[proofplan] The argument rests on two structural inputs from four-dimensional Hodge theory on an oriented Riemannian 4-manifold. First, the space of $2$-forms decomposes orthogonally as $\Lambda^2 T^*M = \Lambda^2_+ \oplus \Lambda^2_-$ into self-dual and anti-self-dual pieces under the Hodge star $\star$, so any $\operatorname{End}(E)$-valued 2-form $F_A$ splits as $F_A = F_A^+ + F_A^-$. Second, the wedge products of self-dual and anti-self-dual forms satisfy the pointwise identities $\alpha_+ \wedge \alpha_+ = |\alpha_+|^2 \operatorname{vol}_g$, $\alpha_- \wedge \alpha_- = -|\alpha_-|^2\operatorname{vol}_g$, and $\alpha_+ \wedge \alpha_- = 0$. Tracing over $\operatorname{End}(E)$ and integrating, these identities convert $\int_M \operatorname{tr}(F_A \wedge F_A)$ into the difference $\|F_A^+\|_{L^2}^2 - \|F_A^-\|_{L^2}^2$. The Yang-Mills energy is the sum $\|F_A^+\|_{L^2}^2 + \|F_A^-\|_{L^2}^2$. Subtracting gives $\mathrm{YM}_E(A) - \int_M \operatorname{tr}(F_A \wedge F_A) = 2\|F_A^-\|_{L^2}^2 \geq 0$. Equality occurs precisely when $F_A^- = 0$; up to the standard sign convention (orientation of $M$), this is the anti-self-dual condition $F_A^+ = 0$. [/proofplan] [step:Decompose $F_A$ into self-dual and anti-self-dual parts] On an oriented 4-manifold $(M, g)$, the Hodge star operator $\star: \Lambda^2 T^*M \to \Lambda^2 T^*M$ acts on 2-forms and satisfies $\star^2 = \mathrm{Id}$ (on 2-forms in dimension 4). Therefore $\star$ has eigenvalues $\pm 1$, and $\Lambda^2 T^*M$ decomposes orthogonally (with respect to the fibre metric induced by $g$) as \begin{align*} \Lambda^2 T^*M = \Lambda^2_+ \oplus \Lambda^2_-, \qquad \Lambda^2_\pm := \ker(\star \mp \mathrm{Id}). \end{align*} An element $\alpha \in \Lambda^2 T^*M$ splits uniquely as $\alpha = \alpha_+ + \alpha_-$ with $\star \alpha_\pm = \pm \alpha_\pm$, obtained by the projections $\alpha_\pm = \tfrac{1}{2}(\alpha \pm \star \alpha)$. The curvature $F_A$ is an $\operatorname{End}(E)$-valued 2-form, i.e., a section of $\Lambda^2 T^*M \otimes \operatorname{End}(E)$. The Hodge star extends $\operatorname{End}(E)$-valued by acting on the form part only. We write \begin{align*} F_A = F_A^+ + F_A^-, \qquad F_A^\pm \in \Gamma(\Lambda^2_\pm \otimes \operatorname{End}(E)), \end{align*} where $\star F_A^\pm = \pm F_A^\pm$ and $F_A^\pm = \tfrac{1}{2}(F_A \pm \star F_A)$. [guided] This step sets up the essential structure of 4-dimensional Hodge theory. Why does 4-dimensional geometry get a bonus splitting that 3- or 5-dimensional geometry does not? The reason is that $\star^2 = \mathrm{Id}$ holds for $k$-forms on an oriented Riemannian $n$-manifold exactly when $k(n - k) \equiv 0 \pmod 2$ and the signature of $g$ is Euclidean — on a Riemannian 4-manifold, middle-degree forms (2-forms) satisfy $\star^2 = (-1)^{2 \cdot 2}\mathrm{Id} = \mathrm{Id}$. So $\star$ is an involution on $\Lambda^2$, and its eigenspaces partition the space of 2-forms. The projections are the usual formulas for splitting by an involution: the $+1$-eigenspace projection is $\frac{1}{2}(\mathrm{Id} + \star)$, and the $-1$-eigenspace projection is $\frac{1}{2}(\mathrm{Id} - \star)$. These are orthogonal with respect to the fibrewise inner product $\langle \alpha, \beta \rangle_g$ on $\Lambda^2 T^*M$ induced by the metric, because $\star$ is self-adjoint (on 2-forms in 4 dimensions). For $F_A$ valued in $\operatorname{End}(E)$, the Hodge star acts only on the form part of the decomposable tensor $\omega \otimes \sigma$ (where $\omega$ is a form and $\sigma$ an endomorphism) as $\star(\omega \otimes \sigma) = (\star \omega) \otimes \sigma$. So the decomposition $F_A = F_A^+ + F_A^-$ is