This course develops the modern theory of fibre bundles from the viewpoint of symmetry, geometry, and differential operators. It focuses on principal bundles as the basic objects encoding local triviality together with a [group action](/page/Group%20Action), then shows how many geometric structures can be expressed through them. The main themes are the passage between local and global descriptions, the role of structure groups, and the way connections organize geometry on bundles and on the spaces associated to them.
The chapters begin by recasting principal bundles as symmetry objects and revisiting transition functions in that setting, then move to frame bundles and associated bundles as central examples. After that, the course studies reductions of structure group and uses them to connect bundle geometry with geometric structures on manifolds. The second half introduces connections, covariant derivatives, curvature, gauge transformations, parallel transport, and holonomy, showing how these ideas interact and how they induce geometry on associated bundles and reductions. Throughout, the final chapters tie the theory together through examples and computations that illustrate how the abstract machinery is used in practice.
# Introduction
This introductory chapter explains what the course is trying to organize. The first course on fibre bundles treats local product structure, transition functions, vector bundles, and sections; this course takes the next step by putting symmetry groups at the centre of the theory. Principal bundles record the moving frames, gauges, and choices of coordinates that often sit behind vector bundles, while connections explain how to compare such choices at nearby points.
The guiding theme is that geometric data can be studied by separating two roles: the base manifold contains the points of the underlying space, and a Lie group records the symmetries of the fibres. Once this separation is made, curvature, parallel transport, holonomy, and gauge transformations become different faces of the same construction. The course develops these ideas without yet entering the characteristic-class theory that they support.
## Why Principal Bundles Are Needed
A vector bundle describes a family of vector spaces over a manifold, but many geometric constructions depend on choices of frames rather than on vectors themselves. If $E \to M$ is a rank $n$ vector bundle, a frame at $p \in M$ is an ordered basis of $E_p$, and the set of all frames carries a natural right action of $GL(n,\mathbb R)$. The group action remembers how a frame changes while leaving the underlying point $p$ fixed.
This motivates the central object of the course: a bundle whose fibres are not vector spaces, but spaces on which a Lie group acts freely and transitively. At the introductory level, a principal bundle is a geometric object in which every fibre is a copy of a symmetry group, with no preferred identity element in any fibre. The later formal definition adds the local product and smoothness hypotheses needed to make this picture precise.
[example: Frames of a Vector Bundle]
Let $\pi:E\to M$ be a smooth real vector bundle of rank $n$. For each $p\in M$, let $\operatorname{Fr}(E)_p$ be the set of ordered bases $v=(v_1,\dots,v_n)$ of the fibre $E_p$, and set
\begin{align*}
\operatorname{Fr}(E)=\bigsqcup_{p\in M}\operatorname{Fr}(E)_p .
\end{align*}
The projection sends $v\in \operatorname{Fr}(E)_p$ to $p$. If $A=(A_{ij})\in GL(n,\mathbb R)$, define $v\cdot A=(w_1,\dots,w_n)$ by
\begin{align*}
w_j=\sum_{i=1}^n v_iA_{ij}.
\end{align*}
The tuple $w_1,\dots,w_n$ is again a basis of $E_p$. Indeed, if $\sum_{j=1}^n c_jw_j=0$, then expanding the definition of $w_j$ gives
\begin{align*}
0=\sum_{j=1}^n c_j\sum_{i=1}^n v_iA_{ij}.
\end{align*}
Reordering the finite sum gives
\begin{align*}
0=\sum_{i=1}^n v_i\left(\sum_{j=1}^n A_{ij}c_j\right).
\end{align*}
Since $v_1,\dots,v_n$ are linearly independent, $\sum_{j=1}^n A_{ij}c_j=0$ for every $i$, which is the matrix equation $Ac=0$. Since $A$ is invertible, $c=0$, so $w_1,\dots,w_n$ are linearly independent; in an $n$-dimensional [vector space](/page/Vector%20Space), $n$ linearly independent vectors form a basis.
This defines a right action. If $A,B\in GL(n,\mathbb R)$, $v\cdot A=w$, and $w\cdot B=x$, then for each $k$,
\begin{align*}
x_k=\sum_{j=1}^n w_jB_{jk}.
\end{align*}
Substituting $w_j=\sum_{i=1}^n v_iA_{ij}$ gives
\begin{align*}
x_k=\sum_{j=1}^n\left(\sum_{i=1}^n v_iA_{ij}\right)B_{jk}.
\end{align*}
Reordering the finite sum gives
\begin{align*}
x_k=\sum_{i=1}^n v_i\left(\sum_{j=1}^n A_{ij}B_{jk}\right).
\end{align*}
By the definition of matrix multiplication, $\sum_{j=1}^n A_{ij}B_{jk}=(AB)_{ik}$, so
\begin{align*}
x_k=\sum_{i=1}^n v_i(AB)_{ik}.
\end{align*}
Thus $(v\cdot A)\cdot B=v\cdot(AB)$. Also $v\cdot I=v$, since
\begin{align*}
\sum_{i=1}^n v_iI_{ij}=v_j.
\end{align*}
On each fibre the action is free. If $v\cdot A=v$, then for every $j$,
\begin{align*}
v_j=\sum_{i=1}^n v_iA_{ij}.
\end{align*}
The coordinates of $v_j$ in the basis $(v_1,\dots,v_n)$ are $\delta_{ij}$, so $A_{ij}=\delta_{ij}$ for all $i,j$, hence $A=I$. The action is also transitive. If $v=(v_1,\dots,v_n)$ and $u=(u_1,\dots,u_n)$ are two frames of $E_p$, then each $u_j$ has unique coordinates in the basis $v$:
\begin{align*}
u_j=\sum_{i=1}^n v_iA_{ij}.
\end{align*}
The matrix $A$ is invertible because the [linear map](/page/Linear%20Map) with coordinate matrix $A$ sends the standard basis of $\mathbb R^n$ to the coordinate vectors of the basis $u_1,\dots,u_n$ in the basis $v_1,\dots,v_n$. Therefore $u=v\cdot A$, and freeness shows that this $A$ is unique.
Thus every fibre $\operatorname{Fr}(E)_p$ is a copy of $GL(n,\mathbb R)$ once a reference frame has been chosen, but no frame is preferred. The frame bundle is the basic model of a principal $GL(n,\mathbb R)$-bundle because it records choices of basis rather than vector-valued quantities.
[/example]
The frame example also shows why principal bundles are not an extra topic added after vector bundles. They are the mechanism by which vector bundles are assembled from symmetry data. In Chapter 4, associated bundles reverse the construction: from a principal $G$-bundle and a representation of $G$, we recover vector bundles and tensor bundles.
[remark: No Preferred Frame]
A fibre of a principal bundle is isomorphic to the Lie group $G$, but it is not canonically equal to $G$. Choosing a point in the fibre identifies the fibre with $G$; choosing a different point changes that identification by right multiplication. Much of the subject is about writing constructions that do not depend on such choices.
[/remark]
This absence of preferred choices creates the main technical problem of the course. Local calculations are unavoidable, but geometric statements must survive changes of local product chart. The next section records the language in which the course will handle this tension.
## Local Data and Global Geometry
How can local descriptions on coordinate patches determine a global geometric object? For ordinary fibre bundles the answer is given by transition functions satisfying a cocycle condition. Principal bundles refine this answer by making the transition functions take values in a Lie group that also acts on the fibres.
Suppose $M$ is covered by open sets $(U_i)_{i\in I}$. A principal $G$-bundle will be locally described over each $U_i$ by a product $U_i\times G$. On an overlap $U_i\cap U_j$, the two local descriptions differ by a smooth map $g_{ij}:U_i\cap U_j\to G$, and compatibility on triple overlaps is expressed by a nonabelian cocycle identity.
[quotetheorem:6233]
[citeproof:6233]
This theorem is a roadmap rather than a substitute for the full reconstruction theorem proved later. The open-cover hypothesis is essential because transition functions only compare local products over regions where both products have been chosen; without such local product domains there is no overlap on which $g_{ij}$ can be evaluated. Smoothness is also necessary: continuous but nonsmooth maps $g_{ij}:U_i\cap U_j\to G$ may glue a topological principal bundle, but they do not provide the smooth compatibility needed for a smooth total space. The cocycle condition is the algebraic point that prevents contradictory identifications; if on a triple overlap one has $g_{ij}(p)g_{jk}(p)\ne g_{ik}(p)$, then passing from chart $k$ to chart $i$ directly gives a different point from passing through chart $j$. What the theorem does not say is that the bundle is globally a product, nor that the transition functions are uniquely determined: changing local product charts replaces them by cohomologous transition functions. It also explains why the subject is nonabelian: the order of multiplication in the cocycle identity matters. As a result, principal bundles are classified by a nonabelian version of Cech cohomology rather than by an ordinary cohomology group in general.
[example: The Hopf Fibration as a Principal Bundle]
Consider the projection
\begin{align*}
q:S^3\longrightarrow \mathbb{CP}^1,\qquad q(z_0,z_1)=[z_0:z_1],
\end{align*}
where
\begin{align*}
S^3=\{(z_0,z_1)\in\mathbb C^2:|z_0|^2+|z_1|^2=1\}.
\end{align*}
The group $U(1)=\{\lambda\in\mathbb C:|\lambda|=1\}$ acts by
\begin{align*}
(z_0,z_1)\cdot \lambda=(\lambda z_0,\lambda z_1).
\end{align*}
This action is free: if $(z_0,z_1)\cdot\lambda=(z_0,z_1)$, then $(\lambda-1)z_0=0$ and $(\lambda-1)z_1=0$. Since $(z_0,z_1)\in S^3$, at least one of $z_0,z_1$ is nonzero, so $\lambda=1$.
The orbits are exactly the fibres of $q$. If $(w_0,w_1)$ and $(z_0,z_1)$ have the same image in $\mathbb{CP}^1$, then they span the same complex line, so there is some $\alpha\in\mathbb C^\times$ with
\begin{align*}
(w_0,w_1)=\alpha(z_0,z_1).
\end{align*}
Because both vectors lie in $S^3$,
\begin{align*}
1=|w_0|^2+|w_1|^2.
\end{align*}
Substituting $w_0=\alpha z_0$ and $w_1=\alpha z_1$ gives
\begin{align*}
1=|\alpha z_0|^2+|\alpha z_1|^2.
\end{align*}
Using $|\alpha z_i|^2=|\alpha|^2|z_i|^2$ gives
\begin{align*}
1=|\alpha|^2(|z_0|^2+|z_1|^2).
\end{align*}
Since $(z_0,z_1)\in S^3$, this becomes
\begin{align*}
1=|\alpha|^2.
\end{align*}
Thus $\alpha\in U(1)$, so $(w_0,w_1)=(z_0,z_1)\cdot\alpha$. Conversely, multiplying by any $\lambda\in U(1)$ does not change the complex line:
\begin{align*}
[\lambda z_0:\lambda z_1]=[z_0:z_1].
\end{align*}
On the standard affine charts
\begin{align*}
U_0=\{[z_0:z_1]:z_0\ne0\}
\end{align*}
and
\begin{align*}
U_1=\{[z_0:z_1]:z_1\ne0\},
\end{align*}
define local sections by
\begin{align*}
s_0([1:t])=\frac{(1,t)}{\sqrt{1+|t|^2}}
\end{align*}
and
\begin{align*}
s_1([s:1])=\frac{(s,1)}{\sqrt{1+|s|^2}}.
\end{align*}
On $U_0\cap U_1$, write $s=1/t$ with $t\ne0$. Then
\begin{align*}
s_1([1:t])=\frac{(1/t,1)}{\sqrt{1+|1/t|^2}}.
\end{align*}
Since $|1/t|^2=1/|t|^2$, this is
\begin{align*}
s_1([1:t])=\frac{(1/t,1)}{\sqrt{1+1/|t|^2}}.
\end{align*}
Because $1+1/|t|^2=(|t|^2+1)/|t|^2$, we get
\begin{align*}
s_1([1:t])=\frac{|t|}{\sqrt{1+|t|^2}}(1/t,1).
\end{align*}
Since $(|t|/t)(1,t)=(|t|/t,|t|)$, the last expression is
\begin{align*}
s_1([1:t])=\frac{|t|/t}{\sqrt{1+|t|^2}}(1,t).
\end{align*}
Therefore
\begin{align*}
s_1([1:t])=s_0([1:t])\cdot\frac{|t|}{t}.
\end{align*}
The transition function on the overlap is
\begin{align*}
g_{01}(t)=\frac{|t|}{t}\in U(1).
\end{align*}
This exhibits the Hopf fibration as a principal $U(1)$-bundle: the action is free, its orbits are the fibres, and the standard local sections differ by a $U(1)$-valued transition function. Along the circle $|t|=1$, the transition function is
\begin{align*}
g_{01}(e^{i\theta})=e^{-i\theta},
\end{align*}
so it winds once around $U(1)$. A global product $S^3\cong\mathbb{CP}^1\times U(1)$ would give a global section and hence would remove this transition function by compatible local $U(1)$-valued changes of section; the winding-one transition cannot be removed in that way. Thus the Hopf fibration is locally a product principal bundle, but not globally a product.
[/example]
[illustration:hopf-fibration-transition]
The Hopf fibration will reappear because it combines topology and geometry in a compact form. It is locally a product, but globally it cannot be identified with $S^2\times U(1)$. Connections in Chapters 6 and 8 add differential information to this topological example through local potentials and curvature.
## Connections as Infinitesimal Comparison
A principal bundle records the possible choices of frame or gauge at each point, but it does not by itself say how to compare choices at nearby points. The new question is infinitesimal: given a tangent vector on the base manifold, which tangent directions upstairs should count as horizontal lifts? A connection answers this question in a way compatible with the group action.
There are two equivalent viewpoints that the course will move between. A connection may be described as a $G$-equivariant horizontal distribution in $TP$, complementary to the vertical tangent spaces. It may also be described by a Lie-algebra-valued $1$-form on $P$, called the connection form, which returns the infinitesimal group element generating the vertical part of a tangent vector.
[definition: Vertical Tangent Space]
Let $\pi:P\to M$ be a smooth fibre bundle and let $u\in P$. The vertical tangent space at $u$ is
\begin{align*}
d\pi_u&:T_uP\longrightarrow T_{\pi(u)}M, &
V_uP&:=\ker d\pi_u\subset T_uP.
\end{align*}
[/definition]
The vertical tangent space consists of directions that move inside a single fibre. To differentiate along the base, we need a complementary family of directions that projects isomorphically onto tangent spaces of $M$. In a principal bundle that complement must also respect the right $G$-action.
[definition: Horizontal Distribution]
Let $\pi:P\to M$ be a smooth principal $G$-bundle. A horizontal distribution on $P$ is a smooth map
\begin{align*}
H:P&\longrightarrow \operatorname{Gr}(\dim M,TP), & u&\longmapsto H_uP\subset T_uP,
\end{align*}
where $\operatorname{Gr}(\dim M,TP)$ denotes the Grassmann bundle of $\dim M$-planes in $TP$, such that
\begin{align*}
T_uP&=H_uP\oplus V_uP, & d(R_g)_u(H_uP)&=H_{u\cdot g}P
\end{align*}
for every $u\in P$ and $g\in G$, where $R_g:P\to P$ denotes right multiplication by $g$.
[/definition]
This definition is geometric: it specifies which tangent vectors are horizontal. The next problem is to encode the same splitting in a form that can be differentiated, pulled back by local sections, and inserted into curvature formulae. The connection-form viewpoint supplies exactly that algebraic package.
[definition: Connection Form]
Let $\pi:P\to M$ be a smooth principal $G$-bundle with Lie algebra $\mathfrak g$. For $\xi\in\mathfrak g$, let $\xi_P:P\to TP$ be the fundamental vector field generated by $\xi$, defined by
\begin{align*}
\xi_P(u)&:=\left.\frac{d}{dt}\right|_{t=0}u\cdot \exp(t\xi)\in T_uP.
\end{align*}
For each $u\in P$, the infinitesimal action map is
\begin{align*}
\kappa_u:\mathfrak g&\longrightarrow V_uP, & \xi&\longmapsto \xi_P(u).
\end{align*}
A connection form on $P$ is a differential form $\omega\in\Omega^1(P;\mathfrak g)$, meaning a smooth family of linear maps
\begin{align*}
\omega_u:T_uP\longrightarrow \mathfrak g,
\end{align*}
such that
\begin{align*}
\omega_u(\xi_P(u))&=\xi, & (R_g)^*\omega&=\operatorname{Ad}_{g^{-1}}\omega
\end{align*}
for all $\xi\in\mathfrak g$, $u\in P$, and $g\in G$.
[/definition]
The first condition says that the form recognizes vertical directions coming from the group action. The second condition records equivariance under changes of frame. Since horizontal distributions and connection forms are meant to describe the same geometric choice, the course needs a precise dictionary between them before using either language in later chapters.
[quotetheorem:6234]
[citeproof:6234]
This equivalence lets the course choose the most efficient language for each calculation. The complement condition is necessary because $\ker d\pi_u$ alone only describes motion inside the fibre; if $H_uP$ fails to complement $V_uP$, a tangent vector on $M$ may have no unique horizontal lift. Equivariance is equally necessary: choosing arbitrary complements in each tangent space can produce horizontal lifts that depend on the chosen representative $u\in P_p$, so right multiplication by $g$ would change the rule rather than transport it. The bijection also has a precise limitation: it identifies two ways of encoding the same connection on a fixed principal bundle, but it does not classify principal bundles and it does not make a curved connection locally gauge-equivalent to the product flat connection. Horizontal distributions are well suited to parallel transport, while connection forms make curvature and gauge transformation formulae compact.
[illustration:connection-horizontal-vertical-splitting]
## Curvature, Parallel Transport, and Holonomy
Once a connection is chosen, the next question is whether horizontal directions integrate consistently around small loops. Curvature measures the failure of horizontal lifts to close up in the way expected from a flat product. Parallel transport turns horizontal lifting into a method for comparing fibres along paths.
For a connection form $\omega$, its curvature is a $\mathfrak g$-valued $2$-form on $P$. The formula uses both the [exterior derivative](/theorems/1525) and the Lie bracket on $\mathfrak g$, reflecting that the structure group may be nonabelian.
[definition: Curvature Form]
Let $\omega\in\Omega^1(P;\mathfrak g)$ be a connection form on a smooth principal $G$-bundle $P\to M$. The curvature form of $\omega$ is
\begin{align*}
\Omega:=d\omega+\frac{1}{2}[\omega\wedge\omega]\in\Omega^2(P;\mathfrak g).
\end{align*}
[/definition]
The curvature form is horizontal and equivariant, so it descends to tensorial data on the base after choosing associated bundles. In local gauge, a Lie-algebra-valued local connection form $A$ has curvature
\begin{align*}
F_A=dA+\frac{1}{2}[A\wedge A].
\end{align*}
[example: Flat Connection on a Product Bundle]
Let $P=M\times G$ with right action $(p,a)\cdot g=(p,ag)$, and write $T_{(p,a)}P\simeq T_pM\oplus T_aG$. The product connection declares the $M$-directions horizontal, so
\begin{align*}
H_{(p,a)}P=T_pM\oplus\{0\}.
\end{align*}
Let $\theta_L\in\Omega^1(G;\mathfrak g)$ be the left Maurer-Cartan form. For a matrix group, this means
\begin{align*}
(\theta_L)_a(Y)=a^{-1}Y
\end{align*}
for $Y\in T_aG$. Define
\begin{align*}
\omega=\operatorname{pr}_G^*\theta_L .
\end{align*}
If $(X,Y)\in T_pM\oplus T_aG$, then $d(\operatorname{pr}_G)_{(p,a)}(X,Y)=Y$, and therefore
\begin{align*}
\omega_{(p,a)}(X,Y)=(\theta_L)_a(Y).
\end{align*}
Since $(\theta_L)_a:T_aG\to\mathfrak g$ is a linear isomorphism, $\omega_{(p,a)}(X,Y)=0$ holds exactly when $Y=0$. Hence
\begin{align*}
\ker\omega_{(p,a)}=T_pM\oplus\{0\}=H_{(p,a)}P.
\end{align*}
For $\xi\in\mathfrak g$, the fundamental vector field for the right action is
\begin{align*}
\xi_P(p,a)=\left(0,\left.\frac{d}{dt}\right|_{t=0}a\exp(t\xi)\right).
\end{align*}
Applying $\omega$ gives
\begin{align*}
\omega_{(p,a)}(\xi_P(p,a))=(\theta_L)_a\left(\left.\frac{d}{dt}\right|_{t=0}a\exp(t\xi)\right).
\end{align*}
For a matrix group, $(\theta_L)_a(Y)=a^{-1}Y$, so
\begin{align*}
\omega_{(p,a)}(\xi_P(p,a))=a^{-1}\left.\frac{d}{dt}\right|_{t=0}a\exp(t\xi).
\end{align*}
Because $a$ is constant in $t$,
\begin{align*}
a^{-1}\left.\frac{d}{dt}\right|_{t=0}a\exp(t\xi)=\left.\frac{d}{dt}\right|_{t=0}\exp(t\xi)=\xi.
\end{align*}
Thus $\omega(\xi_P)=\xi$.
The equivariance identity also follows from the same formula. Since $\operatorname{pr}_G\circ R_g=r_g\circ\operatorname{pr}_G$, where $r_g(a)=ag$, we have
\begin{align*}
(R_g)^*\omega=\operatorname{pr}_G^*(r_g^*\theta_L).
\end{align*}
For a matrix group and $Y\in T_aG$, the differential of $r_g$ sends $Y$ to $Yg\in T_{ag}G$. Therefore
\begin{align*}
(r_g^*\theta_L)_a(Y)=(\theta_L)_{ag}(Yg).
\end{align*}
Using the left Maurer-Cartan formula,
\begin{align*}
(\theta_L)_{ag}(Yg)=(ag)^{-1}Yg.
\end{align*}
Since $(ag)^{-1}=g^{-1}a^{-1}$,
\begin{align*}
(ag)^{-1}Yg=g^{-1}a^{-1}Yg=\operatorname{Ad}_{g^{-1}}\bigl((\theta_L)_a(Y)\bigr).
\end{align*}
Hence
\begin{align*}
(R_g)^*\omega=\operatorname{Ad}_{g^{-1}}\omega.
\end{align*}
The curvature of this connection is
\begin{align*}
\Omega=d\omega+\frac{1}{2}[\omega\wedge\omega].
\end{align*}
Substituting $\omega=\operatorname{pr}_G^*\theta_L$ gives
\begin{align*}
\Omega=d(\operatorname{pr}_G^*\theta_L)+\frac{1}{2}[\operatorname{pr}_G^*\theta_L\wedge\operatorname{pr}_G^*\theta_L].
\end{align*}
Exterior derivative commutes with pullback, and the bracket-wedge product of pulled-back Lie-algebra-valued forms is the pullback of the bracket-wedge product, so
\begin{align*}
\Omega=\operatorname{pr}_G^*(d\theta_L)+\frac{1}{2}\operatorname{pr}_G^*[\theta_L\wedge\theta_L].
\end{align*}
Combining the two pullback terms,
\begin{align*}
\Omega=\operatorname{pr}_G^*\left(d\theta_L+\frac{1}{2}[\theta_L\wedge\theta_L]\right).
\end{align*}
By the left Maurer-Cartan equation,
\begin{align*}
d\theta_L+\frac{1}{2}[\theta_L\wedge\theta_L]=0,
\end{align*}
and therefore
\begin{align*}
\Omega=0.
\end{align*}
A curve $(\gamma(t),a(t))$ in $M\times G$ is horizontal exactly when
\begin{align*}
0=\omega_{(\gamma(t),a(t))}(\gamma'(t),a'(t))=(\theta_L)_{a(t)}(a'(t)).
\end{align*}
Since $(\theta_L)_{a(t)}$ is an isomorphism, this is equivalent to $a'(t)=0$. Thus horizontal transport keeps the $G$-coordinate constant, so in this product flat gauge parallel transport depends only on the endpoints of the base path.
[/example]
The product example shows what it means for curvature to vanish locally, but it does not yet describe what happens after transport around a closed path. A loop can return to the same base point while moving the chosen point of the principal fibre by a group element. The subgroup generated by all such return effects is the next invariant to define.
[definition: Holonomy Group]
Let $\pi:P\to M$ be a smooth principal $G$-bundle with connection, let $p\in M$, and let $u\in P_p$. The holonomy group at $u$ is the subgroup of $G$ consisting of all elements $g\in G$ such that horizontal transport around some piecewise smooth loop based at $p$ sends $u$ to $u\cdot g$.
[/definition]
Holonomy links the infinitesimal and global sides of the course. Curvature gives infinitesimal information about small loops, while holonomy records the accumulated effect along arbitrary loops. Chapter 10 makes this relationship precise through the [Ambrose-Singer theorem](/theorems/6282).
## Gauge Transformations and Associated Bundles
A final organizing question runs through the course: when do two local descriptions represent the same geometry? Gauge transformations are automorphisms of a principal bundle covering the identity on the base. They change the frame or gauge in which connection and curvature are written, but they do not change the underlying geometric object.
[definition: Gauge Transformation]
Let $\pi:P\to M$ be a smooth principal $G$-bundle. A gauge transformation of $P$ is a principal bundle automorphism $\Phi:P\to P$ such that
\begin{align*}
\pi\circ\Phi=\pi.
\end{align*}
[/definition]
Gauge transformations are symmetries of the description rather than symmetries of the base manifold. To compare calculations made in different local sections, the course needs the exact formula for how a connection form changes under a gauge transformation. That formula explains which parts of a connection are coordinate-dependent and prepares the later gauge-covariance results.
[quotetheorem:6235]
[citeproof:6235]
This transformation law is the prototype for gauge covariance throughout the course. The hypothesis that $s'$ has the form $s\cdot h$ is what makes the formula local: it compares two sections over the same [open set](/page/Open%20Set) $U$ using a single smooth map $h:U\to G$. If $s$ is defined on $U$ and $s'$ is defined on $V$, there is no single map $h:U\to G$ comparing them on all of $U$; for instance, the two standard sections of the Hopf fibration over the affine charts of $\mathbb{CP}^1$ are only compared on the overlap $U\cap V$, where the transition map takes values in $U(1)$. If $h$ is not smooth, then $h^*\theta$ is not a smooth $1$-form and the transformed local connection form is not defined in the smooth category; a continuous map $h:\mathbb R\to U(1)$ with a corner gives a concrete failure. The principal-bundle hypothesis also matters: for a general fibre bundle without a free transitive right $G$-action, the expression $s\cdot h$ is not defined and there is no canonical Maurer-Cartan term. Finally, the convention for $\theta$ is part of the statement. Using the right Maurer-Cartan form instead of the left one changes the adjoint placement and sign convention; for a nonabelian matrix group, replacing $h^{-1}dh$ by $dh\,h^{-1}$ gives a different local formula. The inhomogeneous term is necessary: even for the flat connection $A=0$ on a product bundle, changing gauge by a nonconstant map $h:U\to G$ gives $A'=h^*\theta$, not zero. The formula does not say that a global gauge exists putting every connection into a preferred form, and it does not remove curvature. Curvature transforms without the inhomogeneous Maurer-Cartan term, which is why curvature is closer to a tensor than the connection form itself.
[example: Associated Vector Bundle]
Let $P\to M$ be a principal $G$-bundle and let $\rho:G\to GL(V)$ be a finite-dimensional representation. Define
\begin{align*}
P\times_G V:=(P\times V)/\sim,
\end{align*}
where the generating relation is
\begin{align*}
(u\cdot g,v)\sim(u,\rho(g)v).
\end{align*}
Write the equivalence class of $(u,v)$ as $[u,v]$. The projection to $M$ is
\begin{align*}
[u,v]\longmapsto \pi(u),
\end{align*}
because $\pi(u\cdot g)=\pi(u)$ for the right action on a principal bundle.
Let $s:U\to P$ be a local section. Define
\begin{align*}
\Phi_s:U\times V&\longrightarrow (P\times_G V)|_U, &
(p,v)&\longmapsto [s(p),v].
\end{align*}
This map is onto over $U$. Indeed, if $[u,w]$ lies over $p\in U$, then $\pi(u)=p$ and $u$ and $s(p)$ lie in the same principal fibre. Since the right $G$-action on each fibre is transitive, there is some $g\in G$ such that
\begin{align*}
u=s(p)\cdot g.
\end{align*}
Then
\begin{align*}
[u,w]=[s(p)\cdot g,w]=[s(p),\rho(g)w],
\end{align*}
so $[u,w]=\Phi_s(p,\rho(g)w)$.
The map is also one-to-one. Suppose
\begin{align*}
\Phi_s(p,v)=\Phi_s(q,w),
\end{align*}
so
\begin{align*}
[s(p),v]=[s(q),w].
\end{align*}
Equivalent pairs project to the same point of $M$, hence
\begin{align*}
p=\pi(s(p))=\pi(s(q))=q.
\end{align*}
With $p=q$, both first coordinates are $s(p)$. If the [equivalence relation](/page/Equivalence%20Relation) changes $(s(p),v)$ to another pair whose first coordinate is again $s(p)$, then the principal action must use an element $g\in G$ with
\begin{align*}
s(p)\cdot g=s(p).
\end{align*}
Since the action is free, $g=e$. Therefore
\begin{align*}
w=\rho(e)v=v,
\end{align*}
because $\rho(e)=I_V$. Thus $\Phi_s$ identifies the associated bundle over $U$ with $U\times V$.
Now compare two local sections over the same open set, with
\begin{align*}
s'(p)=s(p)\cdot h(p)
\end{align*}
for a smooth map $h:U\to G$. If the same point of the associated bundle has $s'$-coordinates $v'$, then
\begin{align*}
[s'(p),v']=[s(p)\cdot h(p),v'].
\end{align*}
Using the defining relation with $g=h(p)$ gives
\begin{align*}
[s(p)\cdot h(p),v']=[s(p),\rho(h(p))v'].
\end{align*}
Thus its $s$-coordinates are
\begin{align*}
v=\rho(h(p))v',
\end{align*}
or equivalently
\begin{align*}
v'=\rho(h(p))^{-1}v.
\end{align*}
So a principal bundle records the changing choice of frame or gauge, while the representation $\rho$ turns those changes into linear coordinate changes on $V$. This is how principal bundles produce the vector bundles used in differential geometry.
[/example]
Associated bundles close the introductory circle. Principal bundles encode symmetry choices; representations turn those choices into vector spaces; connections on principal bundles induce covariant derivatives on associated vector bundles. The rest of the course develops these constructions in detail and uses them as the common language for frames, curvature, transport, gauge, and holonomy.
Having seen how representations turn principal data into vector bundles and covariant derivatives, we now return to the underlying symmetry object itself. The next chapter begins from the idea that a principal bundle remembers the group action rather than a preferred point in each fibre, and uses that viewpoint to set up the basic language for the rest of the course.
# 1. Principal Bundles as Symmetry Objects
Principal bundles are the form in which a geometric object remembers only its symmetries and not a preferred origin in each fibre. The previous course on fibre bundles treated local products and transition maps; here the fibre is a Lie group $G$, but the group is not merely a model fibre: it acts freely and transitively on every fibre. This chapter sets up that viewpoint, studies maps that respect it, and uses the product bundle $M \times G$ to introduce the Maurer-Cartan form as the first connection-like object.
## From Fibres to Group Actions
The first question is how to encode a family of copies of a Lie group without choosing an identity element in each copy. A principal bundle answers this by replacing fibre coordinates with a right action of $G$; each fibre is a $G$-torsor rather than a group with a distinguished identity.
[definition: Smooth Principal Bundle]
Let $G$ be a Lie group and let $M$ be a smooth manifold. A smooth principal $G$-bundle over $M$ consists of a smooth manifold $P$, a smooth surjective submersion $\pi:P\to M$, and a smooth right action
$P\times G\to P$, $(p,g)\mapsto p\cdot g$, such that:
1. $\pi(p\cdot g)=\pi(p)$ for all $p\in P$ and $g\in G$.
2. For every $x\in M$, the action of $G$ on the fibre $P_x=\pi^{-1}(x)$ is free and transitive.
3. For every $x\in M$, there exists an open neighbourhood $U\subset M$ of $x$ and a $G$-equivariant diffeomorphism
\begin{align*}
\Phi_U:\pi^{-1}(U)\to U\times G
\end{align*}
over $U$, where $G$ acts on $U\times G$ by right multiplication in the second factor.
[/definition]
The freeness condition says that no non-identity group element fixes a point of $P$, while transitivity says that two points in the same fibre differ by a unique group element. Thus, for $p,q\in P$ with $\pi(p)=\pi(q)$, there is a unique $g\in G$ such that $q=p\cdot g$. The local product condition turns this fibrewise symmetry statement into a smooth bundle condition.
[example: Product Principal Bundle]
Let $M$ be a smooth manifold and let $G$ be a Lie group. On $P=M\times G$, define the projection by $\pi(x,h)=x$, and let $G$ act on the right by
\begin{align*}
(x,h)\cdot g=(x,hg).
\end{align*}
This action preserves the projection, since
\begin{align*}
\pi((x,h)\cdot g)=\pi(x,hg)=x=\pi(x,h).
\end{align*}
For a fixed $x\in M$, the fibre is $P_x=\pi^{-1}(x)=\{x\}\times G$. The action on this fibre is free: if $(x,h)\cdot g=(x,h)$, then $(x,hg)=(x,h)$, so $hg=h$ in $G$; multiplying on the left by $h^{-1}$ gives $g=h^{-1}h=e$. The action is transitive: given $(x,h_1),(x,h_2)\in \{x\}\times G$, choose $g=h_1^{-1}h_2$. Then
\begin{align*}
(x,h_1)\cdot g=(x,h_1(h_1^{-1}h_2))=(x,(h_1h_1^{-1})h_2)=(x,eh_2)=(x,h_2).
\end{align*}
The identity map $\operatorname{id}_{M\times G}:M\times G\to M\times G$ is a global product chart over $M$, and it is $G$-equivariant because
\begin{align*}
\operatorname{id}_{M\times G}((x,h)\cdot g)=(x,hg)=\operatorname{id}_{M\times G}(x,h)\cdot g.
\end{align*}
Thus $M\times G\to M$ satisfies the defining conditions of a principal $G$-bundle, and it is the product model used in every local trivialization.
[/example]
This model makes the order of multiplication important. We use right actions because the associated bundle construction and the usual convention for transition functions then match the formulae for frames and gauge transformations used later in the course.
[remark: Torsor Viewpoint]
A fibre $P_x$ of a principal $G$-bundle is not canonically equal to $G$. Choosing $p\in P_x$ identifies $G$ with $P_x$ by $g\mapsto p\cdot g$, but changing $p$ changes this identification by left multiplication on $G$. Principal geometry is the study of constructions that do not depend on making such choices.
[/remark]
The torsor viewpoint leaves an important recognition problem: when does a free symmetry action actually come from a principal bundle over its orbit space? The quotient point of view answers this by adding properness, which gives a well-behaved smooth orbit space and enough local slices to produce the required local product charts.
[quotetheorem:6236]
[citeproof:6236]
This theorem explains why principal bundles appear whenever geometry is formed by taking a smooth space of choices and quotienting by a symmetry group. It is a recognition theorem, not a classification theorem: it does not say that every principal bundle is globally a product, nor does it produce a preferred section of $P\to P/G$.
The hypotheses are not cosmetic. If a free action is not proper, the orbit space can fail to be Hausdorff; the standard irrational-flow type examples show that smooth free orbits may be dense, so the quotient cannot serve as a smooth base manifold. If the action is proper but not free, stabilisers survive: the fibre over an orbit has the form $G/H$ rather than $G$, producing quotient geometries closer to orbifolds or stratified spaces than principal bundles. Thus in practice one checks freeness, properness, and local product structure separately before calling a quotient a principal bundle.
[example: Hopf Fibration]
Let
\begin{align*}
S^3=\{(z_1,z_2)\in\mathbb C^2: |z_1|^2+|z_2|^2=1\},
\end{align*}
and let $U(1)=\{\lambda\in\mathbb C:|\lambda|=1\}$ act on the right by
\begin{align*}
(z_1,z_2)\cdot \lambda=(z_1\lambda,z_2\lambda).
\end{align*}
The action preserves $S^3$, because
\begin{align*}
|z_1\lambda|^2+|z_2\lambda|^2=|z_1|^2|\lambda|^2+|z_2|^2|\lambda|^2=|z_1|^2+|z_2|^2=1.
\end{align*}
It is free: if $(z_1,z_2)\cdot\lambda=(z_1,z_2)$, then $z_1(\lambda-1)=0$ and $z_2(\lambda-1)=0$. Since $(z_1,z_2)\in S^3$, at least one of $z_1,z_2$ is nonzero, so $\lambda-1=0$.
Define
\begin{align*}
q:S^3\to \mathbb CP^1,\qquad q(z_1,z_2)=[z_1:z_2].
\end{align*}
For $\lambda\in U(1)$,
\begin{align*}
q((z_1,z_2)\cdot\lambda)=[z_1\lambda:z_2\lambda]=[z_1:z_2],
\end{align*}
so each $U(1)$-orbit lies in a fibre of $q$. Conversely, if $q(w_1,w_2)=q(z_1,z_2)$, then $(w_1,w_2)=c(z_1,z_2)$ for some $c\in\mathbb C^\times$. Because both vectors lie in $S^3$,
\begin{align*}
1=|w_1|^2+|w_2|^2=|c|^2(|z_1|^2+|z_2|^2)=|c|^2,
\end{align*}
so $c\in U(1)$ and $(w_1,w_2)=(z_1,z_2)\cdot c$. Thus the fibres of $q$ are exactly the $U(1)$-orbits.
On the affine chart $V_1=\{[z_1:z_2]:z_1\ne 0\}$, write $\zeta=z_2/z_1$ and define
\begin{align*}
s_1([1:\zeta])=\frac{(1,\zeta)}{\sqrt{1+|\zeta|^2}}.
\end{align*}
Then
\begin{align*}
\left|\frac{1}{\sqrt{1+|\zeta|^2}}\right|^2+\left|\frac{\zeta}{\sqrt{1+|\zeta|^2}}\right|^2=\frac{1+|\zeta|^2}{1+|\zeta|^2}=1,
\end{align*}
so $s_1$ takes values in $S^3$, and $q(s_1([1:\zeta]))=[1:\zeta]$. If $(z_1,z_2)\in q^{-1}(V_1)$, then $z_1\ne 0$, $\zeta=z_2/z_1$, and
\begin{align*}
1+|\zeta|^2=1+\left|\frac{z_2}{z_1}\right|^2=\frac{|z_1|^2+|z_2|^2}{|z_1|^2}=\frac{1}{|z_1|^2}.
\end{align*}
Hence $(1+|\zeta|^2)^{-1/2}=|z_1|$. With $\lambda=z_1/|z_1|\in U(1)$, we get
\begin{align*}
s_1([1:\zeta])\cdot\lambda=(|z_1|,|z_1|\zeta)\cdot\frac{z_1}{|z_1|}=\left(z_1,|z_1|\frac{z_2}{z_1}\frac{z_1}{|z_1|}\right)=(z_1,z_2).
\end{align*}
Therefore the map
\begin{align*}
V_1\times U(1)\to q^{-1}(V_1),\qquad ([1:\zeta],\lambda)\mapsto s_1([1:\zeta])\cdot\lambda
\end{align*}
is a product chart over $V_1$. On the second affine chart $V_2=\{[z_1:z_2]:z_2\ne 0\}$, the same construction uses
\begin{align*}
s_2([\eta:1])=\frac{(\eta,1)}{\sqrt{1+|\eta|^2}}.
\end{align*}
The charts $V_1$ and $V_2$ cover $\mathbb CP^1$.
The action is proper because the map $S^3\times U(1)\to S^3\times S^3$ given by $((z_1,z_2),\lambda)\mapsto ((z_1,z_2),(z_1\lambda,z_2\lambda))$ has compact domain and Hausdorff target, so the inverse image of every compact subset is compact. Hence $q:S^3\to\mathbb CP^1$ is a principal $U(1)$-bundle. Since $\mathbb CP^1$ is diffeomorphic to $S^2$, this is the Hopf fibration $S^3\to S^2$, the standard non-product principal circle bundle.
[/example]
The Hopf fibration is the first non-product example: locally it looks like $U\times U(1)$, but no global choice of one point in every fibre exists. Chapter 6 introduces its local connection forms, and Chapters 8 and 12 compute the corresponding curvature as the model for magnetic monopole geometry.
[illustration:hopf-local-product-charts]
A second central source of principal bundles comes from frames. Instead of attaching a vector space to every point, we attach all admissible bases of the tangent space; the structure group records how bases change.
[example: Oriented Orthonormal Frame Bundle]
Let $(M,g)$ be an oriented Riemannian $n$-manifold. Define $F_{SO}(M)$ to be the set of all ordered oriented orthonormal bases $(e_1,\dots,e_n)$ of $T_xM$ as $x$ varies over $M$, and let $\pi:F_{SO}(M)\to M$ send a frame to its base point. For $A\in SO(n)$, define the right action by
\begin{align*}
(e_1,\dots,e_n)\cdot A=\left(\sum_{i=1}^n e_i A_{i1},\dots,\sum_{i=1}^n e_i A_{in}\right).
\end{align*}
Writing $f_j=\sum_{i=1}^n e_iA_{ij}$, the new frame has the same base point, so $\pi((e_1,\dots,e_n)\cdot A)=\pi(e_1,\dots,e_n)$. It is still orthonormal because, for each $j,k$,
\begin{align*}
g(f_j,f_k)=g\left(\sum_{i=1}^n e_i A_{ij},\sum_{\ell=1}^n e_\ell A_{\ell k}\right)=\sum_{i=1}^n\sum_{\ell=1}^n A_{ij}A_{\ell k}g(e_i,e_\ell).
\end{align*}
Since $(e_1,\dots,e_n)$ is orthonormal, $g(e_i,e_\ell)=\delta_{i\ell}$, hence
\begin{align*}
g(f_j,f_k)=\sum_{i=1}^n\sum_{\ell=1}^n A_{ij}A_{\ell k}\delta_{i\ell}=\sum_{i=1}^n A_{ij}A_{ik}=(A^\top A)_{jk}.
\end{align*}
Because $A\in SO(n)$, $A^\top A=I$, so $g(f_j,f_k)=\delta_{jk}$. The frame is still oriented because the change-of-basis matrix from $(e_1,\dots,e_n)$ to $(f_1,\dots,f_n)$ is $A$, and $\det A=1$.
For a fixed $x\in M$, the fibre $F_{SO}(M)_x$ is the set of oriented orthonormal bases of $T_xM$. The action on this fibre is free: if $(e_1,\dots,e_n)\cdot A=(e_1,\dots,e_n)$, then for each $j$,
\begin{align*}
\sum_{i=1}^n e_i A_{ij}=e_j=\sum_{i=1}^n e_i\delta_{ij}.
\end{align*}
Since $(e_1,\dots,e_n)$ is a basis, the coefficients agree, so $A_{ij}=\delta_{ij}$ for all $i,j$, and therefore $A=I$. The action is transitive: if $(e_1,\dots,e_n)$ and $(f_1,\dots,f_n)$ are two oriented orthonormal bases of $T_xM$, then each $f_j$ has unique coordinates
\begin{align*}
f_j=\sum_{i=1}^n e_iA_{ij}.
\end{align*}
The orthonormality computation gives $g(f_j,f_k)=(A^\top A)_{jk}$, and since $g(f_j,f_k)=\delta_{jk}$, we get $A^\top A=I$. Thus $(\det A)^2=\det(A^\top A)=\det I=1$. Since both frames are oriented, the change-of-basis determinant is positive, so $\det A=1$. Hence $A\in SO(n)$ and $(f_1,\dots,f_n)=(e_1,\dots,e_n)\cdot A$.
Now choose an oriented orthonormal local frame $(s_1,\dots,s_n)$ over an open set $U\subset M$. Every frame $(e_1,\dots,e_n)$ over $x\in U$ can be written uniquely as
\begin{align*}
e_j=\sum_{i=1}^n s_i(x)A_{ij}
\end{align*}
for a matrix $A\in SO(n)$ by the same coordinate and orientation argument. Therefore
\begin{align*}
F_{SO}(M)|_U\to U\times SO(n),\qquad (e_1,\dots,e_n)\mapsto (x,A)
\end{align*}
is a local product chart, with inverse
\begin{align*}
(x,A)\mapsto \left(\sum_{i=1}^n s_i(x)A_{i1},\dots,\sum_{i=1}^n s_i(x)A_{in}\right).
\end{align*}
In this chart, changing frames by $B\in SO(n)$ sends $A$ to $AB$, since the $j$th vector of the changed frame is
\begin{align*}
\sum_{k=1}^n\left(\sum_{i=1}^n s_i(x)A_{ik}\right)B_{kj}=\sum_{i=1}^n s_i(x)\left(\sum_{k=1}^n A_{ik}B_{kj}\right)=\sum_{i=1}^n s_i(x)(AB)_{ij}.
\end{align*}
Thus $F_{SO}(M)\to M$ is locally $U\times SO(n)\to U$, and the fibre over each point records all oriented orthonormal choices of basis with no preferred one.
[/example]
The frame bundle shows the guiding principle of the course: a connection on a principal bundle will later become a rule for comparing frames at nearby points. The principal-bundle language separates the intrinsic symmetry group from any particular vector representation.
## Equivariant Maps and Pullbacks
Once principal bundles have been defined, the next question is what the correct maps between them should be. A map of total spaces is too loose unless it preserves base points and the right action; the category of principal bundles is built from equivariant maps.
A fibre-preserving map alone is not enough. On a product bundle $M\times G$, the map
\begin{align*}
I:M\times G&\to M\times G, & I(x,h)&=(x,h^{-1})
\end{align*}
preserves the base point but reverses the torsor structure unless $G$ is abelian in a special way; it does not commute with right multiplication. The definition therefore insists that maps preserve the group action, not merely the projection.
[definition: Principal Bundle Morphism]
Let $\pi:P\to M$ and $\pi':P'\to M'$ be smooth principal $G$-bundles. A principal bundle morphism from $P$ to $P'$ is a pair of smooth maps $F:P\to P'$ and $f:M\to M'$ such that
\begin{align*}
\pi'\circ F=f\circ \pi
\end{align*}
and
\begin{align*}
F(p\cdot g)=F(p)\cdot g
\end{align*}
for all $p\in P$ and $g\in G$.
[/definition]
The first condition says that $F$ sends fibres to fibres over the map $f$, and the second says that it preserves the $G$-torsor structure on each fibre. A particularly important question is which such maps are internal symmetries of a fixed principal bundle; these will become the gauge transformations acting on connections.
[definition: Bundle Automorphism]
Let $\pi:P\to M$ be a smooth principal $G$-bundle. A bundle automorphism of $P$ is a principal bundle morphism $F:P\to P$ covering $\operatorname{id}_M$ that is a diffeomorphism.
[/definition]
Bundle [automorphisms form a group](/theorems/4982) under composition. They are also called gauge transformations, and later this group will act on connection forms and curvature.
[example: Gauge Transformations Of A Product Bundle]
For the product bundle $P=M\times G$, let $u:M\to G$ be smooth and define
\begin{align*}
F_u:M\times G\to M\times G,\qquad F_u(x,h)=(x,u(x)h).
\end{align*}
This map covers $\operatorname{id}_M$, since
\begin{align*}
\pi(F_u(x,h))=\pi(x,u(x)h)=x=\operatorname{id}_M(\pi(x,h)).
\end{align*}
It also commutes with the right $G$-action. For $g\in G$,
\begin{align*}
F_u((x,h)\cdot g)=F_u(x,hg).
\end{align*}
By the definition of $F_u$,
\begin{align*}
F_u(x,hg)=(x,u(x)(hg)).
\end{align*}
Associativity in $G$ gives
\begin{align*}
(x,u(x)(hg))=(x,(u(x)h)g).
\end{align*}
Using the right action on $M\times G$ again,
\begin{align*}
(x,(u(x)h)g)=(x,u(x)h)\cdot g=F_u(x,h)\cdot g.
\end{align*}
Thus $F_u((x,h)\cdot g)=F_u(x,h)\cdot g$.
The inverse map is
\begin{align*}
F_u^{-1}(x,h)=(x,u(x)^{-1}h).
\end{align*}
Indeed,
\begin{align*}
F_u^{-1}(F_u(x,h))=F_u^{-1}(x,u(x)h).
\end{align*}
Substituting the formula for $F_u^{-1}$ gives
\begin{align*}
F_u^{-1}(x,u(x)h)=(x,u(x)^{-1}(u(x)h)).
\end{align*}
By associativity,
\begin{align*}
(x,u(x)^{-1}(u(x)h))=(x,(u(x)^{-1}u(x))h).
\end{align*}
Since $u(x)^{-1}u(x)=e$ and $eh=h$,
\begin{align*}
(x,(u(x)^{-1}u(x))h)=(x,eh)=(x,h).
