✗ Failed on benchmark 2026

Davis–Wielandt Certified Residual Blocks

Implementation & benchmark of arXiv:2607.10335 — Geometric Decentralized Stability Certificate of Power Electronics-Dominated Power Systems Covering Variable Operating Points

Usefulness8/10

Difficulty5/10

Novelty7/10

Dynamics Stability Accuracy Optimizer Training dynamics Mlp Transformer Control theory Dynamical systems Linear algebra Spectral theory Geometry

Source paper: Geometric Decentralized Stability Certificate of Power Electronics-Dominated Power Systems Covering Variable Operating Points arXiv:2607.10335 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace unconstrained residual updates with blocks whose Jacobian is monitored through a Davis–Wielandt shell. The shell simultaneously measures directional dissipation and non-normal amplification, yielding a per-block step-size or residual-scale bound that is stronger than checking only the largest eigenvalue or spectral norm.

Formulas

$$\mathcal{DW}(A)=\left\{\left(\operatorname{Re}(x^{*}Ax),\,\|Ax\|_{2}^{2}\right):\|x\|_{2}=1\right\}.$$

$$\frac{\|(I+hA)x\|_{2}^{2}}{\|x\|_{2}^{2}}=1+2h\operatorname{Re}(x^{*}Ax)+h^{2}\|Ax\|_{2}^{2}=1+2hu+h^{2}v.$$

$$\sup_{(u,v)\in\mathcal{DW}(A)}\left(1+2hu+h^{2}v\right)\le 1-\gamma\quad\Longrightarrow\quad \|I+hA\|_{2}^{2}\le 1-\gamma.$$

$$u\le-\alpha<0,\quad v\le\beta\quad\Longrightarrow\quad 0<h<\frac{2\alpha}{\beta},\qquad \rho_{\mathrm{bound}}(h)=1-2h\alpha+h^{2}\beta<1.$$

Mathematical statement

For a real or complex block Jacobian A in C^(d x d) and any unit vector x, define the Davis–Wielandt shell DW(A) as the set of pairs (u,v), where u = Re(x* A x) and v = ||A x||_2^2. For an explicit residual step x_(k+1) = x_k + h A x_k, its directional squared gain is ||x_(k+1)||_2^2 / ||x_k||_2^2 = 1 + 2 h u + h^2 v. Therefore, a sufficient contraction certificate with margin gamma > 0 is 1 + 2 h u + h^2 v <= 1 - gamma for every point (u,v) in DW(A). If u <= -alpha < 0 and v <= beta, a conservative step-size condition is 0 < h < 2 alpha / beta, with contraction factor bounded by 1 - 2 h alpha + h^2 beta. In a neural block, A is the input Jacobian J_f(z) = partial f_theta(z) / partial z and the residual update is z_plus = z + h f_theta(z). The DW shell is estimated using random unit probes x_j, Jacobian-vector products y_j = J_f(z) x_j, u_j = Re(x_j* y_j), and v_j = ||y_j||_2^2.

Implementation notes

1. Integration point: use this in residual MLPs, Transformer feed-forward residual branches, or continuous-depth networks. Let a block be z_plus = z + s_l f_(theta_l)(z), where s_l is a learned or scheduled residual scale. Before each optimizer update, and periodically during training, estimate the DW projection of the block Jacobian on minibatch activations. 2. Pseudocode: for each block and minibatch, draw m normalized random probes x_j; compute y_j = J_f(z) x_j with forward-mode JVPs or reverse-mode vector-Jacobian products; compute u_j = x_j transpose y_j and v_j = ||y_j|| squared; set alpha_hat = -max_j u_j and beta_hat = max_j v_j; choose s_l = min(s_l, 2(1-epsilon) alpha_hat / beta_hat) when alpha_hat > 0, otherwise shrink the block and add the penalty max_j[1 + 2 s_l u_j + s_l squared v_j - 1 + gamma]_+. 3. Computed versus estimated: the residual-gain identity and DW contraction condition are analytic; shell extrema are estimated from probes, minibatch states, and optionally endpoints of a prescribed activation region. Use m = 16 initially and retain exponential-moving maxima of u_j and v_j to reduce false certificates. 4. First cheap experiment: train a 6-layer residual MLP on MNIST or CIFAR-10 with ordinary residual scaling versus DW scaling, using identical Adam settings. Sweep the nominal residual step s and measure activation-norm growth and gradient stability. The quantitative prediction is a divergence boundary near s_c approximately equal to 2 alpha_hat / beta_hat; stable operation should persist below (1-epsilon)s_c and fail or show exploding activation norms above s_c for blocks with approximately fixed Jacobians. Also test the predicted decay law ||delta z_k|| squared less than or approximately equal to (1 - 2 s alpha_hat + s squared beta_hat)^k ||delta z_0|| squared by injecting two nearby inputs and fitting the log-distance slope.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built and numerically verified a Davis–Wielandt residual-block MVP. The gain identity held to 6.1e-16 absolute error; increasing non-normality K from 0 to 1.9 reduced the certified boundary from 2.0 to 0.0184, while the true contraction boundary decreased from 2.0 to 0.1, confirming sensitivity to non-normal amplification. The certificate was conservative rather than tight: for K=1.4, predicted h=0.1626 versus observed operator contraction boundary h=0.6. The certified scale prevented amplification, but no neural-network accuracy or practical training win was tested.

Agent confidence: 8/10

Baseline: Linear residual chain at K=1.4, scale h=0.8: one-step squared operator gain 1.3332 (>1), maximum trajectory norm 1.1546, final norm 1.16e-12 for the tested singular-vector initialization.
Idea: DW-certified scale h=0.1594: one-step squared operator gain 0.9207, DW upper bound 0.9981, maximum trajectory norm 1.0000, final norm 0.1704 after 20 steps. Identity maximum error was 6.1e-16; certified versus observed boundary was 0.1626 versus 0.6000.

Limitations: This is a small exact linear 2x2 NumPy toy experiment, not a trained MNIST/CIFAR residual MLP. Random shell probes were used rather than forward-mode JVPs in a neural block; probe estimates can miss extrema, and no optimizer, activation-region drift, runtime, memory, or accuracy comparison was measured.

How to run: python3 dw_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.10335) Sep 2, 2026

No significant difference vs. the tuned baseline (−-0.0%, p=1, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0141
Idea mean: 0.0141
Effect (Δ): 0 (−-0.0%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 1 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.0%
Mechanism: Not confirmed ✗
Practical effect: No effect
Baseline tuning: swept over 6 configs

Limitations:

The benchmark used 400 training and 200 test examples and 5 epochs after the initial full attempt exceeded 40 minutes; shell probing was periodic with 4 random probes over 8 activations rather than 16 probes at every batch. The intervention was tested only on the registered built-in dynamics track and one residual GRU branch, not on transformers or larger residual networks.

How to run: python3 stage2_dw_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 30, 2026 latest Failed ✗

Agent confidence: 10/10

The registered dynamics/rnn_small benchmark was rerun with the required tuned baseline sweep and 8 paired seeds. The Davis–Wielandt idea produced exactly the same test MSE as the baseline (paired delta 0, permutation p=1), and its trained-model mechanism signature was not confirmed because alpha_hat=0 and the observed local gain exceeded the predicted bound.

Baseline: Best baseline lr=0.006, scale=1.3; mean test MSE 0.014112892575212754 over 8 seeds.
Idea: DW-certified residual GRU lr=0.006, scale=1.3; mean test MSE 0.014112892575212754 over 8 seeds; paired delta 0 and p=1.

Benchmark result