nothing to retain. [example: Normal Form Of A Quadric] Consider the quadric $Q\subset \mathbb C^2$ with coordinates $(z,w)$, where $w=u+iv$, and defining equation \begin{align*}v=|z|^2.\end{align*} Equivalently, $Q=\rho^{-1}(0)$ for \begin{align*}\rho(z,w)=v-|z|^2.\end{align*} At the origin, $|z|^2=z\bar z$ has no constant or linear term, while $v$ is linear in the real coordinates $(\operatorname{Re}z,\operatorname{Im}z,u,v)$. Therefore the linear part of $\rho$ at $0$ is $v$, so the tangent plane is \begin{align*}T_0Q=\{(\delta z,\delta w):\operatorname{Im}\delta w=0\}=\{v=0\}.\end{align*} The complex tangent space is the largest complex-linear subspace contained in $T_0Q$. Let $\delta w=a+ib$ with $a,b\in\mathbb R$. If $(\delta z,\delta w)$ belongs to a complex line contained in $T_0Q$, then $(\delta z,\delta w)\in T_0Q$ and $i(\delta z,\delta w)\in T_0Q$. The first condition gives \begin{align*}\operatorname{Im}\delta w=\operatorname{Im}(a+ib)=b=0.\end{align*} The second condition gives \begin{align*}\operatorname{Im}(i\delta w)=\operatorname{Im}(ia-b)=a=0.\end{align*} Thus $a=b=0$, hence $\delta w=0$, and so \begin{align*}T^{1,0}_0Q=\{(\delta z,\delta w):\delta w=0\}=\{w=0\}.\end{align*} The normal-coordinate form at the origin is \begin{align*}v=|z|^2+F(z,\bar z,u).\end{align*} For the quadric, the defining equation is already \begin{align*}v=|z|^2.\end{align*} Subtracting $|z|^2$ from both equations gives \begin{align*}F(z,\bar z,u)=0.\end{align*} Hence each bihomogeneous component in the Chern-Moser expansion satisfies \begin{align*}F_{kl}(z,\bar z,u)=0.\end{align*} The trace of the zero polynomial is zero, and repeated traces of the zero polynomial are still zero, so \begin{align*}\operatorname{tr}F_{22}=0,\qquad \operatorname{tr}^2F_{23}=0,\qquad \operatorname{tr}^3F_{33}=0.\end{align*} Thus the Chern-Moser normal form of $Q$ is exactly the Levi quadric, with no higher-weight terms remaining; this is the local normal form obtained from the sphere by the Cayley transform described above. [/example] Nonzero coefficients measure the failure of a hypersurface to be equivalent to the model. The first non-removable terms are the local CR curvature quantities. ## CR Curvature And Sphericity The central invariant question is whether the normal-form coefficients are merely coordinate artefacts or genuine curvature. Chern-Moser theory answers this by extracting tensorial combinations of the first surviving coefficients, giving a CR analogue of conformal curvature. [definition: Chern-Moser Curvature Tensor] Let $M \subset \mathbb C^{n+1}$ be a strictly pseudoconvex hypersurface in Chern-Moser normal form at $p$, and let $H_p^{1,0}=T_p^{1,0}M$. For CR dimension $n\ge 2$, the Chern-Moser curvature tensor at $p$ is the trace-free Hermitian tensor \begin{align*} S_p: H_p^{1,0}\times H_p^{1,0}\times \overline{H_p^{1,0}}\times \overline{H_p^{1,0}} \to \mathbb C \end{align*} obtained from the coefficient of bidegree $(2,2)$ in the normal-form expansion by lowering and contracting indices using the Levi form. [/definition] In CR dimension at least $2$, this tensor is the direct analogue of the Weyl tensor in conformal geometry. In CR dimension $1$, the corresponding local obstruction appears at higher weighted order and is Cartan's sixth-order invariant, so the phrase Chern-Moser curvature should be read through that low-dimensional modification. [quotetheorem:9219] [citeproof:9219] This theorem is a rigidity statement. Real analyticity is what turns vanishing of the complete formal curvature data into a genuine local biholomorphic equivalence; for merely smooth hypersurfaces, formal flatness does not by itself guarantee a convergent equivalence map. Strict pseudoconvexity is again the geometric hypothesis that supplies the sphere as the correct flat model and gives a positive Levi form for the normal-form construction. A strictly pseudoconvex hypersurface may be very close to the sphere to finite order, but local sphericity requires all curvature invariants and their covariant derivatives to vanish, not merely the Levi form or a small number of low-order coefficients. [example: Ellipsoid Boundary] Let \begin{align*}E_{a,b}=\left\{(z_1,z_2)\in \mathbb C^2: \frac{|z_1|^2}{a^2}+\frac{|z_2|^2}{b^2}<1\right\},\end{align*} where $a,b>0$, and define \begin{align*}\Phi(z_1,z_2)=\left(\frac{z_1}{a},\frac{z_2}{b}\right).\end{align*} The map is complex-linear because, for $\lambda,\mu\in \mathbb C$ and $z,w\in\mathbb C^2$, \begin{align*}\Phi(\lambda z+\mu w)=\left(\frac{\lambda z_1+\mu w_1}{a},\frac{\lambda z_2+\mu w_2}{b}\right)=\lambda\left(\frac{z_1}{a},\frac{z_2}{b}\right)+\mu\left(\frac{w_1}{a},\frac{w_2}{b}\right)=\lambda\Phi(z)+\mu\Phi(w).\end{align*} Its inverse is \begin{align*}\Phi^{-1}(\zeta_1,\zeta_2)=(a\zeta_1,b\zeta_2),\end{align*} since \begin{align*}\Phi^{-1}(\Phi(z_1,z_2))=\Phi^{-1}\left(\frac{z_1}{a},\frac{z_2}{b}\right)=\left(a\frac{z_1}{a},b\frac{z_2}{b}\right)=(z_1,z_2)\end{align*} and \begin{align*}\Phi(\Phi^{-1}(\zeta_1,\zeta_2))=\Phi(a\zeta_1,b\zeta_2)=\left(\frac{a\zeta_1}{a},\frac{b\zeta_2}{b}\right)=(\zeta_1,\zeta_2).\end{align*} Thus $\Phi$ is a biholomorphism of $\mathbb C^2$. We compute its effect on the boundary. If $(z_1,z_2)\in \partial E_{a,b}$, then \begin{align*}\frac{|z_1|^2}{a^2}+\frac{|z_2|^2}{b^2}=1.\end{align*} Writing $\Phi(z_1,z_2)=(\zeta_1,\zeta_2)$ gives $\zeta_1=z_1/a$ and $\zeta_2=z_2/b$. Since $a,b>0$ are real, \begin{align*}|\zeta_1|^2+|\zeta_2|^2=\left|\frac{z_1}{a}\right|^2+\left|\frac{z_2}{b}\right|^2=\frac{|z_1|^2}{a^2}+\frac{|z_2|^2}{b^2}=1.\end{align*} Hence $\Phi(z_1,z_2)\in S^3$, so $\Phi(\partial E_{a,b})\subset S^3$. Conversely, let $(\zeta_1,\zeta_2)\in S^3$, so \begin{align*}|\zeta_1|^2+|\zeta_2|^2=1.\end{align*} Put $(z_1,z_2)=\Phi^{-1}(\zeta_1,\zeta_2)=(a\zeta_1,b\zeta_2)$. Then \begin{align*}\frac{|z_1|^2}{a^2}+\frac{|z_2|^2}{b^2}=\frac{|a\zeta_1|^2}{a^2}+\frac{|b\zeta_2|^2}{b^2}=\frac{a^2|\zeta_1|^2}{a^2}+\frac{b^2|\zeta_2|^2}{b^2}=|\zeta_1|^2+|\zeta_2|^2=1.\end{align*} Thus $(z_1,z_2)\in\partial E_{a,b}$, and therefore $\Phi(\partial E_{a,b})=S^3$. It remains to check the CR structure. Since $d\Phi$ is the constant complex-linear map \begin{align*}d\Phi_{(z_1,z_2)}(\xi_1,\xi_2)=\left(\frac{\xi_1}{a},\frac{\xi_2}{b}\right),\end{align*} we have \begin{align*}d\Phi(i\xi_1,i\xi_2)=\left(\frac{i\xi_1}{a},\frac{i\xi_2}{b}\right)=i\left(\frac{\xi_1}{a},\frac{\xi_2}{b}\right)=i\,d\Phi(\xi_1,\xi_2).\end{align*} Because $\Phi$ maps $\partial E_{a,b}$ diffeomorphically onto $S^3$, its differential maps real tangent spaces to real tangent spaces; because $d\Phi$ commutes with multiplication by $i$, it maps the maximal complex subspace of each tangent space to the corresponding maximal complex subspace. Hence $\Phi|_{\partial E_{a,b}}$ is a local CR equivalence. The sphere is locally CR-equivalent to the flat quadric, whose Chern-Moser normal-form remainder is zero, so its local CR curvature invariants vanish. Since Chern-Moser curvature invariants are invariants of local CR equivalence, the same invariants vanish on $\partial E_{a,b}$. Thus even when $a\ne b$ and the Euclidean ellipsoid is not round, its boundary is CR-spherical; the CR invariants detect biholomorphic geometry, not Euclidean eccentricity. [/example] The ellipsoid calculation is a useful warning for boundary geometry. Curvature invariants must be invariant under biholomorphic changes of ambient coordinates, so quantities depending on Euclidean principal curvatures are not CR invariants by themselves. ## Webster Pseudo-Hermitian Geometry Normal forms are local and coordinate-based. Webster's pseudo-Hermitian geometry asks for an intrinsic differential-geometric package on a strictly pseudoconvex CR manifold once a contact form has been chosen. The choice of contact form is extra data, but it is exactly the data needed to define a connection and scalar curvature. [definition: Pseudo-Hermitian Structure] Let $M$ be a strictly pseudoconvex CR manifold of hypersurface type with CR distribution $H \subset TM$ and complex structure $J:H\to H$. A pseudo-Hermitian structure on $M$ is a smooth real $1$-form \begin{align*} \theta: TM \to \mathbb R \end{align*} such that $\ker \theta=H$ and the Levi form \begin{align*} L_\theta: H\times H\to \mathbb R, \qquad L_\theta(X,Y)=d\theta(X,JY) \end{align*} is positive definite on $H$. [/definition] A pseudo-Hermitian structure fixes the horizontal geometry but still leaves the transverse direction unnamed. The next object is needed because connection formulas and curvature decompositions require a canonical vector field complementary to $H$. [definition: Reeb Vector Field] Let $(M,H,J,\theta)$ be a pseudo-Hermitian manifold. The Reeb vector field $T$ is the unique vector field satisfying \begin{align*} \theta(T)=1, \qquad d\theta(T,X)=0 \end{align*} for every vector field $X$ on $M$. [/definition] The Reeb field gives the splitting $TM=H\oplus \mathbb R T$, so the Levi form can now be extended from $H$ to the whole tangent bundle. This motivates the Webster metric, the Riemannian representative attached to the chosen contact form. [definition: Webster Metric] Let $(M,H,J,\theta)$ be a pseudo-Hermitian manifold with Reeb vector field $T$. The Webster metric is the smooth symmetric tensor \begin{align*} g_\theta: TM\times TM\to \mathbb R \end{align*} defined by the following conditions: for $X,Y\in H$, \begin{align*} g_\theta(X,Y)=d\theta(X,JY), \end{align*} while \begin{align*} g_\theta(T,T)=1, \qquad g_\theta(T,X)=0 \quad \text{for } X\in H. \end{align*} [/definition] The Webster metric is not a CR invariant independent of choices, because rescaling $\theta$ changes it. Its strength is that it turns CR questions into metric-compatible differentiation problems, which motivates introducing the connection adapted to $H$, $J$, and $\theta$. [definition: Tanaka-Webster Connection] Let $(M,H,J,\theta)$ be a strictly pseudoconvex pseudo-Hermitian manifold. The Tanaka-Webster connection is the linear connection \begin{align*} \nabla: \Gamma(TM)\times \Gamma(TM)\to \Gamma(TM), \qquad (X,Y)\mapsto \nabla_XY, \end{align*} preserving $H$, $J$, $\theta$, and $g_\theta$, whose torsion satisfies the standard pseudo-Hermitian torsion normalisations. [/definition] The connection definition packages the admissible differentiation rules, but a connection with torsion also has components that influence CR curvature identities and [integration by parts](/theorems/210). This motivates defining the Webster torsion as the remaining Reeb-horizontal torsion component after the pseudo-Hermitian normalisations are imposed. [definition: Webster Torsion] Let $\nabla$ be the Tanaka-Webster connection on $(M,H,J,\theta)$. The Webster torsion is the horizontal tensor \begin{align*} A: H\times H\to \mathbb R \end{align*} defined by \begin{align*} A(X,Y)=g_\theta(\operatorname{Tor}_\nabla(T,X),Y) \end{align*} for $X,Y\in H$, where $T$ is the Reeb vector field. [/definition] Torsion measures one part of the pseudo-Hermitian connection, but it does not by itself measure how horizontal directions curve as they are transported around small loops. The obstruction is that the ordinary Riemannian scalar curvature of an adapted metric also sees the Reeb direction, while CR geometry needs an invariant extracted from the horizontal bundle and the Levi form. Thus the curvature operator of the Tanaka-Webster connection must be formed first, and only then can one take the Ricci and scalar contractions relevant to pseudo-Hermitian geometry. These horizontal contractions supply the scalar quantity used in CR Yamabe problems, boundary estimates, and comparisons with Chern-Moser flatness. The formal definitions below fix the distinction between the full connection curvature and the Webster Ricci and scalar curvatures obtained from it. [definition: Webster Curvature And Scalar Curvature] Let $\nabla$ be the Tanaka-Webster connection on a strictly pseudoconvex pseudo-Hermitian manifold. Its curvature operator is the map \begin{align*} R^\nabla: \Gamma(TM)\times \Gamma(TM)\times \Gamma(TM)\to \Gamma(TM) \end{align*} defined by \begin{align*} R^\nabla(X,Y)Z=\nabla_X\nabla_YZ-\nabla_Y\nabla_XZ-\nabla_{[X,Y]}Z. \end{align*} The Webster Ricci tensor and Webster scalar curvature are the Ricci and scalar contractions of $R^\nabla$ over the horizontal CR directions using the Levi form. [/definition] Unlike the Chern-Moser curvature system, Webster curvature depends on the chosen contact form. Under conformal change $\hat\theta=e^{\varphi}\theta$, the torsion and scalar curvature transform by explicit differential formulas, so pseudo-Hermitian geometry is best viewed as the calculus of representatives within a CR conformal class. [example: Standard Sphere With Its Contact Form] On $S^{2n+1}\subset \mathbb C^{n+1}$, write $z_j=x_j+iy_j$ and take \begin{align*} \theta = \frac{i}{2}\sum_{j=1}^{n+1}(\bar z_j\,dz_j-z_j\,d\bar z_j)\bigg|_{S^{2n+1}}. \end{align*} Using $dz_j=dx_j+i\,dy_j$, $d\bar z_j=dx_j-i\,dy_j$, and $\bar z_j=x_j-iy_j$, the $j$th summand expands as \begin{align*} \bar z_j\,dz_j=(x_j-iy_j)(dx_j+i\,dy_j)=x_j\,dx_j+i x_j\,dy_j-i y_j\,dx_j+y_j\,dy_j. \end{align*} Similarly, \begin{align*} z_j\,d\bar z_j=(x_j+iy_j)(dx_j-i\,dy_j)=x_j\,dx_j-i x_j\,dy_j+i y_j\,dx_j+y_j\,dy_j. \end{align*} Subtracting gives \begin{align*} \bar z_j\,dz_j-z_j\,d\bar z_j=2i x_j\,dy_j-2i y_j\,dx_j. \end{align*} Multiplying by $i/2$ gives \begin{align*} \frac{i}{2}(\bar z_j\,dz_j-z_j\,d\bar z_j)=y_j\,dx_j-x_j\,dy_j. \end{align*} Thus \begin{align*} \theta=\sum_{j=1}^{n+1}(y_j\,dx_j-x_j\,dy_j)\bigg|_{S^{2n+1}}. \end{align*} Define \begin{align*} T=\sum_{j=1}^{n+1}\left(y_j\frac{\partial}{\partial x_j}-x_j\frac{\partial}{\partial y_j}\right). \end{align*} Since $|z|^2=\sum_{j=1}^{n+1}(x_j^2+y_j^2)$, we have \begin{align*} T(|z|^2)=\sum_{j=1}^{n+1}\left(y_j\,2x_j-x_j\,2y_j\right)=0. \end{align*} Therefore $T$ is tangent to the level set $|z|^2=1$, hence tangent to $S^{2n+1}$. Evaluating $\theta$ on $T$ gives \begin{align*} \theta(T)=\sum_{j=1}^{n+1}\left(y_j\,dx_j(T)-x_j\,dy_j(T)\right)=\sum_{j=1}^{n+1}(y_j^2+x_j^2)=|z|^2. \end{align*} On $S^{2n+1}$ this becomes \begin{align*} \theta(T)=1. \end{align*} Next, \begin{align*} d(y_j\,dx_j-x_j\,dy_j)=dy_j\wedge dx_j-dx_j\wedge dy_j=-2\,dx_j\wedge dy_j. \end{align*} Hence \begin{align*} d\theta=-2\sum_{j=1}^{n+1}dx_j\wedge dy_j. \end{align*} Let $X=\sum_j(a_j\partial_{x_j}+b_j\partial_{y_j})$ be tangent to the sphere. Tangency to $|z|^2=1$ means \begin{align*} X(|z|^2)=\sum_{j=1}^{n+1}(2x_ja_j+2y_jb_j)=0. \end{align*} Equivalently, \begin{align*} \sum_{j=1}^{n+1}(x_ja_j+y_jb_j)=0. \end{align*} For each $j$, \begin{align*} (dx_j\wedge dy_j)(T,X)=dx_j(T)dy_j(X)-dx_j(X)dy_j(T)=y_jb_j+a_jx_j. \end{align*} Therefore \begin{align*} d\theta(T,X)=-2\sum_{j=1}^{n+1}(y_jb_j+x_ja_j)=0. \end{align*} Thus $T$ satisfies $\theta(T)=1$ and $d\theta(T,X)=0$ for every tangent vector $X$, so $T$ is the Reeb vector field. The flow of $T$ solves \begin{align*} \frac{dx_j}{dt}=y_j,\qquad \frac{dy_j}{dt}=-x_j. \end{align*} For $z_j=x_j+iy_j$ this gives \begin{align*} \frac{dz_j}{dt}=\frac{dx_j}{dt}+i\frac{dy_j}{dt}=y_j-i x_j=-i(x_j+iy_j)=-iz_j. \end{align*} Hence \begin{align*} z_j(t)=e^{-it}z_j(0), \end{align*} so the Reeb flow is the Hopf circle action with this sign convention. The CR distribution is $H=\ker\theta\subset TS^{2n+1}$. If $X=\sum_j(a_j\partial_{x_j}+b_j\partial_{y_j})$ is tangent to the sphere, then the ambient complex structure sends it to \begin{align*} JX=\sum_{j=1}^{n+1}\left(-b_j\frac{\partial}{\partial x_j}+a_j\frac{\partial}{\partial y_j}\right). \end{align*} The vector $JX$ is tangent to the sphere exactly when \begin{align*} \sum_{j=1}^{n+1}(x_j(-b_j)+y_ja_j)=\sum_{j=1}^{n+1}(y_ja_j-x_jb_j)=\theta(X)=0. \end{align*} Thus $H=TS^{2n+1}\cap JTS^{2n+1}$, and the standard complex structure preserves $H$. The contact form is invariant under $U(n+1)$. Indeed, for $\zeta=Uz$ with $U^*U=I$, one has $d\zeta=U\,dz$ and \begin{align*} \sum_{j=1}^{n+1}\bar\zeta_j\,d\zeta_j=\bar z^{\,t}U^*U\,dz=\bar z^{\,t}dz=\sum_{j=1}^{n+1}\bar z_j\,dz_j. \end{align*} The conjugate term is invariant in the same way, so $\theta$ is $U(n+1)$-invariant. For the standard pseudo-Hermitian sphere, the Tanaka-Webster connection is the homogeneous connection induced by this $U(n+1)$-invariant structure. Its Reeb-horizontal torsion component vanishes by the standard torsion-free property of the Sasakian sphere, so the Webster torsion tensor is \begin{align*} A=0. \end{align*} Because $U(n+1)$ acts transitively on $S^{2n+1}$ and preserves $\theta$, it also preserves the Tanaka-Webster connection and its curvature contractions. Therefore the Webster scalar curvature has the same value at every point of $S^{2n+1}$. The standard sphere consequently has vanishing pseudo-Hermitian torsion and constant Webster scalar curvature, matching its role as the homogeneous flat model in CR geometry. [/example] The chapter therefore gives two complementary languages for the same geometry. Chern-Moser normal form is the local classification tool, identifying curvature as the obstruction to reducing a hypersurface to the quadric. Webster geometry is the intrinsic calculus used in estimates, variational problems, and global CR geometry once a contact form has been selected. Once the local invariants are understood, the natural analytic question is how they affect boundary-value problems for the full $\bar\partial$ operator. The $\bar\partial$-Neumann problem brings these geometric features into estimates, regularity, and compactness at the boundary. # 8. The $\bar\partial$-Neumann Problem and Boundary Regularity Chapter 4 studied the boundary operator $\bar{\partial}_b$ and its Kohn Laplacian. This chapter turns the interior $\bar{\partial}$ equation into a boundary-value problem: in the interior of a domain the complex Laplacian behaves like an elliptic operator, but boundary conditions force it to interact with the Levi geometry of $\partial\Omega$. The main question is when the analytic solution operator gains enough regularity to see the boundary smoothly, and when geometry prevents this from happening. The chapter begins with the Hilbert-space formulation of the $\bar{\partial}$-Neumann problem, then specialises to strictly pseudoconvex domains where Kohn's estimate gives a half derivative of subelliptic gain. We then discuss global regularity of the Neumann operator on smooth data, before moving to finite type and the failure of regularity on worm domains. This is the bridge from CR geometry to the analytic machinery behind boundary regularity in several complex variables. ## The Complex Laplacian with Boundary Conditions What boundary condition should replace the Dirichlet or Neumann condition when the unknown is a $(0,q)$-form and the equation is built from $\bar{\partial}$? The answer is not imposed component by component; it is encoded by the Hilbert-space adjoint $\bar{\partial}^*$ and by the requirement that both $u$ and $\bar{\partial}u$ lie in the domain of this adjoint. This formulation makes the boundary condition coordinate-independent and keeps the equation compatible with the $\bar{\partial}$ complex. Let $\Omega \subset \mathbb C^n$ be a bounded domain with $C^\infty$ boundary, and fix $0 \,\le q \le n$. Write $L^2_{0,q}(\Omega)$ for the [Hilbert space](/page/Hilbert%20Space) of square-integrable $(0,q)$-forms on $\Omega$, with [inner product](/page/Inner%20Product) induced by the Euclidean Hermitian metric and Lebesgue measure. [definition: Square-Integrable Forms] A $(0,q)$-form $u$ on $\Omega$ belongs to $L^2_{0,q}(\Omega)$ if it can be written \begin{align*} u = \sum_{|J|=q}' u_J\, d\bar z_J, \end{align*} where each coefficient $u_J$ lies in $L^2(\Omega)$. [/definition] Here the prime means that $J=(j_1<\cdots<j_q)$ ranges over strictly increasing multi-indices, and $d\bar z_J=d\bar z_{j_1}\wedge\cdots\wedge d\bar z_{j_q}$. The Hilbert norm is \begin{align*} \|u\|_{L^2}^2 = \sum_{|J|=q}' \int_\Omega |u_J|^2\, d\mathcal L^{2n}. \end{align*} To define the differential operator next, we need the $L^2$ framework to accept weak derivatives rather than only smooth coefficient functions. [definition: Maximal Dolbeault Operator] The maximal $\bar{\partial}$ operator on $L^2_{0,q}(\Omega)$ is the closed densely-defined operator \begin{align*} \bar{\partial}: \operatorname{Dom}(\bar{\partial}) \subset L^2_{0,q}(\Omega) \to L^2_{0,q+1}(\Omega) \end{align*} defined distributionally by \begin{align*} \bar{\partial}u = \sum_{|J|=q}'\sum_{k=1}^n \frac{\partial u_J}{\partial \bar z_k}\, d\bar z_k \wedge d\bar z_J, \end{align*} with domain consisting of all $u\in L^2_{0,q}(\Omega)$ for which this distribution belongs to $L^2_{0,q+1}(\Omega)$. [/definition] The adjoint $\bar{\partial}^*$ is the Hilbert-space adjoint of this maximal operator. On smooth forms its interior expression is the formal adjoint, but its domain contains boundary conditions. If $\rho$ is a smooth defining function for $\Omega$, then the condition $u \in \operatorname{Dom}(\bar{\partial}^*)$ imposes a contraction condition with $\bar{\partial}\rho$ along $\partial\Omega$. This prepares the next definition, where the operator and its boundary domain are combined into the complex Laplacian. [definition: Complex Laplacian] For $0\le q\le n$, the $\bar{\partial}$-Neumann complex Laplacian is the unbounded operator \begin{align*} \Box_q u = \bar{\partial}\bar{\partial}^*u + \bar{\partial}^*\bar{\partial}u \end{align*} on $L^2_{0,q}(\Omega)$, with domain \begin{align*} \operatorname{Dom}(\Box_q)=\{u\in \operatorname{Dom}(\bar{\partial})\cap \operatorname{Dom}(\bar{\partial}^*):\bar{\partial}u\in \operatorname{Dom}(\bar{\partial}^*),\ \bar{\partial}^*u\in \operatorname{Dom}(\bar{\partial})\}. \end{align*} [/definition] The equation $\Box_q u=f$ is the $\bar{\partial}$-Neumann problem. The word "Neumann" reflects the way the boundary condition appears through an adjoint condition, but it is not the scalar normal derivative condition from the real Laplacian. [definition: Bar Partial-Neumann Problem] Given $f\in L^2_{0,q}(\Omega)$, the $\bar{\partial}$-Neumann problem in degree $q$ asks for $u\in \operatorname{Dom}(\Box_q)$ satisfying \begin{align*} \Box_q u=f. \end{align*} [/definition] The natural energy form is the quadratic form associated to $\Box_q$: \begin{align*} Q_q(u,v)=(\bar{\partial}u,\bar{\partial}v)_{L^2}+ (\bar{\partial}^*u,\bar{\partial}^*v)_{L^2}, \end{align*} for $u,v\in \operatorname{Dom}(\bar{\partial})\cap\operatorname{Dom}(\bar{\partial}^*)$. In particular, \begin{align*} Q_q(u,u)=\|\bar{\partial}u\|_{L^2}^2+\|\bar{\partial}^*u\|_{L^2}^2. \end{align*} This form is the basic object because estimates for $Q_q(u,u)$ control the solvability and regularity of the Neumann problem. [quotetheorem:9220] [citeproof:9220] This identity explains why the problem is controlled by a first-order energy even though $\Box_q$ is second order. Boundary regularity is therefore reduced to finding estimates that recover derivatives of $u$ from $\bar{\partial}u$, $\bar{\partial}^*u$, and lower-order norms. [example: Computation of the Complex Laplacian in Two Variables] Let $\Omega\subset\mathbb C^2$ and let \begin{align*} u=u_1\,d\bar z_1+u_2\,d\bar z_2 \end{align*} be a smooth $(0,1)$-form. We compute the interior second-order expression for $\Box_1u$, so the boundary conditions encoded in $\operatorname{Dom}(\bar{\partial}^*)$ are not used in the local calculation. By the coefficient formula for $\bar{\partial}$, \begin{align*} \bar{\partial}u=\frac{\partial u_1}{\partial\bar z_1}\,d\bar z_1\wedge d\bar z_1+\frac{\partial u_1}{\partial\bar z_2}\,d\bar z_2\wedge d\bar z_1+\frac{\partial u_2}{\partial\bar z_1}\,d\bar z_1\wedge d\bar z_2+\frac{\partial u_2}{\partial\bar z_2}\,d\bar z_2\wedge d\bar z_2. \end{align*} The exterior product is alternating, so $d\bar z_1\wedge d\bar z_1=0$ and $d\bar z_2\wedge d\bar z_2=0$. Also $d\bar z_2\wedge d\bar z_1=-d\bar z_1\wedge d\bar z_2$. Therefore the two nonzero terms combine as \begin{align*} \bar{\partial}u=-\frac{\partial u_1}{\partial\bar z_2}\,d\bar z_1\wedge d\bar z_2+\frac{\partial u_2}{\partial\bar z_1}\,d\bar z_1\wedge d\bar z_2. \end{align*} Thus \begin{align*} \bar{\partial}u=\left(\frac{\partial u_2}{\partial \bar z_1}-\frac{\partial u_1}{\partial \bar z_2}\right)d\bar z_1\wedge d\bar z_2. \end{align*} The interior formal adjoint on smooth $(0,1)$-forms is \begin{align*} \bar{\partial}^*u=-\frac{\partial u_1}{\partial z_1}-\frac{\partial u_2}{\partial z_2}. \end{align*} Applying $\bar{\partial}$ to this scalar function gives \begin{align*} \bar{\partial}\bar{\partial}^*u=\frac{\partial}{\partial\bar z_1}\left(-\frac{\partial u_1}{\partial z_1}-\frac{\partial u_2}{\partial z_2}\right)d\bar z_1+\frac{\partial}{\partial\bar z_2}\left(-\frac{\partial u_1}{\partial z_1}-\frac{\partial u_2}{\partial z_2}\right)d\bar z_2. \end{align*} Distributing the two derivatives across the sums, \begin{align*} \bar{\partial}\bar{\partial}^*u=\left(-\frac{\partial^2u_1}{\partial\bar z_1\partial z_1}-\frac{\partial^2u_2}{\partial\bar z_1\partial z_2}\right)d\bar z_1+\left(-\frac{\partial^2u_1}{\partial\bar z_2\partial z_1}-\frac{\partial^2u_2}{\partial\bar z_2\partial z_2}\right)d\bar z_2. \end{align*} Set \begin{align*} a=\frac{\partial u_2}{\partial \bar z_1}-\frac{\partial u_1}{\partial \bar z_2}. \end{align*} Then $\bar{\partial}u=a\,d\bar z_1\wedge d\bar z_2$. For a smooth $(0,2)$-form in two complex variables, the interior formal adjoint is \begin{align*} \bar{\partial}^*(a\,d\bar z_1\wedge d\bar z_2)=\frac{\partial a}{\partial z_2}\,d\bar z_1-\frac{\partial a}{\partial z_1}\,d\bar z_2. \end{align*} The first coefficient is \begin{align*} \frac{\partial a}{\partial z_2}=\frac{\partial}{\partial z_2}\left(\frac{\partial u_2}{\partial \bar z_1}-\frac{\partial u_1}{\partial \bar z_2}\right)=\frac{\partial^2u_2}{\partial z_2\partial\bar z_1}-\frac{\partial^2u_1}{\partial z_2\partial\bar z_2}. \end{align*} The second coefficient is \begin{align*} \frac{\partial a}{\partial z_1}=\frac{\partial}{\partial z_1}\left(\frac{\partial u_2}{\partial \bar z_1}-\frac{\partial u_1}{\partial \bar z_2}\right)=\frac{\partial^2u_2}{\partial z_1\partial\bar z_1}-\frac{\partial^2u_1}{\partial z_1\partial\bar z_2}. \end{align*} Hence \begin{align*} \bar{\partial}^*\bar{\partial}u=\left(\frac{\partial^2u_2}{\partial z_2\partial\bar z_1}-\frac{\partial^2u_1}{\partial z_2\partial\bar z_2}\right)d\bar z_1-\left(\frac{\partial^2u_2}{\partial z_1\partial\bar z_1}-\frac{\partial^2u_1}{\partial z_1\partial\bar z_2}\right)d\bar z_2. \end{align*} Now add the $d\bar z_1$-coefficients from $\bar{\partial}\bar{\partial}^*u$ and $\bar{\partial}^*\bar{\partial}u$: \begin{align*} -\frac{\partial^2u_1}{\partial\bar z_1\partial z_1}-\frac{\partial^2u_2}{\partial\bar z_1\partial z_2}+\frac{\partial^2u_2}{\partial z_2\partial\bar z_1}-\frac{\partial^2u_1}{\partial z_2\partial\bar z_2}. \end{align*} Since the coefficients are smooth, mixed partial derivatives commute. Thus \begin{align*} -\frac{\partial^2u_2}{\partial\bar z_1\partial z_2}+\frac{\partial^2u_2}{\partial z_2\partial\bar z_1}=0. \end{align*} Also \begin{align*} \frac{\partial^2u_1}{\partial\bar z_1\partial z_1}=\frac{\partial^2u_1}{\partial z_1\partial\bar z_1}. \end{align*} Therefore the $d\bar z_1$-coefficient is \begin{align*} -\frac{\partial^2u_1}{\partial z_1\partial\bar z_1}-\frac{\partial^2u_1}{\partial z_2\partial\bar z_2}. \end{align*} For the $d\bar z_2$-coefficient, the sum is \begin{align*} -\frac{\partial^2u_1}{\partial\bar z_2\partial z_1}-\frac{\partial^2u_2}{\partial\bar z_2\partial z_2}-\frac{\partial^2u_2}{\partial z_1\partial\bar z_1}+\frac{\partial^2u_1}{\partial z_1\partial\bar z_2}. \end{align*} Commutation of mixed partial derivatives gives \begin{align*} -\frac{\partial^2u_1}{\partial\bar z_2\partial z_1}+\frac{\partial^2u_1}{\partial z_1\partial\bar z_2}=0. \end{align*} Also \begin{align*} \frac{\partial^2u_2}{\partial\bar z_2\partial z_2}=\frac{\partial^2u_2}{\partial z_2\partial\bar z_2}. \end{align*} Therefore the $d\bar z_2$-coefficient is \begin{align*} -\frac{\partial^2u_2}{\partial z_1\partial\bar z_1}-\frac{\partial^2u_2}{\partial z_2\partial\bar z_2}. \end{align*} Combining the two coefficients, \begin{align*} \Box_1u=-\sum_{j=1}^2\frac{\partial^2u_1}{\partial z_j\partial\bar z_j}\,d\bar z_1-\sum_{j=1}^2\frac{\partial^2u_2}{\partial z_j\partial\bar z_j}\,d\bar z_2. \end{align*} Writing $z_j=x_j+iy_j$, the Wirtinger derivatives are \begin{align*} \frac{\partial}{\partial z_j}=\frac12\left(\frac{\partial}{\partial x_j}-i\frac{\partial}{\partial y_j}\right),\qquad \frac{\partial}{\partial\bar z_j}=\frac12\left(\frac{\partial}{\partial x_j}+i\frac{\partial}{\partial y_j}\right). \end{align*} Multiplying the operators on a smooth test coefficient $\phi$ gives \begin{align*} \frac{\partial}{\partial z_j}\frac{\partial\phi}{\partial\bar z_j}=\frac14\left(\frac{\partial}{\partial x_j}-i\frac{\partial}{\partial y_j}\right)\left(\frac{\partial\phi}{\partial x_j}+i\frac{\partial\phi}{\partial y_j}\right). \end{align*} Expanding the right side, \begin{align*} \frac{\partial}{\partial z_j}\frac{\partial\phi}{\partial\bar z_j}=\frac14\left(\frac{\partial^2\phi}{\partial x_j^2}+i\frac{\partial^2\phi}{\partial x_j\partial y_j}-i\frac{\partial^2\phi}{\partial y_j\partial x_j}+\frac{\partial^2\phi}{\partial y_j^2}\right). \end{align*} The real mixed partial derivatives commute, so \begin{align*} i\frac{\partial^2\phi}{\partial x_j\partial y_j}-i\frac{\partial^2\phi}{\partial y_j\partial x_j}=0. \end{align*} Thus \begin{align*} \frac{\partial}{\partial z_j}\frac{\partial\phi}{\partial\bar z_j}=\frac14\left(\frac{\partial^2\phi}{\partial x_j^2}+\frac{\partial^2\phi}{\partial y_j^2}\right). \end{align*} Applied to each coefficient $u_1,u_2$, this shows that the interior principal part of $\Box_1$ is $-\frac14\Delta$ on each coefficient. The second-order interior operator separates componentwise, while the $\bar{\partial}^*$ boundary condition still couples the coefficients and is where the Levi geometry enters. [/example] ## Strict Pseudoconvexity and the Half-Derivative Estimate Why should curvature of the boundary give analytic control of derivatives? Near a boundary point, normal derivatives can be recovered from the equation, while tangential derivatives are controlled only through the commutators of complex tangential directions. Strict pseudoconvexity makes those commutators point in the missing normal direction, producing a subelliptic gain rather than a full elliptic estimate. Let $\rho$ be a smooth defining function for $\Omega$, with $\Omega=\{\rho<0\}$ and $d\rho\ne 0$ on $\partial\Omega$. The complex tangent space at $p\in\partial\Omega$ is \begin{align*} T_p^{1,0}(\partial\Omega)=\{L\in T_p^{1,0}\mathbb C^n: L\rho=0\}. \end{align*} The Levi form is the Hermitian form obtained from the complex Hessian of $\rho$ on this complex tangent space. [definition: Strictly Pseudoconvex Boundary] A smoothly bounded domain $\Omega\subset\mathbb C^n$ is strictly pseudoconvex if for some smooth defining function $\rho$, the Levi form \begin{align*} \mathcal L_\rho(p;L)=\sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial \bar z_k}(p)L_j\overline{L_k} \end{align*} is positive definite on $T_p^{1,0}(\partial\Omega)$ for every $p\in\partial\Omega$. [/definition] Strict pseudoconvexity is invariant under changing the defining function by a positive smooth factor. The analytic content is that the Levi form supplies the missing boundary positivity needed for the following Morrey-Kohn-Hörmander identity. [quotetheorem:3685] [citeproof:3685] The identity is the point at which CR geometry enters the estimate: positivity of the Levi form appears as a boundary term rather than as an interior elliptic coefficient. This is why strict pseudoconvexity is the right geometric hypothesis for the $\bar\partial$-Neumann problem. The identity still has a limitation: the boundary condition leaves one characteristic direction uncontrolled, so the estimate cannot simply be the usual full first-derivative elliptic bound. The remaining analytic problem is to convert the positive boundary terms and interior derivatives into a precise fractional Sobolev gain. For this page, the important consequence is conceptual rather than a new theorem card: strict Levi positivity supplies a local subelliptic gain of order $1/2$ for the $\bar\partial$-Neumann problem. That gain is weaker than elliptic regularity because the boundary problem remains characteristic, but it is still genuine boundary regularity. It explains how the Morrey-Kohn-Hörmander identity compensates for the missing normal ellipticity and gives the analytic input needed for the Hilbert-space construction to behave well. This local regularity estimate should now be separated from the Hilbert-space existence statement. The $\bar\partial$-Neumann operator $N_q$ is the bounded inverse, on the orthogonal complement of the harmonic forms, of the complex Laplacian $\Box_q=\bar\partial\bar\partial^*+\bar\partial^*\bar\partial$. Existence of $N_q$ gives the $L^2$ solvability framework; the subelliptic estimate is the additional boundary regularity input that explains why smooth data can be expected to produce solutions with controlled boundary behavior. Thus the local estimate is not a second solvability theorem, a global [regularity theorem](/theorems/2750), or a Bergman projection statement; it is the analytic mechanism used after solvability has been established. [example: Unit Ball Estimate] Let $\Omega=B(0,1)\subset\mathbb C^n$ and set $\rho(z)=|z|^2-1=\sum_{\ell=1}^n z_\ell\bar z_\ell-1$. We first verify that this defining function has positive definite Levi form on the complex tangent space of the boundary. Since $\rho(z)<0$ exactly when $|z|<1$ and $\rho(z)=0$ exactly when $|z|=1$, we have $\Omega=\{\rho<0\}$ and $\partial\Omega=\{\rho=0\}$. Writing $z_\ell=x_\ell+iy_\ell$, we get \begin{align*} \rho(z)=\sum_{\ell=1}^n(x_\ell^2+y_\ell^2)-1. \end{align*} Therefore \begin{align*} \nabla\rho(p)=(2x_1,2y_1,\dots,2x_n,2y_n). \end{align*} If $p\in\partial\Omega$, then $|p|=1$, so \begin{align*} |\nabla\rho(p)|^2=(2x_1)^2+(2y_1)^2+\cdots+(2x_n)^2+(2y_n)^2. \end{align*} Combining the terms gives \begin{align*} |\nabla\rho(p)|^2=4\sum_{\ell=1}^n(x_\ell^2+y_\ell^2)=4|p|^2=4. \end{align*} Thus $|\nabla\rho(p)|=2\ne 0$ on $\partial\Omega$, so $\rho$ is a smooth defining function for the unit ball. For $1\le k\le n$, compute the first antiholomorphic derivative: \begin{align*} \frac{\partial \rho}{\partial \bar z_k}=\sum_{\ell=1}^n \frac{\partial}{\partial \bar z_k}(z_\ell\bar z_\ell)-\frac{\partial}{\partial \bar z_k}(1). \end{align*} The constant term has derivative $0$. Since $z_\ell$ is independent of $\bar z_k$ and $\partial\bar z_\ell/\partial\bar z_k=\delta_{\ell k}$, the product rule gives \begin{align*} \frac{\partial}{\partial \bar z_k}(z_\ell\bar z_\ell)=\frac{\partial z_\ell}{\partial\bar z_k}\bar z_\ell+z_\ell\frac{\partial\bar z_\ell}{\partial\bar z_k}=0\cdot \bar z_\ell+z_\ell\delta_{\ell k}=z_\ell\delta_{\ell k}. \end{align*} Therefore \begin{align*} \frac{\partial \rho}{\partial \bar z_k}=\sum_{\ell=1}^n z_\ell\delta_{\ell k}=z_k. \end{align*} Differentiating this identity with respect to $z_j$ gives \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\bar z_k}=\frac{\partial}{\partial z_j}(z_k)=\delta_{jk}. \end{align*} Thus the complex Hessian matrix of $\rho$ is the identity matrix $(\delta_{jk})$. Now fix $p\in\partial B(0,1)$ and let \begin{align*} L=\sum_{j=1}^n L_j\frac{\partial}{\partial z_j}\in T_p^{1,0}(\partial B(0,1)). \end{align*} Since \begin{align*} \frac{\partial \rho}{\partial z_j}=\bar z_j, \end{align*} the tangency condition $L\rho(p)=0$ is \begin{align*} \sum_{j=1}^n L_j\overline{p_j}=0. \end{align*} The Levi form at $p$ on this tangent vector is \begin{align*} \mathcal L_\rho(p;L)=\sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial\bar z_k}(p)L_j\overline{L_k}. \end{align*} Substituting $\partial^2\rho/\partial z_j\partial\bar z_k=\delta_{jk}$ yields \begin{align*} \mathcal L_\rho(p;L)=\sum_{j,k=1}^n \delta_{jk}L_j\overline{L_k}. \end{align*} The summands with $j\ne k$ vanish because $\delta_{jk}=0$, and the summands with $j=k$ remain because $\delta_{jj}=1$. Hence \begin{align*} \mathcal L_\rho(p;L)=\sum_{j=1}^n L_j\overline{L_j}. \end{align*} Since $L_j\overline{L_j}=|L_j|^2$, this is \begin{align*} \mathcal L_\rho(p;L)=\sum_{j=1}^n |L_j|^2. \end{align*} If $L\ne 0$, then at least one coefficient $L_j$ is nonzero, so at least one term $|L_j|^2$ is positive and every other term is nonnegative. Therefore \begin{align*} \mathcal L_\rho(p;L)>0. \end{align*} Thus the Levi form is positive definite on $T_p^{1,0}(\partial B(0,1))$ for every boundary point $p$, so the unit ball is strictly pseudoconvex. By *[Kohn Subelliptic Estimate for Strongly Pseudoconvex Domains](/theorems/3735)*, for every $1\le q\le n$ there is a constant $C>0$ such that every smooth $(0,q)$-form $u$ satisfying the $\bar{\partial}$-Neumann boundary condition obeys \begin{align*} \|u\|_{H^{1/2}}^2\le C\left(\|\bar{\partial}u\|_{L^2}^2+\|\bar{\partial}^*u\|_{L^2}^2+\|u\|_{L^2}^2\right). \end{align*} For the unit ball, the geometric input in this estimate is the identity matrix $(\delta_{jk})$ obtained as the complex Hessian of $\rho(z)=|z|^2-1$, so the Levi form is simply the squared length $\sum_{j=1}^n |L_j|^2$ on complex tangent vectors. [/example] ## Smoothness of the Neumann Operator up to the Boundary Does the half-derivative estimate upgrade an $L^2$ solution into a solution smooth on $\overline{\Omega}$ when the data are smooth? The difficulty is that differentiating tangentially preserves the boundary condition only up to commutators, while differentiating normally interacts with the equation itself. Kohn's global regularity theorem is the statement that, for strictly pseudoconvex domains, the subelliptic gain can be iterated without losing control of these boundary terms. [definition: Global Regularity of the Neumann Operator] The $\bar{\partial}$-Neumann operator $N_q$ is globally regular on $\Omega$ if for every $m\in\mathbb N$ there exists $C_m>0$ such that \begin{align*} \|N_q f\|_{H^m}\le C_m\|f\|_{H^m} \end{align*} for all smooth $(0,q)$-forms $f$ lying in the domain of $N_q$ and orthogonal to $\ker\Box_q$. [/definition] Sobolev estimates of all orders imply preservation of $C^\infty(\overline{\Omega})$ by Sobolev embedding and standard localisation. The obstruction is that each differentiation of the boundary condition produces commutator terms that must be reabsorbed by the subelliptic estimate. In this course, Kohn global regularity is the assertion that this reabsorption works uniformly at every Sobolev order on strictly pseudoconvex domains: for smooth data orthogonal to the harmonic forms, $N_q$ produces a solution smooth up to $\partial\Omega$. Thus the half-derivative estimate is not being quoted a second time as the theorem; it is the engine behind the global regularity statement used here. [remark: Degree Zero] In degree $q=0$, the $\bar{\partial}$-Neumann problem interacts with holomorphic functions because $\ker\bar{\partial}$ is infinite-dimensional. The standard closed-range and Neumann theory is therefore most naturally stated in positive form degrees, where the strictly pseudoconvex estimates give finite-dimensional harmonic spaces. [/remark] The Neumann operator is only one member of a family of canonical solution operators. For instance, $\bar{\partial}^*N_q$ gives the canonical $L^2$ solution of $\bar{\partial}v=f$ when $f$ is $\bar{\partial}$-closed and orthogonal to obstructions. The key analytic input is not a separate [projection theorem](/theorems/1985) but the boundary estimate that lets normal and tangential derivatives be controlled together. The next formal result records that subelliptic estimate in the strictly pseudoconvex case, where positivity of the Levi form supplies a uniform gain at the boundary. [quotetheorem:3735] [citeproof:3735] The half-derivative gain is the local mechanism behind the later smoothness assertions. Its hypotheses are essential: strict pseudoconvexity excludes the flat complex-tangential directions that can destroy subelliptic control, and the positive-degree condition avoids the infinite-dimensional holomorphic kernel in degree zero. Once this estimate is available at every boundary point, the commutator terms created by differentiating the boundary condition can be absorbed in higher Sobolev estimates. [example: Canonical Solution on the Ball] Let $\Omega=B(0,1)\subset\mathbb C^n$, let $f\in C^\infty_{0,1}(\overline{\Omega})$ satisfy $\bar{\partial}f=0$, and assume $f\perp\ker\Box_1$. Set \begin{align*} u=N_1f,\qquad v=\bar{\partial}^*u. \end{align*} The unit ball is strictly pseudoconvex by the preceding Levi-form computation, so the subelliptic estimate above, iterated through the standard Kohn regularity argument, gives \begin{align*} u=N_1f\in C^\infty_{0,1}(\overline{\Omega}). \end{align*} Write \begin{align*} u=\sum_{j=1}^n u_j\,d\bar z_j. \end{align*} For a smooth $(0,1)$-form, the interior expression for the formal adjoint is \begin{align*} \bar{\partial}^*u=-\sum_{j=1}^n \frac{\partial u_j}{\partial z_j}. \end{align*} Since each $u_j$ is smooth on $\overline{\Omega}$, each derivative $\partial u_j/\partial z_j$ is smooth on $\overline{\Omega}$. Hence \begin{align*} v=\bar{\partial}^*u=-\sum_{j=1}^n \frac{\partial u_j}{\partial z_j}\in C^\infty(\overline{\Omega}). \end{align*} Because $N_1$ is the inverse of $\Box_1$ on $(\ker\Box_1)^\perp$ and $f\perp\ker\Box_1$, we have \begin{align*} \Box_1u=f. \end{align*} By the definition of the complex Laplacian in degree $1$, \begin{align*} f=\Box_1u=\bar{\partial}\bar{\partial}^*u+\bar{\partial}^*\bar{\partial}u. \end{align*} Apply $\bar{\partial}$ to both sides. Linearity gives \begin{align*} \bar{\partial}f=\bar{\partial}\bar{\partial}\bar{\partial}^*u+\bar{\partial}\bar{\partial}^*\bar{\partial}u. \end{align*} The left side is $0$ because $f$ is $\bar{\partial}$-closed. The first term on the right is $0$ because $\bar{\partial}^2=0$. Therefore \begin{align*} 0=\bar{\partial}\bar{\partial}^*\bar{\partial}u. \end{align*} Now pair this identity with $\bar{\partial}u$. Since $u\in\operatorname{Dom}(\Box_1)$, the defining domain condition includes $\bar{\partial}u\in\operatorname{Dom}(\bar{\partial}^*)$, so the Hilbert-space adjoint relation applies: \begin{align*} 0=(\bar{\partial}\bar{\partial}^*\bar{\partial}u,\bar{\partial}u)_{L^2}. \end{align*} Using $(\bar{\partial}A,B)_{L^2}=(A,\bar{\partial}^*B)_{L^2}$ with $A=\bar{\partial}^*\bar{\partial}u$ and $B=\bar{\partial}u$, we get \begin{align*} (\bar{\partial}\bar{\partial}^*\bar{\partial}u,\bar{\partial}u)_{L^2}=(\bar{\partial}^*\bar{\partial}u,\bar{\partial}^*\bar{\partial}u)_{L^2}. \end{align*} Thus \begin{align*} 0=\|\bar{\partial}^*\bar{\partial}u\|_{L^2}^2. \end{align*} A vector in a Hilbert space has norm zero exactly when it is the zero vector, so \begin{align*} \bar{\partial}^*\bar{\partial}u=0. \end{align*} Substituting this into the