The Arbitrary-Placement Problem in Entropy-Minimizing Selection, and a Residual-Entropy Formulation

Alyssa H. Shin; Claire H. Shin

cs.LGOct 5, 2026

The Arbitrary-Placement Problem in Entropy-Minimizing Selection, and a Residual-Entropy Formulation

Authors: Alyssa H. Shin, Claire H. Shin

Organizations: Division of Biology and Biological Engineering California Institute of Technology · Robert Frederick Smith School of Chemical and Biomolecular Engineering Cornell University

Abstract

Entropy-based selection objectives suffer from a fundamental degeneracy: minimizing Shannon entropy $H(p_A)$ rewards confident selection regardless of whether the selected candidate is informative. We address this limitation with the residual entropy $D = H(p_A) - H(p_β)$ , where $p_β$ is induced by candidate trust weights. We prove the exact identity $D = -\mathrm{KL}(p_A\Vert p_β) - Δ$ , where $Δ$ measures whether the score-induced distribution and trust profile favor the same candidates. Boundary cases establish basic safety: under uniform trust, $D\leq0$ automatically, so an equal-trust, non-starving state is never penalized, while at any one-hot limit, $D\to0$ regardless of the selected candidate. For the intermediate regime where selection occurs, we prove that $D\leq0$ when candidate ordering by trust agrees pairwise with ordering by informativeness, and derive a tighter certificate based on the leading candidate's margin over its competitors. These results are independent of the candidate-scoring function and apply to both stationary and dynamically changing information. Experiments with a gradient-based mixture-of-experts router confirm that the ordering conditions can hold during real optimization and show that correct ordering improves downstream performance when candidates are non-interchangeable and selections are used directly rather than averaged. Beyond routing, margin-based reweighting matches or outperforms fixed-strength baselines in a class-imbalance task, while informative selection in a production video-prediction system reduces MSE by approximately 20 $\%$ and transfers to a related species. Residual entropy, therefore, provides a safety criterion for selection and a usable signal for deciding when that selection is informative.

Figures & tables

Figure 1: From entropy degeneracy to residual-entropy correction. Propositions 1–2 characterize the problem and correction and Corollaries 1–4 provide boundary results and sufficient sign certificates.

Study / target	Entropy alone / alternatives	Residual entropy
(i) Initialization: objective separation	0.0047 / 0.0050; ratio 0.93	$-0.3641$ / $-0.0134$ ; ratio 27.11
(ii) Functional form: scoring / normalization	Ratios 0.64–0.93	Rational / log: 27.11 / 23.67; sigmoid / softmax: 1.00 / 1.00
(iii) Alternatives: separation / collapse	Penalty ratios: 0.62–0.88; KL variant: 90/90 retain initial window	$-0.3641$ / $-0.0134$ ; lower objective at informative initialization
(iv) Class imbalance: identification	Selection rate 0.775	0.938 ( $1.21\times$ ; $p=0.003$ )
(v) Full-batch MoE: participation boundary	5/6 lower-bounded at floor $>0$ ; no transition	Threshold 0.07–0.10; 0/6 below, 4/6 above
(vi) Streaming MoE: boundary robustness	5/6 lower-bounded at floor $>0$ ; no transition	Threshold 0.062–0.070; 0/6 below, 4/6 above

Table 1: Seven studies: questions and quantitative evidence. Objective pairs are informative / elsewhere initialization means.

Setting	Uniform	Fixed $5\times$	Conf.-scaled	Hurt (fixed / conf.)	$p$ (conf. vs. fixed)
Digits $T{=}5$ , 8% ( $n{=}80$ )	0.8466	0.8727	0.8732	2/80 / 0/80	0.159
Digits $T{=}5$ , 2% ( $n{=}15$ )	0.6776	0.6909	0.7018	5/15 / 0/15	0.023

Table 2: Confidence-modulated reweighting at both scarcity levels: mean held-out minority-class recall after retraining with each oversampling scheme. “Hurt” counts seeds below uniform training; the final column compares confidence-scaled with fixed

5\times

weighting. The same

s{=}0.25

is used.