\end{align*}
Similarly,
\begin{align*}
F_u(F_u^{-1}(x,h))=F_u(x,u(x)^{-1}h).
\end{align*}
Substituting the formula for $F_u$ gives
\begin{align*}
F_u(x,u(x)^{-1}h)=(x,u(x)(u(x)^{-1}h)).
\end{align*}
By associativity,
\begin{align*}
(x,u(x)(u(x)^{-1}h))=(x,(u(x)u(x)^{-1})h).
\end{align*}
Since $u(x)u(x)^{-1}=e$ and $eh=h$,
\begin{align*}
(x,(u(x)u(x)^{-1})h)=(x,eh)=(x,h).
\end{align*}
Multiplication and inversion in a Lie group are smooth, so both $F_u$ and $F_u^{-1}$ are smooth. Hence $F_u$ is a bundle automorphism.
Conversely, let $F:M\times G\to M\times G$ be a bundle automorphism over $M$. Since $F$ covers $\operatorname{id}_M$, the first component of $F(x,h)$ is $x$. Therefore for each $x\in M$ there is a unique element $u(x)\in G$ such that
\begin{align*}
F(x,e)=(x,u(x)).
\end{align*}
The map $u$ is smooth because it is the second component of the smooth map $x\mapsto F(x,e)$. For any $h\in G$,
\begin{align*}
F(x,h)=F((x,e)\cdot h).
\end{align*}
Equivariance gives
\begin{align*}
F((x,e)\cdot h)=F(x,e)\cdot h.
\end{align*}
Using $F(x,e)=(x,u(x))$ and the product right action,
\begin{align*}
F(x,e)\cdot h=(x,u(x))\cdot h=(x,u(x)h).
\end{align*}
Thus $F(x,h)=(x,u(x)h)$ for all $(x,h)\in M\times G$. Every automorphism of the product bundle over $M$ is therefore exactly one of the maps $F_u$, and $u:M\to G$ records fibrewise left multiplication in the chosen product trivialization.
[/example]
The formula $F_u(x,h)=(x,u(x)h)$ is an early warning that gauge transformations act by left multiplication in a chosen product chart, even though the principal action is on the right. This distinction becomes important in Chapter 6, where local connection forms transform by an adjoint term plus a Maurer-Cartan term.
Pullbacks answer a different mapping problem: given a bundle over $M$ and a smooth map $f:N\to M$, how should the bundle be transported to $N$? The construction keeps exactly the points of $P$ [lying over](/theorems/2876) the image of each point of $N$.
[definition: Pullback Principal Bundle]
Let $\pi:P\to M$ be a smooth principal $G$-bundle and let $f:N\to M$ be a smooth map. The pullback bundle $f^*P$ is
\begin{align*}
f^*P=\{(y,p)\in N\times P: f(y)=\pi(p)\},
\end{align*}
with projection $\pi_N:f^*P\to N$, $\pi_N(y)=y$ on pairs written as $(y,p)$, and right action
\begin{align*}
f^*P\times G&\to f^*P, & ((y,p),g)&\mapsto (y,p)\cdot g=(y,p\cdot g).
\end{align*}
[/definition]
Because local product charts of $P$ pull back along $f$, the object just defined is again a smooth principal $G$-bundle. The next question is how to recognise this construction without coordinates: the universal property says that any equivariant map into $P$ whose base map factors through $f$ factors uniquely through $f^*P$.
[quotetheorem:6237]
[citeproof:6237]
The universal property is the clean way to recognise pullbacks without choosing local coordinates. The compatibility condition $\pi\circ H=f\circ\rho$ is necessary because otherwise the pair $(\rho(q),H(q))$ need not lie in the fibre product $f^*P$. Equivariance is equally necessary: without it, the resulting map could send a point and its right translate to unrelated points in the pulled-back fibre, so it would not be a principal bundle morphism.
The statement is also deliberately limited. It constructs the unique map into the already-defined pullback, but it does not produce a section, trivialise $P$, or identify pullbacks taken along homotopic maps. What it does provide is the exact formal mechanism used in Chapter 6 for connections: once a connection form lives on $P$, the projection $f^*P\to P$ is the canonical map along which that form can be pulled back to a connection over $N$.
[example: Restricting A Principal Bundle]
Let $i:U\hookrightarrow M$ be the inclusion of an open subset and let $\pi:P\to M$ be a principal $G$-bundle. By definition of the pullback,
\begin{align*}
i^*P=\{(x,p)\in U\times P:i(x)=\pi(p)\}.
\end{align*}
Since $i(x)=x$ for the inclusion map, this is
\begin{align*}
i^*P=\{(x,p)\in U\times P:x=\pi(p)\}.
\end{align*}
Define $\Psi:i^*P\to \pi^{-1}(U)$ by $\Psi(x,p)=p$. This is well-defined because $(x,p)\in i^*P$ implies $\pi(p)=x\in U$, hence $p\in\pi^{-1}(U)$. Define $\Phi:\pi^{-1}(U)\to i^*P$ by
\begin{align*}
\Phi(p)=(\pi(p),p).
\end{align*}
If $p\in\pi^{-1}(U)$, then $\pi(p)\in U$ and $i(\pi(p))=\pi(p)$, so $(\pi(p),p)\in i^*P$.
The two maps are inverse to one another. For $p\in\pi^{-1}(U)$,
\begin{align*}
\Psi(\Phi(p))=\Psi(\pi(p),p)=p.
\end{align*}
For $(x,p)\in i^*P$, membership in $i^*P$ gives $\pi(p)=x$, so
\begin{align*}
\Phi(\Psi(x,p))=\Phi(p)=(\pi(p),p)=(x,p).
\end{align*}
Thus $i^*P$ identifies with $\pi^{-1}(U)$.
Under this identification, the pullback projection becomes the restricted projection. Indeed, if $\pi_U:i^*P\to U$ is given by $\pi_U(x,p)=x$, then for $p\in\pi^{-1}(U)$,
\begin{align*}
\pi_U(\Phi(p))=\pi_U(\pi(p),p)=\pi(p).
\end{align*}
The right action is also the restricted right action. For $(x,p)\in i^*P$ and $g\in G$,
\begin{align*}
(x,p)\cdot g=(x,p\cdot g),
\end{align*}
and this pair lies in $i^*P$ because
\begin{align*}
\pi(p\cdot g)=\pi(p)=x.
\end{align*}
Applying $\Psi$ gives
\begin{align*}
\Psi((x,p)\cdot g)=\Psi(x,p\cdot g)=p\cdot g=\Psi(x,p)\cdot g.
\end{align*}
Therefore $i^*P$ is the restricted principal bundle $\pi^{-1}(U)\to U$ with its original right $G$-action. Under this identification, the canonical map $i^*P\to P$ sends $(x,p)$ to $p$, so it is exactly the inclusion $\pi^{-1}(U)\hookrightarrow P$.
[/example]
This example is the local version of pullback used throughout the course. Every statement about a principal bundle can be checked after restricting to a cover, provided the resulting local statements agree on overlaps.
## The Product Bundle and the Maurer-Cartan Form
The final question of the chapter is what principal geometry looks like before twisting is introduced. On the product bundle, all vertical directions come from motion inside the group $G$, and the Lie algebra $\mathfrak g=T_eG$ records the infinitesimal symmetry of each fibre.
[definition: Fundamental Vector Field]
Let $P\to M$ be a smooth principal $G$-bundle and let $\xi\in\mathfrak g$. The fundamental vector field generated by $\xi$ is the vector field $\xi_P\in\mathfrak X(P)$ defined by
\begin{align*}
(\xi_P)_p=\frac{d}{dt}\bigg|_{t=0}p\cdot \exp(t\xi).
\end{align*}
[/definition]
Fundamental vector fields identify Lie algebra elements with vertical tangent vectors. For a fixed $p\in P$, the map $\xi\mapsto (\xi_P)_p$ is an isomorphism from $\mathfrak g$ to the vertical tangent space $\ker d\pi_p$.
[example: Fundamental Fields On A Product Bundle]
Fix $\xi\in\mathfrak g$ and $(x,h)\in M\times G$. By the definition of the fundamental vector field for the right action, we compute
\begin{align*}
(\xi_{M\times G})_{(x,h)}=\frac{d}{dt}\bigg|_{t=0}(x,h)\cdot \exp(t\xi).
\end{align*}
Using the product-bundle action $(x,h)\cdot g=(x,hg)$, this becomes
\begin{align*}
(\xi_{M\times G})_{(x,h)}=\frac{d}{dt}\bigg|_{t=0}(x,h\exp(t\xi)).
\end{align*}
Under the product tangent identification $T_{(x,h)}(M\times G)\cong T_xM\oplus T_hG$, the derivative of a curve $t\mapsto (x(t),h(t))$ is the pair $(x'(0),h'(0))$. In the present curve, the base component is constant:
\begin{align*}
x(t)=x.
\end{align*}
Hence
\begin{align*}
x'(0)=0\in T_xM.
\end{align*}
The group component is
\begin{align*}
h(t)=h\exp(t\xi).
\end{align*}
Since $t\mapsto h\exp(t\xi)$ is the left translate by $h$ of the curve $t\mapsto \exp(t\xi)$, its derivative at $0$ is
\begin{align*}
h'(0)=(dL_h)_e(\xi)\in T_hG.
\end{align*}
Therefore
\begin{align*}
(\xi_{M\times G})_{(x,h)}=(0,(dL_h)_e(\xi))\in T_xM\oplus T_hG.
\end{align*}
The fundamental field on the product bundle has no base component; it records exactly the infinitesimal motion inside the fibre $\{x\}\times G$.
[/example]
To measure tangent vectors on the group itself, there is a canonical Lie-algebra-valued one-form. Ordinary coordinates on $G$ depend on a chart, and tangent vectors at two different group elements lie in different vector spaces $T_gG$ and $T_hG$. Left translation supplies the missing comparison: it moves every tangent vector back to $T_eG=\mathfrak g$ in a way compatible with the group multiplication.
[definition: Maurer-Cartan Form]
Let $G$ be a Lie group with Lie algebra $\mathfrak g=T_eG$. The left Maurer-Cartan form is the $\mathfrak g$-valued one-form $\theta\in\Omega^1(G;\mathfrak g)$ defined by
\begin{align*}
\theta_g(v)=(dL_{g^{-1}})_g(v)
\end{align*}
for $g\in G$ and $v\in T_gG$, where $L_{g^{-1}}:G\to G$ denotes left multiplication by $g^{-1}$.
[/definition]
The form $\theta$ is canonical: it requires no metric, chart, or representation of $G$. It is the prototype for a connection form because it sends the fundamental vector generated by $\xi$ to $\xi$ on the group model.
[example: Maurer-Cartan Form On A Matrix Group]
Let $G\subset GL(n,\mathbb R)$ be a matrix Lie group, and let $g(t)$ be a smooth curve in $G$ with $g(0)=g$ and $g'(0)=v$. Left multiplication by $g^{-1}$ is the map
\begin{align*}
L_{g^{-1}}(a)=g^{-1}a.
\end{align*}
Applying this map to the curve $g(t)$ gives
\begin{align*}
L_{g^{-1}}(g(t))=g^{-1}g(t).
\end{align*}
Since $g^{-1}$ is constant with respect to $t$, differentiating at $t=0$ gives
\begin{align*}
(dL_{g^{-1}})_g(v)=\frac{d}{dt}\bigg|_{t=0}g^{-1}g(t).
\end{align*}
The derivative of the product with the constant matrix $g^{-1}$ is
\begin{align*}
\frac{d}{dt}\bigg|_{t=0}g^{-1}g(t)=g^{-1}g'(0).
\end{align*}
Using $g'(0)=v$, this becomes
\begin{align*}
(dL_{g^{-1}})_g(v)=g^{-1}v.
\end{align*}
By the definition of the left Maurer-Cartan form,
\begin{align*}
\theta_g(v)=g^{-1}v.
\end{align*}
For $G=U(1)$, write a curve as $g(t)=e^{i\alpha(t)}$, with $\alpha(0)=\alpha$ and $\dot\alpha=\alpha'(0)$. Then
\begin{align*}
g(0)=e^{i\alpha}.
\end{align*}
By the chain rule,
\begin{align*}
g'(0)=i\alpha'(0)e^{i\alpha(0)}.
\end{align*}
Substituting $\alpha'(0)=\dot\alpha$ and $\alpha(0)=\alpha$ gives
\begin{align*}
v=g'(0)=i\dot\alpha e^{i\alpha}.
\end{align*}
Since $g^{-1}=e^{-i\alpha}$, the matrix-group formula gives
\begin{align*}
\theta_g(v)=e^{-i\alpha}\left(i\dot\alpha e^{i\alpha}\right).
\end{align*}
Because $i\dot\alpha$ is a scalar and $e^{-i\alpha}e^{i\alpha}=1$, this is
\begin{align*}
\theta_g(v)=i\dot\alpha.
\end{align*}
Thus on $U(1)$ the Maurer-Cartan form removes the current phase $e^{i\alpha}$ and keeps only the infinitesimal angular velocity $i\dot\alpha\in i\mathbb R$.
[/example]
The matrix formula makes the form computable, but it also raises a structural question: what differential identity records the multiplication law of $G$? The answer is the Maurer-Cartan equation, which packages the noncommutativity of the group into a differential equation for a one-form.
[quotetheorem:6238]
[citeproof:6238]
This equation is the flatness equation for the tautological geometry of a Lie group. The sign is tied to the convention chosen above: for the left Maurer-Cartan form $\theta_g=(dL_{g^{-1}})_g$, the equation is $d\theta+\frac{1}{2}[\theta\wedge\theta]=0$ with the stated bracket convention. For the right Maurer-Cartan form defined using right translation back to the identity, the corresponding formula has the opposite sign.
The theorem is special to the Maurer-Cartan form and does not say that every $\mathfrak g$-valued one-form satisfies this equation. A general form $\omega\in\Omega^1(P;\mathfrak g)$ has curvature expression $d\omega+\frac{1}{2}[\omega\wedge\omega]$, and that expression usually does not vanish. Later, a connection form on a principal bundle will use exactly this curvature formula, so the Maurer-Cartan equation identifies the group itself as the model flat case rather than as the generic situation.
On the product principal bundle $M\times G$, the projection to $G$ pulls back the Maurer-Cartan form to a vertical Lie-algebra-valued form. If $\operatorname{pr}_G:M\times G\to G$ is the second projection, then $\operatorname{pr}_G^*\theta$ reads the infinitesimal group component of a tangent vector. This is not yet a general connection on an arbitrary principal bundle, but it is the local normal form from which connection forms are built.
The local normal form from the previous chapter is the bridge to the more rigid cocycle description used here. With transition functions for principal bundles in hand, we can see exactly how local sections change, how the cocycle condition encodes consistency, and how the bundle is reconstructed from that data.
# 2. Transition Functions Revisited for Principal Bundles
This chapter returns to transition functions, now in the more rigid setting of principal bundles. We assume the earlier construction of vector bundle transition functions, the definition of a smooth principal $G$-bundle, and the fact that each fibre is a right $G$-torsor. The goal is to learn how local sections produce $G$-valued transition functions, why these functions satisfy a nonabelian cocycle identity, and how the same data reconstructs and classifies principal bundles over a fixed trivializing cover. This prepares the language used in Chapters 6 and 9 for connections and gauge transformations, where changing a local section changes the formulas but not the underlying geometry.
## Principal Transition Functions and the Cocycle Identity
The first question is how local trivializations of a principal $G$-bundle talk to each other. Since each fibre is a right $G$-torsor rather than a vector space, the transition function should measure the unique group element carrying one local section to another.
Let $\tau_i: \rho^{-1}(U_i) \to U_i \times G$ be a principal trivialization of a principal $G$-bundle $\rho:P\to M$. It is often cleaner to encode $\tau_i$ by its local section $s_i:U_i\to P$, where $s_i(x)$ is the unique point satisfying $\tau_i(s_i(x))=(x,e)$.
[definition: Principal Transition Function]
Let $\rho:P\to M$ be a smooth principal $G$-bundle with local sections $s_i:U_i\to P$ and $s_j:U_j\to P$. On $U_i\cap U_j$, the principal transition function from $j$ to $i$ is the smooth map
\begin{align*}
g_{ij}:U_i\cap U_j\to G
\end{align*}
defined by
\begin{align*}
s_j(x)=s_i(x)g_{ij}(x).
\end{align*}
[/definition]
The direction of the indices matters. With the convention above, $g_{ij}$ is the group element which changes the $i$-section into the $j$-section by the right action. Different books sometimes reverse this convention, so every formula involving products should be checked against the defining equation.
[example: Product Bundle Transition Functions]
Let $P=M\times G$ with right action $(x,h)k=(x,hk)$. A local section over $U_i$ has the form $s_i(x)=(x,a_i(x))$ for a smooth map $a_i:U_i\to G$, because $\rho(x,h)=x$.
On $U_i\cap U_j$, write $s_i(x)=(x,a_i(x))$ and $s_j(x)=(x,a_j(x))$. The defining equation for the principal transition function is
\begin{align*}
s_j(x)=s_i(x)g_{ij}(x).
\end{align*}
Substituting the two sections gives
\begin{align*}
(x,a_j(x))=(x,a_i(x))g_{ij}(x).
\end{align*}
Using the product right action $(x,h)k=(x,hk)$, the right-hand side is
\begin{align*}
(x,a_i(x))g_{ij}(x)=(x,a_i(x)g_{ij}(x)).
\end{align*}
Hence
\begin{align*}
(x,a_j(x))=(x,a_i(x)g_{ij}(x)).
\end{align*}
Equality in $M\times G$ means equality of second coordinates, so
\begin{align*}
a_j(x)=a_i(x)g_{ij}(x).
\end{align*}
Multiplying this equation on the left by $a_i(x)^{-1}$ gives
\begin{align*}
a_i(x)^{-1}a_j(x)=a_i(x)^{-1}\bigl(a_i(x)g_{ij}(x)\bigr).
\end{align*}
By associativity in $G$,
\begin{align*}
a_i(x)^{-1}\bigl(a_i(x)g_{ij}(x)\bigr)=\bigl(a_i(x)^{-1}a_i(x)\bigr)g_{ij}(x).
\end{align*}
Since $a_i(x)^{-1}a_i(x)=e$ and $eg_{ij}(x)=g_{ij}(x)$, we obtain
\begin{align*}
g_{ij}(x)=a_i(x)^{-1}a_j(x).
\end{align*}
Thus even the globally product bundle can have non-identity transition functions: the functions record the chosen local sections, and they are identically $e$ exactly when $a_i(x)=a_j(x)$ on the overlap.
[/example]
The next constraint comes from comparing three sections over a triple overlap. Starting with $s_i$, changing to $s_j$, and then changing to $s_k$ must give the same result as changing directly from $s_i$ to $s_k$.
[quotetheorem:6080]
[citeproof:6080]
This theorem is the principal-bundle version of the familiar vector-bundle cocycle identity. The extra care is noncommutativity: $g_{ij}g_{jk}$ is meaningful in this order because the right action composes on the right. If the order were reversed, the formula would generally assert $g_{jk}g_{ij}=g_{ik}$, which fails for a nonabelian group unless the particular transition values commute. The freeness of the action is also essential: without it, the equation $s_i(x)a=s_i(x)b$ would not force $a=b$, so local comparisons would not determine a unique group-valued function. The theorem does not say that every family of maps is allowed; it singles out precisely the compatibility condition that will make gluing local products transitive in the reconstruction theorem.
[example: Principal $U(1)$-Bundles Over The Circle]
Let $S^1=U_0\cup U_1$, where $U_0$ and $U_1$ are connected arcs and $U_0\cap U_1=V_+\sqcup V_-$ has two connected interval components. Choose local sections $s_0$ and $s_1$, with transition function
\begin{align*}
g_{01}:U_0\cap U_1\to U(1).
\end{align*}
On each interval component $V_\pm$, the map $g_{01}|_{V_\pm}$ has no winding obstruction, because an interval is contractible. Thus, after changing the local section over $U_1$ on the two overlap components and extending smoothly across the connected arc $U_1$, we may arrange that the new transition function is constant on each component:
\begin{align*}
g'_{01}|_{V_+}=c_+,\qquad g'_{01}|_{V_-}=c_-
\end{align*}
for some $c_+,c_-\in U(1)$.
The remaining clutching data is therefore the pair $(c_+,c_-)$, equivalently a map $S^0\to U(1)$. Since $U(1)$ is connected, choose a smooth path $\gamma:[0,1]\to U(1)$ with
\begin{align*}
\gamma(0)=c_+,\qquad \gamma(1)=c_-.
\end{align*}
Using a coordinate on the connected arc $U_0$, extend this path to a smooth map $a_0:U_0\to U(1)$ whose values on the two overlap components are $c_+$ and $c_-$. Keep the section over $U_1$ fixed, so $a_1=1$. By the change-of-trivialization formula, the transformed transition function is
\begin{align*}
g''_{01}(x)=a_0(x)^{-1}g'_{01}(x)a_1(x)=a_0(x)^{-1}g'_{01}(x).
\end{align*}
On $V_+$, this gives
\begin{align*}
g''_{01}(x)=c_+^{-1}c_+=1.
\end{align*}
On $V_-$, this gives
\begin{align*}
g''_{01}(x)=c_-^{-1}c_-=1.
\end{align*}
Thus every smooth principal $U(1)$-bundle over $S^1$ has transition functions coboundary equivalent to the identity cocycle, so it is isomorphic to the product bundle. Nonconstant transition functions can still appear, but only from the chosen local sections, not from a nontrivial bundle over the circle.
[/example]
## Change of Trivialization and Coboundary Equivalence
The next problem is to decide when two sets of transition functions describe the same principal bundle. Local sections are choices, not intrinsic structure, so changing them should modify $g_{ij}$ without changing the isomorphism class of $P$.
Suppose the same bundle is described by another family of local sections $s'_i:U_i\to P$. Since $s_i(x)$ and $s'_i(x)$ lie in the same fibre, there is a unique smooth map $a_i:U_i\to G$ with
\begin{align*}
s'_i(x)=s_i(x)a_i(x).
\end{align*}
The resulting transition functions are related by a nonabelian analogue of adding a coboundary.
[quotetheorem:6239]
[citeproof:6239]
The formula has conjugation on both sides because $G$ need not be abelian. The hypotheses that the two systems of sections live on the same bundle and the same cover are part of the statement, not cosmetic assumptions: otherwise the pointwise comparison $s'_i(x)=s_i(x)a_i(x)$ may not even be defined on a common domain or inside a common fibre. For a concrete failure, take the product principal $U(1)$-bundle over $S^2$ and the Hopf principal $U(1)$-bundle over $S^2$, both described over northern and southern hemispherical neighbourhoods. Each has local sections on those two sets, but a section of the Hopf bundle and a section of the product bundle are points in different total spaces, so the equation $s'_i=s_i a_i$ has no meaning until an identification of the two bundles has already been supplied. The theorem therefore describes the effect of changing local choices inside a fixed bundle with a fixed cover; it does not by itself prove that all cocycles related by this formula give isomorphic reconstructed bundles. That converse is supplied later by the classification theorem.
[definition: Coboundary Equivalent Principal Cocycles]
Let $(U_i)_{i\in I}$ be an open cover of $M$. Two families of smooth maps $g_{ij}:U_i\cap U_j\to G$ and $g'_{ij}:U_i\cap U_j\to G$ satisfying the principal cocycle identity are coboundary equivalent if there exist smooth maps $a_i:U_i\to G$ such that
\begin{align*}
g'_{ij}=a_i^{-1}g_{ij}a_j
\end{align*}
on every $U_i\cap U_j$.
[/definition]
This definition packages exactly the freedom of changing local sections. The word "equivalent" is doing more than algebraic bookkeeping: it anticipates the classification theorem, where equivalent cocycles produce isomorphic principal bundles.
[example: Gauge Change on a Product Bundle]
For $P=M\times G$ with right action $(x,h)k=(x,hk)$, take the standard section $s_i(x)=(x,e)$ on every open set. Its transition function satisfies
\begin{align*}
s_j(x)=s_i(x)g_{ij}(x).
\end{align*}
Since both standard sections are $(x,e)$, this becomes
\begin{align*}
(x,e)=(x,e)g_{ij}(x).
\end{align*}
Using the product right action,
\begin{align*}
(x,e)g_{ij}(x)=(x,eg_{ij}(x))=(x,g_{ij}(x)).
\end{align*}
Equality of second coordinates gives $g_{ij}(x)=e$.
Now replace the standard sections by $s'_i(x)=(x,a_i(x))$ and $s'_j(x)=(x,a_j(x))$, where $a_i:U_i\to G$ and $a_j:U_j\to G$ are smooth. The new transition function is defined by
\begin{align*}
s'_j(x)=s'_i(x)g'_{ij}(x).
\end{align*}
Substituting the sections gives
\begin{align*}
(x,a_j(x))=(x,a_i(x))g'_{ij}(x).
\end{align*}
Using the right action again,
\begin{align*}
(x,a_i(x))g'_{ij}(x)=(x,a_i(x)g'_{ij}(x)).
\end{align*}
Therefore
\begin{align*}
(x,a_j(x))=(x,a_i(x)g'_{ij}(x)).
\end{align*}
Equality of second coordinates gives
\begin{align*}
a_j(x)=a_i(x)g'_{ij}(x).
\end{align*}
Multiplying on the left by $a_i(x)^{-1}$ gives
\begin{align*}
a_i(x)^{-1}a_j(x)=a_i(x)^{-1}\bigl(a_i(x)g'_{ij}(x)\bigr).
\end{align*}
By associativity in $G$,
\begin{align*}
a_i(x)^{-1}\bigl(a_i(x)g'_{ij}(x)\bigr)=\bigl(a_i(x)^{-1}a_i(x)\bigr)g'_{ij}(x).
\end{align*}
Since $a_i(x)^{-1}a_i(x)=e$,
\begin{align*}
\bigl(a_i(x)^{-1}a_i(x)\bigr)g'_{ij}(x)=eg'_{ij}(x).
\end{align*}
Since $e$ is the identity element of $G$,
\begin{align*}
eg'_{ij}(x)=g'_{ij}(x).
\end{align*}
Thus
\begin{align*}
g'_{ij}(x)=a_i(x)^{-1}a_j(x).
\end{align*}
So the same globally split bundle may be represented by a nonconstant cocycle, but this cocycle comes entirely from the change of local sections and is a coboundary.
[/example]
## Reconstruction from a G-Valued Cocycle
The final question reverses the construction. If we are handed maps $g_{ij}:U_i\cap U_j\to G$ satisfying the cocycle identity, can we build a principal bundle whose transition functions are exactly these maps?
The construction starts with the disjoint union of local models $U_i\times G$ and glues them over overlaps. The transition function must identify the point represented in the $j$-chart with the corresponding point in the $i$-chart. Since $s_j=s_i g_{ij}$, the element $h$ in the $j$-chart corresponds to $g_{ij}(x)h$ in the $i$-chart.
[quotetheorem:6240]
[citeproof:6240]
This theorem explains why transition functions are not merely invariants extracted from a bundle. They are construction data: local products plus compatible gluing maps are enough to recover the global principal bundle. The cocycle identity is exactly what makes the quotient relation transitive: without $g_{ij}g_{jk}=g_{ik}$, a point could be glued from the $k$-chart to the $j$-chart and then to the $i$-chart in a way that disagrees with direct gluing from $k$ to $i$. Smoothness is also essential, not only a technical preference. For example, let $M=\mathbb R$, let $G=\mathbb R_{>0}$ under multiplication, and cover $M$ by $U_- = (-\infty,1)$ and $U_+ = (-1,\infty)$. The continuous transition function $g_{-+}(x)=1+|x|$ on $U_-\cap U_+$ glues two local products into a topological principal $G$-bundle, but the overlap coordinate change $(x,h)\mapsto (x,(1+|x|)h)$ is not smooth at $x=0$. Hence this gluing data does not define a smooth principal bundle structure with those charts. Thus the theorem reconstructs smooth principal bundles only from smooth cocycles satisfying the full compatibility conditions.
[example: Principal $U(1)$-Bundles Over The Sphere]
Cover $S^2$ by northern and southern hemispherical neighbourhoods $U_N,U_S$, and identify their overlap with an equatorial annulus carrying angular coordinate $e^{i\theta}\in S^1$. For each $n\in\mathbb Z$, set
\begin{align*}
g_{NS}(e^{i\theta})=e^{in\theta}
\end{align*}
and define
\begin{align*}
g_{SN}(e^{i\theta})=g_{NS}(e^{i\theta})^{-1}=e^{-in\theta}.
\end{align*}
Also set $g_{NN}=g_{SS}=1$. The two nontrivial cocycle checks are
\begin{align*}
g_{NS}(e^{i\theta})g_{SN}(e^{i\theta})=e^{in\theta}e^{-in\theta}=e^0=1=g_{NN}(e^{i\theta})
\end{align*}
and
\begin{align*}
g_{SN}(e^{i\theta})g_{NS}(e^{i\theta})=e^{-in\theta}e^{in\theta}=e^0=1=g_{SS}(e^{i\theta}).
\end{align*}
Thus the reconstruction theorem glues $U_N\times U(1)$ to $U_S\times U(1)$ by
\begin{align*}
(e^{i\theta},h)_S\sim (e^{i\theta},e^{in\theta}h)_N.
\end{align*}
The clutching integer is the [winding number](/page/Winding%20Number) of $g_{NS}:S^1\to U(1)$. Writing $g_{NS}(\theta)=e^{in\theta}$, we compute
\begin{align*}
\frac{d}{d\theta}g_{NS}(\theta)=in e^{in\theta}.
\end{align*}
Hence
\begin{align*}
g_{NS}(\theta)^{-1}\frac{d}{d\theta}g_{NS}(\theta)=e^{-in\theta}in e^{in\theta}=in.
\end{align*}
Therefore
\begin{align*}
\frac{1}{2\pi i}\int_0^{2\pi}g_{NS}(\theta)^{-1}\frac{d}{d\theta}g_{NS}(\theta)\,d\theta=\frac{1}{2\pi i}\int_0^{2\pi}in\,d\theta=\frac{2\pi in}{2\pi i}=n.
\end{align*}
With the standard clutching classification for principal $U(1)$-bundles over $S^2$, this winding number is the Chern number of the reconstructed bundle. When $n=0$, the gluing function is identically $1$, so the two local products glue to the product bundle; when $n=1$, the clutching map is $e^{i\theta}\mapsto e^{i\theta}$, the standard clutching function of the Hopf fibration $S^3\to S^2$.
[/example]
The reconstruction theorem also identifies the correct classification object. Because transition functions are changed by local maps $a_i:U_i\to G$, the remaining problem is to pass from individual reconstructed bundles to a precise bijection between isomorphism classes and cocycles modulo this change of section.
[quotetheorem:6241]
[citeproof:6241]
This classification is usually denoted informally by $\check{H}^1(M,G)$, but for nonabelian $G$ it is a pointed set rather than a group. The base point is the class of the identity cocycle, corresponding to the product principal bundle. The fixed-cover hypothesis is necessary: the displayed set only sees bundles equipped with trivializations over the chosen open cover. If a principal bundle does not split over that cover, it has no representative in this particular cocycle set, even though it may be represented after passing to a refinement. The theorem also does not turn nonabelian Cech classes into a group of bundles; it gives a pointed classification set, with no natural multiplication of isomorphism classes in general.
[example: Frame Bundle Transition Functions of The Two Sphere]
Let $U_N=S^2\setminus\{S\}$ and $U_S=S^2\setminus\{N\}$ be the usual oriented stereographic charts, with complex coordinates $z=x+iy$ on $U_N$ and $w=p+iq$ on $U_S$. On the overlap, use the oriented coordinate change
\begin{align*}
w=\frac{1}{z}.
\end{align*}
If $z=re^{i\theta}$, then
\begin{align*}
w=r^{-1}e^{-i\theta}.
\end{align*}
The differential of the coordinate change is complex multiplication by
\begin{align*}
\frac{dw}{dz}=-\frac{1}{z^2}.
\end{align*}
Substituting $z=re^{i\theta}$ gives
\begin{align*}
-\frac{1}{z^2}=-\frac{1}{r^2e^{2i\theta}}=r^{-2}e^{i(\pi-2\theta)}.
\end{align*}
In stereographic coordinates the round metric is conformal to the Euclidean metric, so the local oriented orthonormal frame is obtained from the coordinate frame by multiplying both coordinate vectors by the same positive scalar. The factor $r^{-2}$ in the differential changes lengths, while the unit complex number
\begin{align*}
e^{i(\pi-2\theta)}
\end{align*}
is the rotational part. Therefore the transition function of the [oriented orthonormal frame bundle](/theorems/6245) along the equatorial overlap is
\begin{align*}
g_{NS}(\theta)=e^{i(\pi-2\theta)}.
\end{align*}
As $\theta$ increases from $0$ to $2\pi$, the angle $\pi-2\theta$ changes from $\pi$ to $\pi-4\pi$, so the total change is
\begin{align*}
(\pi-4\pi)-\pi=-4\pi.
\end{align*}
Thus the transition map winds twice in $SO(2)\cong U(1)$, with sign depending on the chosen orientation convention. This winding number is the frame-bundle manifestation of the Euler number $2$ of $TS^2$.
[/example]
The chapter's lesson is that principal bundles are locally simple and globally encoded by group-valued gluing. The cocycle identity records compatibility on triple overlaps, coboundaries record changes of local section, and reconstruction turns the cocycle into the bundle itself. These three operations form the technical foundation for the frame-bundle constructions of Chapter 3, the associated-bundle construction of Chapter 4, and the connection and gauge formulas of Chapters 6 and 9.
Once transition functions are understood for principal bundles, the frame bundle becomes the most concrete example of the theory. This chapter shows how bases, coordinates, and structure groups fit into a single principal-bundle picture, preparing the way for associated bundles and reductions.
# 3. Frame Bundles and Structure Groups
Principal bundles become concrete when their fibres are interpreted as choices of coordinates, bases, or gauges. The prerequisites for this chapter are smooth vector bundles, local trivialisations and transition functions, and the principal bundles with right actions and transition functions developed in Chapters 1 and 2. We now explain how the transition functions of a vector bundle are already the transition functions of a canonical principal bundle, the frame bundle. This viewpoint lets us translate geometric extra structure, such as metrics, orientations, and Hermitian forms, into reductions of structure group.
## Linear Frames of a Vector Bundle
Given a rank-$r$ vector bundle $E \to M$, the first problem is to package all possible bases of the fibres $E_p$ into a single smooth object. A local vector bundle chart identifies each fibre over an open set with $\mathbb R^r$, so a basis of $E_p$ is the same as an isomorphism $\mathbb R^r \to E_p$. Changing a basis on the right by an invertible matrix is the symmetry that will produce a principal bundle.
[definition: Linear Frame]
Let $\pi:E\to M$ be a smooth real vector bundle of rank $r$. A linear frame of $E$ at $p\in M$ is a linear isomorphism
\begin{align*}
\nu: \mathbb R^r \longrightarrow E_p.
\end{align*}
The frame bundle of $E$ is the set
\begin{align*}
\operatorname{Fr}(E) := \bigsqcup_{p\in M} \operatorname{Iso}(\mathbb R^r,E_p),
\end{align*}
with projection $q:\operatorname{Fr}(E)\to M$ sending $\nu:\mathbb R^r\to E_p$ to $p$.
[/definition]
A frame may also be written as an ordered basis $(e_1,\dots,e_r)$ of $E_p$ by setting $e_i=\nu(e_i^{\mathrm{std}})$, where $(e_1^{\mathrm{std}},\dots,e_r^{\mathrm{std}})$ is the standard basis of $\mathbb R^r$. To turn the set of frames into a principal bundle, we next need the group action that compares two bases of the same vector space.
[definition: Right Action on Frames]
Let $E\to M$ be a real rank-$r$ vector bundle. The group $GL_r(\mathbb R)$ acts on $\operatorname{Fr}(E)$ on the right by
\begin{align*}
\nu \cdot A := \nu \circ A, \qquad \nu\in \operatorname{Iso}(\mathbb R^r,E_p),\ A\in GL_r(\mathbb R).
\end{align*}
[/definition]
This action changes the coordinates of the frame while keeping the base point fixed. In basis language, if the columns of $A=(A_{ij})$ are used to form new vectors, then
\begin{align*}
(\nu\cdot A)(e_j^{\mathrm{std}})=\sum_{i=1}^r A_{ij}\nu(e_i^{\mathrm{std}}).
\end{align*}
The next result checks that these fibrewise actions fit together smoothly over the base manifold.
[quotetheorem:6242]
[citeproof:6242]
This theorem turns vector bundle transition matrices into the transition functions of a principal bundle. The rank hypothesis is essential because the single group $GL_r(\mathbb R)$ only acts on fibres of one fixed dimension; a family of vector spaces with varying dimension cannot have one frame fibre modelled on one Lie group. Smoothness of the vector bundle is also doing work: if local trivialisations were only continuous, the same set of frames would at best be a topological principal bundle, not a smooth one on which connection forms can later live. The theorem does not choose a frame globally; it only packages all possible local frame choices into a principal bundle whose local sections are local frames.
If $\Phi_i:E|_{U_i}\to U_i\times \mathbb R^r$ are local vector bundle charts with vector bundle transition functions $g_{ij}:U_i\cap U_j\to GL_r(\mathbb R)$, then the frame bundle has the same transition functions up to the convention forced by right actions. The tangent bundle is the next example because its transition functions are the Jacobian matrices of coordinate changes, and it shows how the frame bundle detects whether coordinate vector fields glue globally.
[example: Tangent Frame Bundle]
For a smooth $n$-manifold $M$, a point of $\operatorname{Fr}(TM)$ over $p$ is a linear isomorphism $\nu:\mathbb R^n\to T_pM$. In coordinates $(U,\varphi)$ with $\varphi=(x_1,\dots,x_n)$, compare $\nu$ with the coordinate basis $(\partial_{x_1}|_p,\dots,\partial_{x_n}|_p)$. This gives the unique matrix $A=(a_{ij})\in GL_n(\mathbb R)$ whose $j$th column records the coordinates of the $j$th frame vector:
\begin{align*}
\nu(e_j^{\mathrm{std}})=\sum_{i=1}^n a_{ij}\,\partial_{x_i}|_p,\qquad 1\le j\le n.
\end{align*}
The matrix $A$ is invertible because $\nu$ sends the standard basis of $\mathbb R^n$ to a basis of $T_pM$.
On an overlap with another coordinate chart $(V,\psi)$, write $\psi=(y_1,\dots,y_n)$. By the chain rule, each $x$-coordinate basis vector is expressed in the $y$-coordinate basis as
\begin{align*}
\partial_{x_i}|_p=\sum_{k=1}^n \frac{\partial y_k}{\partial x_i}(p)\,\partial_{y_k}|_p.
\end{align*}
Substituting this into the expression for the same frame vector gives
\begin{align*}
\nu(e_j^{\mathrm{std}})=\sum_{i=1}^n a_{ij}\sum_{k=1}^n \frac{\partial y_k}{\partial x_i}(p)\,\partial_{y_k}|_p.
\end{align*}
Reordering the finite sums gives
\begin{align*}
\nu(e_j^{\mathrm{std}})=\sum_{k=1}^n\left(\sum_{i=1}^n \frac{\partial y_k}{\partial x_i}(p)a_{ij}\right)\partial_{y_k}|_p.
\end{align*}
Thus, if $A_x$ is the matrix of the frame in the $x$-basis and $A_y$ is the matrix in the $y$-basis, then
\begin{align*}
A_y=\left(\frac{\partial y_k}{\partial x_i}(p)\right)_{k,i}A_x.
\end{align*}
So the transition matrix for the tangent frame bundle is the Jacobian matrix of the coordinate change, with the convention determined by expressing the same tangent vectors in the new coordinate basis.
For $M=S^1$, the vector field
\begin{align*}
X_{(\cos\theta,\sin\theta)}=(-\sin\theta,\cos\theta)
\end{align*}
is tangent to $S^1$ because its dot product with the radius vector is
\begin{align*}
(\cos\theta,\sin\theta)\cdot(-\sin\theta,\cos\theta)=-\cos\theta\sin\theta+\sin\theta\cos\theta=0,
\end{align*}
and it is nonzero because
\begin{align*}
\|X_{(\cos\theta,\sin\theta)}\|^2=\sin^2\theta+\cos^2\theta=1.
\end{align*}
Hence $s(p)(t)=tX_p$ defines a global section $s:S^1\to \operatorname{Fr}(TS^1)$. Since $GL_1(\mathbb R)=\mathbb R^\times$, every frame of $T_pS^1$ is uniquely of the form $s(p)\cdot a$ for some $a\in \mathbb R^\times$, so $(p,a)\mapsto s(p)\cdot a$ identifies $\operatorname{Fr}(TS^1)$ with the product principal bundle $S^1\times GL_1(\mathbb R)$. For $M=S^2$, a global tangent frame would give a nowhere-zero tangent vector field by taking $p\mapsto \nu_p(e_1^{\mathrm{std}})$, contradicting the *[Hairy Ball Theorem](/theorems/2248)*. Thus tangent frame bundles record whether tangent spaces can be smoothly based all at once, not merely how coordinate bases look locally.
[/example]
The tangent example is the model for much of differential geometry: choosing a local frame is choosing a local section of $\operatorname{Fr}(TM)$. A connection on a principal frame bundle will later become a rule for differentiating tensor fields in a frame-dependent way that transforms correctly under changes of frame.
## Vector Bundles from Principal Linear Bundles
The frame bundle construction should not lose information. The next question is whether every principal $GL_r(\mathbb R)$-bundle arises as the frame bundle of a vector bundle, and whether the original vector bundle can be recovered from its frames. The answer uses the associated bundle construction with the standard representation of $GL_r(\mathbb R)$ on $\mathbb R^r$.
[definition: Standard Associated Vector Bundle]
Let $P\to M$ be a principal $GL_r(\mathbb R)$-bundle. The associated rank-$r$ vector bundle for the standard representation is
\begin{align*}
P\times_{GL_r(\mathbb R)}\mathbb R^r := (P\times \mathbb R^r)/\sim,
\end{align*}
where
\begin{align*}
(u\cdot A,v)\sim (u,Av)
\end{align*}
for $u\in P$, $A\in GL_r(\mathbb R)$, and $v\in \mathbb R^r$.
[/definition]
The quotient relation says that changing the frame by $A$ must also change the coordinate vector by $A$ so that the represented vector in the fibre does not change. This construction leads to the main equivalence: frames and standard associated vector bundles recover one another.
[quotetheorem:6243]
[citeproof:6243]
This result is the bridge between vector-bundle language and principal-bundle language. The use of the standard representation is essential: a different representation of $GL_r(\mathbb R)$ would produce a different associated bundle, such as a tensor, dual, or determinant bundle rather than $E$ itself. The theorem also depends on working with principal $GL_r(\mathbb R)$-bundles; a principal $O(r)$-bundle by itself recovers a metric vector bundle only after extending the structure group back to $GL_r(\mathbb R)$. It does not say that a vector bundle has a preferred frame, only that the totality of all frames contains exactly the same gluing information as the vector bundle.
In Chapters 6 and 7, connections will be defined on principal bundles because symmetry is cleaner there, while covariant derivatives will be recovered on associated vector bundles. The product principal bundle is a useful test case for the quotient relation and for the side on which matrices must act.
[example: Associated Bundle of a Product Principal Bundle]
Let $P=M\times GL_r(\mathbb R)$ be the product principal bundle with right action $(p,A)\cdot B=(p,AB)$. Its standard associated vector bundle is
\begin{align*}
(M\times GL_r(\mathbb R))\times_{GL_r(\mathbb R)}\mathbb R^r,
\end{align*}
with equivalence relation
\begin{align*}
((p,A)\cdot B,v)\sim ((p,A),Bv).
\end{align*}
Define
\begin{align*}
F([(p,A),v])=(p,Av).
\end{align*}
This is independent of the representative because, for $B\in GL_r(\mathbb R)$,
\begin{align*}
F([((p,A)\cdot B),v])=F([(p,AB),v])=(p,(AB)v),
\end{align*}
and associativity of matrix multiplication with the standard vector action gives
\begin{align*}
(p,(AB)v)=(p,A(Bv))=F([(p,A),Bv]).
\end{align*}
Thus the two representatives identified by the associated-bundle relation have the same image under $F$.
The inverse map is
\begin{align*}
G(p,w)=[(p,I),w].
\end{align*}
Then
\begin{align*}
F(G(p,w))=F([(p,I),w])=(p,Iw)=(p,w).
\end{align*}
Conversely, for any class $[(p,A),v]$,
\begin{align*}
G(F([(p,A),v]))=G(p,Av)=[(p,I),Av].
\end{align*}
Since $(p,I)\cdot A=(p,A)$, the relation gives
\begin{align*}
[(p,A),v]=[((p,I)\cdot A),v]=[(p,I),Av].
\end{align*}
Hence $F$ is a vector bundle isomorphism from the associated bundle to $M\times\mathbb R^r$.
If the quotient relation were instead $(u\cdot B,v)\sim (u,B^{-1}v)$ while keeping the same right action on frames, the same formula would fail to descend. Indeed,
\begin{align*}
F([((p,A)\cdot B),v])=(p,ABv),
\end{align*}
whereas
\begin{align*}
F([(p,A),B^{-1}v])=(p,AB^{-1}v).
\end{align*}
Taking $A=I$, $B=2I$, and $v\ne 0$ gives
\begin{align*}
ABv=2v.
\end{align*}
For the other representative,
\begin{align*}
AB^{-1}v=\frac{1}{2}v.
\end{align*}
These are different vectors in $\mathbb R^r$, so equivalent representatives would have different images. The sign check shows that the associated-bundle relation must match the right-frame convention: changing a frame by $B$ changes the coordinate vector by $B$ so that the represented vector $Av$ is unchanged.
[/example]
The non-product case differs only in the gluing data. A principal bundle with transition functions $g_{ij}:U_i\cap U_j\to GL_r(\mathbb R)$ gives a vector bundle whose local vector coordinates are glued by the same matrices acting through the standard representation.
## Orthonormal and Oriented Frames
Many vector bundles carry more than a vector-space structure on each fibre. A Riemannian metric on a real vector bundle singles out bases that are orthonormal, while an orientation singles out bases with positive orientation. The problem is to describe these extra choices without leaving the principal-bundle framework.
[definition: Bundle Metric]
Let $E\to M$ be a smooth real vector bundle of rank $r$. A bundle metric on $E$ is a smooth section
\begin{align*}
h\in \Gamma(\operatorname{Sym}^2(E^*))
\end{align*}
such that, for every $p\in M$, the [bilinear form](/page/Bilinear%20Form)
\begin{align*}
h_p:E_p\times E_p\longrightarrow \mathbb R
\end{align*}
is an [inner product](/page/Inner%20Product).
[/definition]
A general frame of $E_p$ records an arbitrary linear identification $\mathbb R^r\to E_p$, so it forgets the metric information on the fibre. Once $h$ is fixed, the natural problem is to keep only those frames that carry the standard Euclidean inner product on $\mathbb R^r$ to $h_p$. The transition maps between such frames must preserve the Euclidean inner product, so the relevant structure group is reduced from $GL(r,\mathbb R)$ to $O(r)$.
[definition: Orthonormal Frame Bundle]
Let $(E,h)\to M$ be a smooth real rank-$r$ vector bundle with bundle metric. The orthonormal frame bundle is
\begin{align*}
\operatorname{Fr}_O(E,h):=\{\nu\in \operatorname{Fr}(E): h_p(\nu(v),\nu(w))=v\cdot w\text{ for all }v,w\in \mathbb R^r\}.
\end{align*}
It is projected to $M$ by the restriction of $q:\operatorname{Fr}(E)\to M$.
[/definition]
The set just defined is visibly meaningful in each fibre, but a fibrewise subset need not automatically vary smoothly over the base. The missing point is whether local orthonormal frames can be chosen smoothly and whether changing between them uses exactly orthogonal matrices. Positive-definiteness supplies the smooth Gram--Schmidt construction that resolves this issue.
[quotetheorem:6244]
[citeproof:6244]
This theorem says that a metric is equivalent, at the level of frames, to reducing the allowed changes of basis from all invertible matrices to orthogonal matrices. Positive-definiteness is needed because Gram--Schmidt can divide by the length of a nonzero vector; for a degenerate symmetric form, a nonzero vector can have zero length and the construction breaks. Smoothness of $h$ is also needed: a merely pointwise choice of inner products can select orthonormal frames fibre by fibre without producing a smooth subbundle. The theorem does not make arbitrary frames orthonormal; it cuts out exactly those frames that preserve the fixed Euclidean model.
The Riemannian tangent bundle provides the central geometric example, and it also shows why the orthonormal restriction is a genuine loss of freedom rather than notation.