equation for $f$ gives \begin{align*} f=\bar{\partial}\bar{\partial}^*u+\bar{\partial}^*\bar{\partial}u. \end{align*} Using $\bar{\partial}^*\bar{\partial}u=0$, this becomes \begin{align*} f=\bar{\partial}\bar{\partial}^*u+0. \end{align*} Since $v=\bar{\partial}^*u$, we conclude \begin{align*} f=\bar{\partial}(\bar{\partial}^*u)=\bar{\partial}v. \end{align*} Thus $v=\bar{\partial}^*N_1f$ is a smooth solution of $\bar{\partial}v=f$ on $\overline{\Omega}$. It remains to verify the canonical $L^2$ condition. Let $h\in\ker\bar{\partial}$, so $h\in\operatorname{Dom}(\bar{\partial})$ and \begin{align*} \bar{\partial}h=0. \end{align*} Since $u\in\operatorname{Dom}(\bar{\partial}^*)$, the adjoint relation gives \begin{align*} (v,h)_{L^2}=(\bar{\partial}^*u,h)_{L^2}=(u,\bar{\partial}h)_{L^2}. \end{align*} Using $\bar{\partial}h=0$, \begin{align*} (u,\bar{\partial}h)_{L^2}=(u,0)_{L^2}=0. \end{align*} Therefore \begin{align*} (v,h)_{L^2}=0 \end{align*} for every holomorphic $L^2$ function $h$. Hence $v$ is orthogonal to $\ker\bar{\partial}$, so $v=\bar{\partial}^*N_1f$ is the canonical $L^2$ solution of $\bar{\partial}v=f$. [/example] ## Finite Type, Multitype, and Loss of Regularity What remains of the theory when strict pseudoconvexity degenerates? If the Levi form has null directions, commutators of complex tangential fields may still recover the missing normal direction after more steps. Finite type measures how many commutators or how high an order of holomorphic contact is needed, and the resulting subelliptic estimate has an order smaller than $1/2$. [definition: DAngelo Type] Let $M\subset\mathbb C^n$ be a smooth real hypersurface and let $p\in M$. The D'Angelo type of $M$ at $p$ is \begin{align*} \Delta(M,p)=\sup_\gamma \frac{\operatorname{ord}_0(\rho\circ\gamma)}{\operatorname{ord}_0(\gamma)}, \end{align*} where $\rho$ is a smooth defining function for $M$, and the supremum is taken over all nonconstant holomorphic curves $\gamma:(\mathbb C,0)\to(\mathbb C^n,p)$. [/definition] This definition measures the maximum order of contact between the boundary and a holomorphic curve. Strict pseudoconvexity corresponds to type $2$, while higher finite type allows flatter complex tangential directions but still rules out infinite-order complex contact. To capture the different rates of flatness in several tangential directions, we need the multitype definition next. [definition: Catlin Multitype] For a smooth pseudoconvex hypersurface $M\subset\mathbb C^n$ at $p\in M$, Catlin's multitype is an ordered $n$-tuple \begin{align*} \mathcal M(p)=(m_1,\dots,m_n) \end{align*} of weights associated to the weighted orders of defining functions in adapted holomorphic coordinates. [/definition] Catlin's multitype refines scalar finite type by recording different orders of degeneracy in different complex tangential directions. It is designed to match the anisotropic geometry seen by the $\bar{\partial}$-Neumann problem and prepares the finite-type estimate below. [quotetheorem:9222] Catlin's estimate is local and quantitative: finite type prevents infinite-order tangency, but the amount of subelliptic gain depends on the multitype and can be much smaller than in the strictly pseudoconvex case. The theorem does not by itself give global regularity for every pseudoconvex domain, nor does it erase the difference between weak and strict boundary points. Its role here is to show how finite-order geometry is converted into Sobolev gain, which is exactly what the model hypersurface below makes visible. [example: Model Polynomial Hypersurfaces] [claim]For the model hypersurface \begin{align*}M_m=\{(z,w)\in\mathbb C^2:\operatorname{Re}(w)+|z|^{2m}=0\},\end{align*} the D'Angelo type of $M_m$ at $0$ is $2m$.[/claim] [proof]Use the defining function \begin{align*}\rho(z,w)=\operatorname{Re}(w)+|z|^{2m}=\frac12(w+\overline w)+(z\overline z)^m.\end{align*} The curve $\gamma_0(t)=(t,0)$ is holomorphic, nonconstant, and satisfies $\gamma_0(0)=0$. Its order is \begin{align*}\operatorname{ord}_0\gamma_0=\min(\operatorname{ord}_0 t,\operatorname{ord}_0 0)=\min(1,\infty)=1.\end{align*} Along this curve, \begin{align*}\rho(\gamma_0(t))=\operatorname{Re}(0)+|t|^{2m}=0+(t\overline t)^m=t^m\overline t^{\,m}.\end{align*} The monomial $t^m\overline t^{\,m}$ has total degree $m+m=2m$, so \begin{align*}\operatorname{ord}_0(\rho\circ\gamma_0)=2m.\end{align*} Therefore \begin{align*}\Delta(M_m,0)\ge \frac{\operatorname{ord}_0(\rho\circ\gamma_0)}{\operatorname{ord}_0\gamma_0}=\frac{2m}{1}=2m.\end{align*} For the reverse inequality, let $\gamma(t)=(z(t),w(t))$ be any nonconstant holomorphic curve with $\gamma(0)=0$. Set \begin{align*}a=\operatorname{ord}_0 z,\qquad b=\operatorname{ord}_0 w,\qquad r=\operatorname{ord}_0\gamma=\min(a,b),\end{align*} where $\operatorname{ord}_0 0=\infty$. First assume $a<\infty$ and $b<\infty$. Then \begin{align*}z(t)=t^a h(t),\qquad w(t)=t^b g(t),\end{align*} with $h$ and $g$ holomorphic and $h(0)\ne 0$, $g(0)\ne 0$. Since $g(t)=g(0)+t g_1(t)$ for a holomorphic function $g_1$, \begin{align*}w(t)=g(0)t^b+t^{b+1}g_1(t).\end{align*} Taking real parts gives \begin{align*}\operatorname{Re}(w(t))=\frac12g(0)t^b+\frac12\overline{g(0)}\,\overline t^{\,b}+\frac12t^{b+1}g_1(t)+\frac12\overline t^{\,b+1}\overline{g_1(t)}.\end{align*} Thus the lowest nonzero terms of $\operatorname{Re}(w(t))$ have total degree $b$, with bidegrees $(b,0)$ and $(0,b)$. For the other term, \begin{align*}|z(t)|^{2m}=(z(t)\overline{z(t)})^m=(t^a h(t)\overline t^{\,a}\overline{h(t)})^m=t^{ma}\overline t^{\,ma}(h(t)\overline{h(t)})^m.\end{align*} Because $h(0)\ne 0$, \begin{align*}(h(t)\overline{h(t)})^m=|h(0)|^{2m}+\text{terms of positive total degree}.\end{align*} Multiplying by $t^{ma}\overline t^{\,ma}$ gives \begin{align*}|z(t)|^{2m}=|h(0)|^{2m}t^{ma}\overline t^{\,ma}+\text{terms of total degree greater than }2ma.\end{align*} The lowest term from $|z(t)|^{2m}$ has bidegree $(ma,ma)$. Since $a\ge 1$, this bidegree is different from both $(b,0)$ and $(0,b)$, so the degree-$b$ terms from $\operatorname{Re}(w(t))$ cannot cancel the degree-$2ma$ term from $|z(t)|^{2m}$. Therefore \begin{align*}\operatorname{ord}_0(\rho\circ\gamma)=\min(b,2ma).\end{align*} If $a\le b$, then $r=a$, and \begin{align*}\frac{\operatorname{ord}_0(\rho\circ\gamma)}{\operatorname{ord}_0\gamma}=\frac{\min(b,2ma)}{a}\le \frac{2ma}{a}=2m.\end{align*} If $b<a$, then $r=b$, and because $\min(b,2ma)\le b$ while $\min(b,2ma)\ge b$ by definition of the minimum with first entry $b$, we have \begin{align*}\min(b,2ma)=b.\end{align*} Hence \begin{align*}\frac{\operatorname{ord}_0(\rho\circ\gamma)}{\operatorname{ord}_0\gamma}=\frac{b}{b}=1\le 2m.\end{align*} Now suppose one component is identically zero. If $w\equiv 0$, then $b=\infty$, $r=a$, and \begin{align*}\rho(\gamma(t))=|z(t)|^{2m}=|h(0)|^{2m}t^{ma}\overline t^{\,ma}+\text{terms of total degree greater than }2ma.\end{align*} Thus \begin{align*}\frac{\operatorname{ord}_0(\rho\circ\gamma)}{\operatorname{ord}_0\gamma}=\frac{2ma}{a}=2m.\end{align*} If $z\equiv 0$, then $a=\infty$, $r=b$, and \begin{align*}\rho(\gamma(t))=\operatorname{Re}(w(t))=\frac12g(0)t^b+\frac12\overline{g(0)}\,\overline t^{\,b}+\text{terms of total degree greater than }b.\end{align*} Thus \begin{align*}\frac{\operatorname{ord}_0(\rho\circ\gamma)}{\operatorname{ord}_0\gamma}=\frac{b}{b}=1\le 2m.\end{align*} Every nonconstant holomorphic curve through $0$ therefore has contact ratio at most $2m$, and $\gamma_0(t)=(t,0)$ has contact ratio exactly $2m$. By the definition of D'Angelo type, \begin{align*}\Delta(M_m,0)=2m.\end{align*} [/proof] The computation shows that the exponent $2m$ is exactly the order of flatness in the complex tangential $z$-direction. [/example] Finite type is a local condition and gives local subelliptic control. Global regularity asks for estimates across the whole boundary and is more delicate: local finite-type estimates do not automatically assemble into exact Sobolev estimates for the global Neumann operator in every pseudoconvex domain. [definition: Worm Domain] A worm domain is a smoothly bounded pseudoconvex domain in $\mathbb C^2$ whose boundary contains an annular region along which the complex normal direction twists as one moves around the annulus. [/definition] The original Diederich-Fornaess worm domains were constructed to show that pseudoconvexity can behave badly under exhaustion by strictly pseudoconvex domains. In the $\bar{\partial}$-Neumann problem, worm geometry produces a mismatch between local regularity and global boundary regularity. [quotetheorem:9223] [example: Non-Regularity Mechanism in a Worm Domain] [claim]On a worm domain with a Levi-flat annular portion whose complex normal twists around the annulus, pseudoconvexity alone does not force the $\bar{\partial}$-Neumann operator or the Bergman projection to satisfy global Sobolev estimates.[/claim] [proof]Near the annulus, let $\theta$ be the angular variable and write one boundary component in [Fourier series](/page/Fourier%20Series): \begin{align*} g(\theta)=\sum_{k\in\mathbb Z}g_k e^{ik\theta}. \end{align*} For Sobolev order $s$, the tangential Sobolev norm in this variable is \begin{align*} \|g\|_{H^s_\theta}^2=\sum_{k\in\mathbb Z}(1+k^2)^s |g_k|^2. \end{align*} Thus a global Sobolev estimate of order $s$ for the Bergman projection $P$ would imply that there is a constant $C_s>0$ such that \begin{align*} \sum_{k\in\mathbb Z}(1+k^2)^s |(Pf)_k|^2\le C_s\sum_{k\in\mathbb Z}(1+k^2)^s |f_k|^2 \end{align*} for every smooth datum $f$. The twisting of the complex normal changes the boundary condition seen by the $k$th Fourier mode. In the Kiselman-Christ model analysis of worm geometry, this produces smooth data whose Fourier coefficients decay faster than every power, while the projected coefficients decay only at a fixed polynomial rate along an infinite subsequence. Concretely, suppose that for every $N$ there is a constant $C_N>0$ such that \begin{align*} |f_k|\le C_N(1+|k|)^{-N} \end{align*} for all $k\in\mathbb Z$, and that there are constants $c>0$, $\alpha<\infty$, and an infinite subsequence $S\subset\mathbb Z$ such that \begin{align*} |(Pf)_k|\ge c(1+|k|)^{-\alpha} \end{align*} for every $k\in S$. First fix a Sobolev order $s$ and check that $f$ belongs to $H^s_\theta$. Choose $N>s+1$. Since $1+k^2\le (1+|k|)^2$, raising both sides to the power $s$ gives \begin{align*} (1+k^2)^s\le (1+|k|)^{2s}. \end{align*} Using the assumed decay of $f_k$, \begin{align*} (1+k^2)^s|f_k|^2\le (1+|k|)^{2s}C_N^2(1+|k|)^{-2N}. \end{align*} Combining the powers gives \begin{align*} (1+k^2)^s|f_k|^2\le C_N^2(1+|k|)^{2s-2N}. \end{align*} Because $N>s+1$, we have $2s-2N<-2$. The comparison series $\sum_{k\in\mathbb Z}(1+|k|)^{2s-2N}$ converges, so \begin{align*} \sum_{k\in\mathbb Z}(1+k^2)^s|f_k|^2<\infty. \end{align*} Now choose $s>\alpha+\frac12$ and test the projected coefficients. For $|k|\ge 1$, the inequality $1+|k|\le 2|k|$ gives \begin{align*} (1+|k|)^2\le 4k^2. \end{align*} Since $k^2\le 1+k^2$, it follows that \begin{align*} (1+|k|)^2\le 4(1+k^2). \end{align*} Dividing by $4$ gives \begin{align*} 1+k^2\ge \frac14(1+|k|)^2. \end{align*} Raising to the power $s$ gives \begin{align*} (1+k^2)^s\ge 4^{-s}(1+|k|)^{2s}. \end{align*} For $k\in S$ with $|k|\ge 1$, the lower bound for $(Pf)_k$ gives \begin{align*} |(Pf)_k|^2\ge c^2(1+|k|)^{-2\alpha}. \end{align*} Multiplying the last two inequalities, \begin{align*} (1+k^2)^s |(Pf)_k|^2\ge 4^{-s}(1+|k|)^{2s}c^2(1+|k|)^{-2\alpha}. \end{align*} Combining the powers gives \begin{align*} (1+k^2)^s |(Pf)_k|^2\ge c^2 4^{-s}(1+|k|)^{2s-2\alpha}. \end{align*} Since $s>\alpha+\frac12$, we have $2s-2\alpha>1$. The infinite subset $S\subset\mathbb Z$ is unbounded, so along $S$ the lower bound $c^2 4^{-s}(1+|k|)^{2s-2\alpha}$ tends to infinity. Hence the terms $(1+k^2)^s |(Pf)_k|^2$ do not tend to $0$ along $S$. A convergent series must have terms tending to $0$, so \begin{align*} \sum_{k\in\mathbb Z}(1+k^2)^s |(Pf)_k|^2=\infty. \end{align*} Thus $Pf\notin H^s_\theta$, even though $f\in H^s_\theta$ for every $s$. This contradicts the Sobolev estimate for $P$ at the chosen order $s>\alpha+\frac12$. Therefore the Bergman projection cannot satisfy global Sobolev estimates of all orders on such a worm domain. The same obstruction reaches the $\bar{\partial}$-Neumann operator through the identity \begin{align*} P=I-\bar{\partial}^*N_1\bar{\partial} \end{align*} on smooth functions. If $N_1$ preserved Sobolev regularity in all orders, then $\bar{\partial}$ would lower Sobolev order by one, $N_1$ would preserve the resulting regularity, and $\bar{\partial}^*$ would lower Sobolev order by one. Consequently $I-\bar{\partial}^*N_1\bar{\partial}$ would still send smooth functions to smooth functions. Since the Fourier construction gives smooth $f$ with $Pf\notin H^s_\theta$ for some $s$, such global Sobolev estimates for $N_1$ cannot hold. Christ's smooth worm construction shows that the obstruction is not caused by a nonsmooth boundary: the twisting annular geometry remains in the smooth category. Local pseudoconvexity controls the sign of the Levi form, but the global rotation of the complex normal can force high Fourier modes to lose Sobolev decay.[/proof] Thus smooth input may produce a Neumann solution or Bergman projection with boundary singularities, so pseudoconvexity by itself is weaker than global regularity. [/example] The final lesson is that the $\bar{\partial}$-Neumann problem is a boundary-sensitive elliptic theory. Strict pseudoconvexity gives the cleanest outcome: closed range, subelliptic gain, and global smoothness. Finite type keeps a local subelliptic estimate but weakens the exponent. Worm domains show that global regularity can fail when boundary degeneracy and global geometry interact in the wrong way. Chapter 9 uses the same strict pseudoconvex regularity background to construct CR functions through the Szego projector. The $\bar\partial$-Neumann theory shows how regularity depends on both local and global boundary geometry, and that same regularity is the input for constructing CR functions and embeddings. The next chapter uses this analytic foundation to ask when an abstract CR manifold can actually be realized as a hypersurface in complex space. # 9. CR Embeddings and the Boutet de Monvel–Guillemin Theory This chapter turns from local CR analysis to the global question of whether an abstract CR manifold comes from an actual hypersurface in complex Euclidean space. The prerequisites are the Levi form, strictly pseudoconvex boundaries, the tangential Cauchy-Riemann operator $\bar\partial_b$, basic Hilbert-space projections, and the boundary Hardy space. Earlier chapters developed these as intrinsic objects. We now ask when they produce enough CR functions to give coordinates, and how the Szego projector encodes that answer microlocally. ## The CR Embedding Problem The basic problem is not to recognise a hypersurface after it has already been placed in $\mathbb C^N$, but to decide whether an abstract CR manifold has enough CR functions to place it there. Compactness makes this a global question: local CR coordinates may exist in many settings, while a finite collection of global CR functions may fail to separate points or tangent directions. [definition: CR Embedding] Let $M$ be a smooth CR manifold of CR dimension $n-1$. A CR embedding of $M$ into $\mathbb C^N$ is a smooth embedding $F:M\to \mathbb C^N$ such that each component $F_j$ is a CR function on $M$. [/definition] The condition says that the embedding map is built from solutions of the tangential Cauchy-Riemann equations. Thus the embedding problem asks whether the analytic sheaf of global CR functions is rich enough to recover the manifold and its CR structure. [example: Boundary Embedding] Let $\Omega\subset \mathbb C^n$ be a bounded domain with smooth boundary $M=\partial\Omega$, and let $i:M\to\mathbb C^n$ be the inclusion map. Since $M$ is already a smooth embedded hypersurface of $\mathbb C^n$, the map $i$ is a smooth embedding. Its components are the restricted coordinate functions \begin{align*} i_j(p)=z_j(p)\quad\text{for }p\in M,\ 1\le j\le n. \end{align*} To check that $i_j$ is CR, let $\overline L$ be any smooth section of $T^{0,1}M$. Because $T^{0,1}M\subset T^{0,1}\mathbb C^n|_M$, locally \begin{align*} \overline L=\sum_{k=1}^n a_k\,\frac{\partial}{\partial \overline z_k} \end{align*} with smooth coefficients $a_k$ subject to the tangency condition to $M$. Hence \begin{align*} \overline L(i_j)=\sum_{k=1}^n a_k\,\frac{\partial z_j}{\partial \overline z_k}=0, \end{align*} because each holomorphic coordinate $z_j$ is independent of every $\overline z_k$. Thus all components of $i$ are CR functions, so the inclusion $M\hookrightarrow\mathbb C^n$ is a CR embedding. This example shows one direction of the embedding problem: every already embedded smooth boundary supplies global CR coordinate functions, namely $z_1|_M,\dots,z_n|_M$. The reverse direction is subtler. An abstract compact strictly pseudoconvex CR manifold may have local CR functions near each point, but a CR embedding requires a finite global list $f_1,\dots,f_N$ such that the map $p\mapsto(f_1(p),\dots,f_N(p))$ separates distinct points and has injective differential everywhere. In real dimension three, Rossi-type examples show that these global separation conditions can fail even when local CR charts exist. Thus the embedding problem is not merely the local construction of CR functions; it is the global problem of producing enough CR functions to separate points and tangent directions. [/example] The main obstruction is not the formal definition of a CR map, but the production of global CR functions. In analytic terms, this is governed by the closed range and microlocal regularity properties of $\bar\partial_b$ and by the finite-dimensional correction terms appearing in the Szego projector. [definition: Strictly Pseudoconvex CR Manifold] Let $M$ be a smooth CR manifold of real dimension $2n-1$ with CR bundle $T^{1,0}M$. It is strictly pseudoconvex if there exists a contact form $\theta$ annihilating the real contact distribution $H(M)=\operatorname{Re}(T^{1,0}M\oplus T^{0,1}M)$ such that the Levi form \begin{align*} L_{\theta,x}:T^{1,0}_xM\times T^{1,0}_xM\longrightarrow \mathbb C \end{align*} defined by \begin{align*} L_{\theta,x}(Z,W)=-i\,d\theta_x(Z,\overline{W}) \end{align*} is positive definite for every $x\in M$. [/definition] Strict pseudoconvexity is the curvature positivity condition that replaces convexity in this boundary theory. It gives subelliptic estimates for $\bar\partial_b$ in high enough dimension, and those estimates are the analytic input behind the embedding theorem. [remark: Dimension Threshold] For a strictly pseudoconvex CR manifold of real dimension $2n-1$, the condition $2n-1\ge 5$ means $n\ge 3$. The three-dimensional case $n=2$ behaves differently because the relevant Kohn estimates have exceptional low-degree features. [/remark] This threshold is one reason the embedding theorem has a striking form: the natural geometric hypothesis is the same in all dimensions, but the analytic proof separates dimension at the level of the $\bar\partial_b$ complex. The preceding definitions isolate the geometry that an embedded strongly pseudoconvex boundary always has; the remaining question is whether that geometry alone forces the existence of enough global CR functions. The answer is not obtained by constructing an ambient complex space directly, but by using analytic estimates for $\bar\partial_b$ and the Szego projector to manufacture global CR coordinate functions from local ones. This is the point at which strict pseudoconvexity changes from a curvature condition into an embedding mechanism. ## Boutet de Monvel's Embedding Theorem The central question is whether strict pseudoconvexity alone forces the existence of enough global CR functions. Boutet de Monvel's theorem answers this in real dimension at least five and is one of the main bridges between abstract CR geometry and complex analysis on domains. [quotetheorem:9224] The theorem is stated here as a structural result rather than proved in full. It marks the point where strict pseudoconvexity becomes strong enough to guarantee global CR coordinate functions, not merely local CR charts. Each hypothesis in the theorem has a different role. Compactness turns local separation into a finite global list of CR functions and gives a Hilbert-space setting for the Szego projector. Strict pseudoconvexity supplies the positive Levi form and the subelliptic estimates used to correct approximate CR functions. Smoothness is used to run the microlocal construction to all orders and to obtain smooth CR coordinate functions. The dimension condition $2n-1\ge 5$ excludes the low-dimensional case where the relevant $\bar\partial_b$ estimates and global separation mechanism can fail. The theorem gives no canonical minimal value of $N$: the embedding dimension depends on the analytic construction, and the conclusion is qualitative. [explanation: Meaning of the Embedding Theorem] The conclusion is global: one finite CR map must separate points and tangent directions on all of $M$. This is stronger than having local CR coordinates near each point. Compactness, strict pseudoconvexity, smoothness, and the dimension bound together rule out the main obstructions to turning local CR information into a global embedding. [/explanation] The excluded three-dimensional case is not a minor endpoint. It is the first dimension in which strictly pseudoconvex CR geometry is already rich