betaFeedback

WorksEngineering

Quantum Natural Muon: Spectral-Natural Optimization for Variational Quantum Circuits

Abstract

Training parameterized quantum circuits requires optimization methods that account for two distinct structures: the information geometry of quantum states and the matrix or tensor structure induced by circuit parameterization. We introduce Quantum Natural Muon (QNM), a spectral natural-gradient framework that applies polar normalization to circuit-semantic parameter blocks after damped quantum Fisher preconditioning. This ordering is essential: quantum Fisher preconditioning and blockwise polar normalization do not commute in general, so QNM is not an order-free composition of quantum natural gradient (QNG) and Muon. We show that Frobenius-matched polar mixing preserves active singular subspaces while raising a spectral floor and improving active conditioning, and that the resulting direction remains a controlled perturbation of QNG. For its distance-aware realization (DA-QNM), we further establish alignment, descent, and nonconvex stationarity guarantees for the normalized update, together with perturbation bounds under gradient and QFIM estimation error. Experiments on variational quantum eigensolving, weighted QAOA-MaxCut, and supervised quantum classification isolate the roles of quantum geometry, circuit-semantic spectral structure, and radius control. The strongest gains arise when the spectral blocks match the circuit structure, supporting the central principle that quantum-state and structured spectral geometries provide complementary information for variational quantum optimization.

Introduction

Parameterized quantum circuits (PQCs) underlie variational quantum eigensolving, quantum approximate optimization, and quantum learning [4, 2, 16, 7]. Their optimization has two distinct structures. First, Euclidean parameter distance does not generally reflect distance between the induced quantum states: different parameter directions can have very different local state sensitivities [1, 19]. Second, PQC parameters are not unstructured coordinates; circuit architectures organize them by layers, qubits, gate families, Hamiltonian terms, graph edges, or data-reuploading blocks [2, 4, 7, 15]. Variational quantum optimization therefore couples quantum-state information geometry with circuit-induced parameter structure.

Existing optimizers address these structures separately. Adam and related Euclidean methods adapt coordinate-wise scales but do not account for quantum-state geometry [6, 14]. Quantum natural gradient (QNG) instead preconditions gradients with the quantum Fisher information matrix (QFIM), calibrating updates to local state sensitivity [19]; however, the resulting natural-gradient vector does not control how update magnitude is distributed across circuit-structured matrix or tensor blocks. Muon addresses a different geometry by reshaping momentum into matrices and applying polar normalization to their singular spectra [12, 3]. Distance-Aware Muon further separates normalized direction selection from scalar displacement control [5], but neither method incorporates quantum-state geometry.

Exploiting both geometries creates several nontrivial design problems. Spectral normalization can be applied before or after QFIM preconditioning, but these operations do not commute in general and therefore define different update directions. Full polar normalization also equalizes active singular values at the cost of discarding block-scale information, whereas an aggressive spectral modification can move the update far enough from QNG to lose its descent property. Finally, normalizing the resulting direction separates direction selection from displacement magnitude, so the step radius must be controlled independently. The central question is therefore not how to concatenate QNG and Muon, but how to exploit circuit-level spectral structure while preserving the quantum-geometric and descent information of the natural-gradient direction.

We introduce Quantum Natural Muon (QNM), a spectral natural-gradient framework for structured variational quantum optimization. Given gradient gtg_t, QNM first computes the damped natural-gradient vector ut=(Ft+λI)−1gtu_t = (\mathbf{F}_t + \lambda\mathbf{I})^{-1}g_t and accumulates momentum in this quantum-geometric space. It then reshapes the momentum according to circuit semantics and mixes each block with a Frobenius-matched polar factor, balancing its active spectrum without replacing its block-scale information. This ordering is deliberate: because QFIM preconditioning and blockwise polar normalization do not commute, applying spectral normalization before state-sensitivity calibration generally defines a different optimizer. QNM instead performs spectral direction selection after quantum-geometric calibration. Its distance-aware realization, DA-QNM, normalizes the resulting spectral-natural direction and controls the final displacement with an adaptive radius; the radius rule itself is not claimed as new.

Our analysis characterizes both what this transformation changes and what it preserves. The polar factor solves a spectral-norm-constrained local direction-selection problem. With Frobenius-matched scaling, polar mixing preserves the active singular subspaces of a structured natural-gradient block while raising a spectral floor and improving its active spectral conditioning. We then bound the deviation of the resulting direction from QNG and derive sufficient conditions for gradient alignment and descent. For the normalized DA-QNM update, we connect the radius generated by the algorithm directly to descent and establish an O(1/T)O(1/T) nonconvex stationarity guarantee. We additionally bound the sensitivity of the natural-gradient direction to gradient and QFIM estimation errors, quantifying the stabilizing role of damping.

We evaluate QNM on VQE, weighted generalized QAOA-MaxCut, and supervised quantum classification, which expose layer-, gate/graph-, and data-reuploading structure, respectively ([16, 4, 7, 9, 15]). Rather than only comparing final objectives, our ablations isolate quantum preconditioning, spectral normalization, semantic block choice, and radius control. On the weighted n=10n = 10, p=3p = 3 QAOA instance, QNG reduces the Adam optimality gap by 43.0%, while gate-structured QNM reduces it by 96.9% and DA-QNM by 97.0%. In contrast, Euclidean, layer-wise, and global spectral variants do not obtain the same gain, showing that spectral normalization alone is insufficient. Consistent with the predicted mechanism, polar mixing reduces the measured active condition number of a representative structured update from 10.13 to 1.85 while increasing its effective rank from 2.22 to 2.84. The gains are therefore strongest when quantum-state geometry and circuit-semantic spectral structure are jointly informative.

Our main contributions are:

  1. Structured spectral natural gradients. We introduce QNM, which performs circuit-semantic spectral direction selection after damped QFIM preconditioning. Because QFIM preconditioning and polar normalization do not commute, their ordering defines different optimizers; QNM deliberately regularizes the structured spectrum only after state-sensitivity calibration.

  1. Spectral and optimization guarantees. We show that Frobenius-matched polar mixing preserves active singular subspaces, raises a spectral floor, and improves active spectral conditioning while remaining a controlled perturbation of QNG, yielding explicit alignment and descent conditions.

  1. Normalized-update convergence and estimation sensitivity. For DA-QNM, we connect the algorithm’s generated radius to descent and O(1/T)O(1/T) nonconvex stationarity and derive perturbation bounds under gradient and QFIM estimation error.

  1. Mechanism-driven empirical validation. Across VQE, weighted QAOA, and supervised quantum classification, component and block-structure ablations show that the gains arise from jointly exploiting quantum-state geometry and circuit-semantic spectral structure rather than generic normalization or radius control.

Preliminaries

We formalize the two structures used by QNM: quantum-state geometry, represented by the quantum Fisher information matrix (QFIM), and circuit-semantic block structure.

Quantum Models and Circuit Structure

Let U(θ)\mathbf{U}(\theta) be a parameterized quantum circuit with θ∈Rd\theta\in\mathbb{R}^{d} and ∣ψ(θ)⟩=U(θ)∣0⟩⊗n|\psi(\theta)\rangle= \mathbf{U}(\theta)|0\rangle^{\otimes n}. We consider Hamiltonian objectives C(θ)=⟨ψ(θ)∣H∣ψ(θ)⟩C(\theta) = \langle\psi(\theta)|\mathbf{H}|\psi(\theta)\rangle, covering VQE and QAOA-type problems [16, 4, 7, 8], as well as empirical losses for supervised variational quantum models [18, 9, 15].

Across these settings, quantum parameters have circuit-induced structure: hardware-efficient ansatzes organize parameters by layer, qubit, and rotation axis; generalized QAOA by depth, graph edge, and node; and data-reuploading circuits by layer and encoding block. Let S={1,…,S}\mathcal{S}=\{1,\ldots,S\} index parameter blocks. For each s∈Ss\in\mathcal{S}, let Is⊆{1,…,d}\mathcal{I}_s\subseteq\{1,\ldots,d\} index a block, θ(s)=θIs∈Rds\theta^{(s)}=\theta_{\mathcal{I}_s}\in\mathbb{R}^{d_s}, and define a circuit-semantic reshape

Rs:Rds→Ras×bs,M(s)=Rs(θ(s)).(1)\mathcal{R}_s:\mathbb{R}^{d_s}\to\mathbb{R}^{a_s\times b_s},\qquad\mathbf{M}^{(s)}=\mathcal{R}_s(\theta^{(s)}). \tag*{(1)}

The maps Rs\mathcal{R}_s determine which circuit-level correlations are exposed to spectral normalization; task-specific choices are given in § 5 and Appendix § 9.

Quantum Natural Gradient

For pure states, the QFIM is the real part of the quantum geometric tensor [17, 19],

[F(θ)]ij=Re⁡[⟨∂iψ∣∂jψ⟩−⟨∂iψ∣ψ⟩⟨ψ∣∂jψ⟩].(2)[\mathbf{F}(\theta)]_{ij}=\operatorname{Re}\left[\langle\partial_i\psi|\partial_j\psi\rangle-\langle\partial_i\psi|\psi\rangle\langle\psi|\partial_j\psi\rangle\right]. \tag*{(2)}

It measures local quantum-state sensitivity to parameter perturbations. With gt=∇C(θt)\mathbf{g}_t=\nabla C(\theta_t) and damping λ>0\lambda>0, define Gt=Ft+λI\mathbf{G}_t=\mathbf{F}_t+\lambda\mathbf{I} and ut=Gt−1gt\mathbf{u}_t=\mathbf{G}_t^{-1}\mathbf{g}_t. Here ut\mathbf{u}_t is the damped QNG vector and −ut-\mathbf{u}_t is the corresponding descent direction [19]. Additional geometric details and the equivalent local variational characterization are given in Appendix § 7.§ 2.

QNG calibrates parameter motion to quantum-state geometry, but imposes no spectral constraint on the circuit-semantic matrices Rs(ut(s))\mathcal{R}_s(u_t^{(s)}). QFIM preconditioning therefore need not remove spectral anisotropy within the resulting structured update blocks. This is an update-structure property, distinct from conditioning of the objective landscape.

Polar Spectral Normalization

Muon-style methods act instead on matrix-update geometry [12, 3]. For a nonzero block M=UΣVT\mathbf{M}=\mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^{\mathsf{T}} with compact SVD, define Polar⁡(M)=UVT\operatorname{Polar}(\mathbf{M})=\mathbf{U}\mathbf{V}^{\mathsf{T}}. The polar factor preserves the active left and right singular subspaces while setting all active singular values to one. It therefore re-moves active spectral imbalance without rotating those subspaces. Its spectral-norm variational characterization, used in § 4 to interpret QNM as constrained matrix direction selection, is detailed in Appendix § 7.§ 3.

Polar normalization thus addresses a geometry different from QNG: QNG calibrates directions by quantum-state sensitivity, whereas the polar operation controls the active spectrum of a structured matrix update. Full polar normalization, however, replaces all active singular values by one and therefore discards their relative magnitudes and the original block scale. This motivates the scale-preserving polar mixing introduced next. Moreover, because QFIM preconditioning and block-wise polar normalization generally do not commute, their ordering is consequential. § 3 constructs QNM by applying spectral normalization after quantum preconditioning.

Quantum Natural Muon

We now define Quantum Natural Muon (QNM) and its distance-aware realization (DA-QNM). QNM applies spectral normalization to momentum accumulated in natural-gradient space; DA-QNM additionally normalizes the resulting spectral-natural direction and controls its displacement through a distance-aware radius. The update pipeline is

gt→ut=(Ft+λI)−1gt→mt→{Mt(s)}s∈S→{Dt(s)}s∈S→dt→qt→θt+1=θt−rtqt.(3)\mathbf{g}_t\to\mathbf{u}_t=(\mathbf{F}_t+\lambda\mathbf{I})^{-1}\mathbf{g}_t\to\mathbf{m}_t\to\{\mathbf{M}_t^{(s)}\}_{s\in\mathcal{S}}\to\{\mathbf{D}_t^{(s)}\}_{s\in\mathcal{S}}\to\mathbf{d}_t\to\mathbf{q}_t\to\theta_{t+1}=\theta_t-r_t\mathbf{q}_t. \tag*{(3)}

Here gt=∇C(θt)\mathbf{g}_t=\nabla C(\theta_t) is the Euclidean gradient, ut\mathbf{u}_t is the damped natural-gradient vector, mt\mathbf{m}_t is natural-gradient momentum, Mt(s)\mathbf{M}_t^{(s)} are circuit-semantic matrix blocks, Dt(s)\mathbf{D}_t^{(s)} are polar-mixed blocks, dt\mathbf{d}_t is the concatenated direction, qt\mathbf{q}_t is the normalized direction, and rtr_t is the distance-aware radius.

Natural-Gradient Momentum Blocks

Rather than forming Gt−1\mathbf{G}_t^{-1} explicitly, DA-QNM computes the damped natural-gradient vector by solving Gtut=gt\mathbf{G}_t\mathbf{u}_t=\mathbf{g}_t. The final parameter update moves in the negative direction of the transformed vector, so ut\mathbf{u}_t is the natural-gradient vector and −ut-\mathbf{u}_t is the corresponding descent displacement. DA-QNM accumulates momentum in natural-gradient space:

mt=βmt−1+(1−β)ut,β∈[0,1].(4)\mathbf{m}_t=\beta\mathbf{m}_{t-1}+(1-\beta)\mathbf{u}_t,\qquad\beta\in[0,1]. \tag*{(4)}

Thus, the subsequent spectral normalization acts on a quantum-geometrically calibrated direction rather than on the raw Euclidean gradient.

For each block s∈Ss\in\mathcal{S}, let mt(s)∈Rds\mathbf{m}_t^{(s)}\in\mathbb{R}^{d_s} denote the corresponding subvector of mt\mathbf{m}_t. Using the circuit-semantic reshape map Rs\mathcal{R}_s, DA-QNM forms the matrix block

Mt(s)=Rs(mt(s)).(5)\mathbf{M}_t^{(s)}=\mathcal{R}_s\left(\mathbf{m}_t^{(s)}\right). \tag*{(5)}

The choice of Rs\mathcal{R}_s determines which circuit-induced correlations are spectrally regularized. In our experiments, we use layer-wise blocks for VQE, gate-wise cost/mixer blocks for generalized QAOA, and data-reuploading blocks for supervised quantum classifiers.

Blockwise Spectral Normalization

For each structured momentum block Mt(s)\mathbf{M}^{(s)}_t, let

Mt(s)=Ut(s)Σt(s)Vt(s)T.(6)\mathbf{M}^{(s)}_t = \mathbf{U}^{(s)}_t \boldsymbol{\Sigma}^{(s)}_t \mathbf{V}^{(s)\mathsf{T}}_t . \tag*{(6)}

be its compact singular value decomposition. If Mt(s)≠0\mathbf{M}^{(s)}_t \ne0, define its polar factor as

Pt(s)=Polar⁡(Mt(s))=Ut(s)Vt(s)T.(7)\mathbf{P}^{(s)}_t = \operatorname{Polar}(\mathbf{M}^{(s)}_t) = \mathbf{U}^{(s)}_t \mathbf{V}^{(s)\mathsf{T}}_t . \tag*{(7)}

The polar factor preserves the active left and right singular subspaces of Mt(s)\mathbf{M}^{(s)}_t, while setting its nonzero singular values to one. Replacing Mt(s)\mathbf{M}^{(s)}_t entirely by its polar factor can be too aggressive because it discards block magnitude information. DA-QNM therefore uses Frobenius-preserving polar mixing:

Dt(s)=(1−α)Mt(s)+αct(s)Pt(s),α∈[0,1].(8)\mathbf{D}^{(s)}_t = (1-\alpha)\mathbf{M}^{(s)}_t + \alpha c^{(s)}_t\mathbf{P}^{(s)}_t,\qquad\alpha\in[0,1]. \tag*{(8)}

The scalar ct(s)c^{(s)}_t controls the scale of the polar component. With Frobenius-norm preservation, we set

ct(s)=∥Mt(s)∥F∥Pt(s)∥F=∥Mt(s)∥Frank⁡(Mt(s)).(9)c^{(s)}_t = \frac{\lVert\mathbf{M}^{(s)}_t\rVert_F}{\lVert\mathbf{P}^{(s)}_t\rVert_F} = \frac{\lVert\mathbf{M}^{(s)}_t\rVert_F}{\sqrt{\operatorname{rank}(\mathbf{M}^{(s)}_t)}}. \tag*{(9)}

If Mt(s)=0\mathbf{M}^{(s)}_t=0, we set Dt(s)=0\mathbf{D}^{(s)}_t=0. The scaling in (9) matches the Frobenius norm of the polar component to that of the original block while redistributing magnitude across active singular directions. The blockwise matrices are then vectorized and concatenated dt=vec⁡({Dt(s)}s∈S)\mathbf{d}_t=\operatorname{vec}(\{\mathbf{D}^{(s)}_t\}_{s\in\mathcal{S}}). This vector is the fixed-scale spectral-natural direction. The distance-aware step further normalizes dt\mathbf{d}_t and chooses a scalar radius.