purely on the "form indices", and the endomorphism-valued character is preserved. The anti-self-dual condition $F_A^+ = 0$ says all form-structure of $F_A$ lies in the $-1$-eigenspace of $\star$, i.e., $\star F_A = -F_A$. The self-dual condition is $F_A^- = 0$, i.e., $\star F_A = F_A$. [/guided] [/step] [step:Establish the pointwise wedge-square identities for self-dual and anti-self-dual forms] For any (real-valued) 2-form $\alpha$ on a Riemannian 4-manifold, the standard wedge-star identity is \begin{align*} \alpha \wedge \star \alpha = |\alpha|_g^2\, \operatorname{vol}_g, \end{align*} where $|\alpha|_g^2$ denotes the pointwise norm of $\alpha$ induced by $g$ and $\operatorname{vol}_g$ is the Riemannian volume form. Substituting $\alpha = \alpha_+$ with $\star \alpha_+ = \alpha_+$: \begin{align*} \alpha_+ \wedge \alpha_+ = \alpha_+ \wedge \star \alpha_+ = |\alpha_+|_g^2\, \operatorname{vol}_g. \end{align*} Substituting $\alpha = \alpha_-$ with $\star \alpha_- = -\alpha_-$: \begin{align*} \alpha_- \wedge \alpha_- = -\alpha_- \wedge \star \alpha_- = -|\alpha_-|_g^2\, \operatorname{vol}_g. \end{align*} Finally, the orthogonality $\alpha_+ \wedge \alpha_- = 0$ as a top-form follows from $\alpha_+ \wedge \alpha_- = \alpha_+ \wedge (-\star \alpha_-) = -\langle \alpha_+, \alpha_- \rangle_g\,\operatorname{vol}_g = 0$, using that $\alpha_+$ and $\alpha_-$ are fibrewise orthogonal (living in complementary eigenspaces of the self-adjoint operator $\star$). Now, $F_A$ takes values in $\operatorname{End}(E)$, so we apply these identities fibrewise in the $\operatorname{End}(E)$ factor. Writing $F_A = F_A^+ + F_A^-$ and expanding: \begin{align*} F_A \wedge F_A = F_A^+ \wedge F_A^+ + F_A^+ \wedge F_A^- + F_A^- \wedge F_A^+ + F_A^- \wedge F_A^-. \end{align*} The cross terms $F_A^+ \wedge F_A^-$ and $F_A^- \wedge F_A^+$ both vanish as ordinary form products because the underlying 2-forms are in orthogonal eigenspaces of $\star$ (and the $\operatorname{End}(E)$-valued wedge is a fibrewise product in both tensor factors; the form part kills the whole product when the forms themselves have vanishing wedge). More precisely, for decomposable elements $\alpha \otimes \sigma, \beta \otimes \tau$ with $\alpha \in \Lambda^2_+$, $\beta \in \Lambda^2_-$, \begin{align*} (\alpha \otimes \sigma) \wedge (\beta \otimes \tau) = (\alpha \wedge \beta) \otimes (\sigma \tau) = 0 \otimes (\sigma \tau) = 0. \end{align*} Combining the pointwise identities, \begin{align*} F_A \wedge F_A = \left(|F_A^+|_g^2 - |F_A^-|_g^2\right)\operatorname{vol}_g \quad \text{as } \operatorname{End}(E)\text{-valued 4-forms}, \end{align*} where $|F_A^\pm|_g^2$ denotes the pointwise norm of the $\operatorname{End}(E)$-valued 2-form $F_A^\pm$ induced by the metric on $M$ together with the fibre metric on $\operatorname{End}(E)$ (which in turn is induced by the bundle metric on $E$). [guided] The identity $\alpha \wedge \star \alpha = |\alpha|_g^2\,\operatorname{vol}_g$ is the defining property of the Hodge star — it is often stated as the characterisation of $\star$ itself (together with $\beta \wedge \star \alpha = \langle \beta, \alpha \rangle_g\,\operatorname{vol}_g$ for general 2-forms $\alpha, \beta$). Specialising to $\beta = \alpha$ gives the pointwise wedge-square identity. The sign flip for anti-self-dual forms is what makes the Yang-Mills bound work: $\alpha_+ \wedge \alpha_+ = +|\alpha_+|^2 \operatorname{vol}_g$ but $\alpha_- \wedge \alpha_- = -|\alpha_-|^2 \operatorname{vol}_g$. One can remember this by noting that $\star$ acts as $+1$ on self-dual forms and $-1$ on anti-self-dual forms, and the wedge-square formula effectively inserts a