This course explores the profound interplay between analysis and complex geometry through the lens of L² methods applied to the ∂̄-equation in several complex variables. The central narrative revolves around how estimates in weighted L² spaces can guarantee the existence and regularity of solutions to fundamental differential equations in complex geometry, and how these analytic techniques reveal deep geometric and algebraic structure. Rather than treating complex analysis as an abstract subject divorced from applications, the course demonstrates how solving the ∂̄-equation under carefully chosen metrics yields concrete results about extension theorems, approximation, and the geometry of singularities. The first half of the course establishes the conceptual and technical foundations. Chapters 1–2 review the ∂̄-equation and introduce plurisubharmonic functions, which serve as the natural "weight" in L² estimates—they encode geometric rigidity and control solutions at the boundary and at infinity. Chapters 3–4 build toward Hörmander's L² existence theorem, the crown jewel of the analytic approach, which asserts that L² bounds on the ∂̄-operator guarantee solvability. The course then pivots to applications: the Ohsawa–Takegoshi [extension theorem](/theorems/59), Bergman spaces and reproducing kernels, and approximation theory in the second half (Chapters 5–7). The final chapters unify these strands through higher-order structures. Multiplier ideal sheaves (Chapter 8) encode the locus where L² control breaks down, leading to the [Nadel vanishing theorem](/theorems/3718) and refined extension results. Skoda's division theorem (Chapter 9) and the Donnelly–Fefferman theorem (Chapter 10) show how L² methods interact with singular metrics and analytic sets, pushing the theory toward modern algebraic geometry. By Chapter 11, students have seen how weighted L² estimates become a universal tool linking Kähler geometry, algebraic geometry, and complex analysis, with applications extending to multiplicity estimates, effectivity questions, and the geometry of exceptional sets. # Introduction This course develops the $L^2$ method for the $\bar{\partial}$-operator in several complex variables. The guiding theme is that analytic estimates for differential forms can replace more indirect sheaf-theoretic arguments, while also producing quantitative bounds on the solutions. The main technical result is the $L^2$ existence theorem of Hörmander, and the main applications are extension, positivity, and vanishing theorems that shape modern complex analysis and complex geometry. The course assumes a first course in one complex variable, real analysis at the level of $L^p$ spaces and weak compactness, and the basic language of Hilbert spaces. Differential forms on smooth manifolds will be used, but the needed conventions will be recalled as they appear. Basic elliptic PDE is helpful for intuition, especially when comparing the complex Laplacian with the real Laplacian, but the notes keep the analytic estimates self-contained. ## The Central Problem What does it mean to solve the Cauchy-Riemann equations in several complex variables when there is no single-variable integral formula available on a general domain? In one complex variable, the [Cauchy integral formula](/theorems/345) and residues give direct control over holomorphic functions. In several variables, the equation $\bar{\partial}u=f$ becomes a system on differential forms, and the geometry of the domain begins to control whether solutions with estimates exist. The first transition in the course is from holomorphic functions to the Dolbeault complex. If $\Omega \subset \mathbb C^n$ is a domain, a smooth function $u: \Omega \to \mathbb C$ is holomorphic precisely when $\bar{\partial}u=0$. More generally, $\bar{\partial}$ maps $(p,q)$-forms to $(p,q+1)$-forms, and the equation \begin{align*} \bar{\partial}u=f \end{align*} is meaningful for forms as soon as the compatibility condition $\bar{\partial}f=0$ holds. The difficulty is not only algebraic compatibility. On domains with poor geometry, the condition $\bar{\partial}f=0$ does not by itself provide useful norm control on a solution. The course therefore asks for estimates of the form \begin{align*} \int_\Omega |u|^2 e^{-\varphi}\,d\mathcal L^{2n} \le C \int_\Omega |f|^2 e^{-\varphi}\,d\mathcal L^{2n}, \end{align*} where $\varphi: \Omega \to \mathbb R$ is a weight function and the constant $C>0$ is controlled by the positivity of $\varphi$. [example: Solving Versus Estimating] For a smoothly bounded disc $D\subset\mathbb C$, an explicit Cauchy-Green operator can solve the scalar equation $\bar{\partial}u=f\,d\bar z$ under regularity assumptions on $f$. For example, the formula has the schematic form \begin{align*} u(z)=\frac{1}{\pi}\int_D \frac{f(\zeta)}{\zeta-z}\,d\mathcal L^2(\zeta), \end{align*} and the identity $\bar{\partial}u=f\,d\bar z$ comes from the distributional relation $\bar{\partial}_z((\zeta-z)^{-1})=\pi\delta_\zeta$ with the chosen normalization. This proves existence by writing down a kernel. The $L^2$ method reorganizes the problem: instead of first constructing $u$ and then estimating it, one proves an inequality of the form \begin{align*} \|v\|_{H_2}^2 \le C\bigl(\|T_\varphi^*v\|_{H_1}^2+\|Sv\|_{H_3}^2\bigr), \end{align*} where $T=\bar{\partial}$, $S=\bar{\partial}$ on the next bidegree, and $T_\varphi^*$ is the weighted [Hilbert space](/page/Hilbert%20Space) adjoint. Since $S T=\bar{\partial}^2=0$, this estimate turns compatible data $f\in\ker S$ into a solution $Tu=f$ with \begin{align*} \|u\|_{H_1}\le C^{1/2}\|f\|_{H_2}. \end{align*} Thus the several-variable theory is organized around estimates: the same inequality supplies both existence of a solution and quantitative control of its size. [/example] ## Why Weighted $L^2$ Spaces Enter How can a norm detect the complex geometry that governs solvability of $\bar{\partial}u=f$? The answer is to introduce weights whose complex Hessians supply positivity. This shifts the problem from finding an explicit inverse for $\bar{\partial}$ to proving a coercive inequality for $\bar{\partial}$ and its [Hilbert space](/page/Hilbert%20Space) adjoint. For a real-valued weight $\varphi: \Omega \to \mathbb R$, the weighted space $L^2(\Omega,e^{-\varphi})$ has norm \begin{align*} \|u\|_{L^2(\Omega,e^{-\varphi})}^2 = \int_\Omega |u|^2 e^{-\varphi}\,d\mathcal L^{2n}. \end{align*} The exponential weight can penalize growth, localize mass, or encode curvature. In this theory the most important feature is the Levi form of $\varphi$, which is the Hermitian form built from the second derivatives $\partial^2 \varphi/\partial z_i \partial \bar z_j$. The formal adjoint $\bar{\partial}_\varphi^*$ depends on the weight, so the basic [Hilbert space](/page/Hilbert%20Space) inequality compares a form to $\bar{\partial}$ and $\bar{\partial}_\varphi^*$ applied to it. The resulting estimate is a complex analogue of an energy estimate for an elliptic operator, but it is sharper because it keeps track of bidegree and curvature. [quotetheorem:3669] [citeproof:3669] This abstract principle is the [Hilbert space](/page/Hilbert%20Space) skeleton behind the theorem of Hörmander. The hypothesis $S T=0$ is the abstract form of $\bar{\partial}^2=0$: it says that exact data are automatically compatible with the next equation in the complex. The coercive estimate rules out uncontrolled harmonic components in $\ker T^*\cap\ker S$ and gives the closed-range property needed to invert $T$ on compatible data. If such an estimate fails, the obstruction may be analytic rather than algebraic: compatible data can fail to have solutions with uniform norm bounds, and nonzero [harmonic representatives](/theorems/2747) may survive. In the $\bar{\partial}$ problem the constant $C$ is supplied by lower bounds for the Levi form of the weight, so the analytic work of the course is to prove the weighted estimate needed to apply the principle with $T=\bar{\partial}$ and $S=\bar{\partial}$ on consecutive bidegrees. ## Pseudoconvexity as the Correct Geometry Which domains support the $\bar{\partial}$ estimates needed for [holomorphic function](/page/Holomorphic%20Function) theory? Pseudoconvexity is the geometric condition that survives when ordinary convexity is translated into complex analysis. It can be expressed through plurisubharmonic exhaustion functions, through the Levi form of a boundary defining function in the smooth case, or through holomorphic convexity in more global language. In Course I, pseudoconvexity appeared as a boundary condition and as a domain-of-holomorphy condition. Here it reappears in analytic form: a pseudoconvex domain admits plurisubharmonic weights that make the Bochner-Kodaira identity positive in the directions relevant to $\bar{\partial}$. This is why the same geometric hypothesis controls existence of holomorphic functions, solvability of $\bar{\partial}$, and coherence-type consequences. [example: Convex Domains Are Pseudoconvex] If $\Omega=\mathbb C^n$, then $\rho(z)=|z|^2$ is a plurisubharmonic exhaustion, so $\Omega$ is pseudoconvex. Assume now that $\Omega\ne\mathbb C^n$ and put \begin{align*} \delta(z)=\operatorname{dist}(z,\mathbb C^n\setminus\Omega),\qquad z\in\Omega . \end{align*} We first verify the key convexity estimate for $\delta$. Let $x,y\in\Omega$, $0\le t\le 1$, and choose $0<r_x<\delta(x)$, $0<r_y<\delta(y)$. Then $B(x,r_x)$ and $B(y,r_y)$ are contained in $\Omega$. If \begin{align*} w=(1-t)x+ty+\eta,\qquad |\eta|<(1-t)r_x+tr_y, \end{align*} and $R=(1-t)r_x+tr_y$, then \begin{align*} w &=(1-t)\left(x+\frac{r_x}{R}\eta\right) +t\left(y+\frac{r_y}{R}\eta\right), \end{align*} with \begin{align*} \left|\frac{r_x}{R}\eta\right|<r_x,\qquad \left|\frac{r_y}{R}\eta\right|<r_y. \end{align*} Hence both points in parentheses lie in $\Omega$, and convexity of $\Omega$ gives $w\in\Omega$. Therefore \begin{align*} \delta((1-t)x+ty)\ge (1-t)r_x+tr_y. \end{align*} Letting $r_x\uparrow\delta(x)$ and $r_y\uparrow\delta(y)$ gives \begin{align*} \delta((1-t)x+ty)\ge (1-t)\delta(x)+t\delta(y), \end{align*} so $\delta$ is concave on $\Omega$. Since $s\mapsto -\log s$ is decreasing and convex on $(0,\infty)$, the preceding inequality gives \begin{align*} -\log\delta((1-t)x+ty) &\le -\log\bigl((1-t)\delta(x)+t\delta(y)\bigr)\\ &\le (1-t)(-\log\delta(x))+t(-\log\delta(y)). \end{align*} Thus $-\log\delta$ is convex. A convex function on $\mathbb C^n\simeq\mathbb R^{2n}$ is plurisubharmonic: on each complex line, [Jensen's inequality](/theorems/9) applied to the circle average gives the submean inequality. Therefore \begin{align*} \rho(z)=|z|^2-\log\delta(z) \end{align*} is plurisubharmonic on $\Omega$. It remains to check that $\rho$ is an exhaustion. If $z$ approaches $\partial\Omega$, then $\delta(z)\to0$ and hence $-\log\delta(z)\to+\infty$. If $|z|\to\infty$, choose one fixed point $a\in\mathbb C^n\setminus\Omega$; then \begin{align*} \delta(z)\le |z-a|\le |z|+|a|, \end{align*} so \begin{align*} \rho(z)\ge |z|^2-\log(|z|+|a|)\to+\infty. \end{align*} Thus $\rho$ is a plurisubharmonic exhaustion of $\Omega$, which is precisely the exhaustion-function criterion for pseudoconvexity. Convexity enters only through the concavity of the boundary distance, and the resulting exhaustion records that convex domains have the analytic positivity required for the $\bar{\partial}$ estimates. [/example] Pseudoconvexity is not merely a convenient assumption. On non-pseudoconvex domains, holomorphic functions can extend past the boundary in ways that obstruct the desired cohomological and analytic conclusions. A standard model is a Hartogs figure in $\mathbb C^2$: holomorphic functions on the figure extend across the missing compact hole, so the domain cannot support the same separation and cohomological behaviour as a domain of holomorphy. Analytically, this failure is reflected in the loss of the a priori $\bar{\partial}$ estimates that would otherwise produce controlled solutions on all compatible data. The course treats pseudoconvexity as the bridge between the geometry of the domain and the positivity required by $L^2$ estimates. ## The Main Analytic Engine What estimate is strong enough to turn $\bar{\partial}$-closed data into a solution with a useful bound? The theorem of Hörmander answers this by combining weighted Hilbert spaces, plurisubharmonic weights, and an integration-by-parts identity for forms. The estimate is local enough to be flexible and global enough to solve equations on pseudoconvex domains. A representative special case is for $(0,1)$-forms and smooth strictly plurisubharmonic weights. Later lectures will state the full $(p,q)$-form version, track the curvature operator in every bidegree, and remove the smoothness and strict positivity assumptions by approximation. [quotetheorem:3670] [citeproof:3670] The phrase "with estimates" is essential. Many later arguments solve $\bar{\partial}u=f$ not for its own sake, but to correct an approximate holomorphic object while preserving size, curvature, or vanishing order information. The estimate also shows where the hypotheses are used. If pseudoconvexity fails, Hartogs-type extension phenomena and nonclosed-range behaviour can prevent uniform solvability even when $\bar{\partial}f=0$. If $\varphi$ is only plurisubharmonic, the Levi matrix may have zero directions, so the expression involving $M_\varphi^{-1}$ must be interpreted by approximation or by restricting to directions where the curvature is positive. Finally, the displayed $(0,1)$ case is the entry point: for $(p,q)$-forms the curvature term acts on a larger coefficient bundle, and positivity must be checked in the bidegree where the equation is being solved. ## Applications That Organize the Course What do $L^2$ estimates buy beyond existence for a differential equation? They give a flexible method for producing holomorphic functions and sections with controlled norms. Three families of applications structure the second half of the course: extension theorems, positivity of direct images, and analytic vanishing theorems. The Ohsawa-Takegoshi theorem extends holomorphic data from a complex submanifold to the ambient domain with an $L^2$ bound. From the viewpoint of the course, the theorem is a refined correction argument: extend the data approximately, solve a weighted $\bar{\partial}$-equation to remove the error, and choose a singular weight so that the correction vanishes along the submanifold. The theorem of Berndtsson on positivity of direct images uses $L^2$ techniques in a family. The norm of a holomorphic section over one fibre varies with the base parameter, and the theorem proves that this variation has positive curvature. The proof turns the Hörmander estimate into a curvature inequality for a vector bundle of Hilbert spaces. Kodaira-Nakano vanishing appears here in analytic form. Positivity of a line bundle gives curvature, curvature feeds into the Bochner-Kodaira identity, and the resulting estimate forces [harmonic representatives](/theorems/2747) of certain cohomology classes to vanish. This turns differential inequalities into cohomological conclusions. [example: Correcting an Almost Holomorphic Extension] Let $Y=\{s=0\}\subset\Omega$ locally, where $s$ is holomorphic and $ds$ does not vanish on the smooth part of $Y$. Choose a smooth extension $G$ of $g$ such that $G|_Y=g$, and suppose the weight has the form \begin{align*} \varphi=\psi+2\log |s| \end{align*} near $Y$, with $\psi$ locally bounded. If the weighted estimate gives a solution of \begin{align*} \bar{\partial}u=\bar{\partial}G \end{align*} with \begin{align*} \int_\Omega |u|^2 e^{-\varphi}\,d\mathcal L^{2n}<\infty, \end{align*} then define \begin{align*} F=G-u. \end{align*} Using linearity of $\bar{\partial}$, \begin{align*} \bar{\partial}F &=\bar{\partial}(G-u)\\ &=\bar{\partial}G-\bar{\partial}u\\ &=\bar{\partial}G-\bar{\partial}G\\ &=0, \end{align*} so $F$ is holomorphic. It remains to see why $F$ has the same trace as $G$ on $Y$. Near $Y$ the boundedness of $\psi$ gives $e^{-\psi}\ge c>0$ on each small coordinate patch, hence \begin{align*} \int |u|^2 e^{-\varphi}\,d\mathcal L^{2n} =\int |u|^2 e^{-\psi}|s|^{-2}\,d\mathcal L^{2n}<\infty \end{align*} implies \begin{align*} \int |u|^2 |s|^{-2}\,d\mathcal L^{2n}<\infty \end{align*} locally. Since $\bar{\partial}G$ is supported away from the singular set of the weight, the equation $\bar{\partial}u=\bar{\partial}G$ gives $\bar{\partial}u=0$ near $Y$. Thus $u$ is holomorphic near $Y$. If $u$ had a nonzero value at a smooth point of $Y$, then in coordinates with $s=z_1$ the integrand would dominate a positive multiple of $|z_1|^{-2}$ on a small polydisc, and \begin{align*} \int_{|z_1|<\varepsilon} |z_1|^{-2}\,d\mathcal L^2(z_1) =2\pi\int_0^\varepsilon r^{-2}r\,dr =2\pi\int_0^\varepsilon \frac{dr}{r} =\infty, \end{align*} contradicting the weighted $L^2$ bound. Therefore $u|_Y=0$, and \begin{align*} F|_Y=(G-u)|_Y=G|_Y-u|_Y=g-0=g. \end{align*} The singular weight turns the norm estimate into a vanishing condition for the correction term, so solving one $\bar{\partial}$-equation both makes the extension holomorphic and preserves the prescribed data on $Y$. [/example] These applications are not separate tricks. They are variations on the same pattern: encode the desired geometric constraint in a weight, prove the weighted estimate, solve $\bar{\partial}$, and read the estimate as a theorem about holomorphic objects. ## Relation to Earlier and Later Material How does this course fit with the preceding courses in several complex variables? Course I developed the function theory of domains in $\mathbb C^n$, including holomorphic functions, domains of holomorphy, and pseudoconvexity. Course II emphasized sheaves, coherent analytic structures, and cohomological language. This course returns to analysis and shows how much of the same structure can be accessed through quantitative estimates. The analytic viewpoint complements the sheaf-theoretic one. Solving $\bar{\partial}$ with bounds can prove vanishing and extension results that have cohomological interpretations, but it also gives constants, norms, and stability statements. This makes the method especially useful in complex geometry, algebraic geometry, and problems where metric information matters. The later lectures move from the concrete to the geometric. They begin with forms and operators on domains in $\mathbb C^n$, pass through weighted estimates and approximation, then reinterpret the same inequalities for line bundles, vector bundles, and families. By the end, the reader should be able to recognize an $L^2$ method argument even when it is written in geometric language. ## Conventions for These Notes What conventions will keep the analytic and geometric parts of the course aligned? Domains in complex Euclidean space are denoted by $\Omega\subset \mathbb C^n$, and Lebesgue measure on $\mathbb C^n\simeq \mathbb R^{2n}$ is written $d\mathcal L^{2n}$. The operator $\bar{\partial}$ always raises antiholomorphic degree by one, and weighted adjoints are decorated by the relevant weight when ambiguity is possible. Inner products on Hilbert spaces are linear in the first argument. Weighted $L^2$ norms are written with the domain and weight in the subscript when more than one norm is present. A statement called an estimate should be read as both an inequality and a solvability tool, because the [Hilbert space](/page/Hilbert%20Space) argument converts the former into the latter under closed-range hypotheses. Proofs in these notes are included when they are part of the course. Long analytic arguments are first given as proof sketches identifying the estimate, approximation, and compactness steps; later theorem entries can expand those sketches into full step-by-step proofs. Results used only as background are stated with a pointer to the relevant prerequisite theory rather than reproved. The introduction framed the course through proof sketches and focused use of background results; this strategy now materializes in the concrete language of the ∂̄-operator. In one complex variable integral kernels solve the equation explicitly, but in several variables it becomes a system on differential forms whose compatibility conditions cannot be sidestepped. This opening chapter fixes the analytic language—differential forms, the ∂̄-complex, and weighted Hilbert spaces—that will carry all subsequent theorems. # 1. Review and Setup: The $\bar{\partial}$-Equation in Several Variables This opening chapter fixes the analytic language for the rest of the course. In one complex variable, the equation $\bar\partial u=f$ can often be solved by an integral kernel; in several variables the same equation becomes a system on differential forms, and the compatibility conditions become part of the geometry. The central theme is that local exactness is flexible, while global $L^2$ solvability depends on estimates and on pseudoconvexity. The chapter therefore moves from the operator $\bar\partial$ to its Hilbert-space adjoint, then to the first obstructions, and finally to the geometric condition that will support Hörmander's method. ## The Cauchy-Riemann Operator On Forms How should the equation $\bar\partial u=f$ be written when a form has several antiholomorphic directions, and what compatibility does this impose on the data? The answer is to treat $\bar\partial$ as a differential in a complex of $(p,q)$-forms. The index $p$ counts holomorphic differentials $dz_j$, while $q$ counts antiholomorphic differentials $d\bar z_j$. [definition: Complex Differential Form] Let $\Omega \subset \mathbb C^n$ be open, with coordinates $z_j=x_j+iy_j$. A smooth $(p,q)$-form on $\Omega$ is a finite sum \begin{align*} u=\sum_{\substack{|I|=p\\ |J|=q}}' u_{I,J}\, dz_I\wedge d\bar z_J, \end{align*} where $I=(i_1<\cdots<i_p)$, $J=(j_1<\cdots<j_q)$, $u_{I,J}\in C^\infty(\Omega)$, \begin{align*} dz_I=dz_{i_1}\wedge\cdots\wedge dz_{i_p}, \qquad d\bar z_J=d\bar z_{j_1}\wedge\cdots\wedge d\bar z_{j_q}. \end{align*} The space of smooth $(p,q)$-forms is denoted $\Omega^{p,q}(\Omega)$, and the space of compactly supported smooth $(p,q)$-forms is denoted $\Omega_c^{p,q}(\Omega)$. [/definition] The prime on the sum records the antisymmetry of the wedge product: each basis differential is listed once, with increasing indices. The coefficient functions are the analytic part of the form, and the wedge factors record which complex directions are being measured. To turn these forms into a differential complex, we need an operator that differentiates only in the antiholomorphic directions and raises the antiholomorphic degree by one. The Dolbeault operator is the several-variable replacement for the Cauchy-Riemann operator, and its coefficient formula fixes the signs and target bidegree used throughout the chapter. [definition: Dolbeault Operator] For $u\in \Omega^{p,q}(\Omega)$ written as \begin{align*} u=\sum_{I,J}'u_{I,J}\,dz_I\wedge d\bar z_J, \end{align*} the Dolbeault operator is \begin{align*} \bar\partial u =\sum_{I,J}'\sum_{k=1}^n \frac{\partial u_{I,J}}{\partial \bar z_k}\, d\bar z_k\wedge dz_I\wedge d\bar z_J, \qquad \frac{\partial}{\partial \bar z_k} =\frac{1}{2}\left(\frac{\partial}{\partial x_k}+i\frac{\partial}{\partial y_k}\right). \end{align*} The target space is $\Omega^{p,q+1}(\Omega)$. [/definition] For functions, $\bar\partial u=0$ is the Cauchy-Riemann system in all complex coordinate directions. For higher forms, $\bar\partial$ differentiates only the coefficients and then wedges in one additional antiholomorphic differential. The equation $\bar\partial u=f$ can have a solution only if applying $\bar\partial$ to $f$ gives zero. This compatibility condition is meaningful only because repeated antiholomorphic differentiation cancels after the wedge signs are taken into account. The compatibility condition $\bar\partial f=0$ would be unstable if a second application of $\bar\partial$ produced new terms. The obstruction is that mixed antiholomorphic derivatives appear twice with opposite wedge signs, so the operator should cancel itself for algebraic rather than coordinate-dependent reasons. Before introducing cohomology, we need this cancellation as an intrinsic structural identity: it is what makes $\bar\partial$-closed forms form kernels and $\bar\partial$-exact forms form images inside those kernels. [quotetheorem:3409] [citeproof:3409] Nilpotence turns $\bar\partial$ into a cochain differential: every image automatically lies in the next kernel. If this identity failed, the equation $\bar\partial u=f$ would have no intrinsic compatibility condition and the later obstruction theory would have no cohomological meaning. The quotient of closed forms by exact forms is therefore the natural place where failure of solvability is measured. [definition: Dolbeault Complex] For fixed $p$, the Dolbeault complex on $\Omega$ is \begin{align*} 0\longrightarrow \Omega^{p,0}(\Omega) \xrightarrow{\bar\partial}\Omega^{p,1}(\Omega) \xrightarrow{\bar\partial}\cdots \xrightarrow{\bar\partial}\Omega^{p,n}(\Omega) \longrightarrow 0. \end{align*} [/definition] The course focuses on solving at one step of this complex: given a $\bar\partial$-closed $(p,q)$-form $f$, find a $(p,q-1)$-form $u$ with $\bar\partial u=f$. The local theory says that the differential condition is the only obstruction near a point. [quotetheorem:3410] [citeproof:3410] This lemma is a local statement. It does not give a global norm estimate, and it does not account for boundary conditions; those are exactly the issues that $L^2$ methods address. [example: Cauchy Kernel Solution] Let $g\in C_c^\infty(\mathbb C)$ and set $f=g(z)\,d\bar z$. Define \begin{align*} u(z)=\frac{1}{\pi}\int_{\mathbb C}\frac{g(\zeta)}{z-\zeta}\,d\mathcal L^2(\zeta). \end{align*} We show that $\bar\partial u=f$ in the distributional sense, which means $\partial u/\partial\bar z=g$ as distributions. The kernel \begin{align*} K(z)=\frac{1}{\pi z} \end{align*} satisfies the distribution identity $\bar\partial K=\delta_0$, equivalently, for every $\phi\in C_c^\infty(\mathbb C)$, \begin{align*} -\frac{1}{\pi}\int_{\mathbb C}\frac{1}{z}\frac{\partial\phi}{\partial\bar z}(z)\,d\mathcal L^2(z) =\phi(0). \end{align*} Translating this identity by $\zeta$ gives \begin{align*} -\frac{1}{\pi}\int_{\mathbb C}\frac{1}{z-\zeta} \frac{\partial\phi}{\partial\bar z}(z)\,d\mathcal L^2(z) =\phi(\zeta). \end{align*} Therefore [Fubini's theorem](/theorems/2961), applied on the compact set where $g(\zeta)\phi(z)$ is supported, gives \begin{align*} \left\langle \frac{\partial u}{\partial\bar z},\phi\right\rangle &=-\int_{\mathbb C}u(z)\frac{\partial\phi}{\partial\bar z}(z)\,d\mathcal L^2(z)\\ &=-\frac{1}{\pi}\int_{\mathbb C}\int_{\mathbb C} \frac{g(\zeta)}{z-\zeta} \frac{\partial\phi}{\partial\bar z}(z)\,d\mathcal L^2(\zeta)\,d\mathcal L^2(z)\\ &=\int_{\mathbb C}g(\zeta)\left( -\frac{1}{\pi}\int_{\mathbb C} \frac{1}{z-\zeta} \frac{\partial\phi}{\partial\bar z}(z)\,d\mathcal L^2(z) \right)d\mathcal L^2(\zeta)\\ &=\int_{\mathbb C}g(\zeta)\phi(\zeta)\,d\mathcal L^2(\zeta). \end{align*} Hence $\partial u/\partial\bar z=g$ distributionally, so $\bar\partial u=g\,d\bar z=f$. Away from the support of $g$, the kernel is smooth in $z$ and differentiation under the integral gives $\bar\partial u=0$ there; on regions where $g$ is smooth, the displayed distributional identity identifies the smooth derivative with $g$. This is the one-variable right inverse that models the homotopy operators used in the local several-variable lemma. [/example] ## Formal Adjoints And The Complex Laplacian If $\bar\partial$ is the differential, what is the elliptic object whose estimates should produce solutions? The Hilbert-space approach pairs $\bar\partial$ with its adjoint and studies the associated Laplacian. This replaces direct integral formulas by energy inequalities. [definition: $L^2$ Inner Product on Forms] Let $\Omega\subset\mathbb C^n$ be open. An $L^2$ $(p,q)$-form on $\Omega$ is a measurable form \begin{align*} u=\sum_{I,J}'u_{I,J}\,dz_I\wedge d\bar z_J \end{align*} with $u_{I,J}\in L^2(\Omega)$ for all $I,J$. For $u,v\in L^2_{p,q}(\Omega)$, define \begin{align*} (u,v)_{L^2} =\int_\Omega \sum_{I,J}' u_{I,J}\,\overline{v_{I,J}}\,d\mathcal L^{2n}, \qquad \|u\|_{L^2}^2=(u,u)_{L^2}. \end{align*} [/definition] This inner product makes the coefficientwise square-integrable forms into a [Hilbert space](/page/Hilbert%20Space). The Euclidean metric is being used here; later chapters insert a weight $e^{-\varphi}$ into the same integral. Solvability estimates will compare the size of a form with the sizes of its $\bar\partial$ derivative and its adjoint derivative. To state those estimates, we first need the adjoint at the level where no boundary conditions obscure the formula. The compactly supported setting removes boundary terms from integration by parts, so it gives the formal expression that later becomes the Hilbert-space adjoint after domains are imposed. [definition: Formal Adjoint Of The Dolbeault Operator] On compactly supported smooth forms, the formal adjoint $\bar\partial^*$ is the operator \begin{align*} \bar\partial^*: \Omega_c^{p,q}(\Omega)\to \Omega_c^{p,q-1}(\Omega) \end{align*} determined by \begin{align*} (\bar\partial v,u)_{L^2}=(v,\bar\partial^*u)_{L^2}, \qquad v\in\Omega_c^{p,q-1}(\Omega),\quad u\in\Omega_c^{p,q}(\Omega). \end{align*} In Euclidean coordinates, \begin{align*} \bar\partial^*u =-\sum_{j=1}^n \frac{\partial}{\partial z_j}\, \iota_{\partial/\partial\bar z_j}u, \qquad \frac{\partial}{\partial z_j} =\frac{1}{2}\left(\frac{\partial}{\partial x_j}-i\frac{\partial}{\partial y_j}\right), \end{align*} where $\iota_{\partial/\partial\bar z_j}$ denotes contraction with $\partial/\partial\bar z_j$. [/definition] The compact-support hypothesis removes boundary terms. On a bounded domain with boundary, the Hilbert-space adjoint has a domain condition; this domain condition is where the $\bar\partial$-Neumann problem enters. [quotetheorem:3671] [citeproof:3671] On compactly supported forms, this calculation shows that the formal adjoint and the Hilbert-space adjoint agree on the common test domain. Once compact support is removed, boundary terms no longer vanish automatically, and the adjoint domain must encode a boundary condition. The next operator packages the two first-order quantities $\bar\partial u$ and $\bar\partial^*u$ into a single second-order expression. This is useful because vanishing of both first-order terms can then be studied through one nonnegative Laplacian, while later boundary conditions determine its actual Hilbert-space domain. [definition: Complex Laplacian] On smooth forms of bidegree $(p,q)$, the complex Laplacian is \begin{align*} \square=\bar\partial\bar\partial^*+\bar\partial^*\bar\partial. \end{align*} It maps $(p,q)$-forms to $(p,q)$-forms. [/definition] The sign convention is chosen so that $\square$ is nonnegative as an operator on $L^2$ forms. This is the complex analogue of using $-\Delta$ rather than $\Delta$ for the real Dirichlet energy. Before using $\square$ in estimates, we need to know that this convention has the expected analytic scale. On functions, the $\bar\partial$ part should reduce to the ordinary real Laplacian up to the constants forced by Wirtinger derivatives. The next comparison is therefore a normalization test: it fixes the factor relating complex and real second derivatives, justifies the positivity convention, and identifies the constants that will reappear when real-variable elliptic estimates are translated into complex coordinates. [quotetheorem:3672] [citeproof:3672] For functions, the comparison with the real Laplacian is explicit: \begin{align*} \square u =\bar\partial^*\bar\partial u =-\sum_{j=1}^n\frac{\partial^2u}{\partial z_j\partial\bar z_j} =-\frac14\sum_{j=1}^n\left(\frac{\partial^2u}{\partial x_j^2}+\frac{\partial^2u}{\partial y_j^2}\right). \end{align*} Thus $\square$ is, up to the factor $1/4$, the positive real Laplacian on functions. On higher forms it also contains the geometry of the bidegree and, once weights are introduced, curvature terms. ## Solvability And Obstructions Which forms can be $\bar\partial$ of something, and why do formal energy identities not settle the problem on an arbitrary domain? The first obstruction is algebraic, coming from $\bar\partial^2=0$. The deeper obstructions are analytic: closed range, boundary conditions, singular boundary points, and the absence of coercive estimates. [quotetheorem:3673] [citeproof:3673] This obstruction is necessary in every degree. It is not sufficient for global $L^2$ solvability unless the domain and the functional-analytic range of $\bar\partial$ are controlled. [example: Non-Closed Data In Bidegree Zero One] Let $U\subset\mathbb C^2$ be a connected [open set](/page/Open%20Set), and set \begin{align*} f=\bar z_2\,d\bar z_1 \end{align*} on $U$. We show that no smooth function $u\in C^\infty(U)$ can satisfy $\bar\partial u=f$. Using the definition of $\bar\partial$ on a $(0,1)$-form, \begin{align*} \bar\partial f &=\bar\partial(\bar z_2\,d\bar z_1)\\ &=\sum_{k=1}^2 \frac{\partial \bar z_2}{\partial \bar z_k}\, d\bar z_k\wedge d\bar z_1\\ &= \frac{\partial \bar z_2}{\partial \bar z_1}\, d\bar z_1\wedge d\bar z_1 + \frac{\partial \bar z_2}{\partial \bar z_2}\, d\bar z_2\wedge d\bar z_1\\ &=0\cdot d\bar z_1\wedge d\bar z_1 +1\cdot d\bar z_2\wedge d\bar z_1\\ &=d\bar z_2\wedge d\bar z_1\\ &=-d\bar z_1\wedge d\bar z_2. \end{align*} The form $d\bar z_1\wedge d\bar z_2$ is the standard nonzero basis $(0,2)$-form in $\mathbb C^2$, so $\bar\partial f\ne 0$ on every nonempty open subset of $U$. If $\bar\partial u=f$, then \begin{align*} \bar\partial f=\bar\partial(\bar\partial u)=\bar\partial^2u=0 \end{align*} by *Nilpotence Of The Dolbeault Operator*, contradicting the displayed calculation. Thus $f$ fails the closedness condition and cannot be the $\bar\partial$ of a smooth function. [/example] The compatibility condition $\bar\partial f=0$ is only a formal obstruction; it does not by itself construct a primitive or control its norm. The Hilbert-space method asks for an a priori estimate against the adjoint, because such an estimate is exactly what allows a bounded linear functional to be extended and represented by an $L^2$ solution. This turns solvability into a quantitative statement rather than a purely algebraic one. [quotetheorem:3674] [citeproof:3674] This criterion explains why the course spends so much effort proving estimates. The hypotheses are not cosmetic: closed range and an adjoint estimate are the mechanism that converts the compatibility condition into an actual $L^2$ primitive with norm control. Without that quantitative input, local exactness can coexist with global failure of solvability. The next result shows that the obstruction is not just an artifact of the abstract criterion. On domains with missing sets or poor global geometry, a closed form may satisfy the formal condition but still fail to have the desired primitive, so geometry must enter the estimate. [quotetheorem:3675] [citeproof:3675] The point of the example is not that punctures always destroy solvability in every degree. It is that local exactness and formal adjoints do not by themselves produce the estimate needed for a global $L^2$ primitive. [example: Lewy Equation As A Dbar-Type Warning] [claim]The vector field \begin{align*} L=\frac{\partial}{\partial x}+i\frac{\partial}{\partial y}-2i(x+iy)\frac{\partial}{\partial t} \end{align*} is a tangential Cauchy-Riemann vector field on the Heisenberg hypersurface $\operatorname{Im}w=|z|^2$, but there is $h\in C^\infty$ near $0$ for which $Lu=h$ has no local $C^1$ solution.[/claim] [proof]Let $M\subset\mathbb C^2$ have coordinates $(z,w)$ and be given by $\operatorname{Im}w=|z|^2$. Parametrize $M$ by \begin{align*} \Phi(x,y,t)=\bigl(x+iy,\ t+i(x^2+y^2)\bigr). \end{align*} Indeed, \begin{align*} \operatorname{Im}\bigl(t+i(x^2+y^2)\bigr)-|x+iy|^2 &=(x^2+y^2)-(x^2+y^2)\\ &=0. \end{align*} Writing $z=x+iy$ and $w=t+i(x^2+y^2)$, we have \begin{align*} Lz &=\left(\frac{\partial}{\partial x}+i\frac{\partial}{\partial y}-2iz\frac{\partial}{\partial t}\right)(x+iy)\\ &=1+i\cdot i-2iz\cdot 0\\ &=1-1\\ &=0, \end{align*} and \begin{align*} Lw &=\left(\frac{\partial}{\partial x}+i\frac{\partial}{\partial y}-2iz\frac{\partial}{\partial t}\right)\bigl(t+i(x^2+y^2)\bigr)\\ &=2ix+i(2iy)-2i(x+iy)\\ &=2ix-2y-2ix+2y\\ &=0. \end{align*} Hence, for every [holomorphic function](/page/Holomorphic%20Function) $F(z,w)$ near $M$, the chain rule gives \begin{align*} L(F\circ\Phi) &=\frac{\partial F}{\partial z}(\Phi)Lz+\frac{\partial F}{\partial w}(\Phi)Lw\\ &=0, \end{align*} so $L$ is a $(0,1)$ tangential Cauchy-Riemann vector field. The adjoint testing identity is the obstruction mechanism. If $V$ is a small neighbourhood of $0$, $u\in C^1(V)$, and $\psi\in C_c^\infty(V)$, then compact support removes boundary terms and \begin{align*} \int_V (Lu)\overline{\psi}\,dxdydt &=\int_V \left(\frac{\partial u}{\partial x} +i\frac{\partial u}{\partial y} -2iz\frac{\partial u}{\partial t}\right)\overline{\psi}\,dxdydt\\ &=-\int_V u\frac{\partial\overline{\psi}}{\partial x}\,dxdydt -i\int_V u\frac{\partial\overline{\psi}}{\partial y}\,dxdydt +2i\int_V uz\frac{\partial\overline{\psi}}{\partial t}\,dxdydt\\ &=\int_V u\,\overline{\left( -\frac{\partial}{\partial x} +i\frac{\partial}{\partial y} -2i\bar z\frac{\partial}{\partial t} \right)\psi}\,dxdydt. \end{align*} Thus the formal adjoint is \begin{align*} L^* =-\frac{\partial}{\partial x} +i\frac{\partial}{\partial y} -2i\bar z\frac{\partial}{\partial t}. \end{align*} If $Lu=h$ held on $V$, then every [test function](/page/Test%20Function) $\psi\in C_c^\infty(V)$ would satisfy \begin{align*} \int_V h\overline{\psi}\,dxdydt =\int_V u\,\overline{L^*\psi}\,dxdydt. \end{align*} By *Lewy's Nonsolvability Theorem*, for this vector field one can choose $h\in C^\infty$ near $0$ so that this distributional adjoint identity has no local distribution solution $u$ near $0$. In particular, it has no local $C^1$ solution.