Distance-Aware Radius Control

Polar mixing controls the structured direction but partially changes magnitude information. DA-QNM therefore separates direction from step length by defining the normalized spectral-natural direction

qt=dt∥dt∥2+ϵ,ϵ>0,(10)\mathbf{q}_t = \frac{\mathbf{d}_t}{\lVert\mathbf{d}_t\rVert_2+\epsilon},\qquad\epsilon>0, \tag*{(10)}

and updating θt+1=θt−rtqt\theta_{t+1}=\theta_t-r_t\mathbf{q}_t. Distance-aware radius is

rt=min⁡{η∥mt∥2,Δ0+ρmax⁡0≤k≤t∥θk−θ0∥2,Rmax⁡}.(11)r_t=\min\left\{\eta\lVert\mathbf{m}_t\rVert_2,\Delta_0+\rho\max_{0\le k\le t}\lVert\theta_k-\theta_0\rVert_2,R_{\max}\right\}. \tag*{(11)}

The first term ties the step length to the natural-gradient momentum magnitude. The second adapts the radius to the trajectory displacement from initialization. The third prevents extreme normalized steps.

We do not claim that the radius rule itself is new; rather, DA-QNM uses distance-aware scaling to control the step length after QFIM-preconditioned spectral direction selection. Thus, the core contribution is the spectral-natural direction, while distance-aware control stabilizes the normalized displacement produced by that direction. This decomposition separates the three roles of DA-QNM: QFIM preconditioning calibrates the direction to quantum-state geometry, polar mixing balances the active spectrum in circuit-semantic blocks, and distance-aware radius control determines the scalar displacement along the normalized direction.

Why the order matters. DA-QNM applies polar mixing after QFIM preconditioning. To formalize the ordering, consider a single block ss and define the blockwise metric-preconditioning operator

At,s(M)=Rs([Gt−1ιs(vec⁡(M))](s)),(12)\mathcal{A}_{t,s}(\mathbf{M})=\mathcal{R}_s\left(\left[\mathbf{G}_t^{-1}\iota_s(\operatorname{vec}(\mathbf{M}))\right]^{(s)}\right), \tag*{(12)}

where ιs\iota_s embeds a block vector into the full parameter space by zero-padding outside ss, and [⋅](s)[\cdot]^{(s)} extracts block-ss coordinates. The operator At,s\mathcal{A}_{t,s} represents the effect of QFIM preconditioning on a matrix block after embedding it into the full parameter vector. In general,

Polar⁡(At,s(M))≠At,s(Polar⁡(M)).(13)\operatorname{Polar}\left(\mathcal{A}_{t,s}(\mathbf{M})\right)\ne\mathcal{A}_{t,s}\left(\operatorname{Polar}(\mathbf{M})\right). \tag*{(13)}

Remark 1. QFIM preconditioning and blockwise polar normalization do not commute in general. DA-QNM first computes a quantum-geometrically calibrated direction and then regularizes its structured spectrum, producing a spectral natural-gradient update distinct from both QNG and Muon.

Theoretical Analysis

This section formalizes why DA-QNM is a structured spectral natural-gradient method. The results establish four claims: polar factors solve local spectral-norm direction-selection problems; QFIM preconditioning and blockwise polar normalization do not commute in general; polar mixing improves active spectral conditioning; and the resulting update admits descent, stationarity, and estimation-robustness guarantees under standard assumptions. Complete proofs are deferred to Appendix § 8.

Assumption 1 (Smoothness). The objective CC is LL-smooth, i.e., C(θ+δ)≤C(θ)+⟨∇C(θ),δ⟩+L2∥δ∥22C(\theta+\delta)\le C(\theta)+\langle\nabla C(\theta),\delta\rangle+\frac{L}{2}\lVert\delta\rVert_2^2 for all θ,δ∈Rd\theta,\delta\in\mathbb{R}^d.

Assumption 2 (Damped QFIM boundedness). For all iterates, the damped QFIM Gt=Ft+λI\mathbf{G}_t=\mathbf{F}_t+\lambda\mathbf{I}, λ>0\lambda>0 satisfies μI⪯Gt⪯ΛI\mu\mathbf{I}\preceq\mathbf{G}_t\preceq\Lambda\mathbf{I} for constants 0<μ≤Λ<∞0<\mu\le\Lambda<\infty.

Assumption 3 (Block decomposition). The circuit-semantic blocks form a norm-consistent decomposition of the parameter coordinates. Specifically, for any v∈Rd\mathbf{v}\in\mathbb{R}^d, let v(s)∈Rds\mathbf{v}^{(s)}\in\mathbb{R}^{d_s} be its block-ss subvector and define M(s)=Rs(v(s))\mathbf{M}^{(s)}=\mathcal{R}_s(\mathbf{v}^{(s)}). We assume ∑s∈S∥M(s)∥F2=∥v∥22\sum_{s\in\mathcal{S}}\lVert\mathbf{M}^{(s)}\rVert_F^2=\lVert\mathbf{v}\rVert_2^2.

Algorithm 1 Distance-Aware Quantum Natural Muon
Require: $\theta_0$, $\lambda$, $\beta$, $\alpha$, $\eta$, $\Delta_0$, $\rho$, $R_{\max}$, $\epsilon$, $\{R_s\}_{s\in S}$
1: $m_{-1} \leftarrow 0$
2: for $t = 0, 1, 2, \ldots$ do
3: Estimate gradient $g_t = \nabla C(\theta_t)$
4: Compute QFIM $F_t$, and set $G_t = F_t + \lambda I$
5: Solve $G_t u_t = g_t$ for $u_t$
6: $m_t \leftarrow \beta m_{t-1} + (1-\beta)u_t$
7: for each block $s \in S$ do
8: $M_t^{(s)} \leftarrow R_s(m_t^{(s)})$
9: if $M_t^{(s)} = 0$ then
10: $D_t^{(s)} \leftarrow 0$
11: else
12: $P_t^{(s)} \leftarrow \operatorname{Polar}(M_t^{(s)})$
13: $c_t^{(s)} \leftarrow \lVert M_t^{(s)}\rVert_F / \lVert P_t^{(s)}\rVert_F$
14: $D_t^{(s)} \leftarrow (1-\alpha)M_t^{(s)} + \alpha c_t^{(s)}P_t^{(s)}$
15: end if
16: end for
17: $d_t \leftarrow \operatorname{vec}(\{D_t^{(s)}\}_{s\in S})$
18: $q_t \leftarrow d_t/(\lVert d_t\rVert_2 + \epsilon)$
19: $R_t^{\mathrm{dist}} \leftarrow \Delta_0 + \rho\max_{0\leq k\leq t}\lVert\theta_k-\theta_0\rVert_2$
20: $r_t \leftarrow \min\{\eta\lVert m_t\rVert_2, R_t^{\mathrm{dist}}, R_{\max}\}$
21: $\theta_{t+1} \leftarrow \theta_t-r_tq_t$
22: end for

Assumption 4 (Alignment of normalized direction). For the normalized DA-QNM direction qtq_t, there exists χt>0\chi_t > 0 such that ⟨gt,qt⟩≥χt∥gt∥2\langle g_t, q_t\rangle\geq\chi_t\lVert g_t\rVert_2.

Assumption 5 (Lower bounded objective). The objective is bounded below: there exists C⋆>−∞C_\star> -\infty such that C(θ)≥C⋆C(\theta)\geq C_\star for all θ∈Rd\theta\in\mathbb{R}^d.

We first state the local direction-selection property of the polar factor.

Theorem 1 (Spectral-norm steepest direction). Let M∈Ra×bM \in\mathbb{R}^{a\times b} have compact SVD, M=UΣV⊤M=U\Sigma V^\top with rank rr. If M≠0M\ne0, then its compact polar factor Polar⁡(M)=UV⊤\operatorname{Polar}(M)=UV^\top solves Polar⁡(M)∈arg⁡max⁡∥D∥2≤1⟨M,D⟩F\operatorname{Polar}(M) \in\arg\max_{\lVert D\rVert_2\leq1}\langle M,D\rangle_F. The optimal value is ∥M∥∗\lVert M\rVert_*. If M=0M=0, every feasible DD is optimal and the optimal value is zero.

Theorem 1 gives the polar step a local spectral trust-region interpretation: for a structured natural-gradient momentum Mt(s)M_t^{(s)}, the polar factor is the steepest feasible matrix direction under a spectral-norm constraint. Thus, after QFIM preconditioning, QNM performs constrained direction selection in each circuit-semantic block rather than merely normalizing a matrix.

We next formalize the non-commutativity claim introduced in § 3. Recall the blockwise metric-preconditioning operator At,sA_{t,s} from (12).

Proposition 1 (Non-commutativity of metric preconditioning and polar normalization). In general,

Polar⁡(At,s(M))≠At,s(Polar⁡(M)).(14)\operatorname{Polar}\left(A_{t,s}(M)\right) \ne A_{t,s}\left(\operatorname{Polar}(M)\right). \tag*{(14)}

Thus, QFIM preconditioning and blockwise polar normalization do not commute in general. Equality would require restrictive compatibility between Gt−1G_t^{-1} and the block reshape RsR_s, such as isotropic action on the relevant block subspace.

Proposition 1 motivates the algorithmic design: DA-QNM applies spectral normalization after quantum-metric calibration, yielding a spectral natural-gradient direction distinct from QNG and Muon. We next characterize how polar mixing modifies the active singular spectrum. For a nonzero matrix MM of rank rr, define the active condition number κact(M)=σ1(M)σr(M)\kappa_{\mathrm{act}}(M) = \frac{\sigma_1(M)}{\sigma_r(M)}, where σ1(M)≥⋯≥σr(M)>0\sigma_1(M) \geq\cdots\geq\sigma_r(M) > 0 are nonzero singular values.

Proposition 2 (Active spectrum of the polar factor). Let M∈Ra×bM \in\mathbb{R}^{a\times b}. If M≠0M\ne0, let r=rank⁡(M)r=\operatorname{rank}(M) and let P=Polar⁡(M)P=\operatorname{Polar}(M). Then all active singular values of PP equal one. Consequently, κact(P)=1\kappa_{\mathrm{act}}(P)=1.

Proposition 3 (Frobenius-matched spectral balancing). Let M∈Ra×bM \in\mathbb{R}^{a\times b} be nonzero with rank rr, let P=Polar⁡(M)P=\operatorname{Polar}(M) and define c=∥M∥F∥P∥Fc=\frac{\lVert M\rVert_F}{\lVert P\rVert_F}. Then ∥cP∥F=∥M∥F\lVert cP\rVert_F=\lVert M\rVert_F. Moreover, every active singular value of cPcP equals ∥M∥Fr\frac{\lVert M\rVert_F}{\sqrt{r}}. If M=0M=0, we use the convention cP=0cP=0.

Proposition 3 explains the Frobenius-matched scaling used in DA-QNM. The scaling matches the polar component’s Frobenius norm to that of the original block while redistributing magnitude uniformly across active singular directions. We next analyze the mixed block direction DD. The following result shows that polar mixing raises a floor on the active singular spectrum while preserving the active singular subspaces.

Proposition 4 (Spectral floor from polar mixing). Let M∈Ra×bM \in\mathbb{R}^{a\times b} be nonzero with compact SVD, i.e., M=UΣV⊤M=U\Sigma V^\top, Σ=diag⁡(σ1,…,σr)\Sigma=\operatorname{diag}(\sigma_1,\ldots,\sigma_r) and σi>0\sigma_i>0. Let P=Polar⁡(M)=UV⊤P=\operatorname{Polar}(M)=UV^\top and define D=(1−α)M+αcPD=(1-\alpha)M+\alpha cP. Then DD has the same active left and right singular subspaces as MM, and its active singular values are σ~i=(1−α)σi+αc\widetilde{\sigma}_i=(1-\alpha)\sigma_i+\alpha c. In particular, σ~i≥αc\widetilde{\sigma}_i\geq\alpha c for all i=1,…,ri=1,\ldots,r.

Corollary 1 (Active condition-number improvement). Suppose σ1≥⋯≥σr>0\sigma_1\geq\cdots\geq\sigma_r>0. Then,

κact(D)=(1−α)σ1+αc(1−α)σr+αc≤σ1σr=κact(M).(15)\kappa_{\mathrm{act}}(D)=\frac{(1-\alpha)\sigma_1+\alpha c}{(1-\alpha)\sigma_r+\alpha c}\leq\frac{\sigma_1}{\sigma_r}=\kappa_{\mathrm{act}}(M). \tag*{(15)}

The inequality is strict if α>0\alpha>0, c>0c>0, and σ1>σr\sigma_1>\sigma_r.

Proposition 4 and Corollary 1 show that polar mixing lifts weak active singular directions while preserving active singular subspaces. This is the spectral-conditioning mechanism underlying QNM.

We now show that the fixed-scale spectral-natural direction remains close to QNG when the polar mixing coefficient is controlled. This is important because it ensures that spectral normalization does not arbitrarily destroy the natural-gradient direction.

Lemma 1 (Perturbation size of polar mixing). Let M∈Ra×b\mathbf{M} \in\mathbb{R}^{a \times b} be nonzero, let P=Polar⁡(M)\mathbf{P}=\operatorname{Polar}(\mathbf{M}), c=∥M∥F∥P∥Fc=\frac{\lVert\mathbf{M}\rVert_{\mathrm{F}}}{\lVert\mathbf{P}\rVert_{\mathrm{F}}} and define D=(1−α)M+αcP\mathbf{D}=(1-\alpha)\mathbf{M}+\alpha c\mathbf{P}, α∈[0,1]\alpha\in[0,1]. Then, ∥D−M∥F≤2α∥M∥F\lVert\mathbf{D}-\mathbf{M}\rVert_{\mathrm{F}}\le2\alpha\lVert\mathbf{M}\rVert_{\mathrm{F}}. If M=0\mathbf{M}=0, we use D=0\mathbf{D}=0, and the bound holds trivially.

Theorem 2 (Descent for controlled spectral-natural perturbations). Suppose Assumptions 1 and 2 hold, and suppose gt≠0g_t\ne0. Let gt=∇C(θt)g_t=\nabla C(\theta_t), ut=Gt−1gtu_t=G_t^{-1}g_t be the damped natural-gradient vector. Suppose the fixed-scale spectral-natural direction dtd_t satisfies ∥dt−ut∥2≤εt∥ut∥2\lVert d_t-u_t\rVert_2\le\varepsilon_t\lVert u_t\rVert_2. If εt<μΛ\varepsilon_t<\frac{\mu}{\Lambda} then dtd_t is a descent vector, i.e., ⟨gt,dt⟩>0\langle g_t,d_t\rangle>0. Moreover, any step size satisfying 0<γ<2⟨gt,dt⟩L∥dt∥220<\gamma<\frac{2\langle g_t,d_t\rangle}{L\lVert d_t\rVert_2^2} yields C(θt−γdt)<C(θt)C(\theta_t-\gamma d_t)<C(\theta_t).

Theorem 2 states that if the spectral-natural direction is a sufficiently small relative perturbation of QNG, then it preserves descent. The threshold depends on the damped QFIM condition ratio μ/Λ\mu/\Lambda.

Corollary 2 (Sufficient polar-mixing condition). Consider the no-momentum single-block case, and suppose the reshape map is norm-consistent as in Assumption 3. Let M\mathbf{M} be the reshape of utu_t, and D\mathbf{D} the corresponding Frobenius-preserving polar-mixed block with mixing coefficient α\alpha. If 2α<μΛ2\alpha<\frac{\mu}{\Lambda} then the fixed-scale spectral-natural direction dtd_t is guaranteed to be a descent vector.

Corollary 3 (Descent with natural-gradient momentum). Suppose Assumptions 1–3 hold. Let ut=Gt−1gtu_t=G_t^{-1}g_t be the damped natural-gradient vector, and suppose the natural-gradient momentum satisfies ∥mt−ut∥2≤ξt∥ut∥2\lVert m_t-u_t\rVert_2\le\xi_t\lVert u_t\rVert_2 for some ξt≥0\xi_t\ge0. Let dtd_t be obtained by applying Frobenius-preserving polar mixing with coefficient α\alpha to the circuit-semantic block decomposition of mtm_t. Then ∥dt−ut∥2≤(2α(1+ξt)+ξt)∥ut∥2\lVert d_t-u_t\rVert_2\le(2\alpha(1+\xi_t)+\xi_t)\lVert u_t\rVert_2. If 2α(1+ξt)+ξt<μΛ2\alpha(1+\xi_t)+\xi_t<\frac{\mu}{\Lambda} then dtd_t is a descent vector, i.e., ⟨gt,dt⟩>0\langle g_t,d_t\rangle>0.

Corollary 3 connects the analysis to the actual DAQNM update, where polar mixing is applied to natural-gradient momentum mtm_t. When ξt=0\xi_t=0, the result reduces to the no-momentum condition 2α<μ/Λ2\alpha<\mu/\Lambda. When ξt>0\xi_t>0, descent is guaranteed provided the momentum remains close enough to the current damped natural-gradient direction.

Corollary 4 (Alignment of the normalized DA-QNM direction). Suppose Assumption 2 holds and gt≠0g_t\ne0. Let ut=Gt−1gtu_t=G_t^{-1}g_t, and suppose the fixed-scale spectral-natural direction satisfies ∥dt−ut∥2≤ζt∥ut∥2\lVert d_t-u_t\rVert_2\le\zeta_t\lVert u_t\rVert_2 for some ζt<μ/Λ\zeta_t<\mu/\Lambda. Let qt=dt∥dt∥2+ϵq_t=\frac{d_t}{\lVert d_t\rVert_2+\epsilon} where ϵ≤τt∥ut∥2\epsilon\le\tau_t\lVert u_t\rVert_2 for some τt≥0\tau_t\ge0. Then ⟨gt,qt⟩≥χt∥gt∥2\langle g_t,q_t\rangle\ge\chi_t\lVert g_t\rVert_2, χt=μ−ζtΛΛ(1+ζt+τt)>0\chi_t=\frac{\mu-\zeta_t\Lambda}{\Lambda(1+\zeta_t+\tau_t)}>0. In particular, under Corollary 3, one may take ζt=2α(1+ξt)+ξt\zeta_t=2\alpha(1+\xi_t)+\xi_t.

