Method and evidence
We partition a finite training set into strata of masses $\pi_j$ and allocate positive integer batch counts $k_j$ with total $B$. Write $q_j=k_j/B$. The conditional gradient variance multiplied by $B$ is $\sum_j\pi_j^2v_j/q_j$. We minimize the proxy version of this energy plus $\tau D_{\mathrm{KL}}(q\|u)$, with $u$ uniform over strata. Thus larger allocation entropy lowers the objective. The count optimizer is exact by a diminishing-returns allocation proof. Weighting each gradient by $\pi_j/k_j$ preserves the original objective in conditional expectation. Variance estimates reuse the preceding forward passes.
The free-energy decomposition adapts the relative-entropy ledger in Replication and Persistence in Spacetime to computational allocation. The mathematical paper gives the covariance decomposition, a bound for proxy error, and sufficient conditions for an expected energy decrease. The artifact includes the executable PyTorch adapter, exact allocation tests, dataset input checksums, reproducible benchmark scripts, and complete per-seed traces.
Speedup theorems
A nonlinear two-layer ReLU example has an invariant random initialization basin. On known replicated strata, one gradient per stratum reproduces the full-gradient trajectory, using $mt$ versus $Nt$ gradient examples for $t$ updates. We give a centered random initialization for which convergence holds almost surely and a positive-probability basin under Gaussian initialization. The proof also covers independent trainable input and output weights in an active region. A second construction with nonduplicate finite regression records gives an arbitrarily large asymptotic exposure ratio against deterministic-step uniform minibatch SGD. A rare-label theorem applies to adaptive algorithms with iid uniform queries.
Benchmark results
We ran six methods from three matched random starts on 26 tasks. FEAS-Adam reached the validation target in at least two seeds on 24 tasks and full GD on 11. Relative to uniform Adam, FEAS-Adam used fewer gradient examples in seven of 23 jointly reached tasks and tied on 12. An exact-crossing timing audit replays the same trajectories with validation outside the timed path. In the audit, single-tensor full GD is faster than FEAS-Adam on seven of eight jointly reached tasks. The paper explains the effect using update counts, quadratic curvature, and stage timings. In the matched CNN comparisons, FEAS-Adam is faster than uniform Adam on all three tasks, and fixed stratification is faster than FEAS on two. The source includes separate label-pool and extended GD runs.
A second experiment adapts the batch-size and peak-step calibration used in e-means to neural training. It tests the calibrated cosine policy against a fixed step at the same peak, uniform sampling under the same schedule, and a full-gradient control. Results and calibration cost are reported separately from the original protocol. The manuscript gives a batch-size and step-size law for the expected one-step descent bound under a local affine cost model.
The complete manuscript with proofs, tables, and figures is attached as a PDF. The editable source and reproducibility files are attached as a ZIP. The manuscript uses an anonymous byline. Generative AI assisted drafting, implementation, and verification.