[/proof] The example shows that being a Cauchy-Riemann-type first-order operator is not enough for local solvability; the later positivity hypotheses are analytic input, not cosmetic assumptions. [/example] ## Pseudoconvexity And Plurisubharmonic Exhaustions What geometric condition restores the estimates, and how can it be stated in analytic terms? In several complex variables the correct convexity notion is not real convexity of line segments, but subharmonicity along complex lines. This is the form of pseudoconvexity that connects directly to weighted $L^2$ estimates. [definition: Plurisubharmonic Function] Let $\Omega\subset\mathbb C^n$ be open. An upper semicontinuous function $\varphi:\Omega\to[-\infty,\infty)$ is plurisubharmonic if, for every $a\in\Omega$ and every $\xi\in\mathbb C^n$, the function \begin{align*} \lambda\mapsto \varphi(a+\lambda\xi) \end{align*} is subharmonic on each disc in $\mathbb C$ whose image lies in $\Omega$. [/definition] Plurisubharmonic functions are the potentials that supply positivity in the $\bar\partial$ estimate. For smooth functions, this positivity should be visible infinitesimally, not only through restrictions to complex lines. The complex Hessian gives the local matrix whose nonnegativity records exactly the second-derivative form of the plurisubharmonic condition. The line-based definition is geometrically stable, but it is not the form that appears inside an $L^2$ integration by parts calculation. Estimates see the second derivatives of the weight, so the key obstruction is whether positivity along every complex line can be detected by one Hermitian matrix of mixed derivatives. In the smooth case, the following criterion supplies that local test and makes plurisubharmonicity usable in analytic inequalities. [quotetheorem:3403] [citeproof:3403] This criterion is the bridge between geometry and analysis. For merely upper semicontinuous plurisubharmonic functions there need not be a pointwise Hessian, so the line-restriction definition is the stable one and the Levi matrix criterion is the smooth special case. Nonnegativity, rather than strict positivity, is the natural condition because domains and weights often arise as limits of strictly plurisubharmonic approximations; later estimates extract strict positivity by perturbing the weight when needed. [definition: Pseudoconvex Domain] A domain $\Omega\subset\mathbb C^n$ is pseudoconvex if there exists a continuous plurisubharmonic exhaustion function $\rho:\Omega\to\mathbb R$ such that \begin{align*} \{z\in\Omega:\rho(z)<c\}\Subset\Omega \end{align*} for every $c\in\mathbb R$. [/definition] This definition is invariant under biholomorphic coordinate changes and is flexible enough for nonsmooth domains. Smooth bounded domains also have a boundary formulation in terms of the Levi form of a defining function, recalled from the first course. [example: Ball And Polydisc] Let $B(0,1)\subset\mathbb C^n$ be the unit ball and let $P=\{z\in\mathbb C^n:|z_j|<1\text{ for }1\le j\le n\}$ be the unit polydisc. We show that both are pseudoconvex by writing down continuous plurisubharmonic exhaustion functions. For the ball, set \begin{align*} \rho_B(z)=-\log(1-|z|^2), \qquad |z|^2=\sum_{\ell=1}^n z_\ell\bar z_\ell. \end{align*} Writing $s=1-|z|^2$, we have $s>0$ on $B(0,1)$ and \begin{align*} \frac{\partial \rho_B}{\partial z_j} &=-\frac{1}{s}\frac{\partial s}{\partial z_j} =-\frac{1}{s}(-\bar z_j) =\frac{\bar z_j}{s},\\ \frac{\partial^2\rho_B}{\partial z_j\partial\bar z_k} &=\frac{\partial}{\partial\bar z_k}\left(\frac{\bar z_j}{s}\right) =\frac{\delta_{jk}}{s}+\bar z_j\frac{z_k}{s^2}. \end{align*} Hence, for $\xi\in\mathbb C^n$, \begin{align*} \sum_{j,k=1}^n \frac{\partial^2\rho_B}{\partial z_j\partial\bar z_k}(z) \,\xi_j\overline{\xi_k} &=\frac{1}{s}\sum_{j=1}^n|\xi_j|^2 +\frac{1}{s^2} \left(\sum_{j=1}^n\bar z_j\xi_j\right) \left(\sum_{k=1}^n z_k\overline{\xi_k}\right)\\ &=\frac{|\xi|^2}{s} +\frac{\left|\sum_{j=1}^n\bar z_j\xi_j\right|^2}{s^2}\\ &\ge 0. \end{align*} Thus $\rho_B$ is plurisubharmonic. It is continuous, and its sublevel sets are relatively compact: if $c\le 0$, then $\{\rho_B<c\}=\varnothing$, while if $c>0$, \begin{align*} \rho_B(z)<c &\Longleftrightarrow -\log(1-|z|^2)<c\\ &\Longleftrightarrow \log(1-|z|^2)>-c\\ &\Longleftrightarrow 1-|z|^2>e^{-c}\\ &\Longleftrightarrow |z|^2<1-e^{-c}. \end{align*} Since $1-e^{-c}<1$, the closure of this sublevel set is contained in $B(0,1)$. Therefore $\rho_B$ is a plurisubharmonic exhaustion of the ball. For the polydisc, define \begin{align*} \rho_j(z)=-\log(1-|z_j|^2), \qquad \rho_P(z)=\max_{1\le j\le n}\rho_j(z). \end{align*} Each $\rho_j$ is continuous. If $s_j=1-|z_j|^2$, then \begin{align*} \frac{\partial\rho_j}{\partial z_a} &= \begin{cases} \dfrac{\bar z_j}{s_j},&a=j,\\ 0,&a\ne j, \end{cases}\\ \frac{\partial^2\rho_j}{\partial z_a\partial\bar z_b} &= \begin{cases} \dfrac{1}{s_j}+\dfrac{|z_j|^2}{s_j^2} =\dfrac{s_j+|z_j|^2}{s_j^2} =\dfrac{1}{s_j^2},&a=b=j,\\ 0,&\text{otherwise}. \end{cases} \end{align*} Therefore \begin{align*} \sum_{a,b=1}^n \frac{\partial^2\rho_j}{\partial z_a\partial\bar z_b}(z) \,\xi_a\overline{\xi_b} =\frac{|\xi_j|^2}{(1-|z_j|^2)^2} \ge 0, \end{align*} so each $\rho_j$ is plurisubharmonic. The finite maximum $\rho_P$ is also plurisubharmonic: on every complex line, the restriction is the maximum of finitely many continuous subharmonic functions, and if $m=\max_\ell u_\ell$ on a disc and $u_{j_0}$ realizes the maximum at the centre, then \begin{align*} m(0)=u_{j_0}(0) \le \frac{1}{2\pi}\int_0^{2\pi}u_{j_0}(re^{i\theta})\,d\theta \le \frac{1}{2\pi}\int_0^{2\pi}m(re^{i\theta})\,d\theta. \end{align*} Thus $\rho_P$ is continuous and plurisubharmonic. Its sublevel sets satisfy, for $c>0$, \begin{align*} \rho_P(z)<c &\Longleftrightarrow \rho_j(z)<c\text{ for every }j\\ &\Longleftrightarrow |z_j|^2<1-e^{-c}\text{ for every }j, \end{align*} and for $c\le 0$ the sublevel set is empty. Hence the closure of each nonempty sublevel set is contained in the compact polydisc $\{|z_j|^2\le 1-e^{-c}\text{ for all }j\}\Subset P$. Therefore $\rho_P$ is a plurisubharmonic exhaustion of $P$, and both the ball and the polydisc are pseudoconvex. [/example] In one complex variable, every domain has enough holomorphic functions locally, so there is no serious convexity obstruction to solving extension and separation problems. In several variables, a domain can have hidden holomorphic obstructions: boundary geometry may prevent holomorphic functions from separating points or continuing in the expected way. Pseudoconvexity is designed to rule out exactly this obstruction. [quotetheorem:3416] The construction of holomorphic functions with prescribed boundary obstruction belongs to the qualitative theory of domains of holomorphy. In this course we use the analytic side of the equivalence: pseudoconvexity gives plurisubharmonic weights, and those weights give $L^2$ estimates. [definition: Weighted $L^2$ Space of Forms] Let $\varphi:\Omega\to\mathbb R$ be measurable. The weighted space $L^2_{p,q}(\Omega,e^{-\varphi})$ consists of measurable $(p,q)$-forms \begin{align*} u=\sum_{I,J}'u_{I,J}\,dz_I\wedge d\bar z_J \end{align*} such that \begin{align*} \|u\|_{L^2_\varphi}^2 =\int_\Omega \sum_{I,J}' |u_{I,J}|^2 e^{-\varphi}\,d\mathcal L^{2n}<\infty. \end{align*} [/definition] The remaining lectures turn this setup into a theorem: on pseudoconvex domains, carefully chosen plurisubharmonic weights make the adjoint estimate strong enough to solve $\bar\partial u=f$ with quantitative $L^2$ control. The review in this chapter isolates the two ingredients that will recur throughout the course: the complex $\Omega^{p,\bullet}$ and the positivity encoded by plurisubharmonic functions. Chapter 1 showed that ∂̄-solvability depends critically on the choice of weight; this chapter identifies the weights that supply the necessary positivity. Plurisubharmonic functions, whose complex Hessians encode curvature, are the natural class for such weights. We study both their analytic properties—monotonicity under holomorphic maps, local control via subharmonic slices—and their geometric meaning as curvature potentials that feed positivity into L² estimates. # 2. Plurisubharmonic Functions and Weight Geometry Chapter 1 framed the $\bar{\partial}$-equation as an analytic problem on weighted $L^2$ spaces, after the introduction identified weighted estimates as the course's main tool. This chapter introduces the class of weights that make the estimates work: plurisubharmonic functions. They are the several-variable replacement for subharmonic functions, but their defining tests are performed along complex directions, so they detect the complex Hessian rather than the full real Hessian. The guiding idea is that a weight should have non-negative curvature in every holomorphic direction. We first build this idea from line restrictions and mean values, then translate it into the Levi form for smooth weights, and finally explain why nonsmooth psh weights can be replaced by smooth ones when estimates require differentiability. ## Testing Positivity Along Complex Lines How can a scalar function on $\mathbb C^n$ remember the one-variable maximum principle in every complex direction? The answer is to test it on affine complex lines and holomorphic discs. This is the first place where several complex variables differs from ordinary convexity: real line tests are not the right tests for the $\bar{\partial}$-problem. We recall the one-variable notion in the form used throughout the chapter. Let $D(a,r)=\{\zeta\in\mathbb C:|\zeta-a|<r\}$. [definition: Subharmonic Function] Let $U\subset\mathbb C$ be open. A function $u:U\to[-\infty,\infty)$ is subharmonic if $u$ is upper semicontinuous, is not identically $-\infty$ on any connected component of $U$, and for every closed disc $\overline{D}(a,r)\subset U$, \begin{align*} u(a)\leq \frac{1}{2\pi}\int_0^{2\pi}u(a+re^{i\theta})\,d\theta. \end{align*} [/definition] The mean-value inequality is the analytic shadow of positivity. A [holomorphic function](/page/Holomorphic%20Function) has subharmonic logarithmic modulus, and the same one-variable mechanism will be applied after restricting to complex lines in $\mathbb C^n$. For the $\bar\partial$ problem, the weights must behave subharmonically along every holomorphic direction, not merely along real line segments. This requirement leads to the several-variable notion that preserves the one-variable maximum principle after pullback by holomorphic discs. [definition: Plurisubharmonic Function] Let $\Omega\subset\mathbb C^n$ be open. A function $\phi:\Omega\to[-\infty,\infty)$ is plurisubharmonic if $\phi$ is upper semicontinuous, is not identically $-\infty$ on any connected component of $\Omega$, and for every holomorphic map $f:D(0,1)\to\Omega$, the function $\phi\circ f$ is subharmonic on $D(0,1)$. [/definition] We write psh for plurisubharmonic. The definition asks for all holomorphic discs, which can be difficult to check directly because the discs may curve through the domain in many ways. A usable criterion should reduce the test to affine complex lines while still detecting the same holomorphic-direction positivity. [quotetheorem:3676] [citeproof:3676] This theorem is the practical test for psh functions. It reduces many arguments to one complex variable while preserving the holomorphic geometry needed later in $L^2$ estimates. [example: Logarithm Of The Modulus] On $\mathbb C^n$, define $\phi(z)=\log |z|^2$ for $z\neq 0$ and $\phi(0)=-\infty$. This function is upper semicontinuous, since it is continuous on $\mathbb C^n\setminus\{0\}$ and $\log |z|^2\to-\infty$ as $z\to 0$. We show that every nonconstant affine complex-line restriction is subharmonic, so the *Slice Characterisation Of Plurisubharmonicity* gives that $\phi$ is psh. Fix $a\in\mathbb C^n$ and $v\in\mathbb C^n$ with $v\neq 0$. Put \begin{align*} c=\sum_{j=1}^n a_j\overline{v_j},\qquad \alpha=\frac{c}{|v|^2},\qquad p=a-\alpha v. \end{align*} Then \begin{align*} \sum_{j=1}^n p_j\overline{v_j} &=\sum_{j=1}^n a_j\overline{v_j}-\alpha\sum_{j=1}^n v_j\overline{v_j} \\ &=c-\frac{c}{|v|^2}|v|^2 \\ &=0. \end{align*} Hence, for every $\zeta\in\mathbb C$, \begin{align*} a+\zeta v &=p+(\zeta+\alpha)v, \end{align*} and therefore \begin{align*} |a+\zeta v|^2 &=\sum_{j=1}^n |p_j+(\zeta+\alpha)v_j|^2 \\ &=\sum_{j=1}^n\Bigl(|p_j|^2+|\zeta+\alpha|^2|v_j|^2 +p_j\overline{(\zeta+\alpha)v_j}+(\zeta+\alpha)v_j\overline{p_j}\Bigr) \\ &=|p|^2+|\zeta+\alpha|^2|v|^2 +\overline{\zeta+\alpha}\sum_{j=1}^n p_j\overline{v_j} +(\zeta+\alpha)\sum_{j=1}^n v_j\overline{p_j} \\ &=|p|^2+|v|^2|\zeta+\alpha|^2. \end{align*} Let $A=|v|^2>0$ and $B=|p|^2\geq 0$. The line restriction is \begin{align*} u(\zeta)=\phi(a+\zeta v)=\log\bigl(A|\zeta+\alpha|^2+B\bigr), \end{align*} with value $-\infty$ only when $B=0$ and $\zeta=-\alpha$. If $B>0$, write $w=\zeta+\alpha=x+iy$. Then \begin{align*} \frac{\partial}{\partial x}\log(A(x^2+y^2)+B) &=\frac{2Ax}{A(x^2+y^2)+B},\\ \frac{\partial^2}{\partial x^2}\log(A(x^2+y^2)+B) &=\frac{2A(A(x^2+y^2)+B)-4A^2x^2}{(A(x^2+y^2)+B)^2}, \end{align*} and similarly \begin{align*} \frac{\partial^2}{\partial y^2}\log(A(x^2+y^2)+B) &=\frac{2A(A(x^2+y^2)+B)-4A^2y^2}{(A(x^2+y^2)+B)^2}. \end{align*} Adding the two second derivatives gives \begin{align*} \Delta u &=\frac{4A(A(x^2+y^2)+B)-4A^2(x^2+y^2)}{(A(x^2+y^2)+B)^2}\\ &=\frac{4AB}{(A(x^2+y^2)+B)^2}\geq 0. \end{align*} Thus $u$ is subharmonic by the one-variable $C^2$ subharmonic criterion. If $B=0$, define \begin{align*} u_\varepsilon(\zeta)=\log\bigl(A|\zeta+\alpha|^2+\varepsilon\bigr),\qquad \varepsilon>0. \end{align*} The same calculation gives \begin{align*} \Delta u_\varepsilon(\zeta)=\frac{4A\varepsilon}{(A|\zeta+\alpha|^2+\varepsilon)^2}\geq 0, \end{align*} so each $u_\varepsilon$ is subharmonic. As $\varepsilon\downarrow 0$, the functions $u_\varepsilon$ decrease pointwise to $\log(A|\zeta+\alpha|^2)$, with value $-\infty$ at $\zeta=-\alpha$; the decreasing-limit property for subharmonic functions therefore makes this limiting restriction subharmonic. The degenerate case $v=0$ gives a constant restriction. Hence all affine complex-line tests pass, and $\log |z|^2$ is plurisubharmonic on $\mathbb C^n$. [/example] ## Mean Values, Maximum Principles, And Convolution What basic estimates survive after replacing holomorphic functions by psh weights? The most important ones are the sub-mean-value inequality, the maximum principle, and smoothing by positive convolution kernels. These are the tools that let us move between geometric definitions and analytic estimates. [quotetheorem:3677] [citeproof:3677] The inequality says that psh functions cannot have isolated upward spikes: their value is controlled by averages on small Euclidean balls. This is the basic estimate behind compactness arguments for psh families, because it converts pointwise behavior into integral control. It also prepares the maximum principle, where an interior maximum forces the function to lose all room for genuine variation. [quotetheorem:3678] [citeproof:3678] The maximum principle is a rigidity statement rather than a smoothing statement. It explains why boundary behavior governs psh functions on connected domains, but it does not produce the differentiable weights needed in $L^2$ identities. For estimates we therefore need a separate approximation tool that preserves plurisubharmonicity while improving regularity. [quotetheorem:3679] [citeproof:3679] This local smoothing is enough for many computations on compact subsets. It does not by itself solve the global gluing problem, because different local convolutions need not match on overlaps. [example: Corner Weight On A Polydisc] Let \begin{align*} \psi_j(z)=-\log(1-|z_j|^2),\qquad 1\leq j\leq n, \end{align*} so that $\rho=\max_j\psi_j$. We first verify plurisubharmonicity from the definition. For the one-variable model \begin{align*} \chi(w)=-\log(1-|w|^2),\qquad w=x+iy,\ |w|<1, \end{align*} put $s=1-x^2-y^2$. Then \begin{align*} \frac{\partial \chi}{\partial x} &=\frac{2x}{s},& \frac{\partial^2\chi}{\partial x^2} &=\frac{2}{s}+\frac{4x^2}{s^2},\\ \frac{\partial \chi}{\partial y} &=\frac{2y}{s},& \frac{\partial^2\chi}{\partial y^2} &=\frac{2}{s}+\frac{4y^2}{s^2}. \end{align*} Hence \begin{align*} \Delta\chi &=\frac{4}{s}+\frac{4(x^2+y^2)}{s^2}\\ &=\frac{4s+4(x^2+y^2)}{s^2}\\ &=\frac{4}{(1-|w|^2)^2}\geq 0. \end{align*} Thus $\chi$ is subharmonic by the one-variable $C^2$ subharmonic criterion. Now let $f:D(0,1)\to\Delta^n$ be holomorphic. For each $j$, \begin{align*} u_j(\zeta)=\psi_j(f(\zeta))=\chi(f_j(\zeta)). \end{align*} Writing $f_j=p+iq$, the real chain rule and the Cauchy-Riemann equations give \begin{align*} \Delta(\chi\circ f_j) &=\chi_{xx}(f_j)(p_x^2+p_y^2)+2\chi_{xy}(f_j)(p_xq_x+p_yq_y) +\chi_{yy}(f_j)(q_x^2+q_y^2)\\ &\qquad+\chi_x(f_j)\Delta p+\chi_y(f_j)\Delta q\\ &=(\chi_{xx}(f_j)+\chi_{yy}(f_j))(p_x^2+p_y^2)\\ &=(\Delta\chi)(f_j)|f_j'|^2\\ &=\frac{4|f_j'|^2}{(1-|f_j|^2)^2}\geq 0. \end{align*} So each $u_j$ is subharmonic. To check the maximum, fix a closed disc $\overline{D}(\zeta_0,r)\subset D(0,1)$ and choose $m$ with \begin{align*} \max_j u_j(\zeta_0)=u_m(\zeta_0). \end{align*} Then \begin{align*} (\rho\circ f)(\zeta_0) &=u_m(\zeta_0)\\ &\leq \frac{1}{2\pi}\int_0^{2\pi}u_m(\zeta_0+re^{i\theta})\,d\theta\\ &\leq \frac{1}{2\pi}\int_0^{2\pi}\max_j u_j(\zeta_0+re^{i\theta})\,d\theta\\ &=\frac{1}{2\pi}\int_0^{2\pi}(\rho\circ f)(\zeta_0+re^{i\theta})\,d\theta. \end{align*} Thus $\rho\circ f$ is subharmonic for every holomorphic disc $f$, so $\rho$ is psh. Since each $\psi_j$ is continuous and $\rho$ is a finite maximum of the $\psi_j$, $\rho$ is continuous. It remains to check the exhaustion property. If $c\leq 0$, then $\{\rho<c\}=\varnothing$. If $c>0$, then \begin{align*} \rho(z)<c &\Longleftrightarrow -\log(1-|z_j|^2)<c\quad\text{for every }j\\ &\Longleftrightarrow \log(1-|z_j|^2)>-c\quad\text{for every }j\\ &\Longleftrightarrow 1-|z_j|^2>e^{-c}\quad\text{for every }j\\ &\Longleftrightarrow |z_j|^2<1-e^{-c}\quad\text{for every }j. \end{align*} Therefore \begin{align*} \{\rho<c\} =\left\{z\in\Delta^n:|z_j|<\sqrt{1-e^{-c}}\text{ for every }j\right\}, \end{align*} whose closure is compact and still contained in $\Delta^n$. Hence $\rho$ is a continuous psh exhaustion of $\Delta^n$. The nonsmoothness comes from the maximum. Since $r\mapsto-\log(1-r^2)$ is increasing on $[0,1)$, the active terms are exactly those coordinates with $|z_j|=\max_\ell |z_\ell|$. Where a unique coordinate is active, $\rho$ equals one smooth function $\psi_j$ locally. Where two nonzero coordinates $j\neq k$ tie for the maximum, the active gradients are supported in different coordinate directions: \begin{align*} \nabla\psi_j &=\frac{2}{1-|z_j|^2}(0,\ldots,x_j,y_j,\ldots,0),\\ \nabla\psi_k &=\frac{2}{1-|z_k|^2}(0,\ldots,x_k,y_k,\ldots,0). \end{align*} These gradients are distinct, so the directional derivative of $\max(\psi_j,\psi_k)$ is the maximum of two different linear functions rather than a single linear function. Thus $\rho$ has corners along the locus where two or more coordinate terms tie for the maximum. [/example] ## Strict Positivity And The Levi Form How do we recognise psh weights by differentiating them? For smooth functions, the line test becomes a positivity condition on a Hermitian matrix. This matrix is the Levi form, and it is the local curvature object that appears in the Kohn-Morrey identity in Chapter 3 and in the Bochner-Kodaira identity in Chapter 4. [definition: Levi Form] Let $\Omega\subset\mathbb C^n$ be open and let $\phi\in C^2(\Omega)$. The Levi form of $\phi$ at $z\in\Omega$ is the Hermitian form \begin{align*} \mathcal L_\phi(z;\xi)=\sum_{j,k=1}^n \frac{\partial^2\phi}{\partial z_j\partial \bar z_k}(z)\,\xi_j\overline{\xi_k},\qquad \xi\in\mathbb C^n. \end{align*} [/definition] The Levi form measures the second derivative only in complex directions. A function may be convex as a real function without being adapted to holomorphic geometry, and the Levi form is the adapted replacement. [quotetheorem:3403] [citeproof:3403] The criterion turns the qualitative line-test definition into a matrix inequality that can be inserted directly into analytic estimates. For basic positivity, semidefinite Levi forms are stable under limits and therefore fit approximation arguments well. Coercive estimates, however, can fail in Levi-null directions because the weight contributes no lower bound there. To state hypotheses that control every complex tangent direction, we isolate the strictly positive version of the same Levi-form condition. [definition: Strictly Plurisubharmonic Function] Let $\Omega\subset\mathbb C^n$ be open. A function $\phi\in C^2(\Omega)$ is strictly plurisubharmonic if \begin{align*} \mathcal L_\phi(z;\xi)>0 \end{align*} for every $z\in\Omega$ and every non-zero $\xi\in\mathbb C^n$. [/definition] For nonsmooth functions, strictness is usually interpreted locally by subtracting a small multiple of $|z|^2$ and asking the result to remain psh. The smooth definition above is the case used in direct curvature computations. [example: Quadratic Weight] For $\phi(z)=|z|^2=\sum_{\ell=1}^n |z_\ell|^2=\sum_{\ell=1}^n z_\ell\overline{z_\ell}$ on $\mathbb C^n$, compute its Wirtinger derivatives coordinate by coordinate. For each $k$, \begin{align*} \frac{\partial \phi}{\partial \bar z_k} &=\sum_{\ell=1}^n \frac{\partial}{\partial \bar z_k}\bigl(z_\ell\overline{z_\ell}\bigr)\\ &=\sum_{\ell=1}^n z_\ell\frac{\partial \bar z_\ell}{\partial \bar z_k}\\ &=z_k. \end{align*} Therefore, for $1\leq j,k\leq n$, \begin{align*} \frac{\partial^2\phi}{\partial z_j\partial\bar z_k} &=\frac{\partial}{\partial z_j}(z_k)\\ &=\delta_{jk}. \end{align*} Thus the Levi matrix is $(\delta_{jk})$, the identity matrix. For any $\xi=(\xi_1,\ldots,\xi_n)\in\mathbb C^n$, \begin{align*} \mathcal L_\phi(z;\xi) &=\sum_{j,k=1}^n \frac{\partial^2\phi}{\partial z_j\partial\bar z_k}(z)\,\xi_j\overline{\xi_k}\\ &=\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\\ &=|\xi|^2. \end{align*} If $\xi\neq 0$, then at least one $|\xi_j|^2$ is positive, so $\mathcal L_\phi(z;\xi)>0$. Hence $\phi$ is strictly plurisubharmonic, and its curvature is constant in $z$ and equal to one in every unit complex direction. This is the flat positive model behind the Bargmann-Fock weight $e^{-|z|^2}$. [/example] The quadratic weight is the non-degenerate model: its curvature does not vanish in any complex direction. The next radial example shows the complementary phenomenon, where a logarithmic psh weight remains semipositive but loses strictness along a distinguished direction. [example: Logarithmic Radial Weights] *For the logarithmic weight.* Put $t=|z|^2=\sum_{\ell=1}^n z_\ell\overline{z_\ell}$, so $t>0$ on $\mathbb C^n\setminus\{0\}$. For $\phi(z)=\log t$, the first Wirtinger derivatives are \begin{align*} \frac{\partial \phi}{\partial \bar z_k} &=\frac{1}{t}\frac{\partial t}{\partial \bar z_k}\\ &=\frac{z_k}{t}. \end{align*} Differentiating once more gives, for $1\leq j,k\leq n$, \begin{align*} \frac{\partial^2\phi}{\partial z_j\partial \bar z_k} &=\frac{\partial}{\partial z_j}\left(\frac{z_k}{t}\right)\\ &=\frac{\delta_{jk}}{t}+z_k\frac{\partial}{\partial z_j}(t^{-1})\\ &=\frac{\delta_{jk}}{t}-z_k t^{-2}\frac{\partial t}{\partial z_j}\\ &=\frac{\delta_{jk}}{t}-\frac{z_k\overline{z_j}}{t^2}. \end{align*} Hence \begin{align*} \mathcal L_\phi(z;\xi) &=\sum_{j,k=1}^n\left(\frac{\delta_{jk}}{t}-\frac{z_k\overline{z_j}}{t^2}\right)\xi_j\overline{\xi_k}\\ &=\frac{1}{t}\sum_{j=1}^n|\xi_j|^2 -\frac{1}{t^2}\left(\sum_{j=1}^n\overline{z_j}\xi_j\right) \left(\sum_{k=1}^n z_k\overline{\xi_k}\right)\\ &=\frac{t|\xi|^2-\left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2}{t^2}\\ &=\frac{|z|^2|\xi|^2-\left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2}{|z|^4}. \end{align*} The finite-dimensional [Cauchy-Schwarz inequality](/theorems/432) gives \begin{align*} \left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2 \leq \left(\sum_{j=1}^n|z_j|^2\right)\left(\sum_{j=1}^n|\xi_j|^2\right) =|z|^2|\xi|^2, \end{align*} so $\mathcal L_\phi(z;\xi)\geq 0$. By the *Smooth Psh Criterion*, $\log |z|^2$ is psh on $\mathbb C^n\setminus\{0\}$. It is not strictly psh for $n\geq 2$, since choosing the non-zero direction $\xi=z$ gives \begin{align*} \mathcal L_\phi(z;z) &=\frac{|z|^2|z|^2-\left|\sum_{j=1}^n\overline{z_j}z_j\right|^2}{|z|^4}\\ &=\frac{|z|^4-|z|^4}{|z|^4}\\ &=0. \end{align*} Thus the logarithmic weight has non-negative curvature, but its radial complex direction carries zero Levi curvature. *For the modified radial weight.* On $0<|z|<1$, define \begin{align*} \psi(z)=-\log(-\log |z|^2). \end{align*} Again write $t=|z|^2$ and set $L=-\log t>0$. With $h(t)=-\log(-\log t)=-\log L$, we have \begin{align*} h'(t) &=-\frac{1}{L}\frac{dL}{dt}\\ &=-\frac{1}{L}\left(-\frac{1}{t}\right)\\ &=\frac{1}{tL}, \end{align*} and \begin{align*} h''(t) &=\frac{d}{dt}(tL)^{-1}\\ &=-(tL)^{-2}\frac{d(tL)}{dt}\\ &=-\frac{L+tL'}{t^2L^2}\\ &=-\frac{L-1}{t^2L^2}\\ &=\frac{1-L}{t^2L^2}. \end{align*} For $\psi(z)=h(|z|^2)$, \begin{align*} \frac{\partial \psi}{\partial \bar z_k} &=h'(t)z_k, \end{align*} and therefore \begin{align*} \frac{\partial^2\psi}{\partial z_j\partial \bar z_k} &=\frac{\partial}{\partial z_j}\bigl(h'(t)z_k\bigr)\\ &=h''(t)\overline{z_j}z_k+h'(t)\delta_{jk}. \end{align*} Thus \begin{align*} \mathcal L_\psi(z;\xi) &=h'(t)|\xi|^2+h''(t)\left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2. \end{align*} Decompose $\xi$ into its radial and tangential parts by setting \begin{align*} \lambda=\frac{\sum_{j=1}^n\overline{z_j}\xi_j}{t}, \qquad \eta=\xi-\lambda z. \end{align*} Then \begin{align*} \sum_{j=1}^n\overline{z_j}\eta_j &=\sum_{j=1}^n\overline{z_j}\xi_j-\lambda\sum_{j=1}^n|z_j|^2\\ &=0, \end{align*} so \begin{align*} |\xi|^2 &=|\lambda z+\eta|^2\\ &=|\lambda|^2t+|\eta|^2, \end{align*} and \begin{align*} \left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2 &=|\lambda t|^2\\ &=|\lambda|^2t^2. \end{align*} Substituting these identities gives \begin{align*} \mathcal L_\psi(z;\xi) &=h'(t)\bigl(|\lambda|^2t+|\eta|^2\bigr)+h''(t)|\lambda|^2t^2\\ &=h'(t)|\eta|^2+\bigl(th'(t)+t^2h''(t)\bigr)|\lambda|^2. \end{align*} The two coefficients are \begin{align*} h'(t) &=\frac{1}{tL}>0, \end{align*} and \begin{align*} th'(t)+t^2h''(t) &=\frac{1}{L}+\frac{1-L}{L^2}\\ &=\frac{L+1-L}{L^2}\\ &=\frac{1}{L^2}>0. \end{align*} Therefore \begin{align*} \mathcal L_\psi(z;\xi) &=\frac{|\eta|^2}{tL}+\frac{|\lambda|^2}{L^2}. \end{align*} If $\xi\neq 0$, then either $\eta\neq 0$ or $\lambda\neq 0$, so $\mathcal L_\psi(z;\xi)>0$. Hence $-\log(-\log |z|^2)$ is strictly plurisubharmonic on $0<|z|<1$: the logarithmic correction turns the formerly flat radial direction into a positive curvature direction while preserving positivity in the tangential directions. [/example] ## Curvature Forms Attached To Weights How does a scalar psh function become the curvature term in an $L^2$ estimate? The bridge is the $(1,1)$-form $i\partial\bar{\partial}\phi$. When $\phi$ is used in the density $e^{-\phi}$, this form is the curvature of the corresponding Hermitian weight. [definition: Curvature Form Of A Smooth Weight] Let $\Omega\subset\mathbb C^n$ be open and let $\phi\in C^2(\Omega)$. The curvature form attached to $\phi$ is \begin{align*} \omega_\phi=i\partial\bar{\partial}\phi=i\sum_{j,k=1}^n\frac{\partial^2\phi}{\partial z_j\partial\bar z_k}\,dz_j\wedge d\bar z_k. \end{align*} [/definition] Positivity of this form is exactly positivity of the Levi form. If the Levi matrix is positive definite at every point, $\omega_\phi$ is a Kähler form on $\Omega$; if it is only positive semidefinite, it is a semipositive $(1,1)$-form. [quotetheorem:3680] [citeproof:3680] This is the geometric reason psh weights are the natural hypotheses in Hörmander theory. The curvature term in the basic estimate is positive when the weight is psh, and it is quantitatively positive when the weight is strictly psh. [example: Euclidean And Fubini-Study Type Weights] For the Euclidean weight $\phi(z)=|z|^2=\sum_{\ell=1}^n z_\ell\overline{z_\ell}$, the coordinate derivatives are \begin{align*} \frac{\partial \phi}{\partial \bar z_k} &=\sum_{\ell=1}^n \frac{\partial}{\partial \bar z_k}(z_\ell\overline{z_\ell})\\ &=\sum_{\ell=1}^n z_\ell\delta_{\ell k}\\ &=z_k, \end{align*} and hence \begin{align*} \frac{\partial^2\phi}{\partial z_j\partial\bar z_k} &=\frac{\partial z_k}{\partial z_j}\\ &=\delta_{jk}. \end{align*} Therefore \begin{align*} \omega_\phi &=i\sum_{j,k=1}^n\delta_{jk}\,dz_j\wedge d\bar z_k\\ &=i\sum_{j=1}^n dz_j\wedge d\bar z_j, \end{align*} and for every $\xi\in\mathbb C^n$, \begin{align*} \mathcal L_\phi(z;\xi) &=\sum_{j,k=1}^n\delta_{jk}\xi_j\overline{\xi_k}\\ &=\sum_{j=1}^n|\xi_j|^2\\ &=|\xi|^2. \end{align*} Thus the Euclidean weight gives the flat positive curvature model in this normalization. For the Fubini-Study type weight, put \begin{align*} s=1+|z|^2=1+\sum_{\ell=1}^n z_\ell\overline{z_\ell}. \end{align*} For $\phi(z)=\log s$, \begin{align*} \frac{\partial \phi}{\partial \bar z_k} &=\frac{1}{s}\frac{\partial s}{\partial \bar z_k}\\ &=\frac{z_k}{s}. \end{align*} Differentiating once more gives \begin{align*} \frac{\partial^2\phi}{\partial z_j\partial\bar z_k} &=\frac{\partial}{\partial z_j}\left(\frac{z_k}{s}\right)\\ &=\frac{\delta_{jk}}{s}+z_k\frac{\partial}{\partial z_j}(s^{-1})\\ &=\frac{\delta_{jk}}{s}-z_k s^{-2}\frac{\partial s}{\partial z_j}\\ &=\frac{\delta_{jk}}{s}-\frac{z_k\overline{z_j}}{s^2}. \end{align*} Consequently \begin{align*} \mathcal L_\phi(z;\xi) &=\sum_{j,k=1}^n\left(\frac{\delta_{jk}}{s}-\frac{z_k\overline{z_j}}{s^2}\right)\xi_j\overline{\xi_k}\\ &=\frac{1}{s}\sum_{j=1}^n|\xi_j|^2 -\frac{1}{s^2}\left(\sum_{j=1}^n\overline{z_j}\xi_j\right) \left(\sum_{k=1}^n z_k\overline{\xi_k}\right)\\ &=\frac{s|\xi|^2-\left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2}{s^2}\\ &=\frac{(1+|z|^2)|\xi|^2-\left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2}{(1+|z|^2)^2}. \end{align*} This Hermitian form is positive definite. If $z=0$, then the displayed formula gives $\mathcal L_\phi(0;\xi)=|\xi|^2$. If $z\neq 0$, write \begin{align*} \lambda=\frac{\sum_{j=1}^n\overline{z_j}\xi_j}{|z|^2}, \qquad \eta=\xi-\lambda z. \end{align*} Then \begin{align*} \sum_{j=1}^n\overline{z_j}\eta_j &=\sum_{j=1}^n\overline{z_j}\xi_j-\lambda\sum_{j=1}^n|z_j|^2\\ &=0, \end{align*} so \begin{align*} |\xi|^2 &=|\lambda z+\eta|^2\\ &=|\lambda|^2|z|^2+|\eta|^2, \end{align*} and \begin{align*} \left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2 &=\left|\sum_{j=1}^n\overline{z_j}(\lambda z_j+\eta_j)\right|^2\\ &=\left|\lambda |z|^2\right|^2\\ &=|\lambda|^2|z|^4. \end{align*} Substitution into the numerator gives \begin{align*} (1+|z|^2)|\xi|^2-\left|\sum_{j=1}^n\overline{z_j}\xi_j\right|^2 &=(1+|z|^2)\bigl(|\lambda|^2|z|^2+|\eta|^2\bigr)-|\lambda|^2|z|^4\\ &=|\lambda|^2|z|^2+(1+|z|^2)|\eta|^2. \end{align*} For $\xi\neq 0$, the pair $(\lambda,\eta)$ is not $(0,0)$, so this numerator is positive; the denominator $(1+|z|^2)^2$ is also positive. Hence $\log(1+|z|^2)$ has positive definite Levi form, with curvature form \begin{align*} \omega_\phi =i\sum_{j,k=1}^n\left(\frac{\delta_{jk}}{1+|z|^2} -\frac{z_k\overline{z_j}}{(1+|z|^2)^2}\right)dz_j\wedge d\bar z_k. \end{align*} This is the standard affine-coordinate potential for the Fubini-Study curvature on projective space. [/example] ## Richberg Regularisation And Psh Exhaustions How can nonsmooth psh geometry be used in estimates that require smooth differential forms? Local convolution is not enough near boundaries or across coordinate patches, so one needs a smoothing theorem that preserves strict positivity. The Richberg theorem supplies this bridge. [quotetheorem:3681] [citeproof:3681] Richberg regularisation lets us prove estimates with smooth strictly psh weights and then pass back to continuous strictly psh weights. For singular psh weights, the [convolution theorem](/theorems/250) gives decreasing smooth approximations on interior subdomains, and monotone convergence is used to pass limits through weighted $L^2$ norms. This distinction matters because regularity is usually needed only during the estimate, while the geometric condition itself should remain stable under limits. To apply such weights on noncompact domains, we also need a way to keep track of approach to the boundary. An exhaustion function packages the boundary into sublevel sets inside the domain, so estimates can be run on relatively compact pieces without losing plurisubharmonic positivity. [definition: Plurisubharmonic Exhaustion Function] Let $\Omega\subset\mathbb C^n$ be open. A plurisubharmonic exhaustion function on $\Omega$ is a psh function $\rho:\Omega\to\mathbb R$ such that for every $c\in\mathbb R$ the set \begin{align*} \{z\in\Omega:\rho(z)<c\} \end{align*} is relatively compact in $\Omega$. [/definition] Exhaustions encode boundary geometry inside the domain. They let us cut off near the boundary by sublevel sets while retaining psh positivity. The remaining question is whether this analytic device is merely convenient or is actually equivalent to pseudoconvexity. For $L^2$ methods, the needed direction is that a pseudoconvex domain supplies plurisubharmonic functions whose sublevel sets stay away from the boundary, so estimates can be localized on compactly contained regions. The converse is also important conceptually: it shows that the exhaustion formulation captures exactly the same domains as the geometric notion. [quotetheorem:3682] This theorem is the analytic form of pseudoconvexity used in the rest of the course. Its proof belongs to the Levi problem and is recalled rather than reproved here; the direction needed for $L^2$ methods is that pseudoconvex domains provide psh weights that blow up at the boundary. [example: Boundary Blow-Up On A Smooth Pseudoconvex Domain] Let $r_0$ be the original $C^2$ defining function. For $\lambda>0$, set \begin{align*} r_\lambda=e^{\lambda r_0}-1. \end{align*} Then $r_\lambda<0$ exactly where $r_0<0$, $r_\lambda=0$ on $\partial\Omega$, and \begin{align*} dr_\lambda=\lambda e^{\lambda r_0}\,dr_0, \end{align*} so $r_\lambda$ is again a defining function. Its Wirtinger derivatives are \begin{align*} \frac{\partial r_\lambda}{\partial \bar z_k} &=\lambda e^{\lambda r_0}\frac{\partial r_0}{\partial \bar z_k}, \end{align*} and \begin{align*} \frac{\partial^2 r_\lambda}{\partial z_j\partial\bar z_k} &=\frac{\partial}{\partial z_j}\left(\lambda e^{\lambda r_0}\frac{\partial r_0}{\partial \bar z_k}\right)\\ &=\lambda^2 e^{\lambda r_0}\frac{\partial r_0}{\partial z_j}\frac{\partial r_0}{\partial \bar z_k} +\lambda e^{\lambda r_0}\frac{\partial^2 r_0}{\partial z_j\partial\bar z_k}. \end{align*} Therefore \begin{align*} \mathcal L_{r_\lambda}(z;\xi) &=\lambda e^{\lambda r_0(z)}\mathcal L_{r_0}(z;\xi) +\lambda^2 e^{\lambda r_0(z)} \left|\sum_{j=1}^n\frac{\partial r_0}{\partial z_j}(z)\xi_j\right|^2. \end{align*} Strong pseudoconvexity gives positivity of $\mathcal L_{r_0}$ on complex tangential directions at $\partial\Omega$, and the added square term is positive in the complex normal direction. By compactness of a sufficiently small boundary collar, choosing $\lambda$ large makes $\mathcal L_{r_\lambda}(z;\xi)\geq 0$ throughout that collar. Rename this modified defining function $r$. Now define, on the collar inside $\Omega$, \begin{align*} \rho(z)=-\log(-r(z)). \end{align*} Since $r<0$, the function $\rho$ is $C^2$. For each $k$, \begin{align*} \frac{\partial \rho}{\partial\bar z_k} &=-\frac{1}{-r}\frac{\partial(-r)}{\partial\bar z_k}\\ &=-\frac{1}{-r}\left(-\frac{\partial r}{\partial\bar z_k}\right)\\ &=\frac{r_{\bar k}}{-r}. \end{align*} Differentiating once more gives \begin{align*} \frac{\partial^2\rho}{\partial z_j\partial\bar z_k} &=\frac{\partial}{\partial z_j}\left(r_{\bar k}(-r)^{-1}\right)\\ &=r_{j\bar k}(-r)^{-1}+r_{\bar k}\frac{\partial}{\partial z_j}\left((-r)^{-1}\right)\\ &=\frac{r_{j\bar k}}{-r}+r_{\bar k}\left(-(-r)^{-2}(-r_j)\right)\\ &=\frac{r_{j\bar k}}{-r}+\frac{r_jr_{\bar k}}{(-r)^2}. \end{align*} Thus for $\xi\in\mathbb C^n$, \begin{align*} \mathcal L_\rho(z;\xi) &=\sum_{j,k=1}^n\left(\frac{r_{j\bar k}}{-r}+\frac{r_jr_{\bar k}}{(-r)^2}\right)\xi_j\overline{\xi_k}\\ &=\frac{1}{-r}\sum_{j,k=1}^n r_{j\bar k}\xi_j\overline{\xi_k} +\frac{1}{(-r)^2} \left(\sum_{j=1}^n r_j\xi_j\right) \left(\sum_{k=1}^n r_{\bar k}\overline{\xi_k}\right)\\ &=\frac{\mathcal L_r(z;\xi)}{-r} +\frac{\left|\sum_{j=1}^n r_j\xi_j\right|^2}{(-r)^2}\\ &\geq 0. \end{align*} By the *Smooth Psh Criterion*, $\rho$ is plurisubharmonic in the boundary collar. Finally, as $z\to\partial\Omega$ from inside the collar, $r(z)\to 0^{-}$, so $-r(z)\to 0^{+}$ and \begin{align*} \rho(z)=-\log(-r(z))\to+\infty. \end{align*} For any real $c$, \begin{align*} \rho(z)<c &\Longleftrightarrow -\log(-r(z))<c\\ &\Longleftrightarrow \log(-r(z))>-c\\ &\Longleftrightarrow -r(z)>e^{-c}\\ &\Longleftrightarrow r(z)<-e^{-c}. \end{align*} Thus every sublevel set of $\rho$ in the collar stays a positive defining-function distance away from $\partial\Omega$. The logarithmic singularity supplies the boundary blow-up, while the modified defining function supplies the non-negative Levi curvature. [/example] Boundary blow-up examples explain how smooth boundary curvature produces exhaustion weights. The final example returns to nonsmooth geometry and shows how Richberg regularisation turns corners into smooth curvature without changing the underlying exhaustion behaviour. [example: Smoothing A Corner Weight] On the polydisc $\Delta^n$, write \begin{align*} \rho(z)=\max_{1\leq j\leq n}\psi_j(z), \qquad \psi_j(z)=-\log(1-|z_j|^2). \end{align*} The preceding corner-weight computation showed that $\rho$ is a continuous psh exhaustion. Fix $\delta>0$ and set \begin{align*} \phi_\delta(z)=\rho(z)+\delta |z|^2. \end{align*} To see explicitly why this adds strict positivity, let $f=(f_1,\ldots,f_n):D(0,1)\to\Delta^n$ be holomorphic and write $f_\ell=p_\ell+iq_\ell$. Since $p_\ell$ and $q_\ell$ are harmonic and satisfy the Cauchy-Riemann equations, \begin{align*} \Delta |f_\ell|^2 &=\Delta(p_\ell^2+q_\ell^2)\\ &=2(p_{\ell,x}^2+p_{\ell,y}^2+q_{\ell,x}^2+q_{\ell,y}^2) +2p_\ell\Delta p_\ell+2q_\ell\Delta q_\ell\\ &=2(p_{\ell,x}^2+p_{\ell,y}^2+q_{\ell,x}^2+q_{\ell,y}^2)\\ &=4(p_{\ell,x}^2+p_{\ell,y}^2)\\ &=4|f_\ell'|^2\geq 0. \end{align*} Therefore \begin{align*} \Delta |f|^2 &=\Delta\left(\sum_{\ell=1}^n |f_\ell|^2\right)\\ &=\sum_{\ell=1}^n 4|f_\ell'|^2\geq 0. \end{align*} Thus $|z|^2$ is psh, and \begin{align*} \phi_\delta(z)-\frac{\delta}{2}|z|^2 &=\rho(z)+\frac{\delta}{2}|z|^2 \end{align*} is psh as a sum of psh functions. Hence $\phi_\delta$ is continuous and strictly psh in the nonsmooth sense. Given any positive continuous error $\varepsilon$, the *[Richberg Regularisation Theorem](/theorems/3681)* applied to $\phi_\delta$ produces a smooth strictly psh function $\widetilde\rho_{\delta,\varepsilon}$ such that \begin{align*} \rho(z)+\delta |z|^2 \leq \widetilde\rho_{\delta,\varepsilon}(z) \leq \rho(z)+\delta |z|^2+\varepsilon(z) \end{align*} for all $z\in\Delta^n$. Equivalently, \begin{align*} 0 \leq \widetilde\rho_{\delta,\varepsilon}(z)-\bigl(\rho(z)+\delta |z|^2\bigr) \leq \varepsilon(z). \end{align*} Since $\widetilde\rho_{\delta,\varepsilon}$ is smooth and strictly psh, its Levi form satisfies \begin{align*} \mathcal L_{\widetilde\rho_{\delta,\varepsilon}}(z;\xi)>0 \end{align*} for every $z\in\Delta^n$ and every non-zero $\xi\in\mathbb C^n$. The exhaustion behaviour is unchanged. If the error is chosen with $0<\varepsilon\leq\eta$, then $0\leq |z|^2<n$ on $\Delta^n$, so \begin{align*} \rho(z) \leq \widetilde\rho_{\delta,\varepsilon}(z) <\rho(z)+\delta n+\eta. \end{align*} Hence, for every real $c$, \begin{align*} \rho(z)<c-\delta n-\eta &\Longrightarrow \widetilde\rho_{\delta,\varepsilon}(z)<c,\\ \widetilde\rho_{\delta,\varepsilon}(z)<c &\Longrightarrow \rho(z)<c. \end{align*} Thus \begin{align*} \{\rho<c-\delta n-\eta\} \subset \{\widetilde\rho_{\delta,\varepsilon}<c\} \subset \{\rho<c\}. \end{align*} Since the sublevel sets of $\rho$ are relatively compact in $\Delta^n$, the sublevel sets of $\widetilde\rho_{\delta,\varepsilon}$ are also relatively compact. Richberg regularisation therefore replaces the nonsmooth maximum by nearby smooth positive Levi curvature while keeping the boundary blow-up and sublevel-set geometry of the original corner weight. [/example] The chapter therefore leaves us with two interchangeable languages. Psh functions are upper semicontinuous objects controlled by one-variable slices and mean values, while smooth psh weights are curvature potentials with semipositive $i\partial\bar{\partial}$ form. The rest of the course moves between these languages whenever an $L^2$ estimate is proved first for smooth weights and then extended by approximation. Chapter 2 developed the weight geometry that makes L² estimates possible; this geometry now becomes part of a concrete boundary value problem on smooth bounded domains. The ∂̄-equation emerges as a question about closed operators on Hilbert spaces, with the Levi form—the restriction of the complex Hessian to