[example: Riemannian Frame Bundle]
Let $(M,g)$ be a Riemannian $n$-manifold. A point of $\operatorname{Fr}_O(TM,g)$ over $p$ is a linear isomorphism $\nu:\mathbb R^n\to T_pM$ satisfying
\begin{align*}
g_p(\nu(v),\nu(w))=v\cdot w
\end{align*}
for all $v,w\in\mathbb R^n$.
Fix a local $g$-orthonormal frame $(e_1,\dots,e_n)$ on an open set $U$, and let $s_p:\mathbb R^n\to T_pM$ be the frame map defined by $s_p(e_i^{\mathrm{std}})=e_i|_p$. Any tangent frame $\nu:\mathbb R^n\to T_pM$ over $p\in U$ is uniquely of the form
\begin{align*}
\nu=s_p\circ A
\end{align*}
for some $A=(a_{i\alpha})\in GL_n(\mathbb R)$, because each vector $\nu(e_\alpha^{\mathrm{std}})$ has unique coordinates in the basis $(e_1|_p,\dots,e_n|_p)$. Explicitly,
\begin{align*}
Ae_\alpha^{\mathrm{std}}=\sum_{i=1}^n a_{i\alpha}e_i^{\mathrm{std}}.
\end{align*}
Applying $s_p$ gives
\begin{align*}
\nu(e_\alpha^{\mathrm{std}})=s_p(Ae_\alpha^{\mathrm{std}})=\sum_{i=1}^n a_{i\alpha}e_i|_p.
\end{align*}
The frame $\nu$ is orthonormal exactly when the Gram matrix of the vectors $\nu(e_1^{\mathrm{std}}),\dots,\nu(e_n^{\mathrm{std}})$ is $I$. Since $(e_1,\dots,e_n)$ is $g$-orthonormal,
\begin{align*}
g_p(e_i|_p,e_j|_p)=\delta_{ij}.
\end{align*}
Therefore
\begin{align*}
g_p(\nu(e_\alpha^{\mathrm{std}}),\nu(e_\beta^{\mathrm{std}}))=g_p\left(\sum_{i=1}^n a_{i\alpha}e_i|_p,\sum_{j=1}^n a_{j\beta}e_j|_p\right).
\end{align*}
By bilinearity of $g_p$,
\begin{align*}
g_p(\nu(e_\alpha^{\mathrm{std}}),\nu(e_\beta^{\mathrm{std}}))=\sum_{i=1}^n\sum_{j=1}^n a_{i\alpha}a_{j\beta}g_p(e_i|_p,e_j|_p).
\end{align*}
Using $g_p(e_i|_p,e_j|_p)=\delta_{ij}$, this becomes
\begin{align*}
g_p(\nu(e_\alpha^{\mathrm{std}}),\nu(e_\beta^{\mathrm{std}}))=\sum_{i=1}^n\sum_{j=1}^n a_{i\alpha}a_{j\beta}\delta_{ij}.
\end{align*}
The Kronecker delta keeps only the terms with $i=j$, so
\begin{align*}
g_p(\nu(e_\alpha^{\mathrm{std}}),\nu(e_\beta^{\mathrm{std}}))=\sum_{i=1}^n a_{i\alpha}a_{i\beta}.
\end{align*}
By the definition of matrix multiplication,
\begin{align*}
(A^\top A)_{\alpha\beta}=\sum_{i=1}^n a_{i\alpha}a_{i\beta}.
\end{align*}
Hence the Gram matrix of $\nu$ is $A^\top A$, so $\nu$ is orthonormal exactly when
\begin{align*}
A^\top A=I.
\end{align*}
This is precisely the condition $A\in O(n)$. Thus all tangent frames are obtained from the chosen frame by matrices in $GL_n(\mathbb R)$, while the orthonormal tangent frames are exactly those obtained by the smaller group $O(n)$, the matrices preserving lengths and angles.
[/example]
Orientation imposes a different restriction: it keeps only frames compatible with the chosen orientation. This leads to a second reduction, now from $GL_r(\mathbb R)$ to the positive-determinant subgroup.
[definition: Oriented Frame Bundle]
Let $E\to M$ be an oriented smooth real vector bundle of rank $r$. The oriented frame bundle is
\begin{align*}
\operatorname{Fr}^+(E):=\{\nu\in \operatorname{Fr}(E): (\nu(e_1^{\mathrm{std}}),\dots,\nu(e_r^{\mathrm{std}}))\text{ is positively oriented}\}.
\end{align*}
It has projection $q:\operatorname{Fr}^+(E)\to M$ given by restricting the projection $q:\operatorname{Fr}(E)\to M$.
[/definition]
This subbundle has structure group $GL_r^+(\mathbb R)$, the subgroup of matrices with positive determinant. If a metric is also present, neither restriction alone is enough for the geometric frame bundle usually used on an oriented Riemannian vector bundle: $\operatorname{Fr}^+(E)$ still allows non-isometric positive changes of basis, while $\operatorname{Fr}_O(E,h)$ still contains orientation-reversing orthonormal frames. The remaining problem is to impose both tests at the same time, and that problem motivates the definition below. The needed subbundle is the common part where orientation and orthonormality are both satisfied.
[definition: Oriented Orthonormal Frame Bundle]
Let $(E,h)\to M$ be an oriented smooth real rank-$r$ vector bundle with bundle metric. The oriented orthonormal frame bundle is
\begin{align*}
\operatorname{Fr}_{SO}(E,h):=\operatorname{Fr}^+(E)\cap \operatorname{Fr}_O(E,h).
\end{align*}
It has projection $q:\operatorname{Fr}_{SO}(E,h)\to M$ given by restricting the projection $q:\operatorname{Fr}(E)\to M$.
[/definition]
The intersection description expresses both constraints at once: the frame must preserve lengths and angles, and it must preserve the chosen orientation. The remaining verification is that the allowed frame changes are exactly the special orthogonal matrices.
[quotetheorem:6245]
[citeproof:6245]
This gives the principal bundle most often used in Riemannian geometry on oriented manifolds. As in the orientation criterion of Chapter 2, the obstruction is global rather than local. The orientation hypothesis cannot be dropped: on a non-orientable vector bundle there is no consistent choice of the positive component of each orthonormal frame fibre, so the $O(r)$-bundle need not contain an $SO(r)$-subbundle. For example, the Mobius line bundle over $S^1$ admits a fibre metric and hence an $O(1)$-frame bundle, but going once around the circle reverses the sign of a local unit frame, so there is no compatible $SO(1)$-reduction. The metric hypothesis is separate; orientation alone reduces $GL_r(\mathbb R)$ to $GL_r^+(\mathbb R)$ but does not restrict changes of frame to isometries. The construction is the natural home for connection forms of the Levi-Civita connection when written in moving frames, and it is also the starting point for spin structures, which ask whether this $SO(r)$-bundle lifts through $\operatorname{Spin}(r)\to SO(r)$.
The following example records the standard tangent-bundle case and shows the obstruction encoded by the determinant sign.
[example: Special Orthogonal Frames on an Oriented Riemannian Manifold]
Let $(M,g)$ be an oriented Riemannian $n$-manifold. The bundle $\operatorname{Fr}_{SO}(TM,g)$ consists of frames $(e_1,\dots,e_n)$ of $T_pM$ satisfying
\begin{align*}
g_p(e_i,e_j)=\delta_{ij}
\end{align*}
and whose ordered basis agrees with the chosen orientation of $T_pM$.
In dimension $2$, fix one positively oriented $g$-orthonormal frame $(e_1,e_2)$ of $T_pM$. If $(f_1,f_2)$ is another frame of $T_pM$, then there is a unique matrix
\begin{align*}
A=\begin{pmatrix}a&b\cr c&d\end{pmatrix}\in GL_2(\mathbb R)
\end{align*}
such that
\begin{align*}
f_1=ae_1+ce_2
\end{align*}
and
\begin{align*}
f_2=be_1+de_2.
\end{align*}
Using bilinearity of $g_p$ and the identities $g_p(e_1,e_1)=1$, $g_p(e_2,e_2)=1$, and $g_p(e_1,e_2)=g_p(e_2,e_1)=0$, we get
\begin{align*}
g_p(f_1,f_1)=g_p(ae_1+ce_2,ae_1+ce_2)=a^2+c^2,
\end{align*}
\begin{align*}
g_p(f_2,f_2)=g_p(be_1+de_2,be_1+de_2)=b^2+d^2,
\end{align*}
and
\begin{align*}
g_p(f_1,f_2)=g_p(ae_1+ce_2,be_1+de_2)=ab+cd.
\end{align*}
Thus $(f_1,f_2)$ is orthonormal exactly when
\begin{align*}
a^2+c^2=1,
\end{align*}
\begin{align*}
b^2+d^2=1,
\end{align*}
and
\begin{align*}
ab+cd=0.
\end{align*}
These three equations are the entries of the matrix equation
\begin{align*}
A^\top A=\begin{pmatrix}a&c\cr b&d\end{pmatrix}\begin{pmatrix}a&b\cr c&d\end{pmatrix}=\begin{pmatrix}a^2+c^2&ab+cd\cr ab+cd&b^2+d^2\end{pmatrix}=I.
\end{align*}
So changing one orthonormal frame to another is described by a matrix in $O(2)$.
The orientation condition selects the component with positive determinant. Since
\begin{align*}
\det A=ad-bc,
\end{align*}
the frame $(f_1,f_2)$ is positively oriented exactly when $\det A>0$. For an orthogonal $2\times 2$ matrix, $A^\top A=I$, so taking determinants gives
\begin{align*}
\det(A^\top A)=\det I.
\end{align*}
Using multiplicativity of determinant and $\det(A^\top)=\det A$, this becomes
\begin{align*}
(\det A)^2=1.
\end{align*}
Hence $\det A=\pm 1$, and the positively oriented orthonormal changes of frame are exactly those with $\det A=1$, namely the matrices in $SO(2)$.
If no orientation is chosen, matrices with determinant $-1$ are also allowed. For example,
\begin{align*}
R=\begin{pmatrix}1&0\cr 0&-1\end{pmatrix}
\end{align*}
satisfies
\begin{align*}
R^\top R=\begin{pmatrix}1&0\cr 0&-1\end{pmatrix}\begin{pmatrix}1&0\cr 0&-1\end{pmatrix}=\begin{pmatrix}1&0\cr 0&1\end{pmatrix},
\end{align*}
while
\begin{align*}
\det R=-1.
\end{align*}
Thus $R$ preserves lengths and angles but reverses orientation. The metric condition reduces frame changes to $O(2)$, and the orientation condition removes exactly the determinant-$-1$ component, leaving $SO(2)$.
[/example]
## Complex and Unitary Frames
A complex vector bundle can be regarded as a real vector bundle of twice the rank, but ordinary real frames lose the operation of multiplication by $i$. The problem is to restrict frames so that they are complex-linear, not merely real-linear. The relevant coordinate space is $\mathbb C^r$, the full complex structure group is $GL_r(\mathbb C)$, and a Hermitian metric reduces the allowed changes of orthonormal complex frame further to $U(r)$.
[definition: Complex Frame Bundle]
Let $E\to M$ be a smooth complex vector bundle of complex rank $r$. A complex frame at $p\in M$ is a complex-linear isomorphism $\nu:\mathbb C^r\to E_p$. The complex frame bundle is
\begin{align*}
\operatorname{Fr}_{\mathbb C}(E):=\bigsqcup_{p\in M}\operatorname{Iso}_{\mathbb C}(\mathbb C^r,E_p).
\end{align*}
It has projection $q:\operatorname{Fr}_{\mathbb C}(E)\to M$ sending $\nu:\mathbb C^r\to E_p$ to $p$.
[/definition]
The right action is again composition, now by $GL_r(\mathbb C)$. This is not merely a real frame bundle of rank $2r$: complex-linearity restricts the allowed frames and remembers the complex structure. The point that still has to be checked is whether these restricted frames vary locally like a product with $GL_r(\mathbb C)$, rather than just forming a fibrewise subset of the real frame bundle. Local complex trivializations provide such complex frames, and changes between them are exactly complex invertible matrices.
[quotetheorem:6246]
[citeproof:6246]
Complex frames remember complex-linearity but not lengths. The complex-rank hypothesis is essential because the structure group $GL_r(\mathbb C)$ acts on a fixed complex vector space; a real rank-$2r$ bundle without a chosen complex structure has no preferred notion of complex-linear frame. For example, the tangent bundle of $S^2$ has real rank $2$ and admits many real tangent frames locally, but treating those frames as complex frames requires a chosen almost-complex structure on the tangent fibres. The theorem also does not identify $\operatorname{Fr}_{\mathbb C}(E)$ with the full real frame bundle of the underlying real bundle: it is a smaller subbundle consisting of frames that commute with multiplication by $i$. Hermitian geometry is obtained by imposing a compatibility condition on these complex frames: the frame must also preserve a positive-definite Hermitian inner product. To reduce further to unitary frame changes, we next need the fibrewise inner product compatible with the complex structure.
[definition: Hermitian Metric]
Let $E\to M$ be a smooth complex vector bundle. A Hermitian metric on $E$ is a smooth section
\begin{align*}
h\in \Gamma(E^*\otimes \overline{E}^*)
\end{align*}
such that, for every $p\in M$, the map
\begin{align*}
h_p:E_p\times E_p\longrightarrow \mathbb C
\end{align*}
is a positive-definite Hermitian inner product, linear in the first argument and conjugate-linear in the second.
[/definition]
With a Hermitian metric fixed, arbitrary complex frames still allow changes of basis that distort lengths and angles. The remaining problem is to isolate the complex frames that preserve the metric, and this motivates the definition below. Such a frame must identify the given Hermitian inner product on $E_p$ with the standard Hermitian product on $\mathbb C^r$. This condition removes the non-unitary complex changes of basis and leaves precisely the frames adapted to the metric.
[definition: Unitary Frame Bundle]
Let $(E,h)\to M$ be a smooth complex vector bundle of complex rank $r$ with Hermitian metric. The unitary frame bundle is
\begin{align*}
\operatorname{Fr}_U(E,h):=\{\nu\in \operatorname{Fr}_{\mathbb C}(E): h_p(\nu(v),\nu(w))=(v,w)_{\mathbb C^r}\text{ for all }v,w\in \mathbb C^r\}.
\end{align*}
It has projection $q:\operatorname{Fr}_U(E,h)\to M$ given by restricting the projection $q:\operatorname{Fr}_{\mathbb C}(E)\to M$.
[/definition]
The standard Hermitian product on $\mathbb C^r$ is linear in the first argument in these notes. After imposing the unitary condition fibre by fibre, the remaining question is whether this subset really carries the expected principal-bundle structure. The theorem below answers that question by verifying that local smooth unitary frames exist and that the transition from one such frame to another preserves the standard Hermitian product. Those transition maps are exactly the unitary matrices.
[quotetheorem:6247]
[citeproof:6247]
The reduction theorem identifies the principal bundle associated with a Hermitian vector bundle. Positive-definiteness is again essential: for an indefinite Hermitian form the symmetry group would be a pseudo-unitary group rather than $U(r)$, and for a degenerate form there may be no unitary bases at all. Smoothness ensures that the Hermitian Gram--Schmidt process produces local smooth unitary frames, not merely pointwise bases. The theorem does not make the bundle a product; it only replaces arbitrary complex frame changes by unitary ones wherever local frames are chosen.
The basic example is the standard Hermitian product bundle, where the reduction admits a global section, but the same description also shows how a non-product Hermitian bundle can fail to have one.
[example: Unitary Frames of a Hermitian Vector Bundle]
Let $(E,h)\to M$ be a Hermitian complex vector bundle of complex rank $r$. A point of $\operatorname{Fr}_U(E,h)$ over $p$ is a complex-linear isomorphism $\nu:\mathbb C^r\to E_p$ such that
\begin{align*}
h_p(\nu(v),\nu(w))=(v,w)_{\mathbb C^r}
\end{align*}
for all $v,w\in\mathbb C^r$. If $e_i=\nu(e_i^{\mathrm{std}})$, then this is equivalent to
\begin{align*}
h_p(e_i,e_j)=\delta_{ij}
\end{align*}
for all $1\le i,j\le r$. Indeed, for
\begin{align*}
v=\sum_{i=1}^r v_i e_i^{\mathrm{std}}
\end{align*}
and
\begin{align*}
w=\sum_{j=1}^r w_j e_j^{\mathrm{std}},
\end{align*}
complex-linearity gives
\begin{align*}
\nu(v)=\sum_{i=1}^r v_i e_i
\end{align*}
and
\begin{align*}
\nu(w)=\sum_{j=1}^r w_j e_j.
\end{align*}
Since $h_p$ is linear in the first argument and conjugate-linear in the second,
\begin{align*}
h_p(\nu(v),\nu(w))=\sum_{i=1}^r\sum_{j=1}^r v_i\overline{w_j}\,h_p(e_i,e_j).
\end{align*}
When $h_p(e_i,e_j)=\delta_{ij}$, this becomes
\begin{align*}
h_p(\nu(v),\nu(w))=\sum_{i=1}^r\sum_{j=1}^r v_i\overline{w_j}\delta_{ij}.
\end{align*}
The Kronecker delta keeps only the terms with $i=j$, so
\begin{align*}
h_p(\nu(v),\nu(w))=\sum_{i=1}^r v_i\overline{w_i}=(v,w)_{\mathbb C^r}.
\end{align*}
For the product bundle $E=M\times\mathbb C^r$ with the standard Hermitian metric, define
\begin{align*}
s(p)(v)=(p,v).
\end{align*}
Then
\begin{align*}
h_p(s(p)(v),s(p)(w))=(v,w)_{\mathbb C^r},
\end{align*}
so $s$ is a global unitary frame. Every unitary frame $\nu:\mathbb C^r\to \{p\}\times\mathbb C^r$ is uniquely of the form
\begin{align*}
\nu=s(p)\circ A
\end{align*}
for a complex-linear automorphism $A$ of $\mathbb C^r$. The frame $\nu$ is unitary exactly when, for all $v,w\in\mathbb C^r$,
\begin{align*}
(v,w)_{\mathbb C^r}=h_p(\nu(v),\nu(w)).
\end{align*}
Substituting $\nu=s(p)\circ A$ gives
\begin{align*}
h_p(\nu(v),\nu(w))=h_p(s(p)(Av),s(p)(Aw)).
\end{align*}
Because $h$ is the standard Hermitian metric on the product bundle,
\begin{align*}
h_p(s(p)(Av),s(p)(Aw))=(Av,Aw)_{\mathbb C^r}.
\end{align*}
Thus $\nu$ is unitary exactly when
\begin{align*}
(Av,Aw)_{\mathbb C^r}=(v,w)_{\mathbb C^r}
\end{align*}
for all $v,w\in\mathbb C^r$, which is precisely the condition $A\in U(r)$. Hence
\begin{align*}
M\times U(r)\longrightarrow \operatorname{Fr}_U(M\times\mathbb C^r,h),\qquad (p,A)\longmapsto s(p)\circ A
\end{align*}
is a fibrewise bijection and respects the right $U(r)$-action.
In complex rank $1$, a unitary frame over $p$ is a single vector $e\in E_p$ satisfying
\begin{align*}
h_p(e,e)=1.
\end{align*}
If $e$ and $f$ are two unitary vectors in the same complex line $E_p$, then $f=ze$ for a unique $z\in\mathbb C^\times$. Using linearity in the first argument and conjugate-linearity in the second,
\begin{align*}
1=h_p(f,f)=h_p(ze,ze)=z\overline z\,h_p(e,e)=|z|^2.
\end{align*}
Thus $z\in U(1)$. A global unitary frame therefore trivializes the line bundle by
\begin{align*}
M\times\mathbb C\longrightarrow E,\qquad (p,z)\longmapsto z\,e(p),
\end{align*}
while a product Hermitian line bundle has the global unitary frame obtained from its standard unit vector. Changing a local unitary frame of a Hermitian line bundle is exactly multiplication by a smooth function $g:U\to U(1)$, so the unitary frame bundle is the principal $U(1)$-bundle that records the line bundle's gauge freedom.
[/example]
The real and complex stories have the same form: a metric-like structure selects the frames in which the fibrewise geometry has its standard model. The selected frames form a principal subbundle whose group is the symmetry group of that standard model.
## Reductions of Structure Group
The examples above suggest a general principle. A vector bundle initially has transition functions in a large group such as $GL_r(\mathbb R)$ or $GL_r(\mathbb C)$, but extra structure may allow the same bundle to be described using transition functions in a smaller subgroup. The principal-bundle formulation makes this statement precise through reductions of structure group.
[definition: Reduction of Structure Group]
Let $P\to M$ be a principal $G$-bundle and let $H\le G$ be a Lie subgroup. A reduction of structure group of $P$ from $G$ to $H$ is a principal $H$-bundle $Q\to M$ together with an $H$-equivariant embedding $Q\hookrightarrow P$ over $M$ such that the induced map
\begin{align*}
Q\times_H G \longrightarrow P, \qquad [q,g]\longmapsto qg,
\end{align*}
is an isomorphism of principal $G$-bundles.
[/definition]
The associated-bundle condition says that $P$ is recovered from $Q$ by allowing all $G$-frames again. Thus a reduction is not unrelated total-space data; it is a compatible choice of smaller set of frames inside $P$. Without this compatibility condition, an $H$-bundle mapping into $P$ could miss the transition functions of $P$ and would not describe the same underlying principal $G$-bundle.
To use the definition in practice, one needs a way to recognise such compatible choices without first constructing the subbundle $Q$ by hand. In each fibre of $P$, an $H$-reduction selects not a single point but an $H$-orbit of points, since changing an adapted frame by an element of $H$ should not change the reduced structure. The quotient bundle $P/H$ packages these possible orbits over all points of $M$, so a section of $P/H$ is exactly the sort of global datum that can select the reduced frames fibre by fibre. The next theorem formalises this selection principle and is the main test for reductions used throughout the rest of the chapter.
[illustration:frame-bundle-metric-reduction]
The closed-subgroup hypothesis is part of the mechanism, not a technical decoration: it ensures that $G/H$ is a smooth manifold and hence that $P/H$ is an ordinary smooth fibre bundle. With that smooth quotient available, the statement below converts the geometric language of principal subbundles into the more flexible language of sections.
[quotetheorem:6248]
[citeproof:6248]
This theorem is useful because many geometric structures are sections of natural quotient bundles. The closedness of $H$ is not cosmetic: if $H$ is not closed, $G/H$ can fail to be a smooth manifold, so there is no ordinary smooth associated bundle whose sections could classify reductions. The result also does not assert that sections always exist; their nonexistence is precisely how obstruction theory enters, for example in orientability and spin questions. A Riemannian metric is a section of the bundle of positive-definite inner products, which leads to the concrete equivalence between metrics and orthogonal reductions.
[quotetheorem:6249]
[citeproof:6249]
The theorem uses the compact subgroup $O(r)$ because it is exactly the stabiliser of the Euclidean inner product on $\mathbb R^r$. Reducing to a different subgroup would encode different data: for instance, $GL_r^+(\mathbb R)$ records orientation but not lengths, while $SO(r)$ records both orientation and a metric. A concrete way to see the distinction is the product bundle $M\times \mathbb R^r$: its standard orientation gives a $GL_r^+(\mathbb R)$-reduction, but the frames
\begin{align*}
(2e_1,e_2,\dots,e_r)
\end{align*}
are positively oriented and do not preserve the Euclidean inner product, so orientation data alone cannot recover a metric. Without an $O(r)$-reduction, there is no specified class of frames in which vectors have their standard lengths and angles. The theorem does not claim a metric is unique on a vector bundle; different reductions to $O(r)$ generally give different bundle metrics, although paracompactness guarantees that smooth metrics exist on ordinary smooth vector bundles. This metric-reduction dictionary is the prototype for later characteristic classes, where topology is read from the possible reductions and their obstructions.
The same principle applies to orientation and complex structure, though the precise subgroup changes. An orientation is a reduction from $GL_r(\mathbb R)$ to $GL_r^+(\mathbb R)$, and a metric together with orientation gives a reduction to $SO(r)$. The next example spells out the orientation case and shows how the obstruction appears in transition functions.
[explanation: Local Tests for Reductions]
To compute the frame-bundle transition functions of a vector bundle, first choose vector bundle charts $\Phi_i:E|_{U_i}\to U_i\times \mathbb R^r$. On each overlap $U_i\cap U_j$, write the coordinate change in the form
\begin{align*}
\Phi_i\circ \Phi_j^{-1}(p,v)=(p,g_{ij}(p)v).
\end{align*}
The same maps $g_{ij}:U_i\cap U_j\to GL_r(\mathbb R)$, interpreted with the right-action convention, are the transition functions of $\operatorname{Fr}(E)$.
To verify a reduction to $H\le GL_r(\mathbb R)$ from local data, look for new local frames whose transition functions all take values in $H$. Equivalently, find gauge changes $a_i:U_i\to GL_r(\mathbb R)$ such that
\begin{align*}
a_i(p)^{-1}g_{ij}(p)a_j(p)\in H
\end{align*}
on every nonempty overlap. For orientation, this asks whether all determinants can be made positive; for a metric, it asks whether the chosen frames can be made orthonormal so that the resulting transition functions lie in $O(r)$.
[/explanation]
The orientation case is the most economical place to see the local test at work, because the subgroup condition is detected by the sign of a determinant.
[example: Orientation as a Reduction]
Let $E\to M$ be a real rank-$r$ vector bundle, and write
\begin{align*}
GL_r^+(\mathbb R)=\{A\in GL_r(\mathbb R):\det A>0\}.
\end{align*}
A reduction of $\operatorname{Fr}(E)$ to $GL_r^+(\mathbb R)$ is the same as choosing, smoothly in $p$, which frames of $E_p$ are positive: if $\nu$ is positive, then $\nu\cdot A$ is positive exactly when $\det A>0$. In local trivialisations with transition functions $g_{ij}:U_i\cap U_j\to GL_r(\mathbb R)$, this says that after choosing compatible local frames one has
\begin{align*}
\det g_{ij}(p)>0
\end{align*}
for every $p\in U_i\cap U_j$.
The Mobius line bundle over $S^1$ shows what can go wrong. For a line bundle, the structure group is
\begin{align*}
GL_1(\mathbb R)=\mathbb R^\times
\end{align*}
and
\begin{align*}
GL_1^+(\mathbb R)=\mathbb R_{>0}.
\end{align*}
Using the standard two-chart description of the Mobius line bundle, the transition function on one overlap component is multiplication by $1$, while on the other overlap component it is multiplication by $-1$. If gauge changes by nonzero functions $a_0:U_0\to\mathbb R^\times$ and $a_1:U_1\to\mathbb R^\times$ made both transition functions positive, then the new transition functions would be
\begin{align*}
h_{01}=a_0^{-1}g_{01}a_1.
\end{align*}
On the component where $g_{01}=1$, this has sign
\begin{align*}
\operatorname{sgn}(h_{01})=\operatorname{sgn}(a_0)^{-1}\operatorname{sgn}(a_1),
\end{align*}
while on the component where $g_{01}=-1$, it has sign
\begin{align*}
\operatorname{sgn}(h_{01})=-\operatorname{sgn}(a_0)^{-1}\operatorname{sgn}(a_1).
\end{align*}
Since each $a_i$ is continuous and nowhere zero on a connected chart, each $\operatorname{sgn}(a_i)$ is constant, so these two signs are opposite. They cannot both be positive. Thus the Mobius line bundle has transition functions in $GL_1(\mathbb R)$ but no reduction to $GL_1^+(\mathbb R)$.
If a bundle metric is also chosen, the orthonormal frame bundle restricts changes of frame to $O(r)$. Intersecting the positive frames with the orthonormal frames restricts the change matrix $A$ to satisfy both
\begin{align*}
A^\top A=I
\end{align*}
and
\begin{align*}
\det A>0.
\end{align*}
Taking determinants in $A^\top A=I$ gives
\begin{align*}
\det(A^\top A)=\det I,
\end{align*}
hence
\begin{align*}
\det(A^\top)\det(A)=1.
\end{align*}
Since $\det(A^\top)=\det A$, this becomes
\begin{align*}
(\det A)^2=1.
\end{align*}
Together with $\det A>0$, it follows that $\det A=1$, so the combined metric-and-orientation reduction has structure group $SO(r)$.
[/example]
Complex structures on real bundles can also be expressed through reductions, but they require a representation of $GL_r(\mathbb C)$ inside $GL_{2r}(\mathbb R)$. The following example gives the standard reduction associated with a fibrewise complex structure.
[example: Complex Structure as a Reduction]
Let $E\to M$ be a real vector bundle of rank $2r$, and let $J:E\to E$ be a smooth bundle endomorphism satisfying
\begin{align*}
J^2=-\operatorname{id}_E.
\end{align*}
Let $J_0:\mathbb R^{2r}\to\mathbb R^{2r}$ be the standard complex structure obtained from the identification $\mathbb R^{2r}\cong\mathbb C^r$, so $J_0(v)=iv$ in complex notation and
\begin{align*}
J_0^2(v)=i(iv)=-v
\end{align*}
for every $v\in\mathbb R^{2r}$.
A frame $\nu:\mathbb R^{2r}\to E_p$ is adapted to $J$ when it intertwines the two complex structures:
\begin{align*}
J_p\circ \nu=\nu\circ J_0.
\end{align*}
Equivalently, for every $v\in\mathbb R^{2r}$,
\begin{align*}
J_p(\nu(v))=\nu(J_0v).
\end{align*}
Thus $\nu$ is not merely a real-linear frame; it is complex-linear after viewing $\mathbb R^{2r}$ as $\mathbb C^r$ by $J_0$ and viewing $E_p$ as a complex vector space by $J_p$.
Now compare two adapted frames in the same fibre. Since both are real frames, there is a unique $A\in GL_{2r}(\mathbb R)$ such that
\begin{align*}
\mu=\nu\circ A.
\end{align*}
The frame $\mu$ is adapted exactly when
\begin{align*}
J_p\circ \mu=\mu\circ J_0.
\end{align*}
Substituting $\mu=\nu\circ A$ gives
\begin{align*}
J_p\circ \nu\circ A=\nu\circ A\circ J_0.
\end{align*}
Because $\nu$ is adapted, $J_p\circ\nu=\nu\circ J_0$, so the left-hand side becomes
\begin{align*}
\nu\circ J_0\circ A=\nu\circ A\circ J_0.
\end{align*}
Since $\nu$ is an isomorphism, canceling $\nu$ gives
\begin{align*}
J_0\circ A=A\circ J_0.
\end{align*}
Therefore the allowed change-of-frame matrices are precisely the real invertible matrices commuting with $J_0$. Under the identification $\mathbb R^{2r}\cong\mathbb C^r$, this condition says exactly that $A$ is complex-linear, so the structure group is $GL_r(\mathbb C)$.
The adapted frames therefore form
\begin{align*}
\operatorname{Fr}_{\mathbb C}(E,J):=\{\nu\in\operatorname{Fr}(E):J_p\circ\nu=\nu\circ J_0\},
\end{align*}
and the right action is
\begin{align*}
\nu\cdot A=\nu\circ A,\qquad A\in GL_r(\mathbb C).
\end{align*}
Indeed,
\begin{align*}
J_p\circ(\nu\circ A)=(J_p\circ\nu)\circ A=(\nu\circ J_0)\circ A=\nu\circ(J_0A),
\end{align*}
and since $A$ commutes with $J_0$,
\begin{align*}
\nu\circ(J_0A)=\nu\circ(AJ_0)=(\nu\circ A)\circ J_0.
\end{align*}
So $\nu\cdot A$ is again adapted. Conversely, the calculation above shows that every adapted frame in the same fibre is obtained from $\nu$ by a unique such $A$.
If $h$ is a compatible Hermitian metric on $(E,J)$, then one further restricts to adapted frames satisfying
\begin{align*}
h_p(\nu(v),\nu(w))=(v,w)_{\mathbb C^r}.
\end{align*}
If $\mu=\nu\circ A$, then
\begin{align*}
h_p(\mu(v),\mu(w))=h_p(\nu(Av),\nu(Aw))=(Av,Aw)_{\mathbb C^r}.
\end{align*}
Thus $\mu$ is again unitary exactly when
\begin{align*}
(Av,Aw)_{\mathbb C^r}=(v,w)_{\mathbb C^r}
\end{align*}
for all $v,w\in\mathbb C^r$, which is the condition $A\in U(r)$. Hence a fibrewise complex structure reduces the real frame bundle from $GL_{2r}(\mathbb R)$ to $GL_r(\mathbb C)$, and a compatible Hermitian metric reduces it further to $U(r)$.
[/example]
The moral of the chapter is that geometric structure is encoded by restricting the permitted changes of frame. Full frames give $GL$-bundles; metrics, orientations, complex structures, and Hermitian metrics select subbundles with groups $O(r)$, $GL_r^+(\mathbb R)$, $GL_r(\mathbb C)$, $SO(r)$, or $U(r)$. In the next part of the course, connections will be introduced as differential forms on these principal bundles, and reductions will determine which connection forms are compatible with the chosen geometry.
Frame bundles make the abstract symmetry language tangible, but geometry still needs a way to turn that symmetry into tensorial objects. The next chapter explains how associated bundles are built from a principal bundle and a representation, so that vector fields, tensors, and other geometric data can be treated equivariantly.
# 4. Associated Bundles and Equivariant Descriptions
This chapter belongs to the course's development of principal bundles as a language for geometry with changing frames. Chapters 2 and 3 explained how a principal $G$-bundle $P \to M$ records local symmetries, transition functions, and frame choices; the present chapter explains how ordinary geometric objects with fibres such as vector spaces, manifolds, or Lie algebras are recovered from that symmetry data. The construction turns a principal $G$-bundle and a left $G$-space $F$ into a fibre bundle over $M$ with fibre $F$. The central point is that sections of this new bundle are the same thing as $G$-equivariant maps out of $P$, so geometric fields may be described either downstairs on $M$ or upstairs on the symmetry space $P$.
## The Associated Bundle Construction
The first problem is to build a bundle whose transition functions are prescribed by a principal bundle, but whose typical fibre is not the group $G$ itself. A left action of $G$ on a space $F$ tells us how a change of principal frame should transform an element of $F$. Before forming the quotient, we isolate the transformation law that will be imposed on the model fibre.
[definition: Left G-Space]
Let $G$ be a Lie group. A left $G$-space is a smooth manifold $F$ equipped with a smooth map
\begin{align*}
G \times F &\longrightarrow F, & (g,f) &\longmapsto g \cdot f,
\end{align*}
satisfying $e \cdot f = f$ and $(gh) \cdot f = g \cdot (h \cdot f)$ for all $g,h \in G$ and $f \in F$.
[/definition]
This definition supplies the transformation rule on the model fibre. The remaining problem is to combine it with the right action on $P$: moving from $p$ to $pg$ should not change the geometric point of the new bundle, but it must change the fibre coordinate by the corresponding left action. If we merely took $P \times F \to M$, then every choice of principal frame would give a separate copy of the same geometric fibre value, so the construction would depend on redundant frame data. The quotient below removes exactly that redundancy; the inverse in the formula is what makes the right action on $P$ compatible with the left action on $F$.
[definition: Associated Bundle]
Let $\pi:P \to M$ be a smooth principal $G$-bundle and let $F$ be a left $G$-space. The associated bundle with fibre $F$ is
\begin{align*}
P \times_G F := (P \times F)/G,
\end{align*}
where $G$ acts on $P \times F$ on the right by
\begin{align*}
(p,f) \cdot g := (p g, g^{-1} \cdot f).
\end{align*}
The projection is
\begin{align*}
\pi_F : P \times_G F &\longrightarrow M, & [p,f] &\longmapsto \pi(p).
\end{align*}
[/definition]
The notation $[p,f]$ means the $G$-orbit of $(p,f)$. The same equivalence relation may be written as $(pg,f) \sim (p,g \cdot f)$, which is often the most convenient form for computations. Having defined the quotient, the next issue is whether it is genuinely a smooth fibre bundle over $M$ rather than only a set-theoretic quotient.
[quotetheorem:6127]
[citeproof:6127]
The hypotheses are doing real work here. Smoothness of the $G$-action on $F$ is needed because the transition functions of the new bundle are exactly $x \mapsto (f \mapsto g_{ij}(x)\cdot f)$; if the action were not smooth, the quotient could fail to have smooth bundle charts even when $P$ is smooth. The principality of $P$ is also essential: without a free and transitive fibrewise $G$-action, a point $p \in P_x$ would not give a well-defined coordinate on the whole fibre over $x$. The theorem does not say that every fibre bundle with fibre $F$ arises this way from a fixed group action; it says that whenever the structure group has been reduced to $G$ and acts on $F$, the corresponding bundle is obtained by this quotient. This is the practical bridge from principal bundles to ordinary fibre bundles, and it is useful to check first that no new twisting appears when the principal bundle itself is a product.
[example: Associated Bundle Of A Product Principal Bundle]
Let $P=M\times G$ with right action $(x,h)g=(x,hg)$, and let $F$ be a left $G$-space. Define
\begin{align*}
\Phi([(x,h),f])=(x,h\cdot f).
\end{align*}
This is independent of the representative: if we replace $((x,h),f)$ by $((x,h),f)\cdot g=((x,hg),g^{-1}\cdot f)$, then
\begin{align*}
\Phi([(x,hg),g^{-1}\cdot f])=(x,(hg)\cdot(g^{-1}\cdot f)).
\end{align*}
By the left action identity $(ab)\cdot u=a\cdot(b\cdot u)$,
\begin{align*}
(x,(hg)\cdot(g^{-1}\cdot f))=(x,h\cdot(g\cdot(g^{-1}\cdot f))).
\end{align*}
Applying the same identity to $g$ and $g^{-1}$ gives
\begin{align*}
(x,h\cdot(g\cdot(g^{-1}\cdot f)))=(x,h\cdot((gg^{-1})\cdot f)).
\end{align*}
Since $gg^{-1}=e$ and $e\cdot f=f$,
\begin{align*}
(x,h\cdot((gg^{-1})\cdot f))=(x,h\cdot f)=\Phi([(x,h),f]).
\end{align*}
Define
\begin{align*}
\Psi(x,u)=[(x,e),u].
\end{align*}
Then
\begin{align*}
\Phi(\Psi(x,u))=\Phi([(x,e),u])=(x,e\cdot u)=(x,u),
\end{align*}
so $\Phi\circ\Psi=\operatorname{id}_{M\times F}$. Conversely,
\begin{align*}
\Psi(\Phi([(x,h),f]))=\Psi(x,h\cdot f)=[(x,e),h\cdot f].
\end{align*}
The representative $((x,e),h\cdot f)$ is equivalent to $((x,h),f)$ because
\begin{align*}
((x,e),h\cdot f)\cdot h=((x,h),h^{-1}\cdot(h\cdot f)).
\end{align*}
By the left action identity,
\begin{align*}
h^{-1}\cdot(h\cdot f)=(h^{-1}h)\cdot f=e\cdot f=f.
\end{align*}
Hence $[(x,e),h\cdot f]=[(x,h),f]$, so $\Psi\circ\Phi=\operatorname{id}$. Both maps preserve the base point $x$, and therefore the associated bundle of the product principal bundle is the product $F$-bundle $M\times F$.
[/example]
This example fixes the local picture: every associated bundle is locally of this form, and all global information sits in how the copies of $F$ are glued by the $G$-action.
## Sections As Equivariant Maps
The next question is how to describe a section of $P \times_G F$ without choosing local charts. Since a point of the associated fibre is represented by $[p,f]$, a section should assign to each $p \in P$ a value $f \in F$ in a way compatible with changing $p$ along its principal fibre.
[definition: Equivariant Map To The Fibre]
Let $P$ be a right principal $G$-bundle and let $F$ be a left $G$-space. A smooth map $\varphi:P \to F$ is $G$-equivariant for the associated bundle convention if
\begin{align*}
\varphi(pg)=g^{-1}\cdot \varphi(p)
\end{align*}
for all $p \in P$ and $g \in G$.
[/definition]
The inverse in the formula matches the quotient convention used above. It says that if the frame is changed on the right by $g$, the coordinate expression of the same geometric object changes by the inverse left action. This motivates the following theorem, which is the associated-bundle analogue of the descent logic used for transition functions in Chapter 2: the equivariance condition is not an auxiliary constraint, but exactly the descent condition for a map on $P$ to define a section over $M$.
[quotetheorem:6250]
[citeproof:6250]
The equivariance hypothesis is precisely the descent condition. If a smooth map $\varphi:P \to F$ fails to satisfy $\varphi(pg)=g^{-1}\cdot\varphi(p)$, then the formula $s(x)=[p,\varphi(p)]$ depends on the chosen frame $p \in P_x$ and therefore does not define a section over $M$. Conversely, the theorem does not identify all maps $P \to F$ with geometric fields; it identifies only those maps whose variation along each principal fibre is exactly prescribed by the representation convention. This is the main computational reason to use principal bundles: a section over the base can be replaced by an equivariant function upstairs, where changes of frame are recorded by the principal action.
[example: Local Functions And Gauge Transformation]
Let $s_i:U_i \to P$ and $s_j:U_j \to P$ be local sections over an overlap, and let $\varphi:P \to F$ satisfy $\varphi(pg)=g^{-1}\cdot \varphi(p)$. The associated section is represented in the $s_i$-frame by
\begin{align*}
f_i(x)=\varphi(s_i(x)).
\end{align*}
Assume the two local sections are related by
\begin{align*}
s_j(x)=s_i(x)g_{ij}(x).
\end{align*}
Then the representative in the $s_j$-frame is
\begin{align*}
f_j(x)=\varphi(s_j(x)).
\end{align*}
Substituting the transition formula for $s_j$ gives
\begin{align*}
f_j(x)=\varphi(s_i(x)g_{ij}(x)).
\end{align*}
By equivariance of $\varphi$, with $p=s_i(x)$ and $g=g_{ij}(x)$,
\begin{align*}
\varphi(s_i(x)g_{ij}(x))=g_{ij}(x)^{-1}\cdot \varphi(s_i(x)).
\end{align*}
Using the definition of $f_i(x)$, this becomes
\begin{align*}
f_j(x)=g_{ij}(x)^{-1}\cdot f_i(x).
\end{align*}
Thus the local fibre coordinate changes by the inverse of the transition function determined by $s_j=s_i g_{ij}$, exactly matching the associated-bundle equivariance convention.
[/example]
The sign of the inverse depends on whether the transition functions are defined by $s_j=s_i g_{ij}$ or $s_i=s_j g_{ji}$. The invariant statement is the equivariant-map correspondence; local formulas must be read together with the chosen convention.
## Functoriality Of Associated Bundles
Once the associated bundle construction is available, the next question is whether it behaves naturally under maps of fibres and maps of principal bundles. The answer is yes: equivariant maps between model fibres induce bundle maps between associated bundles, and morphisms of principal bundles induce pullback-compatible maps.
[quotetheorem:6251]
[citeproof:6251]
The $G$-equivariance of $a$ is the exact condition that makes the formula independent of representatives. For a concrete failure, take $G=\mathbb R^\times$ acting on $F=F'=\mathbb R$ by scalar multiplication and let $a(v)=v+1$. Then $a(g^{-1}v)=g^{-1}v+1$ is not equal to $g^{-1}a(v)=g^{-1}v+g^{-1}$ when $g\ne 1$, so the equivalent representatives $(p,v)$ and $(pg,g^{-1}v)$ may be sent to non-equivalent points. The theorem does not claim that every bundle map between associated bundles comes from a map of model fibres; it gives a canonical source of such maps when the same principal bundle and a compatible fibre map are fixed. This result turns each principal bundle into a machine for producing fibre bundles from $G$-spaces, and it leads to the companion question of what happens when the principal bundle itself is varied by a principal bundle morphism.
[quotetheorem:6252]
[citeproof:6252]
The principal-bundle equivariance of $\Phi$ is essential here. If $\Phi(pg)$ were not equal to $\Phi(p)g$, then the two representatives $(p,u)$ and $(pg,g^{-1}\cdot u)$ could have images representing different points of $Q \times_G F$. The theorem also does not say that an arbitrary bundle map between $P \times_G F$ and $Q \times_G F$ determines a principal bundle morphism; it only constructs the induced map from a morphism already compatible with the group action. The construction is therefore compatible with both ingredients, and in later chapters this naturality is used when connections, curvature forms, and gauge transformations are passed from a principal bundle to associated vector bundles.
## Associated Vector Bundles From Representations
The most important special case occurs when the model fibre is a vector space and $G$ acts linearly. This is how a principal bundle produces vector bundles, and it explains why frame bundles are universal for vector bundle constructions.
[definition: Associated Vector Bundle]
Let $\rho:G \to GL(V)$ be a finite-dimensional representation of a Lie group $G$ on a real or complex vector space $V$. For a principal $G$-bundle $P \to M$, the associated vector bundle is
\begin{align*}
P \times_\rho V := P \times_G V,
\end{align*}
where the left action of $G$ on $V$ is $g\cdot v=\rho(g)v$.
Its projection is
\begin{align*}
\pi_\rho:P \times_\rho V &\longrightarrow M, & [p,v] &\longmapsto \pi(p).
\end{align*}
[/definition]
Because the action is linear, the fibres are expected to inherit vector space structures from $V$. The next theorem records the needed compatibility: the quotient construction does not merely produce a fibre bundle with fibre a set $V$, but a vector bundle whose transition maps are linear.
[quotetheorem:6127]
[citeproof:6127]
Linearity of the representation is the point that upgrades a fibre bundle to a vector bundle. If $G$ acted on $V$ by nonlinear diffeomorphisms, the same quotient would still be a smooth fibre bundle with fibre $V$, but addition and scalar multiplication would not be preserved on overlaps and hence would not descend to globally defined vector-space operations on fibres. The theorem does not classify all vector bundles by a single representation; it says that a chosen principal $G$-bundle and a chosen representation produce a vector bundle with transition functions obtained by applying $\rho$. This packages many standard constructions: whenever a geometric object has components transforming by a representation of the structure group, it is a section of the associated vector bundle for that representation.
[example: Tensor Bundles From The Frame Bundle]
Let $E \to M$ be a rank $r$ real vector bundle, and write a frame $u \in \operatorname{Fr}(E)_x$ as a linear isomorphism $u:\mathbb R^r \to E_x$. For the standard representation of $GL_r(\mathbb R)$ on $\mathbb R^r$, define
\begin{align*}
[u,v]\longmapsto u(v).
\end{align*}
If the representative is changed by $A\in GL_r(\mathbb R)$, then the associated-bundle relation sends $(u,v)$ to $(uA,A^{-1}v)$. The image of the new representative is
\begin{align*}
(uA)(A^{-1}v)=u(A(A^{-1}v))=u((AA^{-1})v)=u(ev)=u(v).
\end{align*}
Thus the standard representation recovers $E$.
For the dual bundle, use the dual representation on $(\mathbb R^r)^*$:
\begin{align*}
A\cdot \lambda=\lambda\circ A^{-1}.
\end{align*}
The associated map is
\begin{align*}
[u,\lambda]\longmapsto \lambda\circ u^{-1}.
\end{align*}
Changing representatives gives $(u,\lambda)\cdot A=(uA,A^{-1}\cdot \lambda)$, and the dual action gives
\begin{align*}
A^{-1}\cdot\lambda=\lambda\circ A.
\end{align*}
The image of the changed representative is
\begin{align*}
(A^{-1}\cdot\lambda)\circ (uA)^{-1}=(\lambda\circ A)\circ(A^{-1}\circ u^{-1})=\lambda\circ(AA^{-1})\circ u^{-1}=\lambda\circ u^{-1}.
\end{align*}
So the dual representation recovers $E^*$.
The same representative check gives the usual tensor bundles. For instance, on $(\mathbb R^r)^{\otimes k}$ the tensor-power representation is
\begin{align*}
A\cdot(v_1\otimes\cdots\otimes v_k)=(Av_1)\otimes\cdots\otimes(Av_k),
\end{align*}
and the associated map sends
\begin{align*}
[u,v_1\otimes\cdots\otimes v_k]\longmapsto u(v_1)\otimes\cdots\otimes u(v_k).
\end{align*}
If the representative is changed by $A$, the fibre coordinate becomes
\begin{align*}
A^{-1}\cdot(v_1\otimes\cdots\otimes v_k)=(A^{-1}v_1)\otimes\cdots\otimes(A^{-1}v_k).
\end{align*}
Its image under the frame $uA$ is
\begin{align*}
(uA)(A^{-1}v_1)\otimes\cdots\otimes(uA)(A^{-1}v_k)=u(v_1)\otimes\cdots\otimes u(v_k).
\end{align*}
Passing this construction to symmetric tensors and alternating tensors gives $\operatorname{Sym}^k E$ and $\Lambda^k E$. In a local frame, a tensor field is therefore a function with values in the corresponding model tensor space, and replacing the frame by $uA$ changes its components by the inverse associated-bundle coordinate rule for that representation.
[/example]
The previous example is more than a list of constructions: it says that the frame bundle contains all tensorial information of the vector bundle. This raises a basic recovery question: if the frame bundle is treated as the primary principal bundle, can the original vector bundle be reconstructed without choosing a global frame? The obstruction is precisely the absence of global frames on nontrivial bundles, so the reconstruction has to use the associated-bundle quotient that glues local frame coordinates by the inverse transition rule.