Corollary 4 closes the alignment loop for the normalized algorithm. If momentum and polar mixing keep dtd_t close enough to the damped direction utu_t, then the normalized direction qtq_t satisfies Assumption 4 with an explicit positive constant. Thus, the descent and stationarity results below apply to the actual normalized DAQNM update generated by Algorithm 1.

We next analyze the normalized distance-aware update θt+1\theta_{t+1}. Since qt=dt∥dt∥2+ϵq_t=\frac{d_t}{\lVert d_t\rVert_2+\epsilon} we have ∥qt∥2≤1\lVert q_t\rVert_2\le1.

Theorem 3 (Descent of normalized distance-aware QNM). Suppose Assumptions 1 and 4 hold. Let qtq_t be a normalized DA-QNM direction. Then C(θt+1)≤C(θt)−rtχt∥gt∥2+L2rt2C(\theta_{t+1})\le C(\theta_t)-r_t\chi_t\lVert g_t\rVert_2+\frac{L}{2}r_t^2. In particular, the update is descending whenever 0<rt<2χt∥gt∥2L0<r_t<\frac{2\chi_t\lVert g_t\rVert_2}{L}.

Theorem 3 explains why normalized optimizers require explicit radius control: once the direction is normalized, the displacement length is governed by rtr_t, not by the original natural-gradient magnitude. It gives a one-step descent condition for an abstract normalized DA-QNM update. We now connect this condition to the specific radius generated by Algorithm 1. The next result shows that, under uniform alignment, bounded gradients, and a momentum-gradient comparability condition, the generated radius is both small enough for descent and large enough to yield an O(1/T)O(1/T) stationarity bound.

Theorem 4 (Stationarity of normalized DA-QNM with generated radius). Suppose Assumptions 1, 4, and 5 hold. Assume there exist constants χ>0\chi>0, Gmax⁡<∞G_{\max}<\infty, and 0<αm≤bm<∞0<\alpha_m\le b_m<\infty such that, for all tt, ⟨gt,qt⟩≥χ∥gt∥2\langle g_t,q_t\rangle\ge\chi\lVert g_t\rVert_2, ∥gt∥2≤Gmax⁡\lVert g_t\rVert_2\le G_{\max} and αm∥gt∥2≤∥mt∥2≤bm∥gt∥2\alpha_m\lVert g_t\rVert_2\le\lVert m_t\rVert_2\le b_m\lVert g_t\rVert_2. Let η>0\eta>0, Δ0>0\Delta_0>0, ρ≥0\rho\ge0, and Rmax⁡>0R_{\max}>0, and let DA-QNM use the generated radius rtr_t in (11). If ηbm≤χL\eta b_m\le\frac{\chi}{L}, Δ0≥ηamGmax⁡\Delta_0\ge\eta a_mG_{\max}, and Rmax⁡≥ηamGmax⁡R_{\max}\ge\eta a_mG_{\max} then ηam∥gt∥2≤rt≤ηbm∥gt∥2≤χL∥gt∥2\eta a_m\lVert g_t\rVert_2\le r_t\le\eta b_m\lVert g_t\rVert_2\le\frac{\chi}{L}\lVert g_t\rVert_2. Consequently, the normalized update satisfies C(θt+1)≤C(θt)−χηam2∥gt∥22C(\theta_{t+1})\le C(\theta_t)-\frac{\chi\eta a_m}{2}\lVert g_t\rVert_2^2. Therefore, for every T≥1T\ge1,

1T∑t=0T−1∥∇C(θt)∥22≤2(C(θ0)−C⋆)χηamT.(16)\frac{1}{T}\sum_{t=0}^{T-1}\lVert\nabla C(\theta_t)\rVert_2^2\le\frac{2(C(\theta_0)-C_\star)}{\chi\eta a_m T}. \tag*{(16)}

Theorem 4 is the main stationarity guarantee for the actual normalized DA-QNM update. It connects the algorithmic radius rule directly to descent and stationarity, whereas the fixed-scale stationarity result in Appendix § 8.16 applies to the auxiliary update θt+1=θt−γdt\theta_{t+1}=\theta_t-\gamma d_t. Finally, we quantify robustness to gradient and QFIM estimation error.

Theorem 5 (Natural-gradient perturbation under QFIM and gradient error). Let G^=F^+λI\widehat{G}=\widehat{F}+\lambda I be a symmetric estimate of G=F+λI⪰μIG=F+\lambda I\succeq\mu I. Suppose ∥g^−g∥2≤δg\lVert\widehat{g}-g\rVert_2\le\delta_g, ∥G^−G∥2≤δF<μ\lVert\widehat{G}-G\rVert_2\le\delta_F<\mu. Let u=G−1gu=G^{-1}g, u^=G^−1g^\widehat{u}=\widehat{G}^{-1}\widehat{g}. Then, ∥u^−u∥2≤δgμ−δF+δFμ(μ−δF)∥g∥2\lVert\widehat{u}-u\rVert_2\le\frac{\delta_g}{\mu-\delta_F}+\frac{\delta_F}{\mu(\mu-\delta_F)}\lVert g\rVert_2.

Theorem 5 motivates damping and structured QFIM approximations. Without damping, small eigenvalues of GG can amplify gradient and metric-estimation error. With damping, inverse sensitivity is controlled by μ\mu, and hence at least by the damping level λ\lambda.

Overall, the analysis justifies DA-QNM as a geometry-matched optimizer: QFIM preconditioning calibrates directions to local quantum-state geometry, polar mixing performs spectral direction selection in circuit-semantic blocks, and distance-aware scaling controls the displacement after normalization.

Experiments

We evaluate QNM on VQE, weighted QAOA-MaxCut, and supervised quantum classification through component ablations that isolate Euclidean adaptation, matrix spectral normalization, quantum Fisher/QGT preconditioning (including empirical supervised geometry), circuit-semantic block choice, and distance-aware radius control. We compare Adam, Muon, distance-aware Muon (DA-Muon), QNG, and QNM variants as appropriate. Muon isolates spectral normalization without quantum geometry, DA-Muon additionally tests radius control, QNG isolates quantum-geometric preconditioning, and QNM tests spectral normalization after such preconditioning.

For VQE, we report best energy error E(θ)−E⋆E(\theta)-E_\star. For weighted QAOA-MaxCut, with achieved expected cut VmV_m and exact optimum C⋆C_\star, we report approximation ratio Vm/C⋆V_m/C_\star, optimality gap Δm=C⋆−Vm\Delta_m=C_\star-V_m, and relative gap reduction ρm←b=(Δb−Δm)/Δb\rho_{m\leftarrow b}=(\Delta_b-\Delta_m)/\Delta_b. For supervised classification, we report mean test accuracy with standard errors. Finite-shot results are deferred to Appendix § 9.§ 5 as empirical robustness evidence rather than a probabilistic shot-noise guarantee.

To test the predicted spectral mechanism, we track the active condition number κact(M)=σ1(M)/σr(M)\kappa_{\mathrm{act}}(M)=\sigma_1(M)/\sigma_r(M) and effective rank reff(M)=exp⁡(−∑ipilog⁡pi)r_{\mathrm{eff}}(M)=\exp(-\sum_i p_i\log p_i), where pi=σi/∑jσjp_i=\sigma_i/\sum_j\sigma_j, of structured momentum blocks. These diagnostics measure whether polar mixing reduces active spectral imbalance by lifting weak singular directions, as predicted by the spectral-floor analysis.

VQE: Layer-Wise Spectral Structure

We consider VQE for a periodic transverse-field Ising Hamiltonian using a four-layer hardware-efficient ansatz. Because the ansatz reaches near-exact energies, the best error E(θ)−E⋆E(\theta)-E_\star, with E⋆E_\star obtained by classical diagonalization, provides a controlled optimizer diagnostic. As Table 1 shows, QNM improves the best energy error on all three seeds and reduces the mean error by 80.3%. QNM-layer is selected on all three seeds, supporting the role of circuit-semantic block choice: layer-wise reshaping organizes parameters across qubits and rotation axes according to the ansatz structure, allowing polar mixing to regularize the spectrum of QFIM-preconditioned momentum in an aligned block geometry.

SeedAdam errorBest QNM errorVariant
04.832×10−74.832 \times10^{-7}1.529×10−71.529 \times10^{-7}QNM-layer
15.142×10−75.142 \times10^{-7}1.509×10−71.509 \times10^{-7}QNM-layer
21.218×10−61.218 \times10^{-6}1.334×10−71.334 \times10^{-7}QNM-layer
Mean7.385×10−77.385 \times10^{-7}1.457×10−71.457 \times10^{-7}–

Table 1. Four-layer TFIM VQE. “Best QNM” selects among layer-, gate-, and global-block variants; QNM-layer is selected for all seeds.

Weighted QAOA: Gate-Structured QNM

We evaluate generalized QAOA on weighted MaxCut with n=10n=10 nodes, ∣E∣=23|E|=23 edges, depth p=3p=3, and exact optimum C⋆=17.818521C_\star=17.818521, computed by exhaustive enumeration. The cost and mixer parameters form Γ∈R3×23\Gamma\in\mathbb{R}^{3\times23} and B∈R3×10B \in\mathbb{R}^{3\times10}, indexed by depth–edge and depth–node, respectively. Heterogeneous edge weights directly affect the cost generators and can induce anisotropy across the edge-indexed Γ\Gamma update, making this a natural test of semantic block choice. We use this instance as a controlled component diagnostic rather than a broad QAOA benchmark.

Table 2 isolates the three components. QNG reduces Adam’s optimality gap by 43.0%, showing the benefit of quantum geometry. QNM-gate increases this reduction to 96.9%, whereas QNM-layer/global remain near QNG, showing that spectral normalization depends critically on circuit-semantic block choice. DA-QNM-gate further reduces the QNM-gate gap by 5.7%, while DA-Muon performs worse than Adam; neither Euclidean spectral normalization nor distance-aware scaling alone explains the gain. The dominant improvement therefore comes from gate-aligned spectral normalization after QFIM preconditioning, with radius control providing a smaller additional benefit.

MethodVmV_mVm/C⋆V_m/C_\starΔm\Delta_m
Adam17.3559840.9740420.4625
Muon17.0942120.9593510.7243
DA-Muon17.0980260.9595650.7205
QNG17.5551050.9852170.2634
QNM-layer17.5353280.9841070.2832
QNM-global17.5478300.9848080.2707
QNM-gate17.8040150.9991860.01451
DA-QNM-gate17.8048400.9992320.01368

Table 2. Weighted generalized QAOA-MaxCut results.

Figure 1 summarizes the ablation: Muon/DA-Muon worsen the Adam gap, QNG improves it, and the largest gains occur only when QFIM preconditioning is followed by gate-structured spectral normalization.

Optimality-gap reduction relative to Adam on weighted QAOA-MaxCut

Figure 1. Optimality-gap reduction relative to Adam on weighted QAOA-MaxCut. QNM-gate and DA-QNM-gate achieve 96.9% and 97.0% reductions, respectively.

Mechanism Diagnostics: Spectral Regularization

We test whether QNM exhibits the mechanism predicted by the spectral analysis. After QFIM preconditioning, a structured natural-gradient momentum block can remain spectrally anisotropic; polar mixing should lift weak active singular directions, reduce spectral imbalance, and increase effective rank while preserving the active singular subspaces. Figure 2 shows an 81.7% reduction in active condition number and a 27.9% increase in effective rank, consistent with the spectral-floor mechanism in § 4. Thus, the QAOA gain is accompanied by the predicted reduction in spectral concentration of the QFIM-preconditioned structured update, complementing the component and block-structure ablations in Table 2.

Polar mixing reduces the active condition number and increases effective rank

Figure 2. Polar mixing reduces the active condition number from 10.13 to 1.85 and increases effective rank from 2.22 to 2.84.

Supervised Quantum Classification

We test whether QNM transfers to supervised quantum learning using two binary Iris tasks: setosa versus rest and the harder versicolor versus virginica. We use a two-qubit data-reuploading classifier with a 2×2×32 \times2 \times3 parameter tensor organized by layer, qubit, and rotation axis, together with an empirical supervised QGT/QFIM. Muon provides the Euclidean spectral baseline; Diag-QNM uses diagonal quantum-geometric preconditioning before spectral normalization, while DA-QNM additionally applies distance-aware radius control. Data preprocessing, circuit details, and geometry construction are given in Appendix § 9.§ 4.

Table 3 shows that quantum-geometric QNM variants improve mean accuracy over Adam on both tasks. Diag-QNM performs best on setosa versus rest (0.9481 versus 0.9111), while DA-QNM performs best on the harder versicolor-versus-virginica task (0.8762 versus 0.8222), providing a controlled transfer test of the spectral-natural principle beyond VQE and QAOA. We do not interpret this small Iris study as a broad supervised-QML benchmark. Finite-shot results are reported in Appendix § 9.§ 5 as empirical robustness evidence, not a probabilistic shot-noise guarantee.

MethodSetosa vs. restVersicolor vs. virginica
Adam0.9111 ± 0.01280.8222 ± 0.0222
Muon0.9407 ± 0.01480.8333 ± 0.0385
Diag-QNM0.9481 ± 0.01960.8556 ± 0.0401
DA-QNM0.9407 ± 0.01480.8762 ± 0.0222

Table 3. Iris classification after 120 steps: mean test accuracy ± standard error over three train–test splits.

Conclusion

We introduced Quantum Natural Muon (QNM), a structured spectral natural-gradient framework that applies circuit-semantic polar mixing after quantum Fisher calibration. We showed that QFIM preconditioning and polar normalization do not commute, and that Frobenius-matched polar mixing raises weak active singular directions and improves active spectral conditioning while remaining a controlled perturbation of QNG. For the distance-aware realization, we established descent and nonconvex stationarity guarantees. Across VQE, weighted QAOA, and supervised quantum classification, the strongest gains occur when spectral blocks align with circuit structure, supporting the joint use of quantum-state and circuit-structured update geometry. Future work will study scalable QFIM approximations, finite-shot theory, and larger hardware-noisy settings.

Additional Details of Preliminaries

This appendix provides a detail of the two geometric structures that motivate DA-QNM. The first is the information geometry of the quantum state manifold, captured locally by QFIM. The second is the matrix or tensor structure induced by the circuit parameterization, such as layer-by-qubit blocks in hardware-efficient VQE ansatzes or depth-by-edge blocks in generalized QAOA. Existing optimizers typically exploit one of these structures but not both.

Parameterized Quantum Models

Let U(θ)U(\theta) be a parameterized quantum circuit acting on nn qubits, with parameter vector θ∈Rd\theta\in\mathbb{R}^{d}. For a fixed initial state ∣0⟩⊗n\lvert0\rangle^{\otimes n}, the circuit prepares

∣ψ(θ)⟩=U(θ)∣0⟩⊗n.(17)\lvert\psi(\theta)\rangle= U(\theta)\lvert0\rangle^{\otimes n}. \tag*{(17)}

For Hamiltonian objectives, such as those arising in variational quantum eigensolvers, the loss is

C(θ)=⟨ψ(θ)∣H∣ψ(θ)⟩,(18)C(\theta) = \langle\psi(\theta)\lvert\mathbf{H} \rvert\psi(\theta)\rangle, \tag*{(18)}

where H\mathbf{H} is a Hermitian problem Hamiltonian [16, 4]. In QAOA, H\mathbf{H} is typically a cost Hamiltonian encoding a combinatorial objective [7, 8]. In quantum learning, an input xx is encoded through an input-dependent circuit U(x,θ)U(x,\theta), measured to produce a feature vector z(x,θ)z(x,\theta), and passed to a classical readout fθ,w(x)=σTz(x,θ)+bf_{\theta,w}(x) = \sigma^{\mathsf{T}}z(x,\theta) + b with empirical risk

L(θ,w)=1N∑i=1Nℓ(fθ,w(xi),yi).(19)\mathcal{L}(\theta,w) = \frac{1}{N}\sum_{i=1}^{N}\ell(f_{\theta,w}(x_i),y_i). \tag*{(19)}

This setting includes variational quantum classifiers, quantum feature-map models, and data-reuploading classifiers [18, 9, 2, 15].

Although these objectives differ, they share a structural property: the trainable quantum parameters are not arbitrary unstructured coordinates. A hardware-efficient ansatz may have tensor-shaped parameters Θ∈RL×n×r\Theta\in\mathbb{R}^{L \times n \times r} where LL is the number of layers, nn is the number of qubits, and rr is the number of trainable rotation axes per qubit [13, 2]. Generalized QAOA for a graph G=(V,E)G=(V,E) may use separate cost and mixer parameter matrices Γ∈Rp×∣E∣\Gamma\in\mathbb{R}^{p \times\lvert E\rvert}, B∈Rp×∣V∣\mathbf{B} \in\mathbb{R}^{p \times\lvert V\rvert} where pp is the QAOA depth [7, 8]. Data-reuploading classifiers similarly organize parameters by reuploading layer, qubit, and rotation family [15].