the boundary—determining whether solutions can exist. Pseudoconvexity, the semipositive-ness of the Levi form, becomes the key condition guaranteeing that the formal estimates hold. # 3. The $\bar{\partial}$-Neumann Problem and Boundary Conditions Chapter 1 turned the equation $\bar{\partial}u=f$ into a question about closed operators on Hilbert spaces, and Chapter 2 identified the plurisubharmonic weights that supply positivity. This chapter identifies what those operators mean on a bounded smooth domain: the adjoint is not only a differential expression, but also a boundary condition. The main point is that pseudoconvexity is exactly the sign condition that makes the boundary term in the Kohn-Morrey identity usable for $L^2$ estimates. ## The Hilbert Space Problem Behind $\bar{\partial}$-Neumann What has to be specified before the equation $\Box u=f$ becomes a well-posed [Hilbert space](/page/Hilbert%20Space) problem? In the interior, $\Box=\bar{\partial}\bar{\partial}^*+\bar{\partial}^*\bar{\partial}$ resembles a complex Laplacian. On a domain with boundary, the domain of $\bar{\partial}^*$ carries the boundary information, and this is where the Neumann problem enters. [definition: Square-Integrable Forms] Let $\Omega\subset\mathbb C^n$ be a bounded domain and let $0\le q\le n$. The space $L^2_{(0,q)}(\Omega)$ consists of measurable $(0,q)$-forms \begin{align*} u&=\sum_{|J|=q}' u_J\,d\bar z_J \end{align*} with coefficients $u_J\in L^2(\Omega)$ and norm \begin{align*} \|u\|_{L^2}^2&=\sum_{|J|=q}'\int_\Omega |u_J|^2\,d\mathcal L^{2n}. \end{align*} The inner product is \begin{align*} (u,v)_{L^2}&=\sum_{|J|=q}'\int_\Omega u_J\,\overline{v_J}\,d\mathcal L^{2n}. \end{align*} [/definition] The prime on the sum means that $J=(j_1<\cdots<j_q)$ is increasing. This convention avoids counting the same alternating coefficient several times. On $L^2$ forms, coefficients need not have classical derivatives, so the smooth formula for $\bar\partial$ cannot be used pointwise. The operator must instead be interpreted distributionally, and the domain should contain exactly those weakly differentiable forms whose $\bar\partial$ derivative is still square-integrable. This choice is the broad Hilbert-space domain appropriate for weak solvability, before any boundary condition is imposed. [definition: Maximal $\bar{\partial}$ Operator] The maximal $\bar{\partial}$ operator on $L^2_{(0,q)}(\Omega)$ is the closed operator \begin{align*} \bar{\partial}:\operatorname{Dom}(\bar{\partial})\subset L^2_{(0,q)}(\Omega)\longrightarrow L^2_{(0,q+1)}(\Omega) \end{align*} whose domain consists of all $u\in L^2_{(0,q)}(\Omega)$ for which the distributional form $\bar{\partial}u$ belongs to $L^2_{(0,q+1)}(\Omega)$. [/definition] For smooth compactly supported forms this is the usual Cauchy-Riemann operator on coefficients. The maximal definition is the right one for weak solutions, because no boundary value is prescribed for $u$ in the definition of $\operatorname{Dom}(\bar{\partial})$. [definition: Hilbert Adjoint Of $\bar{\partial}$] The Hilbert adjoint \begin{align*} \bar{\partial}^*:\operatorname{Dom}(\bar{\partial}^*)\subset L^2_{(0,q)}(\Omega)\longrightarrow L^2_{(0,q-1)}(\Omega) \end{align*} is defined by the condition that $u\in\operatorname{Dom}(\bar{\partial}^*)$ if there exists $w\in L^2_{(0,q-1)}(\Omega)$ such that \begin{align*} (\bar{\partial}v,u)_{L^2}&=(v,w)_{L^2} \end{align*} for every $v\in\operatorname{Dom}(\bar{\partial})\subset L^2_{(0,q-1)}(\Omega)$. In that case $\bar{\partial}^*u=w$. [/definition] This is an abstract definition, but it is forced by [integration by parts](/theorems/2098). The boundary condition will appear when we compute which smooth forms lie in this adjoint domain. To turn the first-order operators into an estimate, we need a single energy quantity that records both $\bar\partial u$ and $\bar\partial^*u$. The correct domain is therefore the intersection where both terms make sense in $L^2$. This quadratic form is the Hilbert-space object from which the Neumann Laplacian and its inverse will be constructed. [definition: Neumann Form] For $1\le q\le n$, the $\bar{\partial}$-Neumann form domain is \begin{align*} \mathcal D_q&=\operatorname{Dom}(\bar{\partial})\cap\operatorname{Dom}(\bar{\partial}^*)\subset L^2_{(0,q)}(\Omega), \end{align*} and the quadratic form is \begin{align*} Q_q(u,v)&=(\bar{\partial}u,\bar{\partial}v)_{L^2}+(\bar{\partial}^*u,\bar{\partial}^*v)_{L^2},\qquad u,v\in\mathcal D_q. \end{align*} [/definition] The form $Q_q$ measures the energy of a $(0,q)$-form. It is closed and nonnegative, so [Hilbert space](/page/Hilbert%20Space) spectral theory produces a self-adjoint operator associated to it. The next step is to identify the operator represented by this energy form. Formally it should be $\bar\partial\bar\partial^*+\bar\partial^*\bar\partial$, but in $L^2$ that expression is meaningful only on forms for which the two second-order compositions also lie in the correct domains. Naming this operator makes the Neumann problem an equation for a self-adjoint Laplacian rather than a pair of first-order conditions. [definition: $\bar{\partial}$-Neumann Laplacian] The $\bar{\partial}$-Neumann Laplacian in degree $q$ is \begin{align*} \Box_q&=\bar{\partial}\bar{\partial}^*+\bar{\partial}^*\bar{\partial} \end{align*} with domain \begin{align*} \operatorname{Dom}(\Box_q)=\{u\in\mathcal D_q:\bar{\partial}u\in\operatorname{Dom}(\bar{\partial}^*),\ \bar{\partial}^*u\in\operatorname{Dom}(\bar{\partial})\}. \end{align*} [/definition] Solving the $\bar{\partial}$-Neumann problem means inverting $\Box_q$ with this domain. The word "Neumann" refers to the fact that the domain condition is imposed on the adjoint side, much as the real Neumann Laplacian imposes a normal derivative boundary condition. [quotetheorem:3683] [citeproof:3683] This theorem explains why the rest of the chapter focuses on the estimate. Once the estimate is available, existence and boundedness of the Neumann operator follow from functional analysis. The estimate is not just a bound; it removes harmonic obstructions in the relevant degree and makes the range of the operator closed. The analytic content of later arguments is therefore concentrated in proving the inequality under geometric hypotheses such as pseudoconvexity. After that, the Neumann operator packages the weak solution theory into a bounded inverse on the appropriate closed subspace. ## The Boundary Condition In Terms Of The Complex Normal Which boundary condition is hidden inside $\operatorname{Dom}(\bar{\partial}^*)$? The phrase "$\bar{\partial}^*u=0$ on $\partial\Omega$" is shorthand for the vanishing of the boundary term in the adjoint integration-by-parts formula. This condition removes the complex normal component of the form. [definition: Smooth Defining Function] Let $\Omega\subset\mathbb C^n$ be a bounded domain with $C^\infty$ boundary. A smooth defining function for $\Omega$ is a function $\rho\in C^\infty$ on a neighbourhood of $\overline\Omega$ such that \begin{align*} \Omega&=\{z:\rho(z)<0\}, & \partial\Omega&=\{z:\rho(z)=0\}, & d\rho&\ne 0\quad\text{on }\partial\Omega. \end{align*} [/definition] The defining function gives a complex normal direction. Only its complex line matters for the boundary condition, so multiplying $\rho$ by a positive smooth factor does not change the condition. For a $(0,q)$-form, the boundary condition should say that the component containing this complex normal direction vanishes. Since the form has several alternating coefficients, this component is best expressed by contracting the form against the complex normal covector. The following notation records exactly that boundary trace in a way that is independent of the particular coefficient listing. [definition: Complex Normal Contraction] Let $\rho$ be a smooth defining function. Along $\partial\Omega$, define the $(0,1)$ complex normal contraction by \begin{align*} \nu_\rho^{0,1}\lrcorner u &=\sum_{|K|=q-1}'\left(\sum_{j=1}^n \frac{\partial\rho}{\partial z_j}u_{jK}\right)d\bar z_K, \end{align*} where $u=\sum_{|J|=q}'u_Jd\bar z_J$ and $u_{jK}$ denotes the signed coefficient of $d\bar z_j\wedge d\bar z_K$. [/definition] The condition $\nu_\rho^{0,1}\lrcorner u=0$ says that the part of $u$ carrying a $d\bar z$ factor in the complex normal direction vanishes on the boundary. It remains to connect this geometric condition with the abstract adjoint domain defined earlier. The issue is that $\operatorname{Dom}(\bar\partial^*)$ was specified by a boundedness condition against all test forms, not by an explicit boundary equation. Integration by parts shows that, for smooth forms, the only boundary term obstructing adjointness is precisely the complex normal contraction. [quotetheorem:3684] [citeproof:3684] The theorem turns an abstract adjoint domain into a geometric statement. Smooth elements of $\operatorname{Dom}(\bar{\partial}^*)$ are precisely those whose complex normal component is zero at the boundary. [example: Boundary Condition On The Unit Ball] Let $B=\{z\in\mathbb C^n:|z|<1\}$ and choose the defining function $\rho(z)=|z|^2-1$. Since \begin{align*} \rho(z)&=\sum_{\ell=1}^n z_\ell\bar z_\ell-1, \end{align*} treating $z_\ell$ and $\bar z_\ell$ as independent variables gives \begin{align*} \frac{\partial\rho}{\partial z_j}(z) &=\sum_{\ell=1}^n \frac{\partial}{\partial z_j}(z_\ell\bar z_\ell) =\sum_{\ell=1}^n \delta_{j\ell}\bar z_\ell =\bar z_j. \end{align*} For $u=\sum'_{|J|=q}u_Jd\bar z_J$, the complex normal contraction on $\partial B$ is therefore \begin{align*} \nu_\rho^{0,1}\lrcorner u &=\sum_{|K|=q-1}'\left(\sum_{j=1}^n\frac{\partial\rho}{\partial z_j}u_{jK}\right)d\bar z_K \\ &=\sum_{|K|=q-1}'\left(\sum_{j=1}^n\bar z_j u_{jK}\right)d\bar z_K. \end{align*} Thus the $\bar{\partial}$-Neumann boundary condition $\nu_\rho^{0,1}\lrcorner u=0$ is exactly \begin{align*} \sum_{j=1}^n\bar z_j u_{jK}&=0 \end{align*} on $|z|=1$, for every increasing multi-index $K$ with $|K|=q-1$. In degree $q=1$, the only multi-index $K$ of length $0$ is the empty multi-index, and $u_{j\emptyset}=u_j$. The condition becomes \begin{align*} \sum_{j=1}^n\bar z_j u_j&=0. \end{align*} Since the complex tangent space to $\partial B$ at $z$ is \begin{align*} T_z^{1,0}(\partial B) &=\left\{\xi\in\mathbb C^n:\sum_{j=1}^n\bar z_j\xi_j=0\right\}, \end{align*} the coefficient vector $(u_1,\dots,u_n)$ has no complex normal component on the sphere; it is tangential in the Hermitian sense. [/example] This example is the model to keep in mind: the boundary condition does not force the whole form to vanish on $\partial\Omega$. It only removes the forbidden normal component. ## The Kohn-Morrey Identity How does the sign of the boundary geometry enter an $L^2$ estimate? The energy $Q_q(u,u)$ contains both $\bar{\partial}u$ and $\bar{\partial}^*u$, and [integration by parts](/theorems/210) reorganises that energy into coefficient derivatives plus a boundary term. The boundary term is the Levi form of the defining function acting on the tangential part of $u$. [definition: Levi Form Of A Defining Function] Let $\rho$ be a $C^2$ defining function for $\Omega$. At $p\in\partial\Omega$, the complex tangent space is \begin{align*} T_p^{1,0}(\partial\Omega)&=\left\{\xi\in\mathbb C^n:\sum_{j=1}^n\frac{\partial\rho}{\partial z_j}(p)\xi_j=0\right\}. \end{align*} The Levi form of $\rho$ at $p$ is the Hermitian form \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 T_p^{1,0}(\partial\Omega). \end{align*} [/definition] The boundary condition ensures that the coefficient arrays appearing in the boundary term are tangential arrays. Therefore the Levi form is being evaluated in the directions where pseudoconvexity has a sign. [quotetheorem:3685] [citeproof:3685] The identity is a bridge between PDE and complex geometry. Analytic control of $u$ depends on the sign of the Levi form, because the first term on the right is nonnegative and the second term is the only boundary contribution. [example: The Boundary Term On The Unit Ball] For the unit ball $B=\{z\in\mathbb C^n:|z|<1\}$, take $\rho(z)=|z|^2-1=\sum_{\ell=1}^n z_\ell\bar z_\ell-1$. Treating $z_\ell$ and $\bar z_\ell$ as independent variables, \begin{align*} \frac{\partial\rho}{\partial\bar z_k} &=\sum_{\ell=1}^n z_\ell\frac{\partial\bar z_\ell}{\partial\bar z_k} =\sum_{\ell=1}^n z_\ell\delta_{\ell k} =z_k, \end{align*} and therefore \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 boundary contribution in the Kohn-Morrey identity becomes \begin{align*} \sum_{|K|=q-1}'\int_{\partial B}\sum_{j,k=1}^n \frac{\partial^2\rho}{\partial z_j\partial\bar z_k}u_{jK}\overline{u_{kK}}\,dS_\rho &=\sum_{|K|=q-1}'\int_{\partial B}\sum_{j,k=1}^n \delta_{jk}u_{jK}\overline{u_{kK}}\,dS_\rho \\ &=\sum_{|K|=q-1}'\int_{\partial B}\sum_{j=1}^n u_{jK}\overline{u_{jK}}\,dS_\rho \\ &=\sum_{|K|=q-1}'\int_{\partial B}\sum_{j=1}^n |u_{jK}|^2\,dS_\rho. \end{align*} On $\partial B$, the boundary condition is \begin{align*} \sum_{j=1}^n\bar z_j u_{jK}&=0 \end{align*} for each increasing $K$ with $|K|=q-1$, so each vector $(u_{1K},\dots,u_{nK})$ lies in the complex tangent space to the sphere. Since $|\nabla\rho|=2|z|=2$ on $\partial B$, one has $dS_\rho=\frac12\,d\sigma$, and the boundary term is exactly \begin{align*} \frac12\sum_{|K|=q-1}'\int_{\partial B}\sum_{j=1}^n |u_{jK}|^2\,d\sigma. \end{align*} This is a nonnegative multiple of the boundary $L^2$ norm of the tangential coefficient arrays, so the unit ball contributes the favourable boundary sign in every degree $q\ge1$. [/example] The unit ball illustrates the best possible boundary behaviour. For weakly pseudoconvex domains the Levi form may have zero directions, so the boundary term is nonnegative but may not by itself control all components of $u$. ## The Basic Estimate And Pseudoconvexity When does the energy $Q_q(u,u)$ control the full $L^2$ norm of $u$? The Kohn-Morrey identity says that a negative Levi direction would contribute with the wrong sign. Pseudoconvexity is the boundary condition that prevents this obstruction, while a weighted version of the identity supplies the missing coercivity in directions where the Levi form degenerates. [definition: Smooth Pseudoconvex Domain] Let $\Omega\subset\mathbb C^n$ be a bounded domain with $C^2$ boundary and defining function $\rho$. The domain $\Omega$ is pseudoconvex if \begin{align*} \mathcal L_\rho(p;\xi)&\ge0 \end{align*} for every $p\in\partial\Omega$ and every $\xi\in T_p^{1,0}(\partial\Omega)$. [/definition] This definition is independent of the chosen defining function, since changing $\rho$ by a positive smooth factor rescales the Levi form on complex tangent vectors by a positive factor at the boundary. The reason for isolating this boundary condition is that it is exactly the sign hypothesis needed in the integration-by-parts identity for the $\bar\partial$-Neumann problem. Once negative Levi directions are excluded, the boundary contribution no longer destroys coercivity, so the energy form can begin to control the size of a form rather than only its derivatives. [quotetheorem:3686] [citeproof:3686] This estimate is the first bridge from geometry to analysis: pseudoconvexity removes the dangerous boundary term, but degeneracies can still leave too little positivity in the unweighted energy. The next issue is how to manufacture positivity when the boundary alone does not provide it. A weighted norm changes the integration-by-parts identity by adding an interior curvature term, so a strongly convex weight can turn the basic estimate into a coercive one. The model choice $\phi=t|z|^2$ has complex Hessian $tI$, and for large $t$ the interior Hessian term gives a positive multiple of $t\|u\|^2$ in degree $q\ge1$. [quotetheorem:3687] [citeproof:3687] The weighted estimate is useful because it produces closed-range control for the operators that enter the complex Laplacian. Its restriction $q\ge1$ is essential, not technical: in degree $0$, every $L^2$ [holomorphic function](/page/Holomorphic%20Function) satisfies $\bar{\partial}u=0$, so no estimate can control $\|u\|_{L^2}$ on all of $L^2_{(0,0)}(\Omega)$. Once the estimate gives closed range in positive degree, the analytic question changes from bounding forms to solving $\bar\partial u=f$. The $\bar\partial$-Neumann operator is the device that converts the coercive second-order problem into a canonical first-order solution, and the next formula records that conversion. [quotetheorem:3688] [citeproof:3688] Kohn's formula is the operational heart of the chapter. It turns the inverse of the second-order boundary problem into an explicit solution operator for the first-order equation. [example: Hartogs Triangle And Failure Of The Basic Estimate] Let \begin{align*} H&=\{(z_1,z_2)\in\mathbb C^2:|z_1|<|z_2|<1\}. \end{align*} The map \begin{align*} \Phi:\mathbb D\times\mathbb D^*&\longrightarrow H, & \Phi(w_1,w_2)&=(w_1w_2,w_2) \end{align*} is biholomorphic, with inverse \begin{align*} w_2&=z_2, & w_1&=\frac{z_1}{z_2}. \end{align*} Its complex Jacobian matrix is \begin{align*} D\Phi(w) &= \begin{pmatrix} \frac{\partial z_1}{\partial w_1} & \frac{\partial z_1}{\partial w_2}\\ \frac{\partial z_2}{\partial w_1} & \frac{\partial z_2}{\partial w_2} \end{pmatrix} = \begin{pmatrix} w_2 & w_1\\ 0 & 1 \end{pmatrix}, \end{align*} so \begin{align*} \det D\Phi(w)&=w_2, & d\mathcal L^4_z&=|\det D\Phi(w)|^2\,d\mathcal L^4_w =|w_2|^2\,d\mathcal L^4_w. \end{align*} Consider the model coefficient $1/\bar w_2$ on the punctured factor. Its weighted square norm over $\mathbb D\times\mathbb D^*$ is finite because \begin{align*} \int_{\mathbb D\times\mathbb D^*}\left|\frac{1}{\bar w_2}\right|^2|w_2|^2\,d\mathcal L^4_w &=\int_{\mathbb D\times\mathbb D^*}1\,d\mathcal L^4_w \\ &=\left(\int_{\mathbb D}d\mathcal L^2_{w_1}\right) \left(\int_{\mathbb D}d\mathcal L^2_{w_2}\right) \\ &=\pi^2. \end{align*} Thus the singularity at $w_2=0$ is exactly cancelled by the Jacobian factor $|w_2|^2$. Let $\chi_m$ be logarithmic cutoffs on $\mathbb D^*$ which equal $0$ very close to $0$, equal $1$ away from $0$, and satisfy \begin{align*} \int_{\mathbb D^*}|\nabla\chi_m(w_2)|^2\,d\mathcal L^2_{w_2}&\longrightarrow0. \end{align*} For instance, on the annulus $e^{-m^2}<|w_2|<e^{-m}$ one may take \begin{align*} \chi_m(w_2) &=\frac{\log |w_2|+m^2}{m^2-m}. \end{align*} Writing $r=|w_2|$, this gives \begin{align*} \frac{d\chi_m}{dr} &=\frac{1}{(m^2-m)r}, \end{align*} and hence on that annulus \begin{align*} \int |\nabla\chi_m|^2\,d\mathcal L^2 &=2\pi\int_{e^{-m^2}}^{e^{-m}} \frac{1}{(m^2-m)^2r^2}\,r\,dr \\ &=\frac{2\pi}{(m^2-m)^2} \int_{e^{-m^2}}^{e^{-m}}\frac{dr}{r} \\ &=\frac{2\pi}{(m^2-m)^2} \left[\log r\right]_{e^{-m^2}}^{e^{-m}} \\ &=\frac{2\pi}{(m^2-m)^2}(m^2-m) \\ &=\frac{2\pi}{m^2-m}\longrightarrow0. \end{align*} The corresponding pulled-back model forms have the shape \begin{align*} \alpha_m&=c_m\,\chi_m(w_2)\frac{d\bar w_2}{\bar w_2}, \end{align*} where $c_m$ is chosen so that $\|\alpha_m\|_{L^2}=1$ in the pulled-back weighted norm. The preceding norm computation shows that the constants $c_m$ remain bounded above and below after passing to a subsequence, while the only new derivatives come from $\chi_m$. Therefore \begin{align*} \|\bar{\partial}\alpha_m\|_{L^2}^2+\|\bar{\partial}^*\alpha_m\|_{L^2}^2&\longrightarrow0. \end{align*} Pushing $\alpha_m$ forward by $\Phi$ gives normalized $(0,1)$-forms on $H$ with $Q(\alpha_m,\alpha_m)\to0$. If an estimate \begin{align*} \|u\|_{L^2}^2&\le C Q(u,u) \end{align*} held on the Hartogs triangle, then applying it to these forms would give \begin{align*} 1=\|\alpha_m\|_{L^2}^2&\le C Q(\alpha_m,\alpha_m)\longrightarrow0, \end{align*} which is impossible. The failure comes from the singular boundary point at the origin: pseudoconvexity gives the correct Levi sign on smooth boundary pieces, but it does not supply the smooth trace and integration-by-parts framework used in the Kohn-Morrey proof. [/example] The Hartogs triangle separates two issues that coincide in the smooth theory. Pseudoconvexity supplies the correct Levi sign, while smooth boundary supplies the trace theory and [integration by parts](/theorems/2098) needed to convert that sign into a closed-range estimate. [remark: What The Chapter Proves] The $\bar{\partial}$-Neumann problem is the [Hilbert space](/page/Hilbert%20Space) problem attached to the closed form $Q_q$. Its boundary condition is the vanishing of the complex normal contraction. The Kohn-Morrey identity exposes the Levi form as the boundary term in $Q_q$. On smooth bounded pseudoconvex domains, the weighted identity gives the basic estimate, the Neumann operator $N_q$ exists, and Kohn's formula $\bar{\partial}^*N_qf$ gives the minimal-norm solution of $\bar{\partial}u=f$. [/remark] Chapters 1–3 built the formal machinery of ∂̄-identities and boundary conditions under the assumption that solutions exist; this chapter realizes that assumption by proving existence. The new ingredient is a weight $e^{-\varphi}$ whose complex Hessian supplies a positive curvature term in the fundamental L² identity, making the estimate strong enough for solvability. The central question is quantitative: when the right-hand side f is small enough in a weighted L² norm, does the solution u also remain bounded in a weighted norm? # 4. Hörmander's $L^2$ Existence Theorem This chapter turns the formal $\bar\partial$-calculus from the preceding chapters into an existence theorem. The new ingredient is a weight $e^{-\varphi}$ whose complex Hessian contributes a positive curvature term to the $L^2$ identity. The central question is quantitative: when $f$ is $\bar\partial$-closed, can we solve $\bar\partial u=f$ with an a priori norm bound controlled by the positivity of $i\partial\bar\partial\varphi$? The answer is the Hörmander $L^2$ existence theorem. Its proof has three moving parts: a weighted adjoint, a commutator calculation producing curvature, and a [Hilbert space](/page/Hilbert%20Space) lemma that converts an estimate for test forms into a solution of the equation. The theorem is the analytic engine behind many later applications, including Cousin and Mittag-Leffler problems, extension theorems, and vanishing results. ## Weighted $L^2$ Spaces and the Twisted Adjoint How can a choice of weight help solve $\bar\partial u=f$ rather than merely changing the norm? The point is that the adjoint of $\bar\partial$ depends on the weight, so [integration by parts](/theorems/210) produces new zeroth-order terms. These terms are exactly where the positivity of $i\partial\bar\partial\varphi$ enters. [definition: Weighted Space of Forms] Let $\Omega \subset \mathbb C^n$ be a domain and let $\varphi: \Omega \to \mathbb R$ be measurable. For $0 \le q \le n$, the weighted space $L^2_{(0,q)}(\Omega,e^{-\varphi})$ is the space of measurable $(0,q)$-forms \begin{align*} f = \sum_{J\in\mathcal I_q} f_J\, d\bar z_J \end{align*} such that \begin{align*} \|f\|_{e^{-\varphi}}^2 := \int_\Omega |f(z)|^2 e^{-\varphi(z)}\,d\mathcal L^{2n}(z) < \infty. \end{align*} The inner product is \begin{align*} (f,g)_{e^{-\varphi}} := \int_\Omega \sum_{J\in\mathcal I_q} f_J(z)\overline{g_J(z)} e^{-\varphi(z)}\,d\mathcal L^{2n}(z). \end{align*} [/definition] Here $\mathcal I_q$ denotes the strictly increasing multi-indices $J=(j_1<\cdots<j_q)$. The unweighted space is recovered by taking $\varphi=0$, but the weighted space is better adapted to growth problems because $e^{-\varphi}$ penalises or permits growth at prescribed rates. [example: Gaussian Weighted Space on Complex Euclidean Space] For $\Omega=\mathbb C^n$ and $\varphi(z)=|z|^2$, the weighted norm of a function is \begin{align*} \|u\|_{e^{-|z|^2}}^2 = \int_{\mathbb C^n} |u(z)|^2 e^{-|z|^2}\,d\mathcal L^{2n}(z). \end{align*} We first check that every polynomial belongs to this space. If $P$ has degree at most $d$, then for some constant $C>0$, \begin{align*} |P(z)|^2 \le C(1+|z|)^{2d}. \end{align*} Using polar coordinates in $\mathbb R^{2n}$, with $\sigma_{2n-1}$ denoting the surface measure of the unit sphere, each radial monomial term satisfies \begin{align*} \int_{\mathbb C^n} |z|^{2m} e^{-|z|^2}\,d\mathcal L^{2n}(z) &= \sigma_{2n-1}\int_0^\infty r^{2m}e^{-r^2}r^{2n-1}\,dr \\ &= \sigma_{2n-1}\int_0^\infty r^{2m+2n-1}e^{-r^2}\,dr \\ &= \frac{\sigma_{2n-1}}{2}\int_0^\infty t^{m+n-1}e^{-t}\,dt \\ &= \frac{\sigma_{2n-1}}{2}\Gamma(m+n)<\infty, \end{align*} where $t=r^2$ in the third line. Expanding $(1+|z|)^{2d}$ into finitely many powers of $|z|$ therefore gives $\|P\|_{e^{-|z|^2}}<\infty$. For real $a$, set $u(z)=e^{a|z|^2}$. Then \begin{align*} \|u\|_{e^{-|z|^2}}^2 &= \int_{\mathbb C^n} e^{2a|z|^2}e^{-|z|^2}\,d\mathcal L^{2n}(z) \\ &= \int_{\mathbb C^n} e^{-(1-2a)|z|^2}\,d\mathcal L^{2n}(z) \\ &= \sigma_{2n-1}\int_0^\infty e^{-(1-2a)r^2}r^{2n-1}\,dr. \end{align*} If $a<1/2$, then $1-2a>0$, and the substitution $t=(1-2a)r^2$ gives \begin{align*} \sigma_{2n-1}\int_0^\infty e^{-(1-2a)r^2}r^{2n-1}\,dr = \frac{\sigma_{2n-1}}{2(1-2a)^n}\int_0^\infty t^{n-1}e^{-t}\,dt<\infty. \end{align*} If $a=1/2$, the radial integral becomes \begin{align*} \sigma_{2n-1}\int_0^\infty r^{2n-1}\,dr=\infty. \end{align*} If $a>1/2$, then $2a-1>0$, and the radial integral is \begin{align*} \sigma_{2n-1}\int_0^\infty e^{(2a-1)r^2}r^{2n-1}\,dr=\infty. \end{align*} Thus the Gaussian weight makes all polynomial growth square-integrable, but it allows the exponential $e^{a|z|^2}$ exactly up to the threshold $a<1/2$. [/example] The preceding example shows what a weight does to integrability. To use weights in estimates, we also need to know how they change integration by parts. In the unweighted theory, the formal adjoint of $\bar\partial$ is determined by the Lebesgue inner product; after inserting $e^{-\varphi}$, derivatives falling on the weight create new first-order terms. We therefore name the weighted adjoint explicitly before deriving the identities in which its commutator produces curvature. [definition: Twisted Adjoints] Let $\Omega \subset \mathbb C^n$ be a domain and let $\varphi \in C^1(\Omega)$. The weighted adjoint $\bar\partial^*_\varphi$ is the [Hilbert space](/page/Hilbert%20Space) adjoint of \begin{align*} \bar\partial: L^2_{(0,q)}(\Omega,e^{-\varphi}) \longrightarrow L^2_{(0,q+1)}(\Omega,e^{-\varphi}). \end{align*} On compactly supported smooth $(0,q)$-forms it is given by contraction with \begin{align*} \delta_j^\varphi := e^\varphi \frac{\partial}{\partial z_j} e^{-\varphi} = \frac{\partial}{\partial z_j} - \frac{\partial \varphi}{\partial z_j}. \end{align*} [/definition] Thus the weighted adjoint differs from the unweighted formal adjoint by first-order multiplication with derivatives of $\varphi$. This is the first sign that curvature of the weight can affect solvability. The next step is to put this modified adjoint into the same integration-by-parts framework used for the unweighted problem. The goal is not merely to rewrite the energy, but to separate the boundary contribution from the interior curvature contribution. That separation will show exactly where pseudoconvexity and positivity of the weight enter the estimate. [quotetheorem:3689] [citeproof:3689] This identity is the analytic replacement for the [Cauchy integral formula](/theorems/345) in the present theory. It does not produce a solution by itself, but it gives the coercive inequality needed by [Hilbert space](/page/Hilbert%20Space) duality. The hypothesis that $\Omega$ is pseudoconvex is exactly what fixes the sign of $B_{\partial\Omega}(v)$. On a non-pseudoconvex smooth boundary, a complex tangential direction with negative Levi form makes the same integration-by-parts term negative; test forms concentrated near that boundary direction can prevent the identity from giving any global lower bound. For instance, the annulus $B(0,2)\setminus\overline{B}(0,1)$ in $\mathbb C^n$ for $n\ge 2$ has a concave inner boundary, so the boundary contribution has the wrong sign there. The $C^2$ assumption on $\varphi$ is also the natural regularity threshold at this stage: the identity differentiates $\varphi$ twice and uses the entries $\varphi_{j\bar k}$ as pointwise coefficients. Rougher plurisubharmonic weights are handled in Chapter 10 by approximation, after the $C^2$ estimate has established the stable inequality. [illustration:scv-iii-pseudoconvex-levi-boundary] ## The Curvature Commutator Where does the positive term in the estimate come from if $\bar\partial$ itself has no zeroth-order part? It comes from the failure of $\bar\partial$ to commute with the weighted adjoint. The commutator is a multiplication operator given by the complex Hessian of the weight. [quotetheorem:3690] [citeproof:3690] The commutator calculation explains why plurisubharmonicity is the correct convexity notion for weighted $L^2$ estimates. Positivity of the Hermitian matrix $(\varphi_{j\bar k})$ turns an algebraic curvature term into analytic control of the norm of a form. For existence theorems, however, qualitative positivity is not enough: the duality argument needs a uniform numerical lower bound so that the resulting functional has a controlled operator norm. This motivates the following strengthened form of strict plurisubharmonicity. [definition: Uniformly Strictly Plurisubharmonic Weight] Let $\Omega\subset\mathbb C^n$ be a domain. A function $\varphi\in C^2(\Omega)$ satisfies $i\partial\bar\partial\varphi\ge c\omega$ for $c>0$ if \begin{align*} \sum_{j,k=1}^n \varphi_{j\bar k}(z)\xi_j\overline{\xi_k} \ge c|\xi|^2 \end{align*} for every $z\in\Omega$ and every $\xi\in\mathbb C^n$, where $\omega=i\sum_{j=1}^n dz_j\wedge d\bar z_j$ up to the course normalisation. [/definition] With this convention, the curvature term dominates the pointwise squared norm of a $(0,1)$-form. That domination is the source of the constant $1/c$ in the Hörmander theorem. [example: Curvature of the Gaussian Weight] Let $\varphi(z)=|z|^2=\sum_{\ell=1}^n z_\ell\overline{z_\ell}$ on $\mathbb C^n$. Since $\partial \overline{z_\ell}/\partial\bar z_k=\delta_{\ell k}$ and $\partial z_k/\partial z_j=\delta_{jk}$, we have \begin{align*} \frac{\partial\varphi}{\partial\bar z_k} &= \frac{\partial}{\partial\bar z_k}\sum_{\ell=1}^n z_\ell\overline{z_\ell} \\ &= \sum_{\ell=1}^n z_\ell\frac{\partial\overline{z_\ell}}{\partial\bar z_k} \\ &= \sum_{\ell=1}^n z_\ell\delta_{\ell k} \\ &= z_k, \end{align*} and therefore \begin{align*} \varphi_{j\bar k} = \frac{\partial}{\partial z_j}\left(\frac{\partial\varphi}{\partial\bar z_k}\right) = \frac{\partial z_k}{\partial z_j} = \delta_{jk}. \end{align*} Thus the complex Hessian matrix is the identity matrix. For any $\xi=(\xi_1,\ldots,\xi_n)\in\mathbb C^n$, \begin{align*} \sum_{j,k=1}^n \varphi_{j\bar k}\xi_j\overline{\xi_k} &= \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 \\ &= |\xi|^2. \end{align*} Hence $i\partial\bar\partial\varphi\ge \omega$ in the normalisation $\omega=i\sum_{j=1}^n dz_j\wedge d\bar z_j$, with equality at every point. The Gaussian weight is therefore the model case of constant curvature lower bound $c=1$. [/example] This example is the model case for the theorem on $\mathbb C^n$: the curvature is constant, so the final estimate has the same constant at every point. ## Hörmander Main Estimate and Existence Theorem What estimate is strong enough to turn the identity into an actual solution of $\bar\partial u=f$? We need a lower bound for the test form norm in terms of $\bar\partial^*_\varphi$ and $\bar\partial$. Once such an estimate holds, the solution is obtained by applying a [Hilbert space](/page/Hilbert%20Space) criterion to the functional $\bar\partial^*_\varphi v\mapsto (f,v)_{e^{-\varphi}}$. [quotetheorem:3691] [citeproof:3691] The estimate has the form required by functional analysis: it controls every test form $v$ by the operators paired against the equation and its compatibility condition. The strict lower bound $c>0$ matters because it supplies a uniform inverse to the curvature operator. If curvature is merely nonnegative, the inequality may still give qualitative information, but the constant can escape to infinity and the range of $\bar\partial$ need not be closed in the weighted norm. Pseudoconvexity is equally structural: if the Levi boundary term changes sign, the boundary contribution can cancel the interior curvature and the estimate no longer follows from the identity. The next theorem is therefore not just a formal duality statement; it packages the two positivity inputs into a bounded functional whose representing vector is the desired solution. [example: Failure of a Coercive Estimate for a Flat Weight] [claim]No estimate $c\|v\|^2\le \|\bar\partial v\|^2+\|\bar\partial^*v\|^2$ with a fixed $c>0$ can hold for all compactly supported smooth $(0,1)$-forms on $\mathbb C^n$ when $\varphi=0$.[/claim] [proof]Choose $\chi\in C_c^\infty(B(0,2))$ with $\chi=1$ on $B(0,1)$, and set \begin{align*} v_R(z)=\chi(z/R)\,d\bar z_1,\qquad R>0. \end{align*} Since $\varphi=0$, the weighted norm is the ordinary $L^2$ norm. With the change of variables $z=Rw$, we get \begin{align*} \|v_R\|^2 &=\int_{\mathbb C^n}|\chi(z/R)|^2\,d\mathcal L^{2n}(z)\\ &=R^{2n}\int_{\mathbb C^n}|\chi(w)|^2\,d\mathcal L^{2n}(w)\\ &\ge R^{2n}\mathcal L^{2n}(B(0,1)). \end{align*} Write $v_R=\sum_{j=1}^n (v_R)_j\,d\bar z_j$, where $(v_R)_1=\chi(z/R)$ and $(v_R)_j=0$ for $j\ge 2$. Then \begin{align*} \frac{\partial (v_R)_1}{\partial \bar z_k}(z) =\frac{1}{R}\left(\frac{\partial\chi}{\partial \bar z_k}\right)(z/R), \qquad \frac{\partial (v_R)_1}{\partial z_1}(z) =\frac{1}{R}\left(\frac{\partial\chi}{\partial z_1}\right)(z/R). \end{align*} Thus each coefficient of $\bar\partial v_R$ is bounded by $R^{-1}$ times a fixed derivative of $\chi$ evaluated at $z/R$, and the unweighted adjoint satisfies \begin{align*} \bar\partial^*v_R=-\frac{\partial}{\partial z_1}\chi(z/R). \end{align*} Hence, for a constant \begin{align*} A=\sum_{k=1}^n\int_{\mathbb C^n}\left|\frac{\partial\chi}{\partial\bar z_k}(w)\right|^2\,d\mathcal L^{2n}(w) +\int_{\mathbb C^n}\left|\frac{\partial\chi}{\partial z_1}(w)\right|^2\,d\mathcal L^{2n}(w), \end{align*} the same change of variables gives \begin{align*} \|\bar\partial v_R\|^2+\|\bar\partial^*v_R\|^2 &\le \frac{1}{R^2}\int_{\mathbb C^n}\sum_{k=1}^n\left|\frac{\partial\chi}{\partial\bar z_k}(z/R)\right|^2\,d\mathcal L^{2n}(z) +\frac{1}{R^2}\int_{\mathbb C^n}\left|\frac{\partial\chi}{\partial z_1}(z/R)\right|^2\,d\mathcal L^{2n}(z)\\ &=R^{2n-2}A. \end{align*} Therefore \begin{align*} \frac{\|\bar\partial v_R\|^2+\|\bar\partial^*v_R\|^2}{\|v_R\|^2} &\le \frac{R^{2n-2}A}{R^{2n}\mathcal L^{2n}(B(0,1))}\\ &= \frac{A}{\mathcal L^{2n}(B(0,1))}\frac{1}{R^2}. \end{align*} Letting $R\to\infty$ sends this ratio to $0$. If a fixed estimate with constant $c>0$ held for all compactly supported smooth $(0,1)$-forms, applying it to $v_R$ would force this ratio to be at least $c$ for every $R$, a contradiction.