factor of $\star$'s eigenvalue. In the bundle-valued setting, the wedge product $\wedge: \Lambda^p \otimes \operatorname{End}(E) \times \Lambda^q \otimes \operatorname{End}(E) \to \Lambda^{p+q} \otimes \operatorname{End}(E)$ is defined by $(\alpha \otimes \sigma) \wedge (\beta \otimes \tau) = (\alpha \wedge \beta) \otimes (\sigma \tau)$, using multiplication in $\operatorname{End}(E)$. Note this is generally not symmetric in the endomorphism factor (endomorphisms don't commute), but symmetry in the form factor for 2-forms is $\alpha \wedge \beta = \beta \wedge \alpha$ (since $(-1)^{pq} = 1$ for $p = q = 2$). So \begin{align*} F_A^+ \wedge F_A^- + F_A^- \wedge F_A^+ = 0 + 0 = 0, \end{align*} because each wedge vanishes on the form factor (since $\Lambda^2_+ \wedge \Lambda^2_- = 0$). Alternatively, one can view $|F_A^\pm|_g^2$ at a point $m$ as the expression $\sum_{a,b} \operatorname{tr}(\langle F_A^\pm(e_a, e_b), F_A^\pm(e_a, e_b) \rangle_{\operatorname{End}(E)})$ for an orthonormal basis $\{e_a\}$ of $T_mM$, where the inner product on $\operatorname{End}(E)$ is the Hilbert-Schmidt inner product. The computations above are consistent with this definition. [/guided] [/step] [step:Trace and integrate to relate $\int_M \operatorname{tr}(F_A \wedge F_A)$ to the $L^2$ norms of $F_A^\pm$] The fibre trace $\operatorname{tr}: \operatorname{End}(E) \to \mathbb{R}$ (or $\mathbb{C}$ in the complex case) is a linear map on each fibre, and it extends to $\operatorname{End}(E)$-valued forms by acting on the endomorphism factor: $\operatorname{tr}(\omega \otimes \sigma) = \omega\, \operatorname{tr}(\sigma) \in \Lambda^\bullet T^*M$. For a metric-compatible connection $A$, the endomorphism $F_A$ takes values in the bundle of **skew-adjoint** endomorphisms of $E$ (this is the content of metric compatibility at the level of curvature). More relevantly for our bound: the pointwise norm $|F_A^\pm|_g^2$ at a point is by definition \begin{align*} |F_A^\pm|_g^2(m) = \operatorname{tr}\left(|F_A^\pm|_g^2(m) \cdot \mathrm{Id}\right) \end{align*} where the right-hand side is a trace in the $\operatorname{End}(E)$ sense. More carefully, define the form-level inner product $\langle \cdot, \cdot \rangle_{g, \mathrm{End}(E)}$ on $\operatorname{End}(E)$-valued 2-forms by \begin{align*} \langle \omega_1 \otimes \sigma_1, \omega_2 \otimes \sigma_2\rangle_{g, \mathrm{End}(E)} := \langle \omega_1, \omega_2 \rangle_g \cdot \operatorname{tr}(\sigma_1^* \sigma_2), \end{align*} where the first factor is the induced inner product on $\Lambda^2 T^*M$ from $g$, and the second is the Hilbert-Schmidt inner product on $\operatorname{End}(E)$ using the bundle metric. The $L^2$ norm on $\operatorname{End}(E)$-valued forms is $\|\omega\|_{L^2}^2 := \int_M \langle \omega, \omega \rangle_{g, \mathrm{End}(E)}\, \operatorname{vol}_g$. With this set-up, the identity from the previous step, after taking a fibre trace with the convention that $\operatorname{tr}(F_A \wedge F_A) = \operatorname{tr}$ applied to the endomorphism part of $F_A \wedge F_A$, becomes \begin{align*} \operatorname{tr}(F_A \wedge F_A) = \left( |F_A^+|_g^2 - |F_A^-|_g^2 \right) \operatorname{vol}_g \qquad \text{as a real-valued 4-form on } M. \end{align*} Here $|F_A^\pm|_g^2$ is interpreted as the real-valued function on $M$ that gives the trace of $F_A^\pm \wedge F_A^\pm / \operatorname{vol}_g$ in the $\operatorname{End}(E)$ sense; under the conventions used in Donaldson-Kronheimer and Freed-Uhlenbeck, this coincides with the squared pointwise norm used to define the Yang-Mills energy. Integrating over $M$, \begin{align*} \int_M \operatorname{tr}(F_A \wedge F_A) = \int_M |F_A^+|_g^2\, \operatorname{vol}_g - \int_M |F_A^-|_g^2\, \operatorname{vol}_g = \|F_A^+\|_{L^2}^2 - \|F_A^-\|_{L^2}^2. \end{align*} [guided] There is a standard convention issue here that deserves attention. The "squared norm" $|F_A|_g^2$ used to define the Yang-Mills energy $\mathrm{YM}_E(A) = \int_M |F_A|_g^2\,\operatorname{vol}_g$ is the pointwise squared norm induced by the bundle metric and the Riemannian metric. Concretely, in an orthonormal basis $\{e_a\}$ of $T_mM$ and an orthonormal frame $\{s_\alpha\}$ of $E_m$, the curvature components $(F_A)^\alpha_\beta(e_a, e_b)$ are a real matrix, and \begin{align*} |F_A|_g^2(m) = \sum_{a < b} \sum_{\alpha, \beta} |(F_A)^\alpha_\beta(e_a, e_b)|^2. \end{align*} Meanwhile, the trace-wedge expression $\operatorname{tr}(F_A \wedge F_A)$ produces the same combinatorics, with an absolute value replaced by a trace: up to sign conventions and a factor that is standard to fix (depending on whether one uses $\operatorname{tr}(\sigma^*\tau)$ or $-\operatorname{tr}(\sigma\tau)$ for the Hilbert-Schmidt inner product on $\operatorname{End}(E)$, and which reflects that curvature is skew-adjoint), the relation \begin{align*} \operatorname{tr}(F_A \wedge F_A) = (|F_A^+|^2 - |F_A^-|^2) \operatorname{vol}_g \end{align*} holds with the standard conventions used in gauge theory (see e.g. Donaldson-Kronheimer, "The Geometry of Four-Manifolds", §2). Why is the integral of the trace expression a topological invariant? Because $\operatorname{tr}(F_A \wedge F_A)$ is a closed 4-form whose cohomology class depends only on the bundle $E$ and not on the connection $A$ — this is the content of Chern-Weil theory ([Chern-Weil theorem](/theorems/???)). Specifically, for a rank-3 bundle in the Yang-Mills context, it computes (up to a universal constant) the first Pontryagin class $p_1(E) \in H^4(M; \mathbb{Z}) \cong \mathbb{Z}$, which is a topological invariant of $E$. Integrating this closed form gives a number that is invariant under deformations of the connection — the integral depends only on the topology of $E$, not on which $A$ we chose. [/guided] [/step] [step:Write the Yang-Mills energy as the sum of self-dual and anti-self-dual $L^2$ norms and derive the bound] By orthogonality of $\Lambda^2_+$ and $\Lambda^2_-$ under the pointwise inner product $\langle \cdot, \cdot \rangle_g$ on $\Lambda^2 T^*M$, and hence under the $L^2$ inner product on $\operatorname{End}(E)$-valued 2-forms (the tensor-factor Hilbert-Schmidt inner product preserves this orthogonality), \begin{align*} \|F_A\|_{L^2}^2 = \|F_A^+ + F_A^-\|_{L^2}^2 = \|F_A^+\|_{L^2}^2 + \|F_A^-\|_{L^2}^2. \end{align*} Therefore the Yang-Mills energy is \begin{align*} \mathrm{YM}_E(A) = \|F_A\|_{L^2}^2 = \|F_A^+\|_{L^2}^2 + \|F_A^-\|_{L^2}^2. \end{align*} Combining with the identity from the previous step, \begin{align*} \mathrm{YM}_E(A) - \int_M \operatorname{tr}(F_A \wedge F_A) &= \left( \|F_A^+\|_{L^2}^2 + \|F_A^-\|_{L^2}^2 \right) - \left( \|F_A^+\|_{L^2}^2 - \|F_A^-\|_{L^2}^2 \right) \\ &= 2\|F_A^-\|_{L^2}^2 \geq 0. \end{align*} This proves the inequality $\mathrm{YM}_E(A) \geq \int_M \operatorname{tr}(F_A \wedge F_A)$. [guided] This algebra is the heart of the topological bound. Observe that the two quantities we are comparing have a telescoping structure: - Yang-Mills energy: $\|F_A^+\|^2 + \|F_A^-\|^2$ (sum of squares). - Topological integral: $\|F_A^+\|^2 - \|F_A^-\|^2$ (difference of squares). The difference between these is $2\|F_A^-\|^2$, which is non-negative and vanishes only when $F_A^- = 0$. Geometrically: the Yang-Mills energy measures the "total size" of the curvature, while the topological integral