We formalize this structure through circuit-semantic reshape maps. Let S={1,…,S}\mathcal{S}=\{1,\ldots,S\} index a collection of parameter blocks. For each block s∈Ss\in\mathcal{S}, let Is⊆{1,…,d}I_s\subseteq\{1,\ldots,d\} denote the corresponding parameter indices, and let θ(s)=θIs∈Rds\theta^{(s)}=\theta_{I_s}\in\mathbb{R}^{d_s} be the associated parameter subvector. We define a reshape operator

Rs:Rds→Ras×bs,M(s)=Rs(θ(s)).(20)\mathcal{R}_s:\mathbb{R}^{d_s}\to\mathbb{R}^{a_s\times b_s},\qquad\mathbf{M}^{(s)}=\mathcal{R}_s(\theta^{(s)}). \tag*{(20)}

Examples include layer-wise, gate-family, QAOA cost-angle, QAOA mixer-angle, and data-reuploading reshape maps. These maps encode the circuit-level modeling semantics. A central idea of this paper is that optimizer design should exploit this structure rather than treating θ\theta as a featureless vector.

Quantum Natural Gradient

The Euclidean norm of a parameter perturbation does not generally reflect the induced change in the quantum state. Natural-gradient methods address this mismatch by replacing Euclidean geometry with an information-geometric metric [1]. In variational quantum models, the corresponding metric is given by the quantum geometric tensor, whose real part defines the QFIM for pure states [17, 19].

For pure states, define QFIM as

[F(θ)]ij=Re⁡[⟨∂iψ(θ)∣∂jψ(θ)⟩−⟨∂iψ(θ)∣ψ(θ)⟩⟨ψ(θ)∣∂jψ(θ)⟩].(21)[\mathbf{F}(\theta)]_{ij}=\operatorname{Re}\left[\langle\partial_i\psi(\theta)|\partial_j\psi(\theta)\rangle-\langle\partial_i\psi(\theta)|\psi(\theta)\rangle\langle\psi(\theta)|\partial_j\psi(\theta)\rangle\right]. \tag*{(21)}

With this convention, the local infidelity satisfies

1−∣⟨ψ(θ)∣ψ(θ+δ)⟩∣2=δTF(θ)δ+o(∥δ∥22),(22)1-\left|\langle\psi(\theta)|\psi(\theta+\delta)\rangle\right|^2=\delta^{\mathsf{T}}\mathbf{F}(\theta)\delta+o(\lVert\delta\rVert_2^2), \tag*{(22)}

up to convention-dependent constant factors used in the literature. Thus, F(θ)\mathbf{F}(\theta) measures the local sensitivity of the prepared quantum state to parameter perturbations. Let gt=∇C(θt)g_t=\nabla C(\theta_t) denote the Euclidean gradient of the objective with respect to the quantum parameters, and let Ft=F(θt)\mathbf{F}_t=\mathbf{F}(\theta_t) be the QFIM at the current iterate. Given damping λ>0\lambda>0, define Gt=Ft+λI\mathbf{G}_t=\mathbf{F}_t+\lambda\mathbf{I}. The damped natural-gradient vector is

ut=Gt−1gt.(23)u_t=\mathbf{G}_t^{-1}g_t. \tag*{(23)}

Equivalently, the corresponding descent displacement is the solution of the local quadratic model

dtQNG=arg min⁡d∈Rd{⟨gt,d⟩+12dTGtd}=−Gt−1gt.(24)d_t^{\mathrm{QNG}}=\underset{d\in\mathbb{R}^d}{\operatorname{arg\,min}}\left\{\langle g_t,d\rangle+\frac{1}{2}d^{\mathsf{T}}\mathbf{G}_t d\right\}=-\mathbf{G}_t^{-1}g_t. \tag*{(24)}

Therefore, when the update is written as

θt+1=θt−ηut,(25)\theta_{t+1}=\theta_t-\eta u_t, \tag*{(25)}

utu_t is the natural-gradient vector, while −ut-u_t is the corresponding descent displacement.

QNG therefore solves a local state-metric problem: it calibrates the update to the geometry of the quantum state manifold [19]. However, (23) still produces an unstructured vector. It does not specify how update energy should be distributed across circuit-induced matrix blocks Rs(ut(s))R_s(u_t^{(s)}). Thus, even after QFIM preconditioning, a reshaped natural-gradient block may remain spectrally anisotropic and dominated by a few structured directions. This motivates the spectral block normalization introduced in § 3.

Muon and Normalized Matrix Updates

Muon-style optimizers address a different geometry: the spectral geometry of matrix-valued momentum updates [12, 3]. Let M∈Ra×b\mathbf{M}\in\mathbb{R}^{a\times b} be a matrix momentum block with compact singular value decomposition M=UΣVT\mathbf{M}=\mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\mathsf{T}}, rank⁡(M)=r\operatorname{rank}(\mathbf{M})=r. The polar factor is

Polar⁡(M)=UVT.(26)\operatorname{Polar}(\mathbf{M})=\mathbf{U}\mathbf{V}^{\mathsf{T}}. \tag*{(26)}

It preserves the active left and right singular subspaces of M\mathbf{M}, while setting all nonzero singular values to one:

σ1(Polar⁡(M))=⋯=σr(Polar⁡(M))=1.(27)\sigma_1(\operatorname{Polar}(\mathbf{M}))=\cdots=\sigma_r(\operatorname{Polar}(\mathbf{M}))=1. \tag*{(27)}

Thus, polar normalization controls matrix-level spectral anisotropy. If M\mathbf{M} is dominated by a few singular values, Polar⁡(M)\operatorname{Polar}(\mathbf{M}) increases the relative contribution of weaker active directions while preserving the active singular subspaces. In practice, Muon approximates this polar factor using a small number of Newton–Schulz iterations, a standard approach for computing matrix polar factors [10, 11, 12].

The polar factor also has a local optimization interpretation. It is an optimizer of the spectral-norm constrained linear problem

Polar⁡(M)∈arg max⁡∥D∥2≤1⟨M,D⟩F,(28)\operatorname{Polar}(\mathbf{M})\in\underset{\lVert\mathbf{D}\rVert_2\leq1}{\operatorname{arg\,max}}\langle\mathbf{M},\mathbf{D}\rangle_{\mathrm{F}}, \tag*{(28)}

where ⟨A,B⟩F=tr⁡(A⊤B)\langle\mathbf{A}, \mathbf{B} \rangle_{\mathrm{F}} = \operatorname{tr}(\mathbf{A}^{\top}\mathbf{B}). The optimal value is ∥M∥∗\lVert\mathbf{M}\rVert_{*}, the nuclear norm of M\mathbf{M}, by duality between the spectral and nuclear norms. Therefore, polar normalization is not merely a heuristic rescaling: it selects a steepest matrix direction under a spectral-norm trust region.

However, normalized matrix optimizers introduce a separate scale-selection problem. Once the update direction is normalized, part of the original magnitude information is removed, and the scalar step radius becomes a central algorithmic choice. Recent work on Distance-Aware Muon studies this direction–scale separation and proposes adaptive radius control for normalized optimization [5]. This issue is especially relevant in variational quantum circuits, where large normalized steps can overshoot near sharp optima and where gradient or QFIM estimates may be noisy. These observations motivate the distance-aware radius used in DA-QNM.

The Missing Natural Direction with Structured Spectral Control Geometry

The preceding discussion reveals a gap. QNG corrects the quantum-state geometry but does not control the active singular spectrum of circuit-structured update blocks. Muon controls the active singular spectrum of matrix updates but ignores the quantum-state geometry. Distance-aware Muon improves scalar radius selection for normalized updates but still operates in a classical matrix-update geometry.

This paper studies the optimizer design problem that arises when both geometries are present. Given a gradient gt\mathbf{g}_{t}, a damped quantum metric Gt=Ft+λI\mathbf{G}_{t} = \mathbf{F}_{t} + \lambda\mathbf{I}, and circuit-semantic reshape maps {Rs}s∈S\{\mathcal{R}_{s}\}_{s \in\mathcal{S}}, we seek an update satisfying three requirements:

  • Quantum-state calibration: the update should first correct the Euclidean gradient using the local QFIM geometry.

  • Structured spectral balancing: after quantum preconditioning, the update should reduce spectral anisotropy in semantically meaningful matrix blocks.

  • Distance-aware displacement: after normalization, the scalar step radius should be controlled explicitly rather than inherited implicitly from a normalized direction.

This motivates the DA-QNM pipeline developed in § 3. The key design choice is the order rather than applying spectral normalization directly to the raw Euclidean gradient. In general, QFIM preconditioning and blockwise polar normalization do not commute. Thus, DA-QNM is not simply an order-free composition of existing optimizers; it applies spectral normalization in the natural-gradient geometry induced by the quantum model and then controls the resulting normalized displacement through a distance-aware radius.

Additional Theory and Proofs

This appendix provides complete proofs for the theoretical statements in § 4. We use the notation of the main paper. In particular, gt=∇C(θt)\mathbf{g}_{t} = \nabla C(\boldsymbol{\theta}_{t}), Ft\mathbf{F}_{t} is the QFIM, Gt=Ft+λI\mathbf{G}_{t} = \mathbf{F}_{t} + \lambda\mathbf{I}, ut=Gt−1gt\mathbf{u}_{t} = \mathbf{G}_{t}^{-1}\mathbf{g}_{t}, and dt\mathbf{d}_{t} denotes the fixed-scale spectral-natural direction produced before normalization. The normalized direction is

qt=dt∥dt∥2+ϵ.(29)\mathbf{q}_{t} = \frac{\mathbf{d}_{t}}{\lVert\mathbf{d}_{t}\rVert_{2} + \epsilon}. \tag*{(29)}

More details of Assumptions

Remark 2. Since Ft⪰0\mathbf{F}_{t} \succeq0, damping guarantees Gt⪰λI\mathbf{G}_{t} \succeq\lambda\mathbf{I}. Thus, one may always take μ=λ\mu= \lambda as a conservative lower bound, while a larger μ\mu may be used when a sharper uniform bound is available. The upper bound is natural for finite-dimensional parameterized circuits with bounded generators on compact parameter domains. The condition number Λ/μ\Lambda/\mu controls the sensitivity of the natural-gradient direction to perturbations and determines how much spectral mixing can be tolerated while preserving descent.

Assumption 3 holds when the blocks form a disjoint partition of the coordinates and each Rs\mathcal{R}_{s} is a reshape operation. If overlapping blocks are used, the same analysis can be stated with upper and lower norm-equivalence constants, but we use the norm-preserving case for clarity.

Assumption 4 is a standard alignment condition for normalized nonconvex descent. It states that the normalized direction qt\mathbf{q}_{t} has a positive component along the Euclidean gradient gt\mathbf{g}_{t}, so that the update θt+1=θt−rtqt\boldsymbol{\theta}_{t+1} = \boldsymbol{\theta}_{t} - r_{t}\mathbf{q}_{t} is

locally descending for sufficiently small radius rtr_t. Later results provide sufficient conditions under which alignment follows from controlled perturbations of the damped natural-gradient direction.

Assumption 5 is only needed for the stationarity result. It is satisfied by Hamiltonian expectation objectives with bounded spectra, finite QAOA cost objectives, and standard nonnegative empirical losses.

Proof of Theorem 1

Proof. We prove the result in two steps. First, no feasible matrix D\mathbf{D} with ∥D∥2≤1\lVert\mathbf{D}\rVert_2 \le1 can achieve objective value larger than ∥M∥∗\lVert\mathbf{M}\rVert_*. By duality between the spectral norm and the nuclear norm,

max⁡∥D∥2≤1⟨M,D⟩F=∥M∥∗.(30)\max_{\lVert\mathbf{D}\rVert_2 \le1} \langle\mathbf{M}, \mathbf{D}\rangle_F = \lVert\mathbf{M}\rVert_*. \tag*{(30)}

It remains to show that the polar factor attains this upper bound. Let M=UΣVT\mathbf{M} = \mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^{\mathsf{T}} be the compact singular value decomposition of M\mathbf{M}, where U∈Ra×r\mathbf{U} \in\mathbb{R}^{a \times r}, V∈Rb×r\mathbf{V} \in\mathbb{R}^{b \times r}, UTU=Ir\mathbf{U}^{\mathsf{T}}\mathbf{U} = \mathbf{I}_r, VTV=Ir\mathbf{V}^{\mathsf{T}}\mathbf{V} = \mathbf{I}_r, and

Σ=diag⁡(σ1,…,σr),σi>0.(31)\boldsymbol{\Sigma} = \operatorname{diag}(\sigma_1,\ldots,\sigma_r), \qquad\sigma_i > 0. \tag*{(31)}

For any feasible D∈Ra×b\mathbf{D} \in\mathbb{R}^{a \times b} satisfying ∥D∥2≤1\lVert\mathbf{D}\rVert_2 \le1, we have

⟨M,D⟩F=Tr⁡(MTD)=Tr⁡((UΣVT)TD)(32)\langle\mathbf{M}, \mathbf{D}\rangle_F = \operatorname{Tr}(\mathbf{M}^{\mathsf{T}}\mathbf{D}) = \operatorname{Tr}\left((\mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^{\mathsf{T}})^{\mathsf{T}}\mathbf{D}\right) \tag*{(32)}
=Tr⁡(VΣUTD)(33)= \operatorname{Tr}\left(\mathbf{V}\boldsymbol{\Sigma}\mathbf{U}^{\mathsf{T}}\mathbf{D}\right) \tag*{(33)}

Using cyclic invariance of the trace,

Tr⁡(VΣUTD)=Tr⁡(ΣUTDV).(34)\operatorname{Tr}\left(\mathbf{V}\boldsymbol{\Sigma}\mathbf{U}^{\mathsf{T}}\mathbf{D}\right) = \operatorname{Tr}\left(\boldsymbol{\Sigma}\mathbf{U}^{\mathsf{T}}\mathbf{D}\mathbf{V}\right). \tag*{(34)}

Define A=UTDV∈Rr×r\mathbf{A} = \mathbf{U}^{\mathsf{T}}\mathbf{D}\mathbf{V} \in\mathbb{R}^{r \times r}. Then

⟨M,D⟩F=Tr⁡(ΣA)=∑i=1rσiAii.(35)\langle\mathbf{M}, \mathbf{D}\rangle_F = \operatorname{Tr}(\boldsymbol{\Sigma}\mathbf{A}) = \sum_{i=1}^{r}\sigma_i A_{ii}. \tag*{(35)}

Since U\mathbf{U} and V\mathbf{V} have orthonormal columns,

∥A∥2=∥UTDV∥2≤∥UT∥2∥D∥2∥V∥2≤1.(36)\lVert\mathbf{A}\rVert_2 = \lVert\mathbf{U}^{\mathsf{T}}\mathbf{D}\mathbf{V}\rVert_2 \le\lVert\mathbf{U}^{\mathsf{T}}\rVert_2\lVert\mathbf{D}\rVert_2\lVert\mathbf{V}\rVert_2 \le1. \tag*{(36)}

Hence ∣Aii∣≤∥A∥2≤1\lvert A_{ii}\rvert\le\lVert\mathbf{A}\rVert_2 \le1, and therefore Aii≤1A_{ii} \le1 for every ii. Since σi>0\sigma_i > 0,

⟨M,D⟩F=∑i=1rσiAii≤∑i=1rσi=∥M∥∗.(37)\langle\mathbf{M}, \mathbf{D}\rangle_F = \sum_{i=1}^{r}\sigma_i A_{ii} \le\sum_{i=1}^{r}\sigma_i = \lVert\mathbf{M}\rVert_*. \tag*{(37)}

Now set D∗=UVT\mathbf{D}_* = \mathbf{U}\mathbf{V}^{\mathsf{T}}. Because U\mathbf{U} and V\mathbf{V} have orthonormal columns, the nonzero singular values of UVT\mathbf{U}\mathbf{V}^{\mathsf{T}} are all equal to one. Hence ∥D∗∥2=1\lVert\mathbf{D}_*\rVert_2 = 1, so D∗\mathbf{D}_* is feasible. Evaluating the objective at D∗\mathbf{D}_*,

⟨M,D∗⟩F=Tr⁡(MTUVT)(38)\langle\mathbf{M}, \mathbf{D}_*\rangle_F = \operatorname{Tr}\left(\mathbf{M}^{\mathsf{T}}\mathbf{U}\mathbf{V}^{\mathsf{T}}\right) \tag*{(38)}
=Tr⁡((UΣVT)TUVT)(39)= \operatorname{Tr}\left((\mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^{\mathsf{T}})^{\mathsf{T}}\mathbf{U}\mathbf{V}^{\mathsf{T}}\right) \tag*{(39)}
=Tr⁡(Σ)=∑i=1rσi=∥M∥∗.(40)= \operatorname{Tr}(\boldsymbol{\Sigma}) = \sum_{i=1}^{r}\sigma_i = \lVert\mathbf{M}\rVert_*. \tag*{(40)}

Thus, D∗=UVT=Polar⁡(M)\mathbf{D}_* = \mathbf{U}\mathbf{V}^{\mathsf{T}} = \operatorname{Polar}(\mathbf{M}) attains the upper bound. Therefore,

Polar⁡(M)∈arg max⁡∥D∥2≤1 ⟨M,D⟩F.(41)\operatorname{Polar}(\mathbf{M}) \in\underset{\lVert\mathbf{D}\rVert_2 \le1}{\operatorname{arg\,max}}\ \langle\mathbf{M}, \mathbf{D}\rangle_F. \tag*{(41)}

The optimal value is ∥M∥∗\lVert\mathbf{M}\rVert_*. Finally, if M=0\mathbf{M} = 0, then ⟨M,D⟩F=0\langle\mathbf{M}, \mathbf{D}\rangle_F = 0 for every feasible D\mathbf{D}. Hence every D\mathbf{D} satisfying ∥D∥2≤1\lVert\mathbf{D}\rVert_2 \le1 is optimal, and the optimal value is 0=∥M∥∗0 = \lVert\mathbf{M}\rVert_*. This completes the proof.