[/proof] The flat weight has no positive curvature term, so large slowly varying test forms can have large $L^2$ norm while their first-derivative energy becomes negligible in comparison. [/example] The example shows why nonnegative or flat curvature is insufficient for the constant-version theorem. A non-plurisubharmonic weight is worse: the commutator has negative directions, so it cannot supply the missing coercivity. What remains is to convert the uniform curvature lower bound into solvability of $\bar\partial u=f$. Strict positivity turns the estimate into a uniform bound, and that uniform bound is precisely the hypothesis needed to define a continuous functional on the range of the weighted adjoint. The resulting theorem is the weighted existence statement that makes the preceding estimates usable. [quotetheorem:3692] [citeproof:3692] The content of the theorem is that curvature converts closedness of $f$ into an actual global solution with a quantitative norm bound. The hypothesis $i\partial\bar\partial\varphi\ge c\omega$ is doing all the coercive work; without a positive lower bound, the estimate can fail on large regions where the weight is too flat. The theorem is often stated with a variable lower bound: if $(\varphi_{j\bar k})$ is positive definite but not bounded below by a constant, the estimate uses the inverse Hermitian matrix pointwise. The constant version above is the form most useful for this lecture, and it now feeds directly into the local solvability consequence on pseudoconvex domains. [quotetheorem:3493] [citeproof:3493] This is the advertised corollary: pseudoconvexity is not only a boundary condition for estimates, but also the exact hypothesis that permits global $\bar\partial$ solvability in local $L^2$ spaces. [example: Solving the Equation With Gaussian Weight] Let $\Omega=\mathbb C^n$, let $\varphi(z)=|z|^2$, and let $f\in L^2_{(0,1)}(\mathbb C^n,e^{-|z|^2})$ satisfy $\bar\partial f=0$. We verify the curvature constant and then apply *Hörmander L2 Existence Theorem*. Since \begin{align*} \varphi(z)=|z|^2=\sum_{\ell=1}^n z_\ell\overline{z_\ell}, \end{align*} the product rule gives \begin{align*} \frac{\partial\varphi}{\partial\bar z_k} &=\sum_{\ell=1}^n \frac{\partial z_\ell}{\partial\bar z_k}\overline{z_\ell} +\sum_{\ell=1}^n z_\ell\frac{\partial\overline{z_\ell}}{\partial\bar z_k} \\ &=\sum_{\ell=1}^n 0\cdot\overline{z_\ell} +\sum_{\ell=1}^n z_\ell\delta_{\ell k} \\ &=z_k. \end{align*} Therefore \begin{align*} \varphi_{j\bar k} &=\frac{\partial}{\partial z_j}\left(\frac{\partial\varphi}{\partial\bar z_k}\right) \\ &=\frac{\partial z_k}{\partial z_j} \\ &=\delta_{jk}. \end{align*} For every $\xi=(\xi_1,\ldots,\xi_n)\in\mathbb C^n$, \begin{align*} \sum_{j,k=1}^n \varphi_{j\bar k}\xi_j\overline{\xi_k} &=\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 \\ &=|\xi|^2. \end{align*} Thus $i\partial\bar\partial\varphi\ge 1\cdot\omega$. The domain $\mathbb C^n$ is pseudoconvex, for instance because it is exhausted by the plurisubharmonic sublevel sets $\{z:|z|^2<R^2\}$. Applying *Hörmander L2 Existence Theorem* with $c=1$ gives a function $u\in L^2(\mathbb C^n,e^{-|z|^2})$ such that \begin{align*} \bar\partial u=f \end{align*} in the distributional sense and \begin{align*} \|u\|_{e^{-|z|^2}}^2 &\le \frac{1}{1}\|f\|_{e^{-|z|^2}}^2 \\ &=\|f\|_{e^{-|z|^2}}^2. \end{align*} In the Gaussian model, the identity curvature matrix gives exactly the unit solution bound. [/example] On bounded domains the same calculation gives a scaled estimate with the curvature constant visible in the final inequality. The unit ball example is useful because it separates two issues that are hidden on all of $\mathbb C^n$: the geometry of the domain and the size of the curvature constant. The ball is pseudoconvex, so the boundary sign is harmless, while scaling the Gaussian weight by $\lambda$ changes only the Hessian and hence changes the solution bound by exactly $1/\lambda$. This makes the estimate easy to read as a quantitative statement rather than just an existence statement. [example: Sharp Bound on the Unit Ball] [claim]Let $\Omega=B(0,1)\subset\mathbb C^n$, let $\lambda>0$, and set $\varphi(z)=\lambda |z|^2$. If $f\in L^2_{(0,1)}(B(0,1),e^{-\lambda |z|^2})$ satisfies $\bar\partial f=0$, then there is $u\in L^2(B(0,1),e^{-\lambda |z|^2})$ such that $\bar\partial u=f$ and \begin{align*} \|u\|_{e^{-\lambda |z|^2}}^2 \le \frac{1}{\lambda}\|f\|_{e^{-\lambda |z|^2}}^2. \end{align*}[/claim] [proof]First verify the two hypotheses entering *Hörmander L2 Existence Theorem*. For the unit ball, take the defining function \begin{align*} \rho(z)=|z|^2-1=\sum_{\ell=1}^n z_\ell\overline{z_\ell}-1. \end{align*} On $\partial B(0,1)$ we have $z\ne0$, so $d\rho\ne0$. Its complex Hessian is \begin{align*} \frac{\partial \rho}{\partial\bar z_k} &=\sum_{\ell=1}^n \frac{\partial z_\ell}{\partial\bar z_k}\overline{z_\ell} +\sum_{\ell=1}^n z_\ell\frac{\partial\overline{z_\ell}}{\partial\bar z_k} \\ &=\sum_{\ell=1}^n 0\cdot \overline{z_\ell} +\sum_{\ell=1}^n z_\ell\delta_{\ell k} \\ &=z_k, \end{align*} and therefore \begin{align*} \rho_{j\bar k} =\frac{\partial}{\partial z_j}(z_k) =\delta_{jk}. \end{align*} Thus for every $\xi\in\mathbb C^n$, \begin{align*} \sum_{j,k=1}^n \rho_{j\bar k}\xi_j\overline{\xi_k} &=\sum_{j,k=1}^n \delta_{jk}\xi_j\overline{\xi_k} \\ &=\sum_{j=1}^n |\xi_j|^2 \\ &=|\xi|^2\ge0. \end{align*} In particular, the Levi form is nonnegative on complex tangent vectors, so $B(0,1)$ is pseudoconvex. Now compute the curvature of the weight. Since \begin{align*} \varphi(z)=\lambda\sum_{\ell=1}^n z_\ell\overline{z_\ell}, \end{align*} we have \begin{align*} \frac{\partial\varphi}{\partial\bar z_k} &=\lambda\sum_{\ell=1}^n \frac{\partial z_\ell}{\partial\bar z_k}\overline{z_\ell} +\lambda\sum_{\ell=1}^n z_\ell\frac{\partial\overline{z_\ell}}{\partial\bar z_k} \\ &=\lambda\sum_{\ell=1}^n 0\cdot\overline{z_\ell} +\lambda\sum_{\ell=1}^n z_\ell\delta_{\ell k} \\ &=\lambda z_k. \end{align*} Hence \begin{align*} \varphi_{j\bar k} &=\frac{\partial}{\partial z_j}(\lambda z_k) \\ &=\lambda\delta_{jk}. \end{align*} For every $\xi=(\xi_1,\ldots,\xi_n)\in\mathbb C^n$, \begin{align*} \sum_{j,k=1}^n \varphi_{j\bar k}\xi_j\overline{\xi_k} &=\sum_{j,k=1}^n \lambda\delta_{jk}\xi_j\overline{\xi_k} \\ &=\lambda\sum_{j=1}^n \xi_j\overline{\xi_j} \\ &=\lambda\sum_{j=1}^n |\xi_j|^2 \\ &=\lambda|\xi|^2. \end{align*} Thus $i\partial\bar\partial\varphi\ge \lambda\omega$, so the curvature constant is $c=\lambda$. Applying *Hörmander L2 Existence Theorem* with this value of $c$ gives $u$ satisfying $\bar\partial u=f$ and \begin{align*} \|u\|_{e^{-\lambda |z|^2}}^2 &\le \frac{1}{c}\|f\|_{e^{-\lambda |z|^2}}^2 \\ &=\frac{1}{\lambda}\|f\|_{e^{-\lambda |z|^2}}^2. \end{align*}[/proof] The estimate records exactly how scaling the Gaussian weight changes the solution bound: multiplying the curvature by $\lambda$ divides the guaranteed $L^2$ norm bound by $\lambda$. [/example] These examples show how the theorem turns geometric positivity into quantitative solvability. The same mechanism also solves problems where the desired solution has prescribed poles or interpolation behaviour. ## Hilbert Space Mechanism and Approximation Why does an estimate for compactly supported smooth forms suffice on a noncompact or nonsmooth domain? The proof is organised so that analytic identities are first established where [integration by parts](/theorems/2098) is legitimate, then transported to the [Hilbert space](/page/Hilbert%20Space) closures. This separation keeps the PDE calculation local and the existence step functional-analytic. [quotetheorem:3693] [citeproof:3693] For the Hörmander theorem, $T=\bar\partial$ from functions to $(0,1)$-forms and $S=\bar\partial$ from $(0,1)$-forms to $(0,2)$-forms. The condition $S f=0$ is exactly the compatibility condition $\bar\partial f=0$. [explanation: Density and Exhaustion in the Proof] The integration-by-parts identity is first proved for smooth forms on smoothly bounded pseudoconvex domains. If $\Omega$ is an arbitrary pseudoconvex domain, choose an exhaustion by relatively compact domains and insert cutoff functions supported away from the boundary. The cutoffs allow the weighted identity to be applied locally, while the pseudoconvex exhaustion keeps the boundary contribution nonnegative at the approximating level. The passage to the limit uses the graph norm \begin{align*} \|v\|_{\mathrm{graph}}^2=\|v\|_{e^{-\varphi}}^2+\|\bar\partial v\|_{e^{-\varphi}}^2+\|\bar\partial^*_\varphi v\|_{e^{-\varphi}}^2. \end{align*} Smooth compactly supported forms are dense in the relevant graph domains after the standard Friedrichs mollifier argument in coordinate patches. [Lower semicontinuity of the norm](/theorems/215) preserves the estimate when the approximating forms converge. [/explanation] This density step is technical but conceptually important: the theorem is not a statement about smooth data. It gives weak $L^2$ solutions for all distributionally closed data in the weighted [Hilbert space](/page/Hilbert%20Space). [remark: Andreotti Vesentini Proof] Andreotti and Vesentini independently obtained the same existence theorem using a closely related [Hilbert space](/page/Hilbert%20Space) method. Their formulation emphasised complete Kähler metrics and made the result natural on complex manifolds, while the original Hörmander proof gave the [Euclidean domain](/page/Euclidean%20Domain) version used here. [/remark] The common core is the same: construct a positive curvature term, obtain a coercive estimate, and use [Hilbert space](/page/Hilbert%20Space) duality to solve the equation. ## Higher Degree Forms and Vanishing What changes when the datum is a $(0,q)$-form with $q>1$? The same identity holds, but the curvature term acts on all $q$ anti-holomorphic indices. Positivity becomes stronger as $q$ increases, so the theorem solves $\bar\partial u=f$ in every positive degree under the same strict plurisubharmonic lower bound. [quotetheorem:3694] [citeproof:3694] This stronger estimate leads to the higher-degree form of the existence theorem. The constant is often written using the smallest eigenvalue of the curvature operator on $(0,q)$-forms. The factor $q$ appears because a $(0,q)$-form has $q$ anti-holomorphic slots, and the curvature matrix acts once on each slot. A lower bound by $cI$ therefore contributes at least $c$ from each slot, producing $qc\|v\|_{e^{-\varphi}}^2$ rather than only $c\|v\|_{e^{-\varphi}}^2$. At $q=0$ there is no anti-holomorphic slot, so there is no curvature gain of this kind; estimates for functions require different input. For $q\ge 1$, higher degree is qualitatively easier because curvature has more components on which to act. [quotetheorem:3695] [citeproof:3695] The higher-degree theorem is only as strong as the chosen weight. A small lower curvature bound $c$ gives a large solution norm, and if the curvature is only semipositive then the argument loses the bounded inverse needed to force closed-range exactness. The constant $1/(qc)$ should be read as the sharp scale forced by the uniform lower curvature bound in this simplified setting: increasing either the degree or the curvature improves the bound linearly, while a weight whose least curvature eigenvalue is $c$ cannot give a better estimate by this argument. On strictly pseudoconvex domains, these estimates imply the vanishing of positive-degree $L^2$ Dolbeault cohomology with suitable weights. This is the analytic shadow of Cartan-type theorems for Stein domains. [quotetheorem:3696] [citeproof:3696] The word vanishing here means a cohomological group has no nonzero classes, not that the forms themselves are zero. The theorem says every closed form in positive degree can be solved away by one application of $\bar\partial$. [example: Mittag Leffler from Weighted Solvability] [claim]Let $A\subset\mathbb C$ be discrete, and for each $a\in A$ let \begin{align*} p_a(z)=\sum_{r=1}^{N_a} c_{a,r}(z-a)^{-r} \end{align*} be the prescribed principal part at $a$. Then weighted $\bar\partial$ solvability produces a [meromorphic function](/page/Meromorphic%20Function) on $\mathbb C$ whose principal part at each $a$ is $p_a$.[/claim] [proof]Choose radii $r_a>0$ so that the disks $D(a,2r_a)$ are pairwise disjoint, and choose $\chi_a\in C_c^\infty(D(a,2r_a))$ with $\chi_a=1$ on $D(a,r_a)$. On $\mathbb C\setminus A$ define \begin{align*} M(z)=\sum_{a\in A}\chi_a(z)p_a(z), \end{align*} where the sum is locally finite. For $z\in D(a,2r_a)\setminus\{a\}$, \begin{align*} \bar\partial(\chi_a p_a) &=\frac{\partial}{\partial\bar z}\bigl(\chi_a(z)p_a(z)\bigr)\,d\bar z \\ &=\left(\frac{\partial\chi_a}{\partial\bar z}(z)p_a(z) +\chi_a(z)\frac{\partial p_a}{\partial\bar z}(z)\right)d\bar z. \end{align*} Since each $(z-a)^{-r}$ is holomorphic on $D(a,2r_a)\setminus\{a\}$, \begin{align*} \frac{\partial p_a}{\partial\bar z}(z) &=\sum_{r=1}^{N_a}c_{a,r}\frac{\partial}{\partial\bar z}(z-a)^{-r} \\ &=0. \end{align*} Thus \begin{align*} f:=\bar\partial M =\sum_{a\in A}\frac{\partial\chi_a}{\partial\bar z}p_a\,d\bar z. \end{align*} The support of each summand lies in the annulus $D(a,2r_a)\setminus \overline{D(a,r_a)}$, so $f$ is smooth and vanishes near every point of $A$. In one complex variable, if $f=f_{\bar z}\,d\bar z$, then \begin{align*} \bar\partial f =\frac{\partial f_{\bar z}}{\partial\bar z}\,d\bar z\wedge d\bar z =0, \end{align*} because $d\bar z\wedge d\bar z=0$. Choose a smooth increasing convex function $\rho:[0,\infty)\to[0,\infty)$ growing fast enough that, for \begin{align*} \varphi(z)=|z|^2+\rho(|z|^2), \end{align*} we have \begin{align*} \int_{\mathbb C}|f(z)|^2e^{-\varphi(z)}\,d\mathcal L^2(z)<\infty. \end{align*} Its curvature is positive because \begin{align*} \frac{\partial\varphi}{\partial\bar z} &=z+\rho'(|z|^2)z, \\ \varphi_{z\bar z} &=\frac{\partial}{\partial z}\left((1+\rho'(|z|^2))z\right) \\ &=1+\rho'(|z|^2)+\rho''(|z|^2)|z|^2 \\ &\ge 1. \end{align*} Applying *Hörmander L2 Existence Theorem* on $\mathbb C$ gives $u\in L^2(\mathbb C,e^{-\varphi})$ with \begin{align*} \bar\partial u=f. \end{align*} Set \begin{align*} F=M-u \end{align*} on $\mathbb C\setminus A$. Then \begin{align*} \bar\partial F =\bar\partial M-\bar\partial u =f-f =0, \end{align*} so $F$ is holomorphic on $\mathbb C\setminus A$. Near a fixed $a\in A$, the function $M$ equals $p_a$ on $D(a,r_a)$, and $f=0$ there. Hence $\bar\partial u=0$ on $D(a,r_a)$. Since $u\in L^2_{\mathrm{loc}}$, the standard Weyl lemma for $\bar\partial$ makes $u$ holomorphic near $a$, so \begin{align*} u(z)=\sum_{s=0}^\infty b_s(z-a)^s \end{align*} for $z$ close to $a$. Therefore \begin{align*} F(z) &=p_a(z)-u(z) \\ &=\sum_{r=1}^{N_a}c_{a,r}(z-a)^{-r} -\sum_{s=0}^\infty b_s(z-a)^s. \end{align*} The negative-power part of the Laurent expansion of $F$ at $a$ is exactly $p_a$.[/proof] The weight supplies a global $L^2$ correction for the cutoff error, while the correction is holomorphic at the singular centres and therefore cannot change the prescribed principal parts. [/example] This application illustrates the guiding principle of the chapter. Instead of constructing integral kernels by hand, we encode the desired growth or singular behaviour into the weight and let the $L^2$ theorem solve the global correction problem. Chapter 4 proved that the ∂̄-equation is solvable on pseudoconvex domains under suitable weighted conditions; this chapter uses that solvability to solve extension problems. A [holomorphic function](/page/Holomorphic%20Function) is prescribed on a complex hypersurface and must be extended to the ambient space with controlled growth; the weight e^{−φ} encodes this boundary data into the norm. Extension reduces to an L² solvability question, with the weight machinery ensuring that the extended function has the desired properties. # 5. The Ohsawa–Takegoshi Extension Theorem This chapter turns the abstract $L^2$ solution theory for $\bar\partial$ into an [extension theorem](/theorems/59). Earlier chapters showed how plurisubharmonic weights give solvability with estimates on pseudoconvex domains; here the goal is to prescribe a [holomorphic function](/page/Holomorphic%20Function) on a complex hypersurface and extend it to the ambient domain without losing quantitative control. The surprising point is that the estimate depends only on the geometry encoded in pseudoconvexity and the weight, not on an explicit formula for the extension. The model hypersurface is the coordinate slice $H=\Omega\cap\{z_n=0\}$. The proof strategy is local in the normal variable but global in the tangential variables: build any smooth extension, measure its failure to be holomorphic, and then use a singular $\bar\partial$ estimate to correct the error while forcing the correction to vanish on $H$. ## Extending From A Hyperplane Slice The problem is to turn holomorphic data on a lower-dimensional complex submanifold into holomorphic data on the whole domain while preserving an $L^2$ bound. In one complex variable, boundary extension often relies on explicit kernels; in several variables, pseudoconvexity and $\bar\partial$ estimates replace those kernels. [definition: Weighted Bergman Space On A Domain] Let $\Omega\subset\mathbb C^n$ be a domain and let $\varphi:\Omega\to[-\infty,\infty)$ be plurisubharmonic. The weighted Bergman space $A^2(\Omega,e^{-\varphi})$ is \begin{align*} A^2(\Omega,e^{-\varphi})=\left\{F\in\mathcal O(\Omega):\int_\Omega |F|^2e^{-\varphi}\,d\mathcal L^{2n}<\infty\right\}. \end{align*} [/definition] For a hypersurface $H=\Omega\cap\{z_n=0\}$, the same notation means integration with respect to the induced Lebesgue measure on $H\simeq\mathbb C^{n-1}$. The theorem below says that restriction from $\Omega$ to $H$ has a bounded right inverse. [quotetheorem:3697] [citeproof:3697] The theorem is not merely qualitative. It says that the restriction map \begin{align*} A^2(\Omega,e^{-\varphi})\longrightarrow A^2(H,e^{-\varphi}),\qquad F\longmapsto F|_H, \end{align*} is onto and has controlled right inverse. This quantitative form is the feature that later applications use. In a working problem, the weight is usually chosen backwards from the desired estimate: make $|F|^2e^{-\varphi}$ the quantity to be bounded, then verify plurisubharmonicity by checking the Levi form or by writing $\varphi$ as a sum of known plurisubharmonic weights. Constants should be tracked through the normal coordinate, the measure convention on $H$, and the $\bar\partial$ estimate; none of these constants should depend on the prescribed function $f$. [example: Reinhardt Hyperplane Section] Let $\alpha_j\in\mathbb Z_{\ge0}$ and write \begin{align*} f(z')=z_1^{\alpha_1}\cdots z_{n-1}^{\alpha_{n-1}},\qquad z'=(z_1,\dots,z_{n-1}). \end{align*} The [extension theorem](/theorems/59) above guarantees some $L^2$ holomorphic extension; in this Reinhardt situation we can see one explicitly by taking \begin{align*} F(z',z_n)=f(z'). \end{align*} This function is holomorphic on $\Omega$ because it is a polynomial in the first $n-1$ variables, and \begin{align*} F(z',0)=f(z') \end{align*} for every $(z',0)\in H$. For each $z'\in H$, set \begin{align*} \Omega_{z'}=\{w\in\mathbb C:(z',w)\in\Omega\},\qquad R(z')=\sup\{|w|:w\in\Omega_{z'}\}. \end{align*} Since $\Omega$ is complete Reinhardt, $w\in\Omega_{z'}$ and $|\eta|\le |w|$ imply $\eta\in\Omega_{z'}$; since $\Omega$ is Reinhardt, the fibre is rotation-invariant. Hence $\Omega_{z'}$ is the disc $\{w:|w|<R(z')\}$, up to its boundary, and its planar measure is $\pi R(z')^2$. By [Fubini's theorem](/theorems/2961), \begin{align*} \int_\Omega |F(z',z_n)|^2\,d\mathcal L^{2n} &=\int_H\int_{\Omega_{z'}} |f(z')|^2\,d\mathcal L^2(z_n)\,d\mathcal L^{2n-2}(z')\\ &=\int_H |f(z')|^2\,\mathcal L^2(\Omega_{z'})\,d\mathcal L^{2n-2}(z')\\ &=\int_H |f(z')|^2\,\pi R(z')^2\,d\mathcal L^{2n-2}(z'). \end{align*} If $R_0=\sup_{\Omega}|z_n|<\infty$, then $R(z')\le R_0$ for every $z'\in H$, so \begin{align*} \int_\Omega |F|^2\,d\mathcal L^{2n} &\le \pi R_0^2\int_H |f(z')|^2\,d\mathcal L^{2n-2}\\ &=\pi R_0^2\int_H |z_1|^{2\alpha_1}\cdots |z_{n-1}|^{2\alpha_{n-1}}\,d\mathcal L^{2n-2}. \end{align*} The final integral is finite because $f\in A^2(H)$, so $F\in A^2(\Omega)$ and the extension estimate holds here with constant $\pi R_0^2$. This calculation isolates the normal fibre: the only extra cost of extending the monomial constantly in the $z_n$-direction is the area of the Reinhardt slice above each point of $H$. [/example] ## Correcting A Smooth Extension Once the desired boundary value is fixed, the obstruction is not the value on $H$ but holomorphicity away from $H$. The method is to start with a smooth object having the correct restriction and then remove exactly its $\bar\partial$-error. [definition: Smooth Preliminary Extension] Let $\mathcal O(H)$ denote the space of holomorphic functions on $H$. Let $f\in\mathcal O(H)$ and let $\chi\in C_c^\infty(\mathbb C)$ satisfy $\chi=1$ near $0$. A smooth preliminary extension of $f$ is a smooth map $\widetilde F:\Omega\to\mathbb C$ which locally has the form \begin{align*} \widetilde F(z',z_n)=\chi(z_n)f(z') \end{align*} near the hyperplane coordinates $(z',z_n)$. [/definition] The preliminary extension has the correct trace because $\chi(0)=1$. Its failure to be holomorphic is measured by \begin{align*} \bar\partial\widetilde F=f(z')\frac{\partial\chi}{\partial\bar z_n}(z_n)\,d\bar z_n, \end{align*} so the error lives in a thin annulus in the normal direction. The remaining problem is not to build another extension, but to correct this annular $\bar\partial$-error without changing the trace on $H$. That requires a solution $u$ whose restriction to the hyperplane vanishes, so that $\widetilde F-u$ is both holomorphic and still equal to $f$ along $H$. [quotetheorem:3698] [citeproof:3698] The entire [extension theorem](/theorems/59) is therefore reduced to a refined solvability statement: solve the $\bar\partial$ equation in a norm strong enough to force vanishing on $H$. The trace hypothesis is essential, since without $u|_H=0$ the correction could change the prescribed value even though it removes the $\bar\partial$-error; the principle by itself does not produce such a $u$ or any estimate. Pseudoconvexity enters precisely in the next step, where Hörmander solvability supplies the correction, and the unit ball divisor example shows how the abstract reduction is used in a concrete coordinate model before the full divisor language is introduced. [example: Divisor In The Unit Ball] Let $\Omega=B(0,1)\subset\mathbb C^n$, let $H=\Omega\cap\{z_n=0\}$, and let $\varphi=0$. Choose $0<r<1$ and $\chi\in C_c^\infty(\mathbb C)$ with $\chi(w)=1$ for $|w|<r/2$ and $\operatorname{supp}\chi\subset\{|w|<r\}$. For $f\in A^2(H)$, set \begin{align*} \widetilde F(z',z_n)=\chi(z_n)f(z'). \end{align*} Since $f$ is holomorphic in the $z'$ variables and $\chi$ depends only on $z_n$, \begin{align*} \frac{\partial\widetilde F}{\partial\bar z_j}(z',z_n) &=\chi(z_n)\frac{\partial f}{\partial\bar z_j}(z')=0\qquad 1\le j\le n-1,\\ \frac{\partial\widetilde F}{\partial\bar z_n}(z',z_n) &=f(z')\frac{\partial\chi}{\partial\bar w}(z_n). \end{align*} Thus \begin{align*} \bar\partial\widetilde F =f(z')\frac{\partial\chi}{\partial\bar w}(z_n)\,d\bar z_n. \end{align*} Because $\chi$ is constant on $|w|<r/2$ and vanishes for $|w|\ge r$, this error is supported in the annulus $r/2\le |z_n|\le r$, which is disjoint from $H$. Let $M_\chi=\sup_{\mathbb C}|\partial\chi/\partial\bar w|$. For $z'\in H$, write \begin{align*} \Omega_{z'}=\{w\in\mathbb C:|z'|^2+|w|^2<1\}. \end{align*} By [Fubini's theorem](/theorems/2961), \begin{align*} \int_\Omega |\bar\partial\widetilde F|^2|z_n|^{-2}\,d\mathcal L^{2n} &=\int_H |f(z')|^2\int_{\Omega_{z'}}\left|\frac{\partial\chi}{\partial\bar w}(w)\right|^2|w|^{-2}\,d\mathcal L^2(w)\,d\mathcal L^{2n-2}(z')\\ &\le \int_H |f(z')|^2\int_{r/2\le |w|\le r} M_\chi^2|w|^{-2}\,d\mathcal L^2(w)\,d\mathcal L^{2n-2}(z')\\ &\le \frac{4M_\chi^2}{r^2}\int_H |f(z')|^2\,d\mathcal L^{2n-2}(z')\int_{r/2\le |w|\le r}d\mathcal L^2(w)\\ &=\frac{4M_\chi^2}{r^2}\|f\|_{A^2(H)}^2\cdot \pi\left(r^2-\frac{r^2}{4}\right)\\ &=3\pi M_\chi^2\|f\|_{A^2(H)}^2. \end{align*} So the singular factor $|z_n|^{-2}$ does not create a singular estimate for the cut-off error, because the error lives where $|z_n|\ge r/2$. Now suppose a correction $u$ satisfies $\bar\partial u=\bar\partial\widetilde F$ and \begin{align*} \int_\Omega |u|^2|z_n|^{-2}\,d\mathcal L^{2n}<\infty. \end{align*} On the tube $\Omega\cap\{|z_n|<r/2\}$ we have $\bar\partial\widetilde F=0$, hence $\bar\partial u=0$ there and $u$ is holomorphic in that tube. If $u(p',0)=c\ne0$ at some point of $H$, continuity gives $\varepsilon,\rho>0$ such that $B'(p',\varepsilon)\times\{|w|<\rho\}\subset\Omega\cap\{|z_n|<r/2\}$ and $|u(z',w)|\ge |c|/2$ on this product set. Then \begin{align*} \int_\Omega |u|^2|z_n|^{-2}\,d\mathcal L^{2n} &\ge \int_{B'(p',\varepsilon)}\int_{|w|<\rho}\frac{|c|^2}{4}|w|^{-2}\,d\mathcal L^2(w)\,d\mathcal L^{2n-2}(z')\\ &=\frac{|c|^2}{4}\mathcal L^{2n-2}(B'(p',\varepsilon))\int_0^\rho\int_0^{2\pi}s^{-2}s\,d\theta\,ds\\ &=\frac{\pi|c|^2}{2}\mathcal L^{2n-2}(B'(p',\varepsilon))\int_0^\rho \frac{ds}{s}\\ &=\infty, \end{align*} contradicting the finite singular norm. Therefore $u|_H=0$, and the corrected function $F=\widetilde F-u$ satisfies \begin{align*} \bar\partial F=\bar\partial\widetilde F-\bar\partial u=0,\qquad F(z',0)=\chi(0)f(z')-u(z',0)=f(z'). \end{align*} The same weight therefore has two different roles: on the annulus supporting the cut-off error it is bounded by $4r^{-2}$, while at $H$ its radial integral diverges and forces the correction to vanish. [/example] ## Why The Singular Weight Forces Vanishing The key question is why a weighted estimate can impose a pointwise trace condition. The answer is that the weight has a non-integrable pole along the hypersurface, and finite weighted norm then forces the correction to contain a compensating factor of the defining function. [definition: Logarithmic Pole Along The Slice] Let $H=\Omega\cap\{z_n=0\}$. The logarithmic pole weight along $H$ is the function $\psi:\Omega\to[-\infty,\infty)$ defined by \begin{align*} \psi(z)=\log|z_n|^2, \end{align*} with $\psi(z)=-\infty$ on $H$. The function $\psi$ is plurisubharmonic on $\Omega$ and smooth on $\Omega\setminus H$. [/definition] In the extension proof, Hörmander's estimate is applied with the modified weight $\varphi-\psi=\varphi-\log|z_n|^2$. Since $\log|z_n|^2$ is plurisubharmonic, the modified weight has stronger growth in the norm through the factor $|z_n|^{-2}$. This singular factor is introduced to force the correction term to vanish on the slice. The precise local statement needed is that a holomorphic function with finite $|z_n|^{-2}$ weighted mass cannot carry a nonzero trace along $H$, which is the analytic substitute for imposing a boundary condition on the correction. What remains to justify is that this heuristic survives without assuming a power-series expansion or a smooth boundary value on the slice. The needed lemma turns the weighted $L^2$ condition itself into a trace-vanishing conclusion, so it can be inserted directly into the $ar\partial$ correction argument. [quotetheorem:3699] [citeproof:3699] This is the analytic heart of the proof. The singularity is calibrated to the codimension-one geometry: it is strong enough to kill the trace, but it still fits the curvature inequalities needed for the $\bar\partial$ estimate after regularisation. If the pole were weaker, for instance $|z_n|^{-\alpha}$ with $\alpha<2$, the argument would not force vanishing: take $u\equiv 1$ on the unit polydisc and integrate $|z_n|^{-\alpha}$ over the punctured disc $0<|z_n|<1$, getting $2\pi\int_0^1 r^{1-\alpha}\,dr=2\pi/(2-\alpha)<\infty$ whenever $\alpha<2$. So finiteness of the weighted norm is compatible with the constant trace $1$, and the trace would not be forced to vanish. The borderline exponent $\alpha=2$ is exactly the one for which the radial integral $\int_0^1 r^{-1}\,dr$ diverges, which is why $\log|z_n|^2$ is the unique scale-invariant singular weight that produces the trace condition. For higher-codimension subvarieties, the singularity has to involve all normal defining functions, because a single logarithmic pole controls only one normal direction. [illustration:scv-iii-cutoff-normal-annulus] [explanation: Cauchy Formula And The Normal Direction] The same phenomenon can be read through the one-variable Cauchy formula. For a [holomorphic function](/page/Holomorphic%20Function) $g$ on a disc, the value $g(0)$ is recovered by averaging over circles: \begin{align*} g(0)=\frac{1}{2\pi}\int_0^{2\pi}g(re^{i\theta})\,d\theta. \end{align*} If $g$ has finite integral against $|w|^{-2}\,d\mathcal L^2(w)$ near $0$, then the radial integral of the circular means must be finite with measure $dr/r$. A nonzero value at the centre would force these means to stay bounded away from zero along small radii, producing a logarithmic divergence. The $n$-variable argument is this one-dimensional calculation performed in the normal coordinate and then integrated over tangential variables. [/explanation] ## Sharp Constants And Later Refinements After existence is known, the next problem is whether the extension constant is an artefact of the proof. The development from Ohsawa and Takegoshi through Berndtsson, Błocki, and Guan-Zhou shows that the optimal constant is part of the theorem's structure. [quotetheorem:3700] The derivation of the optimal constant theorem is not part of this course. Sharpness matters because many applications compare extremal quantities, such as Bergman kernels, where an unspecified constant gives only a coarse inequality rather than the exact curvature or monotonicity statement. This exact form is one route to results such as the Suita conjecture and to positivity arguments for direct image bundles; in these problems losing a constant can destroy the final equality or optimal lower bound. Historically, Berndtsson obtained sharp forms using positivity methods, Błocki gave a simplified argument for important cases, and Guan-Zhou proved the optimal constant theorem in full generality in 2015. The optimal value $\pi$ is attained, not merely approached, in a concrete model case: take $\Omega=\mathbb D\times\mathbb D\subset\mathbb C^2$ to be the bidisc with $H=\mathbb D\times\{0\}$, weight $\varphi\equiv 0$, and boundary data $f\equiv 1$. The unique extension minimising the left-hand side is $F\equiv 1$, for which the ambient $L^2$ norm equals $\pi$ times the slice norm because the Lebesgue measure of the unit disc in the $z_2$-direction is exactly $\pi$. Any constant smaller than $\pi$ would fail in this example, so $\pi$ cannot be lowered uniformly. The same calculation in a half-plane or in a polydisc shows that the constant is intrinsic to the codimension-one geometry rather than an artefact of the proof. [remark: Normalisation Of Constants] Different sources state the sharp theorem with different numerical constants because they normalise Lebesgue measure, the defining function of $H$, and the curvature inequality differently. The invariant content is the existence of a universal best constant once those conventions are fixed. [/remark] With sharpness in place, the natural next question is how this several-variable extension principle compares to the one-variable extension theorems familiar from classical complex analysis. The example below contrasts the Ohsawa-Takegoshi viewpoint with a model [extension theorem](/theorems/59) on the disc to highlight what the $L^2$ framework adds. [example: Rudin-Carleson Type Extension] For a [closed set](/page/Closed%20Set) $E\subset\partial\mathbb D$ of arclength measure $0$, the *Rudin-Carleson theorem* says that every $h\in C(E)$ is the restriction to $E$ of some $G\in A(\mathbb D)$. This is not an $L^2$ estimate with respect to arclength on $E$: since $m_{\partial\mathbb D}(E)=0$, \begin{align*} \int_E |h|^2\,dm_{\partial\mathbb D}=0 \end{align*} for every continuous $h$ on $E$. If an estimate of the form \begin{align*} \int_{\mathbb D}|G|^2\,d\mathcal L^2\le C\int_E |h|^2\,dm_{\partial\mathbb D} \end{align*} held for all data, then the right-hand side would be $0$, so $\int_{\mathbb D}|G|^2\,d\mathcal L^2=0$. A continuous [holomorphic function](/page/Holomorphic%20Function) with zero area integral must be identically zero, because if $G(p)\ne0$, continuity gives a small disc $B(p,\rho)\subset\mathbb D$ on which $|G|\ge |G(p)|/2$, and then \begin{align*} \int_{\mathbb D}|G|^2\,d\mathcal L^2 &\ge \int_{B(p,\rho)}\frac{|G(p)|^2}{4}\,d\mathcal L^2\\ &=\frac{|G(p)|^2}{4}\pi\rho^2>0. \end{align*} Thus a nonzero boundary datum cannot be controlled by the zero arclength measure of $E$. The $L^2$ extension model replaces the boundary set by an analytic slice carrying its induced measure. Take $\Omega=\mathbb D\times\mathbb D\subset\mathbb C^2$, $H=\mathbb D\times\{0\}$, and $f\in A^2(H)$, identified with a [holomorphic function](/page/Holomorphic%20Function) on $\mathbb D$. Define \begin{align*} F(z,w)=f(z). \end{align*} Then $F(z,0)=f(z)$, and by *[Fubini's theorem](/theorems/2961)*, \begin{align*} \int_{\mathbb D\times\mathbb D}|F(z,w)|^2\,d\mathcal L^4(z,w) &=\int_{\mathbb D}\int_{\mathbb D}|f(z)|^2\,d\mathcal L^2(w)\,d\mathcal L^2(z)\\ &=\int_{\mathbb D}|f(z)|^2\left(\int_{\mathbb D}1\,d\mathcal L^2(w)\right)d\mathcal L^2(z)\\ &=\int_{\mathbb D}|f(z)|^2\left(\int_0^1\int_0^{2\pi}r\,d\theta\,dr\right)d\mathcal L^2(z)\\ &=\int_{\mathbb D}|f(z)|^2\left(2\pi\int_0^1 r\,dr\right)d\mathcal L^2(z)\\ &=\pi\int_{\mathbb D}|f(z)|^2\,d\mathcal L^2(z)\\ &=\pi\int_H |f|^2\,d\mathcal L^2_H. \end{align*} The shared principle is that small sets can carry prescribed holomorphic data, but the $L^2$ hypersurface theorem is quantitative: the analytic slice has an intrinsic measure, and the cost of the constant [normal extension](/page/Normal%20Extension) is exactly the area $\pi$ of the normal unit disc. [/example] ## Extension From Divisors The hyperplane theorem is the local model for extension from a divisor. The question is how to replace the coordinate function $z_n$ by a holomorphic defining section while keeping the same singular-weight mechanism. [definition: Smooth Divisor With Defining Function] Let $\Omega\subset\mathbb C^n$ be a domain. A smooth divisor $D\subset\Omega$ with defining function $s\in\mathcal O(\Omega)$ is a smooth complex hypersurface such that \begin{align*} D=\{z\in\Omega:s(z)=0\},\qquad ds_z\ne0\quad\text{for all }z\in D. \end{align*} [/definition] Near each point of $D$, the holomorphic [implicit function theorem](/page/Implicit%20Function%20Theorem) gives coordinates in which $s$ is a nonvanishing factor times $z_n$. Passing from a coordinate slice to a divisor requires a theorem whose hypotheses are invariant under changing defining functions. The extension estimate must know not only that $s$ cuts out $D$, but also how the induced hypersurface measure and the ambient weight transform. This is the point at which the local trace-killing lemma becomes a global extension statement. The global divisor theorem must package these local normal coordinates into a coordinate-free estimate. Replacing $\log|z_n|^2$ by $\log|s|^2$ keeps the same trace-killing pole, but the statement also has to track the metric constants that compare the defining function, the hypersurface measure, and the ambient weight. The obstruction is that local extensions constructed in different charts need not agree, and their estimates must remain stable under changes of defining function. The divisor form supplies the invariant formulation that lets the slice argument be used on an arbitrary smooth hypersurface. [quotetheorem:3701] [citeproof:3701] This divisor form is the version used most often in geometry. It allows sections defined on a hypersurface to be extended while keeping enough metric control to feed into positivity arguments, for instance in the extension of pluricanonical sections and in Siu's invariance of plurigenera. The smoothness condition $ds\ne0$ is not cosmetic: if $ds$ vanishes, the normal coordinate used to measure the pole breaks down and the hypersurface measure no longer captures the correct local multiplicity. Singular divisors require multiplier ideals or more refined residue measures, so the clean statement above should be read as the smooth model rather than a theorem for arbitrary analytic sets. [example: Ball Divisor With Linear Defining Function] Let $\Omega=B(0,1)\subset\mathbb C^n$ and let \begin{align*} D=\Omega\cap\{s(z)=0\},\qquad s(z)=a_1z_1+\cdots+a_nz_n, \end{align*} where $a=(a_1,\dots,a_n)\in\mathbb C^n$ and $|a|^2=\sum_{j=1}^n |a_j|^2=1$. Since the row vector $a$ has length $1$, it can be completed to a unitary matrix $U\in U(n)$ whose last row is $a$. For $y=Uz$, the last coordinate is \begin{align*} y_n &=\sum_{j=1}^n U_{nj}z_j\\ &=\sum_{j=1}^n a_jz_j\\ &=s(z). \end{align*} Thus \begin{align*} z\in D &\Longleftrightarrow z\in B(0,1)\ \text{and}\ s(z)=0\\ &\Longleftrightarrow Uz\in B(0,1)\ \text{and}\ (Uz)_n=0, \end{align*} because $|Uz|=|z|$ for a unitary matrix. Hence $U(D)=B(0,1)\cap\{y_n=0\}$. Let $\varphi$ be plurisubharmonic on $B(0,1)$ and let $f\in A^2(D,e^{-\varphi})$. Define \begin{align*} \widetilde\varphi(y)=\varphi(U^{-1}y),\qquad g(y',0)=f(U^{-1}(y',0)). \end{align*} The function $\widetilde\varphi$ is plurisubharmonic because composing with the complex-linear biholomorphism $U^{-1}$ sends complex lines to complex lines. Also $g$ is holomorphic on $B(0,1)\cap\{y_n=0\}$ because $f$ is holomorphic on $D$ and $U^{-1}$ is holomorphic. Applying the *[Ohsawa-Takegoshi Extension Theorem](/theorems/3697)* in the $y$-coordinates gives a holomorphic $G$ on $B(0,1)$ with $G|_{\{y_n=0\}}=g$ and \begin{align*} \int_{B(0,1)} |G(y)|^2e^{-\widetilde\varphi(y)}\,d\mathcal L^{2n}(y) \le C\int_{B(0,1)\cap\{y_n=0\}} |g(y')|^2e^{-\widetilde\varphi(y',0)}\,d\mathcal L^{2n-2}(y'). \end{align*} Set $F(z)=G(Uz)$. If $z\in D$, then $(Uz)_n=0$, so \begin{align*} F(z) &=G(Uz)\\ &=g(Uz)\\ &=f(U^{-1}Uz)\\ &=f(z). \end{align*} Thus $F|_D=f$. The estimate keeps the same constant. Since $U$ is unitary, its real Jacobian has absolute determinant $|\det_{\mathbb C}U|^2=1$, so $d\mathcal L^{2n}(Uz)=d\mathcal L^{2n}(z)$. Therefore \begin{align*} \int_{B(0,1)} |F(z)|^2e^{-\varphi(z)}\,d\mathcal L^{2n}(z) &=\int_{B(0,1)} |G(Uz)|^2e^{-\varphi(z)}\,d\mathcal L^{2n}(z)\\ &=\int_{B(0,1)} |G(y)|^2e^{-\varphi(U^{-1}y)}\,d\mathcal L^{2n}(y)\\ &=\int_{B(0,1)} |G(y)|^2e^{-\widetilde\varphi(y)}\,d\mathcal L^{2n}(y). \end{align*} The restriction of $U$ maps $D$ isometrically onto $B(0,1)\cap\{y_n=0\}$, so the induced hypersurface measure is also preserved, and \begin{align*} \int_{B(0,1)\cap\{y_n=0\}} |g(y')|^2e^{-\widetilde\varphi(y',0)}\,d\mathcal L^{2n-2}(y') &=\int_D |f(z)|^2e^{-\varphi(z)}\,d\mathcal H^{2n-2}(z). \end{align*} Combining the displayed inequalities gives \begin{align*} \int_{B(0,1)} |F|^2e^{-\varphi}\,d\mathcal L^{2n} \le C\int_D |f|^2e^{-\varphi}\,d\mathcal H^{2n-2}. \end{align*} So a linear divisor in the ball is not a new geometric case: after a unitary rotation, it is exactly the coordinate hyperplane slice, and the extension constant is unchanged. [/example] The chapter's main lesson is that $L^2$ methods do more than solve $\bar\partial$: by choosing a weight with the right singularity, they encode interpolation conditions into the norm itself. The Ohsawa-Takegoshi theorem is the prototype for this technique, and later applications use it to produce holomorphic functions, compare Bergman kernels, and prove positivity results for families of complex manifolds. Chapters 4–5 developed existence and extension theorems for solving ∂̄; later chapters will return to cohomological vanishing applications. This chapter shifts attention from the equation to the space of holomorphic functions itself—the Bergman space of L²-integrable holomorphic functions. The Bergman kernel, the reproducing kernel of this [Hilbert space](/page/Hilbert%20Space), encodes both domain geometry and weight geometry; its asymptotics reveal how L² estimates produce explicit geometric information. # 6. Bergman Spaces, Kernels, and the $\bar{\partial}$ Connection Chapters 4 and 5 developed $L^2$ estimates for the operator $\bar{\partial}$ and used them to solve analytic extension problems; the vanishing applications return in Chapter 8. This chapter turns the same estimates toward holomorphic functions themselves. The Bergman space is a [Hilbert space](/page/Hilbert%20Space) of square-integrable holomorphic functions, and its reproducing kernel records how point evaluation, biholomorphic change of variables, and intrinsic geometry interact. The main theme is that $\bar{\partial}$ solvability supplies holomorphic functions with controlled $L^2$ norm, while the Bergman kernel packages the optimal such control. ## The Hilbert Space of Square-Integrable Holomorphic Functions Which holomorphic functions remain visible to [Hilbert space](/page/Hilbert%20Space) methods? A [holomorphic function](/page/Holomorphic%20Function) may have strong boundary growth, and the condition of square integrability separates functions whose analytic behavior can be measured by an $L^2$ norm from those that sit outside the Hilbert framework. [definition: Bergman Space] Let $\Omega \subset \mathbb C^n$ be a domain and let $dV=d\mathcal L^{2n}$. Define \begin{align*} A^2(\Omega)&:=\mathcal O(\Omega)\cap L^2(\Omega,dV),\\ (f,g)_{A^2(\Omega)}&:=\int_\Omega f(z)\overline{g(z)}\,dV(z),\\ \|f\|_{A^2(\Omega)}^2&:=(f,f)_{A^2(\Omega)}. \end{align*} [/definition] The notation $\mathcal O(\Omega)$ denotes holomorphic functions on $\Omega$. When $\Omega$ has finite volume, every bounded [holomorphic function](/page/Holomorphic%20Function) belongs to $A^2(\Omega)$; when $\Omega$ has infinite volume, even constant functions may be excluded. The course uses this distinction to keep track of which constructions depend only on local holomorphicity and which require global integrability. [example: Constants on Finite and Infinite Volume Domains] Let $\Delta\subset\mathbb C$ be the unit disc and let $\mathbb H=\{z=x+iy\in\mathbb C:y>0\}$. The constant function $1$ is holomorphic on both domains, so membership in the Bergman space is exactly the question of whether its $L^2$ norm is finite. On $\Delta$, using polar coordinates $z=re^{i\theta}$ and $dA=r\,dr\,d\theta$, \begin{align*} \|1\|_{A^2(\Delta)}^2 &=\int_\Delta |1|^2\,dA\\ &=\int_0^{2\pi}\int_0^1 1\cdot r\,dr\,d\theta\\ &=\int_0^{2\pi}\left[\frac{r^2}{2}\right]_{0}^{1}\,d\theta\\ &=\int_0^{2\pi}\frac12\,d\theta\\ &=\pi. \end{align*} Thus $1\in L^2(\Delta,dA)$, hence $1\in A^2(\Delta)$. On $\mathbb H$, for every $R>0$ the rectangle $Q_R=\{x+iy:0<x<R,\ 0<y<R\}$ is contained in $\mathbb H$, and therefore \begin{align*} \|1\|_{A^2(\mathbb H)}^2 &=\int_{\mathbb H}|1|^2\,dA\\ &\ge \int_{0}^{R}\int_{0}^{R}1\,dy\,dx\\ &=\int_0^R R\,dx\\ &=R^2. \end{align*} Since $R^2$ can be made arbitrarily large, $\int_{\mathbb H}|1|^2\,dA=+\infty$, so $1\notin A^2(\mathbb H)$. Thus the Bergman space detects the global volume of the domain, not only its local complex structure. [/example] The first structural question is whether $A^2(\Omega)$ is complete. The answer is not automatic from the definition, because $A^2(\Omega)$ is a holomorphic subspace sitting inside the larger Hilbert space $L^2(\Omega,dV)$, and a subspace is a Hilbert space only when it is closed. Completeness is needed before [Riesz representation](/theorems/67) can convert bounded point evaluations into kernel functions. [quotetheorem:3702] [citeproof:3702] The point of the proof is that $A^2(\Omega)$ is not complete merely because it is a collection of holomorphic functions; a general linear subspace of $L^2$ can be dense and non-closed, so its Cauchy sequences may converge to functions outside the subspace. The mean-value inequality is the analytic input that prevents this: it turns $L^2$ control on compactly contained balls into uniform control on compact sets, so holomorphicity survives in the limit rather than being lost in an a.e. representative. Completeness then gives access not only to [Riesz representation](/theorems/67) but also to [orthogonal projection](/theorems/437) from $L^2(\Omega,dV)$ onto $A^2(\Omega)$, which later becomes the Bergman projection. The second structural question is whether evaluating a function at a point