# Pauli-Spectrum Natural Gradient

- ID: 2663
- Canonical URL: https://synthcore.org/idea/2663/pauli-spectrum-natural-gradient
- API JSON: https://synthcore.org/api/idea/2663.json
- API Markdown: https://synthcore.org/api/idea/2663.md
- Verification status: unverified
- Source: [arXiv:2608.21437](https://arxiv.org/abs/2608.21437)
- Category: optimization
- Solves: stability, sample-efficiency, accuracy
- ML areas: optimizer, training-dynamics
- Math tags: geometry, information-theory, probability, linear-algebra
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Train a normalized neural quantum state with a natural-gradient preconditioner computed from the Fisher geometry of its labeled Pauli spectrum. Instead of estimating the usual wavefunction quantum Fisher matrix from state derivatives and overlap covariances, estimate Pauli expectations, differentiate their squared values, and use one half of the resulting classical Fisher matrix as the metric.

## Mathematical statement

For an N-qubit pure state \(\rho=|\psi\rangle\langle\psi|\), let \(d=2^N\), let \(\mathcal P_N=\{I,X,Y,Z\}^{\otimes N}\), and define the real Pauli coefficient \(a_P(\theta)=\operatorname{Tr}(P\rho_\theta)=\langle\psi_\theta|P|\psi_\theta\rangle\). Orthogonality \(\operatorname{Tr}(PP')=d\,\delta_{P,P'}\) gives \(\rho=(1/d)\sum_P a_P P\) and \(\operatorname{Tr}(\rho^2)=(1/d)\sum_P a_P^2\). Purity implies that \(\Xi_\theta(P)=a_P(\theta)^2/d\) is a normalized distribution over labeled Pauli strings. The paper's central identity states that its classical Fisher matrix \(F^{\Xi}_{ij}=\sum_{P:\Xi(P)>0}(\partial_i\Xi(P))(\partial_j\Xi(P))/\Xi(P)\) satisfies \(F^{\Xi}_{ij}=2F^Q_{ij}\), where \(F^Q\) is the pure-state quantum Fisher information matrix under the paper's convention. Therefore use \(G=F^{\Xi}/2\) as a quantum-natural-gradient metric. Away from zero coefficients, \(\partial_i\Xi(P)=2a_P\partial_i a_P/d\), so \(F^{\Xi}_{ij}=(4/d)\sum_{P:a_P\neq0}(\partial_i a_P)(\partial_j a_P)\).

## Key formulas

- $$\Xi_{\rho}(P)=\frac{1}{d}[\operatorname{Tr}(P\rho)]^{2}=\frac{a_{P}^{2}}{d},\qquad d=2^N,$$
- $$\rho=\frac{1}{d}\sum_{P\in\mathcal{P}_{N}}a_{P}P,\qquad \operatorname{Tr}(\rho^{2})=\frac{1}{d}\sum_{P\in\mathcal{P}_{N}}a_{P}^{2}=1,$$
- $$F^{\Xi}_{ij}=\sum_{P:\Xi(P)>0}\frac{\partial_i\Xi(P)\,\partial_j\Xi(P)}{\Xi(P)}=\frac{4}{d}\sum_{P:a_P\neq0}(\partial_i a_P)(\partial_j a_P)=2F^{Q}_{ij},$$
- $$\theta_{t+1}=\theta_t-\eta\,(\widehat G_t+\lambda I)^{-1}\widehat g_t,\qquad \widehat G_t=\frac{1}{2}\widehat F^{\Xi}_t,$$

## Implementation notes

Integrate this into the optimizer for a normalized neural quantum state, not into an ordinary unconstrained classifier. Let the network output complex amplitudes \(\tilde\psi_\theta(x)\) for computational-basis strings \(x\in\{0,1\}^N\), and normalize them by \(\psi_\theta=\tilde\psi_\theta/\|\tilde\psi_\theta\|_2\). At each optimizer step, choose a set of Pauli strings \(S_t\subset\mathcal P_N\), either all strings for \(N\leq 6\) or a fresh random subset for larger systems. For every \(P\in S_t\), compute \(a_P=\langle\psi_\theta|P|\psi_\theta\rangle\) by applying the sparse permutation-and-phase action of \(P\) to the amplitude vector; obtain \(J_{P,i}=\partial a_P/\partial\theta_i\) with reverse-mode autodiff or Jacobian-vector products. Form \(\widehat F_{ij}=4\sum_{P\in S_t}w_P J_{P,i}J_{P,j}/d\), where \(w_P\) is an importance weight correcting the Pauli sampling distribution; for uniform sampling over \(M\) strings, set \(w_P=|\mathcal P_N|/M\). Add \(\epsilon I\) before solving the linear system. Pseudocode: compute normalized amplitudes; sample \(S_t\); compute each \(a_P\) and \(J_P\); accumulate \(\widehat F\leftarrow\widehat F+(4w_P/d)J_PJ_P^\top\); set \(\widehat G=\widehat F/2\); solve \((\widehat G+\lambda I)u=g\); update \(\theta\leftarrow\theta-\eta u\). Use \(\lambda\) large enough to handle zero or nearly zero Pauli coefficients, because the unsimplified Fisher expression omits outcomes with exactly zero probability. The exact mathematical quantity is the complete-distribution Fisher matrix; the sampled subset, importance weights, damping, and low-rank linear solver are engineering approximations. First test on 4- to 6-qubit transverse-field Ising ground-state energy minimization with a small autoregressive or MLP wavefunction, comparing Adam, stochastic reconfiguration using the standard wavefunction metric, and this Pauli-spectrum metric at equal Pauli-evaluation cost. Measure energy decrease per optimizer step, variance across random seeds, condition number of the preconditioner, and final energy error. Success would be faster energy descent or lower variance than Adam, with performance approaching standard quantum natural gradient when all Pauli strings are included.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.