but the high-dimensional embedding mechanism no longer automatically supplies enough global CR functions. The next example isolates that dimensional failure before the course turns to separate embeddability and fillability phenomena. [example: Dimension Three and the Open Embedding Phenomenon] In real dimension three we have $2n-1=3$, so \begin{align*} 2n=4 \end{align*} and therefore $n=2$. Hence the dimension hypothesis $2n-1\ge 5$ in Boutet de Monvel's embedding theorem becomes \begin{align*} 3\ge 5, \end{align*} which is false. Thus a compact strictly pseudoconvex CR manifold of real dimension three is not covered by the theorem, even though it satisfies the same formal definition of strict pseudoconvexity. The missing point is analytic rather than definitional: strict pseudoconvexity still means that the Levi form is positive definite on the one-dimensional complex bundle $T^{1,0}M$, but the high-dimensional $\bar\partial_b$ correction argument used to manufacture many global CR functions is no longer available in the same form. Rossi-type examples show that there are compact strictly pseudoconvex three-dimensional CR manifolds for which the global CR functions do not separate points and tangent directions sufficiently to give an embedding into any $\mathbb C^N$. Therefore, in real dimension three, embeddability is an additional global condition, not a consequence of strict pseudoconvexity alone. [/example] This example is important for interpreting the theorem. It says that strict pseudoconvexity is a sufficient global embedding hypothesis in higher dimension, but not a dimension-free replacement for being the boundary of a complex domain. ## The Szego Projector and Hardy Space To make the embedding theorem analytic, we need a projection that extracts the CR part of an $L^2$ boundary function. The Szego projector is the boundary analogue of the Bergman projector in the interior, and its kernel is the object whose singularities record the CR geometry. [definition: Hardy Space of a CR Manifold] Let $M$ be a compact CR manifold with a smooth positive density $d\mu$. The Hardy space $H^2(M)$ is the closed subspace of $L^2(M,d\mu)$ consisting of $L^2$ limits of smooth CR functions on $M$. [/definition] The closure is essential because the natural projection acts on $L^2(M,d\mu)$, while individual CR functions may be produced first in Sobolev or distributional form and then upgraded by regularity. Having identified the closed target space, the next question is how to extract the CR component of an arbitrary boundary datum in a canonical way. [definition: Szego Projector] Let $M$ be a compact CR manifold with Hardy space $H^2(M)\subset L^2(M,d\mu)$. The Szego projector is the [orthogonal projection](/theorems/437) \begin{align*} \Pi:L^2(M,d\mu)\longrightarrow H^2(M). \end{align*} [/definition] Since $\Pi$ is an orthogonal projection, it satisfies $\Pi^2=\Pi$ and $\Pi^*=\Pi$. These algebraic identities are useful, but microlocal analysis needs a local object on $M\times M$; this leads from the operator to the distribution kernel that represents it. [definition: Szego Kernel] Let $M$ be a compact CR manifold with Szego projector $\Pi$. The Szego kernel is the distribution kernel $S\in \mathcal D'(M\times M)$ representing $\Pi$. [/definition] The kernel is smooth away from the diagonal in strictly pseudoconvex settings. Its singularity along the diagonal is the boundary counterpart of the singularity of the Cauchy kernel in one complex variable, so the first task is to compute the model case where the CR symmetry is maximal. [example: Szego Kernel of the Unit Sphere] Let $d\sigma$ be normalised surface measure on $S^{2n-1}$, and write multi-indices as $\alpha=(\alpha_1,\dots,\alpha_n)$, $|\alpha|=\alpha_1+\cdots+\alpha_n$, $z^\alpha=z_1^{\alpha_1}\cdots z_n^{\alpha_n}$. The Hardy space is the $L^2(S^{2n-1},d\sigma)$-closure of the boundary values of holomorphic polynomials, so it is generated by the monomials $z^\alpha$. For $|\alpha|=q$, the standard sphere moment formula gives \begin{align*} \int_{S^{2n-1}} |z^\alpha|^2\,d\sigma(z)=\frac{\alpha!(n-1)!}{(q+n-1)!}. \end{align*} Thus the normalised monomials are \begin{align*} e_\alpha(z)=\left(\frac{(q+n-1)!}{\alpha!(n-1)!}\right)^{1/2}z^\alpha,\qquad |\alpha|=q. \end{align*} The Szego kernel is the reproducing kernel obtained by summing $e_\alpha(z)\overline{e_\alpha(w)}$ over all holomorphic monomials. The degree-$q$ contribution is \begin{align*} \sum_{|\alpha|=q} e_\alpha(z)\overline{e_\alpha(w)}=\sum_{|\alpha|=q}\frac{(q+n-1)!}{\alpha!(n-1)!}z^\alpha\overline{w}^{\alpha}. \end{align*} Since \begin{align*} (z\cdot\overline w)^q=\left(\sum_{j=1}^n z_j\overline{w_j}\right)^q=\sum_{|\alpha|=q}\frac{q!}{\alpha!}z^\alpha\overline{w}^{\alpha}, \end{align*} the degree-$q$ contribution becomes \begin{align*} \binom{q+n-1}{q}(z\cdot\overline w)^q. \end{align*} Summing over $q\ge 0$ and using the binomial series \begin{align*} \sum_{q=0}^{\infty}\binom{q+n-1}{q}t^q=(1-t)^{-n} \end{align*} gives \begin{align*} S(z,w)=\frac{1}{(1-z\cdot\overline w)^n}. \end{align*} The singularity occurs when $z\cdot\overline w=1$, which on the unit sphere forces $z=w$; this is the model diagonal singularity that the general strictly pseudoconvex Szego kernel reproduces microlocally. [/example] The sphere formula is the local normal form for the general strictly pseudoconvex case. General kernels do not have this exact rational expression, so the next question is which part of the sphere singularity survives after replacing the sphere by an arbitrary strictly pseudoconvex boundary. [quotetheorem:9225] [citeproof:9225] This theorem is the analytic heart of the microlocal description. The phase identifies the positive symplectic cone determined by the CR structure, while the symbol carries curvature and lower-order invariants of the boundary. Strict pseudoconvexity is needed because the positive Levi form gives the subelliptic estimates and the positive complex phase; for weakly pseudoconvex boundaries, the characteristic set can have degeneracies and the same homogeneous expansion may acquire different model behaviour or lose uniform symbolic control. Smoothness of the boundary is also part of the statement, since the construction differentiates the defining function repeatedly to solve the eikonal and transport equations; finite regularity gives only a finite-order version of the expansion. The theorem is microlocal near the diagonal and does not assert a global closed formula for $S(x,y)$ or determine the lower-order curvature coefficients without further calculation. It feeds directly into the Hermite calculus below, where the expansion is reinterpreted as saying that the Szego projector is a Fourier integral operator attached to the contact cone. The proof sketch explains why the kernel expansion is not a decorative formula. It is precisely the mechanism that turns geometric positivity into an operator with enough regularity and enough singularity to generate CR functions. ## Fourier Integral Operators and Hermite Operators The remaining problem is to describe the Szego projector in a calculus stable under composition and adjoints. Ordinary pseudodifferential operators are not adapted to the characteristic cone of $\bar\partial_b$, so Boutet de Monvel and Guillemin introduced a Hermite Fourier integral calculus fitted to contact manifolds. [definition: Positive Contact Cone] Let $(M,\theta)$ be a cooriented contact manifold. The positive contact cone is \begin{align*} \Sigma^+=\{(x,r\theta_x)\in T^*M: x\in M,\ r>0\}. \end{align*} [/definition] This cone is the microlocal support of the Szego projector for a strictly pseudoconvex CR manifold. It records the direction in cotangent space where the tangential Cauchy-Riemann operator fails to be elliptic. Because the projector must be composed with itself and with auxiliary parametrices, the cone motivates a class of Fourier integral operators whose symbols and canonical relations are built around this contact geometry. [definition: Hermite Fourier Integral Operator] Let $M$ be a compact contact manifold with positive contact cone $\Sigma^+$. A Hermite Fourier integral operator of order $m$ associated to $\Sigma^+$ is a continuous linear operator \begin{align*} A:C^\infty(M)&\longrightarrow \mathcal D'(M) \end{align*} whose distribution kernel $K_A\in\mathcal D'(M\times M)$ is locally represented by an oscillatory integral with complex phase parametrising the canonical relation determined by $\Sigma^+$ and with amplitude in the Hermite symbol class of order $m$. [/definition] The definition packages the geometric data needed for composition. Operators of order $m$ in this calculus also act continuously between the corresponding Sobolev spaces after the order shift determined by the Hermite calculus. The point is that the Szego projector should not merely resemble such an operator locally; the next theorem asserts that it is an object of this calculus, with the projection identity visible at symbol level. [quotetheorem:9226] [citeproof:9226] This realisation turns the projector into a geometric quantisation of the contact structure. The idempotent symbol condition is the infinitesimal version of the projection identity $\Pi^2=\Pi$. Compactness ensures that the $L^2$ Hardy space is a closed Hilbert-space target and that the local parametrices can be assembled into a global projection modulo smoothing terms; on noncompact manifolds, the same local formula can hold while global closed range or bounded projection properties fail. Strict pseudoconvexity is needed to single out the positive contact cone and to give the subelliptic estimates behind the smoothing remainder; a Levi-flat or weakly pseudoconvex boundary has a different characteristic geometry and need not support the same Hermite idempotent. The coorientation/contact-cone hypothesis is also structural: without a chosen positive cone, the symbol has no distinguished half of the characteristic variety on which to live. The theorem does not classify all Hermite idempotents or determine an embedding by itself, but it supplies the microlocal object used next to compare boundary and interior kernel asymptotics. Once the projector belongs to this calculus, the asymptotic expansion of the kernel becomes functorial under CR equivalence. In particular, contact transformations preserving the CR structure transport the positive cone and hence preserve the form of the leading singularity. [example: Model Projector on the Heisenberg Boundary] Identify $\partial\mathcal S$ of the Siegel upper half-space \begin{align*} \mathcal S=\{(z,\tau)\in\mathbb C^{n-1}\times\mathbb C:\operatorname{Im}\tau>|z|^2\} \end{align*} with $\mathbb H^{n-1}=\mathbb C^{n-1}\times\mathbb R$ by \begin{align*} (z,t)\longmapsto (z,t+i|z|^2). \end{align*} On the boundary, the standard contact form is \begin{align*} \theta=dt+i\sum_{j=1}^{n-1}(\overline z_j\,dz_j-z_j\,d\overline z_j), \end{align*} so the positive contact cone is \begin{align*} \Sigma^+=\{((z,t),r\theta_{(z,t)}):r>0\}. \end{align*} The model Szego kernel can be written, up to the normalising constant determined by the chosen boundary measure, as \begin{align*} S_{\mathbb H}((z,t),(w,s))=c_n\left(\frac{|\!z|\!^2+|\!w|\!^2}{2}-z\cdot\overline w-\frac{i}{2}(t-s)\right)^{-n}. \end{align*} The denominator separates into its real and imaginary parts because \begin{align*} \frac{|z|^2+|w|^2}{2}-z\cdot\overline w-\frac{i}{2}(t-s)=\frac{|z-w|^2}{2}-i\left(\operatorname{Im}(z\cdot\overline w)+\frac{t-s}{2}\right). \end{align*} Thus the kernel is singular precisely at the Heisenberg diagonal, where $z=w$ and $t=s$. Under the Heisenberg dilation $\delta_\lambda(z,t)=(\lambda z,\lambda^2 t)$, the denominator scales by $\lambda^2$: \begin{align*} \frac{|\lambda z|^2+|\lambda w|^2}{2}-(\lambda z)\cdot\overline{\lambda w}-\frac{i}{2}(\lambda^2t-\lambda^2s)=\lambda^2\left(\frac{|z|^2+|w|^2}{2}-z\cdot\overline w-\frac{i}{2}(t-s)\right). \end{align*} Raising to the power $-n$ gives \begin{align*} S_{\mathbb H}(\delta_\lambda(z,t),\delta_\lambda(w,s))=\lambda^{-2n}S_{\mathbb H}((z,t),(w,s)). \end{align*} This homogeneous scaling is the same leading scaling seen by blowing up the sphere kernel near a boundary point, so Heisenberg coordinates isolate the universal principal singularity of the Szego projector near any strictly pseudoconvex boundary point. [/example] The Heisenberg model is the local microscope for the general theory. Curvature and torsion enter through lower-order perturbations of this model, while the principal singularity is fixed by the contact geometry. ## Szego and Bergman Kernels The boundary projector should be compared with the Bergman projector inside the domain. The question is how the holomorphic $L^2$ theory in $\Omega$ sees the same boundary geometry measured by the Szego kernel on $\partial\Omega$. [definition: Bergman Kernel] Let $\Omega\subset \mathbb C^n$ be a bounded domain. The Bergman space $A^2(\Omega)$ is the closed subspace of $L^2(\Omega,d\mathcal L^{2n})$ consisting of holomorphic functions, and the Bergman kernel $K_\Omega(z,w)$ is the distribution kernel of the orthogonal projection $L^2(\Omega)\to A^2(\Omega)$. [/definition] The Bergman kernel is an interior object, but near a strictly pseudoconvex boundary its singularity is controlled by the same Levi geometry that controls the Szego kernel. Since the two kernels project in different Hilbert spaces, the comparison must pass through boundary values, holomorphic extension, and the normal direction. This motivates a theorem relating the leading singularities rather than identifying the kernels themselves. [quotetheorem:9227] [citeproof:9227] This relation is often used as a transfer principle. Boundary microlocal information about $\Pi$ can be converted into interior asymptotics for the Bergman projection, and conversely interior kernel expansions reflect CR invariants of the boundary. Strict pseudoconvexity is needed because the Poisson extension and Szego projector share the same positive contact canonical relation only in the nondegenerate Levi setting; for a polydisc or a weakly pseudoconvex finite-type domain, corners or degenerate Levi directions produce different boundary singularities. Smooth boundary is also essential for the full asymptotic statement, since the normal projection, defining function calculus, and repeated symbol transport require differentiability to all orders. The statement is local near the boundary diagonal and is not a formula for the Bergman kernel deep in the interior or at pairs of points approaching different boundary components. Its role in the chapter is to connect the boundary Hermite calculus with the interior kernel invariants that appear in Fefferman-type expansions. The proof reduces the comparison to a concrete model: the unit ball. There the two kernels are explicit, and the difference between the boundary and interior singularity orders can be read directly from their formulas. [example: Unit Ball Comparison] For the unit ball $B\subset\mathbb C^n$, write \begin{align*} z\cdot\overline w=\sum_{j=1}^n z_j\overline{w_j}. \end{align*} With the usual normalisation of volume measure, the Bergman kernel has the form \begin{align*} K_B(z,w)=\frac{c_n}{(1-z\cdot\overline{w})^{n+1}}, \end{align*} where $c_n$ is the normalising constant determined by the chosen measure. On the boundary sphere $\partial B=S^{2n-1}$, with the normalised surface measure used for the Hardy space, the Szego kernel is \begin{align*} S_{\partial B}(\zeta,\eta)=\frac{1}{(1-\zeta\cdot\overline{\eta})^n}. \end{align*} The two formulas have the same boundary defining factor $1-z\cdot\overline w$, but their exponents differ by one: \begin{align*} (n+1)-n=1. \end{align*} Equivalently, \begin{align*} \frac{K_B(z,w)}{S_{\partial B}(z,w)}=c_n(1-z\cdot\overline w)^{-1} \end{align*} whenever the boundary expression is compared with the same algebraic factor. Thus the Bergman kernel has exactly one additional power of the boundary singular factor. This is the model calculation behind the statement that, near a strictly pseudoconvex boundary, the Bergman singularity is one positive normal order stronger than the Szego singularity. [/example] The chapter's main lesson is that CR embeddability, the Szego projector, and microlocal Fourier integral theory are not separate subjects. The embedding theorem depends on producing CR functions, the Szego projector produces the CR part of boundary data, and the Boutet de Monvel-Guillemin calculus describes exactly how that projection behaves at the characteristic directions of the CR structure. After embedding and the Szego projection reveal how CR functions are assembled globally, the remaining issue is how their singularities and microlocal structure behave. The FBI transform and propagation theory track that information through CR curves, bicharacteristics, and holomorphic wedges. # 10. Propagation of CR Singularities and the FBI Transform Earlier chapters treated the tangential Cauchy-Riemann equations as differential equations intrinsic to a CR manifold. This chapter asks how far local CR information travels: along CR curves on the manifold, along bicharacteristics in phase space, and into ambient holomorphic wedges. The answer depends on the distinction between smooth and real-analytic regularity, and the central tools are the [Baouendi-Treves approximation theorem](/theorems/9229) and the FBI transform. ## Real-analytic CR Functions and Propagation Along CR Curves The first problem is a continuation problem: if a CR function is known near one point, which geometric paths force the same information to hold elsewhere? The tangential Cauchy-Riemann equations differentiate only in complex tangent directions, so the relevant paths are not arbitrary paths in the manifold but paths generated by the CR distribution and its commutators. [definition: CR Curve and CR Orbit] Let $M$ be a $C^\infty$ CR manifold, and let $H(M)\subset TM$ be its real holomorphic tangent bundle. A CR curve in $M$ is a piecewise $C^1$ map $\gamma:[0,1]\to M$ such that \begin{align*} \dot\gamma(t)\in H_{\gamma(t)}(M) \end{align*} for every $t$ at which $\gamma$ is differentiable. The CR orbit $\mathcal O(p)$ of $p\in M$ is the set of all points of $M$ reachable from $p$ by CR curves. [/definition] The orbit is the geometric region on which the CR equations can communicate information by transport. If brackets of CR vector fields generate all tangent directions, then the orbit is open; if the manifold is Levi-flat, the orbits may be lower-dimensional complex leaves. This leads to the following real-analytic unique continuation theorem along such orbits. [quotetheorem:9228] [citeproof:9228] This is the real-analytic version of propagation along CR geometry. The analytic hypothesis matters: smooth CR functions on nonminimal manifolds may carry arbitrary transverse smooth data, while real-analytic functions cannot hide such data from analytic continuation along the orbit. The connectedness restriction in the conclusion is necessary, not cosmetic: a CR orbit can have several connected components, and a real-analytic CR function may vanish on one component while taking nonzero analytic values on another. The proof produces a vanishing chain along a specific CR curve, and CR curves stay inside one [connected component](/page/Connected%20Component) of the orbit by construction. The result also fails in the smooth category in a strong sense: on a smooth nonminimal CR manifold one can construct $C^\infty$ CR functions that vanish on an open set yet do not vanish on a nearby part of the orbit, because the identity theorem for $C^\infty$ functions is unavailable along the integral leaves used in the propagation step. Real-analyticity of $M$ and of $f$ both enter through that identity theorem; weakening either hypothesis to smoothness destroys the argument. The same geometric idea has a microlocal form. Instead of asking where a function vanishes, we ask which cotangent directions can carry singularities of a CR solution. [definition: Characteristic Set of the Tangential Cauchy-Riemann System] Let $M$ be a $C^\infty$ CR manifold. The characteristic set of $\bar\partial_b$ is \begin{align*} \operatorname{Char}(\bar\partial_b) =\{(p,\xi)\in T^*M\setminus 0: \xi(v)=0\text{ for every }v\in H_p(M)\}. \end{align*} [/definition] A covector outside this set sees at least one CR direction. In such a direction the tangential Cauchy-Riemann system is elliptic, so CR solutions have no microlocal singularity there. To state propagation inside the characteristic set, we next need bicharacteristics. [definition: Bicharacteristic] Let $P:\mathcal D'(M)\to\mathcal D'(M)$ be a first-order scalar differential operator with real principal symbol $p\in C^\infty(T^*M\setminus 0;\mathbb R)$, and let $\Sigma\subset T^*M\setminus 0$ be a conic hypersurface on which $p$ vanishes. A bicharacteristic of $P$ in $\Sigma$ is an integral curve $\Gamma:I\to \Sigma$ of the Hamilton vector field $H_p$. [/definition] For the Lewy operator and for scalar reductions of $\bar\partial_b$, these curves are the phase-space paths along which regularity is transported. The Hamilton field $H_p$ is tangent to $\Sigma$ wherever $p$ vanishes, and its integral curves project to base curves in $M$ that are quite different from CR curves: the projection records where the operator is degenerate, not where the equations differentiate. To verify in practice that a covector $(p,\xi)$ lies on a real-principal-type bicharacteristic, compute the principal symbol in local coordinates, check that its real differential is nonzero on $\Sigma$ at $(p,\xi)$, and integrate the Hamilton field with $(p,\xi)$ as initial condition; the segment one obtains over a compact time interval is the curve along which propagation will act. This prepares the propagation theorem below. Two cautions apply to this construction. First, the analogous statement at a point where the principal symbol vanishes to second order is not a propagation along a curve but along a higher-dimensional submanifold, and the Hamilton-flow picture has to be replaced by the symplectic geometry of the doubly characteristic set. Second, even for real-principal-type operators, the bicharacteristic is a curve in phase space; its image in $M$ may self-intersect or fail to embed, and propagation conclusions are phrased on the curve in $T^*M$ rather than on its base projection. These cautions motivate the following regularity propagation theorem in the real-principal-type case. [quotetheorem:8194] [citeproof:8194] For CR functions, $Pu=0$ for the relevant tangential operator, so the theorem says that singularities cannot stop at an interior point of a bicharacteristic. They either fail to appear on the segment or propagate along it. Several features of this theorem deserve emphasis. The real-principal-type hypothesis is essential: at points where the principal symbol vanishes to higher order, the Hamilton flow is no longer a one-dimensional propagation curve, and singularities can spread along a higher-dimensional submanifold rather than along a single bicharacteristic. The use of FBI weights $\exp(-\varepsilon/h)$ in the analytic version is not a cosmetic change from Sobolev weights; polynomial losses in $h$ would only recover $C^\infty$ regularity, while the exponential gain in $h$ is the precise scaling that distinguishes real-analytic regularity from $C^\infty$. Finally, compactness of the bicharacteristic segment cannot be dropped: the finite iteration of the commutator estimate exhausts a finite number of cutoffs, and on a non-compact bicharacteristic the constants can degenerate as the curve runs to infinity in phase space. The model below illustrates the bracket-generating geometry that supplies plenty of real-principal-type bicharacteristics in the simplest non-Levi-flat hypersurface. [example: Heisenberg Hypersurface and Open CR Orbits] Write $z=x+iy$ and use coordinates $(x,y,s)$ on \begin{align*}M=\{(z,w)\in\mathbb C^2:\operatorname{Im}(w)=|z|^2\},\qquad w=s+i(x^2+y^2).