is continuous in the $A^2$ norm. Holomorphicity supplies local averaging, so point values are controlled by nearby $L^2$ mass. [quotetheorem:3703] [citeproof:3703] Point evaluation is a genuinely holomorphic phenomenon here. For a general $L^2$ function, the value at a fixed point is not well-defined because functions are identified up to a.e. equality, and changing a representative at one point does not change the $L^2$ class. The proof uses the plurisubharmonicity of $|f|^2$ to replace pointwise control by average control on a ball around $a$; the constant depends on the radius of such a ball, hence on the distance from $a$ to $\partial\Omega$. As $a$ approaches the boundary, this estimate may deteriorate, which is exactly the behavior later recorded by the growth of $K_\Omega(a,a)$. Bounded point evaluation lets the [Hilbert space](/page/Hilbert%20Space) represent evaluation by an inner product. The next object records all of those representing vectors at once. Instead of treating each evaluation functional separately, the Bergman kernel packages them into a two-variable function whose diagonal measures the size of point evaluation and whose off-diagonal values encode reproducing behavior. [definition: Bergman Kernel] For $a\in\Omega$, let $k_a\in A^2(\Omega)$ be the Riesz vector satisfying \begin{align*} f(a)=(f,k_a)_{A^2(\Omega)} \end{align*} for all $f\in A^2(\Omega)$. The Bergman kernel is \begin{align*} K_\Omega(z,w):=\overline{k_z(w)}. \end{align*} [/definition] The convention above matches the [Hilbert space](/page/Hilbert%20Space) convention that inner products are linear in the first argument. Before using the kernel for estimates, we need its basic structural identities. These identities say that the definition really produces a Hermitian reproducing kernel: it recovers function values, has the expected holomorphic and anti-holomorphic dependence, and has a positive diagonal wherever point evaluation is nontrivial. [quotetheorem:3704] [citeproof:3704] The convention on the inner product determines the mixed holomorphic behavior of the kernel. Because $K_\Omega(z,w)=k_w(z)$ by the reproducing vector identity, it is holomorphic in $z$ for fixed $w$; because $K_\Omega(z,w)=\overline{k_z(w)}$, it is anti-holomorphic in $w$ for fixed $z$. If $K_\Omega(z,w)=0$ for a fixed pair, then the vectors representing evaluation at $z$ and $w$ are orthogonal, so the vanishing records a Hilbert-space relation between two point constraints. The Hermitian symmetry is what makes diagonal quantities real and positive and is used repeatedly when passing from diagonal estimates to full two-variable kernel estimates by polarisation. The diagonal value $K_\Omega(z,z)$ measures the squared norm of point evaluation. If $K_\Omega(z,z)$ is large, then some unit $A^2$ holomorphic functions can have large value at $z$. ## Producing Kernel Functions from $\bar{\partial}$ Estimates How does the $\bar{\partial}$ theory from the previous chapters create Bergman kernel information? The kernel exists by [Hilbert space](/page/Hilbert%20Space) arguments, but $\bar{\partial}$ estimates explain why it is nonzero, why it detects local jets, and why it has boundary meaning on pseudoconvex domains. [quotetheorem:3705] [citeproof:3705] Taking $P=1$ shows that bounded pseudoconvex domains have enough square-integrable holomorphic functions to separate each point from zero. Combined with the reproducing formula, this gives strict positivity of the kernel on the diagonal. Both hypotheses in the theorem are used essentially. Boundedness is needed because unbounded domains can fail to support any nonzero element of $A^2$ at all: on $\Omega=\mathbb C^n$ every nonconstant entire function of polynomial growth is excluded by infinite mass on dyadic shells, and a Liouville-type argument rules out all nonzero $A^2$ functions, so $A^2(\mathbb C^n)=\{0\}$ and no jet can be realised. Pseudoconvexity is needed because the construction relies on the Hörmander $L^2$ estimate for $\bar{\partial}$, which is available precisely when the ambient domain admits a strictly plurisubharmonic exhaustion with the required curvature lower bound. Realising a full jet of order $m$ is also strictly stronger than knowing that some $A^2$ function is nonzero at $a$: the kernel diagonal $K_\Omega(a,a)>0$ records only the zeroth-order extremal value, while jet realisation supplies a finite-dimensional family of $A^2$ functions whose Taylor expansions span all polynomials of degree at most $m$ at $a$. The latter is what controls higher-order kernel derivatives and feeds into the Bergman metric. To compare kernels, however, it is better to replace the reproducing vector by an extremal problem. The upcoming formula identifies the diagonal kernel with the largest possible point value among unit-norm holomorphic functions, turning kernel estimates into concrete function-construction problems. [quotetheorem:3706] [citeproof:3706] This extremal form is often the most useful way to compare kernels under domain variation. Enlarging a domain weakens the $L^2$ condition while increasing the admissible class of functions, and shrinking a domain has the opposite effect; the competition between these effects is central in Ramadanov convergence. Concretely, if $\Omega \subseteq \Omega'$ and $a\in\Omega$, then restriction $A^2(\Omega')\to A^2(\Omega)$ is norm-decreasing on the image, so the extremal supremum increases on the larger domain at points where $A^2$ separates from zero; this gives the monotonicity $K_{\Omega'}(a,a)\le K_\Omega(a,a)$ for $\Omega\subseteq\Omega'$. The hypothesis $K_\Omega(a,a)>0$ is genuine: if $A^2(\Omega)=\{0\}$, as happens for $\Omega=\mathbb C^n$, the diagonal kernel is identically zero and the supremum formula is vacuous because the constraint set $\{\|f\|_{A^2}\le 1\}$ contains only the zero function. The extremal form is also what converts kernel comparison into a function-existence problem: instead of comparing two reproducing kernels directly, the problem reduces to constructing one $A^2$ function with a controlled value at $a$, which is precisely the kind of task that $\bar{\partial}$ estimates solve. [example: Kernel as Optimal Point Control] Fix $a\in\Omega$ and assume $K_\Omega(a,a)>0$. For any $f\in A^2(\Omega)$ with $\|f\|_{A^2(\Omega)}\le 1$, the reproducing vector identity gives \begin{align*} |f(a)| &=|(f,k_a)_{A^2(\Omega)}|\\ &\le \|f\|_{A^2(\Omega)}\|k_a\|_{A^2(\Omega)} \end{align*} by the *[Cauchy-Schwarz inequality](/theorems/432)* in the [Hilbert space](/page/Hilbert%20Space) $A^2(\Omega)$. Since $K_\Omega(a,a)=\|k_a\|_{A^2(\Omega)}^2$, we have $\|k_a\|_{A^2(\Omega)}=K_\Omega(a,a)^{1/2}$, and therefore \begin{align*} |f(a)|\le K_\Omega(a,a)^{1/2}. \end{align*} Now set \begin{align*} e_a:=\frac{k_a}{K_\Omega(a,a)^{1/2}}. \end{align*} Then \begin{align*} \|e_a\|_{A^2(\Omega)}^2 &=\left(\frac{k_a}{K_\Omega(a,a)^{1/2}},\frac{k_a}{K_\Omega(a,a)^{1/2}}\right)_{A^2(\Omega)}\\ &=\frac{1}{K_\Omega(a,a)}\|k_a\|_{A^2(\Omega)}^2\\ &=\frac{K_\Omega(a,a)}{K_\Omega(a,a)}\\ &=1, \end{align*} so $e_a$ is admissible. Its value at $a$ is \begin{align*} e_a(a) &=(e_a,k_a)_{A^2(\Omega)}\\ &=\left(\frac{k_a}{K_\Omega(a,a)^{1/2}},k_a\right)_{A^2(\Omega)}\\ &=\frac{1}{K_\Omega(a,a)^{1/2}}\|k_a\|_{A^2(\Omega)}^2\\ &=\frac{K_\Omega(a,a)}{K_\Omega(a,a)^{1/2}}\\ &=K_\Omega(a,a)^{1/2}. \end{align*} Thus the largest possible value of $|f(a)|$ on the unit ball of $A^2(\Omega)$ is exactly $K_\Omega(a,a)^{1/2}$, attained by the normalised kernel vector $k_a/K_\Omega(a,a)^{1/2}$. The diagonal kernel therefore measures the sharp Hilbert-space cost of forcing a square-integrable [holomorphic function](/page/Holomorphic%20Function) to be large at $a$. [/example] ## Biholomorphic Transformation of the Kernel If two domains are biholomorphic, should their Bergman kernels be the same object in different coordinates? The answer is yes, but the volume form changes by the squared modulus of the complex Jacobian, so the kernel acquires one holomorphic Jacobian factor and one conjugate factor. [quotetheorem:3707] [citeproof:3707] The transformation law makes the Bergman kernel a biholomorphic invariant rather than a coordinate-dependent formula. It is also the quickest route to kernels on domains that are biholomorphic to standard models. Biholomorphism — not mere holomorphic injection — is essential: a non-injective or non-surjective holomorphic map $F:\Omega\to\widetilde\Omega$ only produces an isometric embedding $U:A^2(\widetilde\Omega)\hookrightarrow A^2(\Omega)$, never a unitary equivalence, and the displayed identity degenerates to an inequality between kernels rather than equality. The Jacobian factors $J_F(z)$ and $\overline{J_F(w)}$ must also be tracked as complex-valued, not as moduli: replacing them by $|J_F|$ would discard the unitary phase that makes the right-hand side holomorphic in $z$ and anti-holomorphic in $w$. On the diagonal $z=w$ only the modulus survives because the kernel is real and positive, but off the diagonal the phase information is what enforces the correct mixed holomorphic behavior, and a $|J_F|$ formula would generally fail to be a reproducing kernel at all. [example: Upper Half-Plane Kernel] Let $\mathbb H=\{z=x+iy\in\mathbb C:y>0\}$ and define \begin{align*} \varphi(z)=\frac{z-i}{z+i}. \end{align*} For $z=x+iy\in\mathbb H$, \begin{align*} |z+i|^2-|z-i|^2 &=\bigl(x^2+(y+1)^2\bigr)-\bigl(x^2+(y-1)^2\bigr)\\ &=(y^2+2y+1)-(y^2-2y+1)\\ &=4y>0, \end{align*} so $|\varphi(z)|<1$. Solving $\zeta=(z-i)/(z+i)$ gives \begin{align*} \zeta(z+i)&=z-i,\\ z(1-\zeta)&=i(1+\zeta),\\ z&=i\frac{1+\zeta}{1-\zeta}. \end{align*} If $|\zeta|<1$, then \begin{align*} i\frac{1+\zeta}{1-\zeta} &=i\frac{(1+\zeta)(1-\overline{\zeta})}{|1-\zeta|^2}\\ &=\frac{i(1-|\zeta|^2+\zeta-\overline{\zeta})}{|1-\zeta|^2}, \end{align*} whose imaginary part is $(1-|\zeta|^2)/|1-\zeta|^2>0$. Hence $\varphi$ is a biholomorphism from $\mathbb H$ to $\Delta$. Its derivative is \begin{align*} \varphi'(z) &=\frac{(z+i)\cdot 1-(z-i)\cdot 1}{(z+i)^2}\\ &=\frac{2i}{(z+i)^2}, \end{align*} and therefore \begin{align*} \overline{\varphi'(w)}=\frac{-2i}{(\overline w-i)^2}. \end{align*} The factor appearing in the disc kernel is \begin{align*} 1-\varphi(z)\overline{\varphi(w)} &=1-\frac{z-i}{z+i}\frac{\overline w+i}{\overline w-i}\\ &=\frac{(z+i)(\overline w-i)-(z-i)(\overline w+i)}{(z+i)(\overline w-i)}\\ &=\frac{(z\overline w-iz+i\overline w+1)-(z\overline w+iz-i\overline w+1)}{(z+i)(\overline w-i)}\\ &=\frac{-2iz+2i\overline w}{(z+i)(\overline w-i)}\\ &=\frac{-2i(z-\overline w)}{(z+i)(\overline w-i)}. \end{align*} Thus \begin{align*} \frac{1}{(1-\varphi(z)\overline{\varphi(w)})^2} &=\frac{(z+i)^2(\overline w-i)^2}{(-2i)^2(z-\overline w)^2}\\ &=-\frac{(z+i)^2(\overline w-i)^2}{4(z-\overline w)^2}. \end{align*} Using the disc kernel formula \begin{align*} K_\Delta(\zeta,\eta)=\frac{1}{\pi(1-\zeta\overline{\eta})^2} \end{align*} and the *[Transformation Law for the Bergman Kernel](/theorems/3707)*, we obtain \begin{align*} K_{\mathbb H}(z,w) &=\varphi'(z)K_\Delta(\varphi(z),\varphi(w))\overline{\varphi'(w)}\\ &=\frac{2i}{(z+i)^2}\cdot \frac{1}{\pi(1-\varphi(z)\overline{\varphi(w)})^2}\cdot \frac{-2i}{(\overline w-i)^2}\\ &=\frac{2i}{(z+i)^2}\cdot \frac{1}{\pi}\cdot \left(-\frac{(z+i)^2(\overline w-i)^2}{4(z-\overline w)^2}\right)\cdot \frac{-2i}{(\overline w-i)^2}\\ &=\frac{(2i)(-1)(-2i)}{4\pi}\frac{1}{(z-\overline w)^2}\\ &=-\frac{1}{\pi(z-\overline w)^2}. \end{align*} On the diagonal, since $z-\overline z=2i\,\operatorname{Im}z$, \begin{align*} K_{\mathbb H}(z,z) &=-\frac{1}{\pi(z-\overline z)^2}\\ &=-\frac{1}{\pi(2i\,\operatorname{Im}z)^2}\\ &=-\frac{1}{\pi\bigl(-4(\operatorname{Im}z)^2\bigr)}\\ &=\frac{1}{4\pi(\operatorname{Im}z)^2}. \end{align*} Thus $K_{\mathbb H}(z,z)\to+\infty$ as $\operatorname{Im}z\to 0^+$, so the diagonal kernel records the approach to the real boundary. [/example] [illustration:scv-iii-half-plane-bergman-kernel] This example also shows why the diagonal positivity statement must be understood through the Hermitian structure of the kernel. The formula has a minus sign off the diagonal, but its diagonal value is positive because $(z-\overline z)^2=-4(\operatorname{Im}z)^2$. ## Explicit Kernels on the Ball and the Polydisc What do the kernels look like on the two basic bounded symmetric domains taught first in several complex variables? The unit ball is rotationally symmetric, while the polydisc is a product domain with corners. Their kernels are both explicit, but the formulas encode very different geometry. [quotetheorem:3708] [citeproof:3708] The ball formula depends only on the scalar invariant $\langle z,w\rangle$. The polydisc formula records each coordinate separately, reflecting that the automorphism group of $\Delta^n$ is built from disc automorphisms and coordinate permutations rather than from all unitary rotations. [example: Diagonal Growth on Model Domains] Substituting $w=z$ in the ball formula gives \begin{align*} K_{B_n}(z,z) &=\frac{n!}{\pi^n}\frac{1}{(1-\langle z,z\rangle)^{n+1}}\\ &=\frac{n!}{\pi^n}\frac{1}{\left(1-\sum_{j=1}^n z_j\overline{z_j}\right)^{n+1}}\\ &=\frac{n!}{\pi^n}\frac{1}{\left(1-\sum_{j=1}^n |z_j|^2\right)^{n+1}}\\ &=\frac{n!}{\pi^n}(1-|z|^2)^{-(n+1)}. \end{align*} For the polydisc, \begin{align*} K_{\Delta^n}(z,z) &=\frac{1}{\pi^n}\prod_{j=1}^n\frac{1}{(1-z_j\overline{z_j})^2}\\ &=\frac{1}{\pi^n}\prod_{j=1}^n\frac{1}{(1-|z_j|^2)^2}\\ &=\frac{1}{\pi^n}\prod_{j=1}^n(1-|z_j|^2)^{-2}. \end{align*} On the ball, if $\delta_{B_n}(z)$ denotes the Euclidean distance from $z$ to $\partial B_n$, then $\delta_{B_n}(z)=1-|z|$ and \begin{align*} 1-|z|^2=(1-|z|)(1+|z|)=\delta_{B_n}(z)(1+|z|). \end{align*} Thus the diagonal blow-up is controlled by one boundary defining quantity. On the polydisc, each factor $1-|z_j|^2$ can tend to $0$ independently; for example, if $z^{(m)}=(r_m,\dots,r_m)$ with $r_m\to1^-$, then \begin{align*} K_{\Delta^n}(z^{(m)},z^{(m)}) &=\frac{1}{\pi^n}\prod_{j=1}^n(1-r_m^2)^{-2}\\ &=\frac{1}{\pi^n}(1-r_m^2)^{-2n}. \end{align*} A point approaching a corner therefore accumulates one singular factor from each approaching coordinate, while a point approaching only one face accumulates only that coordinate's factor. This contrast is the first sign that smooth boundary and product boundary lead to different Bergman geometry. [/example] The diagonal formulas record the boundary geometry of the two model domains in different ways. On the ball, the singularity of $K_{B_n}(z,z)$ depends on a single Euclidean quantity $1-|z|^2$, which equals the squared Euclidean distance to $\partial B_n$ up to a smooth factor; the boundary is a smooth real hypersurface, so the blow-up rate is dictated by one boundary defining function. On the polydisc, each factor $1-|z_j|^2$ controls the distance to a separate face $\{|z_j|=1\}$, and the boundary has corners where several faces meet. A point approaching a corner from the interior sees several of the factors decay simultaneously, so the product kernel diverges at the multiplied rates, while a point approaching a smooth face only sees one factor decay. This product-type behavior is more than a curiosity of the formula and motivates the following warning. [illustration:scv-iii-ball-polydisc-kernel-boundary] [remark: Failure of a Ball-Type Formula on the Polydisc] The polydisc has an explicit product formula, but it does not have a ball-type formula controlled only by $\langle z,w\rangle$. Points with the same value of $\langle z,z\rangle$ can have different diagonal kernel values if their coordinate moduli differ. Thus the product structure prevents the kernel from collapsing to a single radial expression. [/remark] The formulas also suggest what is hard on a general domain. Without a large automorphism group or product decomposition, there is usually no closed expression for $K_\Omega(z,w)$, and the $\bar{\partial}$ method becomes the main way to estimate it. In practice the course uses three routes to a kernel. First, on highly symmetric domains, choose an orthonormal basis of monomials and sum $\sum_\alpha e_\alpha(z)\overline{e_\alpha(w)}$. Second, on a domain biholomorphic to a known model, apply the transformation law and keep careful track of the complex Jacobian. Third, on a general pseudoconvex domain, use $\bar{\partial}$ estimates to construct enough holomorphic functions and then use the extremal characterisation to bound the diagonal, with polarisation recovering off-diagonal information when needed. ## Ramadanov Convergence and the Intrinsic Metric Hierarchy If a sequence of domains converges to a limiting domain, do the corresponding kernels converge as well? This question matters because the Bergman kernel is both an analytic object and a source of intrinsic geometry; stability of the kernel gives stability of the geometry. [quotetheorem:3709] [citeproof:3709] The convergence is local in the interior; boundary behavior requires additional hypotheses. This distinction is important in applications because the Bergman kernel often blows up at the boundary even when the domains converge smoothly. [example: Exhausting the Unit Disc] For $j\ge 2$, write $r_j=1-1/j$, so $\Omega_j=\{z\in\mathbb C:|z|<r_j\}=r_j\Delta$. The dilation \begin{align*} F_j:\Omega_j&\longrightarrow \Delta, & F_j(\zeta)&=\frac{\zeta}{r_j} \end{align*} is biholomorphic, and \begin{align*} F_j'(\zeta)=\frac{1}{r_j},\qquad \overline{F_j'(w)}=\frac{1}{r_j}. \end{align*} Using the unit-disc kernel \begin{align*} K_\Delta(\zeta,\eta)=\frac{1}{\pi(1-\zeta\overline{\eta})^2} \end{align*} and the *[Transformation Law for the Bergman Kernel](/theorems/3707)*, for $z,w\in\Omega_j$ we get \begin{align*} K_{\Omega_j}(z,w) &=F_j'(z)K_\Delta(F_j(z),F_j(w))\overline{F_j'(w)}\\ &=\frac{1}{r_j}\cdot \frac{1}{\pi\left(1-\frac{z}{r_j}\overline{\frac{w}{r_j}}\right)^2} \cdot \frac{1}{r_j}\\ &=\frac{1}{\pi r_j^2}\frac{1}{\left(1-\frac{z\overline w}{r_j^2}\right)^2}\\ &=\frac{1}{\pi(1-1/j)^2}\frac{1}{\left(1-\frac{z\overline w}{(1-1/j)^2}\right)^2}. \end{align*} Now fix $z,w\in\Delta$. Since $r_j\to 1$, there is $J$ such that $|z|<r_j$ and $|w|<r_j$ for all $j\ge J$, so the displayed formula applies eventually. Also \begin{align*} r_j^2&\longrightarrow 1,\\ \frac{z\overline w}{r_j^2}&\longrightarrow z\overline w,\\ 1-\frac{z\overline w}{r_j^2}&\longrightarrow 1-z\overline w. \end{align*} Because $|z\overline w|<1$, the limit denominator $1-z\overline w$ is nonzero. Therefore \begin{align*} K_{\Omega_j}(z,w) &=\frac{1}{\pi r_j^2}\frac{1}{\left(1-\frac{z\overline w}{r_j^2}\right)^2}\\ &\longrightarrow \frac{1}{\pi}\frac{1}{(1-z\overline w)^2}\\ &=K_\Delta(z,w). \end{align*} On every compact subset of $\Delta\times\Delta$, the same convergence is uniform because $|z\overline w|$ stays bounded away from $1$. Thus this exhaustion illustrates the interior, local content of Ramadanov convergence, with no claim about the boundary singularity of the kernel. [/example] [illustration:scv-iii-exhausting-unit-disc] The intrinsic metric hierarchy enters through extremal definitions. The Carathéodory metric measures derivatives of bounded holomorphic maps to the disc, the Kobayashi metric measures holomorphic discs mapping into the domain, and the Bergman metric measures derivatives of logarithms of kernel densities. On bounded domains with $K_\Omega(z,z)>0$, the extremal formula gives \begin{align*} c_\Omega(z;v)\le \beta_\Omega(z;v), \end{align*} where $c_\Omega$ is the Carathéodory metric and $\beta_\Omega$ is the Bergman metric, **with both normalised so that on the unit disc $\Delta\subset\mathbb C$ they reduce to the Poincaré metric $|v|/(1-|z|^2)$**. Under this normalisation the inequality holds on every bounded domain without an explicit constant; if instead one uses the kernel-density convention $\beta_\Omega(z;v)^2=\sum g_{i\overline j}v_i\overline{v_j}$ from the definition above without the factor of $2$ that some references include to match the Poincaré metric on $\Delta$, the inequality becomes $c_\Omega\le C_n\beta_\Omega$ for an explicit dimensional constant. We adopt the Poincaré-normalised convention throughout, so all displayed metric inequalities are constant-free. The general comparison $c_\Omega\le k_\Omega$ with the Kobayashi metric $k_\Omega$ remains part of the usual Carathéodory-Kobayashi hierarchy; stronger comparability involving $\beta_\Omega$ needs geometric hypotheses such as strong pseudoconvexity or finite type. Ramadanov convergence implies convergence of Bergman metrics on compact subsets whenever the diagonal kernels stay positive. Indeed, local [uniform convergence](/page/Uniform%20Convergence) of holomorphic kernels gives convergence of all derivatives on smaller compact sets by [Cauchy estimates](/theorems/2571), and the metric is obtained by differentiating $\log K_\Omega(z,z)$. ## Bergman Metric, Curvature, and Boundary Regularity What geometric structure is hidden in the diagonal of the Bergman kernel? A scalar function $K_\Omega(z,z)$ is not yet a metric, and the difficulty is to extract from it a coordinate-invariant positive Hermitian form rather than merely a density depending on coordinates. Once $K_\Omega(z,z)$ is positive, its logarithm acts as a potential; applying mixed complex derivatives converts analytic estimates for point evaluation into curvature information. [definition: Bergman Metric] Let $\Omega\subset\mathbb C^n$ be a domain with $K_\Omega(z,z)>0$. The Bergman metric tensor at $z$ is the Hermitian form on $T_z^{1,0}\Omega\cong\mathbb C^n$ with coefficients \begin{align*} g_{i\overline j}(z):=\frac{\partial^2}{\partial z_i\partial\overline z_j}\log K_\Omega(z,z). \end{align*} For $v\in T_z^{1,0}\Omega\cong\mathbb C^n$, its squared Bergman length is \begin{align*} \beta_\Omega(z;v)^2:=\sum_{i,j=1}^n g_{i\overline j}(z)v_i\overline{v_j}. \end{align*} [/definition] The positivity of this Hermitian form is a consequence of the reproducing kernel extremal problem. In model domains it is positive definite, and in the pseudoconvex settings of this course it is the natural metric attached to the [Hilbert space](/page/Hilbert%20Space) $A^2(\Omega)$. [quotetheorem:3710] [citeproof:3710] This invariance is why the metric is more than a convenient formula derived from a kernel. On a model domain, a single potential computation can be transported to every biholomorphic copy of that domain, while the pluriharmonic Jacobian term disappears from the metric. The next example computes the metric on the ball, where the high symmetry makes the resulting curvature especially rigid. [example: Bergman Metric on the Unit Ball] From the ball kernel formula, putting $w=z$ gives \begin{align*} K_{B_n}(z,z) &=\frac{n!}{\pi^n}\frac{1}{(1-\langle z,z\rangle)^{n+1}}\\ &=\frac{n!}{\pi^n}\frac{1}{\left(1-\sum_{k=1}^n |z_k|^2\right)^{n+1}}. \end{align*} Set $\rho(z)=1-\sum_{k=1}^n |z_k|^2$. Then \begin{align*} \log K_{B_n}(z,z) &=\log\left(\frac{n!}{\pi^n}\rho(z)^{-(n+1)}\right)\\ &=\log\left(\frac{n!}{\pi^n}\right)-(n+1)\log\rho(z). \end{align*} The constant term has zero mixed second derivatives, so the Bergman metric comes from the potential $-(n+1)\log\rho(z)$. For each $i$, \begin{align*} \frac{\partial \rho}{\partial z_i} &=-\overline{z_i},\\ \frac{\partial}{\partial z_i}\log K_{B_n}(z,z) &=-(n+1)\frac{1}{\rho}\frac{\partial \rho}{\partial z_i}\\ &=(n+1)\frac{\overline{z_i}}{\rho}. \end{align*} Differentiating once more with respect to $\overline z_j$ gives \begin{align*} g_{i\overline j}(z) &=\frac{\partial}{\partial\overline z_j}\left((n+1)\frac{\overline{z_i}}{\rho}\right)\\ &=(n+1)\left(\frac{\delta_{ij}}{\rho}+\overline{z_i}\frac{\partial}{\partial\overline z_j}(\rho^{-1})\right)\\ &=(n+1)\left(\frac{\delta_{ij}}{\rho}+\overline{z_i}\frac{z_j}{\rho^2}\right)\\ &=(n+1)\frac{\rho\,\delta_{ij}+\overline{z_i}z_j}{\rho^2}. \end{align*} Hence, for $v\in\mathbb C^n$, \begin{align*} \beta_{B_n}(z;v)^2 &=\sum_{i,j=1}^n g_{i\overline j}(z)v_i\overline{v_j}\\ &=(n+1)\left(\frac{\sum_{i=1}^n |v_i|^2}{\rho} +\frac{\sum_{i,j=1}^n \overline{z_i}z_jv_i\overline{v_j}}{\rho^2}\right)\\ &=(n+1)\left(\frac{|v|^2}{1-|z|^2} +\frac{|\langle v,z\rangle|^2}{(1-|z|^2)^2}\right). \end{align*} In particular, at the origin, \begin{align*} g_{i\overline j}(0)&=(n+1)\delta_{ij},& \beta_{B_n}(0;v)^2&=(n+1)|v|^2. \end{align*} Thus the ball metric is the complex hyperbolic metric multiplied by the constant factor $n+1$ in this normalisation; its holomorphic sectional curvature is negative and independent of the point and direction. This is the explicit model calculation that later becomes the curvature-tensor computation in Course IV. [/example] [Boundary regularity](/theorems/99) is subtler than interior invariance. The final issue is whether biholomorphic equivalence seen by the kernel also controls smooth boundary behavior. The Bergman projection $P_\Omega:L^2(\Omega,dV)\to A^2(\Omega)$ is the [orthogonal projection](/theorems/437), and its behavior on smooth functions up to the boundary is governed by the same $\bar{\partial}$-Neumann regularity theory that appeared earlier in the course. The next quoted theorem records the regularity hypothesis under which this kernel machinery forces smooth extension of biholomorphisms. [quotetheorem:3711] The course quotes this theorem rather than proving it. Its proof belongs to the [boundary regularity](/theorems/99) theory of the $\bar{\partial}$-Neumann problem and uses precise control of how the Bergman projection preserves smooth boundary jets. The transformation law explains why such a statement affects biholomorphisms: if the kernel and its boundary derivatives are controlled, then the Jacobian factors of a biholomorphic map are controlled as well. Without Condition R, the Bergman projection may fail to preserve $C^\infty(\overline\Omega)$ regularity, so kernel derivatives can lose boundary control and a biholomorphism need not be forced by this argument to extend smoothly to the boundary. Strongly pseudoconvex domains satisfy the standard regularity hypotheses, while worm domains show that smooth bounded pseudoconvex domains can have surprising failures of global regularity; they are the model warning that pseudoconvexity alone is not enough for Bell-type boundary extension. The chapter therefore closes a circle. The $\bar{\partial}$ estimates construct holomorphic functions with prescribed local behavior, [Hilbert space](/page/Hilbert%20Space) duality packages optimal point evaluation into $K_\Omega$, explicit models show how geometry appears in formulas, and [boundary regularity](/theorems/99) connects the kernel back to smooth extension of biholomorphic maps. In the next stage of the course, the same point of view is pushed toward curvature and positivity of bundles rather than only functions on domains. Chapter 6 showed how Bergman kernels and L² estimates encode geometry explicitly; this chapter translates that geometry into approximation theorems. When a function lies in the weighted L² closure of holomorphic functions, choosing the right weight and applying the ∂̄-equation yields a uniform holomorphic approximation. Vanishing, extension, Runge approximation, and boundary obstruction all follow from this single technique. # 7. Applications to Approximation Theory The previous chapters built the analytic engine: weighted $L^2$ estimates solve $\bar\partial u=f$ when the ambient domain is pseudoconvex and the weight has enough curvature. This chapter turns that solvability theorem into approximation statements. The guiding idea is to begin with a function that is only locally holomorphic, correct its $\bar\partial$-error by an $L^2$ solution, and choose the weight so that the correction is small on the compact set where approximation is required. The same mechanism also gives a constructive proof of the Cartan-Thullen theorem. Instead of using sheaf cohomology, we build holomorphic functions with prescribed boundary growth by solving $\bar\partial$ with singular weights. Approximation, domains of holomorphy, and Runge pairs are therefore different faces of the same estimate. ## Runge Approximation From a $\bar{\partial}$-Correction Suppose $f$ is holomorphic on a plane domain $\Omega\subset \mathbb C$, and suppose $K\Subset\Omega$ is the compact set on which we want to approximate $f$. When can the approximants be chosen to be polynomials, rather than merely holomorphic functions on a neighbourhood of $K$? The obstruction is topological. If $K$ surrounds a hole, a function with a pole in the hole may be holomorphic near $K$ but cannot be uniformly approximated there by polynomials, because polynomial approximants have no pole available. The correct hypothesis says that the holes filled in by polynomial convexity remain inside the original domain. [definition: Polynomial Hull In One Variable] Let $K\subset\mathbb C$ be compact. The polynomial hull of $K$ is \begin{align*} \widehat K_{\mathrm{poly}} = \{z\in\mathbb C: |p(z)|\leq \sup_{w\in K}|p(w)|\text{ for every }p\in\mathbb C[z]\}. \end{align*} [/definition] In one complex variable, $\widehat K_{\mathrm{poly}}$ is obtained from $K$ by filling in the bounded connected components of $\mathbb C\setminus K$. To state the approximation theorem for domains, we need a condition saying that these filled-in holes never leave the domain where the original function is holomorphic. This is the Runge condition in the plane, and it is designed to remove exactly the topological obstruction detected by polynomial hulls. [definition: Runge Domain In The Plane] A domain $\Omega\subset\mathbb C$ is Runge in $\mathbb C$ if $\widehat K_{\mathrm{poly}}\subset\Omega$ for every compact set $K\Subset\Omega$. [/definition] The definition is arranged so that a cutoff construction has room to place its $\bar\partial$-error away from the polynomial hull of $K$. [example: Annulus As A Non-Runge Domain] Let $\Omega=\{z\in\mathbb C:1<|z|<2\}$ and $K=\{z\in\mathbb C:|z|=3/2\}$. Since $\mathbb C\setminus K$ has bounded component $\{|z|<3/2\}$, the one-variable description of polynomial hulls gives \begin{align*} \widehat K_{\mathrm{poly}}=\{z\in\mathbb C:|z|\leq 3/2\}, \end{align*} so $0\in \widehat K_{\mathrm{poly}}$ and $\widehat K_{\mathrm{poly}}\not\subset\Omega$. The function $f(z)=1/z$ is holomorphic on $\Omega$ because $0\notin\Omega$. We show that $f$ is not uniformly approximable on $K$ by polynomials. Suppose, toward a contradiction, that polynomials $p_j$ satisfy \begin{align*} \sup_{z\in K}|p_j(z)-z^{-1}|\to 0. \end{align*} Orient $K$ by $\gamma(t)=(3/2)e^{it}$ for $0\leq t\leq 2\pi$. Then $\gamma'(t)=(3/2)i e^{it}$ and \begin{align*} \left|\int_K p_j(z)\,dz-\int_K \frac{1}{z}\,dz\right| &=\left|\int_K \left(p_j(z)-\frac{1}{z}\right)\,dz\right|\\ &=\left|\int_0^{2\pi}\left(p_j(\gamma(t))-\frac{1}{\gamma(t)}\right)\gamma'(t)\,dt\right|\\ &\leq \int_0^{2\pi}\sup_{z\in K}\left|p_j(z)-z^{-1}\right|\,|\gamma'(t)|\,dt\\ &=\int_0^{2\pi}\sup_{z\in K}\left|p_j(z)-z^{-1}\right|\,\frac{3}{2}\,dt\\ &=3\pi\sup_{z\in K}\left|p_j(z)-z^{-1}\right|\to 0. \end{align*} Thus $\int_K p_j(z)\,dz\to \int_K z^{-1}\,dz$. For each polynomial $p_j(z)=\sum_{k=0}^{d_j}a_{j,k}z^k$, the integral around $K$ is \begin{align*} \int_K p_j(z)\,dz &=\sum_{k=0}^{d_j}a_{j,k}\int_0^{2\pi}\left(\frac32 e^{it}\right)^k\left(\frac32 i e^{it}\right)\,dt\\ &=\sum_{k=0}^{d_j}a_{j,k}i\left(\frac32\right)^{k+1}\int_0^{2\pi}e^{i(k+1)t}\,dt\\ &=\sum_{k=0}^{d_j}a_{j,k}i\left(\frac32\right)^{k+1} \left[\frac{e^{i(k+1)t}}{i(k+1)}\right]_{0}^{2\pi}\\ &=\sum_{k=0}^{d_j}a_{j,k}i\left(\frac32\right)^{k+1} \frac{e^{2\pi i(k+1)}-1}{i(k+1)}\\ &=0. \end{align*} But \begin{align*} \int_K \frac{1}{z}\,dz &=\int_0^{2\pi}\frac{1}{(3/2)e^{it}}\left(\frac32 i e^{it}\right)\,dt\\ &=\int_0^{2\pi} i\,dt\\ &=2\pi i. \end{align*} The convergence of the integrals would therefore force $0\to 2\pi i$, a contradiction. Hence no sequence of polynomials can converge uniformly to $1/z$ on $K$. [/example] This example identifies the only obstruction that the one-variable theorem must remove: polynomial approximation cannot create singular behaviour inside a bounded component of the complement. The Runge condition says that every such hidden component has already been absorbed into the domain, so a $\bar\partial$ correction can be placed away from the compact set without crossing a forbidden hole. [quotetheorem:3712] The course proves this theorem by translating approximation into a solvability problem for $\bar\partial$. [citeproof:3712] The final Taylor approximation is a one-variable feature of entire functions. In several variables the analogous step is replaced by Oka-Weil approximation on holomorphically convex compact sets. [remark: From L Two To Uniform Approximation] The $L^2$ estimate is naturally weaker than the classical sup-norm statement, but holomorphicity upgrades it on smaller compact sets. If $L\Subset K^\circ$ and $F$ is holomorphic near $K$, the mean-value inequality bounds $\sup_L |F|$ by a constant times $\|F\|_{L^2(K)}$. Thus an $L^2$ approximation on a buffer compact set gives uniform approximation on the inner compact set. [/remark] ## Boundary Growth And The Cartan-Thullen Theorem How can an $L^2$ existence theorem produce a [holomorphic function](/page/Holomorphic%20Function) that refuses to extend across the boundary of a pseudoconvex domain? The approximation proof made the correction small on a compact set; the Cartan-Thullen proof uses singular weights to make the correction vanish at selected interior points approaching the boundary. [definition: Domain Of Holomorphy] A domain $\Omega\subset\mathbb C^n$ is a domain of holomorphy if, for every connected domain $\Omega_2\subset\mathbb C^n$ with $\Omega\subset\Omega_2$ and $\Omega_2\neq\Omega$, there exists $f\in\mathcal O(\Omega)$ that has no holomorphic extension to $\Omega_2$. [/definition] The definition asks for at least one [holomorphic function](/page/Holomorphic%20Function) that detects every attempted enlargement. Cartan-Thullen identifies this function-theoretic condition with convexity measured by holomorphic functions. [definition: Holomorphic Hull] Let $\Omega\subset\mathbb C^n$ be a domain and let $K\Subset\Omega$ be compact. The holomorphic hull of $K$ in $\Omega$ is \begin{align*} \widehat K_\Omega = \{z\in\Omega: |f(z)|\leq \sup_{w\in K}|f(w)|\text{ for every }f\in\mathcal O(\Omega)\}. \end{align*} [/definition] The hull records the points that cannot be separated from $K$ by holomorphic functions on $\Omega$. A single compact set can have a hull larger than itself, and this is harmless as long as the enlargement remains safely inside the domain. The obstruction to a domain of holomorphy is different: hulls can drift toward the boundary, meaning holomorphic functions on $\Omega$ fail to detect an attempted enlargement before it leaves the domain. To compare this separation condition with pseudoconvexity, we need a domain-level condition that rules out this boundary escape for every compact set. [definition: Holomorphically Convex Domain] A domain $\Omega\subset\mathbb C^n$ is holomorphically convex if $\widehat K_\Omega\Subset\Omega$ for every compact set $K\Subset\Omega$. [/definition] Holomorphic convexity prevents compact sets from having hidden holomorphic hulls touching the boundary. The difficult direction is to produce separating holomorphic functions from the analytic hypothesis of pseudoconvexity. This is where $L^2$ solvability enters: it converts plurisubharmonic exhaustion functions into holomorphic functions with prescribed growth near boundary obstructions. At this point the central question is whether the geometric exhaustion condition and the function-theoretic hull condition are actually the same condition on a domain. The bridge theorem below gives that equivalence: pseudoconvexity is not merely a curvature-style boundary property, but is strong enough to manufacture the holomorphic functions that keep compact hulls from escaping. [quotetheorem:3713] [citeproof:3713] The role of pseudoconvexity is not cosmetic: it is what permits the use of plurisubharmonic exhaustion functions as weights. The theorem should be read as a comparison between two languages for the same boundary phenomenon. Holomorphic convexity says that compact hulls cannot escape toward the boundary, while pseudoconvexity says that the boundary admits enough plurisubharmonic control to prevent such escape analytically. This is why domains of holomorphy are detected by growth near the boundary, not merely by local Cauchy-Riemann equations. [example: Boundary Singularity In The Bidisc] Let \begin{align*} \mathbb B_2 &= \{z\in\mathbb C^2: |z_1|^2+|z_2|^2<1\}, & \Delta^2 &= \{z\in\mathbb C^2: |z_1|<1,\ |z_2|<1\}, \end{align*} and set \begin{align*} p=\left(\frac{1}{\sqrt2},\frac{1}{\sqrt2}\right),\qquad f(z)=\frac{1}{1-(z_1+z_2)/\sqrt2}. \end{align*} We first show that $f$ is holomorphic on $\mathbb B_2$. If $z\in\mathbb B_2$, then \begin{align*} |z_1+z_2|^2 &=(z_1+z_2)(\overline z_1+\overline z_2)\\ &=|z_1|^2+|z_2|^2+z_1\overline z_2+z_2\overline z_1\\ &=|z_1|^2+|z_2|^2+2\operatorname{Re}(z_1\overline z_2)\\ &\leq |z_1|^2+|z_2|^2+2|z_1||z_2|\\ &\leq |z_1|^2+|z_2|^2+|z_1|^2+|z_2|^2\\ &=2(|z_1|^2+|z_2|^2)\\ &<2. \end{align*} Thus $|z_1+z_2|<\sqrt2$, so $(z_1+z_2)/\sqrt2\neq 1$ and the denominator of $f$ does not vanish on $\mathbb B_2$. The point $p$ satisfies \begin{align*} |p_1|^2+|p_2|^2=\frac12+\frac12=1, \end{align*} so $p\in\partial\mathbb B_2$, while $|p_1|=|p_2|=1/\sqrt2<1$, so $p\in\Delta^2$. Also \begin{align*} 1-\frac{p_1+p_2}{\sqrt2} =1-\frac{1/\sqrt2+1/\sqrt2}{\sqrt2} =1-\frac{\sqrt2}{\sqrt2} =0, \end{align*} so the singular hyperplane of $f$ passes through an interior point of the polydisc. Suppose, toward a contradiction, that functions $F_j\in\mathcal O(\Delta^2)$ converge uniformly to $f$ on $\mathbb B_2$. Choose $J$ such that \begin{align*} \sup_{z\in\mathbb B_2}|F_J(z)-f(z)|<1. \end{align*} Since $F_J$ is holomorphic on $\Delta^2$, it is continuous near $p$ and hence bounded on some closed ball $\overline B(p,r)\Subset\Delta^2$; write \begin{align*} |F_J(z)|\leq M\qquad \text{for }z\in \overline B(p,r). \end{align*} For $0<t<1$, put $z(t)=tp$. Then \begin{align*} |z_1(t)|^2+|z_2(t)|^2=t^2\left(\frac12+\frac12\right)=t^2<1, \end{align*} so $z(t)\in\mathbb B_2$, and \begin{align*} f(z(t)) &=\frac{1}{1-\bigl(t/\sqrt2+t/\sqrt2\bigr)/\sqrt2}\\ &=\frac{1}{1-t}. \end{align*} As $t\to 1^-$, we have $z(t)\to p$ and $f(z(t))\to\infty$. Choose $t$ close enough to $1$ that $z(t)\in B(p,r)\cap\mathbb B_2$ and $1/(1-t)>M+1$. Then \begin{align*} |F_J(z(t))-f(z(t))| &\geq |f(z(t))|-|F_J(z(t))|\\ &> (M+1)-M\\ &=1, \end{align*} contradicting the choice of $J$. Therefore $f$ cannot be uniformly approximated on all of $\mathbb B_2$ by functions holomorphic on $\Delta^2$. The obstruction is not holomorphicity inside the ball: it is the demand for uniform control up to a boundary point of $\mathbb B_2$ where the approximating functions remain locally bounded but $f$ blows up. [/example] This example separates two approximation regimes. Runge approximation is always a compact-set statement; asking for uniform approximation all the way to a singular boundary point changes the problem. ## Mergelyan-Type Approximation On Pseudoconvex Domains Runge approximation begins with a function holomorphic on an [open set](/page/Open%20Set) around the compact set. What if the function is only continuous on the compact set and holomorphic on its interior? The one-variable answer is Mergelyan theorem. Its proof can still be organised through $\bar\partial$: smooth the boundary data, measure the failure of holomorphicity as a small $\bar\partial$-term near the boundary, solve