Proof of Proposition 1

Consider a single block with reshape map R:Rds→Ra×b\mathcal{R}:\mathbb{R}^{d_s}\to\mathbb{R}^{a\times b}. For a positive definite metric G≻0\mathbf{G}\succ0 on this block, the induced blockwise preconditioning operator is

AG(M)=R(G−1R−1(M)),(42)\mathcal{A}_{\mathbf{G}}(\mathbf{M})=\mathcal{R}\left(\mathbf{G}^{-1}\mathcal{R}^{-1}(\mathbf{M})\right), \tag*{(42)}

where R−1\mathcal{R}^{-1} is vectorization back to the corresponding parameter coordinates. DA-QNM first applies metric preconditioning and then performs polar normalization, M↦Polar⁡(AG(M))\mathbf{M}\mapsto\operatorname{Polar}(\mathcal{A}_{\mathbf{G}}(\mathbf{M})). A Muon-then-QNG alternative would instead polar-normalize first and then apply the metric, M↦AG(Polar⁡(M))\mathbf{M}\mapsto\mathcal{A}_{\mathbf{G}}(\operatorname{Polar}(\mathbf{M})).

Proof. The proposition states that QFIM preconditioning and blockwise polar normalization do not commute in general. It suffices to give a counterexample. This simplified operator is an instance of the blockwise preconditioning map At,s\mathcal{A}_{t,s} from the main text, restricted to one block with a fixed positive definite metric G\mathbf{G}. Let a=b=2a=b=2, ds=4d_s=4, and let R\mathcal{R} be the standard matrix reshape satisfying

R−1(M)=vec⁡(M).(43)\mathcal{R}^{-1}(\mathbf{M})=\operatorname{vec}(\mathbf{M}). \tag*{(43)}

Choose the positive definite matrix A=(2001)\mathbf{A}=\begin{pmatrix}2&0\\0&1\end{pmatrix}. Define the block metric preconditioner by G−1=I2⊗A\mathbf{G}^{-1}=\mathbf{I}_2\otimes\mathbf{A} where ⊗\otimes denotes the Kronecker product. Since A≻0\mathbf{A}\succ0, we have G−1≻0\mathbf{G}^{-1}\succ0, and therefore G≻0\mathbf{G}\succ0. Under column-wise vectorization, vec⁡(AM)=(I2⊗A)vec⁡(M)\operatorname{vec}(\mathbf{A}\mathbf{M})=(\mathbf{I}_2\otimes\mathbf{A})\operatorname{vec}(\mathbf{M}). Hence, the induced blockwise operator satisfies AG(M)=AM\mathcal{A}_{\mathbf{G}}(\mathbf{M})=\mathbf{A}\mathbf{M}. Now take M=I2\mathbf{M}=\mathbf{I}_2. Since M=I2\mathbf{M}=\mathbf{I}_2, its polar factor is Polar⁡(M)=I2\operatorname{Polar}(\mathbf{M})=\mathbf{I}_2. Therefore, the Muon-then-QNG ordering gives

AG(Polar⁡(M))=AG(I2)=AI2=A.(44)\mathcal{A}_{\mathbf{G}}(\operatorname{Polar}(\mathbf{M}))=\mathcal{A}_{\mathbf{G}}(\mathbf{I}_2)=\mathbf{A}\mathbf{I}_2=\mathbf{A}. \tag*{(44)}

On the other hand, the QNG-then-polar ordering yields AG(M)=A\mathcal{A}_{\mathbf{G}}(\mathbf{M})=\mathbf{A}. Because A\mathbf{A} is symmetric positive definite, its polar factor is

Polar⁡(A)=A(ATA)−1/2=A(A2)−1/2=I2.(45)\operatorname{Polar}(\mathbf{A})=\mathbf{A}(\mathbf{A}^{\mathsf{T}}\mathbf{A})^{-1/2}=\mathbf{A}(\mathbf{A}^2)^{-1/2}=\mathbf{I}_2. \tag*{(45)}

Consequently,

Polar⁡(AG(M))=I2.(46)\operatorname{Polar}\left(\mathcal{A}_{\mathbf{G}}(\mathbf{M})\right)=\mathbf{I}_2. \tag*{(46)}

Combining (44) and (46), we obtain

Polar⁡(AG(M))=I2≠A=AG(Polar⁡(M)).(47)\operatorname{Polar}\left(\mathcal{A}_{\mathbf{G}}(\mathbf{M})\right)=\mathbf{I}_2\ne\mathbf{A}=\mathcal{A}_{\mathbf{G}}(\operatorname{Polar}(\mathbf{M})). \tag*{(47)}

Therefore, metric preconditioning and polar normalization do not commute in general.

Proof of Proposition 2

Proof. Let M=UΣVT\mathbf{M}=\mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\mathsf{T}} be the compact singular value decomposition of M\mathbf{M}. The compact polar factor is P=UVT\mathbf{P}=\mathbf{U}\mathbf{V}^{\mathsf{T}}. Then PPT=UVTVUT=UUT\mathbf{P}\mathbf{P}^{\mathsf{T}}=\mathbf{U}\mathbf{V}^{\mathsf{T}}\mathbf{V}\mathbf{U}^{\mathsf{T}}=\mathbf{U}\mathbf{U}^{\mathsf{T}} and PTP=VUTUVT=VVT\mathbf{P}^{\mathsf{T}}\mathbf{P}=\mathbf{V}\mathbf{U}^{\mathsf{T}}\mathbf{U}\mathbf{V}^{\mathsf{T}}=\mathbf{V}\mathbf{V}^{\mathsf{T}}. Both UUT\mathbf{U}\mathbf{U}^{\mathsf{T}} and VVT\mathbf{V}\mathbf{V}^{\mathsf{T}} are orthogonal projectors of rank rr. Therefore, their nonzero eigenvalues are all equal to one. Since the nonzero singular values of P\mathbf{P} are the square roots of the nonzero eigenvalues of PTP\mathbf{P}^{\mathsf{T}}\mathbf{P}, all active singular values of P\mathbf{P} are equal to one. Hence, κact(P)=σ1(P)σr(P)=1\kappa_{\mathrm{act}}(\mathbf{P})=\frac{\sigma_1(\mathbf{P})}{\sigma_r(\mathbf{P})}=1. This completes the proof.

Proof of Proposition 3

Proof. Since P\mathbf{P} has exactly rr active singular values and each is equal to one by Proposition 2,

∥P∥F2=∑i=1rσi(P)2=∑i=1r1=r.(48)\lVert\mathbf{P}\rVert_F^2=\sum_{i=1}^{r}\sigma_i(\mathbf{P})^2=\sum_{i=1}^{r}1=r. \tag*{(48)}

Therefore, ∥P∥F=r\lVert\mathbf{P}\rVert_F=\sqrt{r}. By the definition of cc,

∥cP∥F=c∥P∥F=∥M∥F∥P∥F∥P∥F=∥M∥F.(49)\lVert c\mathbf{P}\rVert_F=c\lVert\mathbf{P}\rVert_F=\frac{\lVert\mathbf{M}\rVert_F}{\lVert\mathbf{P}\rVert_F}\lVert\mathbf{P}\rVert_F=\lVert\mathbf{M}\rVert_F. \tag*{(49)}

Thus, scaling the polar factor by cc preserves the Frobenius norm of the original block.

Finally, the singular values of cPc\mathbf{P} are cc times the singular values of P\mathbf{P}. Since every active singular value of P\mathbf{P} equals one, every active singular value of cPc\mathbf{P} equals

c=∥M∥F∥P∥F=∥M∥Fr.(50)c = \frac{\lVert\mathbf{M}\rVert_F}{\lVert\mathbf{P}\rVert_F} = \frac{\lVert\mathbf{M}\rVert_F}{\sqrt{r}}. \tag*{(50)}

This proves the proposition. □\square

Proof of Proposition 4

Proof. Using the compact SVD of M\mathbf{M} and the definition of P\mathbf{P}, we obtain

D=(1−α)UΣV⊤+αcUV⊤(51)\mathbf{D} = (1-\alpha)\mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^{\top} + \alpha c\mathbf{U}\mathbf{V}^{\top} \tag*{(51)}
=U((1−α)Σ+αcIr)V⊤.(52)= \mathbf{U}\left((1-\alpha)\boldsymbol{\Sigma}+\alpha c\mathbf{I}_r\right)\mathbf{V}^{\top}. \tag*{(52)}

The middle matrix (1−α)Σ+αcIr(1-\alpha)\boldsymbol{\Sigma}+\alpha c\mathbf{I}_r is diagonal with entries (1−α)σi+αc(1-\alpha)\sigma_i+\alpha c, i=1,…,ri=1,\ldots,r. Because α∈[0,1]\alpha\in[0,1], c>0c>0, and σi>0\sigma_i>0, these entries are strictly positive. Therefore, this expression is a compact singular value decomposition of D\mathbf{D} with the same active left and right singular vectors U\mathbf{U} and V\mathbf{V}. The active singular values of D\mathbf{D} are precisely σ~i=(1−α)σi+αc\widetilde{\sigma}_i=(1-\alpha)\sigma_i+\alpha c. The spectral-floor bound follows immediately: σ~i=(1−α)σi+αc≥αc\widetilde{\sigma}_i=(1-\alpha)\sigma_i+\alpha c\ge\alpha c. This completes the proof. □\square

Proof of Corollary 1

Proof. By Proposition 4, the largest and smallest active singular values of D\mathbf{D} are

σ~1=(1−α)σ1+αc,σ~r=(1−α)σr+αc.(53)\widetilde{\sigma}_1=(1-\alpha)\sigma_1+\alpha c,\qquad\widetilde{\sigma}_r=(1-\alpha)\sigma_r+\alpha c. \tag*{(53)}

Therefore,

κact(D)=σ~1σ~r=(1−α)σ1+αc(1−α)σr+αc.(54)\kappa_{\mathrm{act}}(\mathbf{D})=\frac{\widetilde{\sigma}_1}{\widetilde{\sigma}_r}=\frac{(1-\alpha)\sigma_1+\alpha c}{(1-\alpha)\sigma_r+\alpha c}. \tag*{(54)}

To prove κact(D)≤κact(M)\kappa_{\mathrm{act}}(\mathbf{D})\le\kappa_{\mathrm{act}}(\mathbf{M}), it suffices to show

(1−α)σ1+αc(1−α)σr+αc≤σ1σr.(55)\frac{(1-\alpha)\sigma_1+\alpha c}{(1-\alpha)\sigma_r+\alpha c}\le\frac{\sigma_1}{\sigma_r}. \tag*{(55)}

Since all denominators are positive, this is equivalent

((1−α)σ1+αc)σr≤σ1((1−α)σr+αc).(56)\left((1-\alpha)\sigma_1+\alpha c\right)\sigma_r\le\sigma_1\left((1-\alpha)\sigma_r+\alpha c\right). \tag*{(56)}

Expanding both sides yields

(1−α)σ1σr+αcσr≤(1−α)σ1σr+αcσ1.(57)(1-\alpha)\sigma_1\sigma_r+\alpha c\sigma_r\le(1-\alpha)\sigma_1\sigma_r+\alpha c\sigma_1. \tag*{(57)}

Canceling the common term (1−α)σ1σr(1-\alpha)\sigma_1\sigma_r reduces the inequality to αcσr≤αcσ1\alpha c\sigma_r\le\alpha c\sigma_1. This holds because α≥0\alpha\ge0, c>0c>0, and σr≤σ1\sigma_r\le\sigma_1. If α>0\alpha>0 and σ1>σr\sigma_1>\sigma_r, then αcσr<αcσ1\alpha c\sigma_r<\alpha c\sigma_1, so the condition-number inequality is strict. This completes the proof. □\square

This result gives the structural reason polar mixing can help. It raises weak active singular directions without changing the active singular subspaces. Thus, polar mixing directly reduces spectral collapse in a structured update block. With the Frobenius-preserving choice

c=∥M∥F∥Polar⁡(M)∥F,(58)c=\frac{\lVert\mathbf{M}\rVert_F}{\lVert\operatorname{Polar}(\mathbf{M})\rVert_F}, \tag*{(58)}

this improvement is achieved while preserving the Frobenius-scale budget of the block.

Proof of Lemma 1

Proof. Assume first that M≠0M \ne0. Since D=(1−α)M+αcPD = (1-\alpha)M + \alpha cP, it follows that

∥D−M∥F=α∥cP−M∥F(59)\lVert D-M\rVert_F = \alpha\lVert cP-M\rVert_F \tag*{(59)}
≤α(∥cP∥F+∥M∥F),(60)\le\alpha\left(\lVert cP\rVert_F+\lVert M\rVert_F\right), \tag*{(60)}

where the last step uses the triangle inequality. By the Frobenius-preserving choice of cc,

∥cP∥F=c∥P∥F=∥M∥F.(61)\lVert cP\rVert_F = c\lVert P\rVert_F = \lVert M\rVert_F. \tag*{(61)}

Substituting this identity yields

∥D−M∥F≤2α∥M∥F.(62)\lVert D-M\rVert_F \le2\alpha\lVert M\rVert_F. \tag*{(62)}

If M=0M=0, then by convention D=0D=0, so

∥D−M∥F=0≤2α∥M∥F.(63)\lVert D-M\rVert_F = 0 \le2\alpha\lVert M\rVert_F. \tag*{(63)}

This completes the proof.

Proof of Theorem 2

Proof. Since ut=Gt−1gtu_t = G_t^{-1}g_t, we have gt=Gtutg_t = G_tu_t. Because gt≠0g_t \ne0 and GtG_t is positive definite, ut≠0u_t \ne0. By Assumption 2, Gt⪰μIG_t \succeq\mu I, and therefore

⟨gt,ut⟩=utTGtut≥μ∥ut∥22.(64)\langle g_t,u_t\rangle= u_t^{\mathsf{T}}G_tu_t \ge\mu\lVert u_t\rVert_2^2. \tag*{(64)}

Write dt=ut+etd_t=u_t+e_t, and et=dt−ute_t=d_t-u_t. By assumption, ∥et∥2≤εt∥ut∥2\lVert e_t\rVert_2\le\varepsilon_t\lVert u_t\rVert_2. Then

⟨gt,dt⟩=⟨gt,ut⟩+⟨gt,et⟩(65)\langle g_t,d_t\rangle= \langle g_t,u_t\rangle+\langle g_t,e_t\rangle \tag*{(65)}
≥⟨gt,ut⟩−∥gt∥2∥et∥2.(66)\ge\langle g_t,u_t\rangle-\lVert g_t\rVert_2\lVert e_t\rVert_2. \tag*{(66)}

Using (64), we obtain

⟨gt,dt⟩≥μ∥ut∥22−∥gt∥2εt∥ut∥2.(67)\langle g_t,d_t\rangle\ge\mu\lVert u_t\rVert_2^2-\lVert g_t\rVert_2\varepsilon_t\lVert u_t\rVert_2. \tag*{(67)}

Since gt=Gtutg_t=G_tu_t and Gt⪯ΛIG_t\preceq\Lambda I, we have

∥gt∥2=∥Gtut∥2≤Λ∥ut∥2.(68)\lVert g_t\rVert_2=\lVert G_tu_t\rVert_2\le\Lambda\lVert u_t\rVert_2. \tag*{(68)}

Substituting this bound into (67) gives

⟨gt,dt⟩≥(μ−εtΛ)∥ut∥22.(69)\langle g_t,d_t\rangle\ge(\mu-\varepsilon_t\Lambda)\lVert u_t\rVert_2^2. \tag*{(69)}

If εt<μ/Λ\varepsilon_t<\mu/\Lambda, then μ−εtΛ>0\mu-\varepsilon_t\Lambda>0. Since ut≠0u_t\ne0, (69) implies ⟨gt,dt⟩>0\langle g_t,d_t\rangle>0.

It remains to show descent of the objective. By LL-smoothness assumption, we have

C(θt−γdt)≤C(θt)+⟨gt,−γdt⟩+L2∥γdt∥22(70)C(\theta_t-\gamma d_t)\le C(\theta_t)+\langle g_t,-\gamma d_t\rangle+\frac{L}{2}\lVert\gamma d_t\rVert_2^2 \tag*{(70)}
=C(θt)−γ⟨gt,dt⟩+Lγ22∥dt∥22.(71)=C(\theta_t)-\gamma\langle g_t,d_t\rangle+\frac{L\gamma^2}{2}\lVert d_t\rVert_2^2. \tag*{(71)}

The right-hand side is strictly smaller than C(θt)C(\theta_t) whenever −γ⟨gt,dt⟩+Lγ22∥dt∥22<0-\gamma\langle g_t,d_t\rangle+\frac{L\gamma^2}{2}\lVert d_t\rVert_2^2<0. Since γ>0\gamma>0, this condition is equivalent to γ<2⟨gt,dt⟩L∥dt∥22\gamma<\frac{2\langle g_t,d_t\rangle}{L\lVert d_t\rVert_2^2}. Thus every step size satisfying the condition yields

C(θt−γdt)<C(θt).(72)C(\theta_t-\gamma d_t)<C(\theta_t). \tag*{(72)}

This completes the proof.