\end{align*} We show that the CR orbit through every point is open by proving that the real CR directions and one commutator span all of $T_{(x,y,s)}M$. The antiholomorphic CR vector field is \begin{align*}L=\frac{\partial}{\partial \bar z}-iz\frac{\partial}{\partial s}.\end{align*} Since \begin{align*}\frac{\partial}{\partial\bar z}=\frac12\frac{\partial}{\partial x}+\frac{i}{2}\frac{\partial}{\partial y}\end{align*} and \begin{align*}z=x+iy,\end{align*} we get \begin{align*}L=\frac12\frac{\partial}{\partial x}+\frac{i}{2}\frac{\partial}{\partial y}-i(x+iy)\frac{\partial}{\partial s}.\end{align*} Expanding the scalar coefficient, \begin{align*}-i(x+iy)=-ix-i^2y=-ix+y.\end{align*} Therefore \begin{align*}L=\frac12\frac{\partial}{\partial x}+\frac{i}{2}\frac{\partial}{\partial y}+(y-ix)\frac{\partial}{\partial s}.\end{align*} Separating real and imaginary parts gives \begin{align*}L=\left(\frac12\frac{\partial}{\partial x}+y\frac{\partial}{\partial s}\right)+i\left(\frac12\frac{\partial}{\partial y}-x\frac{\partial}{\partial s}\right).\end{align*} Thus the real CR directions are spanned by \begin{align*}X=\operatorname{Re}L=\frac12\frac{\partial}{\partial x}+y\frac{\partial}{\partial s},\qquad Y=\operatorname{Im}L=\frac12\frac{\partial}{\partial y}-x\frac{\partial}{\partial s}.\end{align*} We compute $[X,Y]$ in the coordinate frame $(\partial_x,\partial_y,\partial_s)$ using \begin{align*}[X,Y]^j=X(Y^j)-Y(X^j).\end{align*} The coefficient functions are \begin{align*}X^x=\frac12,\qquad X^y=0,\qquad X^s=y,\qquad Y^x=0,\qquad Y^y=\frac12,\qquad Y^s=-x.\end{align*} For the $\partial_x$ coefficient, \begin{align*}X(Y^x)-Y(X^x)=X(0)-Y\left(\frac12\right)=0-0=0.\end{align*} For the $\partial_y$ coefficient, \begin{align*}X(Y^y)-Y(X^y)=X\left(\frac12\right)-Y(0)=0-0=0.\end{align*} For the $\partial_s$ coefficient, \begin{align*}X(Y^s)-Y(X^s)=X(-x)-Y(y).\end{align*} Now \begin{align*}X(-x)=\left(\frac12\frac{\partial}{\partial x}+y\frac{\partial}{\partial s}\right)(-x)=\frac12(-1)+y\cdot0=-\frac12.\end{align*} Also \begin{align*}Y(y)=\left(\frac12\frac{\partial}{\partial y}-x\frac{\partial}{\partial s}\right)(y)=\frac12(1)-x\cdot0=\frac12.\end{align*} Hence \begin{align*}X(Y^s)-Y(X^s)=-\frac12-\frac12=-1.\end{align*} Therefore \begin{align*}[X,Y]=0\frac{\partial}{\partial x}+0\frac{\partial}{\partial y}-1\frac{\partial}{\partial s}=-\frac{\partial}{\partial s}.\end{align*} This gives the missing $s$-direction: \begin{align*}\frac{\partial}{\partial s}=-[X,Y].\end{align*} Substituting $\partial_s=-[X,Y]$ into \begin{align*}X=\frac12\partial_x+y\partial_s\end{align*} gives \begin{align*}X=\frac12\partial_x-y[X,Y].\end{align*} Multiplying by $2$, \begin{align*}2X=\partial_x-2y[X,Y].\end{align*} Rearranging, \begin{align*}\partial_x=2X+2y[X,Y].\end{align*} Similarly, substituting $\partial_s=-[X,Y]$ into \begin{align*}Y=\frac12\partial_y-x\partial_s\end{align*} gives \begin{align*}Y=\frac12\partial_y+x[X,Y].\end{align*} Multiplying by $2$, \begin{align*}2Y=\partial_y+2x[X,Y].\end{align*} Rearranging, \begin{align*}\partial_y=2Y-2x[X,Y].\end{align*} Thus $\partial_s$, $\partial_x$, and $\partial_y$ all belong to the Lie algebra generated by $X$ and $Y$ at every point. Since \begin{align*}T_{(x,y,s)}M=\operatorname{span}_{\mathbb R}\left\{\frac{\partial}{\partial x},\frac{\partial}{\partial y},\frac{\partial}{\partial s}\right\},\end{align*} we have \begin{align*}T_{(x,y,s)}M\subset \operatorname{Lie}(X,Y)_{(x,y,s)}.\end{align*} The reverse inclusion holds because $X$, $Y$, and every iterated bracket of tangent vector fields are tangent to $M$. Hence \begin{align*}\operatorname{Lie}(X,Y)_{(x,y,s)}=T_{(x,y,s)}M\end{align*} at every point. By the *Orbit theorem*, the CR orbit through every point is an open subset of $M$. This Heisenberg hypersurface therefore has no lower-dimensional transverse CR leaf on which CR regularity can remain isolated. [/example] This example is the local model behind many finite-type hypersurfaces. The failure of Levi-flatness creates commutator directions, and those directions are exactly where propagation results become stronger than leafwise holomorphic continuation. ## The Baouendi-Treves Approximation Theorem The next problem is analytic rather than geometric: why should a function satisfying only tangential equations be approximable by holomorphic functions of the ambient variables? The Baouendi-Treves theorem supplies this bridge, and it is the reason CR functions often behave like boundary values of holomorphic functions even before a wedge has been constructed. [quotetheorem:9229] [citeproof:9229] The theorem is local and does not claim global polynomial approximation on all of $M$. Its strength is that the approximants are holomorphic in the ambient variables, while convergence is measured only on the CR submanifold. The approximation theorem turns uniqueness questions for CR functions into uniqueness questions for holomorphic polynomials. This is especially useful because zeros and analytic continuation of holomorphic functions are much more rigid than zeros of general smooth functions. Genericity of $M$ is not an incidental hypothesis in this statement. It ensures that the CR equations are tangential equations on a real submanifold with enough real directions for ambient holomorphic approximation to detect the local CR data. On a non-generic complex submanifold, such as $M=\{w=0\}\subset\mathbb C^2$, the CR condition is simply holomorphicity in the submanifold variable $z$, and the problem becomes ordinary holomorphic approximation on a complex variety rather than the tangential approximation theorem stated here. The reason for polynomial rather than merely holomorphic approximation is convenience and not depth: the proof first produces ambient holomorphic functions in a fixed Stein neighbourhood of the small piece of $M$, and a final Taylor truncation step on a compact subdomain converts them to polynomials without affecting the $L^2$ limit on $V$. The locality of the conclusion is sharp. Even on a strongly pseudoconvex hypersurface there exist continuous CR functions that are limits of holomorphic polynomials in every small chart but not the restriction of any entire function to all of $M$, because global polynomial approximation would force a global holomorphic extension that the topology of $M$ may obstruct. The smallness of the neighbourhood $U$ enters through the construction of the Gaussian approximate identity in the proof: the variance must be small enough that the Gaussians on $M$ remain confined to a coordinate chart in which $M$ is a graph, and uniform constants for the kernel are available only on such a chart. This sets up the following local unique-continuation consequence. [quotetheorem:9230] [citeproof:9230] The phrase "reachable" records the local nature of the proof: the argument is repeated along a chain of small coordinate patches adapted to CR curves. This result should be read as a local propagation mechanism rather than a global approximation theorem. At each step, genericity makes the chosen patch a uniqueness set for holomorphic approximants, while the $L^2$ convergence identifies the propagated CR function up to null sets. The totally real model below isolates the approximation mechanism in the simplest setting, where there are no positive-dimensional CR directions. [example: Approximation on a Totally Real Model] Let $U\subset\mathbb R^n$ be an open neighbourhood of the compact set $K$, and let $f\in L^2_{\mathrm{loc}}(U)$. Since $M=\mathbb R^n\subset\mathbb C^n$ is totally real, $H(M)=0$, so the CR equations impose no differential condition on $f$. We show that $f|_K$ is an $L^2(K)$ limit of restrictions of holomorphic polynomials on $\mathbb C^n$. Choose $\chi\in C_c^\infty(U)$ with $\chi=1$ on a neighbourhood of $K$, and define $F=\chi f$, extended by $0$ outside $U$. Since $\operatorname{supp}\chi$ is compact in $U$ and $f\in L^2_{\mathrm{loc}}(U)$, the product $\chi f$ belongs to $L^2(U)$ and its zero extension satisfies $F\in L^2(\mathbb R^n)$ with compact support. For $\varepsilon>0$, set \begin{align*} \rho_\varepsilon(x)=(2\pi\varepsilon)^{-n/2}\exp\left(-\frac{|x|^2}{2\varepsilon}\right). \end{align*} Define \begin{align*} F_\varepsilon(x)=(\rho_\varepsilon*F)(x)=\int_{\mathbb R^n}\rho_\varepsilon(x-y)F(y)\,dy. \end{align*} The kernels $\rho_\varepsilon$ are nonnegative and normalized. Indeed, using the one-dimensional [Gaussian integral](/theorems/1140) $\int_{\mathbb R}e^{-t^2/2}\,dt=(2\pi)^{1/2}$ and the change of variables $x=\sqrt\varepsilon\,u$, \begin{align*} \int_{\mathbb R^n}\rho_\varepsilon(x)\,dx=(2\pi\varepsilon)^{-n/2}\varepsilon^{n/2}\int_{\mathbb R^n}e^{-|u|^2/2}\,du=1. \end{align*} For every $\delta>0$, the same change of variables gives \begin{align*} \int_{|x|>\delta}\rho_\varepsilon(x)\,dx=(2\pi)^{-n/2}\int_{|u|>\delta/\sqrt\varepsilon}e^{-|u|^2/2}\,du. \end{align*} As $\varepsilon\to0$, the lower bound $\delta/\sqrt\varepsilon$ tends to $\infty$, so dominated convergence gives \begin{align*} \int_{|x|>\delta}\rho_\varepsilon(x)\,dx\to0. \end{align*} Thus $\rho_\varepsilon$ is an $L^2$ approximate identity, and hence \begin{align*} \|F_\varepsilon-F\|_{L^2(\mathbb R^n)}\to0. \end{align*} Since $\chi=1$ on a neighbourhood of $K$, we have $F=f$ almost everywhere on $K$. Therefore \begin{align*} \|F_\varepsilon-f\|_{L^2(K)}=\|F_\varepsilon-F\|_{L^2(K)}. \end{align*} Because $K\subset\mathbb R^n$, \begin{align*} \|F_\varepsilon-F\|_{L^2(K)}\le \|F_\varepsilon-F\|_{L^2(\mathbb R^n)}. \end{align*} It follows that \begin{align*} \|F_\varepsilon-f\|_{L^2(K)}\to0. \end{align*} For each fixed $\varepsilon$, define \begin{align*} G_\varepsilon(z)=(2\pi\varepsilon)^{-n/2}\int_{\mathbb R^n}\exp\left(-\frac{(z-y)\cdot(z-y)}{2\varepsilon}\right)F(y)\,dy. \end{align*} Here \begin{align*} (z-y)\cdot(z-y)=\sum_{j=1}^n(z_j-y_j)^2. \end{align*} For each fixed $y$, the function of $z$ inside the integral is entire. If $Q\subset\mathbb C^n$ is compact and $\operatorname{supp}F\subset B_R(0)$, then $|z-y|$ is bounded for $z\in Q$ and $y\in\operatorname{supp}F$. Every $z$-derivative of the exponential is a polynomial in $z-y$ times the same exponential, so on $Q\times\operatorname{supp}F$ it is bounded by a constant depending on $Q$, $R$, $\varepsilon$, and the derivative. Since $F$ is integrable on its compact support by Cauchy-Schwarz, differentiation under the integral is justified. Hence $G_\varepsilon$ is entire. For real $x\in\mathbb R^n$, \begin{align*} (x-y)\cdot(x-y)=\sum_{j=1}^n(x_j-y_j)^2=|x-y|^2. \end{align*} Thus \begin{align*} G_\varepsilon(x)=(2\pi\varepsilon)^{-n/2}\int_{\mathbb R^n}\exp\left(-\frac{|x-y|^2}{2\varepsilon}\right)F(y)\,dy. \end{align*} By the definition of $\rho_\varepsilon$, \begin{align*} G_\varepsilon(x)=\int_{\mathbb R^n}\rho_\varepsilon(x-y)F(y)\,dy=F_\varepsilon(x). \end{align*} Choose a compact polydisc $Q\subset\mathbb C^n$ containing $K$ in its interior. Since $G_\varepsilon$ is entire, its Taylor polynomials $P_{\varepsilon,N}$ converge uniformly to $G_\varepsilon$ on $Q$. On $K\subset Q\cap\mathbb R^n$, we have $F_\varepsilon(x)=G_\varepsilon(x)$, so \begin{align*} \|P_{\varepsilon,N}-F_\varepsilon\|_{L^2(K)}=\left(\int_K |P_{\varepsilon,N}(x)-G_\varepsilon(x)|^2\,dx\right)^{1/2}. \end{align*} For every $x\in K$, \begin{align*} |P_{\varepsilon,N}(x)-G_\varepsilon(x)|\le \sup_{q\in K}|P_{\varepsilon,N}(q)-G_\varepsilon(q)|. \end{align*} Therefore \begin{align*} \|P_{\varepsilon,N}-F_\varepsilon\|_{L^2(K)}\le |K|^{1/2}\sup_{q\in K}|P_{\varepsilon,N}(q)-G_\varepsilon(q)|. \end{align*} The supremum tends to $0$ as $N\to\infty$, so \begin{align*} \|P_{\varepsilon,N}-F_\varepsilon\|_{L^2(K)}\to0. \end{align*} Choose any sequence $\varepsilon_j\to0$. For each $j$, choose $N_j$ so large that \begin{align*} \|P_{\varepsilon_j,N_j}-F_{\varepsilon_j}\|_{L^2(K)}<\frac1j. \end{align*} Put $P_j=P_{\varepsilon_j,N_j}$. The triangle inequality gives \begin{align*} \|P_j|_{\mathbb R^n}-f\|_{L^2(K)}\le \|P_j-F_{\varepsilon_j}\|_{L^2(K)}+\|F_{\varepsilon_j}-f\|_{L^2(K)}. \end{align*} The first term is less than $1/j$, and the second term tends to $0$ because $\varepsilon_j\to0$. Hence \begin{align*} \|P_j|_{\mathbb R^n}-f\|_{L^2(K)}\to0. \end{align*} Thus even with no CR differential equations present, the totally real edge is still rich enough for local $L^2$ approximation by ambient holomorphic polynomials. [/example] This model explains why genericity is the correct ambient hypothesis. A generic real submanifold contains enough real directions to be a uniqueness set for holomorphic functions, even when its CR dimension is zero. ## The FBI Transform and Analytic Wave-front Sets The problem now is to locate analytic singularities with both position and direction. The FBI transform does this by testing a distribution against rapidly concentrated complex Gaussians; exponential decay as the semiclassical parameter tends to zero is the signature of real-analytic regularity. [definition: FBI Transform] For each $0<h\le1$, the FBI transform at scale $h$ is the linear map \begin{align*} T_h:\mathcal E'(\mathbb R^m)\to C^\infty(T^*\mathbb R^m;\mathbb C) \end{align*} defined by \begin{align*} T_hu(x,\xi) = c_m h^{-3m/4}\int_{\mathbb R^m} \exp\left(-\frac{|x-y|^2}{2h}-\frac{i}{h}\xi\cdot y\right)u(y)\,d\mathcal L^m(y), \end{align*} for $(x,\xi)\in T^*\mathbb R^m$, with $c_m>0$ fixed by the chosen normalisation. [/definition] The Gaussian localises near $x$, while the oscillatory factor tests frequency near $\xi/h$. Analytic regularity is stronger than smooth regularity, so it is detected by exponential estimates rather than rapid decay in powers of $h$. This motivates the following analytic wave-front definition. [definition: Analytic Wave-front Set] Let $u\in\mathcal D'(X)$ on a real-analytic manifold $X$. A nonzero covector $(x_0,\xi_0)\in T^*X\setminus0$ is absent from $\operatorname{WF}_a(u)$ if there are a real-analytic coordinate chart near $x_0$, a cut-off $\chi$ with $\chi=1$ near $x_0$, a conic neighbourhood $\Omega$ of $(x_0,\xi_0)$, and constants $C>0$, $\varepsilon>0$, $h_0>0$ such that \begin{align*} |T_h(\chi u)(x,\xi)|\le C\exp\left(-\frac{\varepsilon}{h}\right) \end{align*} for all $(x,\xi)\in\Omega$ and all $0<h<h_0$. [/definition] Thus $\operatorname{WF}_a(u)$ records the directions in which no such exponential decay estimate is available. It refines the analytic singular support by remembering covectors, not only base points. This leads to the following characterisation of real-analytic regularity. [quotetheorem:9231] [citeproof:9231] This theorem is the analytic analogue of the familiar statement that a distribution is smooth near $x_0$ exactly when its smooth wave-front set contains no covector over $x_0$. Several aspects of this characterisation are sharp. The exponential weight $\exp(-\varepsilon/h)$ in the definition of the analytic wave-front set is the smallest decay rate compatible with real-analyticity: weakening it to subexponential decay $\exp(-\varepsilon/h^\rho)$ with $0<\rho<1$ characterises Gevrey regularity of order $1/\rho$ rather than real-analyticity, and the proof of the converse direction breaks down because the contour shift no longer fits inside the Gevrey region of regularity of the symbol. The cutoff $\chi$ is harmless in the analytic category for a separate reason: multiplication by an analytic cutoff is microlocal away from the support boundary of $\chi$, and the FBI transform of $(1-\chi)u$ already has exponential decay near any covector over the interior of $\{\chi=1\}$. Hence the choice of $\chi$ does not change $\operatorname{WF}_a(u)$, even though analytic functions of compact support do not exist and the cutoff itself must be taken in a wider regularity class. For CR functions, the tangential Cauchy-Riemann equations eliminate analytic singularities in directions where the system is microlocally elliptic. The obstruction can only remain in characteristic covectors, where the principal symbols of the tangential operators vanish and the equations no longer control analytic decay of the FBI transform. Thus the relevant question is how sharply the analytic wave-front set is confined to the characteristic variety. [quotetheorem:9232] [citeproof:9232] The first statement is microlocal elliptic regularity for the analytic category. The second statement is the FBI characterisation applied in coordinates on $M$. The inclusion says that the CR equations remove analytic singularities in every non-characteristic direction, but it does not say that characteristic singularities must disappear. The inclusion can be strict: on a Levi-flat hypersurface, every conormal direction over a point lies in $\operatorname{Char}(\bar\partial_b)$, but a CR function pulled back from a real-analytic leafwise datum has empty analytic wave-front set. Conversely, the same conormal directions are available for genuinely non-analytic transverse data, and that is the source of the obstruction. On a minimal CR submanifold the situation is markedly different: bracket-generation forces $\operatorname{Char}(\bar\partial_b)$ to project nontrivially under the Hamilton flow, and propagation results combined with wedge extension push the analytic wave-front set onto a thinner conic set than the whole characteristic variety. In the analytic category, hypoellipticity of $\bar\partial_b$ at $p$ is the statement that no CR distribution has $p$ in its analytic singular support; the inclusion above reduces this to the statement that $\operatorname{WF}_a(f)$ has no covector over $p$, which is checked direction by direction along $\operatorname{Char}(\bar\partial_b)$. In practice, computing $\operatorname{WF}_a(f)$ for a candidate CR function proceeds in two steps: discard all non-characteristic covectors using the theorem, then test the remaining conormal directions by evaluating $T_h(\chi f)$ along them and checking whether $|T_h(\chi f)(x,\xi)|$ decays like $\exp(-\varepsilon/h)$ uniformly in a conic neighbourhood. Levi-flat manifolds provide the basic obstruction: transverse smooth data can be invisible to the CR vector fields and still fail to be real-analytic. The next example exhibits exactly such a conormal analytic singularity. [example: Smooth CR Data with Analytic Wave-front] [claim]For each $z_0\in\mathbb C$, the function $f(z,s)=g(s)$ is CR near $(z_0,0)$, and if $g$ is not real-analytic at $0$, then $\operatorname{WF}_a(f)$ contains at least one nonzero conormal covector $\tau_0\,ds$ over $(z_0,0)$.[/claim] [proof]Write $z=x+iy$ and identify \begin{align*}M=\{(z,w)\in\mathbb C^2:\operatorname{Im}w=0\}\end{align*} with $\mathbb R^3_{x,y,s}$ by $w=s$. The antiholomorphic tangent direction is \begin{align*}L=\frac{\partial}{\partial\bar z}=\frac12\left(\frac{\partial}{\partial x}+i\frac{\partial}{\partial y}\right).\end{align*} Since $f(x,y,s)=g(s)$, the variables $x$ and $y$ do not occur in $f$. Hence \begin{align*}\frac{\partial f}{\partial x}=0.\end{align*} Also \begin{align*}\frac{\partial f}{\partial y}=0.