it away, and then invoke Runge approximation. [definition: Mergelyan Algebra] For a compact set $K\subset\mathbb C^n$, define \begin{align*} A(K)=\{f\in C(K): f|_{K^\circ}\in\mathcal O(K^\circ)\}. \end{align*} [/definition] The space $A(K)$ is the natural target for uniform approximation on compact sets with interior. Its definition contains no information about whether holomorphic functions from outside $K$ are dense in it; that is the content of Mergelyan-type theorems. The obstruction is not pointwise regularity of the function, but the geometry of the compact set. Before moving to several variables, the one-variable theorem gives the model criterion: the complement must have no bounded holes where Cauchy data could hide from polynomial approximation. [quotetheorem:3714] In one variable, polynomial convexity of $K$ is the same as connectedness of $\mathbb C\setminus K$. [citeproof:3714] The theorem identifies the exact topological obstruction in the plane. Holomorphic data on the boundary of a hole can carry periods that no polynomial can reproduce uniformly, while connected complement leaves no such hidden component. This interpretation is the bridge to higher dimensions, where topology alone is no longer enough and polynomial or holomorphic convexity replaces connectedness of the complement. The pseudoconvex version has the same architecture, but the Cauchy-Green operator is replaced by Hörmander estimate and the polynomial approximation step is replaced by holomorphic approximation from the ambient pseudoconvex domain. The next result is useful because it states the approximation principle in a form compatible with $L^2$ methods on domains rather than with global polynomials on the plane. [quotetheorem:3715] The statement is deliberately formulated for the class of compact sets where the boundary regularisation and the $L^2$ estimates interact cleanly. [citeproof:3715] The theorem has two separate ingredients: regularity of the boundary lets us manufacture a small $\bar\partial$ error, while convexity ensures that the final holomorphic approximation comes from the desired ambient domain. If the convexity condition is removed, the local smoothing step may still work, but the final approximation step can fail for topological reasons already visible in one complex variable. [example: Why Polynomial Convexity Cannot Be Dropped] Let $K=\{z\in\mathbb C:|z|=1\}$ and define $f(z)=1/z$ on $K$. Since $K^\circ=\varnothing$ and $f$ is continuous on $K$, we have $f\in A(K)$. We show that $f$ is not uniformly approximable on $K$ by polynomials. Suppose, toward a contradiction, that polynomials $p_j$ satisfy \begin{align*} \sup_{z\in K}\left|p_j(z)-\frac{1}{z}\right|\to 0. \end{align*} Orient $K$ by $\gamma(t)=e^{it}$ for $0\leq t\leq 2\pi$. Then $\gamma'(t)=ie^{it}$ and $|\gamma'(t)|=1$, so \begin{align*} \left|\int_K p_j(z)\,dz-\int_K \frac{1}{z}\,dz\right| &=\left|\int_K\left(p_j(z)-\frac{1}{z}\right)\,dz\right|\\ &=\left|\int_0^{2\pi}\left(p_j(e^{it})-e^{-it}\right)ie^{it}\,dt\right|\\ &\leq \int_0^{2\pi}\left|p_j(e^{it})-e^{-it}\right|\,|ie^{it}|\,dt\\ &\leq \int_0^{2\pi}\sup_{z\in K}\left|p_j(z)-\frac{1}{z}\right|\,dt\\ &=2\pi\sup_{z\in K}\left|p_j(z)-\frac{1}{z}\right|\to 0. \end{align*} Thus $\int_K p_j(z)\,dz\to \int_K z^{-1}\,dz$. Write $p_j(z)=\sum_{k=0}^{d_j}a_{j,k}z^k$. Then \begin{align*} \int_K p_j(z)\,dz &=\int_0^{2\pi}\sum_{k=0}^{d_j}a_{j,k}(e^{it})^k\,ie^{it}\,dt\\ &=\sum_{k=0}^{d_j}a_{j,k}i\int_0^{2\pi}e^{i(k+1)t}\,dt\\ &=\sum_{k=0}^{d_j}a_{j,k}i\left[\frac{e^{i(k+1)t}}{i(k+1)}\right]_{0}^{2\pi}\\ &=\sum_{k=0}^{d_j}a_{j,k}\frac{e^{2\pi i(k+1)}-1}{k+1}\\ &=0, \end{align*} because $k+1$ is an integer and $e^{2\pi i(k+1)}=1$. On the other hand, \begin{align*} \int_K \frac{1}{z}\,dz &=\int_0^{2\pi}\frac{1}{e^{it}}\,ie^{it}\,dt\\ &=\int_0^{2\pi}i\,dt\\ &=2\pi i. \end{align*} The convergence of the integrals would force $0\to 2\pi i$, a contradiction. Therefore $1/z$ cannot be uniformly approximated on $K$ by polynomials. The failure comes from the hole inside the unit circle: polynomial approximants have zero contour integral around $K$, while $1/z$ records the missing singularity at $0$. [/example] Mergelyan theorem is therefore not just Runge approximation with weaker hypotheses on $f$. It also requires the compact set to have no polynomially invisible holes. ## Localisation, Partitions Of Unity, And Runge Pairs Approximation arguments are local in construction but global in conclusion. How can locally defined holomorphic approximants be glued without leaving a $\bar\partial$-error behind? The answer is the $\bar\partial$-[partition of unity](/page/Partition%20of%20Unity) method. Ordinary partitions of unity destroy holomorphicity; Hörmander theorem repairs this by solving exactly for the error created by the partition. [definition: Runge Pair] Let $\Omega_0\subset\Omega_1$ be domains in $\mathbb C^n$. The pair $(\Omega_0,\Omega_1)$ is a Runge pair if, for every compact set $K\Subset\Omega_0$ and every $f\in\mathcal O(\Omega_0)$, there exists a sequence $(F_j)_{j=1}^\infty$ in $\mathcal O(\Omega_1)$ such that $F_j\to f$ uniformly on $K$. [/definition] This definition is the relative version of Runge approximation: functions from the larger domain approximate functions from the smaller one. [explanation: The Dbar Partition Of Unity] Let $K\Subset\Omega_0$ and choose open sets $U_1,\dots,U_m\Subset\Omega_0$ covering a neighbourhood of $K$. Suppose we have local holomorphic functions $f_j\in\mathcal O(U_j)$ that almost agree on overlaps. A smooth [partition of unity](/page/Partition%20of%20Unity) $(\chi_j)$ gives the patched function \begin{align*} F_0=\sum_{j=1}^m \chi_j f_j. \end{align*} The price of patching is \begin{align*} \bar\partial F_0=\sum_{j=1}^m f_j\,\bar\partial\chi_j, \end{align*} which is supported where the cutoff functions vary. If this support is separated from $K$ by a plurisubharmonic weight, Hörmander theorem solves $\bar\partial u=\bar\partial F_0$ with $u$ small on $K$. Then $F=F_0-u$ is holomorphic and retains the desired approximation on $K$. [/explanation] The same idea gives a useful criterion for Runge pairs. The question is how to recognize when holomorphic functions on a smaller pseudoconvex domain can be approximated by functions from a larger one. The criterion below expresses exactly the geometric condition needed by the $\bar\partial$ patching argument: holomorphic hulls formed inside the larger domain must stay compactly inside the smaller domain. [quotetheorem:3716] [citeproof:3716] The compact containment of the hull is what prevents approximation from being obstructed by boundary escape. Conceptually, the pair is Runge when the larger domain sees no additional holomorphic shadow of a compact set in the smaller domain. This makes the condition stable under exhaustion arguments and explains why it appears repeatedly in approximation theorems. The next class of examples makes this condition geometric: for Reinhardt domains, holomorphic separation can be read from convex separation after taking logarithms of the coordinate moduli. The reason this symmetry matters is that general domains have no simple way to compare holomorphic hulls with Euclidean convex hulls; rotations in each coordinate force monomials to be the natural test functions and turn absolute values into linear data on logarithmic coordinates. Defining the class isolates exactly the domains where the Runge condition can be checked by convex geometry. [definition: Reinhardt Domain] A domain $\Omega\subset\mathbb C^n$ is Reinhardt if $(e^{i\theta_1}z_1,\dots,e^{i\theta_n}z_n)\in\Omega$ for every $z\in\Omega$ and every $\theta\in\mathbb R^n$. [/definition] Reinhardt domains are a useful test case because their holomorphic functions are organised by monomials and their pseudoconvexity is visible after taking logarithms of moduli. [example: Runge Exhaustion Of A Reinhardt Domain] Let $\Omega\subset(\mathbb C^*)^n$ be a pseudoconvex Reinhardt domain, and write \begin{align*} \Lambda=\operatorname{Log}(\Omega) =\{(\log|z_1|,\dots,\log|z_n|):z\in\Omega\}\subset\mathbb R^n. \end{align*} By the *logarithmic convexity criterion for pseudoconvex Reinhardt domains*, $\Lambda$ is convex. Choose compact convex sets $Q_m\Subset\Lambda$ with $Q_m\subset Q_{m+1}^\circ$ and $\bigcup_m Q_m=\Lambda$, and set \begin{align*} K_m=\{z\in\Omega:\operatorname{Log}(z)\in Q_m\}, \qquad \Omega_m=K_m^\circ=\operatorname{Log}^{-1}(Q_m^\circ). \end{align*} [claim]The domains $\Omega_m$ form a Runge exhaustion of $\Omega$.[/claim] [proof]First, $K_m$ is compact. Indeed, if \begin{align*} a_{m,j}=\min_{x\in Q_m}x_j, \qquad b_{m,j}=\max_{x\in Q_m}x_j, \end{align*} then \begin{align*} K_m\subset \prod_{j=1}^n\{z_j\in\mathbb C:e^{a_{m,j}}\leq |z_j|\leq e^{b_{m,j}}\}, \end{align*} and $K_m=\operatorname{Log}^{-1}(Q_m)$ is closed in this compact product of annuli. Since $Q_m\subset Q_{m+1}^\circ$, we have $\Omega_m\Subset\Omega_{m+1}$. Also, if $z\in\Omega$ and $x=\operatorname{Log}(z)$, then $x\in Q_m$ for some $m$, hence $x\in Q_{m+1}^\circ$, so $z\in\Omega_{m+1}$. Therefore \begin{align*} \bigcup_{m=1}^\infty \Omega_m=\Omega. \end{align*} It remains to prove that each pair $(\Omega_m,\Omega_{m+1})$ is Runge. Fix $m$ and let $L\Subset\Omega_m$ be compact. Put \begin{align*} P=\operatorname{Log}(L)\subset Q_m^\circ, \qquad C=\operatorname{conv}(P). \end{align*} Since $Q_m^\circ$ is convex, $C\subset Q_m^\circ$; since $P$ is compact in finite-dimensional space, $C$ is compact. Thus \begin{align*} \operatorname{Log}^{-1}(C)\Subset\Omega_m. \end{align*} We show that the $\mathcal O(\Omega_{m+1})$-hull of $L$ is contained in $\operatorname{Log}^{-1}(C)$. Let $z\in\Omega_{m+1}$ and set $x=\operatorname{Log}(z)$. Suppose $x\notin C$. By strict convex separation of the point $x$ from the compact convex set $C$, there are $v\in\mathbb R^n$ and $\delta>0$ such that \begin{align*} v\cdot x-v\cdot y\geq \delta \qquad\text{for every }y\in C. \end{align*} Since $P\subset C$, the same inequality holds for every $y\in P$. Let \begin{align*} R=\max_{y\in P}|x-y|. \end{align*} Here $R>0$, because $x\notin C$ and $P\subset C$. Choose $q\in\mathbb Q^n$ with \begin{align*} |q-v|<\frac{\delta}{2R}. \end{align*} Then, for every $y\in P$, \begin{align*} q\cdot x-q\cdot y &=v\cdot x-v\cdot y+(q-v)\cdot(x-y)\\ &\geq \delta-|q-v|\,|x-y|\\ &\geq \delta-|q-v|R\\ &>\delta-\frac{\delta}{2}\\ &=\frac{\delta}{2}. \end{align*} Choose a positive integer $N$ such that $\alpha=Nq\in\mathbb Z^n$. Then \begin{align*} \alpha\cdot x-\alpha\cdot y =N(q\cdot x-q\cdot y)>0 \qquad\text{for every }y\in P. \end{align*} Because $\Omega_{m+1}\subset(\mathbb C^*)^n$, the Laurent monomial \begin{align*} g_\alpha(\zeta)=\zeta^\alpha =\zeta_1^{\alpha_1}\cdots \zeta_n^{\alpha_n} \end{align*} is holomorphic on $\Omega_{m+1}$. For every $\zeta\in L$, \begin{align*} \log|g_\alpha(\zeta)| &=\log\left(|\zeta_1|^{\alpha_1}\cdots |\zeta_n|^{\alpha_n}\right)\\ &=\sum_{j=1}^n \alpha_j\log|\zeta_j|\\ &=\alpha\cdot\operatorname{Log}(\zeta)\\ &\leq \sup_{y\in P}\alpha\cdot y\\ &<\alpha\cdot x\\ &=\log|g_\alpha(z)|. \end{align*} Exponentiating gives \begin{align*} |g_\alpha(z)|>\sup_{\zeta\in L}|g_\alpha(\zeta)|. \end{align*} Hence $z$ is not in the $\mathcal O(\Omega_{m+1})$-hull of $L$. Therefore \begin{align*} \widehat L_{\Omega_{m+1}}\subset \operatorname{Log}^{-1}(C)\Subset\Omega_m. \end{align*} By the same logarithmic convexity criterion, $\Omega_m$ and $\Omega_{m+1}$ are pseudoconvex, because their logarithmic images are the convex domains $Q_m^\circ$ and $Q_{m+1}^\circ$. The *L Two Criterion For Runge Pairs* now applies to the inclusion $\Omega_m\subset\Omega_{m+1}$, so $(\Omega_m,\Omega_{m+1})$ is a Runge pair for every $m$.[/proof] Thus convex separation in logarithmic coordinates is exactly holomorphic separation by Laurent monomials, and the nested domains $\Omega_m$ exhaust $\Omega$ without introducing non-Runge gaps. [/example] The chapter main lesson is that approximation is not an isolated phenomenon. Runge theorem, Mergelyan theorem in one variable, the Cartan-Thullen theorem, and Runge pairs all follow the same pattern: create a controlled $\bar\partial$-error, solve it with a weight that is favourable on the set of interest, and convert the resulting holomorphic correction into the desired approximation or boundary obstruction. Chapter 7 used ∂̄-solvability to prove approximation and vanishing results about holomorphic functions; this chapter introduces the sheaf language that makes singularities precise and algebraic. A plurisubharmonic weight can be too singular to allow all holomorphic functions to be square-integrable against it, and the functions that survive form an ideal sheaf. Multiplier ideals, born from L² estimates, are coherent (compatible with sheaf cohomology) while retaining the analytic information needed to govern membership and vanishing. # 8. Multiplier Ideal Sheaves and the Nadel Vanishing Theorem This chapter turns the analytic estimates from the preceding lectures into a sheaf-theoretic tool for controlling singularities. A plurisubharmonic weight can be too singular for arbitrary holomorphic functions to remain square-integrable against it, and the functions that survive form an ideal sheaf. The main point is that Hörmander's $L^2$ method makes this ideal coherent, so it can be used inside cohomology and vanishing theorems just like an algebraic ideal sheaf. ## Integrability Ideals Attached to Plurisubharmonic Weights Which holomorphic germs are compatible with a given singular metric? If $\rho=e^{-\phi}$ is the local weight of a possibly singular Hermitian metric, then the natural square-integrability condition for a [holomorphic function](/page/Holomorphic%20Function) $f$ is $|f|^2e^{-\phi}\in L^1_{\mathrm{loc}}$. The multiplier ideal packages exactly this condition and records how much vanishing is forced near the singularities of $\phi$. [definition: Multiplier Ideal Sheaf] Let $X$ be a complex manifold and let $\phi:X\to[-\infty,\infty)$ be plurisubharmonic. The multiplier ideal sheaf $\mathcal I(\phi)\subset \mathcal O_X$ is the sheaf whose sections over an [open set](/page/Open%20Set) $U\subset X$ are \begin{align*} \mathcal I(\phi)(U)=\left\{f\in\mathcal O_X(U): |f|^2e^{-\phi}\in L^1_{\mathrm{loc}}(U)\right\}. \end{align*} [/definition] The sheaf property comes from the local nature of both holomorphicity and local integrability. The word "ideal" reflects the estimate $|gf|^2e^{-\phi}\le C|f|^2e^{-\phi}$ on compact coordinate neighbourhoods whenever $g$ is holomorphic. [example: Smooth Weight Gives No Vanishing] Choose a coordinate neighbourhood $U$ of $p$ on which a germ $f\in\mathcal O_{X,p}$ is represented by a [holomorphic function](/page/Holomorphic%20Function) and on which \begin{align*} A\le \phi \le B \end{align*} for some real constants $A,B$. Since $A\le \phi$, we have $-\phi\le -A$, hence \begin{align*} e^{-\phi}\le e^{-A}. \end{align*} After shrinking to a relatively compact coordinate ball $V\Subset U$, holomorphicity gives a constant $C$ with $|f|\le C$ on $V$. Therefore \begin{align*} 0\le |f|^2e^{-\phi}\le C^2e^{-A} \end{align*} on $V$, and the right-hand side is integrable because $V$ has finite Euclidean volume in coordinates. Thus $|f|^2e^{-\phi}\in L^1_{\mathrm{loc}}$ near $p$, so every holomorphic germ belongs to $\mathcal I(\phi)_p$. The reverse inclusion is built into the definition of $\mathcal I(\phi)$ as a subsheaf of $\mathcal O_X$, hence $\mathcal I(\phi)_p=\mathcal O_{X,p}$. [/example] The first non-smooth case is the logarithmic pole at the origin. It gives the model computation used throughout the chapter. [example: Radial Logarithmic Pole] In $\mathbb C^n$, let $\phi(z)=c\log |z|^2$ near $0$, where $c>0$ and $|z|^2=|z_1|^2+\cdots+|z_n|^2$. For $0<|z|<\varepsilon$, \begin{align*} e^{-\phi(z)} =e^{-c\log |z|^2} =(|z|^2)^{-c} =|z|^{-2c}. \end{align*} Let $f\in\mathcal O_{\mathbb C^n,0}$ have vanishing order $m$ at $0$. Write its Taylor expansion as \begin{align*} f(z)=P_m(z)+P_{m+1}(z)+\cdots, \end{align*} where $P_m$ is a nonzero homogeneous polynomial of degree $m$. For $z=r\zeta$ with $\zeta\in S^{2n-1}$, \begin{align*} f(r\zeta) =r^m\left(P_m(\zeta)+rR(r,\zeta)\right) \end{align*} with $R$ bounded for $0<r<\varepsilon$ after shrinking $\varepsilon$. Hence \begin{align*} \int_{S^{2n-1}} |f(r\zeta)|^2\,d\sigma(\zeta) =r^{2m}\int_{S^{2n-1}} |P_m(\zeta)+rR(r,\zeta)|^2\,d\sigma(\zeta). \end{align*} As $r\to0$, the angular integral tends to \begin{align*} \int_{S^{2n-1}} |P_m(\zeta)|^2\,d\sigma(\zeta)>0, \end{align*} because $P_m$ is not identically zero. Therefore the local integrability of $|f|^2e^{-\phi}$ is equivalent to the convergence of \begin{align*} \int_0^\varepsilon r^{2m}r^{-2c}r^{2n-1}\,dr = \int_0^\varepsilon r^{2m-2c+2n-1}\,dr. \end{align*} For $p=2m-2c+2n-1$, the integral $\int_0^\varepsilon r^p\,dr$ converges exactly when $p>-1$, so here the condition is \begin{align*} 2m-2c+2n-1&>-1,\\ 2m-2c+2n&>0,\\ m&>c-n. \end{align*} Since $m+n$ is an integer, $m>c-n$ is equivalent to $m+n>c$, hence to \begin{align*} m+n\ge \lfloor c\rfloor+1, \end{align*} and therefore to \begin{align*} m\ge \lfloor c\rfloor-n+1. \end{align*} Taking into account that every vanishing order satisfies $m\ge0$, the required order is \begin{align*} k=\max\{0,\lfloor c\rfloor-n+1\}. \end{align*} Thus the admissible germs are exactly those vanishing to order at least $k$, so \begin{align*} \mathcal I(c\log |z|^2)_0=\mathfrak m_0^k, \end{align*} where $\mathfrak m_0$ is the maximal ideal of germs vanishing at $0$. [/example] This example shows the general principle: stronger singularities force higher vanishing. Multiplier ideals convert analytic blow-up rates into algebraic conditions on Taylor coefficients. ## Coherence from the $L^2$ Method Why should the integrability condition above define a finitely generated analytic object? The definition is local and analytic, while coherence is an algebraic finiteness statement. Nadel's coherence theorem says that the gap is bridged by solving $\bar\partial$ with estimates: local $L^2$ solutions allow approximation of arbitrary admissible germs by finitely many generators. [quotetheorem:3717] Coherence is much stronger than the sheaf property: it says that each stalk is finitely generated and that those generators persist on a neighbourhood. This finiteness is what lets $\mathcal I(\phi)$ enter Cartan-Serre cohomology and exact-sequence arguments; a sheaf defined only by local integrability would not be enough for Nadel vanishing. The plurisubharmonic hypothesis is essential: for a merely upper-semicontinuous weight, the integrability condition can encode uncontrolled local singularities, and the Hörmander estimate used to produce finite generators is unavailable. The theorem also does not say that $\mathcal I(\phi)$ is globally generated, or that a single set of functions generates it on all of $X$; it is a local finite-generation statement. [citeproof:3717] The theorem is a prototype for how analytic estimates create algebraic finiteness. Its conclusion is existential rather than computational: it guarantees local finite generation of the multiplier ideal, while explicit descriptions still require additional structure such as normal crossings or a resolution. This distinction matters in applications because coherence supplies a finite coefficient sheaf for cohomological arguments, whereas examples identify what that sheaf actually is in concrete coordinates. [remark: Strong Openness] A deeper refinement says that $\mathcal I(\phi)=\bigcup_{\varepsilon>0}\mathcal I((1+\varepsilon)\phi)$. This [strong openness theorem](/theorems/3740) is not needed for the basic Nadel vanishing argument, but it explains why multiplier ideals are stable under a small strengthening of the singularity. [/remark] ## Singular Metrics and Nadel Vanishing How can a positive line bundle still give vanishing when its metric has singularities? The singularities prevent arbitrary holomorphic sections from being used, so the coefficient sheaf must be twisted by the multiplier ideal. After that correction, the curvature positivity still drives the $L^2$ estimate and kills higher cohomology. [definition: Singular Hermitian Metric] Let $L\to X$ be a holomorphic line bundle. A singular Hermitian metric on $L$ is given locally over an [open set](/page/Open%20Set) $U\subset X$, in a holomorphic frame $e$, by \begin{align*} |e|_h^2=e^{-\phi}, \end{align*} where $\phi:U\to[-\infty,\infty)$ is a function in $L^1_{\mathrm{loc}}(U)$. [/definition] If the local weights $\phi$ are plurisubharmonic, the curvature current of $h$ is positive. Locally it is represented by $i\partial\bar\partial\phi$, and changing frames modifies $\phi$ by the logarithm of the squared modulus of a nowhere-vanishing [holomorphic function](/page/Holomorphic%20Function). Because the weight changes under a frame change, the integrability condition must be phrased in a way that is independent of the chosen trivialisation. The next definition packages the local multiplier ideals into an intrinsic sheaf attached to the metric itself, which is the object that can be used in vanishing theorems. [definition: Multiplier Ideal of a Singular Metric] Let $(L,h)$ be a holomorphic line bundle with singular Hermitian metric whose local weight in a frame $e$ is $\phi$. The multiplier ideal sheaf $\mathcal I(h)$ is the sheaf locally equal to $\mathcal I(\phi)$. [/definition] At this point the multiplier ideal has done two jobs: it records the local integrability forced by the singular metric, and coherence makes it legal as a coefficient sheaf. The next theorem explains why this correction is the right one: it is the maximal local restriction needed to run the global $\bar\partial$ estimate without losing positivity. In applications the metric is often singular because the line bundle is big rather than ample, so the vanishing theorem must remember exactly which sections remain $L^2$ near the singular set. [quotetheorem:3718] The theorem says that positivity survives after removing exactly the sections that are too singular for the metric. Compactness is part of the global conclusion: on non-compact Kähler manifolds, Kodaira-type vanishing for sheaf cohomology can fail without extra hypotheses controlling behaviour at infinity. The strict lower bound $\varepsilon\omega$ is also not cosmetic; the Bochner-Kodaira inequality needs a positive spectral gap in degree $q\ge 1$, while semipositivity can leave [harmonic representatives](/theorems/2747). The multiplier ideal cannot be omitted: if $X$ is the blow-up of $\mathbb P^2$ at a point, with $H$ the pullback of a line and $E$ the exceptional divisor, the big line bundle $L=\mathcal O_X(3H+E)$ admits a singular metric with strictly positive curvature current, but $H^1(X,K_X\otimes L)=H^1(X,\mathcal O_X(2E))\ne0$; the singularities of the metric are exactly what $\mathcal I(h)$ records. [citeproof:3718] A useful special case occurs when the metric is smooth. Then $\mathcal I(h)=\mathcal O_X$, and Nadel vanishing reduces to the Kodaira-type vanishing theorem obtained earlier from the smooth Bochner-Kodaira identity. [remark: Relation to Kawamata Viehweg Vanishing] In algebraic geometry, Kawamata-Viehweg vanishing concerns line bundles of the form $K_X\otimes L$ where $L$ is big and nef, often after adding a boundary divisor with controlled coefficients. Analytically, the bigness is encoded by a singular positively curved metric and the boundary contribution is encoded by its multiplier ideal. Nadel vanishing is therefore the analytic framework behind many forms of Kawamata-Viehweg vanishing. [/remark] The simplest algebraic singularities come from divisors, where the multiplier ideal is read from one-variable integrability along each branch. A simple normal crossings divisor is the first case where the local geometry is still visible in coordinates: each branch contributes an independent one-variable pole, and the multiplier ideal records the amount of vanishing needed along that branch. This calculation is the bridge between analytic weights and algebraic boundary coefficients, because the exponents become rounding rules for divisors after choosing coordinates adapted to the crossings. [illustration:scv-iii-snc-divisor-integrability] [example: Divisorial Singularities] Fix $p\in X$. After reindexing, suppose $D_1,\dots,D_r$ are the components passing through $p$; on a simple normal crossings coordinate polydisc with coordinates $z_1,\dots,z_n$, write $g_j=u_jz_j$ for $1\le j\le r$, where each $u_j$ is nowhere zero, and the remaining $g_j$ are nowhere zero. Put \begin{align*} \beta_j=c_ja_j,\qquad k_j=\lfloor \beta_j\rfloor . \end{align*} The functions $\log |u_j|^2$ and the terms coming from components not passing through $p$ are bounded on a smaller polydisc, so multiplying by their exponentials changes $e^{-\phi}$ only by positive bounded factors. Thus local integrability is equivalent to local integrability of \begin{align*} |f(z)|^2\prod_{j=1}^r |z_j|^{-2\beta_j}. \end{align*} Write the Taylor expansion \begin{align*} f(z)=\sum_{\alpha\in\mathbb N^n} b_\alpha z^\alpha . \end{align*} Using polar coordinates $z_j=\rho_je^{i\theta_j}$ and orthogonality of the characters on the torus, the angular integration kills all cross-terms: \begin{align*} \int_{[0,2\pi]^n}\left|\sum_{\alpha}b_\alpha \prod_{j=1}^n \rho_j^{\alpha_j}e^{i\alpha_j\theta_j}\right|^2\,d\theta =(2\pi)^n\sum_{\alpha}|b_\alpha|^2\prod_{j=1}^n \rho_j^{2\alpha_j}. \end{align*} Hence any monomial $z^\alpha$ occurring in $f$ contributes the one-variable factors \begin{align*} \int_0^\varepsilon \rho_j^{2\alpha_j-2\beta_j+1}\,d\rho_j \end{align*} for $1\le j\le r$. This integral converges exactly when \begin{align*} 2\alpha_j-2\beta_j+1&>-1,\\ \alpha_j&>\beta_j-1,\\ \alpha_j&\ge \lfloor\beta_j\rfloor, \end{align*} where the last equivalence uses $\alpha_j\in\mathbb N$. Therefore $|f|^2e^{-\phi}$ is locally integrable exactly when every monomial appearing in $f$ is divisible by $z_1^{k_1}\cdots z_r^{k_r}$. Equivalently, \begin{align*} f=z_1^{k_1}\cdots z_r^{k_r}h \end{align*} for some holomorphic $h$. Conversely, if $f$ has this form, then $h$ is bounded on a smaller polydisc and \begin{align*} |f|^2\prod_{j=1}^r |z_j|^{-2\beta_j} \le C\prod_{j=1}^r |z_j|^{2k_j-2\beta_j}, \end{align*} whose radial factors converge because $k_j>\beta_j-1$. Thus \begin{align*} \mathcal I(\phi)_p = \left(z_1^{\lfloor c_1a_1\rfloor}\cdots z_r^{\lfloor c_ra_r\rfloor}\right) = \mathcal O_X\left(-\sum_j \lfloor c_ja_j\rfloor D_j\right)_p . \end{align*} The rounding down records the exact integer vanishing order needed along each branch of the normal crossings divisor. [/example] ## Lelong Numbers and Jumping of Ideals How do we measure the strength of a plurisubharmonic singularity without choosing a special coordinate formula? The Lelong number extracts the logarithmic rate of growth at a point. It is coarse compared with the full multiplier ideal, but it gives useful thresholds for when the ideal must vanish at that point. [definition: Lelong Number] Let $\phi$ be plurisubharmonic in a neighbourhood of $p\in\mathbb C^n$. The Lelong number of $\phi$ at $p$ is \begin{align*} \nu(\phi,p)=\liminf_{z\to p}\frac{\phi(z)}{\log |z-p|}. \end{align*} [/definition] Since $\log |z-p|$ tends to $-\infty$, a larger Lelong number means a stronger logarithmic pole. For instance, $\phi(z)=c\log |z-p|^2$ has $\nu(\phi,p)=2c$ under this convention. The next question is how this single numerical invariant constrains the multiplier ideal. It cannot determine the whole ideal, because it forgets the direction and shape of the singularity, but it does give useful thresholds for when constant germs remain integrable and when every admissible germ must vanish at the point. [quotetheorem:3719] These inequalities are not sharp in every direction, but they give the key qualitative picture. Low Lelong number allows constants to remain integrable, while high Lelong number forces every admissible holomorphic germ to vanish. The gap between the two thresholds is real: in $\mathbb C^2$, the weights $\phi_1=\log(|z_1|^2+|z_2|^2)$ and $\phi_2=\log |z_1|^2$ both have Lelong number $2$ at $0$, but $\mathcal I(\phi_1)_0=\mathcal O_{\mathbb C^2,0}$ whereas $\mathcal I(\phi_2)_0=(z_1)$. Thus the Lelong number records a logarithmic rate, not the full shape or direction of the singularity. [citeproof:3719] The inequalities show both the strength and the limitation of Lelong numbers. They give immediate integrability consequences, but they do not provide a complete catalogue of multiplier ideals because different singularities can share the same logarithmic rate. To study the whole family attached to one weight, it is better to vary the coefficient multiplying the weight and record the exact parameters where the ideal changes. As $c$ increases, the condition $|f|^2e^{-c\phi}\in L^1_{\mathrm{loc}}$ becomes stricter, so the multiplier ideals form a decreasing family. The important data are not all real values of $c$, but the critical values where a new vanishing condition appears. The next definition names those critical parameters so that the later discussion can treat the multiplier ideals as a discrete filtration rather than an unstructured continuum. [definition: Jumping Number] Let $\phi$ be plurisubharmonic near $p\in X$. A positive real number $c$ is a jumping number of $\phi$ at $p$ if \begin{align*} \mathcal I(c\phi)_p\ne\mathcal I((c-\varepsilon)\phi)_p \end{align*} for every sufficiently small $\varepsilon>0$. [/definition] The family $\mathcal I(c\phi)$ decreases as $c$ increases. Jumping numbers mark the moments when the required vanishing order increases. [example: Jumping for the Radial Pole] For $\phi(z)=\log |z|^2$ on $\mathbb C^n$, the radial pole calculation with coefficient $c$ gives \begin{align*} \mathcal I(c\phi)_0 = \mathfrak m_0^{k(c)}, \qquad k(c)=\max\{0,\lfloor c\rfloor-n+1\}. \end{align*} If $c<n$, then $\lfloor c\rfloor\le n-1$, so \begin{align*} \lfloor c\rfloor-n+1\le 0, \qquad k(c)=0, \qquad \mathcal I(c\phi)_0=\mathcal O_{\mathbb C^n,0}. \end{align*} At $c=n$, however, \begin{align*} k(n)=\max\{0,n-n+1\}=1, \end{align*} while for every $0<\varepsilon<1$, \begin{align*} k(n-\varepsilon) =\max\{0,\lfloor n-\varepsilon\rfloor-n+1\} =\max\{0,(n-1)-n+1\} =0. \end{align*} Thus $\mathcal I(n\phi)_0=\mathfrak m_0$ but $\mathcal I((n-\varepsilon)\phi)_0=\mathcal O_{\mathbb C^n,0}$, so $n$ is a jumping number. More generally, for an integer $\ell\ge0$ and $0<\varepsilon<1$, \begin{align*} k(n+\ell) &=\max\{0,\lfloor n+\ell\rfloor-n+1\} =\ell+1,\\ k(n+\ell-\varepsilon) &=\max\{0,\lfloor n+\ell-\varepsilon\rfloor-n+1\}\\ &=\max\{0,(n+\ell-1)-n+1\} =\ell. \end{align*} Hence \begin{align*} \mathcal I((n+\ell)\phi)_0=\mathfrak m_0^{\ell+1} \ne \mathfrak m_0^\ell = \mathcal I((n+\ell-\varepsilon)\phi)_0, \end{align*} so $n,n+1,n+2,\dots$ are jumping numbers. If $c$ is not one of these integers, then $\lfloor t\rfloor$ is constant for all $t<c$ sufficiently close to $c$, and therefore $k(t)=k(c)$ for such $t$; no jump occurs there. Thus the jumping numbers at $0$ are exactly \begin{align*} n,n+1,n+2,\dots . \end{align*} The dimension shifts the first positive vanishing condition: constants remain integrable against $|z|^{-2c}$ precisely for $c<n$, and at $c=n$ the multiplier ideal first forces vanishing at the origin. [/example] Hypersurface singularities give a less symmetric test case, because the geometry of the zero set changes the resulting ideal. For a smooth divisor, all singularity is concentrated in one coordinate direction, so the one-variable threshold controls the answer. Once the hypersurface is singular, different branches and exceptional divisors can contribute different vanishing requirements, and two equations with comparable orders at the origin may have different multiplier ideals. [example: Hypersurface Singularity Detection] Let $g\in\mathcal O_{\mathbb C^n,0}$ define $D=\{g=0\}$, and set $\phi=2c\log |g|$. Since \begin{align*} e^{-\phi} = e^{-2c\log |g|} = |g|^{-2c}, \end{align*} a germ $f$ belongs to $\mathcal I(\phi)_0$ exactly when $|f|^2|g|^{-2c}$ is locally integrable near $0$. First suppose $D$ is smooth and choose coordinates with $g=z_1$. Write $z=(z_1,z')$ and expand \begin{align*} f(z_1,z')=\sum_{\ell\ge 0}z_1^\ell a_\ell(z'), \qquad a_\ell\in\mathcal O_{\mathbb C^{n-1},0}. \end{align*} On a small polydisc, polar coordinates $z_1=\rho e^{i\theta}$ and orthogonality of the functions $e^{i\ell\theta}$ give \begin{align*} \int_0^{2\pi}|f(\rho e^{i\theta},z')|^2\,d\theta = 2\pi\sum_{\ell\ge0}\rho^{2\ell}|a_\ell(z')|^2 . \end{align*} Thus a nonzero coefficient $a_\ell$ contributes the radial factor \begin{align*} \int_0^\varepsilon \rho^{2\ell-2c+1}\,d\rho . \end{align*} This integral converges exactly when \begin{align*} 2\ell-2c+1&>-1,\\ \ell&>c-1,\\ \ell&\ge \lfloor c\rfloor, \end{align*} where the last step uses $\ell\in\mathbb N$. Therefore all coefficients $a_0,\dots,a_{\lfloor c\rfloor-1}$ must vanish, equivalently \begin{align*} f=z_1^{\lfloor c\rfloor}h \end{align*} for some holomorphic $h$. Conversely, if $f=z_1^{\lfloor c\rfloor}h$, then $h$ is bounded on a smaller polydisc and the remaining one-variable factor has exponent \begin{align*} 2\lfloor c\rfloor-2c+1>-1, \end{align*} so it is integrable. Hence \begin{align*} \mathcal I(2c\log |z_1|)_0=(z_1)^{\lfloor c\rfloor}. \end{align*} For a singular hypersurface, choose a log resolution $\mu:Y\to\mathbb C^n$ near $0$. In local coordinates $w_1,\dots,w_n$ on $Y$, the pullbacks have the form \begin{align*} g\circ\mu &= u(w)\prod_{i=1}^s w_i^{a_i},\\ \operatorname{Jac}_{\mathbb C}(\mu) &= v(w)\prod_{i=1}^s w_i^{b_i}, \end{align*} where $u$ and $v$ are nowhere zero, $a_i\ge1$, and $b_i\ge0$. The real change-of-variables formula gives \begin{align*} |f|^2|g|^{-2c}\,d\mathcal L_z \quad\longleftrightarrow\quad |f\circ\mu|^2 \prod_{i=1}^s |w_i|^{-2ca_i+2b_i}\,d\mathcal L_w \end{align*} up to positive bounded factors coming from $u$ and $v$. If a monomial $w^\alpha$ occurs in $f\circ\mu$, then its $w_i$-radial factor is \begin{align*} \int_0^\varepsilon \rho_i^{2\alpha_i-2ca_i+2b_i+1}\,d\rho_i, \end{align*} which converges exactly when \begin{align*} 2\alpha_i-2ca_i+2b_i+1&>-1,\\ \alpha_i+b_i&>ca_i-1,\\ \alpha_i&\ge \lfloor ca_i\rfloor-b_i . \end{align*} Thus each divisor $w_i=0$ on the resolution imposes the vanishing condition \begin{align*} \operatorname{ord}_{w_i}(f\circ\mu)\ge \lfloor ca_i\rfloor-b_i \end{align*} whenever the right-hand side is positive. The smooth case has only $a_1=1$ and $b_1=0$, giving the single condition $\operatorname{ord}_{z_1}(f)\ge\lfloor c\rfloor$; singular hypersurfaces acquire further conditions from exceptional divisors, so the multiplier ideal records more than the order of $g$ at the origin. [/example] ## Skoda Division and the Effective Nullstellensatz When does a [holomorphic function](/page/Holomorphic%20Function) belong to the ideal generated by given holomorphic functions? Algebra gives membership criteria through syzygies and Gröbner bases, while the $L^2$ method gives an analytic criterion: if the quotient by a suitable power of the generators is square-integrable, then division is possible. A weaker condition such as vanishing on the same zero set is not enough: in $\mathcal O_{\mathbb C,0}$, the germ $f=z$ vanishes on the zero set of $g=z^2$, but $z\notin(z^2)$, and the Skoda density $|z|^2|z^2|^{-2}=|z|^{-2}$ is not locally integrable. The theorem below measures precisely the extra vanishing needed to pass from radical membership to actual division by the generators. [quotetheorem:3720] Skoda's theorem is the division counterpart of Nadel coherence. Its hypotheses say that $f$ vanishes enough relative to the common zero set of the $g_j$ to absorb one copy of the ideal $(g_1,\dots,g_m)$. [citeproof:3720] Skoda division is local and analytic: it gives holomorphic quotients once integrability has encoded enough vanishing near the common zero set. This is stronger than set-theoretic vanishing but still does not by itself explain what happens for polynomial ideals on all of affine space. Passing to the algebraic setting introduces a new obstruction, namely controlling coefficients globally rather than merely solving near one point. The global question is whether a polynomial that vanishes on all common zeros of an ideal must satisfy an actual algebraic membership statement after taking a power. The [algebraic Nullstellensatz](/theorems/3721) gives exactly that global membership principle, and it is the target that analytic division will later recover with growth control for polynomial coefficients. [quotetheorem:3721] This is the familiar Nullstellensatz conclusion, but the method is analytic. The absence of common zeros gives a lower bound for $|F|^2=\sum_j|F_j|^2$ on compact sets and a controlled growth problem at infinity. [citeproof:3721] Global Nullstellensatz arguments give polynomial coefficients, but local analytic problems usually ask for membership near a common zero set before any degree bound is relevant. The local version of Skoda division turns that question into a scalar integrability estimate, so the same multiplier-ideal information can be checked chart by chart and then passed to the sheaf. This is often the practical form used in singularity computations. [example: Membership by an L² Test] Let $p\in\Omega$ and choose a coordinate ball $B\Subset\Omega$ centered at $p$; the ball $B$ is pseudoconvex. Write \begin{align*} |g|^2=|g_1|^2+\cdots+|g_m|^2, \qquad q=\min\{n,m-1\}. \end{align*} We show that the germ $f_p$ belongs to the stalk $\mathfrak a_p$. If $q\ge1$, set \begin{align*} \alpha=1+\frac{\varepsilon}{q}. \end{align*} Then $\alpha>1$ and \begin{align*} \alpha q+1 &=\left(1+\frac{\varepsilon}{q}\right)q+1\\ &=q+\varepsilon+1. \end{align*} Thus the assumed local integrability gives \begin{align*} \int_B |f|^2|g|^{-2(\alpha q+1)}\,d\mathcal L^{2n} = \int_B |f|^2|g|^{-2(q+1+\varepsilon)}\,d\mathcal L^{2n} <\infty . \end{align*} Applying *[Skoda Division Theorem](/theorems/3722)* on $B$ with $\psi=0$, there exist $h_1,\dots,h_m\in\mathcal O(B)$ such that \begin{align*} f=\sum_{j=1}^m g_jh_j \end{align*} on $B$. Hence $f_p\in(g_1,\dots,g_m)_p=\mathfrak a_p$. If $q=0$, then for any $\alpha>1$ the Skoda exponent is \begin{align*} \alpha q+1=\alpha\cdot0+1=1. \end{align*} Since the holomorphic tuple $g$ is bounded on $B$, choose $M>0$ with $|g|\le M$ on $B$. Then \begin{align*} |f|^2|g|^{-2} &=|f|^2|g|^{-2(1+\varepsilon)}|g|^{2\varepsilon}\\ &\le M^{2\varepsilon}|f|^2|g|^{-2(1+\varepsilon)}. \end{align*} The right-hand side is integrable on $B$, so *[Skoda Division Theorem](/theorems/3722)* again gives holomorphic functions $h_j$ on $B$ with \begin{align*} f=\sum_{j=1}^m g_jh_j . \end{align*} Since this holds near every point $p$, the function $f$ lies locally in the ideal sheaf $\mathfrak a$. The weighted $L^2$ condition is therefore a local analytic test for actual division by the generators $g_1,\dots,g_m$. [/example] These applications explain the role of multiplier ideals in the course. They are born from $L^2$ estimates, they are coherent enough for sheaf cohomology, and they retain enough analytic information about singularities to prove algebraic membership and vanishing theorems. Chapter 8 used multiplier ideals and sheaf cohomology to prove vanishing theorems; this chapter applies the L² method to a membership and division problem. Given holomorphic generators and a function f, Skoda's theorem decides whether f lies in their ideal by encoding divisor singularities and generators into a weight. The division problem becomes an L² solvability question, with the weight controlling when holomorphic quotients exist. # 9. Skoda's Division Theorem and Analytic Geometry The preceding chapters developed Hörmander's $L^2$ method as a way to solve $\bar\partial$ with quantitative control. This chapter uses that method for a division problem: a [holomorphic function](/page/Holomorphic%20Function) $f$ is given, together with holomorphic generators $g_1,\dots,g_m$, and the task is to decide whether $f$ can be written as $\sum_j g_j u_j$ with the coefficients $u_j$ still square-integrable. The result is Skoda's division theorem, which turns singular weighted integrability near the common zero set of the $g_j$ into membership in the analytic ideal generated by them. The final applications connect this analytic estimate with algebraic geometry, especially the Briancon-Skoda theorem and a quantitative form of the Nullstellensatz. ## The $L^2$ Division Problem When does an equation of the form $\sum_j g_j u_j=f$ admit holomorphic coefficients with an $L^2$ bound? In one complex variable this is close to asking whether the zeros of $g$ are also zeros of $f$ with sufficient multiplicity. In several variables the common zero set of $g_1,\dots,g_m$ may have complicated geometry, and divisibility is measured by integrability against powers of $|g|^{-1}$ rather than by a single order of vanishing. We use the following notation throughout the chapter. Let $\Omega\subset\mathbb C^n$ be a domain, let $g=(g_1,\dots,g_m)$ be a tuple of holomorphic functions on $\Omega$, and set \begin{align*} |g|^2=\sum_{j=1}^m |g_j|^2. \end{align*} The zero set of $g$ is $Z(g)=\{z\in\Omega: g_1(z)=\cdots=g_m(z)=0\}$. [definition: Bergman Space] Let $\Omega\subset\mathbb C^n$ be a domain. The Bergman space $A^2(\Omega)$ is \begin{align*} A^2(\Omega)=\left\{f:\Omega\to\mathbb C\mid f\text{ is holomorphic and }\int_\Omega |f|^2\,d\mathcal L^{2n}<\infty\right\}. \end{align*} [/definition] The division problem asks not only for holomorphic coefficients, but also for control of their $A^2$ size. Ordinary ideal membership allows arbitrary holomorphic quotients, but an $L^2$ argument only produces quotients whose square integrals are controlled. This creates a separate notion of generated ideal: a function may be algebraically divisible by $g$ while the required quotients fail to lie in the Bergman space, especially near the common zero set or near the boundary of $\Omega$. [definition: Bergman Ideal] Let $g_1,\dots,g_m\in\mathcal O(\Omega)$. The Bergman ideal generated by $g$ is \begin{align*} I_{A^2}(g)=\left\{\sum_{j=1}^m g_j u_j\mid u_1,\dots,u_m\in A^2(\Omega)\right\}. \end{align*} [/definition] The notation hides a real difficulty: multiplication by $g_j$ may improve integrability near $Z(g)$, while division by $g_j$ may create singularities. Skoda's theorem gives an exact quantitative hypothesis ensuring that these singularities remain square-integrable. [example: Vanishing Without Ideal Membership] Near $0\in\mathbb C^2$, let $g_1(z,w)=z^2$ and $g_2(z,w)=w$. Then \begin{align*} Z(g)&=\{(z,w):z^2=0,\ w=0\}\\ &=\{(0,0)\}, \end{align*} so $f(z,w)=z$ satisfies $f|_{Z(g)}=0$. We show that this set-theoretic vanishing does not imply $z\in (z^2,w)\mathcal O_{\mathbb C^2,0}$. Suppose, for contradiction, that there are germs $u,v\in\mathcal O_{\mathbb C^2,0}$ such that \begin{align*} z=z^2u+wv. \end{align*} Restrict this identity to the complex line $\{w=0\}$, using the homomorphism $\mathcal O_{\mathbb C^2,0}\to \mathcal O_{\mathbb C,0}$ given by $h(z,w)\mapsto h(z,0)$. This gives \begin{align*} z &=z^2u(z,0)+0\cdot v(z,0)\\ &=z^2u(z,0). \end{align*} Hence, in the one-variable local ring $\mathcal O_{\mathbb C,0}$, \begin{align*} 0&=z^2u(z,0)-z\\ &=z\bigl(zu(z,0)-1\bigr). \end{align*} The ring $\mathcal O_{\mathbb C,0}$ is the ring of convergent [power series](/page/Power%20Series) in one variable, hence an integral domain. Since the germ $z$ is not zero, the equality above forces \begin{align*} zu(z,0)-1=0. \end{align*} Evaluating this germ at $z=0$ gives \begin{align*} 0\cdot u(0,0)-1=-1, \end{align*} which cannot equal $0$. Therefore $z\notin (z^2,w)\mathcal O_{\mathbb C^2,0}$. The example shows that vanishing on the common zero set records only set-theoretic information, while ideal membership requires the stronger order information measured later by Skoda's singular weight. [/example] The example leaves a precise analytic problem: replace set-theoretic vanishing by an integrability condition strong enough to force actual division by the tuple $g=(g_1,\ldots,g_m)$. Skoda's division theorem supplies that missing quantitative condition by measuring whether $f$ vanishes fast enough, in an $L^2$ sense, for the tentative quotients by the $g_j$ to be corrected into genuine square-integrable holomorphic coefficients. [quotetheorem:3722] [citeproof:3722] The exponent $\alpha q+1$ is the cost of division. The extra $1$ reflects the first non-holomorphic quotient $\overline{g_j}f/|g|^2$, while $q=\min(n,m-1)$ measures the worst rank of the Koszul directions in which the correction must be solved. [example: Ideal Generated By Coordinates In The Ball] Let $\Omega=B(0,1)\subset\mathbb C^2$, let $g_1(z)=z_1$ and $g_2(z)=z_2$, and write $B=B(0,1)$. Then \begin{align*} |g(z)|^2 &=|g_1(z)|^2+|g_2(z)|^2\\ &=|z_1|^2+|z_2|^2\\ &=|z|^2, \end{align*} and, since $n=2$ and $m=2$, \begin{align*} q=\min(n,m-1)=\min(2,1)=1. \end{align*} [claim]For nonnegative integers $a,b$, the monomial $f(z)=z_1^a z_2^b$ lies in the Bergman ideal generated by $z_1,z_2$ exactly when $a+b\ge 1$.