[quotetheorem:6253]
[citeproof:6253]
The use of the frame bundle is essential because its fibre consists of all linear identifications $\mathbb R^r \cong E_x$. For instance, over $M=S^2$, the product principal bundle $S^2\times GL_2(\mathbb R)$ associated through the standard representation gives the product vector bundle $S^2\times \mathbb R^2$, whereas $\operatorname{Fr}(TS^2)$ recovers the tangent bundle $TS^2$, which has no global frame. Thus an arbitrary principal $GL_r(\mathbb R)$-bundle may produce a rank $r$ vector bundle not canonically isomorphic to the chosen $E$. The theorem does not say that choosing a frame is global; indeed, a nontrivial vector bundle may have no global frame, and the principal bundle records exactly how local frames differ on overlaps. The recovery theorem explains why principal bundles are not extra structure added to vector bundles: a vector bundle and its frame bundle encode the same local linear data, while the principal bundle separates the choice of frame from the objects whose components are measured in that frame.
## Geometric Examples
The final question in this chapter is how the construction appears in familiar geometry. Three recurring examples are tautological bundles, tensor bundles, and adjoint bundles.
[example: Tautological Line Bundle From The Hopf Bundle]
Let $S^{2n+1}\to \mathbb{CP}^n$ be the Hopf principal $U(1)$-bundle, where $p\in S^{2n+1}\subset \mathbb C^{n+1}$ maps to the complex line $[p]$ and the right action is $p\lambda$ for $\lambda\in U(1)$. Let $U(1)$ act on $\mathbb C$ by the standard left action $\lambda\cdot z=\lambda z$. We identify the associated bundle with the tautological line bundle
\begin{align*}
\gamma^1=\{([p],v)\in \mathbb{CP}^n\times \mathbb C^{n+1}:v\in \mathbb C p\}
\end{align*}
by defining
\begin{align*}
\Phi([p,z])=([p],zp).
\end{align*}
This formula is independent of the representative. If $\lambda\in U(1)$, then the associated-bundle relation replaces $(p,z)$ by $(p\lambda,\lambda^{-1}z)$. Since $p\lambda$ spans the same complex line as $p$, we have $[p\lambda]=[p]$. Applying $\Phi$ to the new representative gives
\begin{align*}
\Phi([p\lambda,\lambda^{-1}z])=([p\lambda],(\lambda^{-1}z)(p\lambda)).
\end{align*}
Using $[p\lambda]=[p]$, this becomes
\begin{align*}
([p\lambda],(\lambda^{-1}z)(p\lambda))=([p],(\lambda^{-1}z)\lambda p).
\end{align*}
By associativity and commutativity of multiplication in $\mathbb C$,
\begin{align*}
([p],(\lambda^{-1}z)\lambda p)=([p],(\lambda^{-1}\lambda)zp).
\end{align*}
Since $\lambda^{-1}\lambda=1$, we obtain
\begin{align*}
([p],(\lambda^{-1}\lambda)zp)=([p],zp)=\Phi([p,z]).
\end{align*}
Thus $\Phi$ is well-defined.
To construct the inverse, take $([p],v)\in\gamma^1$ and choose a unit representative $p\in S^{2n+1}$ of the line $[p]$. Since $v\in\mathbb C p$ and $p\neq 0$, there is a unique $z\in\mathbb C$ such that $v=zp$. Define
\begin{align*}
\Psi([p],v)=[p,z].
\end{align*}
If the unit representative is changed from $p$ to $p\lambda$, then
\begin{align*}
v=zp=(z\lambda^{-1})(p\lambda).
\end{align*}
Because complex multiplication is commutative, $z\lambda^{-1}=\lambda^{-1}z$, so the new coordinate is $\lambda^{-1}z$. The associated-bundle relation therefore gives
\begin{align*}
[p\lambda,\lambda^{-1}z]=[p,z],
\end{align*}
and hence $\Psi$ is well-defined.
The two maps are inverse to each other. If $v=zp$, then
\begin{align*}
\Phi(\Psi([p],v))=\Phi([p,z])=([p],zp)=([p],v).
\end{align*}
In the other direction,
\begin{align*}
\Psi(\Phi([p,z]))=\Psi([p],zp)=[p,z].
\end{align*}
Therefore $S^{2n+1}\times_{U(1)}\mathbb C$ is the tautological complex line bundle: the principal point $p$ supplies a unit vector on the line, and the associated coordinate $z$ supplies the scalar multiple $zp$.
[/example]
This example turns the quotient definition into geometry: the associated vector remembers a scalar coordinate $z$, while the principal point $p$ supplies the unit vector that gives that coordinate a line in $\mathbb C^{n+1}$. Gauge theory needs a different associated object, where the fibre is the group itself and a change of frame acts by conjugation.
[definition: Adjoint Bundle]
Let $P \to M$ be a principal $G$-bundle. The adjoint bundle is
\begin{align*}
\operatorname{Ad}(P):=P \times_G G,
\end{align*}
where $G$ acts on itself by conjugation, $g\cdot h=ghg^{-1}$.
Its projection is
\begin{align*}
\pi_{\operatorname{Ad}}:\operatorname{Ad}(P) &\longrightarrow M, & [p,h] &\longmapsto \pi(p).
\end{align*}
[/definition]
The adjoint bundle has fibre $G$, not a vector space in general. It is the bundle whose sections describe gauge transformations in a form independent of a choice of chart. This motivates the following definition: for infinitesimal gauge theory and curvature, the group-valued fibre must be replaced by the Lie algebra with the infinitesimal version of conjugation.
[definition: Adjoint Lie Algebra Bundle]
Let $P \to M$ be a principal $G$-bundle with Lie algebra $\mathfrak g$. The adjoint Lie algebra bundle is
\begin{align*}
\operatorname{ad}(P):=P \times_G \mathfrak g,
\end{align*}
where $G$ acts on $\mathfrak g$ by the adjoint representation $\operatorname{Ad}:G \to GL(\mathfrak g)$.
Its projection is
\begin{align*}
\pi_{\operatorname{ad}}:\operatorname{ad}(P) &\longrightarrow M, & [p,X] &\longmapsto \pi(p).
\end{align*}
[/definition]
The vector bundle $\operatorname{ad}(P)$ is the natural home for infinitesimal gauge transformations in Chapter 9 and for the curvature of a principal connection in Chapter 8. Its local sections transform by the adjoint representation, which is why Lie-algebra-valued forms in gauge theory change by conjugation under a gauge change.
[example: Adjoint Bundle Of A Product Principal Bundle]
For $P=M\times G$ with right action $(x,k)g=(x,kg)$, the adjoint bundle is
\begin{align*}
(M\times G)\times_G G,
\end{align*}
where $G$ acts on the fibre $G$ by conjugation, $g\cdot h=ghg^{-1}$. Define
\begin{align*}
\Phi([((x,k),h)])=(x,khk^{-1}).
\end{align*}
This is independent of the representative. If $g\in G$, then the associated-bundle relation replaces $((x,k),h)$ by
\begin{align*}
((x,k),h)\cdot g=((x,kg),g^{-1}hg).
\end{align*}
Applying $\Phi$ to this new representative gives
\begin{align*}
\Phi([((x,kg),g^{-1}hg)])=(x,(kg)(g^{-1}hg)(kg)^{-1}).
\end{align*}
Since $(kg)^{-1}=g^{-1}k^{-1}$, the second coordinate is
\begin{align*}
(kg)(g^{-1}hg)(kg)^{-1}=k g g^{-1} h g g^{-1} k^{-1}.
\end{align*}
Using $gg^{-1}=e$ and $ke=k$, this becomes
\begin{align*}
k g g^{-1} h g g^{-1} k^{-1}=k e h e k^{-1}=khk^{-1}.
\end{align*}
Hence
\begin{align*}
\Phi([((x,kg),g^{-1}hg)])=(x,khk^{-1})=\Phi([((x,k),h)]).
\end{align*}
The inverse map is
\begin{align*}
\Psi(x,u)=[((x,e),u)].
\end{align*}
Then
\begin{align*}
\Phi(\Psi(x,u))=\Phi([((x,e),u)])=(x,eue^{-1})=(x,u).
\end{align*}
Conversely,
\begin{align*}
\Psi(\Phi([((x,k),h)]))=\Psi(x,khk^{-1})=[((x,e),khk^{-1})].
\end{align*}
The representative $((x,e),khk^{-1})$ is equivalent to $((x,k),h)$ because
\begin{align*}
((x,e),khk^{-1})\cdot k=((x,k),k^{-1}(khk^{-1})k).
\end{align*}
The fibre coordinate is
\begin{align*}
k^{-1}(khk^{-1})k=(k^{-1}k)h(k^{-1}k)=ehe=h.
\end{align*}
Thus
\begin{align*}
[((x,e),khk^{-1})]=[((x,k),h)],
\end{align*}
so $\Psi\circ\Phi=\operatorname{id}$ and $\operatorname{Ad}(M\times G)\cong M\times G$ in the chosen product chart.
The dependence on the chart is visible from the same equivalence relation. Let the original section be $s(x)=(x,e)$, and replace it by
\begin{align*}
s_a(x)=s(x)a(x)=(x,a(x))
\end{align*}
for a smooth map $a:M\to G$. If the same adjoint-bundle element has old coordinate $h$, then
\begin{align*}
[s(x),h]=[s(x)a(x),a(x)^{-1}ha(x)]=[s_a(x),a(x)^{-1}ha(x)].
\end{align*}
Therefore the fibre coordinate in the new chart is $a(x)^{-1}ha(x)$.
For the adjoint Lie algebra bundle, the fibre action is the adjoint representation. The same calculation gives
\begin{align*}
[s(x),X]=[s(x)a(x),\operatorname{Ad}_{a(x)^{-1}}X]=[s_a(x),\operatorname{Ad}_{a(x)^{-1}}X].
\end{align*}
Thus a local Lie-algebra coordinate changes from $X$ to $\operatorname{Ad}_{a(x)^{-1}}X$, which is the infinitesimal version of the conjugation rule for the group-valued adjoint bundle.
[/example]
The associated-bundle viewpoint therefore unifies the objects that will be differentiated in the next part of the course. A connection on $P$ will induce covariant derivatives on every associated vector bundle, and its curvature will be read either as an equivariant horizontal form on $P$ or as a form on $M$ with values in $\operatorname{ad}(P)$.
Associated bundles show how representations produce the objects on which geometry acts, but they do not yet tell us when a principal bundle carries extra structure. The next chapter asks when the structure group can be reduced, and how such reductions encode choices like orientations, metrics, or complex structures.
# 5. Reductions of Structure Group
This chapter develops the course's first systematic method for recognizing geometric structure inside a principal bundle. We assume the preceding material on principal bundles, transition functions, associated bundles, and frame bundles, and ask when a bundle with structure group $G$ can be described using a smaller Lie subgroup $H \subset G$. In the preceding chapters, transition functions and associated bundles made the structure group visible as the group in which all changes of local trivialisation take values. This chapter explains how extra geometric structure on the base or on an associated bundle is often the same thing as replacing those transition functions by $H$-valued ones.
Building on the quotient-section viewpoint introduced in Chapter 3 for reductions, the main result is the reduction-section theorem: reducing a principal $G$-bundle to $H$ is equivalent to choosing a section of the quotient bundle $P/H \to M$. This turns questions about orientations, Riemannian metrics, Hermitian metrics, volume forms, and almost complex structures into the common language of sections of homogeneous bundles.
## The Problem of Smaller Symmetry
Suppose a principal $G$-bundle $\pi:P\to M$ is described locally by transition functions $g_{ij}:U_i\cap U_j\to G$. If there is a subgroup $H\subset G$, the natural question is whether the same bundle can be described using only transition functions with values in $H$. Geometrically, this asks whether the fibres carry extra structure whose stabiliser is $H$.
For example, a general frame of an $n$-dimensional real vector space has symmetry group $GL_n(\mathbb R)$, while an orthonormal frame has symmetry group $O(n)$. Choosing a Riemannian metric on a vector bundle should therefore reduce the frame bundle from $GL_n(\mathbb R)$ to $O(n)$.
[definition: Lie Subgroup Reduction]
Let $G$ be a Lie group and let $H\subset G$ be a Lie subgroup. Let $\pi:P\to M$ be a smooth principal $G$-bundle. An $H$-reduction of $P$ is a smooth submanifold $Q\subset P$ such that:
1. $Q$ is preserved by the right action of $H$;
2. the restricted projection $\pi|_Q:Q\to M$ is a smooth principal $H$-bundle;
3. the natural map
\begin{align*}
Q\times_H G &\longrightarrow P, & [q,g]&\longmapsto qg
\end{align*}
is an isomorphism of principal $G$-bundles.
[/definition]
The third condition says that $P$ is recovered from $Q$ by extending the structure group from $H$ to $G$. Thus a reduction is not merely an $H$-invariant subset; it must contain enough points to generate all of $P$ after allowing the larger $G$-action.
[example: Product Bundle Orthonormal Frames]
Let $E=M\times \mathbb R^n$ carry the fibre metric
\begin{align*}
\langle (x,v),(x,w)\rangle_x=v\cdot w.
\end{align*}
A frame of the fibre $E_x=\{x\}\times \mathbb R^n$ is a linear isomorphism $\mathbb R^n\to \mathbb R^n$, so choosing the base point $x$ and the matrix of the isomorphism gives
\begin{align*}
\operatorname{Fr}(E)=M\times GL_n(\mathbb R).
\end{align*}
For a frame represented by $A\in GL_n(\mathbb R)$, the images of the standard basis are orthonormal exactly when
\begin{align*}
(Ae_i)\cdot(Ae_j)=e_i\cdot e_j=\delta_{ij}
\end{align*}
for all $i,j$. Since
\begin{align*}
(Ae_i)\cdot(Ae_j)=e_i^\top A^\top A e_j=(A^\top A)_{ij},
\end{align*}
this condition is equivalent to $A^\top A=I_n$. Hence the orthonormal frames are precisely
\begin{align*}
Q=M\times O(n)\subset M\times GL_n(\mathbb R).
\end{align*}
The right action of $O(n)$ preserves $Q$. Indeed, if $A\in O(n)$ and $B\in O(n)$, then
\begin{align*}
(AB)^\top(AB)=B^\topA^\top AB.
\end{align*}
Using $A^\top A=I_n$, this becomes
\begin{align*}
B^\topA^\top AB=B^\top I_n B=B^\topB.
\end{align*}
Using $B^\topB=I_n$, we get
\begin{align*}
(AB)^\top(AB)=I_n.
\end{align*}
Thus $AB\in O(n)$.
The extension-of-structure-group map is
\begin{align*}
\Phi:Q\times_{O(n)}GL_n(\mathbb R)\longrightarrow M\times GL_n(\mathbb R),\qquad [(x,A),g]\longmapsto (x,Ag).
\end{align*}
It is well-defined because, for $h\in O(n)$,
\begin{align*}
\Phi([(x,Ah),g])=(x,Ahg)=\Phi([(x,A),hg]).
\end{align*}
It is surjective because every $(x,C)\in M\times GL_n(\mathbb R)$ satisfies
\begin{align*}
(x,C)=\Phi([(x,I_n),C]).
\end{align*}
It is injective as follows. If
\begin{align*}
\Phi([(x,A),g])=\Phi([(x,A'),g']),
\end{align*}
then $Ag=A'g'$. Set $h=A^{-1}A'$. Since $A,A'\in O(n)$, we have
\begin{align*}
h^\toph=(A^{-1}A')^\top(A^{-1}A')=(A')^\top(A^{-1})^\topA^{-1}A'.
\end{align*}
For $A\in O(n)$, $A^{-1}=A^\top$, so $(A^{-1})^\topA^{-1}=AA^\top=I_n$. Therefore
\begin{align*}
h^\toph=(A')^\top A'=I_n,
\end{align*}
and hence $h\in O(n)$. Also $A'=Ah$, and from $Ag=A'g'=Ahg'$ we get $g=hg'$ after left-multiplying by $A^{-1}$. Therefore
\begin{align*}
[(x,A'),g']=[(x,Ah),g']=[(x,A),hg']=[(x,A),g].
\end{align*}
So $\Phi$ is a bijective bundle map with inverse obtained by the displayed representative calculation, and hence
\begin{align*}
Q\times_{O(n)}GL_n(\mathbb R)\cong M\times GL_n(\mathbb R)=\operatorname{Fr}(E).
\end{align*}
Thus $Q=M\times O(n)$ is an $O(n)$-reduction of the product frame bundle; it records exactly the frames adapted to the standard fibre metric.
[/example]
This example captures the guiding idea: a reduction selects, in every fibre, those frames adapted to a chosen piece of geometric structure. The next problem is to recognize such selections without first guessing the subbundle $Q$.
## Reductions as Sections of the Quotient Bundle
Given $H\subset G$, the right action of $H$ on $P$ has orbit space $P/H$. The projection $P/H\to M$ has fibre $G/H$, so it is an associated bundle for the principal $G$-bundle $P$. A point of $P/H$ over $x\in M$ records an $H$-orbit inside the $G$-torsor $P_x$, and choosing such an orbit in every fibre should be the same as choosing a reduced $H$-subbundle.
[definition: Quotient Bundle by a Subgroup]
Let $\pi:P\to M$ be a smooth principal $G$-bundle and let $H\subset G$ be a closed Lie subgroup. The quotient bundle $P/H\to M$ is the smooth fibre bundle whose total space is the orbit space of the right $H$-action on $P$, with projection induced by $\pi$.
[/definition]
The closedness hypothesis is part of the definition because it ensures that the homogeneous space $G/H$ is a smooth manifold and that $P/H\to M$ is locally modelled on $U\times G/H$. Without it, the orbit space may fail to be Hausdorff or to carry the intended smooth bundle structure. The question now is whether the fibrewise choice of an $H$-orbit in $P_x$ is exactly the same data as the reduced principal subbundle from the definition.
[quotetheorem:6254]
[citeproof:6254]
The theorem changes the task from finding a subbundle to finding a section. The closedness of $H$ is essential here: if $H$ is not closed, $G/H$ need not be a smooth manifold, so the phrase "smooth section of $P/H$" may not define an object in the category of smooth fibre bundles. For example, a dense one-parameter subgroup of a torus gives a quotient with pathological topology rather than an ordinary homogeneous manifold. The theorem gives a bijection of reduction data, but it does not say that reductions are unique; uniqueness only appears in special geometric situations where the quotient bundle has a distinguished section.
[example: Reduction over a Contractible Base]
Let $M$ be contractible and let $P\to M$ be a principal $G$-bundle. Choose a global product description
\begin{align*}
P\cong M\times G,
\end{align*}
with right action
\begin{align*}
(x,g)\cdot a=(x,ga).
\end{align*}
Quotienting by the right $H$-action gives
\begin{align*}
(M\times G)/H\cong M\times G/H,
\end{align*}
where the isomorphism sends the $H$-orbit of $(x,g)$ to $(x,gH)$.
A smooth map $s:M\to G/H$ therefore defines a section $\sigma:M\to M\times G/H$ by
\begin{align*}
\sigma(x)=(x,s(x)).
\end{align*}
Under the [Reduction Section Theorem](/theorems/Reduction Section Theorem), the corresponding subset of $M\times G$ is
\begin{align*}
Q_s=\{(x,g)\in M\times G:gH=s(x)\}.
\end{align*}
This subset is preserved by the right action of $H$: if $(x,g)\in Q_s$ and $h\in H$, then
\begin{align*}
(gh)H=g(hH)=gH=s(x),
\end{align*}
so $(x,gh)\in Q_s$. On each fibre, if $(x,g),(x,g')\in Q_s$, then
\begin{align*}
gH=s(x)=g'H.
\end{align*}
Thus $g'H=gH$, so $g'=gh$ for some $h\in H$; this $h$ is unique because right multiplication in $G$ is free. Hence $Q_s\to M$ is a principal $H$-bundle, locally the pullback of the principal $H$-bundle $G\to G/H$ along $s$.
For the constant map $s(x)=eH$, the defining condition becomes
\begin{align*}
Q_s=\{(x,g):gH=eH\}.
\end{align*}
Since $gH=eH$ holds exactly when $g\in H$, this gives
\begin{align*}
Q_s=M\times H.
\end{align*}
This is the standard reduction $M\times H\subset M\times G$. Thus, after a global trivialisation, reductions are exactly sections of the product $G/H$-bundle; local existence is automatic, and the real obstruction is whether these locally defined sections glue to a global section.
[/example]
The section viewpoint also explains why reductions interact well with transition functions. The next problem is to translate the existence of a section of $P/H$ back into the atlas language from the previous chapter, where changing local trivialisations changes the cocycle by coboundary terms.
[quotetheorem:6255]
[citeproof:6255]
This criterion is closest to the original definition of structure group. The phrase "after refining the cover and changing local trivialisations" is necessary: a single chosen cocycle may take values in $G\setminus H$ even when an equivalent cocycle, obtained by changing local sections, is $H$-valued. For instance, a trivial real line bundle may be described using local frames whose transition functions are positive or negative depending on the arbitrary signs of the chosen frames, although a global positive frame gives transitions in $GL_1^+(\mathbb R)$. Thus the criterion is not a test of one atlas in isolation; it is a test of whether the principal bundle admits some compatible atlas whose changes of frame preserve the extra structure. This is the form used in the examples below, where orientations, metrics, and complex structures are recognized by choosing better adapted frames.
## Orientations as Reductions
The first major geometric example is orientation. A smooth $n$-manifold has a frame bundle $\operatorname{Fr}(TM)$ with structure group $GL_n(\mathbb R)$. Since orientation-preserving changes of basis form the subgroup $GL_n^+(\mathbb R)$, orienting $M$ should be the same as reducing $\operatorname{Fr}(TM)$ to this subgroup.
[definition: Oriented Frame Bundle]
Let $M$ be a smooth $n$-manifold. A $GL_n^+(\mathbb R)$-reduction of $\operatorname{Fr}(TM)$ is called an oriented frame bundle of $M$.
[/definition]
An oriented frame at $x\in M$ is a frame belonging to one of the two connected components of $\operatorname{Fr}(T_xM)$. The question is whether the usual atlas definition of orientability, phrased through positive Jacobian determinants, is exactly the same as this principal-bundle reduction.
[quotetheorem:6129]
[citeproof:6129]
This theorem packages the usual atlas definition into principal-bundle language. Orientability is a global condition rather than a local one: every coordinate ball admits a positive frame, but these local choices may fail to glue around loops. The Mobius line bundle over $S^1$ gives the rank-one model of this failure, with monodromy $-1$ preventing a reduction to $GL_1^+(\mathbb R)$. In quotient-bundle terms, since $GL_n(\mathbb R)/GL_n^+(\mathbb R)$ has two points, an orientation is a section of a two-sheeted covering associated to the frame bundle. This viewpoint is the entry point to obstruction theory, where the failure to choose such a section is measured topologically.
[example: The Circle and the Mobius Band Boundary]
Write $S^1=\{(\cos \theta,\sin \theta):\theta\in \mathbb R\}$ and define $V_{(\cos \theta,\sin \theta)}=(-\sin \theta,\cos \theta)$. This vector is tangent because it is perpendicular to the radial vector:
\begin{align*}
(-\sin \theta,\cos \theta)\cdot(\cos \theta,\sin \theta)=-\sin\theta\cos\theta+\cos\theta\sin\theta=0.
\end{align*}
It is nowhere zero because
\begin{align*}
\|V_{(\cos \theta,\sin \theta)}\|^2=(-\sin\theta)^2+(\cos\theta)^2=\sin^2\theta+\cos^2\theta=1.
\end{align*}
Thus $V$ is a smooth nowhere-zero tangent vector field. Since each $T_xS^1$ is one-dimensional, every frame $u:\mathbb R\to T_xS^1$ is determined by $u(1)=aV_x$ for a unique $a\in\mathbb R^\times$. The positive frames are exactly those with $a>0$, so they form a $GL_1^+(\mathbb R)=\mathbb R_+$-reduction of $\operatorname{Fr}(TS^1)$ by *Orientation as a Reduction*.
For comparison, model the Mobius line bundle as
\begin{align*}
L=([0,1]\times \mathbb R)/((0,t)\sim(1,-t)).
\end{align*}
Transporting the standard frame $e(t)=1$ once around the base gives $e(1)=-e(0)$, so the monodromy in $GL_1(\mathbb R)=\mathbb R^\times$ is $-1$. If local frames are changed by nonzero functions $\lambda_i$, then the transition functions change by
\begin{align*}
g'_{ij}=\lambda_i^{-1}g_{ij}\lambda_j.
\end{align*}
Multiplying the transformed transitions around the circle gives
\begin{align*}
(\lambda_0^{-1}g_{01}\lambda_1)(\lambda_1^{-1}g_{12}\lambda_2)\cdots(\lambda_k^{-1}g_{k0}\lambda_0)=\lambda_0^{-1}(g_{01}g_{12}\cdots g_{k0})\lambda_0.
\end{align*}
For the Mobius bundle, the original product is $g_{01}g_{12}\cdots g_{k0}=-1$, hence
\begin{align*}
\lambda_0^{-1}(g_{01}g_{12}\cdots g_{k0})\lambda_0=\lambda_0^{-1}(-1)\lambda_0=-1,
\end{align*}
because $GL_1(\mathbb R)$ is abelian. If all transformed transition functions took values in $\mathbb R_+$, their product would lie in $\mathbb R_+$, contradicting the product $-1$. Thus the Mobius bundle has no $GL_1^+(\mathbb R)$-reduction, and the obstruction is exactly the sign picked up after going once around the circle.
[/example]
The obstruction in the example is not caused by local geometry: every small interval admits a choice of positive frame. It comes from the failure of those local choices to agree after travelling around the base.
## Metrics and Orthogonal Reductions
Orientation reduces a frame bundle by keeping track of a component of the space of frames. A Riemannian metric reduces it more strongly: it singles out the frames in which the inner product has the standard Euclidean form. The stabiliser of the standard inner product on $\mathbb R^n$ is $O(n)$.
[definition: Orthogonal Frame Bundle]
Let $E\to M$ be a rank-$n$ real vector bundle with a fibre metric $g$. The orthogonal frame bundle $\operatorname{Fr}_O(E,g)$ is the set of frames $u:\mathbb R^n\to E_x$ such that
\begin{align*}
g_x(u(v),u(w))&=v\cdot w
\end{align*}
for all $v,w\in \mathbb R^n$.
[/definition]
The right action of $O(n)$ preserves this condition, while a general element of $GL_n(\mathbb R)$ changes the Gram matrix of the frame. The question is whether every $O(n)$-reduction arises from a unique fibre metric, not just whether a metric gives examples of such reductions.
[quotetheorem:6256]
[citeproof:6256]
For tangent bundles, this says that a Riemannian metric on $M$ is exactly an $O(n)$-reduction of $\operatorname{Fr}(TM)$. The theorem itself is an equivalence of data, not an existence theorem: it says what an $O(n)$-reduction means once one exists. Existence of Riemannian metrics on paracompact smooth manifolds is a separate result, proved by partitions of unity; without paracompactness, that construction can fail and global metrics need not follow from local positive-definite forms. The reduction language is still useful because it separates the algebraic stabiliser $O(n)$ from the analytic/topological question of whether a global section of the relevant quotient bundle exists.
[example: Constructing a Riemannian Metric]
Let $(U_i,\varphi_i)_{i\in I}$ be a locally finite atlas and let $(\rho_i)_{i\in I}$ be a smooth [partition of unity](/page/Partition%20of%20Unity) subordinate to it, so $\rho_i\ge 0$, $\operatorname{supp}(\rho_i)\subset U_i$, and $\sum_i\rho_i(x)=1$ for every $x\in M$. Write $\varphi_i=(\varphi_i^1,\dots,\varphi_i^n)$. On $U_i$, define the coordinate Euclidean metric by
\begin{align*}
h_{i,x}(v,w)=\sum_{a=1}^n d\varphi_i^a{}_x(v)d\varphi_i^a{}_x(w).
\end{align*}
Since $d\varphi_i{}_x:T_xM\to \mathbb R^n$ is a linear isomorphism, if $v\ne 0$, then $d\varphi_i{}_x(v)\ne 0$, and therefore
\begin{align*}
h_{i,x}(v,v)=\sum_{a=1}^n \bigl(d\varphi_i^a{}_x(v)\bigr)^2>0.
\end{align*}
Extend each tensor $\rho_i h_i$ by zero outside $U_i$, and define
\begin{align*}
g_x(v,w)=\sum_{i\in I}\rho_i(x)h_{i,x}(v,w).
\end{align*}
The sum has only finitely many nonzero terms near each point because the atlas is locally finite, so $g$ is smooth. It is symmetric because each $h_i$ is symmetric, and it is bilinear because each $h_i$ is bilinear and each $\rho_i(x)$ is a scalar. For $v\ne 0$,
\begin{align*}
g_x(v,v)=\sum_{i\in I}\rho_i(x)h_{i,x}(v,v).
\end{align*}
Every summand is nonnegative. Since $\sum_i\rho_i(x)=1$, some index $j$ satisfies $\rho_j(x)>0$, and for that index $h_{j,x}(v,v)>0$. Hence
\begin{align*}
g_x(v,v)\ge \rho_j(x)h_{j,x}(v,v)>0.
\end{align*}
Thus $g$ is a smooth positive-definite fibre metric on $TM$.
Its orthonormal frame bundle is
\begin{align*}
Q_x=\{u:\mathbb R^n\to T_xM:g_x(u(e_a),u(e_b))=\delta_{ab}\text{ for all }a,b\}.
\end{align*}
If $u\in Q_x$ and $A\in O(n)$, then the right action is $u\cdot A=u\circ A$. For the standard basis vectors,
\begin{align*}
(uA)(e_a)=u(Ae_a)=u\left(\sum_c A_{ca}e_c\right).
\end{align*}
Using bilinearity of $g_x$,
\begin{align*}
g_x((uA)(e_a),(uA)(e_b))=\sum_{c,d}A_{ca}A_{db}g_x(u(e_c),u(e_d)).
\end{align*}
Since $u$ is orthonormal,
\begin{align*}
\sum_{c,d}A_{ca}A_{db}g_x(u(e_c),u(e_d))=\sum_{c,d}A_{ca}A_{db}\delta_{cd}.
\end{align*}
The Kronecker delta removes all terms with $c\ne d$, so
\begin{align*}
\sum_{c,d}A_{ca}A_{db}\delta_{cd}=\sum_c A_{ca}A_{cb}.
\end{align*}
By matrix multiplication,
\begin{align*}
\sum_c A_{ca}A_{cb}=(A^\top A)_{ab}.
\end{align*}
Since $A\in O(n)$, $A^\top A=I_n$, so
\begin{align*}
g_x((uA)(e_a),(uA)(e_b))=\delta_{ab}.
\end{align*}
Therefore $Q$ is preserved by the right action of $O(n)$, and by *Metric as Orthogonal Reduction*, these orthonormal frames define an $O(n)$-reduction of $\operatorname{Fr}(TM)$.
[/example]
A metric can be combined with orientation. If $M$ is oriented and Riemannian, the frames that are both positive and orthonormal form an $SO(n)$-reduction of $\operatorname{Fr}(TM)$.
[remark: Combining Reductions]
An $SO(n)$-reduction of $\operatorname{Fr}(TM)$ is equivalent to a Riemannian metric together with an orientation. The subgroup $SO(n)=O(n)\cap GL_n^+(\mathbb R)$ preserves both the Euclidean inner product and the standard orientation of $\mathbb R^n$.
[/remark]
This pattern reappears below for volume and almost complex structures, and in Chapter 11 when connections are required to preserve a chosen reduction.
## Volume Forms and Special Linear Reductions
A volume form does not determine lengths or angles, but it does determine which frames have unit oriented volume. The stabiliser of the standard volume form on $\mathbb R^n$ is $SL_n(\mathbb R)$, or $SL_n^+(\mathbb R)$ if one separates orientation conventions through the connected component.
[definition: Volume Reduction]
Let $M$ be a smooth $n$-manifold. An $SL_n(\mathbb R)$-reduction of $\operatorname{Fr}(TM)$ is a principal $SL_n(\mathbb R)$-subbundle $Q\subset \operatorname{Fr}(TM)$.
[/definition]
Such a reduction chooses frames whose determinant, relative to the induced volume density, is fixed to be $1$. The next question is whether prescribing this unit-determinant frame subbundle is exactly the same thing as prescribing a nowhere-vanishing top-degree form.
[quotetheorem:6257]
[citeproof:6257]
This result separates volume from metric. The nowhere-vanishing hypothesis is essential: an $n$-form that vanishes at some point cannot define a unit-volume frame in that fibre, so it cannot produce an $SL_n(\mathbb R)$-subbundle there. A Riemannian metric gives a volume form after an orientation is chosen, but a volume form alone leaves a large residual symmetry group and does not determine lengths or angles. Topologically, the existence of such a form is tied to orientability, so this example continues the theme that reductions package global section problems rather than merely pointwise linear algebra.
[example: Standard Volume on a Torus]
Let $q:\mathbb R^n\to T^n=\mathbb R^n/\mathbb Z^n$ be the quotient map, and write
\begin{align*}
\Omega=dx_1\wedge\cdots\wedge dx_n
\end{align*}
on $\mathbb R^n$. For $m=(m_1,\dots,m_n)\in\mathbb Z^n$, the deck transformation $\tau_m(x)=x+m$ satisfies
\begin{align*}
\tau_m^*dx_i=d(x_i+m_i)=dx_i.
\end{align*}
Therefore
\begin{align*}
\tau_m^*\Omega=dx_1\wedge\cdots\wedge dx_n=\Omega.
\end{align*}
Since $\Omega$ is invariant under all deck transformations of $q$, it descends to a unique $n$-form $\omega$ on $T^n$ satisfying $q^*\omega=\Omega$. In the induced coordinate vector fields $\partial_1,\dots,\partial_n$, this form has
\begin{align*}
\omega(\partial_1,\dots,\partial_n)=1,
\end{align*}
so $\omega$ is nowhere vanishing.
Now let $u:\mathbb R^n\to T_{[x]}T^n$ be a frame, and write its coordinate matrix as $A=(A_{ij})$, meaning
\begin{align*}
u(e_j)=\sum_{i=1}^n A_{ij}\partial_i.
\end{align*}
Using multilinearity of $\omega$ in all $n$ entries gives
\begin{align*}
\omega(u(e_1),\dots,u(e_n))=\sum_{i_1,\dots,i_n}A_{i_1 1}\cdots A_{i_n n}\,\omega(\partial_{i_1},\dots,\partial_{i_n}).
\end{align*}
By alternation, every term with two equal indices is zero. Hence only permutations of $1,\dots,n$ contribute:
\begin{align*}
\omega(u(e_1),\dots,u(e_n))=\sum_{\sigma\in S_n}A_{\sigma(1)1}\cdots A_{\sigma(n)n}\,\omega(\partial_{\sigma(1)},\dots,\partial_{\sigma(n)}).
\end{align*}
Since $\omega(\partial_{\sigma(1)},\dots,\partial_{\sigma(n)})=\operatorname{sgn}(\sigma)$, this becomes
\begin{align*}
\omega(u(e_1),\dots,u(e_n))=\sum_{\sigma\in S_n}\operatorname{sgn}(\sigma)A_{\sigma(1)1}\cdots A_{\sigma(n)n}.
\end{align*}
By the Leibniz formula for the determinant,
\begin{align*}
\omega(u(e_1),\dots,u(e_n))=\det A.
\end{align*}
Thus the reduction associated to $\omega$ by *Volume Forms as Special Linear Reductions* is
\begin{align*}
Q_{[x]}=\{u:\mathbb R^n\to T_{[x]}T^n:\det A=1\}.
\end{align*}
If $B\in SL_n(\mathbb R)$, then the coordinate matrix of the right-translated frame $u\circ B$ is $AB$, and
\begin{align*}
\det(AB)=\det(A)\det(B)=\det(A).
\end{align*}
Therefore the determinant-one condition is exactly preserved by the right $SL_n(\mathbb R)$-action. This reduction fixes unit oriented volume in each tangent space; for $n\ge 2$, it still does not choose lengths or angles, since different positive-definite inner products can induce the same determinant-one volume form.
[/example]
The contrast between $O(n)$ and $SL_n(\mathbb R)$ reductions is a useful test case: different subgroups encode different parts of linear geometry.
## Complex and Almost Complex Structures
Some reductions arise not from tensors of a fixed type, but from changing the permitted linear algebra on each fibre. A complex structure on a real rank-$2n$ vector bundle reduces the structure group from $GL_{2n}(\mathbb R)$ to $GL_n(\mathbb C)$, while a Hermitian metric on a complex vector bundle reduces from $GL_n(\mathbb C)$ to $U(n)$.
[definition: Almost Complex Structure on a Vector Bundle]
Let $E\to M$ be a real vector bundle. An almost complex structure on $E$ is a smooth bundle endomorphism $J:E\to E$ such that
\begin{align*}
J^2&=-\operatorname{id}_E.
\end{align*}
[/definition]
The equation $J^2=-\operatorname{id}_E$ makes each fibre into a complex vector space by defining multiplication by $i$ to be $J$. The question is whether this fibrewise complex linear algebra is exactly the same as reducing real frames to complex-linear frames.
[quotetheorem:6258]
[citeproof:6258]
For tangent bundles, an almost complex structure on $M$ is therefore a reduction of $\operatorname{Fr}(TM)$ to $GL_n(\mathbb C)$. The rank-$2n$ hypothesis is necessary because a real vector space carrying an endomorphism $J$ with $J^2=-\operatorname{id}$ must have even real dimension. The theorem is also only about almost complex structures: integrability, equivalently the existence of compatible holomorphic coordinate charts, is a separate differential condition and is not encoded by the reduction alone. Thus a $GL_n(\mathbb C)$-reduction removes the pointwise linear obstruction, while further topology and analysis determine whether the structure comes from a genuine complex manifold.
[example: Standard Almost Complex Structure on Complex Space]
View $\mathbb C^n$ as the real vector space $\mathbb R^{2n}$ with real basis $e_1,f_1,\dots,e_n,f_n$, where $e_k$ is the $x_k$-direction and $f_k$ is the $y_k$-direction for $z_k=x_k+iy_k$. Multiplication by $i$ is the real linear map $J$ determined by
\begin{align*}
J(e_k)=f_k
\end{align*}
\begin{align*}
J(f_k)=-e_k.
\end{align*}
Applying $J$ again gives
\begin{align*}
J^2(e_k)=J(f_k)=-e_k
\end{align*}
\begin{align*}
J^2(f_k)=J(-e_k)=-f_k.
\end{align*}
Since $e_1,f_1,\dots,e_n,f_n$ is a real basis, $J^2=-\operatorname{id}_{\mathbb R^{2n}}$.
A real frame $u:\mathbb R^{2n}\to \mathbb C^n$ is adapted to this almost complex structure exactly when it intertwines the standard complex structure on the domain with multiplication by $i$ on the target:
\begin{align*}
u(Jv)=J(u(v))\text{ for all }v\in\mathbb R^{2n}.
\end{align*}
Writing both complex structures as multiplication by $i$, this is
\begin{align*}
u(iv)=i\,u(v)\text{ for all }v\in\mathbb C^n.
\end{align*}
Together with real linearity, this implies complex linearity. For $a,b,c,d\in\mathbb R$ and $v,w\in\mathbb C^n$,
\begin{align*}
u((a+ib)v+(c+id)w)=u(av+biv+cw+diw).
\end{align*}
By real linearity,
\begin{align*}
u(av+biv+cw+diw)=a\,u(v)+b\,u(iv)+c\,u(w)+d\,u(iw).
\end{align*}
Using $u(iv)=i\,u(v)$ and $u(iw)=i\,u(w)$,
\begin{align*}
a\,u(v)+b\,u(iv)+c\,u(w)+d\,u(iw)=(a+ib)u(v)+(c+id)u(w).
\end{align*}
Thus $u$ is complex-linear. Conversely, every complex-linear isomorphism satisfies
\begin{align*}
u(iv)=i\,u(v)
\end{align*}
by taking the complex scalar to be $i$. Hence the adapted real frames are precisely the complex bases of $\mathbb C^n$.
If $u$ is adapted and $A\in GL_n(\mathbb C)$, then the right-translated frame is $u\circ A$. For every $v\in\mathbb C^n$,
\begin{align*}
(u\circ A)(iv)=u(A(iv)).
\end{align*}
Since $A$ is complex-linear, $A(iv)=iA(v)$, so
\begin{align*}
u(A(iv))=u(iA(v)).
\end{align*}
Since $u$ is adapted,
\begin{align*}
u(iA(v))=i\,u(A(v))=i\,(u\circ A)(v).
\end{align*}
Thus $u\circ A$ is adapted.
If $u$ and $u'$ are two adapted frames, set $A=u^{-1}\circ u'$. For every $v\in\mathbb C^n$,
\begin{align*}
A(iv)=u^{-1}(u'(iv)).
\end{align*}
Since $u'$ is adapted, $u'(iv)=i\,u'(v)$, so
\begin{align*}
A(iv)=u^{-1}(i\,u'(v)).
\end{align*}
Since $u$ is adapted, $u^{-1}$ is also complex-linear, and therefore
\begin{align*}
u^{-1}(i\,u'(v))=i\,u^{-1}(u'(v))=iA(v).
\end{align*}
So $A\in GL_n(\mathbb C)$. The adapted frames therefore form a principal $GL_n(\mathbb C)$-subbundle of the real frame bundle. This reduction keeps exactly the complex-linear changes of frame; no inner product has been chosen, so it does not distinguish unitary frames, lengths, or angles.
[/example]
The example separates complex linear structure from metric structure. To impose lengths and angles compatible with complex scalar multiplication, we need a smaller subgroup: the unitary group preserving the standard Hermitian inner product.
[definition: Unitary Frame Bundle]
Let $E\to M$ be a rank-$n$ complex vector bundle with Hermitian metric $h$. The unitary frame bundle $\operatorname{Fr}_U(E,h)$ consists of complex-linear frames $u:\mathbb C^n\to E_x$ satisfying
\begin{align*}
h_x(u(v),u(w))&=v\cdot \overline{w}
\end{align*}
for all $v,w\in \mathbb C^n$.
[/definition]
The right action of $U(n)$ preserves this Hermitian orthonormality condition, while general complex frame changes do not. The remaining question is whether every $U(n)$-reduction arises from a Hermitian metric, matching the real orthogonal story.
[quotetheorem:6259]
[citeproof:6259]
The theorem proves that unitary reductions are not additional mysterious data; they are Hermitian metrics written in principal-bundle language. As in the real case, the theorem is an equivalence of structures, while existence over a paracompact base is supplied by a separate partition-of-unity construction. Without that global construction, local Hermitian metrics do not automatically determine a global reduction. This unitary example also points toward characteristic classes: once a complex bundle is reduced to $U(n)$, Chern-Weil theory later extracts topological invariants from compatible connections.
[example: Reducing a Complex Vector Bundle from GL_n(C) to U(n)]
Let $E\to M$ be a rank-$n$ complex vector bundle over a paracompact smooth manifold. Choose a locally finite trivialising cover $(U_i)_{i\in I}$, a smooth partition of unity $(\rho_i)_{i\in I}$ subordinate to it, and complex-linear trivialisations
\begin{align*}
\Phi_i:E|_{U_i}\longrightarrow U_i\times \mathbb C^n.
\end{align*}
Write $\Phi_{i,x}:E_x\to \mathbb C^n$ for the fibre map. On $U_i$, define
\begin{align*}
h_{i,x}(v,w)=\Phi_{i,x}(v)\cdot \overline{\Phi_{i,x}(w)}.
\end{align*}
Since $\Phi_{i,x}$ is complex-linear and the standard form on $\mathbb C^n$ is Hermitian, each $h_i$ is Hermitian. If $v\ne 0$, then $\Phi_{i,x}(v)\ne 0$, so
\begin{align*}
h_{i,x}(v,v)=\sum_{a=1}^n |\Phi_{i,x}(v)_a|^2>0.
\end{align*}
Extend each tensor $\rho_i h_i$ by zero outside $U_i$, and define
\begin{align*}
h_x(v,w)=\sum_{i\in I}\rho_i(x)h_{i,x}(v,w).
\end{align*}
The cover is locally finite, so near each point only finitely many summands are nonzero; hence $h$ is smooth. For $\lambda,\mu\in\mathbb C$,
\begin{align*}
h_x(\lambda v+\mu v',w)=\sum_i \rho_i(x)h_{i,x}(\lambda v+\mu v',w).
\end{align*}
Because each $h_i$ is complex-linear in the first variable,
\begin{align*}
\sum_i \rho_i(x)h_{i,x}(\lambda v+\mu v',w)=\sum_i \rho_i(x)\bigl(\lambda h_{i,x}(v,w)+\mu h_{i,x}(v',w)\bigr).
\end{align*}
Pulling the scalars out of the finite sum gives
\begin{align*}
h_x(\lambda v+\mu v',w)=\lambda h_x(v,w)+\mu h_x(v',w).
\end{align*}
Similarly, conjugate-linearity in the second variable gives
\begin{align*}
h_x(v,\lambda w+\mu w')=\overline{\lambda}\,h_x(v,w)+\overline{\mu}\,h_x(v,w').
\end{align*}
Hermitian symmetry follows from
\begin{align*}
h_x(w,v)=\sum_i\rho_i(x)h_{i,x}(w,v)=\sum_i\rho_i(x)\overline{h_{i,x}(v,w)}=\overline{h_x(v,w)}.
\end{align*}
If $v\ne 0$, then every term $\rho_i(x)h_{i,x}(v,v)$ is nonnegative. Since $\sum_i\rho_i(x)=1$, some $j$ satisfies $\rho_j(x)>0$, and therefore
\begin{align*}
h_x(v,v)=\sum_i\rho_i(x)h_{i,x}(v,v)\ge \rho_j(x)h_{j,x}(v,v)>0.
\end{align*}
Thus $h$ is a global Hermitian metric on $E$.
The unitary frame bundle of $h$ has fibres
\begin{align*}
Q_x=\{u:\mathbb C^n\to E_x:h_x(u(e_a),u(e_b))=\delta_{ab}\text{ for all }a,b\}.
\end{align*}
If $u\in Q_x$ and $A\in U(n)$, the right action is $u\cdot A=u\circ A$. Since
\begin{align*}
Ae_a=\sum_c A_{ca}e_c,
\end{align*}
we have
\begin{align*}
(uA)(e_a)=u(Ae_a)=\sum_c A_{ca}u(e_c).
\end{align*}
Using linearity in the first variable and conjugate-linearity in the second,
\begin{align*}
h_x((uA)(e_a),(uA)(e_b))=\sum_{c,d}A_{ca}\overline{A_{db}}\,h_x(u(e_c),u(e_d)).
\end{align*}
Since $u$ is unitary, this becomes
\begin{align*}
h_x((uA)(e_a),(uA)(e_b))=\sum_{c,d}A_{ca}\overline{A_{db}}\,\delta_{cd}.
\end{align*}
The Kronecker delta removes the terms with $c\ne d$, so
\begin{align*}
h_x((uA)(e_a),(uA)(e_b))=\sum_c A_{ca}\overline{A_{cb}}.
\end{align*}
Because $A\in U(n)$, its columns are orthonormal for the standard Hermitian form, hence
\begin{align*}
\sum_c A_{ca}\overline{A_{cb}}=\delta_{ab}.
\end{align*}
Therefore $u\circ A\in Q_x$, so the unitary frames are preserved by the right $U(n)$-action. By *Hermitian Metric as Unitary Reduction*, $Q$ is a $U(n)$-reduction of the complex frame bundle. Hence every complex vector bundle over a paracompact smooth manifold admits a unitary reduction.
[/example]
This final example matches the real metric case: compact subgroups such as $O(n)$ and $U(n)$ often arise as structure groups after choosing a metric. Later, connections on reduced bundles will let us ask whether a connection preserves the reduced structure, leading to metric-compatible and unitary connections.
Reductions isolate the geometric content hidden inside a larger symmetry group, but they do not yet say how to compare fibres smoothly. The next chapter introduces principal connections, the horizontal distributions that let us separate motion along the fibres from motion across the base in a compatible way.
# 6. Connections on Principal Bundles
A principal bundle already separates directions tangent to the fibres from directions that move across the base. This chapter asks how to choose, in a smooth and symmetry-compatible way, which tangent vectors on the total space should count as horizontal. The resulting object is a principal connection, and it is the device that turns the local gauge fields of Chapter 9, the frame-bundle connections of Chapter 11, and the monopole potentials computed in Chapter 12 into different presentations of the same geometry.
The progression is from geometry to algebra and then back to local coordinates. We first define a connection as a $G$-equivariant horizontal distribution complementary to the vertical tangent spaces. We then encode the same data by a Lie-algebra-valued $1$-form on the principal bundle, and finally compute the local gauge potentials seen by a section of the bundle.