Noise	Threshold	Clamped acc.	Peaked acc.	Winner
0.0	0.10–0.12	0.960	0.964	peaked ( $+0.3$ pt)
0.5	0.09–0.10	0.924	0.918	clamped ( $+0.6$ pt)
1.0	0.07–0.08	0.803	0.783	clamped ( $+2.0$ pt)
1.5	0.02–0.05	0.660	0.645	clamped ( $+1.5$ pt)
$\geq 2.0$	vanished	–	$0.37$ – $0.54$	n/a (single regime)

Table 3: Floor threshold and accuracy vs. task noise (study (v), residual, fair protocol,

n{=}15

seeds/noise level). “Threshold” separates clamped (

\gamma_{f}\approx 0

) from peaked (

\gamma_{f}\approx 4

); “Winner” has higher test acc. (Appendix D.2 ).

Figure 2: Real-pipeline architecture. An encoder-only upstream model identifies a sweet-spot window among five candidates. That window is then oversampled when training a separate encoder-decoder predictor. The exact upstream score is derived in Appendix F.1 .

Arm	Mean MSE	SD
sweetspot ( $\mu^{\ast}$ , ours)	92.4	82.0
uniform (no oversampling)	104.0	82.5
mc_var (BALD-style)	106.5	91.7
random	116.1	85.3
raw_var	139.9	115.6

Table 4: Bacillus, all five windows pooled (13 evaluation points: 5 held-out synthetic + 5 video1 + 3 video2 windows; lower MSE is better).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Synthetic landscape: objective is blind to real signal under $H(p_{A})$ alone, but not under the residual form.

Family	Objective	obj(info.)	obj(else.)	ratio	$\gamma$ (i/e)
Rational (main)	$H(p_{A})$ alone	+0.0047	+0.0050	0.93	7.43/7.87
Rational (main)	$H(p_{A})-H(p_{\beta})$	$-0.3641$	$-0.0134$	27.11	0.88/6.58
Sigmoid	$H(p_{A})$ alone	+1.1109	+1.3517	0.82	6.52/6.37
Sigmoid	$H(p_{A})-H(p_{\beta})$	$-0.2502$	$-0.2502$	$1.00$	0.16/0.16
Log	$H(p_{A})$ alone	+0.0042	+0.0046	0.91	7.33/7.74
Log	$H(p_{A})-H(p_{\beta})$	$-0.4139$	$-0.0175$	23.67	0.86/7.05

Appendix

Table 5: Full functional-form and aggregation-scheme generality table (study ii).

Objective	obj(info.)	obj(else.)	ratio	$\gamma$ (i/e)
$H(p_{A})$ alone (reference)	+0.0047	+0.0050	0.93	7.43/7.87
$H(p_{A})-H(p_{\beta})$ (ours, reference)	$-0.3641$	$-0.0134$	27.11	0.88/6.58
$H(p_{A})+1.0\,(\gamma/\gamma_{\max})^{2}$	+0.1004	+0.1137	0.88	5.27/5.65
$H(p_{A})$ , tighter $\gamma$ clamp ( $\gamma_{\max}{=}2.0$ vs. $20.0$ )	+0.3830	+0.4543	0.84	2.00/2.00 ‡
$H(p_{A})+5.0\sum_{t}(p_{\beta,t}-1/T)^{2}$	+1.3691	+2.1906	0.62	0.15/5.51
$H(p_{A})-\mathrm{KL}(p_{A}\\|p_{\beta})$	+0.0040	+0.0048	0.83	7.56/7.91

Appendix

Table 6: Full alternative-regularizer table (study iii), including the KL-divergence variant. Same landscape and protocol as Table 5 (30 seeds

\times

\mu

-inits).