\end{align*} Therefore \begin{align*}Lf=\frac12\left(\frac{\partial f}{\partial x}+i\frac{\partial f}{\partial y}\right)=\frac12(0+i0)=0.\end{align*} Thus $f$ is CR. The real holomorphic tangent bundle is \begin{align*}H(M)=\operatorname{span}_{\mathbb R}\left\{\frac{\partial}{\partial x},\frac{\partial}{\partial y}\right\}.\end{align*} Indeed, $J\partial_x=\partial_y$ and $J\partial_y=-\partial_x$ are tangent to $M$, while $J\partial_s$ points in the normal $\operatorname{Im}w$ direction. A covector at a point of $M$ has the form \begin{align*}\eta=\xi_x\,dx+\xi_y\,dy+\tau\,ds.\end{align*} It annihilates $H(M)$ exactly when \begin{align*}\eta(\partial_x)=\xi_x=0\end{align*} and \begin{align*}\eta(\partial_y)=\xi_y=0.\end{align*} Thus the nonzero conormal covectors are precisely $\tau\,ds$ with $\tau\ne0$. Since $g$ is not real-analytic at $0$, the contrapositive of the *FBI Characterisation of Real-analytic Regularity* gives some $\tau_0\ne0$ such that \begin{align*}(0,\tau_0)\in\operatorname{WF}_a(g).\end{align*} Fix $z_0=x_0+iy_0$. Choose nonnegative cutoffs $\alpha,\beta,\theta\in C_c^\infty(\mathbb R)$ with $\alpha=1$ near $x_0$, $\beta=1$ near $y_0$, and $\theta=1$ near $0$, and set \begin{align*}\chi(x,y,s)=\alpha(x)\beta(y)\theta(s).\end{align*} At $(x_0,y_0,0;0,0,\tau)$, the three-dimensional FBI transform is \begin{align*}T_h^{(3)}(\chi f)(x_0,y_0,0;0,0,\tau)=c_3h^{-9/4}\int_{\mathbb R^3}e^{-\frac{(x_0-X)^2+(y_0-Y)^2+S^2}{2h}}e^{-\frac{i}{h}\tau S}\alpha(X)\beta(Y)\theta(S)g(S)\,dX\,dY\,dS.\end{align*} The integrand factors into an $X$-factor, a $Y$-factor, and an $S$-factor, so [Fubini's theorem](/theorems/2961) gives \begin{align*}T_h^{(3)}(\chi f)(x_0,y_0,0;0,0,\tau)=c_3h^{-9/4}A_hB_hC_h(\tau),\end{align*} where \begin{align*}A_h=\int_{\mathbb R}\exp\left(-\frac{(x_0-X)^2}{2h}\right)\alpha(X)\,dX,\end{align*} \begin{align*}B_h=\int_{\mathbb R}\exp\left(-\frac{(y_0-Y)^2}{2h}\right)\beta(Y)\,dY,\end{align*} and \begin{align*}C_h(\tau)=\int_{\mathbb R}\exp\left(-\frac{S^2}{2h}-\frac{i}{h}\tau S\right)\theta(S)g(S)\,dS.\end{align*} The one-dimensional FBI transform of $\theta g$ is \begin{align*}T_h^{(1)}(\theta g)(0,\tau)=c_1h^{-3/4}C_h(\tau).\end{align*} For all sufficiently small $h$, the interval $|X-x_0|<\sqrt h$ lies where $\alpha=1$, and the interval $|Y-y_0|<\sqrt h$ lies where $\beta=1$. Hence \begin{align*}A_h\ge \int_{|X-x_0|<\sqrt h}\exp\left(-\frac{(x_0-X)^2}{2h}\right)\,dX.\end{align*} With $u=(X-x_0)/\sqrt h$, we have $dX=\sqrt h\,du$ and $(x_0-X)^2/h=u^2$, so \begin{align*}\int_{|X-x_0|<\sqrt h}\exp\left(-\frac{(x_0-X)^2}{2h}\right)\,dX=h^{1/2}\int_{-1}^{1}e^{-u^2/2}\,du.\end{align*} The same change of variables gives \begin{align*}B_h\ge h^{1/2}\int_{-1}^{1}e^{-u^2/2}\,du.\end{align*} With \begin{align*}C_0=\left(\int_{-1}^{1}e^{-u^2/2}\,du\right)^2>0,\end{align*} we get \begin{align*}A_hB_h\ge C_0h.\end{align*} Suppose, for contradiction, that $(x_0,y_0,0;0,0,\tau_0)$ were absent from $\operatorname{WF}_a(f)$. After shrinking the supports of the cutoffs inside the corresponding microlocal neighbourhood, there would be constants $C,\varepsilon>0$ and a conic neighbourhood of $\tau_0$ such that \begin{align*}|T_h^{(3)}(\chi f)(x_0,y_0,0;0,0,\tau)|\le C\exp\left(-\frac{\varepsilon}{h}\right)\end{align*} for all sufficiently small $h$ and all $\tau$ in that conic neighbourhood. From the factorisation and the one-dimensional formula, \begin{align*}|T_h^{(1)}(\theta g)(0,\tau)|=\frac{c_1}{c_3}h^{3/2}(A_hB_h)^{-1}|T_h^{(3)}(\chi f)(x_0,y_0,0;0,0,\tau)|.\end{align*} Using $A_hB_h\ge C_0h$, we obtain \begin{align*}|T_h^{(1)}(\theta g)(0,\tau)|\le \frac{c_1C}{c_3C_0}h^{1/2}\exp\left(-\frac{\varepsilon}{h}\right).\end{align*} For $0<h<1$, $h^{1/2}\le1\le \exp(\varepsilon/(2h))$, so \begin{align*}h^{1/2}\exp\left(-\frac{\varepsilon}{h}\right)\le \exp\left(-\frac{\varepsilon}{2h}\right).\end{align*} Therefore \begin{align*}|T_h^{(1)}(\theta g)(0,\tau)|\le C'\exp\left(-\frac{\varepsilon}{2h}\right)\end{align*} for a constant $C'>0$. This is an exponential FBI estimate near $(0,\tau_0)$ for $\theta g$, contradicting $(0,\tau_0)\in\operatorname{WF}_a(g)$. Hence \begin{align*}(x_0,y_0,0;0,0,\tau_0)\in\operatorname{WF}_a(f).\end{align*} [/proof] Thus the CR equation sees only the leaf variables $z$, while the analytic wave-front set detects the non-analytic transverse dependence in the conormal $ds$ direction. [/example] This example is the basic warning for the chapter. Smooth CR regularity is leafwise in a Levi-flat model, while real-analytic regularity is sensitive to transverse variables detected by conormal covectors. ## Wedge Extension and Propagation of Holomorphic Extendibility The final problem is to turn CR boundary data into an actual holomorphic function on one side of the CR manifold. Minimality is the geometric condition that prevents CR functions from being confined to smaller CR submanifolds, and wedges are the natural domains of extension. [definition: Minimal CR Manifold] Let $M$ be a $C^\infty$ CR manifold and let $p\in M$. The manifold $M$ is minimal at $p$ if the CR orbit $\mathcal O(p)$ contains an open neighbourhood of $p$ in $M$. [/definition] Minimality was introduced in Chapter 0 through CR orbits; here it becomes the bracket-generating condition for wedge extension. It is weaker than strict pseudoconvexity and strong enough for local wedge extension. To verify minimality at a concrete point in practice, work in local coordinates in which $M$ is a graph $\operatorname{Im}(w) = \varphi(z,\operatorname{Re}(w))$ over its complex tangent variables, write out a local basis $L_1,\dots,L_n$ of the CR distribution as first-order operators with coefficients depending on $\varphi$, and form iterated commutators $[L_i, \bar L_j]$, $[L_i, [L_j, \bar L_k]]$, and so on. Evaluate the real and imaginary parts of these brackets at the point of interest and assemble the resulting vectors as columns of a real matrix; minimality at the point holds whenever the column span equals $T_p M$. For a strictly pseudoconvex hypersurface, the first-order commutator $[L, \bar L]$ already supplies the missing transverse direction; for a finite-type hypersurface of type $m$, brackets up to depth $m$ are required, and the depth at which the span fills $T_p M$ coincides with the bracket type at $p$. Constructing the analytic discs that enter the wedge-extension proof is similarly explicit on a quadric: parametrise the unit disc as $\zeta\in\overline{B}(0,1)\subset\mathbb C$, write a candidate map $\zeta\mapsto(z(\zeta), w(\zeta))$ with $w(\zeta) = w_0 + \zeta \, h(\zeta)$ for a suitable holomorphic function $h$, and solve the Bishop equation $\operatorname{Im}(w(\zeta)) = |z(\zeta)|^2$ on $|\zeta|=1$ as a fixed-point problem for $h$. Small data $h$ yields discs of small size with boundary in $M$; varying the boundary parameter along CR flows then sweeps the normal cone. [definition: Wedge with Edge] Let $M\subset\mathbb C^N$ be a generic submanifold, let $p\in M$, and let $\Gamma$ be an open cone in the normal space $T_p\mathbb C^N/T_pM$. A wedge with edge $M$ at $p$ and direction cone $\Gamma$ is a set of the form \begin{align*} \mathcal W=\{q+\eta:q\in U,\ \eta\in\Gamma_q,\ |\eta|<\varepsilon\} \end{align*} in a local normal coordinate identification, where $U\subset M$ is a neighbourhood of $p$ and $(\Gamma_q)_{q\in U}$ is a smooth family of open cones with $\Gamma_p=\Gamma$. [/definition] A wedge is not usually a one-sided tubular neighbourhood. Its cone records the normal complex directions supplied by analytic discs attached to the CR manifold, which is exactly what the following wedge-extension theorem constructs. [quotetheorem:9233] [proofunderconstruction:9233] This is the main extension theorem of the chapter. It says that minimality is the geometric mechanism that converts tangential CR equations into ambient holomorphicity. The continuity hypothesis on the CR function is comfortable but not sharp; refinements of the argument extend the theorem to CR distributions, with $L^2$ and distributional boundary values replacing pointwise continuity. The hypothesis of minimality is sharp in a precise sense: if $M$ fails to be minimal at $p$, the CR orbit through $p$ is a proper CR submanifold of strictly lower dimension, and CR functions can be constructed that are arbitrary on the orbit and locally constant in the transverse directions. Such functions admit no wedge extension at $p$, and Levi-flat hypersurfaces supply explicit counterexamples in every dimension. The normal cone $\Gamma$ produced by the proof is not canonical: it records which direction the family of analytic discs penetrates first, and different families yield different but compatible cones. What is canonical is the existence of some open cone of admissible directions; any two wedges obtained at $p$ have intersecting cones whose common refinement still supports the holomorphic extension. A wedge extension obtained at one point still leaves a transport problem: a CR orbit is connected by tangential complex directions, while the wedge lives in normal directions that may twist as the base point moves. The obstruction is that analytic discs attached at nearby points need not enter the same cone, so one must know whether extendibility survives after shrinking and changing the wedge. Propagation results answer this orbitwise question. [quotetheorem:9234] [citeproof:9234] The theorem explains why wedge extension is not a pointwise accident. It is transported by the same CR geometry that transports uniqueness and microlocal regularity. The theorem is local in both the edge and the normal cone: the wedge obtained after propagation may be smaller and may point in a different admissible direction. This flexibility is essential because analytic discs deform continuously along the CR curve, but their normal directions need not remain fixed. The quadric model below shows the propagation and extension mechanism in a finite-type hypersurface where the bracket directions are visible in coordinates. [example: Wedge Extension from a Quadric Submanifold] [claim]Every continuous CR function near the origin on \begin{align*} M=\{(z,w)\in\mathbb C^2:\operatorname{Im}(w)=|z|^2\} \end{align*} extends holomorphically to a wedge whose normal direction contains the side $\operatorname{Im}(w)>|z|^2$.[/claim] [proof]Write $z=x+iy$ and parametrize $M$ by \begin{align*} (x,y,s)\mapsto (z,w)=(x+iy,s+i(x^2+y^2)). \end{align*} Equivalently, \begin{align*} w=s+iz\bar z. \end{align*} Since \begin{align*} \frac{\partial}{\partial\bar z}=\frac12\left(\partial_x+i\partial_y\right), \end{align*} define \begin{align*} L=\frac{\partial}{\partial\bar z}-iz\partial_s. \end{align*} On the restricted ambient coordinate $z$, \begin{align*} Lz=\frac{\partial z}{\partial\bar z}-iz\frac{\partial z}{\partial s}=0-iz\cdot0=0. \end{align*} On the restricted ambient coordinate $w=s+iz\bar z$, \begin{align*} \frac{\partial}{\partial\bar z}(s+iz\bar z)=0+i\left(\frac{\partial z}{\partial\bar z}\bar z+z\frac{\partial\bar z}{\partial\bar z}\right)=i(0\cdot\bar z+z\cdot1)=iz. \end{align*} Also, \begin{align*} \partial_s(s+iz\bar z)=1+i\partial_s(z\bar z)=1+0=1. \end{align*} Therefore \begin{align*} Lw=iz-iz\cdot1=0. \end{align*} Thus $L$ annihilates the restrictions of the ambient holomorphic coordinates $z$ and $w$, so $L$ spans $T^{0,1}M$. Expanding $L$ in real coordinates gives \begin{align*} L=\frac12\partial_x+\frac{i}{2}\partial_y-i(x+iy)\partial_s. \end{align*} Since \begin{align*} -i(x+iy)=-ix-i^2y=-ix+y, \end{align*} we have \begin{align*} L=\left(\frac12\partial_x+y\partial_s\right)+i\left(\frac12\partial_y-x\partial_s\right). \end{align*} Hence the real CR directions are \begin{align*} X=\operatorname{Re}L=\frac12\partial_x+y\partial_s \end{align*} and \begin{align*} Y=\operatorname{Im}L=\frac12\partial_y-x\partial_s. \end{align*} We compute $[X,Y]$ in the coordinate frame $(\partial_x,\partial_y,\partial_s)$ using $[X,Y]^j=X(Y^j)-Y(X^j)$. The coefficients are \begin{align*} X^x=\frac12,\quad X^y=0,\quad X^s=y,\quad Y^x=0,\quad Y^y=\frac12,\quad Y^s=-x. \end{align*} For the $\partial_x$ coefficient, \begin{align*} X(Y^x)-Y(X^x)=X(0)-Y\left(\frac12\right)=0-0=0. \end{align*} For the $\partial_y$ coefficient, \begin{align*} X(Y^y)-Y(X^y)=X\left(\frac12\right)-Y(0)=0-0=0. \end{align*} For the $\partial_s$ coefficient, \begin{align*} X(Y^s)-Y(X^s)=X(-x)-Y(y). \end{align*} Now \begin{align*} X(-x)=\left(\frac12\partial_x+y\partial_s\right)(-x)=\frac12(-1)+y\cdot0=-\frac12. \end{align*} Also, \begin{align*} Y(y)=\left(\frac12\partial_y-x\partial_s\right)(y)=\frac12(1)-x\cdot0=\frac12. \end{align*} Therefore \begin{align*} X(Y^s)-Y(X^s)=-\frac12-\frac12=-1. \end{align*} Thus \begin{align*} [X,Y]=0\partial_x+0\partial_y-\partial_s=-\partial_s. \end{align*} It follows that \begin{align*} \partial_s=-[X,Y]. \end{align*} Substituting this into $X=\frac12\partial_x+y\partial_s$ gives \begin{align*} X=\frac12\partial_x-y[X,Y]. \end{align*} Multiplying by $2$ gives \begin{align*} 2X=\partial_x-2y[X,Y]. \end{align*} Hence \begin{align*} \partial_x=2X+2y[X,Y]. \end{align*} Similarly, substituting $\partial_s=-[X,Y]$ into $Y=\frac12\partial_y-x\partial_s$ gives \begin{align*} Y=\frac12\partial_y+x[X,Y]. \end{align*} Multiplying by $2$ gives \begin{align*} 2Y=\partial_y+2x[X,Y]. \end{align*} Hence \begin{align*} \partial_y=2Y-2x[X,Y]. \end{align*} Thus $\partial_x$, $\partial_y$, and $\partial_s$ all lie in the Lie algebra generated by $X$ and $Y$. Since \begin{align*} T_{(x,y,s)}M=\operatorname{span}_{\mathbb R}\{\partial_x,\partial_y,\partial_s\}, \end{align*} the bracket-generated distribution equals $TM$ at every point. By the *Orbit theorem*, the CR orbit through the origin is open, so $M$ is minimal at the origin. Now set \begin{align*} \rho(z,w)=|z|^2-\operatorname{Im}(w)=z\bar z-\frac{w-\bar w}{2i}. \end{align*} Then $M=\{\rho=0\}$. The inequality $\operatorname{Im}(w)>|z|^2$ is equivalent to $|z|^2-\operatorname{Im}(w)<0$, so that side is $\{\rho<0\}$. Treat $z,w,\bar z,\bar w$ as independent variables for complex differentiation. Then \begin{align*} \frac{\partial\rho}{\partial z}=\frac{\partial}{\partial z}\left(z\bar z-\frac{w-\bar w}{2i}\right)=\bar z. \end{align*} Also, \begin{align*} \frac{\partial\rho}{\partial w}=\frac{\partial}{\partial w}\left(z\bar z-\frac{w-\bar w}{2i}\right)=-\frac1{2i}=\frac{i}{2}. \end{align*} Therefore a $(1,0)$ vector $V=a\partial_z+b\partial_w$ is complex-tangent to $M$ exactly when \begin{align*} d\rho(V)=\bar z\,a+\frac{i}{2}b=0. \end{align*} The second complex derivatives are \begin{align*} \frac{\partial^2\rho}{\partial z\,\partial\bar z}=1, \end{align*} \begin{align*} \frac{\partial^2\rho}{\partial z\,\partial\bar w}=0, \end{align*} \begin{align*} \frac{\partial^2\rho}{\partial w\,\partial\bar z}=0, \end{align*} and \begin{align*} \frac{\partial^2\rho}{\partial w\,\partial\bar w}=0. \end{align*} Thus the Levi form on $V=a\partial_z+b\partial_w$ is \begin{align*} \mathcal L_\rho(V,\overline V)=1\cdot a\bar a+0\cdot a\bar b+0\cdot b\bar a+0\cdot b\bar b=|a|^2. \end{align*} If $V\ne0$ is complex-tangent and $a=0$, then the tangency equation gives \begin{align*} \frac{i}{2}b=0. \end{align*} Multiplying by $2/i$ gives \begin{align*} b=0, \end{align*} contradicting $V\ne0$. Hence $a\ne0$ for every nonzero complex-tangent $V$, and therefore \begin{align*} \mathcal L_\rho(V,\overline V)=|a|^2>0. \end{align*} So $\rho$ is strictly plurisubharmonic in complex-tangent directions, and the strictly pseudoconvex side is $\{\rho<0\}$, namely $\operatorname{Im}(w)>|z|^2$. Since $M$ is minimal at the origin, the *Trepreau-Tumanov Wedge Extension Theorem* gives holomorphic wedge extension for every continuous CR function near the origin. The positive Levi-form computation identifies an admissible normal direction on the side $\rho<0$. After shrinking the neighbourhood and the cone, the wedge can be taken in the model form \begin{align*} \mathcal W_{\delta,\varepsilon}=\{(z,w): |z|<\delta,\ |\operatorname{Re}w|<\delta,\ 0<\operatorname{Im}(w)-|z|^2<\varepsilon\}. \end{align*} For every point of $\mathcal W_{\delta,\varepsilon}$, \begin{align*} 0<\operatorname{Im}(w)-|z|^2<\varepsilon, \end{align*} so \begin{align*} \operatorname{Im}(w)>|z|^2. \end{align*} Thus every continuous CR function near the origin on $M$ extends holomorphically to a wedge pointing into that side.[/proof] The commutator calculation shows that CR motion fills all tangent directions, while the Levi-form calculation identifies the strictly pseudoconvex normal side into which the holomorphic extension enters. [/example] The positive model should not be mistaken for unrestricted continuation. Propagation results move existing regularity through regions where the CR function is defined and the geometric hypotheses remain available; they do not remove genuine singularities already present in the holomorphic variable. The following Levi-flat example marks that boundary. [example: A CR Curve Where Extension Stops at a Pole] [claim]For \begin{align*}M=\{(z,w)\in\mathbb C^2:\operatorname{Im}(w)=0\},\qquad w=s,\end{align*} and \begin{align*}U=\{(z,s)\in M: |z|<2,\ |s|<1,\ z\ne i\},\end{align*} the function $f(z,s)=1/(z-i)$ is CR on $U$, but CR continuation along $\gamma(t)=(it,0)$ for $0\le t<1$ cannot pass through $(i,0)$.[/claim] [proof]Write $z=x+iy$. On $M$, the coordinate $w=s$ is real, so $M$ is parametrized by $(x,y,s)\in\mathbb R^3$. The complex tangent directions are the $z$-directions, and the antiholomorphic CR direction is \begin{align*}L=\frac{\partial}{\partial\bar z}=\frac12\left(\frac{\partial}{\partial x}+i\frac{\partial}{\partial y}\right).\end{align*} Since \begin{align*}z-i=x+iy-i=x+i(y-1),\end{align*} we have \begin{align*}f(x,y,s)=\frac{1}{x+i(y-1)}.\end{align*} On $U$, the condition $z\ne i$ is exactly \begin{align*}x+i(y-1)\ne0,\end{align*} so the following differentiations are valid. Differentiating with respect to $x$ gives \begin{align*}\frac{\partial f}{\partial x}=\frac{\partial}{\partial x}\left(x+i(y-1)\right)^{-1}=-(x+i(y-1))^{-2}\cdot 1=-\frac{1}{(x+i(y-1))^2}.\end{align*} Differentiating with respect to $y$ gives \begin{align*}\frac{\partial f}{\partial y}=\frac{\partial}{\partial y}\left(x+i(y-1)\right)^{-1}=-(x+i(y-1))^{-2}\cdot i=-\frac{i}{(x+i(y-1))^2}.\end{align*} Therefore \begin{align*}Lf=\frac12\left(\frac{\partial f}{\partial x}+i\frac{\partial f}{\partial y}\right)=\frac12\left(-\frac{1}{(x+i(y-1))^2}+i\left(-\frac{i}{(x+i(y-1))^2}\right)\right).\end{align*} Since $i(-i)=-i^2=1$, this becomes \begin{align*}Lf=\frac12\left(-\frac{1}{(x+i(y-1))^2}+\frac{1}{(x+i(y-1))^2}\right).\end{align*} Hence \begin{align*}Lf=\frac12\cdot0=0.\end{align*} Thus $f$ is CR at every point of $U$. The real holomorphic tangent bundle of this Levi-flat hypersurface is \begin{align*}H(M)=\operatorname{span}_{\mathbb R}\left\{\frac{\partial}{\partial x},\frac{\partial}{\partial y}\right\}.\end{align*} Indeed, $J\partial_x=\partial_y$ and $J\partial_y=-\partial_x$ remain tangent to $M$, while $J\partial_s$ points in the normal $\operatorname{Im}(w)$ direction. For $\gamma(t)=(it,0)$, the real coordinates are \begin{align*}x(t)=0,\qquad y(t)=t,\qquad s(t)=0.\end{align*} Differentiating these coordinate functions gives \begin{align*}\dot\gamma(t)=\frac{dx}{dt}\partial_x+\frac{dy}{dt}\partial_y+\frac{ds}{dt}\partial_s=0\cdot\partial_x+1\cdot\partial_y+0\cdot\partial_s=\partial_y.\end{align*} Since $\partial_y\in H_{\gamma(t)}(M)$, the path $\gamma$ is a CR curve for $0\le t<1$. Along this curve, \begin{align*}f(\gamma(t))=f(it,0)=\frac{1}{it-i}.\end{align*} Factoring the denominator gives \begin{align*}it-i=i(t-1).\end{align*} Therefore \begin{align*}f(\gamma(t))=\frac{1}{i(t-1)}.\end{align*} Since $1/i=-i$, we get \begin{align*}f(\gamma(t))=-\frac{i}{t-1}.\end{align*} Because $-(t-1)=1-t$, this is \begin{align*}f(\gamma(t))=\frac{i}{1-t}.\end{align*} For $0\le t<1$, we have $1-t>0$, so \begin{align*}|f(\gamma(t))|=\left|\frac{i}{1-t}\right|=\frac{|i|}{|1-t|}=\frac{1}{1-t}.\end{align*} As $t\to1^-$, the denominator $1-t$ tends to $0$ through positive values, hence \begin{align*}\frac{1}{1-t}\to\infty.\end{align*} Thus $f$ has no finite continuous value at the endpoint $(i,0)$. Consequently it cannot have a smooth or holomorphic continuation through that point, since either kind of continuation would in particular be finite and continuous there.