Proof of Corollary 2

Proof. In the no-momentum single-block case, the block M\mathrm{M} is obtained by reshaping utu_t, and D\mathrm{D} is the corresponding polar-mixed block. By Lemma 1,

∥D−M∥F≤2α∥M∥F.(73)\lVert\mathrm{D} - \mathrm{M} \rVert_F \le2\alpha\lVert\mathrm{M} \rVert_F. \tag*{(73)}

By norm consistency of the reshape,

∥M∥F=∥ut∥2,∥D−M∥F=∥dt−ut∥2.(74)\lVert\mathrm{M} \rVert_F = \lVert u_t \rVert_2,\qquad\lVert\mathrm{D} - \mathrm{M} \rVert_F = \lVert d_t - u_t \rVert_2. \tag*{(74)}

Therefore,

∥dt−ut∥2≤2α∥ut∥2.(75)\lVert d_t - u_t \rVert_2 \le2\alpha\lVert u_t \rVert_2. \tag*{(75)}

Thus, the relative perturbation condition in Theorem 2 holds with εt≤2α\varepsilon_t \le2\alpha. If 2α<μ/Λ2\alpha< \mu/\Lambda, then εt<μ/Λ\varepsilon_t < \mu/\Lambda. The claim follows directly from Theorem 2.

Proof of Corollary 3

Proof. By Lemma 1 applied blockwise, for each block s∈Ss \in\mathcal{S}, we have

∥Dt(s)−Mt(s)∥F≤2α∥Mt(s)∥F.(76)\lVert\mathrm{D}^{(s)}_t - \mathrm{M}^{(s)}_t \rVert_F \le2\alpha\lVert\mathrm{M}^{(s)}_t \rVert_F. \tag*{(76)}

Using the norm-consistent block decomposition in Assumption 3, we obtain

∥dt−mt∥22=∑s∈S∥Dt(s)−Mt(s)∥F2(77)\lVert d_t - m_t \rVert_2^2 = \sum_{s \in\mathcal{S}} \lVert\mathrm{D}^{(s)}_t - \mathrm{M}^{(s)}_t \rVert_F^2 \tag*{(77)}
≤4α2∑s∈S∥Mt(s)∥F2=4α2∥mt∥22.(78)\le4\alpha^2 \sum_{s \in\mathcal{S}} \lVert\mathrm{M}^{(s)}_t \rVert_F^2 = 4\alpha^2 \lVert m_t \rVert_2^2. \tag*{(78)}

Therefore, we obtain

∥dt−mt∥2≤2α∥mt∥2.(79)\lVert d_t - m_t \rVert_2 \le2\alpha\lVert m_t \rVert_2. \tag*{(79)}

By the momentum deviation assumption,

∥mt∥2≤∥ut∥2+∥mt−ut∥2≤(1+ξt)∥ut∥2.(80)\lVert m_t \rVert_2 \le\lVert u_t \rVert_2 + \lVert m_t - u_t \rVert_2 \le(1+\xi_t)\lVert u_t \rVert_2. \tag*{(80)}

Combining this bound with (79) gives

∥dt−mt∥2≤2α(1+ξt)∥ut∥2.(81)\lVert d_t - m_t \rVert_2 \le2\alpha(1+\xi_t)\lVert u_t \rVert_2. \tag*{(81)}

Finally, by the triangle inequality,

∥dt−ut∥2≤∥dt−mt∥2+∥mt−ut∥2(82)\lVert d_t - u_t \rVert_2 \le\lVert d_t - m_t \rVert_2 + \lVert m_t - u_t \rVert_2 \tag*{(82)}
≤(2α(1+ξt)+ξt)∥ut∥2.(83)\le\bigl(2\alpha(1+\xi_t)+\xi_t\bigr)\lVert u_t \rVert_2. \tag*{(83)}

If the momentum descent condition holds, then the perturbation condition in Theorem 2 is satisfied with εt=2α(1+ξt)+ξt\varepsilon_t = 2\alpha(1+\xi_t)+\xi_t. Therefore, Theorem 2 implies ⟨gt,dt⟩>0\langle g_t,d_t\rangle> 0. This completes the proof.

The conditions in Corollaries 2 and 3 are sufficient, not necessary. Corollary 2 isolates the no-momentum single-block case, while Corollary 3 connects the analysis to the actual DA-QNM update, where polar mixing is applied to natural-gradient momentum mtm_t. When ξt=0\xi_t = 0, the momentum result reduces to the no-momentum perturbation condition 2α<μ/Λ2\alpha< \mu/\Lambda. When ξt>0\xi_t > 0, descent can still be guaranteed provided the accumulated momentum remains close enough to the current damped natural-gradient direction. In practice, substantially larger α\alpha may work because the polar perturbation is often aligned with the natural-gradient block. The results nevertheless formalize that the fixed-scale QNM direction continuously interpolates from QNG to a spectrally normalized natural-gradient update while preserving descent for sufficiently small mixing and sufficiently small momentum lag.

Proof of Corollary 4

Proof. Assume

∥dt−ut∥2≤ζt∥ut∥2,ζt<μ/Λ.(84)\lVert d_t-u_t\rVert_2 \leq\zeta_t\lVert u_t\rVert_2,\qquad\zeta_t < \mu/\Lambda. \tag*{(84)}

We write

dt=ut+et,∥et∥2≤ζt∥ut∥2.(85)d_t=u_t+e_t,\qquad\lVert e_t\rVert_2\leq\zeta_t\lVert u_t\rVert_2. \tag*{(85)}

Since gt=Gtutg_t=G_tu_t, Assumption 2 implies

⟨gt,ut⟩=utTGtut≥μ∥ut∥22,∥gt∥2=∥Gtut∥2≤Λ∥ut∥2.(86)\langle g_t,u_t\rangle=u_t^{\mathsf T}G_tu_t\geq\mu\lVert u_t\rVert_2^2,\qquad\lVert g_t\rVert_2=\lVert G_tu_t\rVert_2\leq\Lambda\lVert u_t\rVert_2. \tag*{(86)}

Therefore, by Cauchy–Schwarz we can obtain

⟨gt,dt⟩=⟨gt,ut⟩+⟨gt,et⟩≥μ∥ut∥22−∥gt∥2∥et∥2≥(μ−ζtΛ)∥ut∥22.(87)\langle g_t,d_t\rangle=\langle g_t,u_t\rangle+\langle g_t,e_t\rangle\geq\mu\lVert u_t\rVert_2^2-\lVert g_t\rVert_2\lVert e_t\rVert_2\geq(\mu-\zeta_t\Lambda)\lVert u_t\rVert_2^2. \tag*{(87)}

Since ζt<μ/Λ\zeta_t<\mu/\Lambda, the coefficient μ−ζtΛ\mu-\zeta_t\Lambda is strictly positive. Hence dtd_t has positive inner product with the gradient.

Next, we control the denominator introduced by normalization. By the triangle inequality,

∥dt∥2+ϵ≤∥ut∥2+∥et∥2+ϵ≤(1+ζt+τt)∥ut∥2,(88)\lVert d_t\rVert_2+\epsilon\leq\lVert u_t\rVert_2+\lVert e_t\rVert_2+\epsilon\leq(1+\zeta_t+\tau_t)\lVert u_t\rVert_2, \tag*{(88)}

where ϵ≤τt∥ut∥2\epsilon\leq\tau_t\lVert u_t\rVert_2. Since the numerator lower bound is positive, dividing by this upper bound yields a valid lower bound:

⟨gt,qt⟩=⟨gt,dt∥dt∥2+ϵ⟩=⟨gt,dt⟩∥dt∥2+ϵ≥μ−ζtΛ1+ζt+τt∥ut∥2.(89)\langle g_t,q_t\rangle=\left\langle g_t,\frac{d_t}{\lVert d_t\rVert_2+\epsilon}\right\rangle=\frac{\langle g_t,d_t\rangle}{\lVert d_t\rVert_2+\epsilon}\geq\frac{\mu-\zeta_t\Lambda}{1+\zeta_t+\tau_t}\lVert u_t\rVert_2. \tag*{(89)}

Using ∥gt∥2≤Λ∥ut∥2\lVert g_t\rVert_2\leq\Lambda\lVert u_t\rVert_2, we obtain

⟨gt,qt⟩≥μ−ζtΛΛ(1+ζt+τt)∥gt∥2.(90)\langle g_t,q_t\rangle\geq\frac{\mu-\zeta_t\Lambda}{\Lambda(1+\zeta_t+\tau_t)}\lVert g_t\rVert_2. \tag*{(90)}

Thus Assumption 4 holds with

χt=μ−ζtΛΛ(1+ζt+τt)>0.(91)\chi_t=\frac{\mu-\zeta_t\Lambda}{\Lambda(1+\zeta_t+\tau_t)}>0. \tag*{(91)}

The constant is strictly positive because

ζt<μΛ⟹μ−ζtΛ>0(92)\zeta_t<\frac{\mu}{\Lambda}\quad\Longrightarrow\quad\mu-\zeta_t\Lambda>0 \tag*{(92)}

and the denominator Λ(1+ζt+τt)\Lambda(1+\zeta_t+\tau_t) is positive.

It remains only to connect this bound to the momentum-based DA-QNM construction. Under Corollary 3 one may take ζt=2α(1+ξt)+ξt\zeta_t=2\alpha(1+\xi_t)+\xi_t. If this quantity satisfies

2α(1+ξt)+ξt<μΛ(93)2\alpha(1+\xi_t)+\xi_t<\frac{\mu}{\Lambda} \tag*{(93)}

then the normalized DA-QNM direction satisfies the stated alignment bound. This proves the corollary.

Proof of Theorem 3

Proof. By LL-smoothness, for the update θt+1=θt−rtqt\theta_{t+1}=\theta_t-r_tq_t, with displacement −rtqt-r_tq_t,

C(θt−rtqt)≤C(θt)+⟨gt,−rtqt⟩+L2∥−rtqt∥22(94)C(\theta_t-r_tq_t)\leq C(\theta_t)+\langle g_t,-r_tq_t\rangle+\frac{L}{2}\lVert-r_tq_t\rVert_2^2 \tag*{(94)}
=C(θt)−rt⟨gt,qt⟩+L2rt2∥qt∥22.(95)=C(\theta_t)-r_t\langle g_t,q_t\rangle+\frac{L}{2}r_t^2\lVert q_t\rVert_2^2. \tag*{(95)}

Using Assumption 4, ⟨gt,qt⟩≥χt∥gt∥2\langle g_t,q_t\rangle\ge\chi_t\lVert g_t\rVert_2 and using ∥qt∥2≤1\lVert q_t\rVert_2 \le1, we obtain

C(θt+1)≤C(θt)−rtχt∥gt∥2+L2rt2.(96)C(\theta_{t+1}) \le C(\theta_t) - r_t\chi_t\lVert g_t\rVert_2 + \frac{L}{2}r_t^2. \tag*{(96)}

The right-hand side is strictly smaller than C(θt)C(\theta_t) whenever

−rtχt∥gt∥2+L2rt2<0.(97)-r_t\chi_t\lVert g_t\rVert_2 + \frac{L}{2}r_t^2 < 0. \tag*{(97)}

Since rt>0r_t > 0, this condition is equivalent to

rt<2χt∥gt∥2L.(98)r_t < \frac{2\chi_t\lVert g_t\rVert_2}{L}. \tag*{(98)}

This proves the theorem.

Remark 3. Theorem 3 explains why normalized optimizers require explicit radius control. Once the direction has been normalized, the displacement length is governed by rtr_t, not by the original magnitude of the natural-gradient vector. The distance-aware radius in DA-QNM is therefore a practical mechanism for controlling the scalar displacement along a structured spectral-natural direction.

Proof of Theorem 4

Proof. First, we prove that the radius generated by DA-QNM is gradient-scaled. The generated radius is

rt=min⁡{η∥mt∥2,Δ0+ρmax⁡0≤k≤t∥θk−θ0∥2,Rmax⁡}.(99)r_t = \min\left\{\eta\lVert m_t\rVert_2,\Delta_0+\rho\max_{0\le k\le t}\lVert\theta_k-\theta_0\rVert_2,R_{\max}\right\}. \tag*{(99)}

By the momentum-gradient comparability assumption,

am∥gt∥2≤∥mt∥2≤bm∥gt∥2.(100)a_m\lVert g_t\rVert_2 \le\lVert m_t\rVert_2 \le b_m\lVert g_t\rVert_2. \tag*{(100)}

Therefore, we obtain

η∥mt∥2≤ηbm∥gt∥2.(101)\eta\lVert m_t\rVert_2 \le\eta b_m\lVert g_t\rVert_2. \tag*{(101)}

Since rtr_t is the minimum of three terms,

rt≤η∥mt∥2≤ηbm∥gt∥2.(102)r_t \le\eta\lVert m_t\rVert_2 \le\eta b_m\lVert g_t\rVert_2. \tag*{(102)}

By the assumed condition ηbm≤χ/L\eta b_m \le\chi/L,

rt≤χL∥gt∥2.(103)r_t \le\frac{\chi}{L}\lVert g_t\rVert_2. \tag*{(103)}

Next, we prove the lower bound. The first argument of the minimum satisfies

η∥mt∥2≥ηam∥gt∥2.(104)\eta\lVert m_t\rVert_2 \ge\eta a_m\lVert g_t\rVert_2. \tag*{(104)}

Since ∥gt∥2≤Gmax⁡\lVert g_t\rVert_2 \le G_{\max} and Δ0≥ηamGmax⁡\Delta_0 \ge\eta a_mG_{\max}, we have

Δ0+ρmax⁡0≤k≤t∥θk−θ0∥2≥Δ0≥ηamGmax⁡≥ηam∥gt∥2,(105)\Delta_0+\rho\max_{0\le k\le t}\lVert\theta_k-\theta_0\rVert_2 \ge\Delta_0 \ge\eta a_mG_{\max} \ge\eta a_m\lVert g_t\rVert_2, \tag*{(105)}

where we use ρ≥0\rho\ge0. Similarly,

Rmax⁡≥ηamGmax⁡≥ηam∥gt∥2.(106)R_{\max} \ge\eta a_mG_{\max} \ge\eta a_m\lVert g_t\rVert_2. \tag*{(106)}

All three terms inside the minimum are therefore at least ηam∥gt∥2\eta a_m\lVert g_t\rVert_2. Thus, we obtain

rt≥ηam∥gt∥2.(107)r_t \ge\eta a_m\lVert g_t\rVert_2. \tag*{(107)}

Combining the lower and upper bounds gives

ηam∥gt∥2≤rt≤ηbm∥gt∥2≤χL∥gt∥2.(108)\eta a_m\lVert g_t\rVert_2 \le r_t \le\eta b_m\lVert g_t\rVert_2 \le\frac{\chi}{L}\lVert g_t\rVert_2. \tag*{(108)}

By Theorem 3 with uniform alignment constant χ\chi,

C(θt+1)≤C(θt)−rtχ∥gt∥2+L2rt2.(109)C(\theta_{t+1}) \le C(\theta_t) - r_t\chi\lVert g_t\rVert_2 + \frac{L}{2}r_t^2. \tag*{(109)}

Using rt≤(χ/L)∥gt∥2r_t \le(\chi/L)\lVert g_t\rVert_2, we have

L2rt2≤12rtχ∥gt∥2.(110)\frac{L}{2}r_t^2 \le\frac{1}{2}r_t\chi\lVert g_t\rVert_2. \tag*{(110)}

Therefore,

C(θt+1)≤C(θt)−12rtχ∥gt∥2.(111)C(\theta_{t+1}) \le C(\theta_t) - \frac{1}{2}r_t\chi\lVert g_t\rVert_2. \tag*{(111)}

Using rt≥ηam∥gt∥2r_t \ge\eta a_m\lVert g_t\rVert_2, we get

C(θt+1)≤C(θt)−χηam2∥gt∥22.(112)C(\theta_{t+1}) \le C(\theta_t) - \frac{\chi\eta a_m}{2}\lVert g_t\rVert_2^2. \tag*{(112)}

Summing from t=0t=0 to T−1T-1 yields

χηam2∑t=0T−1∥gt∥22≤C(θ0)−C(θT).(113)\frac{\chi\eta a_m}{2}\sum_{t=0}^{T-1}\lVert g_t\rVert_2^2 \le C(\theta_0)-C(\theta_T). \tag*{(113)}

Since C(θT)≥C⋆C(\theta_T) \ge C_\star, we have

∑t=0T−1∥gt∥22≤2(C(θ0)−C⋆)χηam.(114)\sum_{t=0}^{T-1}\lVert g_t\rVert_2^2 \le\frac{2(C(\theta_0)-C_\star)}{\chi\eta a_m}. \tag*{(114)}

Dividing by TT gives

1T∑t=0T−1∥∇C(θt)∥22≤2(C(θ0)−C⋆)χηamT.(115)\frac{1}{T}\sum_{t=0}^{T-1}\lVert\nabla C(\theta_t)\rVert_2^2 \le\frac{2(C(\theta_0)-C_\star)}{\chi\eta a_m T}. \tag*{(115)}

This proves the theorem.