[/claim] [proof]Put $d=a+b$. With $\psi=0$ and $q=1$, the Skoda integrability condition for $f=z_1^a z_2^b$ is \begin{align*} I_\alpha =\int_B |z_1^a z_2^b|^2 |z|^{-2(\alpha+1)}\,d\mathcal L^4<\infty \end{align*} for some $\alpha>1$. Write $z=r\xi$, where $0<r<1$ and $\xi\in S^3\subset\mathbb C^2$. Then $d\mathcal L^4=r^3\,dr\,d\sigma(\xi)$, and \begin{align*} |z_1^a z_2^b|^2 |z|^{-2(\alpha+1)}\,d\mathcal L^4 &=r^{2a}|\xi_1|^{2a}r^{2b}|\xi_2|^{2b}r^{-2(\alpha+1)}r^3\,dr\,d\sigma(\xi)\\ &=|\xi_1|^{2a}|\xi_2|^{2b}r^{2d-2\alpha+1}\,dr\,d\sigma(\xi). \end{align*} Hence \begin{align*} I_\alpha = \left(\int_{S^3}|\xi_1|^{2a}|\xi_2|^{2b}\,d\sigma(\xi)\right) \left(\int_0^1 r^{2d-2\alpha+1}\,dr\right). \end{align*} The angular integral is finite because $S^3$ is compact and the integrand is continuous. For $\beta=2d-2\alpha+1$, \begin{align*} \int_\varepsilon^1 r^\beta\,dr = \frac{1-\varepsilon^{\beta+1}}{\beta+1} \end{align*} when $\beta\ne -1$, and $\int_\varepsilon^1 r^{-1}\,dr=-\log\varepsilon$ when $\beta=-1$. Therefore the radial integral is finite precisely when \begin{align*} 2d-2\alpha+1&>-1,\\ 2d-2\alpha&>-2,\\ d&>\alpha-1. \end{align*} If $d\ge1$, choose $\alpha=1+d/2$. Then $\alpha>1$ and \begin{align*} d>\frac d2=\alpha-1, \end{align*} so $I_\alpha<\infty$. By the stated *[Skoda Division Theorem](/theorems/3722)*, there exist holomorphic functions $u_1,u_2$ on $B$ such that \begin{align*} z_1^a z_2^b=z_1u_1+z_2u_2 \end{align*} and \begin{align*} \int_B \left(|u_1|^2+|u_2|^2\right)|z|^{-2\alpha}\,d\mathcal L^4<\infty. \end{align*} Since $|z|^{-2\alpha}\ge 1$ for $0<|z|<1$ and the point $z=0$ has measure zero, \begin{align*} \int_B |u_j|^2\,d\mathcal L^4 \le \int_B \left(|u_1|^2+|u_2|^2\right)|z|^{-2\alpha}\,d\mathcal L^4 <\infty \end{align*} for $j=1,2$. Thus $u_1,u_2\in A^2(B)$, so $z_1^a z_2^b$ lies in the Bergman ideal generated by $z_1,z_2$. Conversely, if $d=0$, then $f=1$. If $1$ belonged to the Bergman ideal, there would be $u_1,u_2\in A^2(B)$ with \begin{align*} 1=z_1u_1+z_2u_2. \end{align*} Since $u_1$ and $u_2$ are holomorphic, evaluating at $z=0$ gives \begin{align*} 1=0\cdot u_1(0)+0\cdot u_2(0)=0, \end{align*} which is impossible. Therefore the constant monomial is not in the ideal.[/proof] The computation shows that, for the coordinate ideal in the ball, the singular weight detects exactly whether the monomial vanishes at the common zero of $z_1$ and $z_2$. [/example] This example shows the role of codimension two: membership is detected by integrability at the common zero, not by division by either generator separately. The theorem packages that local test into a global $L^2$ construction. ## The Weighted $\bar{\partial}$ Estimate What feature of the $\bar\partial$ method makes division possible instead of merely solving an unrelated equation? The point is that the correction term must stay in the Koszul kernel $K_z$, and the singular logarithmic weight must produce enough positive curvature to compensate for the motion of this kernel. Let $K\subset\Omega\times\mathbb C^m$ denote the holomorphic vector subbundle over $\Omega\setminus Z(g)$ given by \begin{align*} K_z=\left\{a=(a_1,\dots,a_m)\in\mathbb C^m\mid \sum_{j=1}^m g_j(z)a_j=0\right\}. \end{align*} The proof works first on $\Omega\setminus Z(g)$ and then removes the zero set by approximation. The singularity at $Z(g)$ is not an obstacle; it is the source of the integrability condition in Skoda's theorem. [illustration:scv-iii-koszul-kernel-degeneration] [quotetheorem:3723] [citeproof:3723] This estimate is the analytic heart of the chapter. It turns algebraic information about the tuple $g$ into a closed range estimate for the constrained $\bar\partial$ problem. [explanation: Hilbert Space Interpretation] Let $T$ be multiplication by $g$, so that $T(u_1,\dots,u_m)=\sum_j g_j u_j$, and let $S=\bar\partial$ acting on vector-valued functions. Since $T\circ S=0$ on vectors in the Koszul kernel, the division problem is governed by a two-step Hilbert complex. The weighted estimate bounds the functional $a\mapsto(a,\bar\partial v^0)_\Phi$ by the graph norm of $S$ and $S^*$, which gives a solution to $Sw=\bar\partial v^0$ with $Tw=0$. Setting $u=v^0-w$ then places the solution back in the holomorphic kernel of $S$ while preserving $Tu=f$. [/explanation] The proof therefore has the same architecture as Hörmander's theorem, but with a new algebraic constraint. The price of that constraint is exactly the Skoda exponent. ## Strictly Pseudoconvex Domains and Closure of Ideals How does the $L^2$ division theorem compare with ordinary holomorphic ideal membership on a bounded domain? On strictly pseudoconvex domains, the answer is that the analytic $A^2$ ideal agrees with the algebraic holomorphic ideal after intersecting with the functions under consideration. [definition: Holomorphic Algebra On A Bounded Domain] Let $\Omega\subset\mathbb C^n$ be bounded. The algebra $A(\Omega)$ is \begin{align*} A(\Omega)=\mathcal O(\Omega)\cap C(\overline{\Omega}). \end{align*} [/definition] For $g_1,\dots,g_m\in A(\Omega)$, write \begin{align*} I_{\mathrm{alg}}(g)=\left\{\sum_{j=1}^m g_j h_j\mid h_1,\dots,h_m\in A(\Omega)\right\}. \end{align*} The Bergman ideal $I_{A^2}(g)$ permits coefficients in $A^2(\Omega)$ instead. The comparison is delicate because algebraic coefficients continuous on $\overline\Omega$ are much more rigid than arbitrary square-integrable holomorphic coefficients, while local ideal membership only sees germs near each point. The obstruction is to turn local divisibility data into global $A^2$ quotients without introducing boundary singularities. The next issue is to compare two kinds of membership: local membership in each germ ring and global membership with Bergman-space coefficients. Strict pseudoconvexity supplies the analytic control needed for this passage, and Skoda's corollary identifies the resulting $A^2$ ideal with the intersection of the local analytic ideals. [quotetheorem:3724] [citeproof:3724] The statement is useful because local analytic geometry is often expressed in terms of germs, while the $\bar\partial$ method produces global square-integrable functions. Strict pseudoconvexity is the bridge that lets local membership data be promoted to global $L^2$ membership. [remark: Role Of Strict Pseudoconvexity] Strict pseudoconvexity is not a decorative boundary assumption. It gives the exhaustion and boundary positivity needed for uniform weighted estimates, and it prevents boundary singularities from interfering with the ideal-theoretic statement. [/remark] ## Effective Briancon-Skoda And The Nullstellensatz Can Skoda division turn vanishing on a common zero set into an explicit algebraic power membership statement? This is the bridge from $L^2$ analysis to the Nullstellensatz: vanishing gives integrability of a sufficiently high power, and the division theorem converts that integrability into membership in the ideal. [definition: Integral Closure Condition For An Ideal] Let $g_1,\dots,g_m\in\mathcal O(\Omega)$ and $f\in\mathcal O(\Omega)$. The function $f$ satisfies the integral closure bound for $(g_1,\dots,g_m)$ on a compact set $K\subset\Omega$ if there is a constant $C_K>0$ such that \begin{align*} |f(z)|\le C_K |g(z)| \end{align*} for all $z\in K$. [/definition] This analytic inequality is a quantitative version of saying that $f$ vanishes along $Z(g)$ at least as strongly as the generators do. The point of introducing it is that it is stronger than set-theoretic vanishing but still checkable by estimates on compact subsets. The remaining analytic question is whether this metric vanishing is enough to perform division by the tuple $g$ without losing holomorphicity at the common zero set. The obstruction is the possible blow-up of the naive quotients near $Z(g)$; Skoda's estimate controls that singularity and converts the bound into actual ideal membership for a power of $f$. [quotetheorem:3725] [citeproof:3725] The phrase "sufficiently close to $1$" has a concrete meaning in the proof. The integral closure bound gives $|f|^{2(q+1)}\le C|g|^{2(q+1)}$ on compact subsets, so the Skoda integrand is bounded by a constant times $|g|^{2q(1-\alpha)}$. For $\alpha>1$ with $\alpha-1$ small, this is only a mild negative power of $|g|$, and the standard local integrability of small negative powers of holomorphic functions gives finiteness on $\Omega'$. The estimate therefore records not just membership but quantitative control of the coefficients. This theorem should not be read as saying that vanishing on $Z(g)$ already gives $f\in(g_1,\dots,g_m)$. The earlier example $(z^2,w)$ shows that set-theoretic vanishing can be too weak; the integral closure bound is a metric replacement for the missing order information. Nor does the theorem usually put $f$ itself in the ideal: it puts the power $f^{q+1}$ in the ideal. The exponent $q+1$ is the analytic cost of dividing by a tuple of generators, and in regular local models it is the sharp Briancon-Skoda exponent. The remaining problem is to start only from ordinary vanishing on the common zero set and still reach ideal membership after taking a power. The Lojasiewicz inequality supplies the missing metric estimate: if $f$ vanishes on $Z(g)$, then on compact subsets some power of $|f|$ is bounded by a constant times $|g|$. Combining that estimate with the Briancon-Skoda theorem gives a concrete exponent $N$ with $f^N\in(g_1,\dots,g_m)$. [quotetheorem:3726] [citeproof:3726] This is the analytic route to an algebraic conclusion. Instead of constructing syzygies directly, the proof obtains weighted estimates for $\bar\partial$ and reads off ideal membership from the resulting holomorphic division. [example: A Plane Curve Ideal] Let $\Omega$ be a small ball about $0$ in $\mathbb C^2$, and put \begin{align*} g_1(z,w)=z^2,\qquad g_2(z,w)=w^3. \end{align*} Then \begin{align*} |g(z,w)|^2 &=|g_1(z,w)|^2+|g_2(z,w)|^2\\ &=|z^2|^2+|w^3|^2\\ &=|z|^4+|w|^6, \end{align*} and \begin{align*} Z(g) &=\{(z,w):z^2=0,\ w^3=0\}\\ &=\{(0,0)\}. \end{align*} [claim]If, near $0$, there are an integer $k\ge 1$ and a constant $C>0$ such that \begin{align*} |f(z,w)|^k\le C\bigl(|z|^4+|w|^6\bigr)^{1/2}, \end{align*} then \begin{align*} f^{2k}\in (z^2,w^3)\mathcal O_{\mathbb C^2,0}. \end{align*} [/claim] [proof]The displayed inequality can be rewritten using $|g(z,w)|=\bigl(|z|^4+|w|^6\bigr)^{1/2}$ as \begin{align*} |f(z,w)^k|\le C|g(z,w)|. \end{align*} Thus $f^k$ satisfies the integral closure bound for the ideal generated by $z^2$ and $w^3$. Here $n=2$ and $m=2$, so \begin{align*} q=\min(n,m-1)=\min(2,1)=1. \end{align*} Applying the *[Skoda Nullstellensatz Consequence](/theorems/3726)* to the function $f^k$ gives \begin{align*} (f^k)^{q+1}\in (z^2,w^3)\mathcal O_{\mathbb C^2,0}. \end{align*} Since $q+1=2$, this is \begin{align*} (f^k)^2=f^{2k}\in (z^2,w^3)\mathcal O_{\mathbb C^2,0}. \end{align*} Equivalently, after possibly shrinking $\Omega$, there are holomorphic functions $u_1,u_2$ on $\Omega$ such that \begin{align*} f^{2k}=z^2u_1+w^3u_2. \end{align*} The division theorem first produces these coefficients with an $A^2$ estimate on smaller pseudoconvex balls, and the same theorem includes their holomorphicity.[/proof] Thus the exponent is explicit: a Lojasiewicz comparison with exponent $k$ forces the power $f^{2k}$ into the plane curve ideal $(z^2,w^3)$. [/example] The example illustrates why the theorem is called effective. The exponent is not merely existential; it is tied to a measurable vanishing comparison. ## Weierstrass Division As A One-Generator Case What remains of Skoda's theorem when the ideal has one generator? The Koszul kernel disappears, the rank parameter is $q=0$, and the theorem reduces to controlled division by a single [holomorphic function](/page/Holomorphic%20Function). For $m=1$, the hypothesis in Skoda's theorem becomes \begin{align*} \int_\Omega |f|^2 |g_1|^{-2}e^{-\psi}\,d\mathcal L^{2n}<\infty. \end{align*} The conclusion is that $f/g_1$ is holomorphic and square-integrable with the expected bound. In one complex variable this is exactly the statement that the zeros of $f$ dominate the zeros of $g_1$ with the required multiplicities. [example: Weierstrass Division from $L^2$ Division] Let \begin{align*} P(z,w)=z^d+c_{d-1}(w)z^{d-1}+\cdots+c_1(w)z+c_0(w) \end{align*} be a Weierstrass polynomial in $z$, and let $F$ be holomorphic near $(0,0)\in\mathbb C\times\mathbb C^{n-1}$. Choose $\rho>0$ and shrink the parameter neighbourhood $W$ of $0$ so that $F$ and the coefficients $c_\nu$ are holomorphic for $|z|\le\rho$, $w\in W$, and \begin{align*} P(\zeta,w)\ne0\qquad\text{whenever }|\zeta|=\rho,\ w\in W. \end{align*} This is possible because $P(\zeta,0)=\zeta^d$ is nonzero on the circle $|\zeta|=\rho$. [claim]After the polynomial remainder $R$ is chosen by the one-variable Cauchy construction in the $z$ variable, Skoda's one-generator division gives a holomorphic quotient $Q$ with \begin{align*} F(z,w)=Q(z,w)P(z,w)+R(z,w),\qquad R(z,w)=\sum_{\ell=0}^{d-1}a_\ell(w)z^\ell. \end{align*} [/claim] [proof]For $|z|<\rho$ and $w\in W$, define \begin{align*} R(z,w) = \frac{1}{2\pi i} \int_{|\zeta|=\rho} F(\zeta,w) \frac{P(\zeta,w)-P(z,w)}{P(\zeta,w)(\zeta-z)} \,d\zeta. \end{align*} The numerator quotient is a polynomial in $z$ of degree at most $d-1$, because \begin{align*} \frac{P(\zeta,w)-P(z,w)}{\zeta-z} &= \frac{\zeta^d-z^d}{\zeta-z} + \sum_{\nu=1}^{d-1}c_\nu(w)\frac{\zeta^\nu-z^\nu}{\zeta-z}\\ &= \sum_{r=0}^{d-1}\zeta^{d-1-r}z^r + \sum_{\nu=1}^{d-1}c_\nu(w)\sum_{r=0}^{\nu-1}\zeta^{\nu-1-r}z^r. \end{align*} Hence \begin{align*} R(z,w)=\sum_{\ell=0}^{d-1}a_\ell(w)z^\ell, \end{align*} where \begin{align*} a_\ell(w) = \frac{1}{2\pi i} \int_{|\zeta|=\rho} \frac{F(\zeta,w)}{P(\zeta,w)} \left( \zeta^{d-1-\ell} + \sum_{\nu=\ell+1}^{d-1}c_\nu(w)\zeta^{\nu-1-\ell} \right) \,d\zeta. \end{align*} Each $a_\ell$ is holomorphic in $w$, since the contour is fixed and the integrand is holomorphic in $w$ on a neighbourhood of $|\zeta|=\rho$. Now set \begin{align*} Q_0(z,w) = \frac{1}{2\pi i} \int_{|\zeta|=\rho} \frac{F(\zeta,w)}{P(\zeta,w)(\zeta-z)} \,d\zeta. \end{align*} Then $Q_0$ is holomorphic for $|z|<\rho$, $w\in W$, and \begin{align*} P(z,w)Q_0(z,w)+R(z,w) &= \frac{1}{2\pi i} \int_{|\zeta|=\rho} F(\zeta,w) \left( \frac{P(z,w)}{P(\zeta,w)(\zeta-z)} + \frac{P(\zeta,w)-P(z,w)}{P(\zeta,w)(\zeta-z)} \right) \,d\zeta\\ &= \frac{1}{2\pi i} \int_{|\zeta|=\rho} F(\zeta,w) \frac{P(\zeta,w)}{P(\zeta,w)(\zeta-z)} \,d\zeta\\ &= \frac{1}{2\pi i} \int_{|\zeta|=\rho} \frac{F(\zeta,w)}{\zeta-z}\,d\zeta\\ &=F(z,w), \end{align*} by the one-variable *[Cauchy Integral Formula](/theorems/345)*. Thus \begin{align*} F-R=P Q_0. \end{align*} If $\lambda$ is a zero of $P(\cdot,w)$ of multiplicity $s$, then \begin{align*} P(z,w)=(z-\lambda)^s h(z,w) \end{align*} with $h(\lambda,w)\ne0$, and therefore \begin{align*} F(z,w)-R(z,w)=(z-\lambda)^s h(z,w)Q_0(z,w). \end{align*} So $F-R$ vanishes along the divisor $P=0$ with the multiplicities encoded by $P$. On every smaller polydisc $U\Subset\{|z|<\rho,\ w\in W\}$, the [holomorphic function](/page/Holomorphic%20Function) $Q_0$ is bounded, say $|Q_0|\le M_U$. Since $(F-R)/P=Q_0$ away from $P=0$ and $Q_0$ extends holomorphically across $P=0$, \begin{align*} \int_U |F-R|^2|P|^{-2}\,d\mathcal L^{2n} = \int_U |Q_0|^2\,d\mathcal L^{2n} \le M_U^2\mathcal L^{2n}(U) <\infty. \end{align*} For one generator, $m=1$ and hence \begin{align*} q=\min(n,m-1)=\min(n,0)=0. \end{align*} Applying the $m=1$ case of the *[Skoda Division Theorem](/theorems/3722)* to $F-R$ and $g_1=P$ gives a holomorphic $Q\in A^2_{\mathrm{loc}}$ such that \begin{align*} F-R=PQ. \end{align*} Therefore \begin{align*} F=QP+R. \end{align*} Since $P(Q-Q_0)=0$ and $P$ is not identically zero, the quotient $Q$ agrees with the Cauchy quotient $Q_0$ on the dense set where $P\ne0$, and hence everywhere by continuity.[/proof] Thus the Cauchy construction supplies the finite polynomial remainder, while the one-generator $L^2$ division theorem supplies the holomorphic quotient after the divisor singularity has been cancelled. [/example] The comparison is useful pedagogically. Weierstrass division is local and algebraic in flavour, while Skoda division is global and metric; the one-generator case shows that the $L^2$ theorem extends classical divisibility rather than replacing it. ## Summary Of The Chapter The main question of the chapter was how to divide a holomorphic $L^2$ function by several holomorphic generators without losing control of the coefficients. Skoda's theorem answers this by imposing a weighted integrability condition against $|g|^{-2(\alpha q+1)}$ and producing coefficients with a sharp weighted $L^2$ estimate. The proof is a constrained Hörmander argument: start from the formal quotient $\overline{g_j}f/|g|^2$, solve a $\bar\partial$ correction inside the Koszul kernel, and use the logarithmic singularity of $|g|$ to win the needed curvature. The consequences explain why this analytic theorem belongs in analytic geometry: it identifies Bergman and algebraic ideal membership on strictly pseudoconvex domains and gives an effective Briancon-Skoda route to the Nullstellensatz. Chapters 4–9 assumed weights were either smooth or handled singularities through integral currents; this chapter removes the smoothness assumption. Singular metrics with logarithmic poles appear naturally in extension problems, and the Donnelly–Fefferman theorem extends the core ∂̄-estimates to such singularities. This final analytic step closes the loop from pure solvability theory back to the geometric singularities that motivated the entire study. # 10. The Donnelly–Fefferman Theorem and Singular Metrics The preceding chapters developed Hörmander-type estimates for smooth weights and then used them to prove extension and vanishing results. The final analytic step is to remove the artificial smoothness assumption: the weights that appear naturally in extension problems have logarithmic poles, and the metrics on line bundles may be singular. This chapter explains how approximation of plurisubharmonic weights, the Donnelly-Fefferman estimate on complete Kahler manifolds, and its twisted variants give a flexible $L^2$ theory for singular metrics. The central theme is that curvature positivity may be weak or distributional, but the $\bar{\partial}$-method still sees it through approximation and weak compactness. We then compare the resulting theory with classical compact Hodge theory and finish with the sharp-constant questions that led to the solution of the Suita conjecture. ## Singular Plurisubharmonic Weights and Approximation What breaks when the weight in an $L^2$ estimate is plurisubharmonic but not smooth? The Bochner-Kodaira calculation needs derivatives of the weight, while the applications produce functions such as $\log |g|^2$ or Green functions, which have poles and define curvature only as currents. As a concrete instance, the weight $\varphi = \log|g|^2$ for a nonconstant $g \in \mathcal O(\Omega)$ is not in $C^2(\Omega)$ near the zero set $\{g=0\}$: the second derivatives that enter the Bochner-Kodaira identity blow up, and $i\partial\bar{\partial}\varphi$ is the integration current along the divisor of $g$ rather than a pointwise form. The smooth $L^2$ machinery does not assign a meaning to $\|\bar{\partial}^*\alpha\|^2_{L^2(e^{-\varphi})}$ in the classical sense, so an approximation step is required before any quantitative estimate can be invoked. The approximation lemma is the device that lets us prove estimates for smooth weights and then pass to singular limits. A singular metric is best understood locally as an exponential weight multiplying a smooth metric. [definition: Singular Hermitian Metric] Let $X$ be a complex manifold, let $L \to X$ be a holomorphic line bundle, and let $h_0$ be a smooth Hermitian metric on $L$. A singular Hermitian metric on $L$ is a metric of the form \begin{align*} h = e^{-\varphi}h_0, \end{align*} where $\varphi \in L^1_{\mathrm{loc}}(X)$ is locally integrable and is not identically $-\infty$ on any connected component of $X$. [/definition] The weight $\varphi$ records the singularity. In a local holomorphic frame $e$ with $|e|_{h_0}^2=e^{-\psi_0}$, a section $s=fe$ has pointwise norm $|s|_h^2=|f|^2e^{-(\psi_0+\varphi)}$, so integrability of sections becomes a precise condition on vanishing along the singular set. To use such metrics in Bochner-Kodaira estimates, their positivity must still have a curvature meaning even when $\varphi$ is not twice differentiable. The correct definition replaces pointwise curvature forms by currents, allowing logarithmic poles while preserving the positivity condition needed for approximation and $L^2$ estimates. [definition: Curvature Current of a Singular Metric] Let $h=e^{-\varphi}h_0$ be a singular Hermitian metric on a holomorphic line bundle $L \to X$. The curvature current of $h$ is \begin{align*} i\Theta_h(L) = i\Theta_{h_0}(L) + i\partial\bar{\partial}\varphi. \end{align*} The metric $h$ has semipositive curvature if $i\Theta_h(L)$ is a positive $(1,1)$-current. [/definition] This definition turns plurisubharmonicity into positivity of curvature. The phrase "positive current" is deliberately weak enough to include logarithmic poles, and strong enough to survive monotone limits. [example: Logarithmic Pole Metric] Let $\Omega \subset \mathbb C^n$ be a domain, let $g\in\mathcal O(\Omega)$ be nonzero, and put $\varphi=c\log |g|^2$ with $c>0$ on the product line bundle $\Omega\times\mathbb C$. Since \begin{align*} \varphi=c\log |g|^2=2c\log |g|, \end{align*} and the logarithm of the modulus of a [holomorphic function](/page/Holomorphic%20Function) is plurisubharmonic, $i\partial\bar{\partial}\varphi$ is a positive $(1,1)$-current. In the standard global frame of $\Omega\times\mathbb C$, the reference metric is flat, so its curvature is $0$, and the curvature current of $h=e^{-\varphi}$ is \begin{align*} i\Theta_h = 0+i\partial\bar{\partial}\varphi = c\,i\partial\bar{\partial}\log |g|^2 \ge 0. \end{align*} Thus $h$ has semipositive curvature. Write a holomorphic section in the standard frame as $s=f$. Its squared pointwise norm is \begin{align*} |s|_h^2 &= |f|^2e^{-\varphi} \\ &= |f|^2e^{-c\log |g|^2} \\ &= |f|^2|g|^{-2c}. \end{align*} Near a point where the divisor of $g$ has simple normal crossings, choose local coordinates $z=(z_1,\ldots,z_n)$ such that \begin{align*} g(z)=u(z)z_1^{a_1}\cdots z_k^{a_k}, \end{align*} where $u$ is holomorphic and nonvanishing. If \begin{align*} f(z)=v(z)z_1^{b_1}\cdots z_k^{b_k}, \end{align*} with $v$ holomorphic and nonvanishing, then on a smaller coordinate polydisc there are constants $0<A\le B$ such that \begin{align*} A\prod_{\ell=1}^k |z_\ell|^{2b_\ell-2ca_\ell} \le |f|^2|g|^{-2c} \le B\prod_{\ell=1}^k |z_\ell|^{2b_\ell-2ca_\ell}. \end{align*} Thus local integrability is determined by the one-variable factors. For one variable, \begin{align*} \int_{|z|<\varepsilon}|z|^{2b-2ca}\,dA(z) &= \int_0^{2\pi}\int_0^\varepsilon r^{2b-2ca}r\,dr\,d\theta \\ &= 2\pi\int_0^\varepsilon r^{2b-2ca+1}\,dr, \end{align*} and this integral is finite exactly when $2b-2ca+1>-1$, equivalently $b>ca-1$. Hence $s$ is locally $L^2$ near this normal-crossings point only when \begin{align*} b_\ell>ca_\ell-1 \end{align*} for each $\ell=1,\ldots,k$. The logarithmic pole contributes semipositive curvature as a current, while the weighted $L^2$ condition forces holomorphic sections to vanish along each divisor component of $g$ with order controlled by $c$. [/example] The approximation lemma is the local analytic input behind the passage from smooth to singular weights. It does two things at once: it produces a sequence of genuinely smooth plurisubharmonic functions to which the Bochner-Kodaira machinery applies verbatim, and it arranges those functions to dominate the singular weight pointwise, so that the weighted norms $\int |\cdot|^2 e^{-\varphi_j}$ increase monotonically to $\int|\cdot|^2 e^{-\varphi}$. The first property feeds the smooth estimate; the second is what allows weak limits and lower semicontinuity to recover a singular solution from the smooth ones. Crucially, the construction is local, so the manifold version requires a partition argument that we record in passing. [quotetheorem:3727] The theorem says that singular positivity can be reached from above by smooth positivity on compactly contained coordinate patches. [citeproof:3727] For $L^2$ estimates, decreasing approximation is important because the weighted norms increase as the singularity appears: \begin{align*} e^{-\varphi_j} \uparrow e^{-\varphi}. \end{align*} This monotonicity gives uniform control on compact levels and then [weak convergence](/page/Weak%20Convergence) in the relevant [Hilbert space](/page/Hilbert%20Space). It also explains why approximation alone is not enough: one must know that solutions obtained for the smooth weights converge to a genuine solution for the singular weight, with the same estimate surviving in the limit. The obstruction is that solving the smooth problems separately gives a sequence of solutions living in changing weighted spaces. Without a compactness principle tied to the monotone weights, the limiting form may fail to exist or may lose the estimate exactly at the singular set. The following stability statement supplies that missing passage from smooth approximants to a genuine singular-weight solution. [quotetheorem:3728] This is the technical bridge between the smooth theory of the earlier chapters and the singular metrics needed for extension. Two features of the hypothesis deserve emphasis. First, monotone decrease $\varphi_j \downarrow \varphi$ is essential, not a convenience: the weighted norms $\|u\|^2_{e^{-\varphi_j}}$ are then increasing in $j$, so a uniform bound at one level controls all later levels, and lower semicontinuity transfers the bound to the limit weight. If one only assumes $\varphi_j \to \varphi$ in $L^1_{\mathrm{loc}}$ without the monotonicity, the weights can oscillate from above and below, and a sequence of solutions $u_j$ to $\bar{\partial} u_j = f$ may have bounded $L^2(e^{-\varphi_j})$ norms while no subsequence converges in any $L^2(e^{-\varphi})$ ball — for instance, $\varphi_j = \varphi + (-1)^j \chi_K$ for a cutoff $\chi_K$ on a compact set $K$ breaks the lower semicontinuity argument even though the smooth weights converge in $L^1$. Second, the limit $u$ is not unique: any $u + h$ with $h$ holomorphic and in the relevant weighted $L^2$ space is a valid solution. Uniqueness can be restored by imposing minimality (orthogonality to holomorphic functions in the weighted [Hilbert space](/page/Hilbert%20Space)), and that minimal solution is the one captured by the weak-compactness argument. [citeproof:3728] The rest of the chapter takes this principle out of the pseudoconvex domain setting and into a global geometric one. The role played here by pseudoconvexity — supplying boundary terms that vanish or have the right sign — will be played by completeness of an ambient Kahler metric. On a complete Kähler manifold there is no boundary at all in the analytic sense: the compactly supported smooth forms are dense in the graph domain of $\bar{\partial}$, so any identity proved with compact support persists under closure. The decreasing approximation lemma combines with this density to make singular metrics on holomorphic line bundles, rather than smooth weights on domains, the natural object of the next section's estimates. ## The Donnelly-Fefferman Estimate How can we solve $\bar{\partial}$ on a noncompact complex manifold without imposing boundary conditions? The answer is to work on a complete Kahler manifold, where compactly supported cutoffs approximate global $L^2$ forms and the Bochner-Kodaira identity has no boundary contribution in the limit. Donnelly and Fefferman observed that a lower Ricci curvature bound can be offset by sufficiently positive twisting curvature, giving a quantitative estimate. [definition: Complete Kahler Manifold] A Kahler manifold $(X,\omega)$ is complete if the Riemannian metric associated to $\omega$ is complete as a [metric space](/page/Metric%20Space). [/definition] Completeness is the analytic substitute for compactness in the integration-by-parts argument. It allows the formal adjoint $\bar{\partial}^*$ to coincide with the Hilbert-space adjoint on the natural maximal domain. The estimate also needs a numerical way to say that the twisting line bundle contributes enough positive curvature to dominate any negative Ricci term. The following definition fixes that comparison against the Kahler form $\omega$, and it applies equally to smooth metrics and singular metrics whose curvature is a current. [definition: Curvature Lower Bound for a Line Bundle] Let $(X,\omega)$ be a Kahler manifold and let $(L,h)$ be a Hermitian holomorphic line bundle on $X$. For a real number $q$, the inequality \begin{align*} i\Theta_h(L) \ge q\omega \end{align*} means that $i\Theta_h(L)-q\omega$ is a semipositive $(1,1)$-form or current. [/definition] The symbol $q$ in this section denotes a curvature constant, not the antiholomorphic degree of a form. The solvability question is whether the positive curvature lower bound for $L$ can overcome the Ricci lower bound of the ambient complete Kahler metric. Donnelly-Fefferman's estimate answers this by making the usable gain the explicit gap between the twisting curvature and the negative Ricci bound. [quotetheorem:3729] The estimate is a global solvability theorem with an explicit constant. It says that negative Ricci curvature costs $K$, while the twisting line bundle contributes $q$; the usable positivity is the difference $q-K$. [citeproof:3729] The next examples show how the numerical curvature gap is read in practice. They also indicate why the hypotheses are not decorative: at $q=K$ the coercive lower bound disappears, and without completeness the cutoff argument can leave uncontrolled boundary terms. [illustration:scv-iii-poincare-disk-weight-gap] [example: Poincare Disk Estimate] Let $\Delta \subset \mathbb C$ carry the complete Poincare Kahler form $\omega_P$ normalized by $\operatorname{Ric}(\omega_P)=-\omega_P$. Then \begin{align*} \operatorname{Ric}(\omega_P) = -\omega_P = -1\cdot \omega_P, \end{align*} so the Ricci lower-bound constant in *Donnelly-Fefferman Estimate* is $K=1$. On the product line bundle $L=\Delta\times\mathbb C$, write $h=e^{-\varphi}$ with $\varphi=q\psi$. The standard product metric is flat, hence $i\Theta_{h_0}(L)=0$, and the curvature of $h$ is \begin{align*} i\Theta_h(L) &= i\Theta_{h_0}(L)+i\partial\bar{\partial}\varphi \\ &= 0+i\partial\bar{\partial}(q\psi) \\ &= q\,i\partial\bar{\partial}\psi \\ &= q\omega_P. \end{align*} Thus the line-bundle curvature lower bound is $i\Theta_h(L)\ge q\omega_P$. Since $q>1=K$, the curvature gap is \begin{align*} q-K=q-1>0. \end{align*} Let $f$ be a $\bar{\partial}$-closed $L^2$ $L$-valued $(0,1)$-form on $(\Delta,\omega_P)$. Applying *Donnelly-Fefferman Estimate* with $K=1$ gives a solution $u$ of $\bar{\partial}u=f$ such that \begin{align*} \int_\Delta |u|_h^2\,dV_{\omega_P} &\le \frac{1}{q-K} \int_\Delta |f|_{\omega_P,h}^2\,dV_{\omega_P} \\ &= \frac{1}{q-1} \int_\Delta |f|_{\omega_P,h}^2\,dV_{\omega_P}. \end{align*} In the standard frame of the product bundle, \begin{align*} |u|_h^2=|u|^2e^{-q\psi}, \qquad |f|_{\omega_P,h}^2=|f|_{\omega_P}^2e^{-q\psi}. \end{align*} Substituting these identities into the estimate gives \begin{align*} \int_\Delta |u|^2e^{-q\psi}\,dV_{\omega_P} \le \frac{1}{q-1} \int_\Delta |f|_{\omega_P}^2e^{-q\psi}\,dV_{\omega_P}. \end{align*} The example shows that the Poincare disk's negative Ricci curvature costs exactly $1$, while the weight curvature contributes $q$, leaving the usable $L^2$ solvability gap $q-1$. [/example] The Poincare disk example is the model case where the negative Ricci curvature of a complete metric is overpowered by a more positive weight; the gap $q-1$ is the only thing that survives in the estimate. The next example, by contrast, illustrates the high-positivity regime on a compact manifold, where the bundle itself can be made arbitrarily positive by raising it to a power, and the Ricci defect $K$ of the background metric becomes negligible. [example: High Powers of a Positive Line Bundle] Let $(M,\omega)$ be a compact Kahler manifold, and let $(A,h_A)$ satisfy \begin{align*} i\Theta_{h_A}(A)\ge c\omega \end{align*} for some $c>0$. Equip $A^m$ with the tensor-power metric $h_A^m$. In a local holomorphic frame with $h_A=e^{-\phi_A}$, the induced metric on $A^m$ is $h_A^m=e^{-m\phi_A}$, so \begin{align*} i\Theta_{h_A^m}(A^m) &= i\partial\bar{\partial}(m\phi_A) \\ &= m\,i\partial\bar{\partial}\phi_A \\ &= m\,i\Theta_{h_A}(A) \\ &\ge mc\,\omega. \end{align*} Thus the curvature lower-bound constant for $A^m$ is $q=mc$. Since $M$ is compact, the Kahler metric associated to $\omega$ is complete. Suppose \begin{align*} \operatorname{Ric}(\omega)\ge -K\omega \end{align*} and choose $m$ so that \begin{align*} mc>K. \end{align*} Then the curvature gap in the *Donnelly-Fefferman Estimate* is \begin{align*} q-K=mc-K>0. \end{align*} For every $\bar{\partial}$-closed $A^m$-valued $(0,1)$-form $f$ with finite $L^2$ norm, the estimate applied to $L=A^m$ gives a section $u$ of $A^m$ satisfying $\bar{\partial}u=f$ and \begin{align*} \int_M |u|_{h_A^m}^2\,dV_\omega &\le \frac{1}{q-K} \int_M |f|_{\omega,h_A^m}^2\,dV_\omega \\ &= \frac{1}{mc-K} \int_M |f|_{\omega,h_A^m}^2\,dV_\omega. \end{align*} As $m$ increases, the denominator $mc-K$ increases linearly, so the solution constant $(mc-K)^{-1}$ decreases; high tensor powers make the positive curvature of $A^m$ dominate the fixed Ricci defect of the background metric. [/example] The examples show the two typical roles of the theorem: complete negatively curved metrics are allowed, and high positivity of a line bundle dominates ambient curvature. ## Complete Kahler Hodge Theory What does the Donnelly-Fefferman estimate