## Horizontal Distributions and Vertical Tangent Spaces
The first problem is that the tangent bundle $TP$ of a principal bundle $\nu:P\to M$ contains directions that do not move the base point at all. To differentiate along $M$, or to lift paths from $M$ to $P$, we need a principled complement to those vertical directions.
[definition: Vertical Tangent Space]
Let $\nu:P\to M$ be a smooth principal $G$-bundle. For $p\in P$, the vertical tangent space at $p$ is
\begin{align*}
V_pP:=\ker(d\nu_p)\subset T_pP.
\end{align*}
The vertical tangent bundle is $VP:=\ker(d\nu)\subset TP$.
[/definition]
The vertical tangent space is intrinsic: it is the tangent space to the fibre through $p$. To compare vertical vectors at different points of the same fibre, we need a uniform language that does not depend on a chosen chart. The principal action supplies this language, because every fibre direction comes from infinitesimal motion in the structure group.
[definition: Fundamental Vector Field]
Let $P$ be a right principal $G$-bundle for a Lie group $G$, and let $\xi\in\mathfrak g$. For each $p\in P$, define the orbit curve $c_{p,\xi}:\mathbb R\to P$ by
\begin{align*}
c_{p,\xi}(t):=p\cdot\exp(t\xi).
\end{align*}
The fundamental vector field generated by $\xi$ is the vector field $\xi_P\in\mathfrak X(P)$ defined by
\begin{align*}
(\xi_P)_p:=\dot c_{p,\xi}(0)=\frac{d}{dt}\Big|_{t=0}p\cdot\exp(t\xi).
\end{align*}
[/definition]
This construction identifies an element of $\mathfrak g$ with a vertical vector at every point of $P$. The next question is whether this identification misses any vertical vectors, or whether two different Lie algebra elements can generate the same one. The answer is the fibrewise infinitesimal form of freeness and transitivity of the principal action.
[quotetheorem:6260]
[citeproof:6260]
The theorem says that no vertical direction is mysterious: it is generated by a unique Lie algebra element. Both parts of the principal-bundle hypothesis are being used here. If a group action on a fibre were not free, a nonzero $\xi\in\mathfrak g$ could generate the zero vertical vector at a point, so the map $\mathfrak g\to V_pP$ would fail to be injective. If the action were not transitive on the fibre, infinitesimal group motion would only see the tangent directions along a smaller orbit, so it would fail to reach all of $V_pP$.
The identification is also fibrewise rather than a choice of coordinates on $P$: it labels vertical vectors by Lie algebra elements at a fixed point $p$, but it does not by itself choose any complement to $V_pP$ inside $T_pP$. A connection now amounts to adding horizontal directions in a way that respects the principal symmetry.
[definition: Principal Connection as Horizontal Distribution]
Let $\nu:P\to M$ be a smooth principal $G$-bundle. A principal connection on $P$ is a smooth subbundle $HP\subset TP$ such that, for every $p\in P$,
\begin{align*}
T_pP=H_pP\oplus V_pP,
\end{align*}
and for every $g\in G$,
\begin{align*}
dR_g(H_pP)=H_{p\cdot g}P,
\end{align*}
where $R_g:P\to P$ is right multiplication by $g$.
[/definition]
The complement condition makes $d\nu_p:H_pP\to T_{\nu(p)}M$ an isomorphism, so base tangent vectors have unique horizontal lifts once a connection has been chosen. The equivariance condition is what makes the construction principal rather than merely a choice of complement on the total space.
[example: Product Horizontal Distribution]
Let $P=M\times G$ be the product right principal $G$-bundle, with projection $\nu(x,h)=x$ and right action $(x,h)\cdot g=(x,hg)$. Under the standard tangent identification
\begin{align*}
T_{(x,h)}P=T_xM\oplus T_hG,
\end{align*}
the differential of the projection is
\begin{align*}
d\nu_{(x,h)}(X,Y)=X,
\end{align*}
so
\begin{align*}
V_{(x,h)}P
=\ker(d\nu_{(x,h)})
=\{(X,Y)\in T_xM\oplus T_hG:X=0\}
=\{0\}\times T_hG.
\end{align*}
Define
\begin{align*}
H_{(x,h)}P:=T_xM\times\{0\}.
\end{align*}
Then every tangent vector $(X,Y)\in T_xM\oplus T_hG$ decomposes as
\begin{align*}
(X,Y)=(X,0)+(0,Y),
\end{align*}
with $(X,0)\in H_{(x,h)}P$ and $(0,Y)\in V_{(x,h)}P$. If $(X,0)=(0,Y)$, then comparing the two components gives $X=0$ and $Y=0$, hence
\begin{align*}
T_{(x,h)}P=H_{(x,h)}P\oplus V_{(x,h)}P.
\end{align*}
It remains to check compatibility with the right action. For fixed $g\in G$, right multiplication is
\begin{align*}
R_g(x,h)=(x,hg).
\end{align*}
Its differential sends
\begin{align*}
d(R_g)_{(x,h)}(X,Y)=(X,d(r_g)_hY),
\end{align*}
where $r_g:G\to G$ is right multiplication by $g$. Therefore, for a horizontal vector $(X,0)$,
\begin{align*}
d(R_g)_{(x,h)}(X,0)=(X,d(r_g)_h0)=(X,0)\in T_xM\times\{0\}=H_{(x,hg)}P.
\end{align*}
Conversely, every vector $(X,0)\in H_{(x,hg)}P$ is the image of $(X,0)\in H_{(x,h)}P$, so
\begin{align*}
dR_g(H_{(x,h)}P)=H_{(x,hg)}P=H_{(x,h)\cdot g}P.
\end{align*}
Thus the product horizontal spaces give a principal connection: horizontal motion changes only the base point, while vertical motion changes only the group coordinate.
[/example]
This example is the flat reference model. More general connections on a product bundle tilt the horizontal subspaces by a Lie-algebra-valued form on the base, which becomes the local gauge potential introduced below in this chapter.
## Connection Forms on the Total Space
The distribution definition is geometric, but computations with curvature and gauge transformations are better expressed by a form. The question is how to encode the vertical projection associated to $TP=HP\oplus VP$ without choosing coordinates on the fibres.
[definition: Connection One Form]
Let $\nu:P\to M$ be a smooth principal $G$-bundle. A connection $1$-form is a form $\omega\in\Omega^1(P;\mathfrak g)$ such that:
\begin{align*}
\omega(\xi_P)=\xi \quad \text{for all }\xi\in\mathfrak g,
\end{align*}
and
\begin{align*}
(R_g)^*\omega=\operatorname{Ad}(g^{-1})\omega \quad \text{for all }g\in G.
\end{align*}
[/definition]
The first condition says that $\omega$ reproduces the infinitesimal generator of a vertical vector. The second says that changing the point in the same principal fibre transforms the Lie algebra label by the adjoint action. The next theorem is needed because it proves that this form is not extra structure: it is exactly the horizontal distribution expressed as a vertical projection.
[quotetheorem:6261]
[citeproof:6261]
The theorem lets us move between pictures depending on the calculation. The complement hypothesis is essential: if $HP$ met $VP$ nontrivially, then a tangent vector could have more than one vertical part, so the formula defining $\omega$ from vertical projection would not be well-defined. If $HP$ failed to span together with $VP$, then some tangent vectors would have no horizontal-plus-vertical decomposition, so no connection form could be recovered from the distribution.
The equivariance and reproduction axioms are equally structural. Without reproduction, $\ker\omega_p$ might intersect the vertical space nontrivially and therefore would not define a horizontal complement. Without equivariance, the kernels $\ker\omega_p$ would not be carried to $\ker\omega_{p\cdot g}$ by right translation, so the resulting connection would depend on a representative point in the fibre rather than on the principal geometry. What the bijection does not assert is that a connection is unique or canonical; it says only that once a compatible horizontal distribution or a compatible $1$-form has been chosen, the other description is determined.
Horizontal lifts are most natural from $HP$, while local gauge formulas and curvature are most compact in terms of $\omega$.
[remark: Transformation of Fundamental Fields]
For a right principal action, the fundamental fields satisfy
\begin{align*}
dR_g((\xi_P)_p)=((\operatorname{Ad}(g^{-1})\xi)_P)_{p\cdot g}.
\end{align*}
This inverse adjoint is the source of the inverse in the equivariance law $(R_g)^*\omega=\operatorname{Ad}(g^{-1})\omega$.
[/remark]
The next example shows how the familiar expression $d+A$ appears from a connection form. Here $d$ is the ordinary exterior derivative on the base, while $A$ is the additional Lie-algebra-valued term describing the horizontal tilt.
[example: Product Bundle Connection]
Let $P=M\times G$, let $A\in\Omega^1(M;\mathfrak g)$, and write $\theta\in\Omega^1(G;\mathfrak g)$ for the left Maurer-Cartan form, so $\theta_h(Y)=d(L_{h^{-1}})_hY$ for $Y\in T_hG$. Define $\omega\in\Omega^1(M\times G;\mathfrak g)$ by
\begin{align*}
\omega_{(x,h)}(X,Y)=\operatorname{Ad}(h^{-1})A_x(X)+\theta_h(Y)
\end{align*}
for $(X,Y)\in T_xM\oplus T_hG$.
We verify the two connection-form axioms. For $\xi\in\mathfrak g$, the fundamental vector at $(x,h)$ is represented by the curve $t\mapsto (x,h\exp(t\xi))$, hence
\begin{align*}
(\xi_P)_{(x,h)}=\left(0,\frac{d}{dt}\Big|_{t=0}h\exp(t\xi)\right)
\end{align*}
and therefore
\begin{align*}
(\xi_P)_{(x,h)}=(0,d(L_h)_e\xi).
\end{align*}
Applying $\omega$ gives
\begin{align*}
\omega_{(x,h)}((\xi_P)_{(x,h)})=\operatorname{Ad}(h^{-1})A_x(0)+\theta_h(d(L_h)_e\xi).
\end{align*}
Since $A_x$ is linear, $A_x(0)=0$, and by the definition of $\theta$,
\begin{align*}
\theta_h(d(L_h)_e\xi)=d(L_{h^{-1}})_h d(L_h)_e\xi=\xi.
\end{align*}
Thus
\begin{align*}
\omega_{(x,h)}((\xi_P)_{(x,h)})=\xi.
\end{align*}
Now fix $g\in G$. Right multiplication on $P$ is $R_g(x,h)=(x,hg)$, so
\begin{align*}
d(R_g)_{(x,h)}(X,Y)=(X,d(r_g)_hY),
\end{align*}
where $r_g:G\to G$ is right multiplication by $g$. Hence
\begin{align*}
((R_g)^*\omega)_{(x,h)}(X,Y)=\omega_{(x,hg)}(X,d(r_g)_hY).
\end{align*}
Using the definition of $\omega$ at $(x,hg)$, this becomes
\begin{align*}
((R_g)^*\omega)_{(x,h)}(X,Y)=\operatorname{Ad}((hg)^{-1})A_x(X)+\theta_{hg}(d(r_g)_hY).
\end{align*}
Because $(hg)^{-1}=g^{-1}h^{-1}$, the first term is
\begin{align*}
\operatorname{Ad}((hg)^{-1})A_x(X)=\operatorname{Ad}(g^{-1})\operatorname{Ad}(h^{-1})A_x(X).
\end{align*}
For the Maurer-Cartan term,
\begin{align*}
\theta_{hg}(d(r_g)_hY)=d(L_{(hg)^{-1}})_{hg}d(r_g)_hY.
\end{align*}
Since $L_{(hg)^{-1}}\circ r_g=L_{g^{-1}h^{-1}}\circ r_g=c_{g^{-1}}\circ L_{h^{-1}}$, where $c_{g^{-1}}(k)=g^{-1}kg$, differentiating gives
\begin{align*}
d(L_{(hg)^{-1}})_{hg}d(r_g)_hY=d(c_{g^{-1}})_e d(L_{h^{-1}})_hY.
\end{align*}
The differential of conjugation at the identity is the adjoint action, so
\begin{align*}
\theta_{hg}(d(r_g)_hY)=\operatorname{Ad}(g^{-1})\theta_h(Y).
\end{align*}
Combining the two terms gives
\begin{align*}
((R_g)^*\omega)_{(x,h)}(X,Y)=\operatorname{Ad}(g^{-1})\omega_{(x,h)}(X,Y).
\end{align*}
Thus $(R_g)^*\omega=\operatorname{Ad}(g^{-1})\omega$, so $\omega$ is a connection $1$-form.
Along the identity section $s(x)=(x,e)$, we have $ds_x(X)=(X,0)$. Therefore
\begin{align*}
(s^*\omega)_x(X)=\omega_{(x,e)}(X,0).
\end{align*}
Substituting $h=e$ gives
\begin{align*}
(s^*\omega)_x(X)=\operatorname{Ad}(e^{-1})A_x(X)+\theta_e(0).
\end{align*}
Since $\operatorname{Ad}(e^{-1})$ is the identity and $\theta_e(0)=0$,
\begin{align*}
(s^*\omega)_x(X)=A_x(X).
\end{align*}
So $s^*\omega=A$: in the identity gauge, the local connection term added to the ordinary exterior derivative is exactly $A$, which is why the covariant derivative is written schematically as $d+A$.
[/example]
This formula already contains the two ingredients that every local gauge potential has: the pulled-back base form and the Maurer-Cartan term from motion in the group direction. The global form $\omega$ is invariantly defined on $P$, while $A$ depends on the chosen local section.
## Local Gauge Potentials and Change of Gauge
A principal connection lives on $P$, but most calculations are made over an open set $U_i\subset M$ after choosing a local section. The local section converts the global connection form into an ordinary $\mathfrak g$-valued $1$-form on $U_i$, and changing the section gives the gauge transformation law.
[definition: Local Gauge Potential]
Let $\nu:P\to M$ be a smooth principal $G$-bundle with connection form $\omega$. If $s_i:U_i\to P$ is a smooth local section, the local gauge potential determined by $s_i$ is
\begin{align*}
A_i:=s_i^*\omega\in\Omega^1(U_i;\mathfrak g).
\end{align*}
[/definition]
The potential $A_i$ is not a new connection on its own; it is the connection form seen through the local gauge $s_i$. On overlaps, two gauges differ by a transition function, and a global connection should give compatible local formulas. The next theorem is needed to identify the exact compatibility rule, including the vertical term produced by differentiating the transition function.
[quotetheorem:6262]
[citeproof:6262]
The formula says that gauge potentials do not patch as ordinary $1$-forms. The stated relation between the sections is essential: if $s_i$ and $s_j$ are not related by a smooth transition function on the overlap, there is no single $g_{ij}$ whose differential can supply the correction term. The connection-form axioms are also used exactly where the two terms appear: without equivariance the right-translated $ds_i$ term would not become $\operatorname{Ad}(g_{ij}^{-1})A_i$, and without reproduction the vertical derivative of $g_{ij}$ would not become the Maurer-Cartan term.
Thus the theorem is not saying that $A_i$ is globally defined, or that any arbitrary collection of local $\mathfrak g$-valued $1$-forms patches to a connection. It gives the compatibility law forced by a global connection and chosen local sections: an affine rule consisting of a homogeneous adjoint part plus an inhomogeneous Maurer-Cartan correction.
[example: Abelian Gauge Transformation]
On a simply connected overlap $U_i\cap U_j$, write the transition function as $g_{ij}=e^{i\chi_{ij}}$, where $\chi_{ij}:U_i\cap U_j\to\mathbb R$ is smooth. Since $U(1)$ is abelian, conjugation by every element of $U(1)$ is the identity, so its differential at the identity is the identity on $\mathfrak u(1)$. Hence $\operatorname{Ad}(g_{ij}^{-1})=\operatorname{id}_{\mathfrak u(1)}$.
For $X\in T_x(U_i\cap U_j)$, choose a curve $\gamma(t)$ in the overlap with $\gamma(0)=x$ and $\dot\gamma(0)=X$. Then
\begin{align*}
(dg_{ij})_x(X)=\frac{d}{dt}\Big|_{t=0}e^{i\chi_{ij}(\gamma(t))}.
\end{align*}
By the ordinary chain rule applied to $t\mapsto e^{i\chi_{ij}(\gamma(t))}$,
\begin{align*}
(dg_{ij})_x(X)=e^{i\chi_{ij}(x)}\,i\,\frac{d}{dt}\Big|_{t=0}\chi_{ij}(\gamma(t)).
\end{align*}
Since $\frac{d}{dt}\big|_{t=0}\chi_{ij}(\gamma(t))=(d\chi_{ij})_x(X)$ and $g_{ij}(x)=e^{i\chi_{ij}(x)}$, this becomes
\begin{align*}
(dg_{ij})_x(X)=g_{ij}(x)\,i\,(d\chi_{ij})_x(X).
\end{align*}
Multiplying by $g_{ij}(x)^{-1}=e^{-i\chi_{ij}(x)}$ gives
\begin{align*}
(g_{ij}^{-1}dg_{ij})_x(X)=e^{-i\chi_{ij}(x)}e^{i\chi_{ij}(x)}\,i\,(d\chi_{ij})_x(X).
\end{align*}
Since $e^{-i\chi_{ij}(x)}e^{i\chi_{ij}(x)}=1$, we get
\begin{align*}
(g_{ij}^{-1}dg_{ij})_x(X)=i\,(d\chi_{ij})_x(X).
\end{align*}
This holds for every tangent vector $X$, so
\begin{align*}
g_{ij}^{-1}dg_{ij}=i\,d\chi_{ij}.
\end{align*}
Substituting this identity and $\operatorname{Ad}(g_{ij}^{-1})=\operatorname{id}_{\mathfrak u(1)}$ into the *Gauge Transformation Law for Local Potentials* gives
\begin{align*}
A_j=A_i+i\,d\chi_{ij}.
\end{align*}
Thus in the abelian case a change of gauge adds the exact $\mathfrak u(1)$-valued $1$-form $i\,d\chi_{ij}$, which is the electromagnetic gauge transformation in differential-form notation.
[/example]
This abelian case is a useful check on signs and conventions. In nonabelian gauge theory the adjoint term cannot be dropped, so changing gauge also conjugates the Lie-algebra component of the potential.
## Frame Bundles and the Levi-Civita Connection
The general definition is designed to include connections arising from Riemannian geometry. The guiding question is: where does the Levi-Civita covariant derivative live when expressed as principal-bundle geometry?
[definition: Orthonormal Frame Bundle]
Let $(M,g)$ be an $n$-dimensional Riemannian manifold. The orthonormal frame bundle $F_{\mathrm{SO}}(M)$ is the principal $SO(n)$-bundle whose fibre at $x\in M$ consists of oriented linear isometries $u:\mathbb R^n\to T_xM$. The right action is change of frame by composition with elements of $SO(n)$.
[/definition]
Here $u$ denotes a frame, while $\pi:F_{\mathrm{SO}}(M)\to M$ denotes the bundle projection. A Riemannian connection gives a way to move frames along curves while preserving orthonormality.
[example: Levi-Civita Connection on the Frame Bundle]
Let $(M,g)$ be an oriented $n$-dimensional Riemannian manifold with Levi-Civita connection $\nabla$, and let $u:\mathbb R^n\to T_xM$ be an oriented orthonormal frame. For a smooth curve of frames $u(t)$ with $u(0)=u$ and base curve $\gamma(t)=\pi(u(t))$, write
\begin{align*}
E_a(t):=u(t)e_a,\qquad a=1,\dots,n.
\end{align*}
We declare $\dot u(0)$ to be horizontal when
\begin{align*}
\nabla_{\dot\gamma(t)}E_a(t)=0
\end{align*}
at $t=0$ for every $a$. Thus a horizontal tangent vector is exactly an infinitesimal motion of the frame in which all frame vectors are parallel along the base curve.
For each $X\in T_xM$, choose a curve $\gamma$ with $\gamma(0)=x$ and $\dot\gamma(0)=X$. Parallel transport along $\gamma$ gives unique vector fields $E_a(t)$ satisfying
\begin{align*}
E_a(0)=u e_a,\qquad \nabla_{\dot\gamma(t)}E_a(t)=0.
\end{align*}
Since $\nabla$ is metric-compatible,
\begin{align*}
\frac{d}{dt}g(E_a(t),E_b(t))
=g(\nabla_{\dot\gamma(t)}E_a(t),E_b(t))
+g(E_a(t),\nabla_{\dot\gamma(t)}E_b(t))
=0+0=0,
\end{align*}
so the vectors $E_1(t),\dots,E_n(t)$ remain orthonormal. Parallel transport is orientation-preserving along a small interval because it varies continuously and starts in the oriented component. Hence $u_X(t)e_a:=E_a(t)$ is a curve in $F_{\mathrm{SO}}(M)$, and $\dot u_X(0)$ is horizontal with
\begin{align*}
d\pi_u(\dot u_X(0))=\dot\gamma(0)=X.
\end{align*}
If a horizontal vector satisfies $d\pi_u(\dot u(0))=0$, then its base curve has $\dot\gamma(0)=0$, and the condition $\nabla_{\dot\gamma}E_a=0$ at $0$ gives $\dot E_a(0)=0$ in the fixed vector space $T_xM$ for every $a$. Therefore $\dot u(0)=0$. Thus $d\pi_u$ identifies the horizontal space with $T_xM$, and the horizontal space intersects the vertical space only in $\{0\}$.
Since $F_{\mathrm{SO}}(M)$ has fibre $SO(n)$, the vertical tangent space at $u$ has dimension $\dim SO(n)=n(n-1)/2$, while the horizontal space has dimension $\dim M=n$. Therefore
\begin{align*}
\dim H_uF_{\mathrm{SO}}(M)+\dim V_uF_{\mathrm{SO}}(M)
=n+\frac{n(n-1)}{2}
=\frac{n(n+1)}{2}
=\dim F_{\mathrm{SO}}(M),
\end{align*}
and the zero intersection gives
\begin{align*}
T_uF_{\mathrm{SO}}(M)=H_uF_{\mathrm{SO}}(M)\oplus V_uF_{\mathrm{SO}}(M).
\end{align*}
The construction is $SO(n)$-equivariant. If $A\in SO(n)$ and $u(t)$ is horizontal, then
\begin{align*}
(u(t)\cdot A)e_a=u(t)(Ae_a)=\sum_{b=1}^n A_{ba}\,u(t)e_b.
\end{align*}
The coefficients $A_{ba}$ are constant, so
\begin{align*}
\nabla_{\dot\gamma(t)}\bigl((u(t)\cdot A)e_a\bigr)
=\sum_{b=1}^n A_{ba}\nabla_{\dot\gamma(t)}(u(t)e_b)
=\sum_{b=1}^n A_{ba}\,0
=0.
\end{align*}
Hence $dR_A(H_uF_{\mathrm{SO}}(M))=H_{u\cdot A}F_{\mathrm{SO}}(M)$.
The corresponding connection form records the vertical infinitesimal rotation of a moving orthonormal frame. If $\Omega=(\Omega_{ab})$ denotes this matrix, metric compatibility forces it to be skew-symmetric: differentiating $g(E_a(t),E_b(t))=\delta_{ab}$ gives
\begin{align*}
0
=\frac{d}{dt}g(E_a(t),E_b(t))
=\Omega_{ba}+\Omega_{ab}.
\end{align*}
Thus $\Omega_{ab}=-\Omega_{ba}$, so the connection form takes values in $\mathfrak{so}(n)$. This is the principal-bundle form of the Levi-Civita connection: horizontal frame motion is precisely parallel transport of orthonormal frames.
[/example]
This example explains why principal connections are not an abstraction detached from classical geometry. The usual Christoffel-symbol connection coefficients are the local gauge potentials of this principal connection after choosing a local oriented orthonormal frame.
[remark: Local Connection Forms from Frames]
Let $e=(e_1,\dots,e_n)$ be a local oriented orthonormal frame on $U\subset M$. The pullback of the Levi-Civita connection form to $U$ is the matrix of $1$-forms $A=(A_{ab})$ determined by
\begin{align*}
\nabla_X e_b=\sum_{a=1}^n A_{ab}(X)e_a.
\end{align*}
Metric compatibility gives $A_{ab}+A_{ba}=0$, so $A$ takes values in $\mathfrak{so}(n)$.
[/remark]
The local matrix $A$ changes under a new orthonormal frame by the same gauge transformation law as before. Thus Riemannian connection coefficients are gauge-dependent local representatives of an invariant principal connection.
## The Dirac Monopole as a Local Connection
The final example shows why local potentials are sometimes unavoidable. A non-product principal bundle may admit no global section, so a connection must be described by compatible local forms rather than by one global potential on the base.
[example: Dirac Monopole Potentials on the Two-Sphere]
View the Hopf fibration $S^3\to S^2$ as a principal $U(1)$-bundle. Cover $S^2$ by $U_N=S^2\setminus\{\text{south pole}\}$ and $U_S=S^2\setminus\{\text{north pole}\}$, with spherical coordinates $(\theta,\phi)$ on $U_N\cap U_S$. A standard charge-one monopole connection is represented locally by
\begin{align*}
A_N=\frac{i}{2}(1-\cos\theta)\,d\phi
\end{align*}
and
\begin{align*}
A_S=-\frac{i}{2}(1+\cos\theta)\,d\phi.
\end{align*}
On the overlap, subtracting the two forms gives
\begin{align*}
A_N-A_S=\frac{i}{2}(1-\cos\theta)\,d\phi-\left(-\frac{i}{2}(1+\cos\theta)\,d\phi\right).
\end{align*}
The minus sign distributes through the second term, so
\begin{align*}
A_N-A_S=\frac{i}{2}(1-\cos\theta)\,d\phi+\frac{i}{2}(1+\cos\theta)\,d\phi.
\end{align*}
Factoring out $\frac{i}{2}d\phi$ gives
\begin{align*}
A_N-A_S=\frac{i}{2}\bigl((1-\cos\theta)+(1+\cos\theta)\bigr)\,d\phi.
\end{align*}
Since $(1-\cos\theta)+(1+\cos\theta)=2$, this becomes
\begin{align*}
A_N-A_S=i\,d\phi.
\end{align*}
Thus
\begin{align*}
A_N=A_S+i\,d\phi.
\end{align*}
Now take the transition function from the southern gauge to the northern gauge to be
\begin{align*}
g_{SN}=e^{i\phi}.
\end{align*}
For $X\in T_x(U_N\cap U_S)$, choose a curve $\gamma$ with $\gamma(0)=x$ and $\dot\gamma(0)=X$. By the chain rule,
\begin{align*}
(dg_{SN})_x(X)=\frac{d}{dt}\Big|_{t=0}e^{i\phi(\gamma(t))}=e^{i\phi(x)}\,i\,(d\phi)_x(X).
\end{align*}
Multiplying by $g_{SN}(x)^{-1}=e^{-i\phi(x)}$ gives
\begin{align*}
(g_{SN}^{-1}dg_{SN})_x(X)=e^{-i\phi(x)}e^{i\phi(x)}\,i\,(d\phi)_x(X).
\end{align*}
Since $e^{-i\phi(x)}e^{i\phi(x)}=1$,
\begin{align*}
(g_{SN}^{-1}dg_{SN})_x(X)=i\,(d\phi)_x(X).
\end{align*}
This holds for every tangent vector $X$, hence
\begin{align*}
g_{SN}^{-1}dg_{SN}=i\,d\phi.
\end{align*}
Therefore the relation $A_N=A_S+i\,d\phi$ is exactly the abelian gauge transformation formula
\begin{align*}
A_N=A_S+g_{SN}^{-1}dg_{SN}
\end{align*}
for the convention $s_N=s_S\cdot g_{SN}$.
The two local potentials also determine the same curvature on the overlap. Using the product rule for exterior derivatives, $d(d\phi)=0$, and $d(\cos\theta)=-\sin\theta\,d\theta$, we compute
\begin{align*}
dA_N=d\left(\frac{i}{2}(1-\cos\theta)\,d\phi\right).
\end{align*}
Since $\frac{i}{2}$ is constant,
\begin{align*}
dA_N=\frac{i}{2}\,d(1-\cos\theta)\wedge d\phi+\frac{i}{2}(1-\cos\theta)\,d(d\phi).
\end{align*}
Now $d(1-\cos\theta)=0-d(\cos\theta)=\sin\theta\,d\theta$ and $d(d\phi)=0$, so
\begin{align*}
dA_N=\frac{i}{2}\sin\theta\,d\theta\wedge d\phi.
\end{align*}
Similarly,
\begin{align*}
dA_S=d\left(-\frac{i}{2}(1+\cos\theta)\,d\phi\right).
\end{align*}
Pulling out the constant $-\frac{i}{2}$ gives
\begin{align*}
dA_S=-\frac{i}{2}\,d(1+\cos\theta)\wedge d\phi-\frac{i}{2}(1+\cos\theta)\,d(d\phi).
\end{align*}
Since $d(1+\cos\theta)=0+d(\cos\theta)=-\sin\theta\,d\theta$ and $d(d\phi)=0$, we get
\begin{align*}
dA_S=\frac{i}{2}\sin\theta\,d\theta\wedge d\phi.
\end{align*}
Thus both charts give the same curvature form
\begin{align*}
F=dA_N=dA_S=\frac{i}{2}\sin\theta\,d\theta\wedge d\phi.
\end{align*}
The connection is global on the Hopf bundle, while the two local potentials differ by the transition term forced by changing from the southern section to the northern section.
[/example]
The monopole illustrates the distinction between a connection and its potentials. The connection is global on the principal bundle, while each potential is tied to a section and may develop coordinate singularities when forced onto a chart.
[remark: Why the Monopole Needs Two Potentials]
The curvature form above integrates to a nonzero multiple of $2\pi i$ over $S^2$, so it cannot be the exterior derivative of a globally defined $i\mathbb R$-valued $1$-form on $S^2$. The obstruction is the non-[product topology](/page/Product%20Topology) of the underlying $U(1)$-bundle. Local potentials avoid this obstruction because each chart is contractible enough to admit a section.
[/remark]
The chapter has now assembled the three descriptions used throughout gauge geometry. A connection is a horizontal distribution on $P$, a connection form $\omega$ on $P$, and a family of local gauge potentials $A_i$ on charts satisfying the affine transformation law. The next stage of the course studies the curvature of such a connection and shows how it measures the failure of horizontal directions to close under Lie bracket.
A principal connection gives the invariant differential geometry that underlies all later constructions, but its meaning becomes clearest after translation to associated bundles. The next chapter explains how the same connection induces covariant derivatives on vector bundles, so that sections can be differentiated in a way that respects the bundle symmetry.
# 7. Covariant Derivatives on Associated Bundles
This chapter explains how a connection on a principal $G$-bundle produces covariant differentiation on every vector bundle associated to it. The previous chapters treated a principal connection as a horizontal distribution, or as a $\mathfrak g$-valued connection form. We now translate that invariant principal-bundle data into the operator that differentiates sections of vector bundles, tensors, dual bundles, and endomorphism bundles.
## From Principal Connections to Associated Covariant Derivatives
A section of an associated vector bundle is not just an ordinary function on the base: it is represented by an equivariant function on the total space of the principal bundle. If one differentiates this function in all directions on $P$, vertical directions record how the chosen representative changes under the $G$-action rather than how the section changes over $M$. For instance, moving from $p$ to $pg$ stays over the same base point, but an equivariant representative changes by $\rho(g^{-1})$; without a horizontal distribution this spurious change is mixed with genuine base-direction variation. The first question is how a horizontal distribution on $P$ differentiates such equivariant functions without depending on arbitrary choices of lifts.
[definition: Associated Vector Bundle]
Let $G$ be a Lie group, let $\pi:P\to M$ be a smooth principal $G$-bundle, and let $\rho:G\to GL(V)$ be a smooth finite-dimensional representation. The associated vector bundle is
\begin{align*}
P\times_G V := (P\times V)/\sim,
\end{align*}
where $(p,v)\sim (pg,\rho(g^{-1})v)$ for $g\in G$.
[/definition]
Sections of $P\times_G V$ can be described in two equivalent languages. A section $s\in\Gamma(P\times_G V)$ corresponds to a smooth function $f_s:P\to V$ satisfying
\begin{align*}
f_s(pg)=\rho(g^{-1})f_s(p).
\end{align*}
This description is the bridge from principal connections to vector-bundle connections, because the connection tells us which derivatives of $f_s$ count as horizontal derivatives. The next result is needed to show that this horizontal derivative descends from $P$ to a well-defined covariant derivative on the quotient bundle $P\times_G V$.
[quotetheorem:6263]
[citeproof:6263]
The point of the theorem is that horizontal differentiation removes the vertical ambiguity in differentiating an equivariant function on $P$. Two features are especially important for using it: the value is independent of the chosen representative $p$ in the fibre, and the resulting operator satisfies the Leibniz rule expected of a covariant derivative.
The hypotheses are all doing work. If the distribution were not $G$-equivariant, replacing $p$ by $pg$ could send a horizontal lift to a non-horizontal vector, so the expression would depend on the chosen point of the fibre. If the representation were not smooth, the vertical behaviour of the equivariant representative would not differentiate to an infinitesimal action, and the later local formula would have no meaning. The theorem does not say that every connection on every vector bundle has been constructed yet; it only constructs the covariant derivative attached to a specified principal connection and representation. This theorem applies the section-as-equivariant-map correspondence from Chapter 4 to the principal connections of Chapter 6, and the following example shows that the ordinary-looking formula for differentiating vector fields is already an associated-bundle construction.
[example: Covariant Derivative On Tangent Bundles]
Let $M$ be an $n$-dimensional smooth manifold and let $P=\operatorname{Fr}(TM)$, so a point $u\in P_x$ is a linear isomorphism $u:\mathbb R^n\to T_xM$. For the standard representation of $GL(n,\mathbb R)$ on $\mathbb R^n$, define $\Phi:P\times_{GL(n,\mathbb R)}\mathbb R^n\to TM$ by
\begin{align*}
\Phi([u,a])=u(a).
\end{align*}
This is well-defined because the associated-bundle relation gives $(u,a)\sim (ug,g^{-1}a)$, and
\begin{align*}
(ug)(g^{-1}a)=u(g(g^{-1}a))=u(a).
\end{align*}
Thus $TM\cong P\times_{GL(n,\mathbb R)}\mathbb R^n$, so any principal connection on the frame bundle induces a covariant derivative on vector fields.
Choose a local frame $e=(e_1,\ldots,e_n)$ over $U$, and write a vector field as
\begin{align*}
X=\sum_{j=1}^n s^j e_j,
\end{align*}
where $s=(s^1,\ldots,s^n)^\top:U\to\mathbb R^n$. If the local connection matrix is $A=(A^i{}_j)$, then the local formula for the induced connection gives the column of $1$-forms
\begin{align*}
(\nabla X)^i=ds^i+\sum_{j=1}^n A^i{}_j s^j.
\end{align*}
Equivalently,
\begin{align*}
\nabla X=\sum_{i=1}^n\left(ds^i+\sum_{j=1}^n A^i{}_j s^j\right)e_i.
\end{align*}
For a tangent vector $Y\in T_xM$, evaluating the $1$-form on $Y$ gives
\begin{align*}
\nabla_Y X=\sum_{i=1}^n\left(Y(s^i)+\sum_{j=1}^n A^i{}_j(Y)s^j\right)e_i(x).
\end{align*}
Thus the compact frame formula is $\nabla X=ds+As$: the term $ds$ differentiates the coordinate functions of the vector field, while $As$ records the correction coming from the connection on the moving frame.
[/example]
## The Local Formula
The global definition is invariant, but computations require a local expression. Without choosing a local section, the formula $[p,df_s(\widetilde X_p)]$ is well-defined but not directly usable: it lives upstairs on $P$ and hides the correction term needed downstairs on $M$. A poor local choice can also make the ordinary derivative $ds_i$ misleading, because changing frames introduces derivatives of the transition functions. Given a local section $\sigma_i:U_i\to P$, we want to see exactly how the principal connection form becomes the matrix-valued $1$-form that compensates for this frame dependence in covariant differentiation.
[definition: Local Connection Form]
Let $\omega\in\Omega^1(P;\mathfrak g)$ be a principal connection and let $\sigma_i:U_i\to P$ be a local section. The local connection form on $U_i$ is
\begin{align*}
A_i:=\sigma_i^*\omega\in\Omega^1(U_i;\mathfrak g).
\end{align*}
[/definition]
The local connection form is what a gauge choice sees of the global connection. If a section of $E$ is written locally as a function $s_i:U_i\to V$, the horizontal derivative can be computed without mentioning $P$ again.
[quotetheorem:6264]
[citeproof:6264]
This formula is the computational form of horizontal differentiation. The connection term is the vertical correction needed to turn the derivative seen in a local section into an invariant horizontal derivative.
The sign in the formula is determined by the convention for the associated-bundle quotient. If the quotient had been written using $\rho(g)$ instead of $\rho(g^{-1})$, the vertical derivative term would have the opposite sign, so this is not a cosmetic convention. The formula also shows the limitation of a single gauge: $A_i$ and $s_i$ are local objects, and on overlaps they must transform together for $ds_i+\rho_*(A_i)s_i$ to define a global section of $T^*M\otimes E$. This local expression is the point at which connections enter gauge theory: changing the section $\sigma_i$ is a gauge transformation, while the covariant derivative remains the invariant object. The next example illustrates the formula in the smallest non-real setting, where compatibility with a metric is visible at once.
[example: Unitary Connections On Hermitian Line Bundles]
Let $L=P\times_{U(1)}\mathbb C$ be the Hermitian line bundle associated to the standard representation of $U(1)$ on $\mathbb C$. Since
\begin{align*}
\mathfrak u(1)=\{i\alpha:\alpha\in\mathbb R\},
\end{align*}
a local connection form has the shape $A_i=i\alpha_i$ for a real-valued $1$-form $\alpha_i$. For the standard representation, the infinitesimal action is scalar multiplication:
\begin{align*}
\rho_*(i\alpha)(z)=i\alpha z.
\end{align*}
Thus the local formula from *Local Formula For The Induced Covariant Derivative* gives, for a local section represented by $f:U_i\to\mathbb C$,
\begin{align*}
\nabla f=df+A_i f.
\end{align*}
Now let $f,g:U_i\to\mathbb C$ be local sections. The Hermitian metric is $h(f,g)=f\overline{g}$. Because $A_i$ is imaginary-valued, $\overline{A_i}=-A_i$. Expanding the two metric terms gives
\begin{align*}
h(\nabla f,g)=(df+A_i f)\overline{g}=df\,\overline{g}+A_i f\overline{g}.
\end{align*}
Similarly,
\begin{align*}
h(f,\nabla g)=f\,\overline{dg+A_i g}=f\,d\overline{g}+f\,\overline{A_i}\,\overline{g}.
\end{align*}
Adding these two identities gives
\begin{align*}
h(\nabla f,g)+h(f,\nabla g)=df\,\overline{g}+f\,d\overline{g}+(A_i+\overline{A_i})f\overline{g}.
\end{align*}
Since $A_i+\overline{A_i}=0$, this reduces to
\begin{align*}
h(\nabla f,g)+h(f,\nabla g)=df\,\overline{g}+f\,d\overline{g}.
\end{align*}
By the ordinary Leibniz rule,
\begin{align*}
d(h(f,g))=d(f\overline{g})=df\,\overline{g}+f\,d\overline{g}.
\end{align*}
Therefore $d(h(f,g))=h(\nabla f,g)+h(f,\nabla g)$, so the imaginary-valued condition on $A_i$ is precisely the cancellation that makes the induced connection preserve the Hermitian metric.
[/example]
## Tensor Operations and Induced Connections
A connection should differentiate not only sections of one associated bundle, but also tensors built from them. If this compatibility failed, differentiating a contraction such as $\alpha(s)$ would depend on whether we first evaluated the scalar function or first differentiated the two factors separately. That would make tensor calculus unusable, because raising indices, contracting indices, and applying bundle endomorphisms would not commute with differentiation in a controlled way. The representation-theoretic nature of the construction answers this: tensor products, duals, exterior powers, and endomorphism bundles are associated to representations built functorially from the original ones.
[quotetheorem:6265]
[citeproof:6265]
The formula states that associated-bundle connections are functorial: every representation-theoretic construction carries its own differentiated version. Locally, this is the familiar product rule together with the differentiated tensor and dual representations.
The tensor and dual hypotheses are essential because the connections on the constructed bundles come from the differentiated tensor and dual representations, not from arbitrary independent choices. If one put unrelated connections on $E$, $F$, and $E\otimes F$, the displayed Leibniz rule could fail even when each connection separately satisfies its own scalar Leibniz rule. The theorem also does not assert flatness or any vanishing of curvature; curvature is transported through the same functorial constructions and later appears in formulas for characteristic classes and gauge fields. This compatibility is the reason a single connection can differentiate tensors of every type. Endomorphisms give a useful case because their connection can be recognized by how they commute with differentiating sections.
[example: Induced Connection On End E]
Let $E=P\times_G V$, and identify $\operatorname{End}(E)$ with the associated bundle for the conjugation representation on $\operatorname{End}(V)$:
\begin{align*}
g\cdot S=\rho(g)S\rho(g)^{-1}.
\end{align*}
For $\xi\in\mathfrak g$, the infinitesimal action is obtained by differentiating along $g(t)=\exp(t\xi)$. Since $\rho(\exp(t\xi))$ has derivative $\rho_*(\xi)$ at $t=0$, and since the derivative of the inverse curve is $-\rho_*(\xi)$, we get
\begin{align*}
\rho_{\operatorname{End},*}(\xi)(S)=\rho_*(\xi)S-S\rho_*(\xi).
\end{align*}
Thus, in a local frame with connection matrix $A_i$, if $B=\rho_*(A_i)$, the local formula for the induced connection gives
\begin{align*}
\nabla^{\operatorname{End}(E)}T=dT+BT-TB.
\end{align*}
Equivalently,
\begin{align*}
\nabla^{\operatorname{End}(E)}T=dT+[B,T].
\end{align*}
This is the formula forced by compatibility with applying an endomorphism to a section. If $s$ is represented locally by a column vector, then the induced connection on $E$ is
\begin{align*}
\nabla^E s=ds+Bs.
\end{align*}
Applying this to the section $Ts$ gives
\begin{align*}
\nabla^E(Ts)=d(Ts)+B(Ts).
\end{align*}
By the ordinary product rule for the matrix-valued function $T$ and the vector-valued function $s$,
\begin{align*}
d(Ts)=(dT)s+Tds.
\end{align*}
Therefore
\begin{align*}
\nabla^E(Ts)=(dT)s+Tds+BTs.
\end{align*}
On the other hand,
\begin{align*}
(\nabla^{\operatorname{End}(E)}T)s+T(\nabla^E s)=(dT+BT-TB)s+T(ds+Bs).
\end{align*}
Expanding the right-hand side gives
\begin{align*}
(\nabla^{\operatorname{End}(E)}T)s+T(\nabla^E s)=(dT)s+BTs-TBs+Tds+TBs.
\end{align*}
The two middle terms cancel:
\begin{align*}
-TBs+TBs=0.
\end{align*}
Hence
\begin{align*}
(\nabla^{\operatorname{End}(E)}T)s+T(\nabla^E s)=(dT)s+Tds+BTs.
\end{align*}
Comparing this with the expression for $\nabla^E(Ts)$ gives
\begin{align*}
\nabla^E(Ts)=(\nabla^{\operatorname{End}(E)}T)s+T(\nabla^E s).
\end{align*}
The commutator term is exactly the correction that makes differentiating an endomorphism commute with the Leibniz rule for applying it to sections.
[/example]
The endomorphism example shows how tensorial constructions inherit connection formulas from their representations. The next theorem is needed to treat preserved fibre tensors, such as Riemannian and Hermitian metrics, with the same associated-bundle method. It states that if the structure group preserves the tensor algebraically, then the induced covariant derivative preserves it differentially.
[quotetheorem:6266]
[citeproof:6266]
Thus a reduction of structure group to an orthogonal, symplectic, or unitary group automatically builds the expected compatibility into the induced covariant derivative. The differential compatibility is the infinitesimal shadow of invariance under the structure group.
The preservation hypothesis cannot be dropped. If $B$ is not $G$-invariant, the infinitesimal identity for $\rho_*(\xi)$ gains extra terms, and the derivative of $B_E(s,t)$ is no longer forced to split into the two covariant-derivative terms. The theorem also gives compatibility only for tensors preserved by the chosen structure group; a general bilinear form on $E$ need not be parallel for an arbitrary associated connection. This is the associated-bundle version of the familiar fact that orthogonal and unitary gauge fields preserve the metrics built into their defining representations.
## Vector-Bundle Connections and Frame Bundles
The constructions above start from a principal connection and produce vector-bundle connections. The converse question is whether an ordinary covariant derivative on a rank-$r$ vector bundle is already the same data as a principal connection on its frame bundle.
[definition: Frame Bundle]
Let $E\to M$ be a smooth real vector bundle of rank $r$. The frame bundle is the smooth fibre bundle
\begin{align*}
\pi_{\operatorname{Fr}}:\operatorname{Fr}(E)\to M
\end{align*}
whose fibre over $x\in M$ is the set of linear isomorphisms $u:\mathbb R^r\to E_x$.
[/definition]
A frame is a choice of basis in a fibre, so the frame bundle records all local frames at once. Precomposition defines a right action of $GL(r,\mathbb R)$ by
\begin{align*}
u\cdot g=u\circ g,
\end{align*}
and with this action $\operatorname{Fr}(E)$ is a principal $GL(r,\mathbb R)$-bundle. Since $E\cong\operatorname{Fr}(E)\times_{GL(r,\mathbb R)}\mathbb R^r$, any principal connection on the frame bundle gives a covariant derivative on $E$. The remaining issue is whether every covariant derivative arises in this way and whether the correspondence loses any information.
[quotetheorem:6267]
[citeproof:6267]
This theorem explains why the two languages can be used interchangeably. A vector-bundle connection differentiates sections, while the corresponding principal connection differentiates frames.
The rank and smoothness hypotheses ensure that the frame bundle is a smooth principal $GL(r,\mathbb R)$-bundle and that local connection matrices transform smoothly on overlaps. The statement is a bijection of connection data, not merely a way to build examples: if two principal connections give the same covariant derivative in every local frame, their local connection forms coincide and hence the principal connections are the same. It also explains a limitation of working only with matrices in a chosen frame: the gauge transformation law is not extra bookkeeping but the condition that the local matrices glue to a global principal connection. This equivalence also clarifies reductions of structure group. If $E$ has a metric, then metric-compatible covariant derivatives correspond to principal connections on the orthonormal frame bundle rather than on the full frame bundle.
[example: Tangent Bundle And Levi-Civita Connection]
Let $(M,g)$ be an $n$-dimensional Riemannian manifold and let $P=\operatorname{OFr}(TM)$ be its orthonormal frame bundle. A point $u\in P_x$ is a linear isometry $u:\mathbb R^n\to T_xM$, where $\mathbb R^n$ has its standard inner product. For the standard representation of $O(n)$ on $\mathbb R^n$, the associated-bundle map $P\times_{O(n)}\mathbb R^n\to TM$ defined by $[u,a]\mapsto u(a)$ is well-defined: if $(u,a)\sim(uh,h^{-1}a)$ with $h\in O(n)$, then
\begin{align*}
(uh)(h^{-1}a)=u(h(h^{-1}a))=u(a).
\end{align*}
Thus $P\times_{O(n)}\mathbb R^n\cong TM$.
Choose a local orthonormal frame $e=(e_1,\ldots,e_n)$ and write
\begin{align*}
X=\sum_{j=1}^n s^j e_j.
\end{align*}
If $A=(A^i{}_j)$ is the local connection matrix of the Levi-Civita connection in this frame, the induced associated-bundle formula gives
\begin{align*}
\nabla X=\sum_{i=1}^n\left(ds^i+\sum_{j=1}^n A^i{}_j s^j\right)e_i.
\end{align*}
The matrix $A$ is skew-symmetric because the frame is orthonormal and the Levi-Civita connection preserves $g$. Since $g(e_i,e_j)=\delta_{ij}$ is constant, its exterior derivative is
\begin{align*}
d(g(e_i,e_j))=d(\delta_{ij})=0.
\end{align*}
Metric compatibility gives
\begin{align*}
d(g(e_i,e_j))=g(\nabla e_i,e_j)+g(e_i,\nabla e_j).
\end{align*}
Therefore
\begin{align*}
0=g(\nabla e_i,e_j)+g(e_i,\nabla e_j).
\end{align*}
Writing
\begin{align*}
\nabla e_i=\sum_{k=1}^n A^k{}_i e_k
\end{align*}
and using bilinearity of $g$ together with $g(e_k,e_j)=\delta_{kj}$ gives
\begin{align*}
g(\nabla e_i,e_j)=g\left(\sum_{k=1}^n A^k{}_i e_k,e_j\right)=\sum_{k=1}^n A^k{}_i g(e_k,e_j)=\sum_{k=1}^n A^k{}_i\delta_{kj}=A^j{}_i.