Theorem 4 closes the gap between the descent condition for an abstract normalized step and the radius actually produced by Algorithm 1. The upper momentum-gradient comparability condition ensures that the first term η∥mt∥2\eta\lVert m_t\rVert_2 makes the generated radius small enough for descent, while the lower comparability condition and the lower bounds on Δ0\Delta_0 and Rmax⁡R_{\max} prevent the distance-aware and cap terms from making the radius too small to yield a stationarity rate. Thus, the generated DA-QNM radius is both safe and large enough to produce an O(1/T)O(1/T) stationarity bound.

Proof of Theorem 5

Proof. Let G=F+λI⪰μIG=F+\lambda I\succeq\mu I, G^=F^+λI\widehat{G}=\widehat{F}+\lambda I and suppose ∥G^−G∥2≤δF<μ\lVert\widehat{G}-G\rVert_2\le\delta_F<\mu. By Weyl’s inequality, G^⪰(μ−δF)I\widehat{G}\succeq(\mu-\delta_F)I and hence

∥G^−1∥2≤1μ−δF,∥G−1∥2≤1μ.(116)\lVert\widehat{G}^{-1}\rVert_2\le\frac{1}{\mu-\delta_F},\qquad\lVert G^{-1}\rVert_2\le\frac{1}{\mu}. \tag*{(116)}

Let u=G−1gu=G^{-1}g, u^=G^−1g^\widehat{u}=\widehat{G}^{-1}\widehat{g}. Then

u^−u=G^−1(g^−g)+(G^−1−G−1)g.(117)\widehat{u}-u=\widehat{G}^{-1}(\widehat{g}-g)+\left(\widehat{G}^{-1}-G^{-1}\right)g. \tag*{(117)}

The first term satisfies

∥G^−1(g^−g)∥2≤δgμ−δF.(118)\left\|\widehat{G}^{-1}(\widehat{g}-g)\right\|_2 \le\frac{\delta_g}{\mu-\delta_F}. \tag*{(118)}

For the second term, use the inverse-difference identity

G^−1−G−1=G^−1(G−G^)G−1.(119)\widehat{G}^{-1}-G^{-1}=\widehat{G}^{-1}(G-\widehat{G})G^{-1}. \tag*{(119)}

Therefore,

∥G^−1−G−1∥2≤1μ−δFδF1μ=δFμ(μ−δF).(120)\left\|\widehat{G}^{-1}-G^{-1}\right\|_2 \le\frac{1}{\mu-\delta_F}\delta_F\frac{1}{\mu}=\frac{\delta_F}{\mu(\mu-\delta_F)}. \tag*{(120)}

Combining the two bounds gives

∥u^−u∥2≤δgμ−δF+δFμ(μ−δF)∥g∥2.(121)\left\|\widehat{u}-u\right\|_2 \le\frac{\delta_g}{\mu-\delta_F}+\frac{\delta_F}{\mu(\mu-\delta_F)}\|g\|_2. \tag*{(121)}

This proves the theorem. □

Auxiliary Fixed-Scale Stationarity Result

The main text analyzes the actual normalized distance-aware update used by DA-QNM, θt+1=θt−rtqt\theta_{t+1}=\theta_t-r_tq_t. For completeness, we also record a standard fixed-scale stationarity result for the auxiliary update θt+1=θt−γdt\theta_{t+1}=\theta_t-\gamma d_t. This result clarifies the descent mechanism for fixed-scale spectral-natural directions, but it is not the main stationarity guarantee for the implemented distance-aware algorithm.

Theorem 6 (Fixed-scale nonconvex stationarity). Assume CC is LL-smooth and lower bounded by C⋆C_\star. Consider updates of the form θt+1=θt−γdt\theta_{t+1}=\theta_t-\gamma d_t. Suppose the update directions satisfy

⟨gt,dt⟩≥ν∥gt∥22,∥dt∥2≤B∥gt∥2(122)\langle g_t,d_t\rangle\ge\nu\|g_t\|_2^2,\qquad\|d_t\|_2\le B\|g_t\|_2 \tag*{(122)}

for constants ν>0\nu>0 and B<∞B<\infty. If

0<γ<2νLB2,(123)0<\gamma<\frac{2\nu}{LB^2}, \tag*{(123)}

then

1T∑t=0T−1∥∇C(θt)∥22≤C(θ0)−C⋆γ(ν−LγB22)T.(124)\frac{1}{T}\sum_{t=0}^{T-1}\|\nabla C(\theta_t)\|_2^2\le\frac{C(\theta_0)-C_\star}{\gamma\left(\nu-\frac{L\gamma B^2}{2}\right)T}. \tag*{(124)}

Consequently,

min⁡0≤t<T∥∇C(θt)∥22=O(1T).(125)\min_{0\le t<T}\|\nabla C(\theta_t)\|_2^2=O\left(\frac{1}{T}\right). \tag*{(125)}

Proof. By LL-smoothness,

C(θt+1)=C(θt−γdt)(126)C(\theta_{t+1})=C(\theta_t-\gamma d_t) \tag*{(126)}
≤C(θt)−γ⟨gt,dt⟩+Lγ22∥dt∥22.(127)\le C(\theta_t)-\gamma\langle g_t,d_t\rangle+\frac{L\gamma^2}{2}\|d_t\|_2^2. \tag*{(127)}

Using the alignment and norm bounds in (122),

C(θt+1)≤C(θt)−γν∥gt∥22+Lγ2B22∥gt∥22(128)C(\theta_{t+1})\le C(\theta_t)-\gamma\nu\|g_t\|_2^2+\frac{L\gamma^2B^2}{2}\|g_t\|_2^2 \tag*{(128)}
=C(θt)−γ(ν−LγB22)∥gt∥22.(129)=C(\theta_t)-\gamma\left(\nu-\frac{L\gamma B^2}{2}\right)\|g_t\|_2^2. \tag*{(129)}

The step-size condition ensures ν−LγB22>0\nu- \frac{L\gamma B^2}{2} > 0. Rearranging gives

γ(ν−LγB22)∥gt∥22≤C(θt)−C(θt+1).(130)\gamma\left(\nu- \frac{L\gamma B^2}{2}\right) \lVert g_t\rVert_2^2 \le C(\theta_t) - C(\theta_{t+1}). \tag*{(130)}

Summing over t=0,…,T−1t=0,\ldots,T-1, the right-hand side telescopes:

γ(ν−LγB22)∑t=0T−1∥gt∥22≤C(θ0)−C(θT)(131)\gamma\left(\nu- \frac{L\gamma B^2}{2}\right) \sum_{t=0}^{T-1} \lVert g_t\rVert_2^2 \le C(\theta_0) - C(\theta_T) \tag*{(131)}
≤C(θ0)−C⋆.(132)\le C(\theta_0) - C_\star. \tag*{(132)}

Dividing by γ(ν−LγB22)T\gamma\left(\nu-\frac{L\gamma B^2}{2}\right)T yields (124). Finally,

min⁡0≤t<T∥∇C(θt)∥22≤1T∑t=0T−1∥∇C(θt)∥22,(133)\min_{0\le t<T}\lVert\nabla C(\theta_t)\rVert_2^2 \le\frac{1}{T}\sum_{t=0}^{T-1}\lVert\nabla C(\theta_t)\rVert_2^2, \tag*{(133)}

so the minimum stationarity measure is also O(1/T)O(1/T).

Theorem 6 is auxiliary. The main stationarity guarantee for the implemented algorithm is Theorem 4, which applies to the normalized distance-aware update generated by Algorithm 1.

Additional Experimental Details

This section provides supplementary implementation details for the experiments in § 5. The main text reports the primary numerical tables and mechanism diagnostics, while this appendix records setup details, architecture diagrams, learning-curve evidence, and finite-shot robustness results. These details improve reproducibility and clarify how the optimizer variants are instantiated across VQE, weighted QAOA, and supervised quantum classification.

VQE Setup

Figure 3 summarizes the VQE experimental workflow used in our simulations, including state preparation, Hamiltonian expectation evaluation, optimizer update, and the DA-QNM-specific QFIM and spectral normalization steps.

VQE experimental architecture

Figure 3. VQE experimental architecture. A parameterized quantum circuit prepares the trial state ∣ψ(θ)⟩|\psi(\theta)\rangle, the Hamiltonian expectation E(θ)=⟨ψ(θ)∣H∣ψ(θ)⟩E(\theta)=\langle\psi(\theta)|\mathbf{H}|\psi(\theta)\rangle is evaluated, and the optimizer updates the circuit parameters. In the DA-QNM variants, gradients are first preconditioned by the damped QFIM, then reshaped into circuit-semantic blocks for spectral normalization and distance-aware radius control.

The VQE experiment uses a transverse-field Ising Hamiltonian with periodic boundary conditions. The variational objective is E(θ)=⟨ψ(θ)∣H∣ψ(θ)⟩E(\theta)=\langle\psi(\theta)|H|\psi(\theta)\rangle. The exact ground-state energy E⋆E_\star is computed by classical diagonalization, and the reported metric is the best energy error achieved over the optimization trajectory, min⁡t(E(θt)−E⋆)\min_t(E(\theta_t)-E_\star). We use a four-layer hardware-efficient ansatz. The QNM variants differ only in the circuit-semantic reshape used for spectral normalization: layer-wise blocks group parameters within each variational layer, gate-wise blocks group parameters by gate family, and the global variant applies a single larger reshape to the full parameter vector. The main text reports the best QNM variant per seed; in the current runs, QNM-layer is selected for all three tested seeds. This supports the interpretation that layer-wise spectral structure is the most compatible block geometry for this VQE ansatz.

Weighted QAOA Setup

Figure 4 summarizes the generalized QAOA workflow used in the weighted MaxCut experiment and highlights the parameter blocks on which gate-structured QNM operates.

Weighted QAOA experimental architecture

Figure 4. Weighted QAOA experimental architecture. The circuit alternates between cost unitaries parameterized by Γ∈R3×23\boldsymbol{\Gamma} \in\mathbb{R}^{3 \times23} and mixer unitaries parameterized by B∈R3×10\mathbf{B} \in\mathbb{R}^{3 \times10}. Gate-structured QNM applies damped QFIM preconditioning followed by spectral normalization on circuit-semantic blocks aligned with the cost-angle and mixer-angle parameter tensors.

The weighted QAOA experiment uses a MaxCut instance with n=10n=10, ∣E∣=23|E|=23, p=3p=3. The exact optimal cut value is computed by exhaustive enumeration: C⋆=17.818521C_\star=17.818521. The generalized QAOA parameters decompose into a cost-angle block and a mixer-angle block, Γ∈R3×10\Gamma\in\mathbb{R}^{3\times10}, B∈R3×10\mathbf{B}\in\mathbb{R}^{3\times10}. Here Γ\Gamma is indexed by depth and edge, while B\mathbf{B} is indexed by depth and node. This circuit-semantic decomposition makes weighted QAOA a natural testbed for gate-structured QNM: heterogeneous edge weights affect the cost-unitary generators and can induce anisotropic columns in the natural-gradient update associated with Γ\Gamma. All methods are evaluated on the same weighted instance and the same initialization seed. The purpose of this experiment is not to claim broad QAOA dominance from one instance, but to isolate whether QFIM preconditioning, gate-structured spectral normalization, and distance-aware radius control improve optimization on a controlled circuit-structured problem.

Additional Mechanism Figures

This section provides an additional spectral diagnostic supporting the mechanism discussed in § 5. The main text reports summary statistics such as the active condition number and effective rank. Here, we show the corresponding active singular spectrum directly, which makes visible how polar mixing changes the geometry of a representative QNM-gate momentum block.

Figure 5 complements the main spectral summary by displaying the singular values themselves. The observed flattening is consistent with the theoretical role of polar mixing: it preserves the active singular subspaces of the structured momentum block while raising weak active directions through the additive spectral floor. This provides empirical evidence that QNM mitigates spectral collapse in circuit-semantic update blocks, consistent with the predicted mechanism.

Active singular spectrum of a representative QNM-gate momentum block before and after Frobenius-preserving polar mixing

Figure 5. Active singular spectrum of a representative QNM-gate momentum block before and after Frobenius-preserving polar mixing. Before mixing, the spectrum is anisotropic, with a few dominant directions and several weak active singular directions. After mixing, the spectrum becomes flatter: weak active directions are lifted while the dominant structured directions are preserved. This visualizes the spectral-regularization mechanism established in Proposition 4 and Corollary 1.

Supervised Classification Setup

The supervised classification experiments use the Fisher Iris dataset as a controlled transfer diagnostic for the proposed optimizer. We consider two binary classification tasks. The first task separates setosa from the other two classes, while the second task separates versicolor from virginica after removing setosa examples. These two tasks provide different levels of difficulty: setosa is nearly linearly separable in standard feature space, whereas versicolor versus virginica is a harder binary discrimination task. Each train–test split is stratified so that the class proportions are preserved. Let xi∈R4x_i \in\mathbb{R}^{4}, yi∈{0,1}y_i \in\{0,1\} denote an input feature vector and its binary label. To avoid data leakage, PCA is fit only on the training split. The same fitted PCA transformation is then applied to the corresponding test split. Specifically, if WPCA∈R4×2\mathbf{W}_{\mathrm{PCA}} \in\mathbb{R}^{4 \times2} denotes the two-dimensional PCA projection learned from the training data, then the reduced feature is zi=WPCA(xi−xˉtrain)∈R2z_i = \mathbf{W}_{\mathrm{PCA}}(x_i-\bar{x}_{\mathrm{train}}) \in\mathbb{R}^{2}. The two principal coordinates are then rescaled to quantum rotation angles. We define the resulting encoded angle vector by ϕi=scale⁡(zi)∈R2\phi_i = \operatorname{scale}(z_i) \in\mathbb{R}^{2}. The scaling map is fit on the training split and then reused on the test split, again to prevent test information from entering the training pipeline. The classifier is a two-qubit data-reuploading variational circuit. For an encoded input ϕi\phi_i, the circuit prepares a state

∣ψ(ϕi;θ)⟩=U(ϕi;θ)∣00⟩,(134)\lvert\psi(\phi_i;\theta)\rangle= \mathbf{U}(\phi_i;\theta)\lvert00\rangle, \tag*{(134)}

where U(ϕi;θ)\mathbf{U}(\phi_i;\theta) alternates data-encoding rotations, trainable single-qubit rotations, and entangling gates. The trainable parameters are organized as a tensor Θ∈R2×2×3\boldsymbol{\Theta} \in\mathbb{R}^{2 \times2 \times3} where the three axes correspond to variational layers, qubits, and trainable rotation axes. Equivalently, the flattened parameter vector θ\theta has dimension d=2⋅2⋅3=12d = 2 \cdot2 \cdot3 = 12. This tensor structure provides the circuit-semantic block geometry used by QNM-style spectral normalization. The model output is obtained from a measured Pauli expectation value. In the binary setting, we use a scalar logit of the form hθ(ϕi)=a⟨Z0⟩ϕi,θ+bh_\theta(\phi_i) = a \langle Z_0 \rangle_{\phi_i,\theta} + b where aa and bb are trainable or fitted readout parameters and ⟨Z0⟩ϕi,θ\langle Z_0 \rangle_{\phi_i,\theta} denotes the expectation of the Pauli-Z observable on the readout qubit. The predicted probability is pθ(yi=1∣ϕi)=σ(hθ(ϕi))p_\theta(y_i = 1 \mid\phi_i) = \sigma(h_\theta(\phi_i)) where σ(⋅)\sigma(\cdot) is the logistic sigmoid. Training minimizes the empirical binary cross-entropy loss

C(θ)=1Ntrain∑i=1Ntrain[−yilog⁡pθ(yi=1∣ϕi)−(1−yi)log⁡(1−pθ(yi=1∣ϕi))].(135)C(\theta) = \frac{1}{N_{\mathrm{train}}} \sum_{i=1}^{N_{\mathrm{train}}} \left[-y_i \log p_\theta(y_i = 1 \mid\phi_i) - (1-y_i)\log(1-p_\theta(y_i = 1 \mid\phi_i))\right]. \tag*{(135)}

For supervised classification, the QNM variants use an empirical supervised QGT/QFIM constructed from the circuit sensitivities over the training mini-batch. At iteration tt, let gt=∇C(θt)g_t = \nabla C(\theta_t) denote the loss gradient and let Gtsup\mathbf{G}^{\mathrm{sup}}_t denote the empirical supervised geometry matrix. The damped natural-gradient direction is computed from

Gtsup=Ftsup+λI,ut=(Gtsup)−1gt.(136)\mathbf{G}^{\mathrm{sup}}_t = \mathbf{F}^{\mathrm{sup}}_t + \lambda\mathbf{I}, \qquad u_t = \left(\mathbf{G}^{\mathrm{sup}}_t\right)^{-1}g_t. \tag*{(136)}

We specify the empirical supervised geometry used in the classification experiments. Let Bt\mathcal{B}_t denote the training mini-batch at iteration tt. For each encoded input ϕi∈Bt\phi_i \in\mathcal{B}_t, define the input-conditioned variational state

∣ψi(θ)⟩=∣ψ(ϕi;θ)⟩=U(ϕi;θ)∣00⟩.(137)\lvert\psi_i(\theta)\rangle= \lvert\psi(\phi_i;\theta)\rangle= \mathbf{U}(\phi_i;\theta)\lvert00\rangle. \tag*{(137)}

