△ Mechanism confirmed, baseline not beaten 2026

Adjoint-Weak Fractional Residuals

Implementation & benchmark of arXiv:2608.12879 — Robust data-driven discovery of fractional differential equations via weak formulations and Pareto-based subset selection

Usefulness7/10

Difficulty4/10

Novelty5/10

Training Stability Accuracy Sample efficiency Training Loss World model Fractional analysis Functional analysis Numerical analysis

Source paper: Robust data-driven discovery of fractional differential equations via weak formulations and Pareto-based subset selection arXiv:2608.12879 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace pointwise fractional derivatives of noisy trajectories in a neural PDE or neural dynamics loss with weak projections in which the fractional operator acts on smooth test functions. The network is trained to match integral residuals over local space-time windows, making the residual insensitive to high-frequency measurement noise while retaining sensitivity to the underlying fractional dynamics.

Formulas

$$\int_a^b \phi(z)\,({}_{a}D_{z}^{\alpha}f)(z)\,dz = \int_a^b f(z)\,({}_{z}D_{b}^{\alpha}\phi)(z)\,dz$$

$$W_{j,\alpha}^{t}[u]=\int_{0}^{T}\!\int_{\Omega} \phi_j(t,x)\,\mathcal{T}_{\alpha}u(t,x)\,dx\,dt=\int_{0}^{T}\!\int_{\Omega}u(t,x)\,\mathcal{T}_{\alpha}^{*}\phi_j(t,x)\,dx\,dt$$

$$R_j(\theta)=W_{j,\alpha}^{t}[u_\theta]-c_xW_{j,\beta}^{x}[u_\theta]-c_0W_j[u_\theta]-W_j[F_\theta],\qquad \mathcal{L}_{\mathrm{weak}}(\theta)=\frac{1}{J}\sum_{j=1}^{J}R_j(\theta)^2+\lambda\|\theta\|_2^2$$

$$\operatorname{Var}\!\left[\sum_i q_i\varepsilon_i\right]=\sigma^2\sum_iq_i^2$$

Mathematical statement

The construction uses the fractional integration-by-parts or adjoint identity: for a sufficiently smooth test function phi with boundary terms eliminated, the integral of phi(z) times the left Riemann-Liouville derivative of f equals the integral of f(z) times the right adjoint fractional derivative of phi. Here D_{a+}^alpha is the left fractional derivative of order alpha on [a,b], D_{b-}^alpha is its right adjoint, f is the observed or predicted field, and phi is a compactly supported smooth test function. For a neural PDE u_theta(t,x), define a weak residual by integrating each candidate equation term against phi_j and moving derivatives from noisy u to analytically or numerically precomputed derivatives of phi_j. The smoothing asset is that the resulting feature is an integral against a fixed smooth kernel: for independent zero-mean grid noise epsilon_i with variance sigma^2, its variance is sigma^2 times the squared norm of the quadrature weights, rather than the amplified variance produced by pointwise fractional differentiation.

Implementation notes

(1) Integration point: modify the loss of a PINN, neural operator, or neural ODE/world model that predicts u_theta(t,x). Do not compute fractional derivatives of noisy training observations with automatic differentiation. Instead, use local space-time patches and fixed smooth test functions phi_j, such as compactly supported cubic B-splines or Gaussian bumps, with J functions per minibatch. For each fractional order alpha or beta, precompute T_alpha^* phi_j and X_beta^* phi_j on the quadrature grid. (2) Pseudocode: sample a minibatch of windows Q_b and quadrature points (t_i,x_i); evaluate y_i=u_theta(t_i,x_i); for every test function j compute w_time[j]=sum_i y_i*(T_alpha_star_phi[j])[i]*quadrature_weight[i]; compute w_space[j] analogously; compute w_source[j]=sum_i F_theta(t_i,x_i)*phi_j[i]*quadrature_weight[i]; set residual[j]=w_time[j]-c_x*w_space[j]-c_0*w_identity[j]-w_source[j]; backpropagate mean(residual**2) plus the ordinary data-fit loss. The adjoint identity determines the precomputed kernels and the displayed residual equation determines the loss. (3) Compute the adjoint fractional derivative of each phi_j analytically or with high-order quadrature; estimate no derivative from noisy labels. Tune window size, test-function width, and the weak/data-loss mixing coefficient on a validation set. Use several alpha values if the order is unknown, but initially fix alpha to the ground-truth order to isolate robustness. (4) First experiment: train a small MLP PINN on one-dimensional fractional advection-diffusion or Burgers trajectories with 1%, 5%, and 10% multiplicative Gaussian noise. Compare strong-form autodiff residuals against weak residuals using the same network, optimizer, collocation points, and FLOPs. Measure held-out trajectory error, residual variance across noise seeds, and convergence failures. Success is lower validation error and a smoother loss curve at larger alpha and noise, with weak-feature variance decreasing when quadrature resolution is doubled.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

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 9/10

Built `experiment.py`, which uses a Grünwald fractional-derivative matrix and its exact weighted transpose as the adjoint weak kernel. The discrete integration-by-parts identity held to relative error 4.84e-16; weak noise variance was proportional to sigma^2 with coefficient 0.01056–0.01139. Under each grid doubling, weak squared kernel norm ratios were 0.503–0.506 versus the predicted 0.5, while strong pointwise derivative norms grew by 3.031 versus the predicted 2^(2*0.8)=3.031. In the matched coefficient toy, weak noise-induced RMSE was 20.1x, 23.8x, 26.0x, and 47.4x lower than the strong estimate for sigma 0.01, 0.03, 0.1, and 0.3, respectively; the mechanism is therefore a real observed win in this setting.

Agent confidence: 9/10

Baseline: Strong local Grünwald fractional derivative: noise variances at sigma=.01/.03/.1/.3 = 1.20, 10.57, 117.54, 1069.14; coefficient RMSE = 0.1474, 0.4890, 1.7304, 10.3612; resolution norm ratios = 3.031.
Idea: Exact-transpose weak projection: noise variances = 1.10e-6, 9.76e-6, 1.14e-4, 9.51e-4; coefficient RMSE = 0.00733, 0.02056, 0.06643, 0.21873; resolution norm ratios = 0.503, 0.504, 0.506, close to predicted 0.5.

Limitations: No neural network was trained and no fractional PDE trajectory benchmark, data-fit tradeoff, FLOP comparison, validation trajectory error, GPU execution, or convergence study was tested. The strong-resolution scaling is specific to the implemented Grünwald discretization, and the coefficient experiment compares noise-induced error relative to each method's own noiseless estimate rather than absolute model-identification bias.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.12879) Sep 2, 2026

No significant difference vs. the tuned baseline (+0.5%, p=0.5236, wins 5 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0007
Idea mean: 0.0007
Effect (Δ): 0 (+0.5%; negative = idea better)
Wins: 5 / 8 paired seeds
p-value: 0.5236 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±2.0%
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 24 configs

Limitations:

Only the registered built-in actuated-pendulum dynamics track was tested. No fractional-PDE custom track, explicit noisy-observation sweep, larger model/data regime, rollout metric beyond harness MSE, FLOP/speed comparison, or alpha sweep beyond the fixed alpha=0.8 idea setting was evaluated.

How to run: python3 stage2_bench.py

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

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented adjoint-weak fractional residual training on the registered dynamics track with the shared rnn_small architecture and identical 8-seed, 16-epoch budgets. The tuned baseline achieved mean test MSE 0.0006985420 at lr=0.01, while the idea achieved 0.0007022070 at lr=0.01, for paired delta_mean +3.66498e-06 and permutation p=0.52365. The trained-model mechanism signature was confirmed directionally (weak correlation 0.9372 versus strong correlation 0.7030), but the idea did not produce a significant task-metric win.

Baseline: Registered dynamics/rnn_small; lr sweep {0.001, 0.003, 0.01}; best lr=0.01; 8-seed mean test MSE=0.0006985420404816978, std=0.00016862358035770679.
Idea: Adjoint-weak residual with alpha=0.8; same registered dynamics/rnn_small model, paired seeds, epochs, and shared lr sweep; best lr=0.01; 8-seed mean test MSE=0.0007022070203674957, std=0.00016781088891265367.
Paired delta mean: 3.6649798857979476e-06
Permutation p value: 0.52365

Benchmark result