\end{align*}
Similarly, using $g(e_i,e_k)=\delta_{ik}$,
\begin{align*}
g(e_i,\nabla e_j)=g\left(e_i,\sum_{k=1}^n A^k{}_j e_k\right)=\sum_{k=1}^n A^k{}_j g(e_i,e_k)=\sum_{k=1}^n A^k{}_j\delta_{ik}=A^i{}_j.
\end{align*}
Substituting these two identities into the metric-compatibility equation gives
\begin{align*}
0=A^j{}_i+A^i{}_j.
\end{align*}
Hence
\begin{align*}
A^i{}_j=-A^j{}_i,
\end{align*}
so $A$ takes values in $\mathfrak o(n)$. The usual Levi-Civita covariant derivative of vector fields is therefore exactly the covariant derivative induced from a principal connection on the orthonormal frame bundle.
[/example]
The chapter therefore closes the circle between the principal and vector viewpoints. Principal connections are the invariant global objects; covariant derivatives on associated bundles are their concrete action on fields, tensors, metrics, and endomorphisms.
Covariant derivatives express how a connection acts on fields, but the next question is what happens when we compare directions in different orders. The next chapter develops curvature as the obstruction to flatness, linking the connection-form formalism to the failure of horizontal lifts to commute.
# 8. Curvature
This chapter develops curvature for principal connections, building on the connection forms and horizontal distributions introduced in Chapter 6 and the covariant-derivative viewpoint of Chapter 7. The guiding question is how to measure the failure of horizontal directions to behave like the flat horizontal directions in a product bundle, and how that measurement appears after choosing local gauges. The answer is a horizontal and equivariant $\mathfrak{g}$-valued $2$-form whose local representative is
\begin{align*}
dA + \frac{1}{2}[A,A].
\end{align*}
This formula is the bridge between the invariant principal-bundle definition and the gauge-theoretic computations used later.
## The Structure Equation for Principal Curvature
A connection form tells us which tangent vectors are vertical and which are horizontal, but it does not yet measure whether horizontal directions fit together to form local horizontal submanifolds. The first problem is to define a $2$-form that records the failure of horizontal vector fields to be closed under Lie bracket.
[definition: Curvature Form]
Let $\pi:P\to M$ be a smooth principal $G$-bundle with connection form $\omega\in \Omega^1(P;\mathfrak{g})$. The curvature form of $\omega$ is the $\mathfrak{g}$-valued $2$-form $\Omega\in \Omega^2(P;\mathfrak{g})$ defined by
\begin{align*}
\Omega(X,Y) = d\omega(hX,hY),
\end{align*}
where $h:T_pP\to H_pP$ denotes projection onto the horizontal subspace determined by $\omega$.
[/definition]
The projection in this definition is important: curvature is intended to measure horizontal geometry, not the tautological vertical behaviour already fixed by the Lie algebra action. For computations, however, the definition with $hX$ and $hY$ is inconvenient; the next theorem rewrites it using only the exterior derivative and the Lie bracket on $\mathfrak{g}$.
[quotetheorem:6268]
[citeproof:6268]
The formula is called a structure equation because it combines the differential structure of $P$ with the algebraic structure of $G$. The hypotheses are doing real work: $\omega$ must be a principal connection form, not just an arbitrary $\mathfrak{g}$-valued $1$-form, because the cancellation on vertical directions uses both $\omega(\xi_P)=\xi$ and right-equivariance. If a $\mathfrak{g}$-valued $1$-form on $P$ fails these connection-form axioms, the expression
\begin{align*}
d\omega+\frac{1}{2}[\omega,\omega]
\end{align*}
can still be written down, but it need not agree with the horizontal projection definition and need not represent curvature of a principal connection. For a concrete failure, take the additive group $G=\mathbb R$ and $P=\mathbb R^2\times \mathbb R\to \mathbb R^2$ with fibre coordinate $t$. The $\mathbb R$-valued form $\alpha=t\,dx$ is not a principal connection form, and the abelian expression $d\alpha=dt\wedge dx$ has vertical sensitivity, while its product-horizontal projection vanishes on horizontal pairs.
The theorem also does not say that curvature is simply the exterior derivative of the connection form. The bracket term is absent only for abelian structure groups or in special gauges where the relevant Lie brackets vanish. The factor $\frac{1}{2}$ compensates for the alternating bracket convention for Lie-algebra-valued forms. The model case to test the formula is the product connection, where the horizontal distribution has no twisting.
[example: Flat Product Connection]
Consider the product principal bundle $P=M\times G\to M$, write $\operatorname{pr}_G:M\times G\to G$, and let $\omega=\operatorname{pr}_G^*\theta$, where $\theta$ is the Maurer-Cartan form on $G$. For a tangent vector $(v,\eta)\in T_xM\oplus T_gG$, the definition of pullback gives
\begin{align*}
\omega_{(x,g)}(v,\eta)=\theta_g\big((d\operatorname{pr}_G)_{(x,g)}(v,\eta)\big)=\theta_g(\eta).
\end{align*}
The product horizontal distribution is $H_{(x,g)}P=T_xM\times\{0\}$, so every horizontal vector has the form $(v,0)$, and therefore
\begin{align*}
\omega_{(x,g)}(v,0)=\theta_g(0)=0.
\end{align*}
Now extend two horizontal tangent vectors locally to product-horizontal vector fields $\widetilde X=(X,0)$ and $\widetilde Y=(Y,0)$. Since $\omega(\widetilde X)=0$ and $\omega(\widetilde Y)=0$, the exterior derivative formula for a $1$-form gives
\begin{align*}
d\omega(\widetilde X,\widetilde Y)=\widetilde X\big(\omega(\widetilde Y)\big)-\widetilde Y\big(\omega(\widetilde X)\big)-\omega([\widetilde X,\widetilde Y]).
\end{align*}
Here $[\widetilde X,\widetilde Y]=([X,Y],0)$ because the $G$-component of both product vector fields is zero. Substituting the three terms gives
\begin{align*}
d\omega(\widetilde X,\widetilde Y)=0-0-\omega([X,Y],0)=0-0-\theta_g(0)=0.
\end{align*}
Thus, for horizontal vectors $(v,0)$ and $(w,0)$,
\begin{align*}
\Omega((v,0),(w,0))=d\omega((v,0),(w,0))=0.
\end{align*}
For arbitrary tangent vectors, the definition of curvature first projects both arguments to their horizontal parts, so the same calculation shows $\Omega=0$ everywhere.
Equivalently, since pullback commutes with exterior differentiation and with the pointwise Lie bracket of Lie-algebra-valued forms,
\begin{align*}
d\omega+\frac{1}{2}[\omega,\omega]=\operatorname{pr}_G^*d\theta+\frac{1}{2}\operatorname{pr}_G^*[\theta,\theta]=\operatorname{pr}_G^*\left(d\theta+\frac{1}{2}[\theta,\theta]\right).
\end{align*}
By the Maurer-Cartan identity $d\theta+\frac{1}{2}[\theta,\theta]=0$, this expression is zero. The product bundle therefore carries a flat product connection, showing that curvature depends on the chosen connection, not only on the underlying principal bundle.
[/example]
The vanishing in the product case motivates the terminology: a flat connection is a connection with zero curvature. Chapter 10 provides the global refinement of flatness through holonomy, since a flat connection may still have nonzero parallel transport around loops when the base is not simply connected.
[definition: Flat Principal Connection]
Let $\pi:P\to M$ be a smooth principal $G$-bundle with connection form $\omega$ and curvature form $\Omega$. The connection is flat if
\begin{align*}
\Omega=0.
\end{align*}
[/definition]
## Horizontality and Equivariance of Curvature
The structure equation gives a computable formula, but it does not yet explain why curvature is a geometric object on the base rather than an arbitrary $2$-form upstairs on $P$. The next question is how $\Omega$ behaves under vertical insertion and under the principal right action.
[quotetheorem:6269]
[citeproof:6269]
Horizontality says that curvature ignores vertical directions, while equivariance says that changing the point in a fibre only changes the Lie algebra value by the adjoint representation. Both hypotheses are necessary for descent. Without horizontality, inserting a vertical vector would produce data depending on motion along the fibre rather than on tangent vectors to $M$; without equivariance, two representatives $p$ and $pg$ of the same base point would give incompatible Lie algebra values.
A concrete failure of horizontality is already visible on $P=\mathbb R^2\times\mathbb R\to \mathbb R^2$ for the additive group $\mathbb R$. The form $\beta=dt\wedge dx$ is a well-defined upstairs $2$-form, but it gives different values on the two lifts $\partial_y$ and $\partial_y+b\partial_t$ of the same base tangent vector, since
\begin{align*}
\beta(\partial_x,\partial_y+b\partial_t)=-b.
\end{align*}
Thus $\beta$ cannot define a $2$-form on the base.
A separate failure of equivariance occurs even for horizontal forms. Take the product principal $SU(2)$-bundle $P=\mathbb R^2\times SU(2)\to \mathbb R^2$, choose a nonzero element $\xi\in\mathfrak{su}(2)$ that is not fixed by the adjoint action, and set
\begin{align*}
\gamma=\xi\, dx\wedge dy.
\end{align*}
This form vanishes whenever a vertical vector is inserted, so it is horizontal. However, right translation by $g\in SU(2)$ gives $(R_g)^*\gamma=\xi\,dx\wedge dy$, while equivariance would require $\operatorname{Ad}_{g^{-1}}\xi\,dx\wedge dy$. Choosing $g$ with $\operatorname{Ad}_{g^{-1}}\xi\ne \xi$ shows that the two representatives $(x,e)$ and $(x,g)$ assign incompatible adjoint-bundle values to the same base $2$-plane. The theorem therefore does not say that every $\mathfrak{g}$-valued form on $P$ descends, only that the curvature form has exactly the two properties required to descend with adjoint-bundle coefficients. This is why the next definition packages curvature downstairs.
[definition: Adjoint Bundle Valued Curvature]
Let $\pi:P\to M$ be a smooth principal $G$-bundle with curvature form $\Omega$. The associated adjoint-bundle-valued curvature is the $2$-form $F_\nabla\in \Omega^2(M;\operatorname{Ad}(P))$ represented by $\Omega$ under the correspondence between horizontal $\operatorname{Ad}$-equivariant $\mathfrak{g}$-valued forms on $P$ and $\operatorname{Ad}(P)$-valued forms on $M$.
[/definition]
This is the invariant version of the local curvature forms introduced below. The notation $F_\nabla$ emphasizes that the same curvature appears when the principal connection induces covariant derivatives on associated vector bundles.
[remark: Abelian Structure Group]
If $G$ is abelian, then $\operatorname{Ad}$ is the identity representation and $[\omega,\omega]=0$. In that case the structure equation reduces to $\Omega=d\omega$, and the curvature form is invariant under the right action rather than merely equivariant.
[/remark]
The abelian case is the bridge to electromagnetism and line bundles: curvature is then an ordinary differential form locally expressed as the exterior derivative of a potential. The nonabelian case differs because the bracket term records the self-interaction of the gauge field.
[example: Curvature of a U One Connection]
Let $P\to M$ be a principal $U(1)$-bundle. The Lie algebra $\mathfrak{u}(1)\cong i\mathbb R$ is abelian, so for any $\mathfrak{u}(1)$-valued $1$-form $\alpha$ one has
\begin{align*}
[\alpha,\alpha](X,Y)=2[\alpha(X),\alpha(Y)]=0.
\end{align*}
Hence the structure equation for a connection form $\omega$ reduces to
\begin{align*}
\Omega=d\omega+\frac{1}{2}[\omega,\omega]=d\omega.
\end{align*}
Choose a local section $s_i:U_i\to P$ and set
\begin{align*}
A_i=s_i^*\omega\in \Omega^1(U_i;i\mathbb R).
\end{align*}
The local curvature form is
\begin{align*}
F_i=s_i^*\Omega=s_i^*(d\omega)=d(s_i^*\omega)=dA_i,
\end{align*}
where pullback commutes with exterior differentiation.
Now replace the local gauge by a map $g:U_i\to U(1)$, so the local connection form changes to
\begin{align*}
A_i'=A_i+g^{-1}dg.
\end{align*}
On any open set where $g=e^{i\phi}$ for a smooth real-valued function $\phi$, we have
\begin{align*}
dg=d(e^{i\phi})=i e^{i\phi}\,d\phi,
\end{align*}
and therefore
\begin{align*}
g^{-1}dg=e^{-i\phi}\,i e^{i\phi}\,d\phi=i\,d\phi.
\end{align*}
Taking exterior derivatives gives
\begin{align*}
d(g^{-1}dg)=d(i\,d\phi)=i\,d^2\phi=0.
\end{align*}
Thus the transformed local curvature is
\begin{align*}
F_i'=dA_i'=d(A_i+g^{-1}dg)=dA_i+d(g^{-1}dg)=dA_i=F_i.
\end{align*}
For a $U(1)$-connection, the local curvature is therefore gauge invariant, not merely gauge covariant.
[/example]
## Local Curvature Forms and Gauge Covariance
Calculations on a principal bundle are usually performed after choosing local sections, so the next problem is to translate $\Omega$ into gauge-dependent forms on open subsets of $M$. The result should remember the global equivariance of $\Omega$ by transforming covariantly on overlaps.
[definition: Local Connection and Curvature Forms]
Let $\pi:P\to M$ be a smooth principal $G$-bundle with connection form $\omega$, and let $s_i:U_i\to P$ be a local section. The local connection form and local curvature form in the gauge $s_i$ are
\begin{align*}
A_i = s_i^*\omega\in \Omega^1(U_i;\mathfrak{g}),
\end{align*}
and
\begin{align*}
F_i = s_i^*\Omega\in \Omega^2(U_i;\mathfrak{g}).
\end{align*}
[/definition]
The local forms are the objects used in calculations, but the definition by pullback does not yet give a formula in terms of the gauge potential $A_i$ alone. Pulling back the structure equation answers that computational problem and produces the curvature formula used in gauge theory.
[quotetheorem:1540]
[citeproof:1540]
The expression
\begin{align*}
dA_i+\frac{1}{2}[A_i,A_i]
\end{align*}
resembles the curvature of a covariant derivative in a local frame. The local section hypothesis is essential: $A_i$ and $F_i$ are not global forms unless the bundle has a global section or the connection is being studied on a local product chart. On a principal bundle without a global product presentation, replacing $s_i$ by another local section changes $A_i$, so the theorem does not assert that
\begin{align*}
dA_i+\frac{1}{2}[A_i,A_i]
\end{align*}
is a globally defined $\mathfrak{g}$-valued form on $M$.
For a concrete failure, use the Hopf principal $U(1)$-bundle $S^3\to S^2$ with a connection whose curvature represents a nonzero multiple of the generator of $H^2(S^2;\mathbb R)$. If the local structure equation could be upgraded to a single global potential $A\in\Omega^1(S^2;i\mathbb R)$ with $F=dA$, then $F$ would be exact and hence would have zero integral over $S^2$ by [Stokes' theorem](/theorems/1530). The Hopf connection has nonzero curvature integral, so no such global $A$ exists. The local sections over the northern and southern charts give local potentials $A_N$ and $A_S$, but on the equatorial overlap they are related by the transition function rather than by equality.
The boundary case is a product bundle with one global gauge, where the local formula can be used globally. In general, the missing information is how these local formulae agree on overlaps. The next theorem supplies precisely that compatibility, showing that the local curvatures glue to the adjoint-bundle-valued curvature rather than to an ordinary $\mathfrak{g}$-valued $2$-form.
[quotetheorem:6270]
[citeproof:6270]
This theorem is the reason curvature is described as gauge covariant, not gauge invariant. The overlap hypothesis is essential: the formula compares two gauges only where both local sections are defined and related by a transition function. It does not say that the matrix entries of $F_i$ are independent of gauge; for a nonabelian group, conjugation can change those entries even though it preserves adjoint-invariant quantities such as traces of powers in matrix representations.
A useful failure case is trying to glue the local $\mathfrak{g}$-valued forms $F_i$ by equality on overlaps. That works for a globally fixed product presentation or for abelian groups with identity adjoint action, but it fails for a general nonabelian principal bundle because the correct comparison is by $\operatorname{Ad}_{g_{ij}^{-1}}$. This is why adjoint-invariant polynomials applied to curvature produce globally defined ordinary differential forms.
[example: Matrix Gauge Fields]
Let $G\subset GL(n,\mathbb C)$ be a matrix Lie group, and write the local connection form as a matrix $A_i=(A_{ab})$ of $1$-forms. For tangent vectors $X,Y\in T_xU_i$, the wedge product of matrix-valued forms is defined by alternating ordinary matrix multiplication of the values:
\begin{align*}
(A_i\wedge A_i)(X,Y)=A_i(X)A_i(Y)-A_i(Y)A_i(X).
\end{align*}
The bracket convention for Lie-algebra-valued $1$-forms gives
\begin{align*}
[A_i,A_i](X,Y)=2[A_i(X),A_i(Y)].
\end{align*}
Since the Lie bracket in a matrix Lie algebra is the matrix commutator,
\begin{align*}
[A_i(X),A_i(Y)]=A_i(X)A_i(Y)-A_i(Y)A_i(X).
\end{align*}
Therefore
\begin{align*}
[A_i,A_i](X,Y)=2(A_i\wedge A_i)(X,Y).
\end{align*}
Hence
\begin{align*}
\frac{1}{2}[A_i,A_i](X,Y)=(A_i\wedge A_i)(X,Y).
\end{align*}
Because this holds for every pair $X,Y$, the *Local Structure Equation* becomes
\begin{align*}
F_i=dA_i+\frac{1}{2}[A_i,A_i]=dA_i+A_i\wedge A_i.
\end{align*}
On an overlap $U_i\cap U_j$, the *Gauge Covariance of Local Curvature* gives
\begin{align*}
F_j=\operatorname{Ad}_{g_{ij}^{-1}}F_i.
\end{align*}
For a matrix group, the adjoint action is conjugation. Thus, at $x\in U_i\cap U_j$ and for $X,Y\in T_x(U_i\cap U_j)$,
\begin{align*}
F_j(X,Y)=g_{ij}(x)^{-1}F_i(X,Y)g_{ij}(x).
\end{align*}
Equivalently, as a matrix-valued $2$-form,
\begin{align*}
F_j=g_{ij}^{-1}F_i g_{ij}.
\end{align*}
This is the local gauge-field formula: the connection matrix changes by a gauge-dependent rule, while its curvature transforms by conjugation.
[/example]
## The Bianchi Identity
After constructing curvature, the next structural question is whether it satisfies a differential identity. The ordinary exterior derivative is not the right operator because $\Omega$ is adjoint-equivariant rather than invariant; the connection supplies the corrected derivative.
[definition: Exterior Covariant Derivative on Equivariant Forms]
Let $\pi:P\to M$ be a principal $G$-bundle with connection form $\omega$. Let $\Omega^k_{\mathrm{hor},\operatorname{Ad}}(P;\mathfrak{g})$ denote the space of horizontal $\operatorname{Ad}$-equivariant $\mathfrak{g}$-valued $k$-forms on $P$. The exterior covariant derivative sends
\begin{align*}
D:\Omega^k_{\mathrm{hor},\operatorname{Ad}}(P;\mathfrak{g})\to \Omega^{k+1}_{\mathrm{hor},\operatorname{Ad}}(P;\mathfrak{g}).
\end{align*}
It is defined by
\begin{align*}
D\alpha=h^*(d\alpha),
\end{align*}
where $h^*(d\alpha)$ means that $d\alpha$ is evaluated after projecting all tangent arguments onto the horizontal distribution.
[/definition]
This derivative is the principal-bundle version of the covariant exterior derivative on $\operatorname{Ad}(P)$-valued forms. The remaining question is what happens when the corrected derivative is applied to the curvature itself; the answer is a universal identity for every connection.
[quotetheorem:1541]
[citeproof:1541]
The Bianchi identity is not an equation of motion; it is a compatibility identity forced by the definition of curvature from a connection. The connection hypothesis is essential: an arbitrary adjoint-bundle-valued $2$-form need not satisfy $DF=0$ for a chosen connection. Even for an ordinary vector bundle, a randomly chosen endomorphism-valued $2$-form generally has nonzero covariant exterior derivative.
For a concrete counterexample, take the product principal $U(1)$-bundle over $\mathbb R^3$ with the flat connection $A=0$. Since $U(1)$ is abelian, the exterior covariant derivative is just $d$. The ordinary $i\mathbb R$-valued $2$-form
\begin{align*}
F=i x_1\, dx_2\wedge dx_3
\end{align*}
is horizontal and gauge invariant in this product gauge, but
\begin{align*}
dF=i\,dx_1\wedge dx_2\wedge dx_3\ne 0.
\end{align*}
So $F$ cannot be the curvature of the flat connection, and it does not satisfy the Bianchi identity for that connection. Thus the theorem does not say that every gauge-covariant $2$-form is a curvature form. It says that once $F$ is built from a connection by the structure equation, its covariant derivative automatically vanishes. In later applications this identity will be paired with extra equations, such as Yang-Mills equations, which impose analytic conditions on the same curvature form.
[remark: Local Meaning of the Covariant Term]
The term $[A_i,F_i]$ is the correction that makes $dF_i+[A_i,F_i]$ transform covariantly on overlaps. Without it, $dF_i$ would acquire derivatives of the transition function and would not represent a global adjoint-bundle-valued $3$-form.
[/remark]
## Curvature as an Operator on Associated Bundles
Principal curvature becomes more familiar after passing to associated vector bundles. The final problem in this chapter is to relate the $2$-form $\Omega$ to the commutator of covariant derivatives.
[quotetheorem:6271]
[citeproof:6271]
This theorem explains why the same word curvature appears in principal bundles, vector bundles, and Riemannian geometry. The representation hypothesis is necessary because a principal curvature value lies in $\mathfrak{g}$, while curvature on a vector bundle must act as an endomorphism of the fibre. The differential $d\rho$ is the mechanism that converts Lie algebra elements into linear operators.
The theorem does not identify principal curvature with every possible vector-bundle curvature operator. It applies to vector bundles arising as associated bundles from the chosen principal bundle and representation. A specific limitation appears for the constant representation $\rho:U(1)\to GL(\mathbb C)$, $\rho(z)=1$. Then $d\rho=0$, so every principal curvature value induces the zero endomorphism on the associated line bundle. On the Hopf bundle with a connection of nonzero curvature, the principal curvature is nonzero, but the curvature operator on the line bundle associated by this constant representation is zero. Thus the representation hypothesis supplies the map into endomorphisms, and a representation with kernel can discard curvature information.
This distinction prepares the next stage of the course: characteristic forms and holonomy often use associated representations, but the principal curvature remains the object that tracks all directions in the structure Lie algebra before a representation is chosen.
[example: Levi Civita Curvature from the Frame Bundle]
Let $(M,g)$ be a Riemannian manifold and let $P=O(M)$ be its orthonormal frame bundle. A point $u\in O(M)_x$ is an isometry $u:\mathbb R^n\to T_xM$, and the standard representation of $O(n)$ on $\mathbb R^n$ identifies $P\times_{O(n)}\mathbb R^n$ with $TM$ by sending $[u,a]$ to $u(a)$. This is well-defined because $(u,a)\sim(uh,h^{-1}a)$ gives $(uh)(h^{-1}a)=u(a)$.
Choose a local orthonormal frame $e=(e_1,\dots,e_n)$ and write
\begin{align*}
Z=\sum_a z^a e_a.
\end{align*}
The Levi-Civita connection is represented in this frame by the matrix-valued $1$-form $A=(A^a_{\ b})$ defined by
\begin{align*}
\nabla_X e_b=\sum_a A^a_{\ b}(X)e_a.
\end{align*}
Using the Leibniz rule,
\begin{align*}
\nabla_X Z=\sum_b X(z^b)e_b+\sum_b z^b\nabla_Xe_b.
\end{align*}
Substituting the definition of $A$ gives
\begin{align*}
\nabla_X Z=\sum_b X(z^b)e_b+\sum_{a,b}z^bA^a_{\ b}(X)e_a.
\end{align*}
Collecting the coefficient of $e_a$ gives
\begin{align*}
\nabla_X Z=\sum_a\left(X(z^a)+\sum_b A^a_{\ b}(X)z^b\right)e_a.
\end{align*}
Thus, in the frame $e$, the covariant derivative has the matrix form
\begin{align*}
\nabla_X z=X(z)+A(X)z,
\end{align*}
where $z$ is the column vector of component functions.
Now compute the curvature operator in the same frame. Applying the preceding formula twice gives
\begin{align*}
\nabla_X\nabla_Yz=X\big(Y(z)+A(Y)z\big)+A(X)\big(Y(z)+A(Y)z\big).
\end{align*}
Expanding by the product rule,
\begin{align*}
\nabla_X\nabla_Yz=X(Y(z))+X(A(Y))z+A(Y)X(z)+A(X)Y(z)+A(X)A(Y)z.
\end{align*}
Similarly,
\begin{align*}
\nabla_Y\nabla_Xz=Y(X(z))+Y(A(X))z+A(X)Y(z)+A(Y)X(z)+A(Y)A(X)z.
\end{align*}
Also,
\begin{align*}
\nabla_{[X,Y]}z=[X,Y](z)+A([X,Y])z.
\end{align*}
Therefore
\begin{align*}
\nabla_X\nabla_Yz-\nabla_Y\nabla_Xz-\nabla_{[X,Y]}z=\big(X(Y(z))-Y(X(z))-[X,Y](z)\big)+\big(X(A(Y))-Y(A(X))-A([X,Y])\big)z+\big(A(X)A(Y)-A(Y)A(X)\big)z.
\end{align*}
The identity $[X,Y](z)=X(Y(z))-Y(X(z))$ cancels the first parenthesis. The mixed first-derivative terms $A(Y)X(z)$ and $A(X)Y(z)$ have already cancelled in the subtraction. Hence
\begin{align*}
R(X,Y)Z=\left(X(A(Y))-Y(A(X))-A([X,Y])+A(X)A(Y)-A(Y)A(X)\right)z.
\end{align*}
For a matrix-valued $1$-form,
\begin{align*}
dA(X,Y)=X(A(Y))-Y(A(X))-A([X,Y]).
\end{align*}
Also,
\begin{align*}
(A\wedge A)(X,Y)=A(X)A(Y)-A(Y)A(X).
\end{align*}
Thus
\begin{align*}
R(X,Y)Z=\big(dA+A\wedge A\big)(X,Y)z.
\end{align*}
For the standard representation of $O(n)$, the differential $d\rho:\mathfrak{o}(n)\to\mathfrak{gl}(\mathbb R^n)$ is the inclusion map. The local form of the principal curvature of the Levi-Civita connection is
\begin{align*}
F=dA+A\wedge A.
\end{align*}
By the associated-bundle curvature formula from *Curvature Operator on an Associated Vector Bundle*,
\begin{align*}
R(X,Y)Z=d\rho(F(X,Y))Z=F(X,Y)z.
\end{align*}
So the usual Riemann curvature operator is the principal curvature of the Levi-Civita connection on the orthonormal frame bundle, applied to tangent vectors through the standard representation.
[/example]
Curvature measures the local failure of a connection to be flat, and the next step is to understand how this curvature behaves under changes of local frame. The next chapter introduces gauge transformations, showing how connection forms and curvature transform while preserving the underlying principal geometry.
# 9. Gauge Transformations and Gauge Fields
This chapter turns from connections as geometric choices to the symmetries that change their local representatives. It assumes the earlier material on principal bundles from Chapters 1 and 2, associated and adjoint bundles from Chapter 4, connection forms from Chapter 6, curvature from Chapter 8, and the adjoint action of a Lie group on its Lie algebra. A principal bundle has automorphisms that move points inside the fibres while leaving the base point fixed, and these automorphisms are the gauge transformations of the bundle. The central questions are how to describe these transformations intrinsically, how they act on connection forms and curvature, and how the local transformation law
\begin{align*}
A \mapsto u^{-1}Au+u^{-1}du
\end{align*}
arises from the global principal-bundle language.
## Automorphisms Over the Identity
The first problem is to isolate the automorphisms that are invisible on the base manifold. General principal bundle maps may cover non-identity diffeomorphisms of $M$, but gauge theory studies changes of internal frame at each point of $M$ while the point of spacetime or the base manifold remains fixed.
[definition: Gauge Transformation]
Let $\pi:P\to M$ be a smooth principal $G$-bundle. A gauge transformation of $P$ is a principal bundle automorphism $\Phi:P\to P$ such that $\pi\circ \Phi=\pi$.
[/definition]
Thus $\Phi$ maps each fibre $P_x$ to itself and respects the right $G$-action: $\Phi(pg)=\Phi(p)g$ for all $p\in P$ and $g\in G$. The next object records all such transformations with composition as its operation.
[definition: Gauge Group]
The gauge group of $P$ is
\begin{align*}
\operatorname{Gau}(P):=\{\Phi:P\to P \mid \Phi \text{ is a principal bundle automorphism and } \pi\circ\Phi=\pi\}.
\end{align*}
Its group operation is composition of maps.
[/definition]
This definition is geometric but not yet calculational. To compute with a gauge transformation, we compare $\Phi(p)$ with $p$ inside the same principal homogeneous $G$-space, which leads to an equivariant function on the total space.
[quotetheorem:6272]
[citeproof:6272]
The theorem uses each principal-bundle hypothesis. Freeness is needed for uniqueness: if a right action had a non-identity element $k$ fixing some $p$, then $p\gamma=p(k\gamma)$, so the comparison element between $p$ and $\Phi(p)$ would not be determined. Equivariance of $\Phi$ is also needed: on the product bundle $M\times G$, the map $(x,h)\mapsto (x,h^2)$ stays over $M$ but is not a principal bundle automorphism, and its comparison function would not satisfy the conjugation rule. The condition $\pi\circ\Phi=\pi$ is needed as well, since an automorphism covering a non-identity diffeomorphism sends $p$ and $\Phi(p)$ to different fibres, where no group element compares them. Thus the associated bundle viewpoint below records the precise frame-dependence forced by the theorem.
[example: Product Bundle Gauge Functions]
Let $P=M\times G$ with right action $(x,h)g=(x,hg)$. Given a smooth map $u:M\to G$, define
\begin{align*}
\Phi_u(x,h)=(x,u(x)h).
\end{align*}
This map stays over the same base point because $\pi(\Phi_u(x,h))=x=\pi(x,h)$. It respects the right action: for $g\in G$,
\begin{align*}
\Phi_u((x,h)g)=\Phi_u(x,hg).
\end{align*}
By the definition of $\Phi_u$,
\begin{align*}
\Phi_u(x,hg)=(x,u(x)hg).
\end{align*}
Using the product right action again,
\begin{align*}
(x,u(x)hg)=(x,u(x)h)g=\Phi_u(x,h)g.
\end{align*}
Thus $\Phi_u$ is a gauge transformation in this product model, with inverse given by $\Phi_{u^{-1}}(x,h)=(x,u(x)^{-1}h)$.
To express the same transformation in the equivariant-function convention, require
\begin{align*}
\Phi_u(x,h)=(x,h)\gamma(x,h).
\end{align*}
Since $(x,h)\gamma(x,h)=(x,h\gamma(x,h))$ while $\Phi_u(x,h)=(x,u(x)h)$, equality of the second components gives
\begin{align*}
h\gamma(x,h)=u(x)h.
\end{align*}
Multiplying this equation on the left by $h^{-1}$ gives
\begin{align*}
\gamma(x,h)=h^{-1}u(x)h.
\end{align*}
For $g\in G$, evaluate this function at $(x,h)g=(x,hg)$:
\begin{align*}
\gamma((x,h)g)=\gamma(x,hg).
\end{align*}
Substituting $hg$ for $h$ in the formula for $\gamma$ gives
\begin{align*}
\gamma(x,hg)=(hg)^{-1}u(x)(hg).
\end{align*}
Using $(hg)^{-1}=g^{-1}h^{-1}$, this becomes
\begin{align*}
(hg)^{-1}u(x)(hg)=g^{-1}h^{-1}u(x)hg.
\end{align*}
Since $\gamma(x,h)=h^{-1}u(x)h$, the last expression is
\begin{align*}
g^{-1}h^{-1}u(x)hg=g^{-1}\gamma(x,h)g.
\end{align*}
Hence $\gamma((x,h)g)=g^{-1}\gamma(x,h)g$, so the base function $u$ becomes an equivariant function on the total space by conjugating it into the frame determined by $h$.
[/example]
This example explains why physicists often write a gauge transformation as a $G$-valued function on the base. That notation is globally available only after choosing a global product model; on a general bundle the invariant replacement is a section of an associated bundle.
## Gauge Transformations as Sections of the Adjoint Group Bundle
The next question is how to package the equivariant function $\gamma:P\to G$ as an object living directly over $M$. Simply trying to define a function $u:M\to G$ by $u(x)=\gamma(p)$ fails on a nontrivial bundle, because replacing $p$ by $pg$ changes the value by conjugation. The conjugation law in the previous theorem is exactly the descent condition for the associated bundle formed using the conjugation action of $G$ on itself.
[definition: Adjoint Group Bundle]
Let $P\to M$ be a principal $G$-bundle. The adjoint group bundle is
\begin{align*}
\operatorname{Ad}(P):=P\times_G G,
\end{align*}
where $G$ acts on itself by conjugation, $g\cdot h=ghg^{-1}$. Its fibre over $x\in M$ is the group associated to $P_x$ by this conjugation action.
Its bundle projection is
\begin{align*}
\pi_{\operatorname{Ad}}:\operatorname{Ad}(P)\to M,\qquad [p,h]\mapsto \pi(p).
\end{align*}
The smooth sections form the set $\Gamma^\infty(\operatorname{Ad}(P))$ of smooth maps $\sigma:M\to \operatorname{Ad}(P)$ satisfying $\pi_{\operatorname{Ad}}\circ\sigma=\operatorname{id}_M$, with pointwise multiplication in the fibres.
[/definition]
A section of $\operatorname{Ad}(P)$ assigns to each base point an element of the structure group expressed in the local frame of the chosen point of $P_x$. Changing the frame conjugates the representative, so the preceding definition is designed to make the equivariant-function description descend to a global section; this motivates the identification theorem.
[quotetheorem:6273]
[citeproof:6273]
This identification depends on the conjugation action: if the associated bundle were built using the trivial action, the formula $x\mapsto[p,\gamma(p)]$ would fail to be independent of $p$ whenever $G$ is nonabelian. For instance, in a product bundle with $G=SU(2)$ and a noncentral value $a\in SU(2)$, replacing $(x,e)$ by $(x,g)$ changes the representative from $a$ to $g^{-1}ag$, which is not the same element of a trivial-action associated bundle. Smoothness is also part of the statement: a discontinuous choice of $u:M\to G$ on a product bundle gives a set-theoretic fibre map but not a smooth principal automorphism. The theorem does not trivialise $\operatorname{Ad}(P)$; echoing the associated-bundle construction of Chapter 4, it only says that gauge transformations are sections of this generally nontrivial bundle. Locally, after choosing a frame, those sections become the familiar matrix-valued functions.
[example: Gauge Group of the Product $SU(n)$-Bundle]
Let $P=M\times SU(n)$, and choose the product section $s_0(x)=(x,e)$. This section trivialises the adjoint group bundle by the map $\Theta:\operatorname{Ad}(P)\to M\times SU(n)$ defined by
\begin{align*}
\Theta([(x,h),a])=(x,hah^{-1}).
\end{align*}
To check that this does not depend on the representative, replace $(x,h)$ by $(x,hg)$ and $a$ by $g^{-1}ag$. Then
\begin{align*}
(hg)(g^{-1}ag)(hg)^{-1}=hgg^{-1}agg^{-1}h^{-1}=hah^{-1}.
\end{align*}
The inverse map sends $(x,b)$ to $[(x,e),b]$, since
\begin{align*}
\Theta([(x,e),b])=(x,ebe^{-1})=(x,b).
\end{align*}
For the other composition, start with $[(x,h),a]$. Applying $\Theta$ gives $(x,hah^{-1})$, and then the inverse gives $[(x,e),hah^{-1}]$. By the associated-bundle equivalence relation,
\begin{align*}
[(x,e),hah^{-1}]=[(x,h),h^{-1}(hah^{-1})h]=[(x,h),a].
\end{align*}
Thus smooth sections of $\operatorname{Ad}(P)$ are exactly smooth maps $u:M\to SU(n)$, and the gauge group identification gives
\begin{align*}
\operatorname{Gau}(P)\cong C^\infty(M,SU(n)).
\end{align*}
Under this identification, a function $u:M\to SU(n)$ corresponds to the vertical automorphism
\begin{align*}
\Phi_u(x,h)=(x,u(x)h).
\end{align*}
Writing the same automorphism in the equivariant-function convention means requiring $\Phi_u(x,h)=(x,h)\gamma(x,h)$. Since $(x,h)\gamma(x,h)=(x,h\gamma(x,h))$, equality with $(x,u(x)h)$ gives
\begin{align*}
h\gamma(x,h)=u(x)h.
\end{align*}
Multiplying on the left by $h^{-1}$ gives
\begin{align*}
\gamma(x,h)=h^{-1}u(x)h.
\end{align*}
The corresponding section of the adjoint group bundle is therefore
\begin{align*}
x\mapsto [(x,h),h^{-1}u(x)h]=[(x,e),u(x)],
\end{align*}
where the equality again uses the associated-bundle equivalence relation.
If the product section is changed to $s_v(x)=(x,v(x))$, the same automorphism is represented by the unique function $u_v:M\to SU(n)$ satisfying
\begin{align*}
\Phi_u(s_v(x))=s_v(x)u_v(x).
\end{align*}
Substituting the definitions gives
\begin{align*}
(x,u(x)v(x))=(x,v(x)u_v(x)).
\end{align*}
Equality of second components gives $u(x)v(x)=v(x)u_v(x)$, and multiplying on the left by $v(x)^{-1}$ gives
\begin{align*}
u_v(x)=v(x)^{-1}u(x)v(x).
\end{align*}
Thus the base function depends on the chosen section; even on a trivial bundle, the intrinsic object is the section of the adjoint group bundle, while local representatives change by conjugation.
[/example]
For infinitesimal questions the group bundle must be replaced by a Lie algebra bundle. Differentiating a family of representatives $u_t$ in two different local frames gives Lie-algebra elements related by the adjoint representation, not by equality. Thus ordinary $\mathfrak g$-valued functions on $M$ are available only after a global trivialisation, while the following associated bundle supplies the global home for infinitesimal gauge parameters.
[definition: Adjoint Lie Algebra Bundle]
Let $P\to M$ be a principal $G$-bundle with Lie algebra $\mathfrak g$. The adjoint Lie algebra bundle is
\begin{align*}
\operatorname{ad}(P):=P\times_G \mathfrak g,
\end{align*}
where $G$ acts on $\mathfrak g$ by the adjoint representation.
It is a vector bundle over $M$ with projection
\begin{align*}
\pi_{\operatorname{ad}}:\operatorname{ad}(P)\to M,\qquad [p,\xi]\mapsto \pi(p).
\end{align*}
Its smooth sections are the maps $\sigma:M\to \operatorname{ad}(P)$ satisfying $\pi_{\operatorname{ad}}\circ\sigma=\operatorname{id}_M$, denoted $\Gamma^\infty(\operatorname{ad}(P))$.
[/definition]
Sections of $\operatorname{ad}(P)$ may be represented by smooth functions $\xi:P\to\mathfrak g$ satisfying $\xi(pg)=\operatorname{Ad}_{g^{-1}}\xi(p)$. They are the infinitesimal gauge parameters.
[example: Infinitesimal Gauge Parameters]
Let $u_t:M\to G$ be a smooth path of gauge transformations on the product bundle $M\times G$ with $u_0(x)=e$. For each fixed $x\in M$, the curve $t\mapsto u_t(x)$ passes through the identity element of $G$, so its velocity at $t=0$ lies in $T_eG=\mathfrak g$:
\begin{align*}
\xi(x)=\frac{d}{dt}\Big|_{t=0}u_t(x)\in\mathfrak g.
\end{align*}
The corresponding vertical automorphism is
\begin{align*}
\Phi_t(x,h)=(x,u_t(x)h).
\end{align*}
In the equivariant-function convention, $\Phi_t(x,h)=(x,h)\gamma_t(x,h)$. Since $(x,h)\gamma_t(x,h)=(x,h\gamma_t(x,h))$, equality with $(x,u_t(x)h)$ gives
\begin{align*}
h\gamma_t(x,h)=u_t(x)h.
\end{align*}
Multiplying on the left by $h^{-1}$ gives
\begin{align*}
\gamma_t(x,h)=h^{-1}u_t(x)h.
\end{align*}
Differentiating at $t=0$ gives
\begin{align*}
\frac{d}{dt}\Big|_{t=0}\gamma_t(x,h)=h^{-1}\xi(x)h=\operatorname{Ad}_{h^{-1}}\xi(x).
\end{align*}
For a general principal bundle, let $\Phi_t$ be a smooth path of gauge transformations with $\Phi_0=\operatorname{id}_P$, and write
\begin{align*}
\Phi_t(p)=p\gamma_t(p).
\end{align*}
Since $\Phi_0(p)=p$, freeness of the right action gives $\gamma_0(p)=e$. Define
\begin{align*}
\xi(p)=\frac{d}{dt}\Big|_{t=0}\gamma_t(p)\in\mathfrak g.
\end{align*}
The equivariance law for $\gamma_t$ is
\begin{align*}
\gamma_t(pg)=g^{-1}\gamma_t(p)g.
\end{align*}
Differentiating this identity at $t=0$ gives
\begin{align*}
\xi(pg)=g^{-1}\xi(p)g=\operatorname{Ad}_{g^{-1}}\xi(p).
\end{align*}
Thus $\xi$ is an equivariant representative of a section of $\operatorname{ad}(P)$. In a local section $s_i:U_i\to P$, the section is represented by $\xi_i(x)=\xi(s_i(x))$. If $s_j(x)=s_i(x)g_{ij}(x)$ on an overlap, then
\begin{align*}
\xi_j(x)=\xi(s_i(x)g_{ij}(x))=\operatorname{Ad}_{g_{ij}(x)^{-1}}\xi_i(x).
\end{align*}
So infinitesimal gauge parameters are not globally ordinary $\mathfrak g$-valued functions unless a trivialisation has been chosen; intrinsically, they are sections of the adjoint Lie algebra bundle.
[/example]
## Connections as an Affine Space
Before studying the gauge action, we need to know what kind of space it acts on. Connections do not form a vector space, because there is no preferred zero connection on a general principal bundle, but the difference between two connections is tensorial and hence lives in a vector space.
[quotetheorem:6274]
[citeproof:6274]
The theorem uses the connection axioms, not only the fact that $\omega_1-\omega_0$ is a difference of forms. If a $\mathfrak g$-valued $1$-form $\alpha$ is not horizontal, then $\omega_0+\alpha$ can fail to reproduce fundamental vertical fields correctly; on a product bundle, adding the fibrewise Maurer-Cartan form doubles the vertical value instead of leaving it equal to the generator. If $\alpha$ is horizontal but not $G$-equivariant, then $\omega_0+\alpha$ fails the principal equivariance law, so it is not a connection form. The base connection $\omega_0$ is needed only to identify the affine space with a vector space of differences; choosing a different base connection translates the vector-space coordinate by an element of $\Omega^1(M;\operatorname{ad}(P))$. There is no intrinsic zero element of $\mathcal A(P)$, because the zero form itself violates vertical reproduction unless $\mathfrak g=0$. In a local section $s:U\to P$, a connection becomes a Lie-algebra-valued $1$-form $A=s^*\omega$, and the affine statement says that changing the connection changes $A$ by an $\operatorname{ad}(P)$-valued $1$-form.
[example: Electromagnetic Potentials]
For a principal $U(1)$-bundle $P\to M$, the Lie algebra is
\begin{align*}
\mathfrak u(1)=T_1U(1)=i\mathbb R.
\end{align*}
Because $U(1)$ is abelian, for every $g,h\in U(1)$ one has $gh=hg$, hence
\begin{align*}
ghg^{-1}=hgg^{-1}=h.
\end{align*}
Therefore the conjugation action of $U(1)$ on itself is trivial, and differentiating this trivial conjugation action at the identity gives the trivial adjoint action on $i\mathbb R$:
\begin{align*}
\operatorname{Ad}_g(\xi)=\xi
\end{align*}
for all $g\in U(1)$ and $\xi\in i\mathbb R$.
Thus
\begin{align*}
\operatorname{ad}(P)=P\times_{U(1)} i\mathbb R
\end{align*}
is formed using the trivial action. If two local sections satisfy $s_j=s_i g_{ij}$ on an overlap, then an adjoint-valued local representative transforms by
\begin{align*}
\xi_j=\operatorname{Ad}_{g_{ij}^{-1}}\xi_i=\xi_i.
\end{align*}
The same calculation applies to differences of local connection forms: if $A_1$ and $A_0$ are two local electromagnetic potentials, then
\begin{align*}
A_1-A_0\in \Omega^1(U;i\mathbb R)
\end{align*}
and on overlaps its representative is unchanged by conjugation. Hence local differences of electromagnetic potentials glue as ordinary $i\mathbb R$-valued $1$-forms. This is exactly the abelian boundary case: for nonabelian structure groups the overlap rule contains the nontrivial factor $\operatorname{Ad}_{g_{ij}^{-1}}$, so the same differences are genuinely adjoint-bundle-valued rather than ordinary Lie-algebra-valued forms.
[/example]
This abelian example also foreshadows a special simplification in curvature: conjugation becomes invisible for $U(1)$, so the field strength is unchanged by gauge transformations.
## Gauge Action on Connections and Curvature
The main computational question is how a vertical automorphism changes a connection. The global rule is pullback of the connection form by the principal bundle automorphism; local formulas then come from expressing that pullback in a chosen section.
[definition: Gauge Action on Connections]
Let $P\to M$ be a smooth principal $G$-bundle. The gauge action on connections is the map
\begin{align*}
\operatorname{Gau}(P)\times \mathcal A(P)\to \mathcal A(P),\qquad (\Phi,\omega)\mapsto \Phi\cdot\omega:=\Phi^*\omega.
\end{align*}
[/definition]
This convention is contravariant at the level of forms. With the usual composition order in $\operatorname{Gau}(P)$, it is a right action written on the left, since $(\Phi_2\circ\Phi_1)^*\omega=\Phi_1^*(\Phi_2^*\omega)$. To obtain a genuine left action, some sources define $\Phi\cdot\omega=(\Phi^{-1})^*\omega$ instead. To use the global rule in calculations, we now need its expression in a local section; the next theorem computes the extra term produced when the section is moved vertically. With the present convention, if $\Phi(s(x))=s(x)u(x)$ in a local section, then the local representative of $\Phi\cdot\omega$ in the same section is also the representative of $\omega$ in the changed section $s'=su$.
[quotetheorem:6275]
[citeproof:6275]
The hypotheses in the local formula prevent several common misreadings. The section $s$ must be fixed when defining $u$: if one changes to an unrelated section $t=sv$ and still uses the same function $u$, the expression describes a different local representative rather than the original gauge action. The automorphism must be vertical; an automorphism covering a diffeomorphism $f:U\to U$ would also pull back the base $1$-form by $f$, so the displayed formula would miss the base-change term. Smoothness of $u$ is needed because $u^{-1}du$ appears; a merely continuous gauge representative on a product bundle has no local connection transformation formula of this kind. The formula also does not say that connection forms transform tensorially; the extra Maurer-Cartan term is precisely the failure of tensoriality. Curvature will repair this defect by eliminating the inhomogeneous derivative term.
[example: Electromagnetic Gauge Transformation]
Let $G=U(1)$, let $A\in\Omega^1(U;i\mathbb R)$ be a local connection form, and write the local gauge representative as $u=e^{i\chi}$ for a smooth function $\chi:U\to\mathbb R$. The local gauge transformation formula gives
\begin{align*}
A'=u^{-1}Au+u^{-1}du.
\end{align*}
Because $u(x)\in U(1)$ is a scalar complex number, it commutes with every element of $i\mathbb R$. Hence, for each $X\in T_xU$,
\begin{align*}
(u^{-1}Au)_x(X)=u(x)^{-1}A_x(X)u(x)=u(x)^{-1}u(x)A_x(X)=A_x(X).
\end{align*}
Thus $u^{-1}Au=A$.
It remains to compute the inhomogeneous term. Since $u=e^{i\chi}$, the chain rule gives
\begin{align*}
du=d(e^{i\chi})=i e^{i\chi}\,d\chi.
\end{align*}
Using $u^{-1}=e^{-i\chi}$, we get
\begin{align*}
u^{-1}du=e^{-i\chi}\bigl(i e^{i\chi}\,d\chi\bigr)=i e^{-i\chi}e^{i\chi}\,d\chi=i\,d\chi.
\end{align*}
Substituting these two computations into the transformation formula gives
\begin{align*}
A'=u^{-1}Au+u^{-1}du=A+i\,d\chi.