For circuit parameters θj\theta_j and θk\theta_k, the sample-wise quantum geometric tensor is

Qi(θ)jk=⟨∂jψi(θ)∣∂kψi(θ)⟩−⟨∂jψi(θ)∣ψi(θ)⟩⟨ψi(θ)∣∂kψi(θ)⟩,(138)Q_i(\theta)_{jk} = \langle\partial_j\psi_i(\theta)\mid\partial_k\psi_i(\theta)\rangle- \langle\partial_j\psi_i(\theta)\mid\psi_i(\theta)\rangle\langle\psi_i(\theta)\mid\partial_k\psi_i(\theta)\rangle, \tag*{(138)}

where ∂jψi(θ)\partial_j\psi_i(\theta) denotes ∂∣ψi(θ)⟩/∂θj\partial\lvert\psi_i(\theta)\rangle/\partial\theta_j. The empirical supervised QFIM used by QNM is the mini-batch average of the real part of this input-conditioned QGT:

Ftsup=1∣Bt∣∑i∈BtRe⁡(Qi(θt)).(139)\mathbf{F}^{\mathrm{sup}}_t = \frac{1}{\lvert B_t\rvert}\sum_{i\in B_t}\operatorname{Re}(\mathbf{Q}_i(\theta_t)). \tag*{(139)}

Equivalently, defining the projected tangent vector

∣ξi,j(θ)⟩=(I−∣ψi(θ)⟩⟨ψi(θ)∣)∂j∣ψi(θ)⟩,(140)\lvert\xi_{i,j}(\theta)\rangle= \left(\mathbf{I}-\lvert\psi_i(\theta)\rangle\langle\psi_i(\theta)\rvert\right)\partial_j\lvert\psi_i(\theta)\rangle, \tag*{(140)}

we have

Ftsup[j,k]=1∣Bt∣∑i∈BtRe⁡⟨ξi,j(θt)∣ξi,k(θt)⟩.(141)\mathbf{F}^{\mathrm{sup}}_t[j,k] = \frac{1}{\lvert B_t\rvert}\sum_{i\in B_t}\operatorname{Re}\langle\xi_{i,j}(\theta_t)\rvert\xi_{i,k}(\theta_t)\rangle. \tag*{(141)}

Thus, Ftsup\mathbf{F}^{\mathrm{sup}}_t measures the average local sensitivity of the data-dependent quantum states over the supervised training mini-batch. The diagonal variant uses Ft,diagsup=diag⁡(Ftsup)\mathbf{F}^{\mathrm{sup}}_{t,\mathrm{diag}}=\operatorname{diag}(\mathbf{F}^{\mathrm{sup}}_t) while the damped geometry used for preconditioning is Gtsup=Ftsup+λI\mathbf{G}^{\mathrm{sup}}_t=\mathbf{F}^{\mathrm{sup}}_t+\lambda\mathbf{I} or, for Diag-QNM, Gt,diagsup=Ft,diagsup+λI\mathbf{G}^{\mathrm{sup}}_{t,\mathrm{diag}}=\mathbf{F}^{\mathrm{sup}}_{t,\mathrm{diag}}+\lambda\mathbf{I}. The readout parameters aa and bb, when optimized, are treated as ordinary Euclidean readout parameters and are not included in the circuit QGT block. The QNM spectral step is applied to the variational circuit parameter tensor Θ\boldsymbol{\Theta}. Using this empirical supervised geometry, the same QNM terminology used in the main paper is applied to the supervised setting:

  • Diag-QNM uses a diagonal approximation of Ftsup\mathbf{F}^{\mathrm{sup}}_t, followed by damped preconditioning and Muon-style spectral normalization on the trainable parameter tensor.

  • DA-QNM applies distance-aware radius control to the normalized supervised QNM direction, using the same radius mechanism as in the main algorithm.

This experiment is intentionally small-scale. Its purpose is not to claim state-of-the-art supervised quantum classification performance on Iris, but to test whether the optimizer components transfer from variational quantum optimization objectives to a supervised empirical-loss setting. The exact training results in the main paper evaluate the deterministic optimizer behavior, while the finite-shot variants in Appendix § 9.§ 5 test robustness when expectation values and gradients are estimated from sampled measurements.

Finite-Shot Supervised Results

The main paper focuses on deterministic descent, perturbation stability, and stationarity of the normalized distance-aware update. Finite-shot supervised classification is therefore reported here as an empirical robustness study, rather than as a theoretically characterized probabilistic shot-noise guarantee. The goal is to test whether the proposed geometric updates remain useful when expectations, gradients, and empirical geometry estimates are computed from sampled measurements.

Training is performed either with exact expectation values or with finite-shot sampled observables. After training, final test metrics are evaluated exactly. Thus, the reported accuracies isolate the effect of finite-shot sampling noise on the optimization trajectory, rather than confounding it with sampling noise at evaluation time. All methods are trained for 70 optimization steps and evaluated over five independent seeds. We report mean test accuracy in percent together with the standard error across seeds.

Table 4 reports the finite-shot supervised classification sweep. The exact-training rows first show that QNM-style quantum-geometric updates remain competitive beyond variational energy optimization and weighted QAOA. On setosa versus rest, Diag-QNM and DA-QNM both achieve 93.33% test accuracy, compared with 91.85% for Adam. On the harder versicolor-versus-virginica task, DA-QNM obtains the best exact-training result, reaching 86.67%, compared with 80.56% for Adam and 83.33% for Diag-QNM. These results support the interpretation that QGT/QFIM-informed spectral updates can transfer to supervised empirical-loss objectives, although this Iris experiment should be viewed as a controlled diagnostic rather than a broad supervised-QML benchmark.

TaskShotsAdamDiag-QNMDA-QNM
Setosa vs. restExact91.85±0.7491.85 \pm0.7493.33±1.2893.33 \pm1.2893.33±1.2893.33 \pm1.28
Setosa vs. rest3243.70±11.6443.70 \pm11.6452.59±9.9752.59 \pm9.9744.44±6.7944.44 \pm6.79
Setosa vs. rest12851.11±5.8851.11 \pm5.8855.56±3.8555.56 \pm3.8553.33±4.6353.33 \pm4.63
Setosa vs. rest51254.81±6.0654.81 \pm6.0660.00±7.8060.00 \pm7.8068.15±4.5168.15 \pm4.51
Versicolor vs. virginicaExact80.56±4.0180.56 \pm4.0183.33±3.8583.33 \pm3.8586.67±1.9286.67 \pm1.92
Versicolor vs. virginica3245.56±7.2945.56 \pm7.2934.44±7.7834.44 \pm7.7841.11±8.0141.11 \pm8.01
Versicolor vs. virginica12860.00±13.3360.00 \pm13.3348.89±4.0148.89 \pm4.0152.22±4.8452.22 \pm4.84
Versicolor vs. virginica51256.67±5.0956.67 \pm5.0948.89±10.9448.89 \pm10.9463.33±8.3963.33 \pm8.39

Table 4. Finite-shot training sweep on the two Iris binary classification tasks. Training uses either exact expectation values or finite-shot sampled observables, while final test metrics are evaluated exactly after 70 optimization steps. Results are mean test accuracy in percent ±\pm standard error over five seeds.

The finite-shot rows reveal a more nuanced pattern. At low and intermediate shot budgets, richer quantum-geometric updates are not uniformly beneficial. On setosa versus rest, Diag-QNM is strongest at 32 and 128 shots, reaching 52.59% and 55.56%, respectively. This suggests that a diagonal empirical geometry can provide useful stabilization when the measurement budget is limited, because it introduces quantum-geometric preconditioning while avoiding the higher estimation burden of richer geometry. On the harder versicolor-versus-virginica task, Adam remains the strongest method at 32 and 128 shots, reaching 45.56% and 60.00%, respectively. This behavior is consistent with the fact that finite-shot noise affects not only the gradient estimate but also the empirical geometry used by QNM-style updates.

At the largest finite-shot budget, DA-QNM becomes the strongest method on both tasks. With 512 shots, DA-QNM reaches 68.15% on setosa versus rest, compared with 54.81% for Adam and 60.00% for Diag-QNM. On versicolor versus virginica, DA-QNM reaches 63.33%, compared with 56.67% for Adam and 48.89% for Diag-QNM. Thus, at the most informative finite-shot setting tested here, DA-QNM improves over Adam by 13.34 percentage points on setosa versus rest and by 6.66 percentage points on versicolor versus virginica.

Overall, the finite-shot experiment suggests that the benefit of distance-aware quantum-geometric spectral updates depends on the quality of the sampled gradient and geometry estimates. When the shot budget is too small, simpler diagonal geometry or even Adam can be more robust. As the shot budget increases, the empirical QGT/QFIM and gradient estimates become more informative, allowing the richer DA-QNM update to exploit preconditioning, spectral normalization, and radius control more effectively. We therefore interpret Table 4 as finite-shot robustness evidence for the proposed optimizer, while leaving a full high-probability convergence analysis under shot noise to future work.

Figure 6a and Figure 6b visualize the shot-budget dependence reported in Table 4. The setosa task shows that diagonal empirical geometry can be useful when the measurement budget is limited: Diag-QNM performs best at 32 and 128 shots. In contrast, the richer distance-aware update becomes most effective at 512 shots, where DA-QNM achieves the highest accuracy. The harder versicolor-versus-virginica task shows a similar transition: Adam is competitive when the shot budget is small, but DA-QNM becomes the strongest method once the finite-shot estimates are sufficiently informative. These patterns support the interpretation that QFIM preconditioning, spectral normalization, and distance-aware radius control are complementary, but that their combined benefit depends on the quality of the sampled gradient and geometry estimates.

Finite-shot Iris classification test accuracy for Setosa-versus-rest and Versicolor-versus-virginica tasks

Figure 6. Finite-shot Iris classification after 70 optimization steps. Training uses either exact expectation values or finite-shot sampled observables, while final test metrics are evaluated exactly. Error bars denote the standard error over five seeds. (a) Diag-QNM performs best at low and intermediate shot budgets, whereas DA-QNM achieves the highest accuracy at 512 shots. (b) Adam remains competitive at low and intermediate shot budgets, while DA-QNM achieves the highest accuracy in the exact and 512-shot settings.

Implementation and Reproducibility

All exact-expectation experiments use statevector simulation. Finite-shot supervised experiments use sampled observables during training, while final test metrics are evaluated exactly after training. This protocol isolates optimization noise due to finite-shot estimation from sampling noise at evaluation time. The VQE ground-truth energies are computed by exact diagonalization, and the weighted QAOA optimum is computed by exhaustive enumeration for the n=10n = 10 weighted MaxCut instance. Unless otherwise stated, results are reported as mean ±\pm standard error over the indicated seeds or splits.

Implementation protocol. All optimizers are evaluated on matched task instances, matched random seeds, and the same initialization whenever applicable. For deterministic statevector experiments, observed differences therefore reflect optimizer behavior rather than measurement noise. We report best-achieved performance when the goal is to test whether an optimizer reaches a good basin, and final performance when the goal is to assess convergence stability.

Task-specific implementation details. For the VQE experiment, we use a four-layer hardware-efficient ansatz for a transverse-field Ising Hamiltonian with periodic boundary conditions. The reported metric is the best energy error achieved during training, min⁡t(E(θt)−E⋆)\min_t (E(\theta_t) - E_\star) where E⋆E_\star is obtained by exact diagonalization. The VQE results are reported over seeds 0, 1, 2, and the selected best QNM variant is QNM-layer for all three seeds. For the weighted QAOA experiment, the instance has n=10n = 10, ∣E∣=23|E| = 23, p=3p = 3. The representative weighted QAOA run used matched initialization across optimizers. The DA-QAOA implementation used CPU-based double-precision simulation, seed 0, 180 optimization steps, learning rate 0.04, Adam learning rate 0.05, momentum coefficient β=0.9\beta= 0.9, QFIM damping λ=0.05\lambda= 0.05, QFIM refresh period 3, maximum update norm 0.05, and polar-mixing coefficient α=0.3\alpha= 0.3. For supervised quantum classification, we use the Fisher Iris dataset with two binary tasks: setosa versus rest and versicolor versus virginica. Exact training results are reported over stratified train–test splits. The finite-shot supervised experiments train each method for 70 optimization steps and evaluate over five independent seeds. Training uses either exact expectation values or finite-shot sampled observables, while final metrics are evaluated exactly.

References

  1. [1]Shun-ichi Amari. Natural gradient works efficiently in learning. Neural Computation, 10(2):251–276, Feb. 1998. doi: 10.1162/089976698300017746. URL https://doi.org/10.1162/089976698300017746.
  2. [2]Marcello Benedetti, Erika Lloyd, Stefan Sack, and Mattia Fiorentini. Parameterized quantum circuits as machine learning models. Quantum Science and Technology, 4(4):043001, Nov. 2019. doi: 10.1088/2058-9565/ab4eb5. URL https://doi.org/10.1088/2058-9565/ab4eb5.
  3. [3]Jeremy Bernstein and Laker Newhouse. Modular duality in deep learning. In Proceedings of the 42nd International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=hErdfTsLu.arxiv.org/abs/2410.21265
  4. [4]Marco Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C. Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R. McClean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio, and Patrick J. Coles. Variational quantum algorithms. Nature Reviews Physics, 3:625–644, Sep. 2021. doi: 10.1038/s42254-021-00348-9. URL https://doi.org/10.1038/s42254-021-00348-9.
  5. [5]Yury Demidovich, Abhishek Chakraborty, Grigory Malinovsky, Angelia Nedić, and Peter Richtárik. Distance-aware Muon: Adaptive step scaling for normalized optimization, 2026. URL https://arxiv.org/abs/2605.18999.
  6. [6]John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(61):2121–2159, 2011. URL https://jmlr.org/papers/v12/duchi11a.html.
  7. [7]Edward Farhi, Jeffrey Goldstone, and Sam Gutmann. A quantum approximate optimization algorithm, 2014. URL https://arxiv.org/abs/1411.4028.
  8. [8]Stuart Hadfield, Zhihui Wang, Bryan O’Gorman, Eleanor G. Rieffel, Davide Venturelli, and Rupak Biswas. From the quantum approximate optimization algorithm to a quantum alternating operator ansatz. Algorithms, 12(2):34, Feb. 2019. doi: 10.3390/a12020034. URL https://doi.org/10.3390/a12020034.
  9. [9]Vojtěch Havlíček, Antonio D. Córcoles, Kristan Temme, Aram W. Harrow, Abhinav Kandala, Jerry M. Chow, and Jay M. Gambetta. Supervised learning with quantum-enhanced feature spaces. Nature, 567(7747):209–212, Mar. 2019. doi: 10.1038/s41586-019-0980-2. URL https://doi.org/10.1038/s41586-019-0980-2.
  10. [10]Nicholas J. Higham. Computing the polar decomposition—with applications. SIAM Journal on Scientific and Statistical Computing, 7(4):1160–1174, 1986. doi: 10.1137/090707. URL https://doi.org/10.1137/0907079.
  11. [11]Nicholas J. Higham. Functions of Matrices: Theory and Computation. Society for Industrial and Applied Mathematics (SIAM), 2008. doi: 10.1137/1.9780898717778. URL https://doi.org/10.1137/1.9780898717778.
  12. [12]Keller Jordan, Yuchen Jin, Vlado Boza, Jiacheng You, Franz Cesista, Laker Newhouse, and Jeremy Bernstein. Muon: An optimizer for hidden layers in neural networks. Online, 2024. URL https://kellerjordan.github.io/posts/muon/.
  13. [13]Abhinav Kandala, Antonio Mezzacapo, Kristan Temme, Maika Takita, Markus Brink, Jerry M. Chow, and Jay M. Gambetta. Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets. Nature, 549(7671):242–246, Sep. 2017. doi: 10.1038/nature23879. URL https://doi.org/10.1038/nature23879.
  14. [14]Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. URL https://arxiv.org/abs/1412.6980.
  15. [15]Adrián Pérez-Salinas, Alba Cervera-Lierta, Elies Gil-Fuster, and José I. Latorre. Data re-uploading for a universal quantum classifier. Quantum, 4:226, Feb. 2020. doi: 10.22331/q-2020-02-06-226. URL https://doi.org/10.22331/q-2020-02-06-226.
  16. [16]Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man-Hong Yung, Xiao-Qi Zhou, Peter J. Love, Alán Aspuru-Guzik, and Jeremy L. O’Brien. A variational eigenvalue solver on a photonic quantum processor. Nature Communications, 5:4213, Jul. 2014. doi: 10.1038/ncomms5213. URL https://doi.org/10.1038/ncomms5213.
  17. [17]Jean Pierre Provost and Guillaume Vallee. Riemannian structure on manifolds of quantum states. Communications in Mathematical Physics, 76(3):289–301, 1980. doi: 10.1007/BF02193559. URL https://doi.org/10.1007/BF02193559.
  18. [18]Maria Schuld and Nathan Killoran. Quantum machine learning in feature hilbert spaces. Phys. Rev. Lett., 122(4):040504, Feb. 2019. doi: 10.1103/PhysRevLett.122.040504. URL https://doi.org/10.1103/PhysRevLett.122.040504.
  19. [19]James Stokes, Josh Izaac, Nathan Killoran, and Giuseppe Carleo. Quantum natural gradient. Quantum, 4:269, May 2020. doi: 10.22331/q-2020-05-25-269. URL https://doi.org/10.22331/q-2020-05-25-269.

Paper details

vlnc:W_gsmgsku2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents