△ Mechanism confirmed, baseline not beaten 2026

Mittag-Leffler Memory Diagnostic and Gating

Implementation & benchmark of arXiv:2607.08588 — Mittag-Leffler-Type Forecast-Error Growth as a Diagnostic Indicator of Fractional Dynamics

Usefulness7/10

Difficulty6/10

Novelty7/10

Architecture Accuracy Stability Sample efficiency Ssm Rnn Optimizer Training dynamics Dynamical systems Statistics Stochastic processes Pde Numerical analysis

Source paper: Mittag-Leffler-Type Forecast-Error Growth as a Diagnostic Indicator of Fractional Dynamics arXiv:2607.08588 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Measure how validation forecast error grows with prediction horizon and fit exponential and Mittag-Leffler models. When the Mittag-Leffler fit is decisively better, activate a fractional-memory SSM or long-memory residual branch and use its fitted effective order to set the branch's kernel decay and horizon-loss weights; otherwise retain a conventional recurrent or finite-memory branch.

Formulas

$$E_\alpha(z)=\sum_{k=0}^{\infty}\frac{z^k}{\Gamma(\alpha k+1)},\qquad e_{\exp}(h)=A e^{\lambda h}+C,$$

$$e_{\mathrm{ML}}(h)=A E_\alpha(\lambda h^\alpha)+C\quad\text{or, for stable decay,}\quad e_{\mathrm{ML}}(h)=A E_\alpha(-\lambda h^\alpha)+C,$$

$$q(h)=\frac{d\log(e(h)-C)}{d\log h}=\frac{h\,e'(h)}{e(h)-C},$$

$$\hat{\alpha},\hat{\lambda},\hat A,\hat C=\arg\min_{\alpha,\lambda,A,C}\sum_{h\in\mathcal H}\left[e_{\mathrm{val}}(h)-A E_\alpha(\lambda h^\alpha)-C\right]^2.$$

Mathematical statement

For a scalar observable or neural predictor, let $e(h)$ be the mean or median forecast error at horizon $h$, and let $E_\alpha(z)=\sum_{k=0}^{\infty}z^k/\Gamma(\alpha k+1)$ be the one-parameter Mittag-Leffler function. Compare $e_{\exp}(h)=A\exp(\lambda h)+C$ with $e_{\mathrm{ML}}(h)=A E_\alpha(\lambda h^\alpha)+C$, where $A>0$ is an amplitude, $\lambda$ is a growth or decay rate, $C\geq0$ is an error floor, and $0<\alpha\leq1$ is an effective shape/order parameter. For stable relaxation use $e_{\mathrm{ML}}(h)=A E_\alpha(-\lambda h^\alpha)+C$ with $\lambda>0$. Select the fractional branch only when cross-validated RMSE or an information criterion for the Mittag-Leffler fit is lower by a preset margin and bootstrap confidence intervals for the improvement exclude zero. The paper's kNN contraction analogue is $e_{kNN}(h)=\operatorname{median}_{i}\|\hat{x}_{i+h}-x_{i+h}\|$, estimated from nearest-neighbor forecasts; its role is to diagnose whether errors contract, remain bounded, or grow. The transferable mechanism is the shape of $e(h)$ and its local logarithmic slope $q(h)=d\log(e(h)-C)/d\log h$, not an assertion that $\alpha$ exactly equals the physical fractional derivative order.

Implementation notes

1. Exact integration point: train a two-branch sequence predictor. Branch B is a standard GRU or diagonal SSM; branch F is a fractional-memory SSM whose state input uses a truncated power-law convolution, $m_t=\sum_{j=0}^{L-1}w_j u_{t-j}$ with $w_j\propto(j+1)^{\alpha-1}$, followed by a learned linear state update. Start with both branches available, but stop gradients through the diagnostic gate during each diagnostic window. 2. Pseudocode: collect validation errors for horizons $\mathcal H=\{1,2,4,8,16,32,64\}$; compute $e(h)$ over identical examples; fit exponential and Mittag-Leffler curves by bounded nonlinear least squares with $0.1\leq\alpha\leq1.5$, $A\geq0$, unrestricted $\lambda$, and $C\in[0,\min_h e(h)]$; bootstrap examples to obtain $\Delta=\mathrm{RMSE}_{\exp}-\mathrm{RMSE}_{\mathrm{ML}}$; if $P(\Delta>0)>0.95$, set gate $g=1$ and use the fractional branch, otherwise set $g=0$ and use the conventional branch. For training, weight horizon $h$ by $w_h\propto1/[e_{\mathrm{ML}}(h)+\epsilon]$ for stable tasks or by $w_h\propto e_{\mathrm{ML}}(h)^{-1/2}$ for growing-error tasks, then normalize $\sum_h w_h=1$. 3. Computed versus estimated: evaluate $E_\alpha$ numerically using a special-function library or truncated series; estimate all curve parameters, the error floor, gate, and convolution scale from held-out errors. Refit every 1,000 to 10,000 updates to avoid noisy gating. 4. First cheap experiment: generate 10,000 sequences from a scalar Caputo relaxation process and a fractional chaotic benchmark at $\alpha\in\{0.5,0.8,1.0\}$, and train equal-size GRU, diagonal SSM, and gated two-branch models. Use a standard GRU as baseline and reserve 20% of trajectories for diagnostics. The falsifiable signature is that fractional data produce lower Mittag-Leffler fit error with bootstrap $P(\Delta>0)>0.95$, whereas integer-order data do not; the gated model should improve 64-step rollout RMSE by at least 10% only in the fractional regime. As the generating order approaches $\alpha=1$, the fitted effective order should approach one and the fractionality gate should switch off. A failure is no separation in fit quality or activation on integer-order data above 5% of trials.

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 8/10

Built a Mittag-Leffler memory diagnostic MVP with numerical evaluation, bounded exponential/ML fitting, bootstrap-style gating, and saved results. The three mechanism checks matched predictions: short-time slope tracked alpha, half-life scaled as lambda^(-1/alpha), and long-time relaxation slope tracked -alpha; the gate activated for alpha=0.5/0.8 with 22.0%/24.8% relative fit gains and correctly rejected alpha=1.0 at a 2.9% gain. This demonstrates the diagnostic mechanism, but not an end-to-end neural SSM accuracy win.

Agent confidence: 8/10

Baseline: Exponential-fit validation RMSE: alpha=0.5 0.012586, alpha=0.8 0.013642, alpha=1.0 0.010481
Idea: Mittag-Leffler-fit validation RMSE: alpha=0.5 0.009818 (21.99% gain, gate=1), alpha=0.8 0.010265 (24.76% gain, gate=1), alpha=1.0 0.010175 (2.92% gain, gate=0). Short-time slopes observed/predicted: alpha=.25 0.2497/.25, .5 0.5000/.5, .8 0.8000/.8, 1 1.0000/1. Half-life lambda slopes observed/predicted: alpha=.5 -2.000/-2, .8 -1.250/-1.25, 1 -1.000/-1. Tail q observed/predicted: alpha=.4 -0.399/-0.4, .6 -0.600/-0.6, .8 -0.800/-0.8.

Limitations: No GRU/SSM branch was trained, no real trajectory dataset or 64-step neural rollout was evaluated, and the bootstrap used repeated synthetic noisy validation curves rather than independent held-out trajectories. The evaluator is specialized to stable negative-argument Mittag-Leffler relaxation and is not a general high-precision implementation for all complex or positive arguments.

How to run: python3 ml_memory_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.08588) Sep 2, 2026

Ahead of the baseline (−0.1%) but not statistically significant (p=0.9929, wins 2 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0004
Idea mean: 0.0004
Effect (Δ): -0 (−0.1%; negative = idea better)
Wins: 2 / 8 paired seeds
p-value: 0.9929 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±11.1%
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

Tested only the built-in 8-step dynamics track, 10 epochs, and a small alpha/lr sweep; no 64-step rollout, explicit validation-error curve fitting, bootstrap gate refitting, sequence track, or larger training budget was evaluated.

How to run: python3 bench_mittag_gate.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: 9/10

Implemented the Mittag-Leffler-inspired fractional-memory RNN on the structurally matched controlled-pendulum dynamics track, using the same GRU/head, paired seeds, equal learning-rate sweep, and canonical trainer. The trained model retained measurable sensitivity to the earliest lag (0.12893), confirming the proposed long-memory mechanism, but the task metric showed essentially no improvement: delta_mean=-4.42e-7 with permutation p=0.9929. Therefore the idea did not produce a significant benchmark win.

Baseline: Best lr=0.006; 8-seed mean MSE=0.0003969427780, std=0.0000941708275
Idea: Best lr=0.006, alpha=0.5; 8-seed mean MSE=0.0003965004107, std=0.0000824931315; paired delta=-0.0000004423673; p=0.9929; 2/8 per-seed wins
Mechanism signature: Trained fractional model earliest-lag prediction sensitivity=0.1289317906; predicted nonzero long-lag influence; confirmed=true

Benchmark result