\end{align*}
After identifying $i\mathbb R$ with $\mathbb R$ by removing the common factor $i$, this is the usual electromagnetic gauge replacement $A\mapsto A+d\chi$.
[/example]
The electromagnetic case shows that a physically important quantity should depend on the connection through its curvature rather than through the choice of potential. For nonabelian structure groups, the next theorem gives the exact covariance law and explains why conjugation-invariant expressions are the right observables.
[quotetheorem:6276]
[citeproof:6276]
The curvature formula depends on using the same convention as the connection formula; reversing the gauge action reverses the adjoint factor. The transformation law for $A'$ is also needed: if an arbitrary $\mathfrak g$-valued $1$-form is substituted for $A'$ without the Maurer-Cartan term, the cancellation in the proof need not occur. For example, replacing $A$ by $u^{-1}Au$ alone gives extra derivatives of $u$ in $dA'$. Smoothness of $u$ is again essential because both $du$ and the Maurer-Cartan equation are used. The theorem does not make $F_A$ gauge-invariant as a Lie-algebra-valued form unless the adjoint action is trivial; for $G=SU(2)$, a constant noncentral gauge transformation sends a noncentral curvature value $X$ to $u^{-1}Xu\ne X$. What survives generally are conjugation-invariant expressions built from curvature, such as traces in matrix representations.
[example: Gauge-Invariant Trace Forms]
Let $G\subset GL(n,\mathbb C)$ be a matrix Lie group, and let the local curvature be represented by a matrix-valued $2$-form $F_A$. Under a gauge transformation, the curvature transformation law gives
\begin{align*}
F_{A'}=u^{-1}F_Au.
\end{align*}
Here $u$ and $u^{-1}$ are matrix-valued $0$-forms, so multiplying by them does not introduce any graded signs in wedge products. Therefore
\begin{align*}
F_{A'}\wedge F_{A'}=(u^{-1}F_Au)\wedge(u^{-1}F_Au).
\end{align*}
Associativity of matrix multiplication together with $uu^{-1}=I$ gives
\begin{align*}
(u^{-1}F_Au)\wedge(u^{-1}F_Au)=u^{-1}F_A\wedge(uu^{-1})F_Au.
\end{align*}
Substituting $uu^{-1}=I$ and using that $I$ acts as the identity matrix gives
\begin{align*}
u^{-1}F_A\wedge(uu^{-1})F_Au=u^{-1}F_A\wedge F_Au.
\end{align*}
Since the exterior product is taken entrywise while the matrix factors multiply in order, this is
\begin{align*}
u^{-1}F_A\wedge F_Au=u^{-1}(F_A\wedge F_A)u.
\end{align*}
Taking the trace, we get
\begin{align*}
\operatorname{tr}(F_{A'}\wedge F_{A'})=\operatorname{tr}\bigl(u^{-1}(F_A\wedge F_A)u\bigr).
\end{align*}
By cyclicity of trace for the ordinary matrix factors $u$ and $u^{-1}$,
\begin{align*}
\operatorname{tr}\bigl(u^{-1}(F_A\wedge F_A)u\bigr)=\operatorname{tr}\bigl((F_A\wedge F_A)uu^{-1}\bigr).
\end{align*}
Using $uu^{-1}=I$ once more gives
\begin{align*}
\operatorname{tr}\bigl((F_A\wedge F_A)uu^{-1}\bigr)=\operatorname{tr}(F_A\wedge F_A).
\end{align*}
Thus $\operatorname{tr}(F_A\wedge F_A)$ is unchanged by gauge transformation; the trace removes the remaining conjugation, which is the local algebraic mechanism behind gauge-invariance of Chern-Weil forms.
[/example]
## Infinitesimal Gauge Action
The last question is what the gauge action looks like to first order near the identity transformation. This linearisation is the starting point for deformation theory, moduli spaces of connections, and the differential equations used to impose gauge-fixing conditions.
[definition: Covariant Exterior Derivative on Adjoint-Valued Functions]
Let $A\in\Omega^1(U;\mathfrak g)$ be a local connection form. The covariant exterior derivative on adjoint-valued functions is the map
\begin{align*}
d_A:C^\infty(U,\mathfrak g)\to \Omega^1(U;\mathfrak g),\qquad \xi\mapsto d_A\xi:=d\xi+[A,\xi].
\end{align*}
[/definition]
The bracket term corrects the ordinary derivative so that the expression transforms according to the adjoint bundle on overlaps. Since infinitesimal gauge parameters are sections of $\operatorname{ad}(P)$, this derivative is the expected candidate for the tangent vector to a gauge orbit. We can see the formula directly by differentiating the finite local gauge action.
[explanation: Infinitesimal Gauge Action in a Local Gauge]
Let $u_t:U\to G$ be a smooth path of gauge transformations with $u_0=e$ and
\begin{align*}
\left.\frac{d}{dt}\right|_{t=0}u_t=\xi\in C^\infty(U,\mathfrak g).
\end{align*}
For the local action
\begin{align*}
A_t=\operatorname{Ad}_{u_t^{-1}}A+u_t^{-1}du_t,
\end{align*}
the first term differentiates to $[A,\xi]$: indeed $u_t^{-1}=\exp(-t\xi)+O(t^2)$ pointwise, so conjugating $A$ by $u_t^{-1}$ contributes $-[\xi,A]=[A,\xi]$ to first order. The second term differentiates to $d\xi$, because $u_0=e$ and the derivative of $u_t^{-1}du_t$ at $t=0$ is the exterior derivative of the infinitesimal parameter. Therefore
\begin{align*}
\left.\frac{d}{dt}\right|_{t=0}A_t=d\xi+[A,\xi]=d_A\xi.
\end{align*}
Thus the tangent space to the gauge orbit through $A$ is locally described by the image of
\begin{align*}
d_A:\Omega^0(U;\mathfrak g)\to\Omega^1(U;\mathfrak g).
\end{align*}
[/explanation]
The sign in this formula is tied to both choices already fixed: right principal action and pullback action $\Phi\cdot\omega=\Phi^*\omega$. The path must pass through the identity, since differentiating a path with $u_0\ne e$ linearises at a different point of the gauge group and first conjugates the base connection. The parameter must be a smooth adjoint-valued function; if $\xi$ is only continuous, $d\xi$ is not a smooth $1$-form and the displayed tangent vector is not defined in $\Omega^1(U;\mathfrak g)$. The formula does not describe every tangent vector in the affine space $\mathcal A(P)$: for a connection with stabiliser, even nonzero $\xi$ may give $d_A\xi=0$, while a general adjoint-valued $1$-form need not lie in the image of $d_A$. Thus a connection is not studied in isolation: gauge-equivalent connections represent the same principal geometry expressed in different internal frames, while curvature transforms by conjugation and carries the gauge-covariant field strength.
Gauge transformations explain how the same connection can look different in different local frames, but the geometry becomes especially vivid when we move around actual paths. The next chapter develops parallel transport and holonomy, turning the connection into a rule for moving data along curves and extracting global information from loops.
# 10. Parallel Transport and Holonomy
This chapter belongs to the part of the course where principal connections become geometric tools rather than only differential forms on the total space. The main goals are to construct parallel transport from the horizontal distribution, package transport around loops as holonomy, and explain why flat connections are classified by monodromy representations. The prerequisites are the connection form and horizontal distributions from Chapter 6, associated bundles from Chapter 4, curvature from Chapter 8, Frobenius integrability, and the fundamental group. In the previous chapters, the connection form and curvature measured horizontal directions and their failure to close under brackets; here we integrate those horizontal directions along paths and compare the resulting global data with the topology of the base.
## Horizontal Lifts and Parallel Transport
A connection on a principal bundle tells us which tangent vectors are horizontal, but a geometric comparison between two fibres needs a pathwise construction. The first question is whether a path in the base has a unique horizontal lift once we choose its starting point upstairs.
[definition: Horizontal Lift of a Path]
Let $\pi:P\to M$ be a smooth principal $G$-bundle with connection form $\omega\in \Omega^1(P;\mathfrak g)$. Let $\gamma:[0,1]\to M$ be a smooth path. A horizontal lift of $\gamma$ is a smooth path $\tilde\gamma:[0,1]\to P$ such that
\begin{align*}
\pi\circ \tilde\gamma &= \gamma, & \omega_{\tilde\gamma(t)}(\dot{\tilde\gamma}(t)) &= 0 \quad \text{for all } t\in[0,1].
\end{align*}
[/definition]
The definition says that the lifted velocity remains in the horizontal distribution at every time. The connection is doing more than selecting a complement at each point: along a curve it gives an ordinary differential equation on the total space.
[quotetheorem:6277]
[citeproof:6277]
This result packages a connection as a transport rule: a starting point in the initial fibre determines an endpoint in the terminal fibre. The smoothness of the bundle, connection, and path are used to reduce horizontality to an ordinary differential equation; if the horizontal distribution were only a nonsmooth choice of complements, the ODE argument could fail and the endpoint assignment need not be well-defined. Compactness of $[0,1]$ is also used when local solutions are patched over finitely many trivializing intervals. The theorem does not say that transport depends only on endpoints: on the round sphere, Levi-Civita transport between two tangent planes along different arcs may give different rotations. To use this endpoint assignment repeatedly, we need to name it as a map between fibres and record its compatibility with the principal right action.
[definition: Principal Parallel Transport]
Let $\pi:P\to M$ be a principal $G$-bundle with connection, and let $\gamma:[0,1]\to M$ be a smooth path. The parallel transport along $\gamma$ is the map
\begin{align*}
\operatorname{PT}_{\gamma}:P_{\gamma(0)}\to P_{\gamma(1)},\qquad \operatorname{PT}_{\gamma}(p_0)=\tilde\gamma(1),
\end{align*}
where $\tilde\gamma$ is the horizontal lift of $\gamma$ starting at $p_0$.
[/definition]
Because the connection is principal, $R_g$ sends horizontal vectors to horizontal vectors. Hence $\operatorname{PT}_{\gamma}(p g)=\operatorname{PT}_{\gamma}(p)g$, so parallel transport is an isomorphism of right $G$-torsors between fibres.
[example: Parallel Transport in a Trivial Bundle]
Let $P=M\times G$, choose the global section $s(x)=(x,e)$, and write the connection in this trivialization as $A\in\Omega^1(M;\mathfrak g)$. For a path $\gamma:[0,1]\to M$, set
\begin{align*}
a(t)=A_{\gamma(t)}(\dot\gamma(t))\in\mathfrak g.
\end{align*}
A lift through $(\gamma(0),g_0)$ has the form $\tilde\gamma(t)=(\gamma(t),g(t))$. In this trivialization, the horizontal condition $\omega_{\tilde\gamma(t)}(\dot{\tilde\gamma}(t))=0$ becomes
\begin{align*}
g'(t)g(t)^{-1}+a(t)=0.
\end{align*}
Thus the fibre coordinate is determined by
\begin{align*}
g'(t)g(t)^{-1}=-A_{\gamma(t)}(\dot\gamma(t)),\qquad g(0)=g_0.
\end{align*}
When $G$ is abelian, the Lie algebra elements $a(t)$ commute under the exponential calculation. Define
\begin{align*}
B(t)=-\int_0^t a(\tau)\,d\tau.
\end{align*}
Then $B'(t)=-a(t)$. Put
\begin{align*}
h(t)=\exp(B(t)),\qquad g(t)=h(t)g_0.
\end{align*}
Since the abelian exponential satisfies $h'(t)=B'(t)\exp(B(t))$, we get
\begin{align*}
g'(t)g(t)^{-1}=h'(t)g_0g_0^{-1}h(t)^{-1}.
\end{align*}
Cancelling $g_0g_0^{-1}=e$ gives
\begin{align*}
g'(t)g(t)^{-1}=h'(t)h(t)^{-1}.
\end{align*}
Substituting $h'(t)=B'(t)\exp(B(t))$ and $h(t)^{-1}=\exp(-B(t))$ gives
\begin{align*}
g'(t)g(t)^{-1}=B'(t)\exp(B(t))\exp(-B(t))=B'(t)=-a(t).
\end{align*}
Also $g(0)=\exp(0)g_0=g_0$. Therefore, in the abelian case,
\begin{align*}
\operatorname{PT}_{\gamma}(\gamma(0),g_0)=\left(\gamma(1),\exp\left(-\int_0^1 A_{\gamma(t)}(\dot\gamma(t))\,dt\right)g_0\right).
\end{align*}
For nonabelian $G$, the values $a(t)$ sampled at different times need not commute, so the endpoint is governed by the time-ordered solution of the differential equation rather than by a single ordinary exponential of the total integral.
[/example]
The example shows why parallel transport is the global shadow of the connection form. Reversing and concatenating paths now give the basic calculus of transport maps.
[quotetheorem:6278]
[citeproof:6278]
The constant, inverse, and concatenation rules are the pathwise algebra that later becomes the algebra of holonomy. The shared initial and terminal points in the concatenation hypothesis are essential: without $\gamma_1(1)=\gamma_2(0)$, the two transport maps cannot be composed. The inverse statement also depends on reversing the same path; a different return path may give a nonidentity loop transport. These operations are stated on the principal bundle, but many applications involve vector bundles built from representations. A failure mode to keep in mind is a vector bundle whose frames return rotated after a closed path: the principal transport has a clean $G$-equivariant description, while the vector transport depends on the chosen representation of $G$. We therefore need the associated transport rule, because it is the bridge from principal connections to the covariant transport of vectors and tensors.
[definition: Associated Parallel Transport]
Let $E=P\times_G V$ be the vector bundle associated to a principal $G$-bundle $P$ and a representation $\rho:G\to GL(V)$. For a path $\gamma:[0,1]\to M$, the associated parallel transport is the map
\begin{align*}
\operatorname{PT}^E_{\gamma}:E_{\gamma(0)}\to E_{\gamma(1)},\qquad \operatorname{PT}^E_{\gamma}([p,v])=[\operatorname{PT}_{\gamma}(p),v],
\end{align*}
where $p\in P_{\gamma(0)}$ and $v\in V$.
[/definition]
This is well-defined because principal parallel transport is $G$-equivariant. Thus the principal-bundle construction recovers the usual notion of parallel transport for vector bundles with covariant derivative.
## Holonomy Groups
Transport around a closed loop starts and ends in the same fibre. Endpoint comparison along open paths is not canonical, because different terminal fibres have no preferred identification. For a loop, that obstruction disappears: the starting and ending fibres are the same right $G$-torsor, so the remaining defect is a single element of the structure group. The collection of all such elements measures the twisting detected by the connection.
[definition: Holonomy Group at a Point]
Let $\pi:P\to M$ be a principal $G$-bundle with connection, and let $p\in P$ with $x=\pi(p)$. The holonomy group at $p$ is
\begin{align*}
\operatorname{Hol}_p(\omega)=\{g\in G: \operatorname{PT}_{\gamma}(p)=p g \text{ for some piecewise smooth loop } \gamma \text{ based at } x\}.
\end{align*}
[/definition]
The torsor property gives a unique element $g$ for each loop and starting point $p$. To treat holonomy as an invariant rather than a list of loop values, we need to know that these elements form a subgroup of $G$.
[quotetheorem:6279]
[citeproof:6279]
The theorem justifies the word group in the definition of holonomy. Each hypothesis has a role: loops are needed so that all transports land back in the same fibre, the principal right action is needed to translate endpoints into elements of $G$, and concatenation of piecewise smooth loops supplies the multiplication. If open paths were allowed, the endpoint would lie in a different fibre and there would be no subgroup of $G$ attached to a single point. The order in the proof is important: for right principal actions, the group element from the later transported loop appears on the left of the element already attached to the starting point. Since the definition still used a chosen point $p$ in the fibre, the next issue is how the subgroup changes when that choice is replaced by another point of the same fibre.
[quotetheorem:6280]
[citeproof:6280]
Changing the point in a fibre accounts for one kind of choice, but a connected base allows loops based at different points. The conjugacy formula says that the subgroup itself is not canonically attached to the base point alone unless the structure group is abelian or a frame has been fixed; for instance, changing a frame in an $SO(n)$-frame bundle conjugates the rotation matrices that record holonomy. The condition $h\in G$ also matters because all other points of the same fibre have the form $p h$ and the right action is free and transitive. This explains why holonomy is usually recorded as a [conjugacy class](/page/Conjugacy%20Class) rather than as a literal subgroup of $G$. To compare holonomy at different basepoints, we need a path between them, and the transport along that path supplies the conjugating element.
[quotetheorem:6281]
[citeproof:6281]
The subgroup therefore depends on auxiliary choices, but its conjugacy class is a global invariant of the connection on a connected base. Connectedness is not cosmetic: over a disconnected base there may be no path from $x$ to $y$, and holonomy on different components is independent data. The formula also depends on the chosen path $\alpha$; replacing $\alpha$ by another path can conjugate by an element already coming from holonomy, so the invariant statement is still a conjugacy class.
[example: Holonomy of a Mobius-Type O(1)-Bundle]
Model the frame bundle as
\begin{align*}
P=(\mathbb R\times O(1))/\sim,\qquad (t+1,\varepsilon)\sim (t,-\varepsilon),
\end{align*}
with projection $[t,\varepsilon]\mapsto [t]\in \mathbb R/\mathbb Z=S^1$ and right action
\begin{align*}
[t,\varepsilon]\cdot \delta=[t,\varepsilon\delta],\qquad \delta\in O(1)=\{\pm 1\}.
\end{align*}
Because $O(1)$ is discrete, the fibres have zero tangent space, so the horizontal lift of the positively oriented generator $\gamma(t)=[t]$ starting at $p=[0,1]$ is the lifted path
\begin{align*}
\tilde\gamma(t)=[t,1].
\end{align*}
Its endpoint is
\begin{align*}
\tilde\gamma(1)=[1,1]=[0,-1]=[0,1]\cdot(-1)=p(-1),
\end{align*}
where the middle equality is exactly the identification defining the nontrivial bundle. Thus the generator of $\pi_1(S^1)$ contributes the holonomy element $-1$, while the constant loop contributes
\begin{align*}
p=p\cdot 1.
\end{align*}
Since every holonomy element lies in $O(1)=\{1,-1\}$ and both elements occur, we get
\begin{align*}
\operatorname{Hol}_p(\omega)=\{1,-1\}=O(1).
\end{align*}
The nontrivial loop therefore flips a frame, which is precisely the obstruction to choosing a globally consistent nonzero frame of the underlying real line bundle.
[/example]
This example is flat in the curvature sense because the structure group is discrete, yet the holonomy is nontrivial. Curvature controls infinitesimal holonomy, while global topology can still contribute through loops not contractible to a point.
[example: Rotations from Levi-Civita Transport on the Sphere]
Let $S^2$ be the unit round sphere, and let $P=SO(S^2)$ be the oriented orthonormal frame bundle with its Levi-Civita connection. Fix an oriented geodesic triangle $T$ with interior angles $\alpha,\beta,\chi$ and positively oriented boundary $\partial T$. Along each side, the boundary tangent is parallel because each side is a geodesic, so the only turning of the boundary tangent occurs at the three corners. The signed exterior turns at the corners are $\pi-\alpha$, $\pi-\beta$, and $\pi-\chi$, so the total signed boundary turning is
\begin{align*}
(\pi-\alpha)+(\pi-\beta)+(\pi-\chi)=3\pi-(\alpha+\beta+\chi).
\end{align*}
On the unit sphere, *Gauss-Bonnet for a geodesic triangle* gives
\begin{align*}
\operatorname{area}(T)=\alpha+\beta+\chi-\pi.
\end{align*}
Equivalently,
\begin{align*}
\alpha+\beta+\chi=\pi+\operatorname{area}(T).
\end{align*}
Substituting this into the turning formula gives
\begin{align*}
3\pi-(\alpha+\beta+\chi)=3\pi-\bigl(\pi+\operatorname{area}(T)\bigr)=2\pi-\operatorname{area}(T).
\end{align*}
Modulo a full rotation by $2\pi$, this boundary turning is the same as $-\operatorname{area}(T)$.
A parallel transported tangent vector has zero covariant derivative along each side, so its angle relative to a parallel frame stays constant on that side. Therefore, when the boundary tangent accumulates signed turning $2\pi-\operatorname{area}(T)$, the transported frame returns rotated by the opposite signed angle:
\begin{align*}
-\bigl(2\pi-\operatorname{area}(T)\bigr)\equiv \operatorname{area}(T)\pmod{2\pi}.
\end{align*}
Thus parallel transport around $\partial T$ acts on the fibre by rotation through the signed spherical area enclosed by the triangle. Since each fibre of $SO(S^2)\to S^2$ consists of oriented orthonormal frames of a tangent plane, the holonomy transformations lie in $SO(2)$; loops enclosing different signed areas can therefore produce different rotations.
[/example]
The sphere example motivates the infinitesimal theorem relating curvature and holonomy. In this course we state the result as a guidepost rather than prove it, because the proof requires a longer analysis of families of horizontal lifts and local generation of Lie subgroups.
[quotetheorem:6282]
This formulation avoids choosing a separate gauge at the basepoint: the reachable set $P(p)$ is the holonomy bundle through $p$, and the curvature form already takes values in the fixed Lie algebra $\mathfrak g$. The theorem concerns restricted holonomy, so it captures the part generated by contractible loops; a flat connection on $S^1$ shows why this restriction matters, since curvature can vanish while full holonomy still contains nontrivial monodromy from $\pi_1(S^1)$. This result points back to the gauge-theoretic transformation laws of Chapter 9 and forward to the computations of Chapter 12, where flatness and curvature constraints are studied through local gauges and holonomy representations.
[illustration:holonomy-horizontal-lift]
## Flat Connections and Monodromy
A connection is flat when its curvature vanishes. The local failure mode being removed is the small-loop effect: if curvature is nonzero, transporting around arbitrarily small rectangles can change a frame to first order in the enclosed area. The global question is what remains after this infinitesimal obstruction has vanished, because noncontractible loops can still carry transport data.
[definition: Flat Connection]
Let $\pi:P\to M$ be a principal $G$-bundle with connection form $\omega$ and curvature form $\Omega$. The connection is flat if
\begin{align*}
\Omega=0.
\end{align*}
[/definition]
Flatness means the horizontal distribution is involutive. The next issue is local path dependence: we need to know whether vanishing curvature makes nearby transport depend only on endpoints. Without shrinking the neighbourhood, this statement can fail even for flat connections: on $S^1$, a flat $U(1)$-connection may have nontrivial transport around the whole circle, although every sufficiently small interval has endpoint-only transport.
[quotetheorem:6283]
[citeproof:6283]
This theorem explains why flat connections are locally gauge-equivalent to the product connection. The path-connected neighbourhood matters because the statement compares paths with fixed endpoints inside a single local horizontal leaf; if $U$ were not path-connected, the assertion would not compare points in different components. The theorem is local only: it does not rule out nontrivial transport around loops that leave the neighbourhood or represent nontrivial elements of the fundamental group. The remaining data is global: loops may move from one local horizontal sheet to another by an element of $G$.
[definition: Monodromy Representation of a Flat Connection]
Let $M$ be connected, let $P\to M$ be a principal $G$-bundle with a flat connection, and choose $p\in P_x$. The monodromy representation based at $p$ is the homomorphism
\begin{align*}
\rho_p:\pi_1(M,x)\to G,
\end{align*}
where $\rho_p([\gamma])$ is defined by $\operatorname{PT}_{\gamma}(p)=p\rho_p([\gamma])$.
[/definition]
For a flat connection, homotopic loops with fixed endpoints give the same transport. This is the point at which topology replaces curvature: a contractible small loop gives no local contribution, while a loop representing a nonzero element of $\pi_1(M,x)$ may still move the fibre. To classify flat bundles, we need the converse construction as well: a representation of the fundamental group should rebuild the bundle and its flat connection.
[quotetheorem:6284]
[citeproof:6284]
The correspondence says that flat principal bundles are locally constant geometric objects, in the same sense that local systems are vector bundles whose transition functions are locally constant. The connected smooth manifold hypothesis keeps a single universal cover and a single fundamental group representation in control of the whole bundle; over a disconnected base, the data would be chosen component by component. The theorem classifies flat bundles with their flat connections, not arbitrary principal bundles, and it uses the usual universal-cover framework for smooth manifolds. The conjugacy is exactly the loss of the chosen point in the reference fibre. Their classification is not governed by curvature, but by how the fibre is reidentified after going around loops in the base; this is the form in which the same idea appears in gauge theory as the passage from flat connections to representation varieties.
[example: Flat U(1)-Connections on S1]
Let $M=S^1=\mathbb R/\mathbb Z$ and $G=U(1)$. A flat principal $U(1)$-bundle with connection determines a monodromy homomorphism
\begin{align*}
\rho:\pi_1(S^1)\to U(1).
\end{align*}
Under $\pi_1(S^1)\cong \mathbb Z$, this homomorphism is determined by the single value
\begin{align*}
u=\rho(1).
\end{align*}
Indeed, for $n>0$,
\begin{align*}
\rho(n)=\rho(\underbrace{1+\cdots+1}_{n\text{ times}})=\rho(1)^n=u^n.
\end{align*}
Also $\rho(0)=1$, and
\begin{align*}
1=\rho(0)=\rho(n+(-n))=\rho(n)\rho(-n).
\end{align*}
Multiplying by $\rho(n)^{-1}$ gives
\begin{align*}
\rho(-n)=\rho(n)^{-1}=u^{-n}.
\end{align*}
On the trivial bundle $S^1\times U(1)$, take the constant connection form represented by
\begin{align*}
A=i\theta\,dt,
\end{align*}
where $t$ has period $1$. For the positively oriented generator $\gamma(t)=[t]$, $0\le t\le 1$, we have
\begin{align*}
A_{\gamma(t)}(\dot\gamma(t))=i\theta\,dt\left(\frac{d}{dt}\right)=i\theta.
\end{align*}
Writing a lifted path as $(\gamma(t),g(t))$, the parallel transport equation in this trivialization is
\begin{align*}
g'(t)g(t)^{-1}=-i\theta.
\end{align*}
Since $U(1)$ is abelian, set
\begin{align*}
g(t)=\exp(-i\theta t)g_0.
\end{align*}
Then
\begin{align*}
g'(t)=(-i\theta)\exp(-i\theta t)g_0.
\end{align*}
Also
\begin{align*}
g(t)^{-1}=g_0^{-1}\exp(i\theta t).
\end{align*}
Therefore
\begin{align*}
g'(t)g(t)^{-1}=(-i\theta)\exp(-i\theta t)g_0g_0^{-1}\exp(i\theta t).
\end{align*}
Using $g_0g_0^{-1}=1$, this becomes
\begin{align*}
g'(t)g(t)^{-1}=(-i\theta)\exp(-i\theta t)\exp(i\theta t).
\end{align*}
Since $\exp(-i\theta t)\exp(i\theta t)=1$, we get
\begin{align*}
g'(t)g(t)^{-1}=-i\theta.
\end{align*}
At $t=1$,
\begin{align*}
g(1)=\exp(-i\theta)g_0.
\end{align*}
Thus the holonomy around the positively oriented generator is $e^{-i\theta}$.
The curvature is zero: first,
\begin{align*}
dA=d(i\theta\,dt)=i\theta\,d(dt)=0.
\end{align*}
Second, $U(1)$ is abelian, so the bracket term in the curvature vanishes. Hence flat $U(1)$-connections on $S^1$ can still have arbitrary monodromy $u\in U(1)$; choosing $\theta$ with $e^{-i\theta}=u$ realizes that monodromy on the trivial bundle.
[/example]
This final example is the model for the general picture. Curvature detects infinitesimal failure of parallel transport to close up, while flat monodromy records the remaining representation-theoretic data carried by noncontractible loops.
Parallel transport and holonomy show how local differential data produces global monodromy, but that picture becomes richer when a connection must also respect a chosen reduction. The next chapter studies precisely that compatibility problem, asking when induced geometry on a reduced bundle is preserved by the ambient connection.
# 11. Connections on Reductions and Induced Geometry
Chapters 1 through 6 developed principal bundles, reductions of structure group, and principal connections as separate pieces of geometry. This chapter studies what happens when these pieces are present at the same time: a connection on a principal $G$-bundle may or may not respect a reduction to a closed Lie subgroup $H \le G$. The guiding question is whether parallel transport preserves the smaller geometric structure encoded by the reduction, rather than moving it inside the larger bundle.
The main prerequisites are principal connection forms, horizontal distributions, fundamental vector fields, associated bundles, and the frame bundle of a smooth manifold. For frame bundles, preservation of a reduction becomes the familiar question of when a linear connection preserves a metric, an orientation, or both. The chapter then uses the solder form to pass from principal-bundle data to torsion on $TM$, ending with the frame-bundle formulation of the Levi-Civita theorem.
## Connections Preserving Reductions
Suppose a principal $G$-bundle $P \to M$ admits an $H$-reduction $Q \subset P$. A connection on $P$ gives horizontal subspaces in $TP$, but these horizontal subspaces need not be tangent to $Q$. The first problem is to characterize when the connection transports points of $Q$ through points of $Q$, so that the reduced geometric structure is preserved along curves in the base.
[definition: Preserved Reduction]
Let $H \le G$ be a closed Lie subgroup, let $Q \subset P$ be an $H$-reduction of a principal $G$-bundle $\pi:P \to M$, and let $A \subset TP$ be a principal connection on $P$ with horizontal distribution $A_p \subset T_pP$. The reduction $Q$ is preserved by $A$ if
\begin{align*}
A_q \subset T_qQ
\end{align*}
for every $q \in Q$.
[/definition]
This definition says that every horizontal lift starting from the reduced bundle begins tangent to the reduced bundle. To use it in computations, we need an equivalent criterion stated directly in terms of the connection form restricted to $TQ$.
[quotetheorem:6285]
[citeproof:6285]
The theorem packages the geometric preservation condition into a test on the connection form. The closedness of $H$ matters because it ensures that an $H$-reduction is a genuine principal subbundle with Lie algebra $\mathfrak h$, so that vertical tangent vectors in $Q$ are exactly the fundamental vector fields generated by elements of $\mathfrak h$. A boundary case is a dense non-closed subgroup $H \subset S^1$: its Lie algebra agrees with that of $S^1$, but its orbits do not form the embedded principal subbundles required by the argument. The reduction hypothesis also matters. For instance, a hypersurface in $P$ that meets each fibre but is not stable under the right $H$-action may carry a restricted $\mathfrak g$-valued form, yet this restriction has no principal equivariance law and is not a connection on a principal $H$-bundle.
The connection-form axioms are doing real work here. Reproduction on fundamental vector fields identifies the vertical part, equivariance guarantees compatibility with the right action, and the kernel of the form recovers the horizontal distribution. Without the $\mathfrak h$-valued restriction, horizontal lifts from $Q$ can immediately point out of $Q$; in the frame-bundle case this is exactly what happens when a linear connection transports orthonormal frames to non-orthonormal frames. The theorem does not say that a connection preserving $Q$ preserves every submanifold contained in $P$, nor does it imply that an arbitrary tensorial condition on $P$ gives a reduction. Its role is narrower and stronger: once a genuine reduction is present, preservation can be checked by the restricted connection form, which is the entry point for decomposing a non-preserving connection into compatible and obstructing pieces.
[example: Metric Reduction of the Frame Bundle]
Let $M$ be an $n$-manifold, let $F(M)$ be its full frame bundle with structure group $GL(n,\mathbb R)$, and let $O_g(M)\subset F(M)$ be the orthonormal frame bundle determined by a Riemannian metric $g$. By the *Preservation Criterion for Reductions*, a $GL(n,\mathbb R)$-connection form $\omega\in\Omega^1(F(M);\mathfrak{gl}(n,\mathbb R))$ preserves $O_g(M)$ exactly when $\omega(v)\in\mathfrak{o}(n)$ for every $v\in TO_g(M)$.
The Lie algebra condition is skew-symmetry. If $A(t)\in O(n)$ and $A(0)=I$, then $A(t)^\topA(t)=I$. Differentiating the matrix identity at $t=0$ gives
\begin{align*}
\dot A(0)^\topA(0)+A(0)^\top\dot A(0)=0.
\end{align*}
Since $A(0)=I$, this becomes
\begin{align*}
\dot A(0)^\top+\dot A(0)=0.
\end{align*}
Thus
\begin{align*}
\mathfrak{o}(n)=\{B\in\mathfrak{gl}(n,\mathbb R):B^\top+B=0\}.
\end{align*}
In a local orthonormal frame $e_1,\ldots,e_n$, write the pulled-back connection matrix as $\alpha=(\alpha_{ij})$, so that
\begin{align*}
\nabla_Xe_j=\sum_{i=1}^n\alpha_{ij}(X)e_i.
\end{align*}
Metric compatibility says that $X(g(e_i,e_j))=g(\nabla_Xe_i,e_j)+g(e_i,\nabla_Xe_j)$. Since $g(e_i,e_j)=\delta_{ij}$ is constant in an orthonormal frame, the left side is $0$, and substituting the connection coefficients gives
\begin{align*}
0=g\left(\sum_{k=1}^n\alpha_{ki}(X)e_k,e_j\right)+g\left(e_i,\sum_{k=1}^n\alpha_{kj}(X)e_k\right).
\end{align*}
Using $g(e_k,e_j)=\delta_{kj}$ and $g(e_i,e_k)=\delta_{ik}$, this becomes
\begin{align*}
0=\sum_{k=1}^n\alpha_{ki}(X)\delta_{kj}+\sum_{k=1}^n\alpha_{kj}(X)\delta_{ik}.
\end{align*}
Evaluating the Kronecker deltas gives
\begin{align*}
0=\alpha_{ji}(X)+\alpha_{ij}(X).
\end{align*}
Therefore $\alpha(X)^\top+\alpha(X)=0$ for every vector field $X$, so the local connection matrix is skew-symmetric exactly when the connection form restricted to $O_g(M)$ is $\mathfrak{o}(n)$-valued. This is the frame-bundle expression of preserving the metric under parallel transport.
[/example]
A reduction can also encode orientation. Combining metric and orientation leads from $O(n)$ to $SO(n)$, and the same criterion explains when a connection descends further.
[example: Oriented Orthonormal Frames]
Let $(M,g)$ be an oriented Riemannian $n$-manifold and let $SO_g(M)\subset O_g(M)$ be the oriented orthonormal frame bundle. Along a curve, write parallel transport in an oriented orthonormal frame as a matrix $A(t)$ with $A(0)=I$. For a metric connection, the infinitesimal matrix
\begin{align*}
B(t)=A(t)^{-1}\dot A(t)
\end{align*}
lies in $\mathfrak{so}(n)$, so $B(t)^\top+B(t)=0$. Looking at the $i$th diagonal entry gives
\begin{align*}
B_{ii}(t)+B_{ii}(t)=0.
\end{align*}
Thus $B_{ii}(t)=0$ for every $i$, and therefore
\begin{align*}
\operatorname{tr}(B(t))=\sum_{i=1}^n B_{ii}(t)=0.
\end{align*}
By *Jacobi's determinant identity*,
\begin{align*}
\frac{d}{dt}\det A(t)=\det A(t)\operatorname{tr}\bigl(A(t)^{-1}\dot A(t)\bigr).
\end{align*}
Substituting $B(t)=A(t)^{-1}\dot A(t)$ gives
\begin{align*}
\frac{d}{dt}\det A(t)=\det A(t)\operatorname{tr}(B(t)).
\end{align*}
Since $\operatorname{tr}(B(t))=0$, this becomes
\begin{align*}
\frac{d}{dt}\det A(t)=0.
\end{align*}
Hence $\det A(t)$ is constant along the curve. Because $A(0)=I$ and $\det I=1$, we get
\begin{align*}
\det A(t)=1
\end{align*}
for all $t$.
Thus metric parallel transport that starts in an oriented orthonormal frame stays in an oriented orthonormal frame, so the connection restricts to the $SO(n)$-reduction $SO_g(M)$. Without metric compatibility, the infinitesimal matrix need not be skew-symmetric and may have nonzero trace; the same determinant equation then measures the failure of parallel transport to preserve oriented orthonormal frames.
[/example]
## Splitting the Connection Form Along a Reduction
The preservation criterion is exact but sometimes too rigid: a connection may fail to preserve an $H$-reduction in a measured way. The obstruction is not just that $\omega|_{TQ}$ has values outside $\mathfrak h$; one also needs that outside part to transform tensorially under the reduced structure group. A random vector-space complement to $\mathfrak h$ may be mixed with $\mathfrak h$ by the $H$-action, so its projected component would depend on the chosen reduced frame in the wrong way. To isolate the failure in a geometrically meaningful form, choose an $\operatorname{Ad}_H$-compatible splitting of Lie algebras and decompose the connection form into its $\mathfrak h$-part and complementary part.
[definition: Reductive Splitting]
Let $H \le G$ be a closed Lie subgroup. A reductive splitting of $\mathfrak g$ relative to $H$ is a vector-space decomposition
\begin{align*}
\mathfrak g = \mathfrak h \oplus \mathfrak m
\end{align*}
such that $\operatorname{Ad}_h(\mathfrak m) \subset \mathfrak m$ for every $h \in H$.
[/definition]
The $\operatorname{Ad}_H$-invariance condition makes the two projected pieces transform correctly under the reduced structure group. Without this compatibility, projecting a connection form to $\mathfrak h$ or $\mathfrak m$ would depend on the chosen reduced frame and would not define geometric data on the reduction. The question is therefore whether an invariant splitting turns the restriction of $\omega$ to $Q$ into two well-defined objects: a genuine connection on the $H$-bundle and a tensorial complementary obstruction.
[quotetheorem:6286]
[citeproof:6286]
The theorem separates an induced reduced connection from a tensorial obstruction. The $\operatorname{Ad}_H$-invariance of the splitting is essential: it ensures that the projections to $\mathfrak h$ and $\mathfrak m$ commute with the right $H$-action, so $\omega_{\mathfrak h}$ is again a principal connection form and $\omega_{\mathfrak m}$ is a tensorial form. A concrete failure occurs when a complement is not invariant. Take $H$ to be the diagonal subgroup of $GL(2,\mathbb R)$ with matrices $\operatorname{diag}(a,1)$, and let $\mathfrak m$ be a non-invariant line complement to $\mathfrak h$, such as the span of $E_{12}+E_{11}$ together with enough additional directions to complete a vector-space complement. Conjugation by elements of $H$ changes the $E_{12}$ component by a different weight from the $E_{11}$ component, so projection to this chosen complement does not commute with the right $H$-action. The projected $\mathfrak m$-part then depends on the reduced frame rather than defining an associated-bundle tensor.
The theorem also does not make the splitting canonical. When a preferred complement exists, as for orthogonal decompositions in matrix groups or reductive homogeneous spaces, the obstruction has a natural interpretation. When no preferred complement is specified, different splittings can give different $\mathfrak h$-connections and different obstruction forms. In many geometric examples, $\omega_{\mathfrak m}$ is therefore not merely an error term but a named geometric object, such as a second fundamental form or intrinsic torsion, once the relevant class of compatible connections has been fixed.
[example: Cartan Style Decomposition for an SO n Reduction]
Let $P$ be a principal $GL(n,\mathbb R)$-bundle and let $Q \subset P$ be an $SO(n)$-reduction. For any matrix $B\in\mathfrak{gl}(n,\mathbb R)$, define
\begin{align*}
B_{\mathfrak{so}}=\frac{1}{2}(B-B^\top),
\qquad
B_{\operatorname{sym}}=\frac{1}{2}(B+B^\top).
\end{align*}
Then
\begin{align*}
B_{\mathfrak{so}}^\top
=\frac{1}{2}(B^\top-B)
=-B_{\mathfrak{so}},
\qquad
B_{\operatorname{sym}}^\top
=\frac{1}{2}(B^\top+B)
=B_{\operatorname{sym}},
\end{align*}
so $B_{\mathfrak{so}}\in\mathfrak{so}(n)$ and $B_{\operatorname{sym}}\in\operatorname{Sym}(n)$. Also
\begin{align*}
B_{\mathfrak{so}}+B_{\operatorname{sym}}
=\frac{1}{2}(B-B^\top)+\frac{1}{2}(B+B^\top)
=B.
\end{align*}
If a matrix is both skew-symmetric and symmetric, then $C^\top=-C$ and $C^\top=C$, hence $C=-C$, so $2C=0$ and $C=0$. Therefore
\begin{align*}
\mathfrak{gl}(n,\mathbb R)=\mathfrak{so}(n)\oplus \operatorname{Sym}(n).
\end{align*}
This splitting is compatible with the $SO(n)$-action. If $h\in SO(n)$, then $h^{-1}=h^\top$. For $S\in\operatorname{Sym}(n)$,
\begin{align*}
(hSh^{-1})^\top
=(hSh^\top)^\top
=hS^\toph^\top
=hSh^\top
=hSh^{-1},
\end{align*}
so $\operatorname{Ad}_hS=hSh^{-1}$ is again symmetric. Thus the complement $\operatorname{Sym}(n)$ is $\operatorname{Ad}_{SO(n)}$-invariant.
Let $\omega\in\Omega^1(P;\mathfrak{gl}(n,\mathbb R))$ be a $GL(n,\mathbb R)$-connection form and write $\beta=\omega|_{TQ}$. For $v\in T_qQ$, decompose
\begin{align*}
\beta(v)=\beta_{\mathfrak{so}}(v)+\beta_{\operatorname{sym}}(v),
\end{align*}
where
\begin{align*}
\beta_{\mathfrak{so}}(v)=\frac{1}{2}\bigl(\beta(v)-\beta(v)^\top\bigr),
\qquad
\beta_{\operatorname{sym}}(v)=\frac{1}{2}\bigl(\beta(v)+\beta(v)^\top\bigr).
\end{align*}
By the *Decomposition of a Connection Along a Reductive Reduction*, $\beta_{\mathfrak{so}}$ is the induced principal $SO(n)$-connection form on $Q$, while $\beta_{\operatorname{sym}}$ is horizontal and $SO(n)$-equivariant.
The original connection preserves the $SO(n)$-reduction exactly when $\omega|_{TQ}$ is $\mathfrak{so}(n)$-valued, by the *Preservation Criterion for Reductions*. In the displayed decomposition, this is equivalent to
\begin{align*}
\beta_{\operatorname{sym}}(v)=0
\end{align*}
for every $v\in TQ$. Thus the symmetric part is precisely the obstruction to preserving the orthonormal reduction: it vanishes exactly when parallel transport stays inside $Q$.
[/example]
The same mechanism applies after passing to associated bundles. The $\mathfrak h$-part differentiates sections in a way compatible with the reduced structure group, while the $\mathfrak m$-part records how the unreduced connection moves the reduction inside the larger bundle.
[remark: Intrinsic Torsion Viewpoint]
For a general $G$-structure $Q \subset F(M)$, the tensorial component of a connection form relative to a splitting often represents the intrinsic torsion of the $G$-structure after quotienting by the possible changes of compatible connection. This chapter uses only the connection-level decomposition, but the same bookkeeping is the starting point for special holonomy and Cartan-geometric constructions.
[/remark]
## The Solder Form and Torsion on the Frame Bundle
The frame bundle has extra structure not present on an arbitrary principal bundle: each frame identifies the tangent space of $M$ with $\mathbb R^n$. This tautological identification is encoded by the solder form, and it allows a connection on the frame bundle to have torsion.
[definition: Solder Form]
Let $F(M)$ be the frame bundle of a smooth $n$-manifold $M$. At a frame $u:\mathbb R^n \to T_xM$, the solder form is the $\mathbb R^n$-valued $1$-form $\theta \in \Omega^1(F(M);\mathbb R^n)$ defined by
\begin{align*}
\theta_u(v)=u^{-1}(d\pi_u(v))
\end{align*}
for $v \in T_uF(M)$.
[/definition]
The solder form vanishes on vertical vectors, because vertical motion changes the frame over a fixed base point. Since it records base velocity in frame coordinates, combining it with a connection should measure how transported frames fail to bracket vector fields correctly; this motivates the torsion form.
[definition: Torsion Form]
Let $\omega \in \Omega^1(F(M);\mathfrak{gl}(n,\mathbb R))$ be a connection form and let $\theta \in \Omega^1(F(M);\mathbb R^n)$ be the solder form. The torsion form of $\omega$ is the $\mathbb R^n$-valued $2$-form
\begin{align*}
T = d\theta + \omega \wedge \theta,
\end{align*}
where $(\omega \wedge \theta)(X,Y)=\omega(X)\theta(Y)-\omega(Y)\theta(X)$.
[/definition]
This is the first Cartan structure equation. Since $T$ is defined on the total space $F(M)$, it is not automatically a tensor on $M$: vertical directions could contribute, and changing the frame could change the components by the wrong rule. The remaining question is whether the Cartan expression has exactly the horizontality and equivariance needed to descend. The theorem below answers this descent problem and identifies the descended form with the usual torsion tensor.
[quotetheorem:6287]
[citeproof:6287]
The structure equation explains why torsion is natural on frame bundles but not on arbitrary principal bundles. The connection form alone differentiates frames vertically, but it does not identify a tangent vector on the base with a vector in a fixed model space. The solder form supplies precisely that identification; without it there is no canonical expression like $d\theta+\omega\wedge\theta$, and hence no intrinsic torsion tensor attached to an arbitrary principal bundle. For example, a principal $U(1)$-bundle over $M$ has connection and curvature forms, but no canonical map from $TM$ into the associated complex line that would play the role of the solder form.
Horizontality and equivariance are the descent conditions, and each condition is needed in a different way. Without horizontality, a vertical vector can change the value of the form even though it projects to zero in $T_xM$; the first example below shows this obstruction on $F(\mathbb R^2)$. Without the equivariance law $(R_g)^*T=g^{-1}T$, changing frame would transform the components by the wrong rule, so the descended object would not be a tangent-bundle-valued $2$-form. Even when both descent conditions hold for some tensorial form, the solder form remains essential for torsion: it is the tautological map that converts base velocities into frame coordinates, and the expression $d\theta+\omega\wedge\theta$ has no canonical analogue on a general principal bundle. For instance, a principal $U(1)$-bundle may have connection and curvature forms, but no preferred identification between $TM$ and an associated copy of $\mathbb C$. The theorem therefore identifies torsion only for linear connections on $TM$ equipped with the tautological solder form. This limitation points directly to the Levi-Civita problem, where the same torsion equation is imposed after restricting from all frames to orthonormal frames.
[example: A Non-Horizontal Form Does Not Descend]
On $F(\mathbb R^2)$ with its standard flat connection, let $\theta=(\theta_1,\theta_2)$ be the solder form and let $\omega=(\omega_{ij})$ be the connection form. Choose the standard coordinate frame $u:\mathbb R^2\to T_x\mathbb R^2$ with $u(e_i)=\partial_{x_i}|_x$. Let $V$ be the fundamental vertical vector at $u$ generated by $E_{11}$, and let $W$ be the horizontal lift at $u$ of $\partial_{x_1}|_x$.
By reproduction of the connection form on fundamental vertical vectors,
\begin{align*}
\omega(V)=E_{11}.
\end{align*}
Hence
\begin{align*}
\omega_{11}(V)=1.
\end{align*}
Since $W$ is horizontal,
\begin{align*}
\omega(W)=0.
\end{align*}
Thus
\begin{align*}
\omega_{11}(W)=0.
\end{align*}
The solder form vanishes on vertical vectors because $d\pi_u(V)=0$, so
\begin{align*}
\theta_1(V)=0.
\end{align*}
For $W$, the definition of the solder form gives
\begin{align*}
\theta_u(W)=u^{-1}(d\pi_u(W)).
\end{align*}
Since $W$ is a lift of $\partial_{x_1}|_x$,
\begin{align*}
d\pi_u(W)=\partial_{x_1}|_x.
\end{align*}
Since $u(e_1)=\partial_{x_1}|_x$, this gives
\begin{align*}
\theta_u(W)=u^{-1}(u(e_1))=e_1.
\end{align*}
Therefore
\begin{align*}
\theta_1(W)=1.
\end{align*}
Now set $\Phi=\omega_{11}\wedge\theta_1$. By the definition of the wedge product of two $1$-forms,
\begin{align*}
\Phi(V,W)=\omega_{11}(V)\theta_1(W)-\omega_{11}(W)\theta_1(V).
\end{align*}
Substituting the four values computed above gives
\begin{align*}
\Phi(V,W)=(1)(1)-(0)(0)=1.
\end{align*}
If $\Phi$ descended from a $2$-form $\varphi$ on $\mathbb R^2$, then $\Phi=\pi^*\varphi$. Evaluating on $(V,W)$ would give
\begin{align*}
\Phi(V,W)=(\pi^*\varphi)(V,W).
\end{align*}
By the definition of pullback,
\begin{align*}
(\pi^*\varphi)(V,W)=\varphi(d\pi_u(V),d\pi_u(W)).