[/proof] The obstruction is the actual pole of the one-variable holomorphic function $z\mapsto 1/(z-i)$ on the Levi-flat leaf $s=0$, so the curve reaches the boundary of the domain where the CR function is defined rather than contradicting propagation. [/example] The chapter therefore links three versions of the same phenomenon. CR curves describe where the equations move information on the manifold, bicharacteristics describe where singularities move in phase space, and wedges describe where CR boundary values become holomorphic functions in the ambient complex manifold. The final chapter changes scale from this local propagation picture to global embedding, filling, and curvature questions. The local propagation picture finally points beyond individual singularities to the broader structure of the theory. The last chapter gathers these themes into global questions of embedding, filling, and curvature, showing how the local CR tools developed so far fit into the larger geometry. # 11. Global Topics and Further Directions The preceding chapters built the local language of CR geometry: Levi forms, finite type, normal forms, tangential Cauchy-Riemann operators, extension theorems, and subelliptic estimates. This final chapter changes scale. Instead of asking only what happens near one boundary point, we ask which compact CR manifolds exist globally, which of them bound complex manifolds, and which scalar invariants survive under CR equivalence. The main theme is that the boundary viewpoint of several complex variables is both analytic and geometric. A strictly pseudoconvex boundary carries a contact form, a Webster connection, curvature, and CR invariants; at the same time, its CR functions and kernel asymptotics remember the holomorphic geometry of the domain it bounds. ## Global Embedding and Fillability in CR Dimension One Which abstract CR 3-manifolds arise as actual boundaries of complex surfaces, and how can a global obstruction detect failure? Chapter 2 gave local CR functions in the real-analytic setting, and Chapter 9 gave enough global CR functions to embed in favourable high dimensions. Real dimension $3$ is the borderline case: the tangential Cauchy-Riemann operator is underdetermined enough to have many local phenomena, but global embeddability and fillability can still fail. [definition: Global CR Embedding] Let $(M,T^{1,0}M)$ be a compact CR manifold. A global CR embedding into a complex manifold $X$ is a smooth embedding $\Phi:M\to X$ such that $d\Phi(T^{0,1}M)\subseteq T^{0,1}X$ and the CR structure induced on $\Phi(M)$ agrees with the original CR structure on $M$. [/definition] The definition packages two demands. The map must embed the underlying smooth manifold, and its components must be CR functions in the sense of the tangential Cauchy-Riemann equations. Thus global embeddability is partly a function theory question: do global CR functions separate points and tangent directions? [definition: Strictly Pseudoconvex Filling] Let $(M,T^{1,0}M)$ be a compact strictly pseudoconvex CR manifold. A strictly pseudoconvex filling of $M$ is a compact complex manifold $X$ with smooth boundary together with a CR diffeomorphism $\partial X\to M$ such that the boundary CR structure induced by $X$ is the given structure on $M$. [/definition] Fillability is stronger than embeddability into some complex manifold. It asks that $M$ occur as the boundary of a complex manifold on the pseudoconvex side, so it remembers global holomorphic extension, not only the existence of CR coordinates. [example: The Standard Three Sphere Is Fillable] Let $S^3=\{(z_1,z_2)\in\mathbb C^2: |z_1|^2+|z_2|^2=1\}$ and set $\rho(z)=|z_1|^2+|z_2|^2-1$. Then $B^2=\{\rho<0\}$ and $\partial B^2=\{\rho=0\}=S^3$. We show that the inclusion $\iota:S^3\hookrightarrow\mathbb C^2$ is a global CR embedding and that $\overline{B^2}$ is a strictly pseudoconvex filling. At $p=(z_1,z_2)\in S^3$, an ambient $(1,0)$ vector has the form \begin{align*} V=a_1\frac{\partial}{\partial z_1}+a_2\frac{\partial}{\partial z_2}. \end{align*} It is complex tangent to the level set $\rho=0$ exactly when $V\rho=0$. Since \begin{align*} \rho=z_1\bar z_1+z_2\bar z_2-1, \end{align*} we have \begin{align*} \frac{\partial\rho}{\partial z_1}=\bar z_1,\qquad \frac{\partial\rho}{\partial z_2}=\bar z_2. \end{align*} Therefore \begin{align*} V\rho=a_1\bar z_1+a_2\bar z_2. \end{align*} Thus \begin{align*} T^{1,0}_pS^3=\left\{a_1\frac{\partial}{\partial z_1}+a_2\frac{\partial}{\partial z_2}:\bar z_1a_1+\bar z_2a_2=0\right\}. \end{align*} The vector \begin{align*} L_p=\bar z_2\frac{\partial}{\partial z_1}-\bar z_1\frac{\partial}{\partial z_2} \end{align*} lies in this space because \begin{align*} L_p\rho=\bar z_2\bar z_1-\bar z_1\bar z_2=0. \end{align*} Since $p\in S^3$, the pair $(z_1,z_2)$ is not $(0,0)$, so the linear functional $(a_1,a_2)\mapsto \bar z_1a_1+\bar z_2a_2$ is nonzero. Its kernel in $\mathbb C^2$ has complex dimension $1$, and $L_p\ne0$, so $L_p$ spans $T^{1,0}_pS^3$. The conjugate vector \begin{align*} \bar L_p=z_2\frac{\partial}{\partial \bar z_1}-z_1\frac{\partial}{\partial \bar z_2} \end{align*} spans $T^{0,1}_pS^3$. The coordinate functions of the inclusion are CR functions because \begin{align*} \bar L_p z_1=z_2\frac{\partial z_1}{\partial\bar z_1}-z_1\frac{\partial z_1}{\partial\bar z_2}=z_2\cdot0-z_1\cdot0=0. \end{align*} Similarly, \begin{align*} \bar L_p z_2=z_2\frac{\partial z_2}{\partial\bar z_1}-z_1\frac{\partial z_2}{\partial\bar z_2}=z_2\cdot0-z_1\cdot0=0. \end{align*} The underlying map $\iota$ is the ordinary smooth inclusion, hence is an embedding. Also \begin{align*} d\iota_p(\bar L_p)=\bar L_p\in T^{0,1}_p\mathbb C^2. \end{align*} So $\iota$ sends $T^{0,1}S^3$ into $T^{0,1}\mathbb C^2$ and induces exactly the CR structure displayed above. Therefore $\iota$ is a global CR embedding. For strict pseudoconvexity, compute the complex Hessian of $\rho$. Since \begin{align*} \frac{\partial\rho}{\partial z_j}=\bar z_j \end{align*} for $j=1,2$, we get \begin{align*} \frac{\partial^2\rho}{\partial z_j\,\partial\bar z_k}=\frac{\partial\bar z_j}{\partial\bar z_k}=\delta_{jk}. \end{align*} If $V=a_1\partial_{z_1}+a_2\partial_{z_2}\in T^{1,0}_pS^3$, then the Levi form associated to $\rho$ is \begin{align*} \mathcal L_\rho(V,\bar V)=\sum_{j,k=1}^2\frac{\partial^2\rho}{\partial z_j\,\partial\bar z_k}a_j\bar a_k. \end{align*} Substituting the Hessian entries gives \begin{align*} \mathcal L_\rho(V,\bar V)=\delta_{11}a_1\bar a_1+\delta_{12}a_1\bar a_2+\delta_{21}a_2\bar a_1+\delta_{22}a_2\bar a_2. \end{align*} Since $\delta_{11}=\delta_{22}=1$ and $\delta_{12}=\delta_{21}=0$, this becomes \begin{align*} \mathcal L_\rho(V,\bar V)=|a_1|^2+|a_2|^2. \end{align*} This is positive whenever $V\ne0$. In particular, for $L_p$ the coefficients are $a_1=\bar z_2$ and $a_2=-\bar z_1$, so \begin{align*} \mathcal L_\rho(L_p,\bar L_p)=|\bar z_2|^2+|-\bar z_1|^2. \end{align*} Since $|\bar z_2|^2=|z_2|^2$ and $|-\bar z_1|^2=|z_1|^2$, we get \begin{align*} \mathcal L_\rho(L_p,\bar L_p)=|z_2|^2+|z_1|^2. \end{align*} Because $p\in S^3$, $|z_1|^2+|z_2|^2=1$, hence \begin{align*} \mathcal L_\rho(L_p,\bar L_p)=1. \end{align*} Thus $S^3=\partial B^2$ is strictly pseudoconvex, with pseudoconvex side $\rho<0$. The closure $\overline{B^2}$ is a compact complex manifold with smooth boundary $S^3$. The identity map $\partial\overline{B^2}\to S^3$ is a CR diffeomorphism because both sides carry the CR structure induced from the same hypersurface in $\mathbb C^2$. Therefore the standard three-sphere is both globally CR embedded and strictly pseudoconvex fillable; in this model case, the same ambient ball supplies the CR coordinates and the filling. [/example] The global embedding problem in real dimension $3$ is delicate because strict pseudoconvexity supplies a contact distribution but not enough global CR functions by itself. Small deformations of the standard sphere remain embeddable under additional hypotheses, while other strictly pseudoconvex CR structures fail to be fillable. [remark: Burns Epstein Invariant] For certain compact strictly pseudoconvex CR $3$-manifolds with a choice of pseudo-Einstein contact form, Burns and Epstein constructed a global numerical invariant from Chern-Simons type pseudohermitian data. In families of CR structures, this invariant can obstruct compatibility with a proposed strictly pseudoconvex filling. It is a global obstruction: it is invisible in the pointwise Levi form and depends on how the local CR geometry is assembled over all of $M$. [/remark] The lesson is that the Levi form starts the theory but does not end it. Two CR $3$-manifolds may look locally like strictly pseudoconvex hypersurfaces while having different global embedding or filling behaviour. ## The CR Yamabe Problem Given a strictly pseudoconvex CR manifold, can we choose the contact form so that its Webster scalar curvature is constant? This is the CR analogue of the Riemannian Yamabe problem, with the conformal class replaced by the pseudohermitian conformal class $\hat\theta=u^{2/n}\theta$ on a CR manifold of real dimension $2n+1$. [definition: Pseudohermitian Conformal Class] Let $(M^{2n+1},T^{1,0}M,\theta)$ be a strictly pseudoconvex pseudohermitian CR manifold. The pseudohermitian conformal class of $\theta$ is the collection of contact forms \begin{align*} [\theta]_{\mathrm{CR}}=\{u^{2/n}\theta: u\in C^\infty(M),\ u>0\}. \end{align*} [/definition] The exponent $2/n$ is chosen so that the Webster scalar curvature transformation law has the same critical Sobolev structure as the conformal Laplacian in Riemannian geometry. The relevant volume form is $\theta\wedge(d\theta)^n$. [definition: CR Yamabe Functional] Let $(M^{2n+1},T^{1,0}M,\theta)$ be compact and strictly pseudoconvex, let $R_\theta$ be the Webster scalar curvature, and let $\nabla_b$ be the horizontal gradient. Define the functional $Q_\theta:\{u\in C^\infty(M):u>0\}\to\mathbb R$ by \begin{align*} Q_\theta[u] &= \frac{\int_M\left(\frac{2(n+1)}{n}|\nabla_b u|^2+R_\theta u^2\right)\,\theta\wedge(d\theta)^n} {\left(\int_M u^{2+2/n}\,\theta\wedge(d\theta)^n\right)^{n/(n+1)}}. \end{align*} [/definition] The denominator is scale invariant, so minimising $Q_\theta$ is a constrained variational problem rather than an ordinary energy minimisation. The obstruction is that changing the contact form by a positive factor changes both the Webster scalar curvature and the pseudohermitian volume, so the critical point equation must balance the sublaplacian term against the nonlinear volume constraint. The variational theorem identifies that balance as the constant Webster scalar curvature equation. [quotetheorem:9235] [citeproof:9235] This theorem is the calculation behind the CR Yamabe problem: it converts the geometric search for a preferred contact form into a nonlinear subelliptic equation. The hypotheses matter because the functional is built from the pseudohermitian representative $\theta$, while the resulting equation must describe the whole conformal class of contact forms $u^{2/n}\theta$. The calculation alone does not guarantee a minimizer; it only identifies what a minimizer would satisfy. Existence is deeper because the exponent $2+2/n$ is critical for the Folland-Stein Sobolev embedding, so compactness can fail through concentration at points. The next quoted result addresses that compactness problem rather than merely differentiating the functional. [quotetheorem:9236] This existence theorem addresses the compactness gap left by the Euler-Lagrange calculation. In high enough dimension, and away from the spherical obstruction case, concentration can be controlled so that the CR Yamabe functional has a minimizer. The dimension hypothesis is part of the analytic mechanism rather than a cosmetic assumption, and the result does not prescribe a unique contact form. It says that the conformal class contains at least one pseudohermitian scale with constant Webster scalar curvature, linking the variational equation above to an actual geometric representative. [example: The CR Sphere As Yamabe Model] Let $S^{2n+1}\subset\mathbb C^{n+1}$ carry the standard pseudohermitian contact form \begin{align*} \theta_0=\frac{i}{2}\sum_{j=1}^{n+1}(\bar z_j\,dz_j-z_j\,d\bar z_j)\bigg|_{S^{2n+1}}. \end{align*} We compute its Webster scalar curvature and show that, in this normalization, it is the positive constant $n(n+1)$. First check that $\theta_0$ is invariant under the unitary action. Let $U=(U_{jk})\in U(n+1)$, so $(Uz)_j=\sum_{k=1}^{n+1}U_{jk}z_k$ and $d(Uz)_j=\sum_{\ell=1}^{n+1}U_{j\ell}\,dz_\ell$. Then \begin{align*} \sum_{j=1}^{n+1}\overline{(Uz)_j}\,d(Uz)_j=\sum_{j=1}^{n+1}\left(\sum_{k=1}^{n+1}\overline{U_{jk}}\,\bar z_k\right)\left(\sum_{\ell=1}^{n+1}U_{j\ell}\,dz_\ell\right). \end{align*} Expanding the finite product gives \begin{align*} \sum_{j=1}^{n+1}\overline{(Uz)_j}\,d(Uz)_j=\sum_{j,k,\ell}\overline{U_{jk}}U_{j\ell}\,\bar z_k\,dz_\ell. \end{align*} Regrouping by $k$ and $\ell$ gives \begin{align*} \sum_{j=1}^{n+1}\overline{(Uz)_j}\,d(Uz)_j=\sum_{k,\ell}\left(\sum_{j=1}^{n+1}\overline{U_{jk}}U_{j\ell}\right)\bar z_k\,dz_\ell. \end{align*} Since $U^*U=I$, the inner sum is $\delta_{k\ell}$, hence \begin{align*} \sum_{j=1}^{n+1}\overline{(Uz)_j}\,d(Uz)_j=\sum_{k,\ell}\delta_{k\ell}\bar z_k\,dz_\ell. \end{align*} Only the terms with $k=\ell$ remain, so \begin{align*} \sum_{j=1}^{n+1}\overline{(Uz)_j}\,d(Uz)_j=\sum_{k=1}^{n+1}\bar z_k\,dz_k. \end{align*} Taking conjugates in the same calculation gives \begin{align*} \sum_{j=1}^{n+1}(Uz)_j\,d\overline{(Uz)_j}=\sum_{k=1}^{n+1}z_k\,d\bar z_k. \end{align*} Substituting these two identities into the formula for $\theta_0$ gives $U^*\theta_0=\theta_0$. Since $U(n+1)$ acts transitively on $S^{2n+1}$ by extending any unit vector to a unitary basis, and Webster scalar curvature is preserved by pseudohermitian automorphisms, $R_{\theta_0}$ is constant on the sphere. It remains to identify the constant. In the affine chart $z_1\ne0$, write \begin{align*} w_\alpha=\frac{z_{\alpha+1}}{z_1},\qquad s=1+\sum_{\gamma=1}^n |w_\gamma|^2. \end{align*} For the Hopf pseudohermitian structure, the Tanaka-Webster Ricci tensor agrees with the Ricci tensor of the transverse Fubini-Study metric with Kähler potential $\log s$. Thus \begin{align*} h_{\alpha\bar\beta}=\frac{\partial^2}{\partial w_\alpha\,\partial\bar w_\beta}\log s. \end{align*} Since $\partial s/\partial\bar w_\beta=w_\beta$, the first derivative is \begin{align*} \frac{\partial}{\partial\bar w_\beta}\log s=\frac{w_\beta}{s}. \end{align*} Differentiating with respect to $w_\alpha$ gives \begin{align*} h_{\alpha\bar\beta}=\frac{\partial}{\partial w_\alpha}\left(\frac{w_\beta}{s}\right). \end{align*} Using $\partial w_\beta/\partial w_\alpha=\delta_{\alpha\beta}$ and $\partial s/\partial w_\alpha=\bar w_\alpha$, the quotient rule gives \begin{align*} h_{\alpha\bar\beta}=\frac{\delta_{\alpha\beta}s-w_\beta\bar w_\alpha}{s^2}. \end{align*} Equivalently, \begin{align*} h_{\alpha\bar\beta}=s^{-1}\left(\delta_{\alpha\beta}-s^{-1}\bar w_\alpha w_\beta\right). \end{align*} To compute the determinant, set $v=(\bar w_1,\ldots,\bar w_n)^\top$ and $u=(w_1,\ldots,w_n)^\top$. Then $(\bar w_\alpha w_\beta)_{\alpha,\beta}=vu^\top$, and factoring $s^{-1}$ from each of the $n$ rows gives \begin{align*} \det(h_{\alpha\bar\beta})=s^{-n}\det(I-s^{-1}vu^\top). \end{align*} If $x$ satisfies $u^\topx=0$, then $(s^{-1}vu^\top)x=s^{-1}v(u^\topx)=0$. Also \begin{align*} (s^{-1}vu^\top)v=s^{-1}v(u^\topv)=s^{-1}\left(\sum_{\alpha=1}^n|w_\alpha|^2\right)v. \end{align*} Therefore the eigenvalues of $I-s^{-1}vu^\top$ are $1$ with multiplicity $n-1$ and $1-s^{-1}\sum_{\alpha=1}^n|w_\alpha|^2$ in the remaining direction. Hence \begin{align*} \det(I-s^{-1}vu^\top)=1-\frac{1}{s}\sum_{\alpha=1}^n|w_\alpha|^2. \end{align*} Since $\sum_{\alpha=1}^n|w_\alpha|^2=s-1$, this becomes \begin{align*} \det(I-s^{-1}vu^\top)=1-\frac{s-1}{s}. \end{align*} The right side is \begin{align*} 1-\frac{s-1}{s}=\frac{s-(s-1)}{s}=\frac1s. \end{align*} Thus \begin{align*} \det(h_{\alpha\bar\beta})=s^{-n}s^{-1}=s^{-(n+1)}. \end{align*} The Ricci tensor of this transverse Kähler metric is \begin{align*} \operatorname{Ric}_{\alpha\bar\beta}=-\frac{\partial^2}{\partial w_\alpha\,\partial\bar w_\beta}\log\det(h_{\gamma\bar\delta}). \end{align*} Using $\det(h_{\gamma\bar\delta})=s^{-(n+1)}$, we have \begin{align*} \log\det(h_{\gamma\bar\delta})=-(n+1)\log s. \end{align*} Therefore \begin{align*} \operatorname{Ric}_{\alpha\bar\beta}=-\frac{\partial^2}{\partial w_\alpha\,\partial\bar w_\beta}\bigl(-(n+1)\log s\bigr). \end{align*} Pulling out the constant gives \begin{align*} \operatorname{Ric}_{\alpha\bar\beta}=(n+1)\frac{\partial^2}{\partial w_\alpha\,\partial\bar w_\beta}\log s. \end{align*} Since $h_{\alpha\bar\beta}=\partial_\alpha\partial_{\bar\beta}\log s$, this is \begin{align*} \operatorname{Ric}_{\alpha\bar\beta}=(n+1)h_{\alpha\bar\beta}. \end{align*} Taking the trace with the inverse matrix $h^{\alpha\bar\beta}$ gives \begin{align*} R_{\theta_0}=\sum_{\alpha,\beta=1}^n h^{\alpha\bar\beta}\operatorname{Ric}_{\alpha\bar\beta}. \end{align*} Substituting $\operatorname{Ric}_{\alpha\bar\beta}=(n+1)h_{\alpha\bar\beta}$ gives \begin{align*} R_{\theta_0}=(n+1)\sum_{\alpha,\beta=1}^n h^{\alpha\bar\beta}h_{\alpha\bar\beta}. \end{align*} Because $(h^{\alpha\bar\beta})$ is the inverse of $(h_{\alpha\bar\beta})$, the matrix with entries $\sum_\beta h^{\alpha\bar\beta}h_{\beta\bar\gamma}$ is the identity matrix. Its trace is $n$, so \begin{align*} \sum_{\alpha,\beta=1}^n h^{\alpha\bar\beta}h_{\alpha\bar\beta}=n. \end{align*} Hence \begin{align*} R_{\theta_0}=n(n+1)>0. \end{align*} For the CR Yamabe functional, the constant function $u\equiv1$ has horizontal gradient $0$: \begin{align*} \nabla_b1=0. \end{align*} The sub-Laplacian of a constant is also $0$, so \begin{align*} \Delta_b1=0. \end{align*} Substituting $u=1$ and $R_{\theta_0}=n(n+1)$ into the [CR Yamabe Euler-Lagrange equation](/theorems/9235) gives \begin{align*} -\frac{2(n+1)}{n}\Delta_b1+R_{\theta_0}\cdot1=-\frac{2(n+1)}{n}\cdot0+n(n+1)=n(n+1). \end{align*} The right side of the same equation is \begin{align*} \lambda\cdot1^{1+2/n}=\lambda. \end{align*} Thus $\lambda=n(n+1)$, and $\theta_0$ is already a constant Webster scalar curvature representative of its pseudohermitian conformal class. The Cayley transform identifies $S^{2n+1}$ minus one point with the Heisenberg group $\mathbb H^n$. With $r^2=|\zeta|^2$ and $D=1+r^2+it$, use \begin{align*} C(\zeta,t)=\left(\frac{2\zeta}{D},\frac{1-r^2-it}{D}\right). \end{align*} To verify that $C(\zeta,t)$ lies on the sphere, first compute \begin{align*} |D|^2=(1+r^2+it)(1+r^2-it)=(1+r^2)^2+t^2. \end{align*} The squared norm of the first component is \begin{align*} \left|\frac{2\zeta}{D}\right|^2=\frac{|2\zeta|^2}{|D|^2}=\frac{4r^2}{(1+r^2)^2+t^2}. \end{align*} The squared norm of the second component is \begin{align*} \left|\frac{1-r^2-it}{D}\right|^2=\frac{(1-r^2-it)(1-r^2+it)}{|D|^2}=\frac{(1-r^2)^2+t^2}{(1+r^2)^2+t^2}. \end{align*} Adding them gives \begin{align*} \left|\frac{2\zeta}{D}\right|^2+\left|\frac{1-r^2-it}{D}\right|^2=\frac{4r^2+(1-r^2)^2+t^2}{(1+r^2)^2+t^2}. \end{align*} Expanding the numerator gives \begin{align*} 4r^2+(1-r^2)^2+t^2=4r^2+1-2r^2+r^4+t^2. \end{align*} Combining like terms gives \begin{align*} 4r^2+1-2r^2+r^4+t^2=1+2r^2+r^4+t^2. \end{align*} Since \begin{align*} (1+r^2)^2+t^2=1+2r^2+r^4+t^2, \end{align*} the numerator equals the denominator, and therefore \begin{align*} \left|\frac{2\zeta}{D}\right|^2+\left|\frac{1-r^2-it}{D}\right|^2=1. \end{align*} In this normalization the pulled-back contact form is \begin{align*} C^*\theta_0=\frac{2}{(1+|\zeta|^2)^2+t^2}\,\theta_{\mathbb H}. \end{align*} The pseudohermitian conformal change has the form $\hat\theta=u^{2/n}\theta$, so the conformal factor corresponding to this identity is determined by \begin{align*} u_0(\zeta,t)^{2/n}=\frac{2}{(1+|\zeta|^2)^2+t^2}. \end{align*} Raising both sides to the power $n/2$ gives \begin{align*} u_0(\zeta,t)=\left(\frac{2}{(1+|\zeta|^2)^2+t^2}\right)^{n/2}. \end{align*} Thus the positive-curvature sphere and the flat Heisenberg group are two conformal pictures of the same sharp CR Yamabe model. In real dimension $3$, where $n=1$, the formula gives \begin{align*} R_{\theta_0}=1\cdot2=2 \end{align*} on $S^3\subset\mathbb C^2$. [/example] The analogy with the Riemannian Yamabe problem is strong but not literal. The sub-Laplacian is not elliptic, the natural dimension in Sobolev estimates is the homogeneous dimension $2n+2$, and the automorphism group of the CR sphere creates the same concentration phenomena that [conformal maps](/page/Conformal%20Maps) create in Riemannian geometry. ## Paneitz Operators and Q-Curvature After prescribing scalar curvature, which higher-order CR invariants control changes of pseudohermitian scale? The answer parallels conformal geometry: fourth-order and critical-order operators encode curvature quantities whose transformation laws are rigid enough to produce global invariants. [definition: CR Paneitz Operator] Let $(M^3,T^{1,0}M,\theta)$ be a strictly pseudoconvex pseudohermitian CR manifold, let $T$ be the Reeb vector field, let $\Delta_b$ be the sub-Laplacian, and let $A_{\alpha\beta}$ be the pseudohermitian torsion. In the convention used in this course, the CR Paneitz operator is the differential operator $P_4:C^\infty(M)\to C^\infty(M)$ defined by \begin{align*} P_4 f=\Delta_b^2 f+T^2 f-4\,\operatorname{Im}\left(\nabla^\alpha(A_{\alpha\beta}\nabla^\beta f)\right). \end{align*} [/definition] The operator $P_4$ is built to be CR covariant rather than merely pseudohermitian. Its lower-order torsion correction is what allows it to transform naturally when the contact form is rescaled. [definition: CR Q Curvature] Let $(M^{2n+1},T^{1,0}M,\theta)$ be a strictly pseudoconvex pseudohermitian CR manifold. A CR Q-curvature is a scalar pseudohermitian invariant $Q_\theta$ associated to a critical CR covariant operator $P_\theta$ such that, for $\hat\theta=e^\Upsilon\theta$, \begin{align*} e^{(n+1)\Upsilon}Q_{\hat\theta}=Q_\theta+P_\theta\Upsilon. \end{align*} [/definition] The transformation law suggests a global invariant, but there is a possible obstruction: after rescaling the contact form, the added term $P_\theta\Upsilon$ could change the total integral. On a compact manifold this term disappears only when the critical operator has the right formal self-adjointness and no boundary contribution is present. The invariant theorem isolates exactly the hypotheses under which the total Q-curvature is independent of the pseudohermitian representative. [quotetheorem:9237] [citeproof:9237] The construction of CR Q-curvature through the Fefferman ambient metric and its refinements is due to work of