say about cohomology rather than about a single equation? On compact manifolds, Hodge theory identifies Dolbeault cohomology with harmonic forms by elliptic theory. On complete noncompact manifolds, the same formal identity exists, but closed range and finite-dimensionality may fail unless an estimate supplies a spectral gap. [definition: $L^2$ Harmonic Form] Let $(X,\omega)$ be a complete Kahler manifold and let $(L,h)$ be a Hermitian holomorphic line bundle. An $L^2$ $L$-valued $(p,r)$-form $\alpha$ is $L^2$ harmonic for the $\bar{\partial}$-Laplacian if \begin{align*} \bar{\partial}\alpha=0, \qquad \bar{\partial}^*\alpha=0 \end{align*} in the Hilbert-space sense. [/definition] The definition uses the maximal closed extensions of $\bar{\partial}$ and $\bar{\partial}^*$. Completeness ensures that these extensions agree with the closure obtained from compactly supported smooth forms. The cohomological problem is whether every reduced $L^2$ Dolbeault class has a canonical representative satisfying these two first-order equations. On a noncompact manifold this is not automatic, because exact forms may have nonclosed range and the Laplacian can have continuous spectrum near zero. The Hodge statement below identifies precisely what survives once one uses reduced cohomology. [quotetheorem:3730] This is the $L^2$ analogue of the compact Hodge theorem, but the hypothesis is analytic rather than topological. [citeproof:3730] This is the point where complete $L^2$ theory differs most visibly from compact Hodge theory. The word reduced means that one quotients closed forms by the closure of exact forms; if the range of $\bar{\partial}$ is not closed, unreduced cohomology can remember non-Hausdorff analytic pathologies that harmonic forms do not see. [remark: Comparison with Compact Hodge Theory] On a compact Kahler manifold with smooth metric, elliptic regularity and compactness of the resolvent imply finite-dimensional harmonic spaces. On a complete noncompact Kahler manifold, the Laplacian may have continuous spectrum, so finite-dimensionality and closed range are additional conclusions only when estimates provide them. Donnelly-Fefferman gives a concrete sufficient condition through curvature positivity. [/remark] This perspective explains why $L^2$ methods are powerful in several complex variables: analytic curvature inequalities replace compactness assumptions. The trade is favourable because curvature inequalities can be checked at a single point and propagated by sign, while compactness is a global topological condition that almost never holds in the situations of interest — Stein manifolds, complements of subvarieties, total spaces of families. The next section pushes this trade further by allowing the curvature lower bound to depend on auxiliary smooth functions, which gives the flexibility needed for extension problems with prescribed jets along a subvariety. ## Twisted Estimates and Direct Images How can the estimate be modified when a single curvature lower bound is too rigid for extension problems? The McNeal-Varolin refinement introduces auxiliary positive functions into the Bochner identity. These functions let the estimate absorb terms produced by singular cutoffs and are especially suited to sharp versions of Ohsawa-Takegoshi extension. [definition: Twisting Data] Let $(X,\omega)$ be a Kahler manifold and let $(L,h=e^{-\varphi})$ be a Hermitian holomorphic line bundle. Twisting data consist of smooth positive functions $\eta,\lambda:X\to (0,\infty)$ used in the weighted Bochner expression \begin{align*} \eta\, i\Theta_h(L)-i\partial\bar{\partial}\eta-\lambda^{-1}i\partial\eta\wedge\bar{\partial}\eta. \end{align*} [/definition] The extra negative terms are not errors; they are the exact cost of commuting the weight $\eta$ through the adjoint in the Bochner identity. Concretely, when one tries to bring $\sqrt{\eta}$ inside $\bar{\partial}^*$, the commutator throws off a first-order term involving $\bar{\partial}\eta$; the only way to absorb it back into a quadratic form is to split it via Cauchy-Schwarz with a parameter, and that parameter is exactly $\lambda$. The term $-\lambda^{-1}\, i\partial\eta\wedge\bar{\partial}\eta$ is the residue left over by this split. The combined expression in the definition is therefore the genuine pointwise curvature seen by twisted $L^2$ sections, and it is what must be positive for an estimate to follow. The estimate will not use the twisted curvature expression merely as a scalar lower bound. Positivity has to be tested after the form acts on $(p,r)$-forms through the Lefschetz commutator, so the obstruction is an operator-valued one rather than a scalar inequality. The next definition names that endomorphism because its inverse is exactly what appears in the $L^2$ norm on the right-hand side of the twisted theorem. [definition: Twisted Curvature Operator] Let $(X,\omega)$ be a Kahler manifold, let $(L,h=e^{-\varphi})$ be a Hermitian holomorphic line bundle, and let $\eta,\lambda:X\to(0,\infty)$ be twisting data. For each $x\in X$ and each form bidegree $(p,r)$, the twisted curvature operator \begin{align*} B_{\eta,\lambda}(x) : \Lambda^{p,r} T^*_x X \otimes L_x \to \Lambda^{p,r} T^*_x X \otimes L_x \end{align*} is the Hermitian endomorphism on $L$-valued $(p,r)$-forms at $x$ defined by the commutator \begin{align*} B_{\eta,\lambda}=\big[\eta\, i\Theta_h(L)-i\partial\bar{\partial}\eta-\lambda^{-1}i\partial\eta\wedge\bar{\partial}\eta,\Lambda_\omega\big], \end{align*} where $\Lambda_\omega$ is the adjoint of wedging with $\omega$ with respect to the pointwise Hermitian inner product induced by $\omega$ and $h$. [/definition] The analytic difficulty is to turn positivity of this pointwise endomorphism into a global solution estimate while the auxiliary weight $\eta$ is varying. The terms involving $\partial\eta$ have already been charged to the curvature operator, so the theorem must show that no uncontrolled commutator remains after integration by parts on the complete manifold. [quotetheorem:3731] This theorem is a flexible form of Donnelly-Fefferman. The untwisted estimate is recovered by taking $\eta$ constant and letting the curvature operator be bounded below by a scalar. [citeproof:3731] Each piece of the twisting data plays a non-removable role. If $\lambda$ is sent to $+\infty$, the term $-\lambda^{-1} i\partial\eta\wedge\bar{\partial}\eta$ formally disappears, but so does the only mechanism for absorbing the first-order $\bar{\partial}\eta$ commutator; the resulting "estimate" is not closed under [integration by parts](/theorems/210) because the surviving cross term has the wrong sign on regions where $\bar{\partial}\eta$ is large. Choosing $\lambda$ to be a finite positive function is therefore unavoidable whenever $\eta$ is nonconstant. If $\eta$ is taken constant, $\bar{\partial}\eta$ vanishes and the estimate collapses back to the untwisted Donnelly-Fefferman bound — useful, but blind to the singular geometry that twisting is designed to detect. If completeness of $(X,\omega)$ fails, the cutoff approximation by compactly supported forms in the last step can leave a boundary contribution involving $\eta$ that no choice of $\lambda$ controls; an explicit pathology is a punctured polydisc with the flat Euclidean metric, where the natural twisted estimate fails for the [holomorphic function](/page/Holomorphic%20Function) $1/z_1$ even though every curvature expression is formally nonnegative. The takeaway is that $\eta$, $\lambda$, and completeness are coupled hypotheses: the twisted estimate degenerates if any one is dropped. With these necessity remarks in place, it is worth comparing the McNeal-Varolin estimate to a parallel positivity statement that uses the same Bochner-Kodaira machinery in a different direction. Berndtsson's theorem on direct images shows that the curvature positivity preserved by Hormander-type estimates is also preserved by passage to fibrewise spaces of holomorphic sections. The full proof is substantially harder than the estimates developed in this course: it requires the second-variation formula for the fibrewise $L^2$ metric, [harmonic representatives](/theorems/2747) in a varying family, Kodaira-Spencer terms, and the Bochner-Kodaira-Nakano identity on the fibres. We record the full statement for orientation, but leave its proof outside the scope of these notes. [quotetheorem:3668] The conceptual link is direct: the McNeal-Varolin estimate controls a singular extension on a single manifold, while Berndtsson's theorem controls how those extensions vary in families. Both rest on positivity of a curvature expression. For later use, the needed takeaway is not another theorem card but the mechanism by which the curvature formula becomes nonnegative. In Berndtsson's calculation, the curvature of the fibrewise $L^2$ metric decomposes into terms whose signs are controlled by the plurisubharmonic weight and by the same Hilbert-space projection structure that appears in Hörmander theory. Once that formula is established, positivity follows by checking that each summand has the required sign. The point of recalling this mechanism is that it closes a triangle of related statements: the untwisted Donnelly-Fefferman estimate, the McNeal-Varolin twisted refinement on a single manifold, and Berndtsson's family version on the parameter space $Y$. All three are quantitative incarnations of the slogan that curvature positivity propagates through $L^2$ constructions. The remark below isolates the specific way in which twisting is used inside Ohsawa-Takegoshi extension, which is the application that the rest of the chapter builds toward. [remark: Why Twisting Matters for Extension] In Ohsawa-Takegoshi extension, the data are often cut off near a subvariety, and differentiating the cutoff produces singular error terms. The twisted estimate lets $\eta$ and $\lambda$ be chosen so that these errors are measured by the same curvature expression that controls the solution. This is the analytic origin of the sharp constants in later refinements. [/remark] ## Optimal Constants and the Suita Conjecture What remains after solvability is known? The modern theory asks for the best possible constants in $L^2$ extension and for geometric interpretations of equality. This shift from existence to optimality led to one of the most striking applications of the method: the solution of the Suita conjecture. [definition: Bergman Kernel on a Planar Domain] Let $\Omega \subset \mathbb C$ be a domain for which the [Hilbert space](/page/Hilbert%20Space) $A^2(\Omega)$ of square-integrable holomorphic functions is nonzero. The Bergman kernel on the diagonal is \begin{align*} K_\Omega(z)=\sup\{ |f(z)|^2 : f\in A^2(\Omega),\ \|f\|_{L^2(\Omega)}\le 1\}. \end{align*} [/definition] The Bergman kernel measures how efficiently an $L^2$ [holomorphic function](/page/Holomorphic%20Function) can concentrate at a point. To compare this Hilbert-space extremal quantity with planar geometry, one needs a second local invariant that measures the size of the same point from potential theory. The Green function has a universal logarithmic singularity at its pole; the finite residual constant left after subtracting that singularity is the capacity term entering Suita's inequality. [definition: Suita Capacity] Let $\Omega \subset \mathbb C$ be a hyperbolic planar domain and let $G_\Omega(\cdot,z)$ be the Green function with pole at $z\in\Omega$. The Suita capacity at $z$ is \begin{align*} c_\Omega(z)=\exp\left(\lim_{\zeta\to z}\big(G_\Omega(\zeta,z)-\log|\zeta-z|\big)\right). \end{align*} [/definition] Capacity is potential-theoretic, while $K_\Omega$ is Hilbert-space analytic. The problem is that they are defined by different extremal procedures: one by Green-function asymptotics and the other by minimal $L^2$ norm of holomorphic functions with prescribed value. Suita's inequality asserts that optimal extension is strong enough to bridge these two notions with the sharp numerical constant. [quotetheorem:3733] The result was proved using optimal $L^2$ extension methods, with decisive contributions by Blocki and by Guan-Zhou. The proof is outside the scope of these lectures, but the statement shows why constants in Donnelly-Fefferman and Ohsawa-Takegoshi estimates are not cosmetic. [explanation: From Donnelly-Fefferman to Sharp Extension] The Donnelly-Fefferman theorem gives a model estimate: curvature positivity minus geometric negativity produces an inverse spectral-gap constant. The McNeal-Varolin theorem refines the model by allowing twisting functions that track the singular geometry of extension from a subvariety. In the optimal extension problem, the goal is to choose these weights so that every loss in the Bochner identity is accounted for and no spare constant remains. The Suita inequality can be viewed as a one-dimensional shadow of this principle. Extending the value of a [holomorphic function](/page/Holomorphic%20Function) from a point with minimal $L^2$ norm is exactly the extremal problem defining the Bergman kernel. The Green function supplies the singular weight whose residual constant is the logarithmic capacity. Sharp extension converts the singular weighted estimate into the inequality $\pi K_\Omega(z)\ge c_\Omega(z)^2$. [/explanation] Seen this way, the final theorem is not an isolated application but the endpoint of the course's main analytic thread: curvature inequalities become norm inequalities, and norm inequalities become geometric extremal statements. [remark: Open Directions] Current questions around $L^2$ extension ask for optimal constants in higher codimension, equality characterisations for singular metrics, and versions over families where direct image positivity and extension estimates interact. The guiding problem is no longer whether $\bar{\partial}$ can be solved, but how precisely curvature, singularities, and volume determine the best possible solution norm. [/remark] The chapter closes the course by returning to the original message of Hormander methods. Positivity of curvature is an analytic resource: once it is expressed in the correct Hilbert-space estimate, it gives existence, cohomological vanishing, extension theorems, and sharp geometric inequalities. The course developed the pieces—the ∂̄-operator, pseudoconvexity, plurisubharmonic weights, Hörmander existence, extension, approximation, Bergman kernels, multiplier ideals, and vanishing—as separate chapters. This final chapter assembles them into a unified framework, showing how positivity of curvature (encoded in weights and complex Hessians) flows through L² estimates to yield existence theorems, extension results, approximation statements, cohomological vanishing, and sharp geometric inequalities. # 11. Synthesis and Further Directions The final chapter assembles the course into a single analytic picture. Earlier chapters developed the $\bar{\partial}$-operator, pseudoconvexity, weighted $L^2$ estimates, extension, division, multiplier ideals, and vanishing theorems. Here the point is to understand how those results depend on one another, what they prove about holomorphic functions, and where the method stops being the whole story. The unifying theme is that curvature of a weight replaces explicit integral formulas. Once a plurisubharmonic function supplies positivity, [Hilbert space](/page/Hilbert%20Space) methods turn the formal identity $\bar{\partial}^2=0$ into existence theorems with norms. The same mechanism explains both the power and the limitations of the subject. ## The Chain From Pseudoconvexity To Holomorphic Separation What is the shortest route from a geometric condition on a domain to the existence of many holomorphic functions? The course's central chain is \begin{align*} \text{pseudoconvexity} &\Longrightarrow \text{plurisubharmonic exhaustion} \\ &\Longrightarrow \text{Hörmander }L^2\text{ estimate} \\ &\Longrightarrow \bar{\partial}\text{-solvability} \\ &\Longrightarrow \text{domain of holomorphy}. \end{align*} Each arrow hides a substantial theorem, but the logic is now compact: geometry gives weights, weights give estimates, estimates give solutions, and solutions manufacture holomorphic functions. [definition: Plurisubharmonic Exhaustion] Let $\Omega \subset \mathbb C^n$ be a domain. A plurisubharmonic exhaustion of $\Omega$ is a plurisubharmonic function $\rho: \Omega \to \mathbb R$ such that $\{z \in \Omega : \rho(z) < c\}$ is relatively compact in $\Omega$ for every $c \in \mathbb R$. [/definition] A plurisubharmonic exhaustion is the analytic form of pseudoconvexity used by the $L^2$ method. It prevents the weight from losing control near the boundary and lets one solve on increasing relatively compact subdomains. The converse direction in the Levi problem requires a language for saying that holomorphic functions genuinely see the boundary of the domain. The obstruction is analytic continuation across a boundary piece: if every holomorphic function extends past that piece, then estimates have not produced a domain maximal for holomorphic function theory. The next definition names this maximality property. [definition: Domain Of Holomorphy] A domain $\Omega \subset \mathbb C^n$ is a domain of holomorphy if there do not exist domains $\Omega_0,\Omega_1 \subset \mathbb C^n$ with $\varnothing \ne \Omega_0 \subset \Omega \cap \Omega_1$ and $\Omega_1 \not\subset \Omega$ such that every $f \in \mathcal O(\Omega)$ admits $F \in \mathcal O(\Omega_1)$ satisfying $F|_{\Omega_0}=f|_{\Omega_0}$. [/definition] This definition says that the holomorphic functions on $\Omega$ detect the boundary. A domain of holomorphy cannot be enlarged across any boundary piece without losing at least one [holomorphic function](/page/Holomorphic%20Function). The remaining issue is how to manufacture such boundary-detecting functions from the geometric hypothesis of pseudoconvexity. The $L^2$ method does this by solving a $\bar\partial$ problem with weights that blow up near the chosen boundary point, correcting a local holomorphic model into a global holomorphic function that cannot extend across that point. [quotetheorem:3734] [citeproof:3734] The theorem is best read as a construction scheme. If the desired holomorphic object can be expressed as a smooth approximate object plus a $\bar{\partial}$-error, Hörmander's estimate supplies the correction term. The pseudoconvexity hypothesis is not decorative: on non-pseudoconvex domains the Hartogs phenomenon can force holomorphic functions to extend across apparent holes, so boundary points may fail to be detected by holomorphic functions. The result also has deliberate limits. It proves existence with a quantitative $L^2$ bound, but it does not by itself give [boundary regularity](/theorems/99), optimal constants, or compactness of the solution operator. | Theorem or tool | Weight condition | Geometric consequence | |---|---|---| | Hörmander estimate | $i\partial\bar{\partial}\varphi \ge c\omega$ | Solvability of $\bar{\partial}u=f$ with an explicit $L^2$ bound | | Ohsawa--Takegoshi extension | Logarithmic singularity transverse to a submanifold | Extension of square-integrable holomorphic functions from lower-dimensional sets | | Skoda division | Singular weight involving $\log |g|^2$ | Membership in ideals generated by holomorphic functions | | Nadel and Kodaira--Nakano vanishing | Positive curvature of a singular Hermitian metric | Vanishing of cohomology groups | | Berndtsson positivity | Plurisubharmonic variation of fibre weights | Positivity of direct image bundles and convexity of Bergman kernels | The table shows why the same estimate reappears in many forms. The first row is the basic engine, while the later rows change the geometry encoded by the weight: transverse logarithmic poles produce extension, logarithmic poles along common zero sets produce division, and curvature of singular metrics produces cohomological vanishing. Thus the course is not using several unrelated tricks; it is repeatedly choosing a weight whose curvature inequality remembers the geometric problem to be solved. [illustration:scv-iii-egg-domain-flat-boundary] [example: Egg Domain] Let \begin{align*} \Omega=\{(z_1,z_2)\in \mathbb C^2: |z_1|^2+|z_2|^4<1\}, \end{align*} and put $s(z)=|z_1|^2+|z_2|^4$. For the standard Kähler form $\omega=i\partial\bar{\partial}(|z_1|^2+|z_2|^2)$, we show that $\Omega$ is pseudoconvex and that the weights $\varphi_A=A(|z_1|^2+|z_2|^2-\log(1-s))$ have curvature at least $A\omega$. First compute the Levi form of $s$. Since $|z_1|^2=z_1\bar z_1$, \begin{align*} \partial\bar{\partial}|z_1|^2 = \partial(z_1\,d\bar z_1) = dz_1\wedge d\bar z_1. \end{align*} For $r=|z_2|^2=z_2\bar z_2$, we have \begin{align*} \partial r=\bar z_2\,dz_2,\qquad \bar{\partial}r=z_2\,d\bar z_2,\qquad \partial\bar{\partial}r=dz_2\wedge d\bar z_2, \end{align*} and therefore \begin{align*} \partial\bar{\partial}(r^2) &= \partial(2r\,\bar{\partial}r)\\ &= 2\,\partial r\wedge\bar{\partial}r+2r\,\partial\bar{\partial}r\\ &= 2(\bar z_2\,dz_2)\wedge(z_2\,d\bar z_2) +2|z_2|^2\,dz_2\wedge d\bar z_2\\ &= 4|z_2|^2\,dz_2\wedge d\bar z_2. \end{align*} Thus \begin{align*} i\partial\bar{\partial}s = i\,dz_1\wedge d\bar z_1 + 4|z_2|^2\,i\,dz_2\wedge d\bar z_2 \ge 0, \end{align*} so $s$ is plurisubharmonic. Now define \begin{align*} \rho(z)=-\log(1-s(z)). \end{align*} On $\Omega$ we have $0\le s<1$. Since $h(t)=-\log(1-t)$ satisfies \begin{align*} h'(t)=\frac{1}{1-t},\qquad h''(t)=\frac{1}{(1-t)^2}, \end{align*} the chain rule gives \begin{align*} i\partial\bar{\partial}\rho &= \frac{i\partial\bar{\partial}s}{1-s} + \frac{i\,\partial s\wedge\bar{\partial}s}{(1-s)^2} \ge 0. \end{align*} Also $\rho(z)\to+\infty$ as $s(z)\to 1$, and for $c>0$, \begin{align*} \{\rho<c\} = \{-\log(1-s)<c\} = \{s<1-e^{-c}\}, \end{align*} whose closure lies in $\Omega$. Hence $\rho$ is a plurisubharmonic exhaustion, so $\Omega$ is pseudoconvex in the plurisubharmonic-exhaustion sense. For $A>0$, set \begin{align*} \varphi_A=A(|z_1|^2+|z_2|^2+\rho). \end{align*} Then \begin{align*} i\partial\bar{\partial}\varphi_A &= A\,i\partial\bar{\partial}(|z_1|^2+|z_2|^2) + A\,i\partial\bar{\partial}\rho\\ &= A\omega+A\,i\partial\bar{\partial}\rho\\ &\ge A\omega. \end{align*} Applying *Hörmander's $L^2$ estimate* with curvature constant $c=A$, every smooth $\bar{\partial}$-closed $(0,1)$-form $f$ with finite weighted norm admits a solution $\bar{\partial}u=f$ satisfying \begin{align*} \int_\Omega |u|^2e^{-\varphi_A}\,d\mathcal L^4 \le \frac{1}{A}\int_\Omega |f|_\omega^2e^{-\varphi_A}\,d\mathcal L^4. \end{align*} The factor $|z_2|^2$ in $i\partial\bar{\partial}s$ shows the boundary degeneracy in the $z_2$ direction, while the added Euclidean term in $\varphi_A$ supplies the uniform lower bound needed for the $L^2$ estimate. [/example] This example also illustrates why weakly pseudoconvex boundaries are natural in the course. The boundary is not strongly pseudoconvex along the complex-tangential degeneracy coming from $|z_2|^4$, but the plurisubharmonic weight still gives solvability. ## Boundary Regularity And Sobolev Information Which analytic questions remain after existence in $L^2$ has been proved? Hörmander's theorem gives a solution and a norm estimate, but it does not by itself describe boundary smoothness, exact Sobolev gain, compactness of the solution operator, or regularity of the Bergman projection. [definition: Subelliptic Estimate] Let $\Omega \subset \mathbb C^n$ be a smoothly bounded domain. A subelliptic estimate of order $\varepsilon>0$ in degree $q$ for the $\bar{\partial}$-Neumann problem is an inequality \begin{align*} \|u\|_{H^\varepsilon(\Omega)}^2 \le C\big(\|\bar{\partial}u\|_{L^2(\Omega)}^2+\|\bar{\partial}^*u\|_{L^2(\Omega)}^2+\|u\|_{L^2(\Omega)}^2\big) \end{align*} for all smooth $(0,q)$-forms $u$ in the domains of $\bar{\partial}$ and $\bar{\partial}^*$. [/definition] A subelliptic estimate measures regularity, not only solvability. It says that the complex Laplacian controls a positive fractional derivative of $u$ up to lower-order terms. The natural boundary question is when the geometry of $\partial\Omega$ forces such a gain uniformly for the $\bar\partial$-Neumann problem. Strong pseudoconvexity is the model case: the Levi form supplies positivity in the boundary directions where elliptic interior estimates alone do not see enough derivatives. [quotetheorem:3735] This result is quoted here as advanced reading rather than proved in the course. Its proof uses boundary microlocal analysis, the Levi form, and the special structure of the $\bar{\partial}$-Neumann boundary conditions. The word strongly is essential: strict positivity of the Levi form gives a uniform gain of one half of a derivative, whereas weakly pseudoconvex finite-type points can have smaller subelliptic gain depending on the type. Degree $q=0$ is a related but different regularity question, usually phrased through the Bergman projection and [holomorphic function](/page/Holomorphic%20Function) estimates rather than the same Neumann boundary condition in positive degree. [example: Boundary Regularity Gap] Let $\Omega\subset \mathbb C^n$ be smoothly bounded and pseudoconvex, and let \begin{align*} f=\sum_{j=1}^n f_j\,d\bar z_j \end{align*} be smooth on $\overline{\Omega}$ with $\bar{\partial}f=0$. Choose $A>0$ and set $\varphi_A=A|z|^2$. Since \begin{align*} i\partial\bar{\partial}\varphi_A = A\,i\partial\bar{\partial}|z|^2 = A\omega, \end{align*} *Hörmander's weighted $L^2$ estimate* gives a solution $u\in L^2(\Omega,e^{-\varphi_A})$ of $\bar{\partial}u=f$ satisfying \begin{align*} \int_\Omega |u|^2e^{-\varphi_A}\,d\mathcal L^{2n} \le \frac{1}{A} \int_\Omega |f|_\omega^2e^{-\varphi_A}\,d\mathcal L^{2n}. \end{align*} If $R=\sup_{\overline{\Omega}}|z|$, then $0\le \varphi_A\le AR^2$, so \begin{align*} \|u\|_{L^2(\Omega)}^2 &= \int_\Omega |u|^2\,d\mathcal L^{2n} \\ &\le e^{AR^2}\int_\Omega |u|^2e^{-\varphi_A}\,d\mathcal L^{2n} \\ &\le \frac{e^{AR^2}}{A} \int_\Omega |f|_\omega^2e^{-\varphi_A}\,d\mathcal L^{2n} \\ &\le \frac{e^{AR^2}}{A}\|f\|_{L^2(\Omega)}^2. \end{align*} Thus the theorem supplies an $L^2$ solution with an $L^2$ bound. What is missing is any estimate for derivatives. For example, if $\chi\in C_c^\infty(\Omega)$ is real-valued and nonzero and $v_N=\chi e^{iNx_1}$ in a local real coordinate $x_1$, then \begin{align*} \|v_N\|_{L^2(\Omega)}^2 &= \int_\Omega |\chi|^2\,d\mathcal L^{2n}, \end{align*} while \begin{align*} \|\partial_{x_1}v_N\|_{L^2(\Omega)}^2 &= \int_\Omega |(\partial_{x_1}\chi)e^{iNx_1}+iN\chi e^{iNx_1}|^2\,d\mathcal L^{2n} \\ &= \int_\Omega |\partial_{x_1}\chi+iN\chi|^2\,d\mathcal L^{2n} \\ &= \int_\Omega \big((\partial_{x_1}\chi)^2+N^2\chi^2\big)\,d\mathcal L^{2n}, \end{align*} because the cross term has real part $0$. Hence a uniform $L^2$ bound does not control even one first derivative, and a fortiori does not imply smoothness up to $\partial\Omega$. [Boundary regularity](/theorems/99) requires additional estimates involving the domains of $\bar{\partial}$ and $\bar{\partial}^*$; in strongly pseudoconvex cases, *Kohn's subelliptic estimate* supplies precisely that extra Sobolev control. [/example] [illustration:scv-iii-worm-domain-weak-region] The Diederich--Fornaess worm domain is the standard warning example in this direction. It is smoothly bounded and pseudoconvex but not strongly pseudoconvex, so it lies outside Kohn's theorem rather than contradicting it. Its failures of global regularity for operators associated with the Bergman projection and the $\bar{\partial}$-Neumann problem show that smooth pseudoconvexity alone is not a substitute for the stronger boundary hypotheses used in subelliptic theory. ## Serre Duality From The Complex Laplacian How does the [Hilbert space](/page/Hilbert%20Space) picture recover the cohomological dualities from sheaf theory? The bridge is the complex Laplacian: cohomology classes are represented by harmonic forms, and the Hodge star converts [harmonic representatives](/theorems/2747) in complementary bidegrees into dual functionals. [definition: Dolbeault Cohomology With Coefficients] Let $X$ be a complex manifold and let $E\to X$ be a holomorphic vector bundle. The Dolbeault cohomology group in bidegree $(p,q)$ with coefficients in $E$ is \begin{align*} H^{p,q}_{\bar{\partial}}(X,E) = \frac{\ker(\bar{\partial}:\mathcal A^{p,q}(X,E)\to \mathcal A^{p,q+1}(X,E))} {\operatorname{im}(\bar{\partial}:\mathcal A^{p,q-1}(X,E)\to \mathcal A^{p,q}(X,E))}. \end{align*} [/definition] The $L^2$ theory replaces abstract quotient spaces by orthogonal decompositions whenever the relevant ranges are closed. This is why the Neumann operator enters next: it provides an analytic inverse to the complex Laplacian on the orthogonal complement of harmonic forms. Once such an inverse exists, a cohomology class can be split into its harmonic representative and exact or coexact pieces, turning the formal equation $\bar{\partial}u=f$ into an operator formula. The homotopy formula is therefore the precise bridge between [Hilbert space](/page/Hilbert%20Space) estimates and Dolbeault cohomology. [quotetheorem:3736] [citeproof:3736] The formula is the analytic version of a homotopy operator for the Dolbeault complex. It explains why estimates for $N_q$ immediately become estimates for solutions of $\bar{\partial}$. The closed-range hypothesis is a serious analytic input, not a formal property of the complex: it can fail on rough or non-smooth pseudoconvex domains, while smooth finite-type hypotheses give standard sufficient conditions in many settings. Smoothness of the boundary is also part of the operator theory, because the domains of $\bar{\partial}^*$ and the Neumann boundary conditions are defined through [integration by parts](/theorems/2098) at the boundary. After the Neumann operator has turned estimates into cohomological decompositions, the next structural question is how those cohomology groups pair with the complementary degrees. On compact complex manifolds the absence of boundary terms and finite-dimensional Hodge theory make this pairing nondegenerate, giving the analytic form of Serre duality. [quotetheorem:3737] [citeproof:3737] This is the analytic shadow of the coherent-sheaf statement from Course II. [Dolbeault's theorem](/theorems/3389) identifies $H^{p,q}_{\bar{\partial}}(X,E)$ with sheaf cohomology $H^q(X,\Omega_X^p\otimes E)$, so the harmonic pairing recovers the usual Serre duality pairing. Compactness is essential in this formulation: it gives finite-dimensional harmonic spaces and removes boundary terms, while non-compact manifolds require extra hypotheses such as Steinness, compact support, or relative boundary conditions. The Hermitian metric is also not cosmetic; smooth metric data are used to define the adjoint, the Laplacian, and the Hodge star, and a merely continuous metric is not enough for the same elliptic Hodge theory. ## Division, Ideals, And The Analytic Nullstellensatz How can an estimate for $\bar{\partial}$ prove that a [holomorphic function](/page/Holomorphic%20Function) belongs to an ideal? The strategy is to write the desired division formula away from the common zero set, measure the singularity near that zero set with a logarithmic weight, and solve the resulting $\bar{\partial}$-error. [definition: Multiplier Ideal Sheaf] Let $X$ be a complex manifold and let $\varphi$ be a plurisubharmonic function on $X$. The multiplier ideal sheaf $\mathcal I(\varphi)$ assigns to each [open set](/page/Open%20Set) $U\subset X$ the ideal \begin{align*} \mathcal I(\varphi)(U)=\{f\in \mathcal O_X(U): |f|^2e^{-\varphi}\text{ is locally integrable on }U\}. \end{align*} [/definition] Multiplier ideals convert analytic growth conditions into algebraic vanishing conditions. The singularity of $\varphi$ records how much vanishing is required for integrability. The division problem now asks for the converse direction: when does an integrability condition force actual membership in an ideal generated by holomorphic functions? The obstruction is that near the common zero set of the generators, the quotients $f/g_j$ may have poles. Skoda's theorem supplies the precise weighted $L^2$ hypothesis strong enough to solve away those poles and produce holomorphic coefficients. [quotetheorem:3738] [citeproof:3738] Skoda division is the analytic source of several algebraic-looking consequences. It gives effective ideal membership from integrability rather than from explicit generators of syzygies. Once ideal membership can be forced by an $L^2$ condition, the natural next question is whether analytic vanishing on a zero set also forces algebraic membership after taking a power. That is the local Nullstellensatz problem. [quotetheorem:3739] [citeproof:3739] This theorem is the meeting point with algebraic geometry. The analytic local ring is not a [polynomial ring](/page/Polynomial%20Ring), but the vanishing-set dictionary still holds after passing to radicals. [example: Multiplier Ideal Threshold At The Origin] Let $\varphi_c=c\log |z|$ near $0$, and let $f\in\mathcal O_{\mathbb C^n,0}$ have vanishing order $k$. Thus the Taylor expansion has the form \begin{align*} f(z)=P_k(z)+P_{k+1}(z)+P_{k+2}(z)+\cdots, \end{align*} where $P_j$ is homogeneous of degree $j$ and $P_k\ne 0$. We compute when \begin{align*} |f|^2e^{-\varphi_c}=|f(z)|^2|z|^{-c} \end{align*} is locally integrable at $0$. Since $f$ vanishes to order $k$, there are constants $C>0$ and $\varepsilon>0$ such that \begin{align*} |f(z)|\le C|z|^k \end{align*} for $|z|<\varepsilon$. Hence \begin{align*} |f(z)|^2|z|^{-c} \le C^2|z|^{2k-c}. \end{align*} Using polar coordinates in $\mathbb R^{2n}$, with $z=r\zeta$, $0<r<\varepsilon$, and $\zeta\in S^{2n-1}$, \begin{align*} \int_{|z|<\varepsilon} |z|^{2k-c}\,d\mathcal L^{2n}(z) &= \int_0^\varepsilon\int_{S^{2n-1}} r^{2k-c}r^{2n-1}\,d\sigma(\zeta)\,dr \\ &= \sigma(S^{2n-1})\int_0^\varepsilon r^{2k-c+2n-1}\,dr. \end{align*} For $\alpha=2k-c+2n-1$, \begin{align*} \int_0^\varepsilon r^\alpha\,dr = \begin{cases} \dfrac{\varepsilon^{\alpha+1}}{\alpha+1},& \alpha>-1,\\ +\infty,& \alpha\le -1, \end{cases} \end{align*} so the upper bound proves local integrability when \begin{align*} 2k-c+2n-1>-1, \end{align*} equivalently \begin{align*} k>\frac c2-n. \end{align*} Conversely, choose $\zeta_0\in S^{2n-1}$ with $P_k(\zeta_0)\ne 0$. By continuity there are an [open set](/page/Open%20Set) $U\subset S^{2n-1}$ containing $\zeta_0$ and a number $a>0$ such that $|P_k(\zeta)|\ge a$ for $\zeta\in U$. Since \begin{align*} f(r\zeta)=r^kP_k(\zeta)+r^{k+1}P_{k+1}(\zeta)+\cdots, \end{align*} shrinking $\varepsilon$ gives \begin{align*} |f(r\zeta)|\ge \frac a2 r^k \end{align*} for $0<r<\varepsilon$ and $\zeta\in U$. Therefore \begin{align*} \int_{|z|<\varepsilon}|f(z)|^2|z|^{-c}\,d\mathcal L^{2n}(z) &\ge \int_0^\varepsilon\int_U \frac{a^2}{4}r^{2k-c}r^{2n-1}\,d\sigma(\zeta)\,dr \\ &= \frac{a^2}{4}\sigma(U)\int_0^\varepsilon r^{2k-c+2n-1}\,dr, \end{align*} which diverges when $2k-c+2n\le 0$. Hence \begin{align*} f\in\mathcal I(\varphi_c)_0 \quad\Longleftrightarrow\quad k>\frac c2-n. \end{align*} For $c=2n$, the condition becomes $k>0$, so precisely the germs vanishing at $0$ are allowed: \begin{align*} \mathcal I(2n\log |z|)_0=\mathfrak m_0. \end{align*} For $c=2(n-1)$, the condition becomes $k>-1$, which holds for every holomorphic germ because $k\ge 0$. Thus \begin{align*} \mathcal I(2(n-1)\log |z|)_0=\mathcal O_{\mathbb C^n,0}. \end{align*} The calculation shows that increasing the logarithmic coefficient by $2$ raises the required vanishing order at the origin by one. [/example] The computation is a useful calibration for logarithmic weights. A shift of $2$ in the coefficient of $\log |z|$ changes the multiplier ideal at the origin. ## Open Problems Around Positivity And Singular Weights Which parts of modern complex analysis still resist the package developed in this course? The most active questions keep the same vocabulary, positivity, singular metrics, extension, and curvature, but ask for sharper versions than the basic $L^2$ method supplies by itself. [quotetheorem:3740] This theorem was proved by Guan--Zhou using sharp extension ideas in the spirit of Ohsawa--Takegoshi. It is quoted here without proof because the proof requires a more refined optimal-constant theory than the course develops. [definition: Griffiths Positive Vector Bundle] Let $E\to X$ be a holomorphic Hermitian vector bundle with Chern curvature tensor $\Theta_E$. The bundle $E$ is Griffiths positive if \begin{align*} (i\Theta_E(\xi,\bar{\xi})v,v)_{E_x}>0 \end{align*} for every $x\in X$, every nonzero $\xi\in T_xX$, and every nonzero $v\in E_x$. [/definition] The Griffiths conjecture predicts that every ample holomorphic vector bundle on a compact complex manifold admits a Griffiths positive Hermitian metric. Line bundles are governed by plurisubharmonic weights, but higher-rank bundles introduce matrix-valued curvature and make the positivity condition much harder to construct. [definition: Invariant Jet Differential] Let $X$ be a complex manifold. An invariant jet differential of order $k$ and weighted degree $m$ is a polynomial expression in the derivatives up to order $k$ of germs $f:(\mathbb C,0)\to X$, homogeneous of weighted degree $m$, and invariant under reparametrisations of $(\mathbb C,0)$. [/definition] Demailly's jet differential program seeks enough global jet differentials on projective varieties of general type to constrain entire holomorphic curves $\mathbb C\to X$. The connection with this course is again analytic: singular Hermitian metrics, curvature currents, and $L^2$ extension methods are used to produce sections with controlled vanishing. | Direction | Analytic input from the course | What remains beyond the course | |---|---|---| | Strong openness | Multiplier ideals and sharp extension | Optimal constants and delicate singularity analysis | | Griffiths conjecture | Curvature positivity and Berndtsson-type ideas | Constructing positive metrics on higher-rank ample bundles | | Jet differentials | Singular metrics and vanishing theorems | Producing enough invariant differential equations for entire curves | | [Boundary regularity](/theorems/99) | $\bar{\partial}$-Neumann estimates | Microlocal analysis and finite-type boundary geometry | The common pattern is that $L^2$ methods turn positivity into existence. The rows differ in what is being made to exist: sections, metrics, differential equations for curves, or regular solution operators. Read together, they show that the modern questions are refinements of the same mechanism rather than departures from it. The frontier asks how sharp that existence can be, how regular the objects are, and how much geometry is encoded in the best constants. ## Reading Map Beyond The Course Where should these notes lead next? Hörmander's book remains the canonical source for the original $L^2$ estimates and their use in the Levi problem. Demailly's text develops the modern package of singular metrics, multiplier ideals, Nadel vanishing, and Ohsawa--Takegoshi extension. The MSRI volume edited by Boas and Straube gives entry points into the $\bar{\partial}$-Neumann problem and Bergman theory. Berndtsson's notes give a concise route from the basic estimate to positivity of direct images and the geometric meaning of varying Bergman spaces. The course should now read as the mechanism introduced in Chapter 0 and developed through the intervening chapters rather than as a list of unrelated theorems. Choose the weight, prove the curvature inequality, solve $\bar{\partial}$ with control, and interpret the solution as geometry. ## References