\end{align*}
Here $d\pi_u(V)=0$ and $d\pi_u(W)=\partial_{x_1}|_x$, so
\begin{align*}
(\pi^*\varphi)(V,W)=\varphi(0,\partial_{x_1}|_x)=0,
\end{align*}
because a $2$-form is multilinear and vanishes when one argument is zero. This contradicts $\Phi(V,W)=1$, so $\omega_{11}\wedge\theta_1$ is not the pullback of any $2$-form on $\mathbb R^2$; its vertical dependence prevents it from defining a tensor on the base.
[/example]
The previous example isolates the descent failure caused by vertical dependence. Once the form is known to descend, the remaining question is how the descended tensor appears in ordinary coordinate notation; this is where the structure equation recovers the familiar antisymmetry of Christoffel symbols.
[example: Coordinate Expression for Torsion]
Let $(U,x)$ be a coordinate chart and let $e_i=\partial_{x_i}$ be the coordinate frame over $U$. Suppose the induced connection satisfies
\begin{align*}
\nabla_{e_i}e_j=\sum_{k=1}^n \Gamma_{ij,k} e_k.
\end{align*}
Using the torsion formula with $X=e_i$ and $Y=e_j$ gives
\begin{align*}
\operatorname{Tor}_{\nabla}(e_i,e_j)=\nabla_{e_i}e_j-\nabla_{e_j}e_i-[e_i,e_j].
\end{align*}
The coordinate vector fields commute. Indeed, for every smooth function $f$,
\begin{align*}
[e_i,e_j](f)=e_i(e_jf)-e_j(e_if).
\end{align*}
Since $e_i=\partial_{x_i}$ and $e_j=\partial_{x_j}$, this becomes
\begin{align*}
[e_i,e_j](f)=\frac{\partial}{\partial x_i}\left(\frac{\partial f}{\partial x_j}\right)-\frac{\partial}{\partial x_j}\left(\frac{\partial f}{\partial x_i}\right).
\end{align*}
Equivalently,
\begin{align*}
[e_i,e_j](f)=\frac{\partial^2 f}{\partial x_i\partial x_j}-\frac{\partial^2 f}{\partial x_j\partial x_i}.
\end{align*}
Equality of mixed partial derivatives gives
\begin{align*}
[e_i,e_j](f)=0.
\end{align*}
Thus $[e_i,e_j]=0$.
Substituting the connection coefficients into the torsion formula gives
\begin{align*}
\operatorname{Tor}_{\nabla}(e_i,e_j)=\sum_{k=1}^n \Gamma_{ij,k}e_k-\sum_{k=1}^n \Gamma_{ji,k}e_k-0.
\end{align*}
Combining the two sums over the same coordinate frame,
\begin{align*}
\operatorname{Tor}_{\nabla}(e_i,e_j)=\sum_{k=1}^n \left(\Gamma_{ij,k}-\Gamma_{ji,k}\right)e_k.
\end{align*}
Therefore the torsion tensor vanishes on all coordinate pairs $e_i,e_j$ exactly when $\Gamma_{ij,k}=\Gamma_{ji,k}$ for all $i,j,k$, so torsion-free means symmetry of the Christoffel symbols in the two lower coordinate indices.
[/example]
For orthonormal frames the same equation is written with an $\mathfrak{so}(n)$-valued connection form. The metric compatibility condition is already built into the reduction, so the torsion equation isolates the remaining freedom.
## Levi-Civita Connection in Frame-Bundle Language
A Riemannian metric gives an $O(n)$-reduction of the frame bundle. The [fundamental theorem of Riemannian geometry](/theorems/1552) asks whether there is a connection on this reduced bundle that has no torsion, and whether this requirement determines it.
[quotetheorem:1552]
[citeproof:1552]
This theorem says that the two natural requirements on a Riemannian connection, preservation of the metric reduction and vanishing torsion, determine the connection. Both requirements are necessary. If metric compatibility is dropped, there are many torsion-free linear connections whose parallel transport does not preserve lengths or angles. On $\mathbb R$ with metric $dx^2$, the connection determined by $\nabla_{\partial_x}\partial_x=f\,\partial_x$ is torsion-free for every smooth function $f$, but it is metric only when $f=0$. If torsion-freeness is dropped, there are many metric connections on $O_g(M)$: starting from the Levi-Civita connection, adding a tensorial $\mathfrak{o}(n)$-valued $1$-form preserves the metric reduction but usually introduces torsion.
In the frame-bundle picture, metric compatibility is reduction to $O_g(M)$ and torsion-freeness is the first structure equation with $T=0$. The same argument has pseudo-Riemannian analogues with $O(p,q)$ replacing $O(n)$, but the theorem as stated is the Riemannian positive-definite case. It also does not say that every torsion-free connection is Levi-Civita for a given metric; the connection must preserve the metric reduction as well. Nor does it choose a connection for an arbitrary $G$-structure: special holonomy and $G$-structure problems often begin by asking whether analogous compatibility and torsion conditions are even solvable. The oriented restriction below is the simplest next step, because it changes the structure group without changing the Lie algebra or the torsion equation.
[example: Levi-Civita Connection as the Unique Torsion-Free Metric Connection]
[claim]The Koszul formula defines the unique covariant derivative that is both metric-compatible and torsion-free.[/claim]
[proof]Define $\nabla$ by requiring, for all local vector fields $X,Y,Z$,
\begin{align*}
2g(\nabla_XY,Z)=Xg(Y,Z)+Yg(Z,X)-Zg(X,Y)+g([X,Y],Z)-g([Y,Z],X)+g([Z,X],Y).
\end{align*}
To check metric compatibility, apply the same formula with $Y$ and $Z$ interchanged:
\begin{align*}
2g(\nabla_XZ,Y)=Xg(Z,Y)+Zg(Y,X)-Yg(X,Z)+g([X,Z],Y)-g([Z,Y],X)+g([Y,X],Z).
\end{align*}
Add the two identities. Since $g(Y,Z)=g(Z,Y)$, $g(Z,X)=g(X,Z)$, $g(Y,X)=g(X,Y)$, and the Lie bracket is skew-symmetric, the terms $Yg(Z,X)$ and $-Yg(X,Z)$ cancel, as do $-Zg(X,Y)$ and $Zg(Y,X)$. The bracket terms cancel in pairs:
\begin{align*}
g([X,Y],Z)+g([Y,X],Z)=g([X,Y],Z)-g([X,Y],Z)=0.
\end{align*}
\begin{align*}
-g([Y,Z],X)-g([Z,Y],X)=-g([Y,Z],X)+g([Y,Z],X)=0.
\end{align*}
\begin{align*}
g([Z,X],Y)+g([X,Z],Y)=g([Z,X],Y)-g([Z,X],Y)=0.
\end{align*}
Therefore
\begin{align*}
2g(\nabla_XY,Z)+2g(\nabla_XZ,Y)=Xg(Y,Z)+Xg(Z,Y)=2Xg(Y,Z).
\end{align*}
Dividing by $2$ and using symmetry of $g$ gives
\begin{align*}
Xg(Y,Z)=g(\nabla_XY,Z)+g(Y,\nabla_XZ),
\end{align*}
so $\nabla g=0$.
For torsion, write the Koszul formula with $X$ and $Y$ exchanged:
\begin{align*}
2g(\nabla_YX,Z)=Yg(X,Z)+Xg(Z,Y)-Zg(Y,X)+g([Y,X],Z)-g([X,Z],Y)+g([Z,Y],X).
\end{align*}
Subtract this from the defining formula. The metric-derivative terms cancel as before, and the bracket terms become
\begin{align*}
g([X,Y],Z)-g([Y,X],Z)=g([X,Y],Z)+g([X,Y],Z)=2g([X,Y],Z).
\end{align*}
\begin{align*}
-g([Y,Z],X)-g([Z,Y],X)=-g([Y,Z],X)+g([Y,Z],X)=0.
\end{align*}
\begin{align*}
g([Z,X],Y)+g([X,Z],Y)=g([Z,X],Y)-g([Z,X],Y)=0.
\end{align*}
Hence
\begin{align*}
2g(\nabla_XY,Z)-2g(\nabla_YX,Z)=2g([X,Y],Z).
\end{align*}
Equivalently,
\begin{align*}
g(\nabla_XY-\nabla_YX-[X,Y],Z)=0
\end{align*}
for every $Z$. Nondegeneracy of $g$ gives
\begin{align*}
\nabla_XY-\nabla_YX-[X,Y]=0,
\end{align*}
so the torsion vanishes.
Finally, let $D$ be any metric-compatible torsion-free connection. Metric compatibility gives
\begin{align*}
Xg(Y,Z)=g(D_XY,Z)+g(Y,D_XZ).
\end{align*}
\begin{align*}
Yg(Z,X)=g(D_YZ,X)+g(Z,D_YX).
\end{align*}
\begin{align*}
Zg(X,Y)=g(D_ZX,Y)+g(X,D_ZY).
\end{align*}
Torsion-freeness gives
\begin{align*}
[X,Y]=D_XY-D_YX.
\end{align*}
\begin{align*}
[Y,Z]=D_YZ-D_ZY.
\end{align*}
\begin{align*}
[Z,X]=D_ZX-D_XZ.
\end{align*}
Substitute these six identities into the right-hand side of the Koszul formula. The metric-compatibility part contributes
\begin{align*}
g(D_XY,Z)+g(Y,D_XZ)+g(D_YZ,X)+g(Z,D_YX)-g(D_ZX,Y)-g(X,D_ZY).
\end{align*}
The bracket part contributes
\begin{align*}
g(D_XY-D_YX,Z)-g(D_YZ-D_ZY,X)+g(D_ZX-D_XZ,Y).
\end{align*}
Using symmetry of $g$, the terms $g(Y,D_XZ)$ and $-g(D_XZ,Y)$ cancel, the terms $g(D_YZ,X)$ and $-g(D_YZ,X)$ cancel, the terms $g(Z,D_YX)$ and $-g(D_YX,Z)$ cancel, the terms $-g(D_ZX,Y)$ and $g(D_ZX,Y)$ cancel, and the terms $-g(X,D_ZY)$ and $g(D_ZY,X)$ cancel. The two remaining terms are
\begin{align*}
g(D_XY,Z)+g(D_XY,Z)=2g(D_XY,Z).
\end{align*}
Thus the Koszul formula gives
\begin{align*}
2g(\nabla_XY,Z)=2g(D_XY,Z)
\end{align*}
for every $Z$. Dividing by $2$ and using nondegeneracy of $g$, we get $\nabla_XY=D_XY$. Therefore the Koszul formula determines the unique torsion-free metric connection.[/proof]
The corresponding principal connection form on $O_g(M)$ is exactly the Levi-Civita connection form: metric compatibility means it is $\mathfrak{o}(n)$-valued, and the torsion-free calculation is the vanishing of $d\theta+\omega\wedge\theta$.
[/example]
The oriented case is obtained by restricting the same connection to the $SO(n)$-reduction when the manifold is oriented. No new torsion equation is needed, since the Lie algebra of $SO(n)$ is still $\mathfrak{so}(n)$.
[example: Induced Connection on the Oriented Orthonormal Frame Bundle]
Let $(M,g)$ be oriented, let $SO_g(M)\subset O_g(M)$ be the oriented orthonormal frame bundle, and let $\omega\in\Omega^1(O_g(M);\mathfrak{o}(n))$ be the Levi-Civita connection form. We show that $\omega$ restricts to a principal $SO(n)$-connection on $SO_g(M)$ and that the torsion equation remains zero after restriction.
In an oriented orthonormal frame, write parallel transport along a curve as a matrix $A(t)\in O(n)$ with $A(0)=I$. Its infinitesimal matrix is
\begin{align*}
B(t)=A(t)^{-1}\dot A(t).
\end{align*}
Because the Levi-Civita connection form is $\mathfrak{o}(n)$-valued, $B(t)$ is skew-symmetric:
\begin{align*}
B(t)^\top+B(t)=0.
\end{align*}
Taking the $i$th diagonal entry gives
\begin{align*}
B_{ii}(t)+B_{ii}(t)=0.
\end{align*}
Thus
\begin{align*}
B_{ii}(t)=0.
\end{align*}
Summing over $i$ gives
\begin{align*}
\operatorname{tr}B(t)=\sum_{i=1}^n B_{ii}(t)=0.
\end{align*}
By *Jacobi's determinant identity*,
\begin{align*}
\frac{d}{dt}\det A(t)=\det A(t)\operatorname{tr}\bigl(A(t)^{-1}\dot A(t)\bigr).
\end{align*}
Substituting $B(t)=A(t)^{-1}\dot A(t)$ gives
\begin{align*}
\frac{d}{dt}\det A(t)=\det A(t)\operatorname{tr}B(t).
\end{align*}
Since $\operatorname{tr}B(t)=0$, this becomes
\begin{align*}
\frac{d}{dt}\det A(t)=0.
\end{align*}
Hence $\det A(t)$ is constant. Since $A(0)=I$ and $\det I=1$,
\begin{align*}
\det A(t)=1
\end{align*}
for all $t$. Therefore parallel transport starting from an oriented orthonormal frame remains in $SO_g(M)$.
The restricted form
\begin{align*}
\omega_{SO}=\omega|_{TSO_g(M)}
\end{align*}
is still $\mathfrak{so}(n)$-valued. On vertical fundamental vector fields for the $SO(n)$-action it reproduces the generating element, because the same reproduction axiom holds for $\omega$ on $O_g(M)$. Its equivariance under the right $SO(n)$-action is the restriction of the $O(n)$-equivariance of $\omega$. Hence $\omega_{SO}$ is a principal $SO(n)$-connection form.
Let $i:SO_g(M)\hookrightarrow O_g(M)$ be the inclusion. The solder form and connection form on $SO_g(M)$ are
\begin{align*}
\theta_{SO}=i^*\theta
\end{align*}
and
\begin{align*}
\omega_{SO}=i^*\omega.
\end{align*}
The restricted torsion form is
\begin{align*}
T_{SO}=d\theta_{SO}+\omega_{SO}\wedge\theta_{SO}.
\end{align*}
Substituting the two pullback identities gives
\begin{align*}
T_{SO}=d(i^*\theta)+(i^*\omega)\wedge(i^*\theta).
\end{align*}
Exterior derivative commutes with pullback, so
\begin{align*}
d(i^*\theta)=i^*(d\theta).
\end{align*}
Pullback also commutes with wedge products of vector-valued forms in this expression, so
\begin{align*}
(i^*\omega)\wedge(i^*\theta)=i^*(\omega\wedge\theta).
\end{align*}
Therefore
\begin{align*}
T_{SO}=i^*(d\theta)+i^*(\omega\wedge\theta).
\end{align*}
By linearity of pullback,
\begin{align*}
T_{SO}=i^*(d\theta+\omega\wedge\theta).
\end{align*}
The Levi-Civita torsion form on $O_g(M)$ satisfies
\begin{align*}
d\theta+\omega\wedge\theta=0.
\end{align*}
Hence
\begin{align*}
T_{SO}=i^*0=0.
\end{align*}
Thus passing from $O_g(M)$ to the oriented reduction $SO_g(M)$ changes the structure group but not the local connection forms or the torsion equation.
[/example]
The chapter closes the circle between reductions and induced geometry. A connection preserving a reduction becomes a connection on the reduced bundle; a connection failing to preserve it decomposes into a compatible part and a tensorial obstruction; and on the frame bundle, the solder form turns the compatible connection into geometric data on $TM$. For Riemannian geometry, these principles recover the Levi-Civita connection as the unique connection that preserves the orthonormal reduction and has zero torsion. The same preservation-and-obstruction pattern also appears in gauge theory: reducing an $SU(2)$-bundle to a $U(1)$-subbundle separates a connection into an electromagnetic-looking $U(1)$ component and charged off-diagonal components, and the latter record how parallel transport leaves the reduced constraint.
Connections that respect a reduction reveal how the same geometric object can split into constrained and unconstrained parts. The final chapter uses this viewpoint in concrete computations, combining transition functions, connection forms, curvature, and reductions to show how topology and local formulas interact in examples.
# 12. Integrative Examples and Computations
This final chapter gathers the main computational techniques from the course and applies them in settings where the topology of the bundle affects the formulas seen in a chart. The prerequisites are principal bundles and local sections from Chapters 1 and 2, connection forms from Chapter 6, curvature from Chapter 8, gauge transformations from Chapter 9, and holonomy from Chapter 10. The goal is to see how the abstract definitions behave in explicit examples: the Hopf fibration, surface frame bundles, matrix-valued gauge fields, and flat local normal forms.
## The Hopf Connection and Monopole Curvature
The first problem is to compute a connection on a non-product principal bundle without hiding the topology. The Hopf fibration $S^3 \to S^2$ is the model case: it is simple enough for explicit coordinates, but its curvature already records the first Chern class of the associated complex line bundle.
We regard $S^3 \subset \mathbb C^2$ as the unit sphere with coordinates $(z_1,z_2)$ and let $U(1)$ act on the right by simultaneous scalar multiplication,
\begin{align*}
(z_1,z_2)\cdot e^{i\theta}=(z_1e^{i\theta},z_2e^{i\theta}).
\end{align*}
The quotient is $S^2$, identified with $\mathbb{C}P^1$ or with the unit sphere in $\mathbb R^3$ by the Hopf map $h:S^3\to S^2\subset \mathbb R^3$ defined by
\begin{align*}
h(z_1,z_2)=(2\operatorname{Re}(z_1\bar z_2),2\operatorname{Im}(z_1\bar z_2),|z_1|^2-|z_2|^2).
\end{align*}
[illustration:hopf-chart-transition-sections]
[definition: Hopf Connection Form]
On the principal $U(1)$-bundle $S^3 \to S^2$, the Hopf connection form is the form $\omega\in \Omega^1(S^3;i\mathbb R)$ given by
\begin{align*}
\omega=\bar z_1\,dz_1+\bar z_2\,dz_2
\end{align*}
restricted to $TS^3 \subset T\mathbb C^2$.
[/definition]
Equivalently, for each $p\in S^3$, the form is a fibrewise linear map $\omega_p:T_pS^3\to i\mathbb R$. The formula is compact, but it encodes the two defining features of a principal connection: it evaluates to the generator on vertical vectors and it is equivariant under the right $U(1)$-action. The next task is to see what this connection measures on the base, so we compute its local potential and curvature in a concrete chart.
[quotetheorem:6288]
[citeproof:6288]
The chart and local-section hypotheses are part of the computation. The formula for $A$ is not a global $1$-form on $S^2$; it is the pullback of the global connection form along the particular section $s(w)=(w,1)/(1+|w|^2)^{1/2}$, which exists only where $z_2\neq 0$. If the chart hypothesis is dropped and the same expression is treated as a form on all of $S^2$, the coordinate $w=z_1/z_2$ breaks down at the point $[1:0]$, so the alleged potential has no value there. More strongly, if a single global potential $A_{\mathrm{glob}}$ existed on $S^2$ with $dA_{\mathrm{glob}}=F_A$, then Stokes' theorem would force $\int_{S^2}F_A=0$, while the Hopf curvature has nonzero integral after the standard normalisation. The theorem therefore computes curvature through one gauge, but it does not produce a global potential on the base.
This calculation is the geometric source of the magnetic monopole form on $S^2$. The curvature is not the differential of a single globally defined $1$-form on $S^2$, but it is assembled from compatible local potentials on the two standard charts. The next example makes that assembly visible by comparing the two standard gauges rather than hiding the transition function.
[example: Monopole Potentials on Two Charts]
On the chart $z_2\neq 0$, write $w=z_1/z_2$ and set
\begin{align*}
A_N=\frac{1}{2}\frac{\bar w\,dw-w\,d\bar w}{1+|w|^2}.
\end{align*}
Here the subscript $N$ follows the course convention that this chart is the northern affine gauge, even though $w=z_1/z_2$ is undefined at $[1:0]$ under some Hopf-sphere naming conventions. Put $q=1+|w|^2=1+w\bar w$. Since $dq=\bar w\,dw+w\,d\bar w$, we have
\begin{align*}
d(\bar w\,dw-w\,d\bar w)=d\bar w\wedge dw-dw\wedge d\bar w=2\,d\bar w\wedge dw.
\end{align*}
Also,
\begin{align*}
dq\wedge(\bar w\,dw-w\,d\bar w)=(\bar w\,dw+w\,d\bar w)\wedge(\bar w\,dw-w\,d\bar w).
\end{align*}
The terms containing $dw\wedge dw$ and $d\bar w\wedge d\bar w$ vanish, so
\begin{align*}
dq\wedge(\bar w\,dw-w\,d\bar w)=-|w|^2\,dw\wedge d\bar w+|w|^2\,d\bar w\wedge dw=2|w|^2\,d\bar w\wedge dw.
\end{align*}
Therefore
\begin{align*}
dA_N=\frac{1}{2}\left(-q^{-2}\,dq\wedge(\bar w\,dw-w\,d\bar w)+q^{-1}\,d(\bar w\,dw-w\,d\bar w)\right).
\end{align*}
Substituting the two displayed computations gives
\begin{align*}
dA_N=\frac{1}{2}\left(-\frac{2|w|^2}{q^2}+\frac{2}{q}\right)d\bar w\wedge dw=\frac{d\bar w\wedge dw}{(1+|w|^2)^2}.
\end{align*}
On the chart $z_1\neq 0$, write $v=z_2/z_1=w^{-1}$ and use the corresponding potential
\begin{align*}
A_S=\frac{1}{2}\frac{\bar v\,dv-v\,d\bar v}{1+|v|^2}.
\end{align*}
On the overlap $w\neq 0$,
\begin{align*}
dv=-w^{-2}\,dw.
\end{align*}
Similarly,
\begin{align*}
d\bar v=-\bar w^{-2}\,d\bar w.
\end{align*}
Thus
\begin{align*}
\bar v\,dv-v\,d\bar v=-\bar w^{-1}w^{-2}\,dw+w^{-1}\bar w^{-2}\,d\bar w.
\end{align*}
Since $1+|v|^2=1+|w|^{-2}=(1+|w|^2)/|w|^2$, this becomes
\begin{align*}
A_S=\frac{1}{2}\left(-\frac{dw}{w}+\frac{d\bar w}{\bar w}\right)\frac{1}{1+|w|^2}.
\end{align*}
The two local sections satisfy
\begin{align*}
s_S(w^{-1})=\frac{(1,w^{-1})}{(1+|w|^{-2})^{1/2}}.
\end{align*}
Because $(1+|w|^{-2})^{1/2}=(1+|w|^2)^{1/2}/|w|$, this gives
\begin{align*}
s_S(w^{-1})=\frac{|w|(1,w^{-1})}{(1+|w|^2)^{1/2}}.
\end{align*}
Multiplying the vector $(w,1)$ by $|w|/w$ gives $|w|(1,w^{-1})$, hence
\begin{align*}
s_S(w^{-1})=s_N(w)\,\frac{|w|}{w}.
\end{align*}
So the transition function under the convention $s_j=s_i g_{ij}$ is $g_{NS}=|w|/w$. Its logarithmic derivative is
\begin{align*}
g_{NS}^{-1}dg_{NS}=d\log\left(\frac{|w|}{w}\right).
\end{align*}
Using $\log |w|=\frac{1}{2}\log(w\bar w)$ on the overlap,
\begin{align*}
g_{NS}^{-1}dg_{NS}=\frac{1}{2}\left(\frac{d\bar w}{\bar w}-\frac{dw}{w}\right).
\end{align*}
Now
\begin{align*}
A_N+g_{NS}^{-1}dg_{NS}=\frac{1}{2}\frac{\bar w\,dw-w\,d\bar w}{1+|w|^2}+\frac{1}{2}\left(\frac{d\bar w}{\bar w}-\frac{dw}{w}\right).
\end{align*}
Putting the terms over the common denominator $1+|w|^2$ gives
\begin{align*}
A_N+g_{NS}^{-1}dg_{NS}=\frac{1}{2}\left(-\frac{dw}{w}+\frac{d\bar w}{\bar w}\right)\frac{1}{1+|w|^2}=A_S.
\end{align*}
Finally, the two curvature computations agree on the overlap. From $dv=-w^{-2}dw$ and $d\bar v=-\bar w^{-2}d\bar w$,
\begin{align*}
d\bar v\wedge dv=\bar w^{-2}w^{-2}\,d\bar w\wedge dw.
\end{align*}
Also,
\begin{align*}
(1+|v|^2)^2=\frac{(1+|w|^2)^2}{|w|^4}.
\end{align*}
Therefore
\begin{align*}
dA_S=\frac{d\bar v\wedge dv}{(1+|v|^2)^2}=\frac{d\bar w\wedge dw}{(1+|w|^2)^2}=dA_N.
\end{align*}
The potentials differ by the transition-function term, but their curvature forms agree on the overlap, so they are two local gauges for the same monopole curvature.
[/example]
The example illustrates a principle used throughout gauge theory: a connection is rarely a single global potential, while curvature is a global horizontal equivariant object. To justify that principle beyond the abelian Hopf case, we need the general overlap formula explaining why the derivative term in the gauge change disappears from curvature.
[quotetheorem:6289]
[citeproof:6289]
The hypotheses ensure that the two local forms really come from the same global principal connection. If arbitrary Lie-algebra-valued forms $A_i$ and $A_j$ are chosen on overlapping charts without the transition rule, a concrete failure occurs already on a product bundle over an annulus: take $A_i=0$ in one gauge and choose $A_j$ to be any nonzero $1$-form not equal to $g^{-1}dg$ for the selected transition function; then the two curvatures need not agree on the overlap and cannot represent one global curvature form. The Lie group hypothesis supplies the adjoint action and the Maurer-Cartan form $g^{-1}dg$; for a mere topological structure group these differential expressions are not available. In the nonabelian case the bracket term in $F_i=dA_i+\frac{1}{2}[A_i\wedge A_i]$ is essential, because differentiating only $dA_j$ leaves terms involving $dg_{ij}$ that do not combine into conjugation. The theorem also does not say that every collection of local forms defines a connection; it gives the compatibility condition forced by a global equivariant connection form.
This theorem explains why curvature forms define a global section of the associated bundle $\Lambda^2T^*M\otimes \operatorname{Ad}(P)$. It is also the computational rule that lets us pass from the Hopf monopole to matrix-valued gauge fields, where conjugation changes the coefficients but invariant measurements of curvature can remain well defined.
## The Levi-Civita Connection on an Oriented Surface
The next problem is to recover a familiar Riemannian invariant, Gaussian curvature, from the language of principal connections. On an oriented Riemannian surface, the oriented orthonormal frame bundle has structure group $SO(2)$, and the Levi-Civita connection becomes an $\mathfrak{so}(2)$-valued principal connection.
Let $(\Sigma,g)$ be an oriented Riemannian surface. The fibre of its oriented orthonormal frame bundle $F_{SO(2)}\Sigma\to \Sigma$ over $p\in \Sigma$ consists of ordered bases $(e_1,e_2)$ of $T_p\Sigma$ satisfying $g_p(e_i,e_j)=\delta_{ij}$ and compatible with the orientation.
[definition: Surface Connection One-Form]
For a local oriented orthonormal frame $(e_1,e_2)$ on $U\subset \Sigma$, the Levi-Civita connection $1$-form is the real-valued $1$-form $\alpha\in \Omega^1(U)$ defined by
\begin{align*}
\alpha_p &: T_p\Sigma \to \mathbb R
\end{align*}
for $p\in U$, with
\begin{align*}
\nabla_X e_1=\alpha(X)e_2,\qquad
\nabla_X e_2=-\alpha(X)e_1
\end{align*}
for every vector field $X$ on $U$.
[/definition]
The skew-symmetry in the definition is the metric compatibility of the Levi-Civita connection written in a moving frame. Since $\mathfrak{so}(2)$ is one-dimensional, the matrix-valued connection form is represented by the single scalar form $\alpha$.
[quotetheorem:6290]
[citeproof:6290]
The assumptions explain why this reduces to one scalar $1$-form. Orthonormality reduces the connection matrix to values in $\mathfrak{so}(2)$; if instead an arbitrary local frame is used, the connection matrix is $\mathfrak{gl}(2,\mathbb R)$-valued and its four entries do not collapse to the single form $\alpha$. Orientation chooses the connected structure group $SO(2)$ rather than $O(2)$; on a non-orientable surface such as the Mobius band with a Riemannian metric, the scalar $SO(2)$ description can be used only on oriented patches and must be compared across orientation-reversing transitions. The sign convention is also part of the statement: replacing $\nabla_Xe_1=\alpha(X)e_2$ by $\nabla_Xe_1=-\alpha(X)e_2$ reverses the displayed sign in the formula relating $d\alpha$ to $K\,\theta_1\wedge\theta_2$. The theorem gives a local representative of curvature, not a preferred global frame or a globally defined connection $1$-form on $\Sigma$.
This theorem turns a classical invariant into a principal-bundle curvature computation. It also explains why changing the local oriented orthonormal frame changes the connection form by a gauge term, while the curvature is globally meaningful. The round sphere example now serves as a calibration: it checks the sign convention and shows how the unavoidable singularity of a local frame is separate from the smoothness of the curvature.
[example: The Tangent Frame Bundle of the Round Sphere]
Use spherical coordinates on $0<\vartheta<\pi$:
\begin{align*}
X(\vartheta,\varphi)=(\sin\vartheta\cos\varphi,\sin\vartheta\sin\varphi,\cos\vartheta).
\end{align*}
For the round metric,
\begin{align*}
g=d\vartheta^2+\sin^2\vartheta\,d\varphi^2.
\end{align*}
Thus $|\partial_\vartheta|=1$, $|\partial_\varphi|=\sin\vartheta$, and $g(\partial_\vartheta,\partial_\varphi)=0$. On the complement of the poles, take the oriented orthonormal frame
\begin{align*}
e_1=\partial_\vartheta,\qquad e_2=\frac{1}{\sin\vartheta}\partial_\varphi.
\end{align*}
The Christoffel symbol needed for $\nabla_{\partial_\varphi}e_1$ is
\begin{align*}
\Gamma^\varphi_{\varphi\vartheta}=\frac{1}{2}g^{\varphi\varphi}\partial_\vartheta g_{\varphi\varphi}.
\end{align*}
Since $g_{\varphi\varphi}=\sin^2\vartheta$ and $g^{\varphi\varphi}=\sin^{-2}\vartheta$,
\begin{align*}
\Gamma^\varphi_{\varphi\vartheta}=\frac{1}{2}\frac{1}{\sin^2\vartheta}(2\sin\vartheta\cos\vartheta)=\cot\vartheta.
\end{align*}
Also $\Gamma^\vartheta_{\vartheta\vartheta}=0$ and $\Gamma^\varphi_{\vartheta\vartheta}=0$, so
\begin{align*}
\nabla_{\partial_\vartheta}e_1=\nabla_{\partial_\vartheta}\partial_\vartheta=0.
\end{align*}
For the other coordinate direction,
\begin{align*}
\nabla_{\partial_\varphi}e_1=\nabla_{\partial_\varphi}\partial_\vartheta=\Gamma^\varphi_{\varphi\vartheta}\partial_\varphi.
\end{align*}
Substituting $\Gamma^\varphi_{\varphi\vartheta}=\cot\vartheta$ and $\partial_\varphi=\sin\vartheta\,e_2$ gives
\begin{align*}
\nabla_{\partial_\varphi}e_1=\cot\vartheta\,\sin\vartheta\,e_2=\cos\vartheta\,e_2.
\end{align*}
Because the surface connection form is defined by $\nabla_Xe_1=\alpha(X)e_2$, these evaluations give $\alpha(\partial_\vartheta)=0$ and $\alpha(\partial_\varphi)=\cos\vartheta$. Hence
\begin{align*}
\alpha=\cos\vartheta\,d\varphi.
\end{align*}
Now compute the curvature:
\begin{align*}
d\alpha=d(\cos\vartheta\,d\varphi).
\end{align*}
Using the exterior product rule and $d(d\varphi)=0$,
\begin{align*}
d\alpha=d(\cos\vartheta)\wedge d\varphi+\cos\vartheta\,d(d\varphi).
\end{align*}
Since $d(\cos\vartheta)=-\sin\vartheta\,d\vartheta$, this becomes
\begin{align*}
d\alpha=-\sin\vartheta\,d\vartheta\wedge d\varphi.
\end{align*}
The dual coframe is $\theta_1=d\vartheta$ and $\theta_2=\sin\vartheta\,d\varphi$, so
\begin{align*}
\theta_1\wedge\theta_2=\sin\vartheta\,d\vartheta\wedge d\varphi.
\end{align*}
Therefore
\begin{align*}
d\alpha=-\theta_1\wedge\theta_2.
\end{align*}
This matches the convention $d\alpha=-K\,\theta_1\wedge\theta_2$ with $K=1$. The chosen frame is singular at the poles because $\partial_\varphi$ vanishes there, but the curvature is the smooth negative of the round area form in this orientation convention.
[/example]
The sphere example is the surface analogue of the Hopf monopole computation from the previous section. Both use two local gauges near unavoidable singularities, and both produce curvature forms whose integrals detect global geometry.
[remark: Local Frames and Gauge Changes]
If a local oriented orthonormal frame is rotated by an angle function $\psi:U\to \mathbb R$, then the connection form changes by $\alpha\mapsto \alpha+d\psi$, while $d\alpha$ is unchanged. This is the $SO(2)$ version of the principal connection transformation law, using the identification $SO(2)\cong U(1)$ at the level of abelian Lie groups.
[/remark]
The remark makes the local-to-global consistency explicit in the simplest non-product setting. It is the same mechanism that lets a Riemannian metric reduce the full frame bundle from $GL(2,\mathbb R)$ to $O(2)$, and orientation reduces it further to $SO(2)$.
## Matrix-Valued Gauge Potentials and the Yang-Mills Curvature Norm
The final problem is to understand which scalar quantities can be formed from curvature without choosing a gauge. The Yang-Mills functional is the basic answer: take the pointwise norm squared of curvature, integrate it over the base, and use an invariant inner product on the Lie algebra to make the result independent of local coordinates on the bundle.
Let $P\to M$ be a principal $G$-bundle over an oriented Riemannian manifold $(M,g)$, and suppose $G$ is compact with an $\operatorname{Ad}$-invariant inner product $(\cdot,\cdot)_{\mathfrak g}$ on its Lie algebra $\mathfrak g$.
[definition: Yang-Mills Functional]
Let $\operatorname{Conn}(P)$ denote the space of smooth connections on $P$. The Yang-Mills functional is the map $\operatorname{YM}:\operatorname{Conn}(P)\to [0,\infty]$ defined by
\begin{align*}
\operatorname{YM}(A)=\frac{1}{2}\int_M |F_A|^2\,d\operatorname{vol}_g
\end{align*}
for $A\in \operatorname{Conn}(P)$, where $F_A\in \Omega^2(M;\operatorname{Ad}(P))$ is the curvature of $A$, and the norm is induced by $g$ on $\Lambda^2T^*M$ and by $(\cdot,\cdot)_{\mathfrak g}$ on $\operatorname{Ad}(P)$.
[/definition]
For the computations in this chapter, the Yang-Mills functional packages curvature into a natural global norm. The remaining issue is to check that the displayed integral is not a coordinate artifact, so we prove that the local curvature norms agree on overlaps. This is exactly where the $\operatorname{Ad}$-invariance assumption enters: if an arbitrary inner product on $\mathfrak g$ changed lengths under conjugation, then two gauges could assign different pointwise norms to the same global curvature.
[quotetheorem:6291]
[citeproof:6291]
Gauge invariance is the reason the Yang-Mills functional is geometrically meaningful. Compactness is a convenient structural hypothesis because compact Lie algebras admit natural $\operatorname{Ad}$-invariant inner products, for instance the negative trace form on matrix subalgebras such as $\mathfrak{su}(2)$. It is stronger than the theorem needs: the proof only uses the existence of the chosen $\operatorname{Ad}$-invariant inner product. If instead we used a non-invariant pairing on $\mathfrak g$, conjugating a curvature coefficient could change its length, so the local functions called $|F_A|^2$ in two gauges would fail to agree on the overlap.
The theorem does not say that the connection form itself is gauge-invariant, nor that the curvature components are unchanged as $\mathfrak g$-valued functions in a chosen trivialization. They usually change by conjugation. It also does not assert that every scalar expression built from $F_A$ is well defined on the base; the scalar expression must use the Riemannian metric on forms and an $\operatorname{Ad}$-invariant pairing on the Lie algebra. The local potential $A$ is a choice of coordinates on the space of connections, while the norm of curvature is an intrinsic measurement of how far the connection is from being flat.
[example: A Product SU(2)-Bundle over Euclidean Four-Space]
Consider the product principal $SU(2)$-bundle $\mathbb R^4\times SU(2)\to \mathbb R^4$. In the product trivialization, write the connection as an $\mathfrak{su}(2)$-valued $1$-form
\begin{align*}
A=\sum_{i=1}^4 A_i(x)\,dx_i,
\end{align*}
where each $A_i:\mathbb R^4\to \mathfrak{su}(2)$ is smooth. Since $d(dx_i)=0$, the exterior derivative is
\begin{align*}
dA=\sum_{i=1}^4 dA_i\wedge dx_i.
\end{align*}
For each coefficient function $A_i$, we have
\begin{align*}
dA_i=\sum_{j=1}^4 \partial_{x_j}A_i\,dx_j.
\end{align*}
Therefore
\begin{align*}
dA=\sum_{i=1}^4\sum_{j=1}^4 \partial_{x_j}A_i\,dx_j\wedge dx_i.
\end{align*}
The terms with $i=j$ vanish because $dx_i\wedge dx_i=0$. For $i<j$, the two terms involving $dx_i\wedge dx_j$ are
\begin{align*}
\partial_{x_i}A_j\,dx_i\wedge dx_j+\partial_{x_j}A_i\,dx_j\wedge dx_i.
\end{align*}
Since $dx_j\wedge dx_i=-dx_i\wedge dx_j$, this becomes
\begin{align*}
\left(\partial_{x_i}A_j-\partial_{x_j}A_i\right)dx_i\wedge dx_j.
\end{align*}
Hence
\begin{align*}
dA=\sum_{i<j}\left(\partial_{x_i}A_j-\partial_{x_j}A_i\right)dx_i\wedge dx_j.
\end{align*}
For the bracket term, the definition of the Lie-bracket wedge product gives
\begin{align*}
[A\wedge A]=\sum_{i,j=1}^4 [A_i,A_j]\,dx_i\wedge dx_j.
\end{align*}
Again the terms with $i=j$ vanish. For $i<j$, the paired ordered terms are
\begin{align*}
[A_i,A_j]\,dx_i\wedge dx_j+[A_j,A_i]\,dx_j\wedge dx_i.
\end{align*}
Using $[A_j,A_i]=-[A_i,A_j]$ and $dx_j\wedge dx_i=-dx_i\wedge dx_j$, this equals
\begin{align*}
2[A_i,A_j]\,dx_i\wedge dx_j.
\end{align*}
Thus
\begin{align*}
\frac{1}{2}[A\wedge A]=\sum_{i<j}[A_i,A_j]\,dx_i\wedge dx_j.
\end{align*}
Combining the two pieces gives the curvature formula
\begin{align*}
F_A=dA+\frac{1}{2}[A\wedge A]=\sum_{i<j}\left(\partial_{x_i}A_j-\partial_{x_j}A_i+[A_i,A_j]\right)dx_i\wedge dx_j.
\end{align*}
A gauge transformation $u:\mathbb R^4\to SU(2)$ changes the local potential to
\begin{align*}
A^u=u^{-1}Au+u^{-1}du.
\end{align*}
The curvature transforms by conjugation:
\begin{align*}
F_{A^u}=u^{-1}F_Au.
\end{align*}
If
\begin{align*}
F_A=\sum_{i<j}F_{ij}\,dx_i\wedge dx_j,
\end{align*}
then
\begin{align*}
F_{A^u}=\sum_{i<j}u^{-1}F_{ij}u\,dx_i\wedge dx_j.
\end{align*}
Now use the negative trace inner product on $\mathfrak{su}(2)$,
\begin{align*}
\langle X,Y\rangle=-\operatorname{tr}(XY).
\end{align*}
For $u\in SU(2)$ and $X,Y\in\mathfrak{su}(2)$,
\begin{align*}
\langle u^{-1}Xu,u^{-1}Yu\rangle=-\operatorname{tr}\left((u^{-1}Xu)(u^{-1}Yu)\right).
\end{align*}
Multiplying inside the trace gives
\begin{align*}
-\operatorname{tr}\left((u^{-1}Xu)(u^{-1}Yu)\right)=-\operatorname{tr}(u^{-1}XYu).
\end{align*}
By cyclic invariance of trace,
\begin{align*}
-\operatorname{tr}(u^{-1}XYu)=-\operatorname{tr}(XYuu^{-1}).
\end{align*}
Since $uu^{-1}=I$, this is
\begin{align*}
-\operatorname{tr}(XYuu^{-1})=-\operatorname{tr}(XY)=\langle X,Y\rangle.
\end{align*}
Thus conjugation preserves every Lie-algebra coefficient pairing. The Euclidean metric on the basis $2$-forms $dx_i\wedge dx_j$ is not changed by a gauge transformation, so the pointwise norm $|F_A|^2$ is unchanged. Consequently the Yang-Mills density
\begin{align*}
\frac{1}{2}|F_A|^2\,dx_1\wedge dx_2\wedge dx_3\wedge dx_4
\end{align*}
has the same value in the original gauge and in the transformed gauge.
[/example]
This example shows why the bracket term is not optional in nonabelian gauge theory. It is exactly what makes the curvature transform by conjugation under the gauge action, and that conjugation law is the input needed for the Yang-Mills norm. The natural next question is what happens at the opposite extreme, when the curvature itself vanishes.
[quotetheorem:6292]
[citeproof:6292]
The flatness criterion closes the course by tying together parallel transport, curvature, gauge transformations, and local descriptions. The locality is essential: solving $du=-Au$ by parallel transport requires path-independence, which is guaranteed on a sufficiently small simply connected neighbourhood but can fail globally through holonomy. Thus flat curvature does not imply that the connection is globally gauge-equivalent to the product connection on an arbitrary base; it only removes the local obstruction. The Yang-Mills functional measures the size of this curvature obstruction in a gauge-invariant way.
## Local-to-Global Consistency Across the Examples
The recurring difficulty in this chapter is that the objects available for computation are local, while the geometry they describe is global. A connection is given locally by forms $A_i$, the overlap rule contains an inhomogeneous term $g_{ij}^{-1}dg_{ij}$, and the curvature removes that inhomogeneous term by the Maurer-Cartan identity.
[explanation: The Common Pattern]
For the Hopf fibration, the potentials on the northern and southern charts differ by the logarithmic derivative of the $U(1)$ transition function. For an oriented surface, rotating a local orthonormal frame adds $d\psi$ to the scalar connection form. For an $SU(2)$ gauge field, changing gauge adds $u^{-1}du$ and conjugates the old potential.
In all three cases, curvature transforms without an added derivative term. This is why the curvature descends to a global object with values in an associated adjoint bundle, and why scalar quantities formed using invariant pairings can be integrated over the base. The examples also show the limits of local formulas: singularities in a gauge may reflect a bad local section rather than a singular geometric connection.
[/explanation]
## Beyond and Connections
This course began with principal bundles as symmetry spaces over manifolds. It ends with the same idea expressed analytically: connections are ways to differentiate in the presence of symmetry, curvature records the failure of horizontal directions to close up, and gauge-invariant curvature norms are the bridge from bundle geometry to geometric analysis. The local formulas throughout the course should now be read as computational representatives of global objects, not as substitutes for them.
The most direct next path is characteristic classes. The Hopf fibration and Dirac monopole computation already contain the first Chern class in local form: the curvature of a unitary connection represents a de Rham cohomology class after the standard normalization, and its integral records the topological twisting of the circle bundle. The same principle extends to Chern-Weil theory, where invariant polynomials on $\mathfrak g$ turn curvature into closed differential forms whose cohomology classes are independent of the chosen connection.
A second path leads back to Riemannian geometry. The orthonormal frame bundle packages the Levi-Civita connection as a principal $O(n)$- or $SO(n)$-connection, while the solder form and torsion explain why the tangent bundle has extra structure not present on an arbitrary associated vector bundle. From this viewpoint, curvature tensors, sectional curvature, and the Bianchi identities are the associated-bundle expressions of the same principal curvature form studied here.
A third path is holonomy. Parallel transport turns a connection into a family of group elements attached to paths, and curvature gives the infinitesimal obstruction to path independence. The Ambrose-Singer viewpoint connects these two levels: local curvature data generates the Lie algebra of the holonomy group, while global topology can still contribute monodromy even when the curvature vanishes.
A fourth path is gauge theory and geometric analysis. The Yang-Mills functional introduced above is the starting point for studying critical connections, instantons, moduli spaces, and compactness phenomena. In that setting, the distinction between a local potential $A$, a gauge transformation $u$, and a global curvature $F_A$ becomes essential: equations must be written locally to compute, but their meaning must survive a change of gauge.
Finally, associated bundles provide the bridge to fields and equations on manifolds. Sections of vector bundles, covariant derivatives, tensor operations, and induced connections all arise from the same equivariant construction. This is why principal bundles are not merely a reformulation of vector bundles: they isolate the symmetry group, making reduction, connection, curvature, holonomy, and gauge transformation part of one unified language.
## References
- Kobayashi, S., and Nomizu, K. *Foundations of Differential Geometry*, Vol. I. Wiley, 1963.
- Steenrod, N. *The Topology of Fibre Bundles*. Princeton University Press, 1951.
- Milnor, J., and Stasheff, J. *Characteristic Classes*. Princeton University Press, 1974.
- Donaldson, S. K., and Kronheimer, P. B. *The Geometry of Four-Manifolds*. Oxford University Press, 1990.
Contents
- Introduction
- Why Principal Bundles Are Needed
- Local Data and Global Geometry
- Connections as Infinitesimal Comparison
- Curvature, Parallel Transport, and Holonomy
- Gauge Transformations and Associated Bundles
- 1. Principal Bundles as Symmetry Objects
- From Fibres to Group Actions
- Equivariant Maps and Pullbacks
- The Product Bundle and the Maurer-Cartan Form
- 2. Transition Functions Revisited for Principal Bundles
- Principal Transition Functions and the Cocycle Identity
- Change of Trivialization and Coboundary Equivalence
- Reconstruction from a G-Valued Cocycle
- 3. Frame Bundles and Structure Groups
- Linear Frames of a Vector Bundle
- Vector Bundles from Principal Linear Bundles
- Orthonormal and Oriented Frames
- Complex and Unitary Frames
- Reductions of Structure Group
- 4. Associated Bundles and Equivariant Descriptions
- The Associated Bundle Construction
- Sections As Equivariant Maps
- Functoriality Of Associated Bundles
- Associated Vector Bundles From Representations
- Geometric Examples
- 5. Reductions of Structure Group
- The Problem of Smaller Symmetry
- Reductions as Sections of the Quotient Bundle
- Orientations as Reductions
- Metrics and Orthogonal Reductions
- Volume Forms and Special Linear Reductions
- Complex and Almost Complex Structures
- 6. Connections on Principal Bundles
- Horizontal Distributions and Vertical Tangent Spaces
- Connection Forms on the Total Space
- Local Gauge Potentials and Change of Gauge
- Frame Bundles and the Levi-Civita Connection
- The Dirac Monopole as a Local Connection
- 7. Covariant Derivatives on Associated Bundles
- From Principal Connections to Associated Covariant Derivatives
- The Local Formula
- Tensor Operations and Induced Connections
- Vector-Bundle Connections and Frame Bundles
- 8. Curvature
- The Structure Equation for Principal Curvature
- Horizontality and Equivariance of Curvature
- Local Curvature Forms and Gauge Covariance
- The Bianchi Identity
- Curvature as an Operator on Associated Bundles
- 9. Gauge Transformations and Gauge Fields
- Automorphisms Over the Identity
- Gauge Transformations as Sections of the Adjoint Group Bundle
- Connections as an Affine Space
- Gauge Action on Connections and Curvature
- Infinitesimal Gauge Action
- 10. Parallel Transport and Holonomy
- Horizontal Lifts and Parallel Transport
- Holonomy Groups
- Flat Connections and Monodromy
- 11. Connections on Reductions and Induced Geometry
- Connections Preserving Reductions
- Splitting the Connection Form Along a Reduction
- The Solder Form and Torsion on the Frame Bundle
- Levi-Civita Connection in Frame-Bundle Language
- 12. Integrative Examples and Computations
- The Hopf Connection and Monopole Curvature
- The Levi-Civita Connection on an Oriented Surface
- Matrix-Valued Gauge Potentials and the Yang-Mills Curvature Norm
- Local-to-Global Consistency Across the Examples
- Beyond and Connections
- References
Fibre Bundles II: Principal Bundles and Connections
Content
Problems
History
Created by admin on 6/11/2026 | Last updated on 6/11/2026
Prerequisites (0/6 completed)
Log in to track your prerequisite progress.
Prerequisites Graph
Interactive dependency map showing prerequisite concepts
Loading dependency graph...
Theorem
Definition
Current
Requires
Rate this page
★
★
★
★
★
Poor
Excellent