Graham-Lee, Hirachi, and others. For these notes, the key point is its role as a bridge: it connects pseudohermitian curvature, invariant differential operators, and the logarithmic terms in kernel asymptotics. [example: The Siegel Boundary Has Vanishing Webster Curvature] Let \begin{align*} \mathcal U^{n+1}=\{(z,w)\in\mathbb C^n\times\mathbb C:\operatorname{Im}w>|z|^2\}. \end{align*} On $\partial\mathcal U^{n+1}$ write $w=t+i|z|^2$ and use coordinates $(z,t)\in\mathbb C^n\times\mathbb R$. For the standard Heisenberg pseudohermitian form, take \begin{align*} \theta=\frac12\left(dt+i\sum_{m=1}^n(z_m\,d\bar z_m-\bar z_m\,dz_m)\right). \end{align*} Set \begin{align*} Z_j=\frac{\partial}{\partial z_j}+i\bar z_j\frac{\partial}{\partial t},\qquad \bar Z_j=\frac{\partial}{\partial\bar z_j}-iz_j\frac{\partial}{\partial t}. \end{align*} We first identify the contact distribution. Since $dt(Z_j)=i\bar z_j$, $d\bar z_m(Z_j)=0$, and $dz_m(Z_j)=\delta_{mj}$, substitution into $\theta$ gives \begin{align*} \theta(Z_j)=\frac12\left(i\bar z_j+i\sum_{m=1}^n(z_m\cdot0-\bar z_m\delta_{mj})\right). \end{align*} The sum has only the $m=j$ term: \begin{align*} \sum_{m=1}^n(z_m\cdot0-\bar z_m\delta_{mj})=-\bar z_j. \end{align*} Hence \begin{align*} \theta(Z_j)=\frac12(i\bar z_j-i\bar z_j)=0. \end{align*} Similarly, $dt(\bar Z_j)=-iz_j$, $d\bar z_m(\bar Z_j)=\delta_{mj}$, and $dz_m(\bar Z_j)=0$, so \begin{align*} \theta(\bar Z_j)=\frac12\left(-iz_j+i\sum_{m=1}^n(z_m\delta_{mj}-\bar z_m\cdot0)\right). \end{align*} Again only the $m=j$ term remains: \begin{align*} \sum_{m=1}^n(z_m\delta_{mj}-\bar z_m\cdot0)=z_j. \end{align*} Therefore \begin{align*} \theta(\bar Z_j)=\frac12(-iz_j+iz_j)=0. \end{align*} Thus the $Z_j$ span the $(1,0)$ contact distribution and the $\bar Z_j$ span the $(0,1)$ contact distribution. Next compute the Levi form. Since $d(dt)=0$, $d(z_m\,d\bar z_m)=dz_m\wedge d\bar z_m$, and $d(\bar z_m\,dz_m)=d\bar z_m\wedge dz_m$, we get \begin{align*} d\theta=\frac{i}{2}\sum_{m=1}^n(dz_m\wedge d\bar z_m-d\bar z_m\wedge dz_m). \end{align*} Using $d\bar z_m\wedge dz_m=-dz_m\wedge d\bar z_m$ gives \begin{align*} d\theta=i\sum_{m=1}^n dz_m\wedge d\bar z_m. \end{align*} For $Z_j$ and $\bar Z_k$, \begin{align*} (dz_m\wedge d\bar z_m)(Z_j,\bar Z_k)=dz_m(Z_j)d\bar z_m(\bar Z_k)-dz_m(\bar Z_k)d\bar z_m(Z_j). \end{align*} Substituting $dz_m(Z_j)=\delta_{mj}$, $d\bar z_m(\bar Z_k)=\delta_{mk}$, $dz_m(\bar Z_k)=0$, and $d\bar z_m(Z_j)=0$ gives \begin{align*} (dz_m\wedge d\bar z_m)(Z_j,\bar Z_k)=\delta_{mj}\delta_{mk}. \end{align*} Therefore \begin{align*} d\theta(Z_j,\bar Z_k)=i\sum_{m=1}^n\delta_{mj}\delta_{mk}. \end{align*} The sum equals $\delta_{jk}$ because it is $1$ when $j=k$ and $0$ otherwise, so \begin{align*} d\theta(Z_j,\bar Z_k)=i\delta_{jk}. \end{align*} With the convention $h_{j\bar k}=-i\,d\theta(Z_j,\bar Z_k)$, this gives \begin{align*} h_{j\bar k}=-i(i\delta_{jk})=\delta_{jk}. \end{align*} Hence, for $V=\sum_{j=1}^n a_jZ_j$, \begin{align*} \mathcal L_\theta(V,\bar V)=\sum_{j,k=1}^n h_{j\bar k}a_j\bar a_k. \end{align*} Substituting $h_{j\bar k}=\delta_{jk}$ gives \begin{align*} \mathcal L_\theta(V,\bar V)=\sum_{j,k=1}^n\delta_{jk}a_j\bar a_k. \end{align*} Only the terms with $j=k$ remain, so \begin{align*} \mathcal L_\theta(V,\bar V)=\sum_{j=1}^n |a_j|^2. \end{align*} Thus the Levi form is exactly the standard Hermitian form on $\mathbb C^n$. The Reeb field is $T=2\partial_t$. Indeed, \begin{align*} \theta(2\partial_t)=\frac12\,dt(2\partial_t)=1, \end{align*} and $d\theta=i\sum_m dz_m\wedge d\bar z_m$ has no $dt$ term, so \begin{align*} d\theta(2\partial_t,\cdot)=0. \end{align*} The commutator $[Z_j,\bar Z_k]$ has only a $\partial_t$ component, since the coefficients of $\partial_{z_j}$ and $\partial_{\bar z_k}$ are constant. Its $\partial_t$ coefficient is \begin{align*} \frac{\partial}{\partial z_j}(-iz_k)-\frac{\partial}{\partial\bar z_k}(i\bar z_j). \end{align*} Since $\partial z_k/\partial z_j=\delta_{jk}$ and $\partial\bar z_j/\partial\bar z_k=\delta_{jk}$, this coefficient is \begin{align*} -i\delta_{jk}-i\delta_{jk}=-2i\delta_{jk}. \end{align*} Therefore \begin{align*} [Z_j,\bar Z_k]=-2i\delta_{jk}\frac{\partial}{\partial t}. \end{align*} Because $T=2\partial_t$, this is \begin{align*} [Z_j,\bar Z_k]=-i\delta_{jk}T. \end{align*} For $[Z_j,Z_k]$, the $\partial_t$ coefficient is \begin{align*} \frac{\partial}{\partial z_j}(i\bar z_k)-\frac{\partial}{\partial z_k}(i\bar z_j)=0-0=0. \end{align*} There are no $\partial_{z_\ell}$ terms because the coefficients of $\partial_{z_j}$ and $\partial_{z_k}$ are constant, so \begin{align*} [Z_j,Z_k]=0. \end{align*} Also the coefficients of $Z_j$ are independent of $t$, hence \begin{align*} [T,Z_j]=0. \end{align*} Let $\theta^j=dz_j$. Then \begin{align*} \theta^j(Z_k)=dz_j\left(\frac{\partial}{\partial z_k}+i\bar z_k\frac{\partial}{\partial t}\right)=\delta_{jk}. \end{align*} Also \begin{align*} \theta^j(\bar Z_k)=dz_j\left(\frac{\partial}{\partial\bar z_k}-iz_k\frac{\partial}{\partial t}\right)=0. \end{align*} And \begin{align*} \theta^j(T)=dz_j(2\partial_t)=0. \end{align*} Since $d^2=0$, \begin{align*} d\theta^j=d(dz_j)=0. \end{align*} The Tanaka-Webster structure equations are \begin{align*} d\theta^\beta=\sum_{\alpha=1}^n\theta^\alpha\wedge\omega_\alpha{}^\beta+\theta\wedge\tau^\beta. \end{align*} Taking \begin{align*} \omega_\alpha{}^\beta=0,\qquad \tau^\beta=0 \end{align*} makes the right-hand side equal to \begin{align*} \sum_{\alpha=1}^n\theta^\alpha\wedge0+\theta\wedge0=0. \end{align*} This agrees with $d\theta^\beta=0$. The Levi matrix $h_{\alpha\bar\beta}=\delta_{\alpha\beta}$ is constant, so these same zero connection forms also preserve the Levi matrix. Thus the Tanaka-Webster connection forms and torsion forms in this coframe are \begin{align*} \omega_\alpha{}^\beta=0,\qquad \tau^\beta=0. \end{align*} The curvature forms are \begin{align*} \Omega_\alpha{}^\beta=d\omega_\alpha{}^\beta-\sum_{\gamma=1}^n\omega_\alpha{}^\gamma\wedge\omega_\gamma{}^\beta. \end{align*} Substituting $\omega_\alpha{}^\beta=0$ gives \begin{align*} \Omega_\alpha{}^\beta=d0-\sum_{\gamma=1}^n0\wedge0=0. \end{align*} Thus every Webster Ricci component is zero: \begin{align*} R_{\alpha\bar\beta}=0. \end{align*} Since $h_{\alpha\bar\beta}=\delta_{\alpha\beta}$, its inverse is $h^{\alpha\bar\beta}=\delta_{\alpha\beta}$. Therefore the Webster scalar curvature is \begin{align*} R_\theta=\sum_{\alpha,\beta=1}^n h^{\alpha\bar\beta}R_{\alpha\bar\beta}. \end{align*} Substituting $R_{\alpha\bar\beta}=0$ gives \begin{align*} R_\theta=\sum_{\alpha,\beta=1}^n h^{\alpha\bar\beta}\cdot0=0. \end{align*} So the Siegel boundary has the standard positive Levi form but vanishing Webster scalar curvature; it is the flat pseudohermitian model corresponding, under the Cayley transform, to the positively curved CR sphere. [/example] Flatness here is pseudohermitian, not Euclidean. The boundary is curved as a real hypersurface in $\mathbb C^{n+1}$, but its left-invariant CR geometry is the local model for strongly pseudoconvex boundaries after scaling. ## Boundary Regularity and Kernel Invariants How much of a biholomorphism between domains is controlled by the CR geometry of the boundary? For strictly pseudoconvex domains with smooth boundary, the answer is remarkably strong: the holomorphic map cannot remain only an interior object. [quotetheorem:3711] [citeproof:3711] This result is quoted as a major theorem from several complex variables. Its hypotheses combine pseudoconvexity with Condition R, the global regularity of the Bergman projection, so the regularity input is stated analytically rather than as strict pseudoconvexity. The assumption $n\ge2$ is where genuine CR boundary geometry enters; in one complex variable, boundary extension is governed by conformal mapping theory rather than a hypersurface CR structure. Strictly pseudoconvex domains form an important neighboring class where strong boundary geometry often supplies the regularity estimates used later in the chapter, but the Bell-Ligocka statement above is organized around Condition R. The $C^\infty$ boundary hypothesis matches the conclusion, since the kernel regularity and $\bar\partial$-Neumann estimates must be differentiated to all orders; finite boundary smoothness generally yields only finite regularity and may lose derivatives. The next object is the Bergman kernel itself. Bell-Ligocka regularity uses the kernel through the Bergman projection, while the following Fefferman expansion studies the kernel's boundary singularity directly. Defining the kernel now lets us pass from smooth extension of biholomorphisms to the finer question of which boundary invariants are encoded in that singularity. [definition: Bergman Kernel] Let $\Omega\subset\mathbb C^n$ be a bounded domain, and set \begin{align*} A^2(\Omega)=\{f:\Omega\to\mathbb C: f\text{ is holomorphic and }\int_\Omega |f|^2\,d\mathcal L^{2n}<\infty\}. \end{align*} The Bergman kernel is the function $K_\Omega:\Omega\times\Omega\to\mathbb C$, written $K_\Omega(z,\bar w)$, which is the reproducing kernel for the Hilbert space $A^2(\Omega)$ with respect to Lebesgue measure. [/definition] The Bergman kernel is invariantly attached to the complex domain, but its boundary behaviour is singular rather than smooth. The problem is to separate the universal blow-up forced by strict pseudoconvexity from the lower-order terms that remember the CR geometry of the boundary. Fefferman's expansion gives that separation, including the logarithmic term where the first obstruction-type boundary invariant appears. [quotetheorem:9238] The expansion is stated here without proof. Strict pseudoconvexity is what forces the universal leading pole $(-\rho)^{-(n+1)}$; for weakly pseudoconvex domains the Levi form degenerates, and the boundary asymptotics can involve different weights, finite-type data, or fail to have this smooth pole-plus-log form. The condition $\phi|_{\partial\Omega}\ne0$ rules out cancellation of the leading singularity, so the kernel really blows up at the strictly pseudoconvex rate near every boundary point. The coefficient $\psi$ is the obstruction-theoretic part of the expansion: after normalising the defining function, its boundary value is a CR invariant, and vanishing of the log term is the model behaviour seen on the ball and spherical CR boundaries. [example: Boundary Invariants From The Ball] Let $B^n=\{z\in\mathbb C^n:|z|^2<1\}$ and $\rho(z)=|z|^2-1$. For a multi-index $\alpha=(\alpha_1,\ldots,\alpha_n)$, write $|\alpha|=\alpha_1+\cdots+\alpha_n$, $\alpha!=\alpha_1!\cdots\alpha_n!$, and $z^\alpha=z_1^{\alpha_1}\cdots z_n^{\alpha_n}$. We compute the diagonal Bergman kernel of $B^n$ and compare it with the pole-plus-log boundary expansion. First compute the $L^2$ norm of $z^\alpha$. In polar coordinates $z=r\xi$, with $0\le r<1$ and $\xi\in S^{2n-1}$, the Euclidean volume form is $dV=r^{2n-1}\,dr\,d\sigma(\xi)$ and $|z^\alpha|^2=r^{2|\alpha|}|\xi^\alpha|^2$. Therefore \begin{align*} \int_{B^n}|z^\alpha|^2\,dV=\int_0^1 r^{2|\alpha|+2n-1}\,dr\int_{S^{2n-1}}|\xi^\alpha|^2\,d\sigma(\xi). \end{align*} The radial factor is \begin{align*} \int_0^1 r^{2|\alpha|+2n-1}\,dr=\left[\frac{r^{2|\alpha|+2n}}{2|\alpha|+2n}\right]_0^1=\frac{1}{2(|\alpha|+n)}. \end{align*} For the spherical factor, compare the same Gaussian integral in rectangular and polar coordinates. On one hand, \begin{align*} \int_{\mathbb C^n}e^{-|z|^2}|z^\alpha|^2\,dV=\prod_{j=1}^n\int_{\mathbb C}e^{-|z_j|^2}|z_j|^{2\alpha_j}\,dA(z_j). \end{align*} In one complex variable, $z_j=re^{i\theta}$ gives \begin{align*} \int_{\mathbb C}e^{-|z_j|^2}|z_j|^{2\alpha_j}\,dA=2\pi\int_0^\infty e^{-r^2}r^{2\alpha_j+1}\,dr. \end{align*} With $u=r^2$, so $du=2r\,dr$, this becomes \begin{align*} 2\pi\int_0^\infty e^{-r^2}r^{2\alpha_j+1}\,dr=\pi\int_0^\infty e^{-u}u^{\alpha_j}\,du=\pi\alpha_j!. \end{align*} Thus \begin{align*} \int_{\mathbb C^n}e^{-|z|^2}|z^\alpha|^2\,dV=\pi^n\alpha!. \end{align*} On the other hand, polar coordinates give \begin{align*} \int_{\mathbb C^n}e^{-|z|^2}|z^\alpha|^2\,dV=\int_0^\infty e^{-r^2}r^{2|\alpha|+2n-1}\,dr\int_{S^{2n-1}}|\xi^\alpha|^2\,d\sigma(\xi). \end{align*} Again putting $u=r^2$ gives \begin{align*} \int_0^\infty e^{-r^2}r^{2|\alpha|+2n-1}\,dr=\frac12\int_0^\infty e^{-u}u^{|\alpha|+n-1}\,du=\frac12(|\alpha|+n-1)!. \end{align*} Therefore \begin{align*} \pi^n\alpha!=\frac12(|\alpha|+n-1)!\int_{S^{2n-1}}|\xi^\alpha|^2\,d\sigma(\xi). \end{align*} Solving for the spherical integral gives \begin{align*} \int_{S^{2n-1}}|\xi^\alpha|^2\,d\sigma(\xi)=\frac{2\pi^n\alpha!}{(|\alpha|+n-1)!}. \end{align*} Multiplying the radial and spherical factors, \begin{align*} \int_{B^n}|z^\alpha|^2\,dV=\frac{1}{2(|\alpha|+n)}\cdot\frac{2\pi^n\alpha!}{(|\alpha|+n-1)!}. \end{align*} Since $(|\alpha|+n)(|\alpha|+n-1)!=(|\alpha|+n)!$, this is \begin{align*} \int_{B^n}|z^\alpha|^2\,dV=\frac{\pi^n\alpha!}{(|\alpha|+n)!}. \end{align*} If $\alpha\ne\beta$, choose $j$ with $\alpha_j\ne\beta_j$. Let \begin{align*} I_{\alpha\beta}=\int_{B^n}z^\alpha\overline{z^\beta}\,dV. \end{align*} Rotating only the $j$th coordinate by $e^{i\theta}$ preserves $B^n$ and $dV$, while the integrand is multiplied by $e^{i(\alpha_j-\beta_j)\theta}$. Hence \begin{align*} I_{\alpha\beta}=e^{i(\alpha_j-\beta_j)\theta}I_{\alpha\beta}. \end{align*} Choose $\theta$ so that $e^{i(\alpha_j-\beta_j)\theta}\ne1$. Then \begin{align*} \left(1-e^{i(\alpha_j-\beta_j)\theta}\right)I_{\alpha\beta}=0. \end{align*} The scalar factor is nonzero, so $I_{\alpha\beta}=0$. Therefore the normalized monomials \begin{align*} e_\alpha(z)=\left(\frac{(|\alpha|+n)!}{\pi^n\alpha!}\right)^{1/2}z^\alpha \end{align*} are orthonormal. Since holomorphic polynomials are dense in $A^2(B^n)$ and every $A^2$ holomorphic function has a Taylor expansion, these monomials form an [orthonormal basis](/page/Orthonormal%20Basis). The reproducing kernel of a Hilbert space with orthonormal basis $\{e_\alpha\}$ is $\sum_\alpha e_\alpha(z)\overline{e_\alpha(w)}$, so \begin{align*} K_{B^n}(z,\bar w)=\sum_{\alpha\in\mathbb N^n}e_\alpha(z)\overline{e_\alpha(w)}. \end{align*} Substituting the formula for $e_\alpha$ gives \begin{align*} K_{B^n}(z,\bar w)=\frac{1}{\pi^n}\sum_{\alpha\in\mathbb N^n}\frac{(|\alpha|+n)!}{\alpha!}z^\alpha\bar w^\alpha. \end{align*} Group the terms by $m=|\alpha|$. For such $\alpha$, \begin{align*} \frac{(m+n)!}{\alpha!}=n!\binom{n+m}{m}\frac{m!}{\alpha!}. \end{align*} Indeed, \begin{align*} n!\binom{n+m}{m}\frac{m!}{\alpha!}=n!\frac{(n+m)!}{n!\,m!}\frac{m!}{\alpha!}=\frac{(n+m)!}{\alpha!}. \end{align*} Thus \begin{align*} K_{B^n}(z,\bar w)=\frac{n!}{\pi^n}\sum_{m=0}^{\infty}\binom{n+m}{m}\sum_{|\alpha|=m}\frac{m!}{\alpha!}z^\alpha\bar w^\alpha. \end{align*} By the multinomial theorem, \begin{align*} \sum_{|\alpha|=m}\frac{m!}{\alpha!}z^\alpha\bar w^\alpha=\left(\sum_{j=1}^n z_j\bar w_j\right)^m. \end{align*} Therefore \begin{align*} K_{B^n}(z,\bar w)=\frac{n!}{\pi^n}\sum_{m=0}^{\infty}\binom{n+m}{m}\left(\sum_{j=1}^n z_j\bar w_j\right)^m. \end{align*} For $z,w\in B^n$, Cauchy-Schwarz gives $|\sum_j z_j\bar w_j|\le |z|\,|w|<1$, so the binomial series applies: \begin{align*} (1-x)^{-(n+1)}=\sum_{m=0}^{\infty}\binom{n+m}{m}x^m. \end{align*} Taking $x=\sum_{j=1}^n z_j\bar w_j$ gives \begin{align*} K_{B^n}(z,\bar w)=\frac{n!}{\pi^n}\frac{1}{\left(1-\sum_{j=1}^n z_j\bar w_j\right)^{n+1}}. \end{align*} On the diagonal, $\sum_{j=1}^n z_j\bar z_j=|z|^2$, hence \begin{align*} K_{B^n}(z,\bar z)=\frac{n!}{\pi^n}\frac{1}{(1-|z|^2)^{n+1}}. \end{align*} Since $\rho(z)=|z|^2-1$, we have $-\rho(z)=1-|z|^2$, and therefore \begin{align*} K_{B^n}(z,\bar z)=\frac{n!}{\pi^n}\frac{1}{(-\rho(z))^{n+1}}. \end{align*} Comparing this identity with \begin{align*} K_{B^n}(z,\bar z)=\frac{\phi(z)}{(-\rho(z))^{n+1}}+\psi(z)\log(-\rho(z)), \end{align*} we may take \begin{align*} \phi(z)\equiv\frac{n!}{\pi^n}. \end{align*} There is no logarithmic summand in the explicit formula, so \begin{align*} \psi(z)\equiv0. \end{align*} Thus the diagonal Bergman kernel of the ball has exactly the leading strictly pseudoconvex pole and no logarithmic term. This reflects the model geometry of the boundary sphere: it is the homogeneous flat model for Chern-Moser curvature, even though its standard pseudohermitian scale has positive Webster scalar curvature. [/example] The regularity theorem and the kernel expansion explain why CR equivalence is the right boundary notion for biholomorphic equivalence. A biholomorphism identifies Bergman kernels, the kernel expansion identifies boundary invariants, and Bell-Ligocka boundary regularity upgrades the interior map to a smooth CR map on the boundary. ## Further Directions From CR Geometry What should be taken from this course into later work in analysis and geometry? The common pattern is that analytic estimates produce CR functions and boundary regularity, while pseudohermitian geometry packages the quantities preserved by those analytic processes. A practical workflow is to match the question to its invariant. To test a candidate CR $3$-manifold for fillability, first look for enough global CR functions to embed it, then compare secondary invariants such as the Burns-Epstein invariant with the proposed filling. To prescribe Webster scalar curvature, choose a pseudohermitian scale, compute $R_\theta$, and study the CR Yamabe functional in that scale. To extract CR invariants from a strictly pseudoconvex domain, compute a defining function, examine the Bergman kernel expansion, and read the logarithmic coefficient $\psi$ as the obstruction-sensitive term. [explanation: Three Global Lessons] First, local strict pseudoconvexity does not settle global existence questions. Embedding and fillability depend on global CR functions, topology, and secondary invariants such as the Burns-Epstein invariant. Second, CR conformal geometry has its own curvature prescription problems. The CR Yamabe problem, Paneitz operators, and Q-curvature parallel Riemannian conformal geometry but must use the subelliptic calculus of the contact distribution. Third, several complex variables recovers boundary geometry from analytic kernels. Bell-Ligocka boundary regularity and the Bergman kernel expansion show that smooth boundaries with the relevant regularity hypotheses are not passive edges of domains; they carry the invariants that control biholomorphic maps. [/explanation] These directions lead naturally to microlocal analysis of the $\bar\partial_b$ complex, CR pluripotential theory, ACH Einstein metrics, and modern work on non-embeddable or non-fillable CR manifolds. The boundary remains the meeting point: it is where holomorphic extension, PDE estimates, contact geometry, and global invariants all become part of the same problem. ## Beyond and Connected Topics This note connects most directly to [Several Complex Variables IV: Complex Geometry and Curvature](/page/Several%20Complex%20Variables%20IV%3A%20Complex%20Geometry%20and%20Curvature), where CR hypersurfaces appear as real boundaries inside holomorphic geometry, and to [Androma Several Complex Variables I: Domains and Holomorphy](/page/Androma%20Several%20Complex%20Variables%20I%3A%20Domains%20and%20Holomorphy), where pseudoconvexity and domains of holomorphy first organize the subject. The analytic estimates in Chapter 8 also point toward [Elliptic Operator](/page/Elliptic%20Operator) and [Sobolev Space](/page/Sobolev%20Space), because boundary regularity is measured by derivative gain rather than by formal solvability alone. On the geometric side, the Levi form links this note to contact and symplectic methods, while CR automorphisms and normal forms connect it to local equivalence problems and the structure theory reflected in [Structural Isomorphisms of Compact Lie Groups](/theorems/2480). Readers who want the cohomological viewpoint should next compare the tangential $\bar\partial_b$ complex with de Rham cohomology through [Differential Forms and de Rham Cohomology](/page/Differential%20Forms%20and%20de%20Rham%20Cohomology), and with sheaf-theoretic methods in [Several Complex Variables II: Sheaves and Stein Theory](/page/Several%20Complex%20Variables%20II%3A%20Sheaves%20and%20Stein%20Theory). ## References - Androma, [Androma Several Complex Variables I: Domains and Holomorphy](/page/Androma%20Several%20Complex%20Variables%20I%3A%20Domains%20and%20Holomorphy). - Androma, [Several Complex Variables II: Sheaves and Stein Theory](/page/Several%20Complex%20Variables%20II%3A%20Sheaves%20and%20Stein%20Theory). - Androma, [Several Complex Variables III: L² Methods and Applications](/page/Several%20Complex%20Variables%20III%3A%20L%C2%B2%20Methods%20and%20Applications). - Androma, [Several Complex Variables IV: Complex Geometry and Curvature](/page/Several%20Complex%20Variables%20IV%3A%20Complex%20Geometry%20and%20Curvature). - Androma, [Differential Forms and de Rham Cohomology](/page/Differential%20Forms%20and%20de%20Rham%20Cohomology). - Androma, [Sobolev Space](/page/Sobolev%20Space). - Androma, [Elliptic Operator](/page/Elliptic%20Operator). - Hörmander, Lars. *An Introduction to Complex Analysis in Several Variables*. North-Holland, 1990. - Folland, G. B., and Kohn, J. J. *The Neumann Problem for the Cauchy-Riemann Complex*. Princeton University Press, 1972. - Range, R. Michael. *Holomorphic Functions and Integral Representations in Several Complex Variables*. Springer, 1986. - Krantz, Steven G. *Function Theory of Several Complex Variables*. AMS Chelsea Publishing, 2001. - Chen, So-Chin, and Shaw, Mei-Chi. *Partial Differential Equations in Several Complex Variables*. AMS/IP, 2001.