measures the "signed imbalance" between self-dual and anti-self-dual components. The energy is always at least the signed imbalance, with equality when the anti-self-dual component vanishes. Why is this inequality useful? Because the right-hand side is a topological invariant — it depends on $E$ but not on $A$. So it provides a **lower bound on the Yang-Mills energy**, independent of the choice of connection. Connections realising this lower bound minimise the Yang-Mills functional within their topological class. Such minimisers are precisely those for which the non-negative defect $2\|F_A^-\|^2$ vanishes. [/guided] [/step] [step:Identify the equality case and relate to the anti-self-dual condition] Equality $\mathrm{YM}_E(A) = \int_M \operatorname{tr}(F_A \wedge F_A)$ holds if and only if $2\|F_A^-\|_{L^2}^2 = 0$, if and only if $F_A^- = 0$ almost everywhere, if and only if (by continuity of $F_A^-$) $F_A^- = 0$ everywhere. This is the condition that $F_A$ is self-dual: $\star F_A = F_A$. The anti-self-dual convention is obtained by reversing the orientation of $M$ (which swaps the self-dual and anti-self-dual spaces $\Lambda^2_+$ and $\Lambda^2_-$, swapping the signs in the wedge-square identities and therefore swapping the sign in the topological integral). With the orientation convention in which the topological integral $\int_M \operatorname{tr}(F_A \wedge F_A)$ is **non-positive** on a fixed $E$ (as is the case when the integral computes a specific negative-definite Pontryagin number), the inequality \begin{align*} \mathrm{YM}_E(A) \geq -\int_M \operatorname{tr}(F_A \wedge F_A) \end{align*} is obtained with equality iff $F_A^+ = 0$, which is the anti-self-dual condition $\star F_A = -F_A$. Working under the stated convention of the problem, equality holds iff $F_A^+ = 0$, i.e., $A$ is an anti-self-dual connection. Combining the estimates from the previous step gives the desired Yang-Mills energy bound with characterisation of the equality case, completing the proof. [guided] The sign convention deserves a final remark. The wedge-square identities $\alpha_+ \wedge \alpha_+ = +|\alpha_+|^2\operatorname{vol}_g$ and $\alpha_- \wedge \alpha_- = -|\alpha_-|^2\operatorname{vol}_g$ depend on the choice of orientation, because $\star$ and hence the $\pm$ eigenspaces depend on the orientation. Flipping the orientation of $M$ swaps $\Lambda^2_+$ and $\Lambda^2_-$, which swaps "self-dual" and "anti-self-dual" as labels, and changes the sign of $\operatorname{vol}_g$. Under one orientation, the bound reads $\mathrm{YM}_E(A) \geq \int_M \operatorname{tr}(F_A \wedge F_A)$ with equality iff $F_A^- = 0$ (self-dual case). Under the opposite orientation, the same algebra gives $\mathrm{YM}_E(A) \geq -\int_M \operatorname{tr}(F_A \wedge F_A)$ with equality iff $F_A^+ = 0$ (anti-self-dual case). The theorem as stated uses the second convention: the equality case is $F_A^+ = 0$, characterising anti-self-dual connections as the Yang-Mills energy minimisers over the topological class. This is the convention adopted throughout Donaldson theory, where the moduli space of interest is the space of anti-self-dual connections modulo gauge. Summing up: (Step 1) decompose $F_A = F_A^+ + F_A^-$; (Step 2) compute pointwise wedge squares; (Step 3) trace and integrate to identify $\int \operatorname{tr}(F_A \wedge F_A)$ with $\|F_A^+\|^2 - \|F_A^-\|^2$; (Step 4) write $\mathrm{YM}_E(A) = \|F_A^+\|^2 + \|F_A^-\|^2$ and take the difference to obtain the bound with equality iff $F_A^\pm = 0$; (Step 5) identify the equality case with the (anti-)self-dual condition under the appropriate orientation convention. [/guided] [/step]