Two-proportion $z$ -test on raw selection rate: $z=2.929$ , $p=0.003$ (significant)
Phase 1 (outer loop)	$H(p_{A})$ alone	Residual (ours)
Relative selection improvement (true minority class)	$1.00\times$ (0.775)	$1.21\times$ (0.938)
Phase 2 (downstream)	Minority recall	Accuracy
Uniform (no emphasis)	0.8466	0.9635
Residual-selected class, fixed $5\times$	0.8727	0.9687
$H(p_{A})$ -selected class, fixed $5\times$	0.8670	0.9675

Appendix

Table 7: Digits (

T{=}5

; minority class retained at 8%, approximately 10 training images; 80 seeds and five

\mu

-initializations per objective). Phase 1 reports the true-minority selection rate relative to

H(p_{A})

alone. Phase 2 reports minority recall and overall accuracy after separately retraining with a fixed

5\times

weight on the class selected by each objective. The 2% variant appears in Appendix C.3 .

floor	lower bdd. (of 6)	mean $\gamma_{f}$	lower bdd. (of 6)	mean $\gamma_{f}$	clip effect
	$H(p_{A})$ alone		Residual (ours)
$0.00$	$4.00$	$6.76$	$0.00$	$0.01$	literal freeze
$0.05$ – $0.07$	$5.00$	$6.77$	$0.00$	$0.01$	trickle
$0.10$ – $0.30$	$5.00$	$6.79$	$4.00$	$\approx 4.06$	trickle
$1.00$	$5.00$	$6.82$	$4.00$	$4.08$	trickle

Appendix

Table 8: Study (v), floor sweep under the fair protocol (one

\mu

-init per candidate window, best-of-6 by lowest final objective, 9 seeds per floor,

E_{\text{TRUE}}=2

, noisy hard-task variant). An expert is lower-bounded when its learning-rate multiplier

\mathrm{lr\_mult}_{e}=\mathrm{clip}(\beta_{e}/\bar{\beta},\,\mathrm{floor},\,5.0)

sits at the lower bound, i.e.

\mathrm{lr\_mult}_{e}=\mathrm{floor}

. Only floor

=0

is literal starvation (exact zero gradient). For floor

>0

these experts still train, rate-capped at floor.

Noise	static_low	static_high	low $\to$ high@75	high $\to$ low@75	Winner
$0.0$	$0.957$	$0.963$	$0.959$	$0.963$	high ( $+0.5$ pt)
$0.5$	$0.922$	$0.916$	$0.922$	$0.916$	low ( $+0.5$ pt)
$1.0$	$0.804$	$0.783$	$0.804$	$0.783$	low ( $+2.1$ pt)
$1.5$	$0.644$	$0.643$	$0.644$	$0.643$	low ( $+0.1$ pt)
$2.0$	$0.537$	$0.537$	$0.537$	$0.537$	tie
$2.5$	$0.442$	$0.441$	$0.442$	$0.441$	low ( $+0.1$ pt)

Appendix

Table 9: Study (v), floor-adjustment strategy comparison across seven noise levels (residual objective, fair protocol,

n{=}25

per cell). floor

{}_{\text{low}}=0.05

, floor

{}_{\text{high}}=1.0

. Both schedules switch at step

75

150

	Lower bdd. (of 6)	Ensemble test acc.
$H(p_{A})$ alone	5.00	$0.9628\pm 0.0092$
Switch-style aux ( $\alpha{=}0.01$ )	5.00	$0.9631\pm 0.0094$
Residual (ours)	0.00	$0.9570\pm 0.0103$

Appendix

Table 10: Study (v) extended with a Switch-Transformer-style load-balancing loss (App. D.3 ), 6-candidate fair protocol (25 seeds, best-of-6 by lowest final objective), same protocol as Table 8 .

$\alpha$	Lower bdd. (of 6)
0.01	5.00
0.1	5.00
1.0	0.00
10.0	0.00

Appendix

Table 11: Load-balancing coefficient sweep, 6-candidate fair protocol (10 seeds each, best-of-6 by lowest final objective), floor

=0.05

(never

0

, so counts below are lower-bounded, not literal starvation). Lower-bounding is fixed monotonically by a large enough

\alpha

floor	lower bdd. (of 6)	mean $\gamma_{f}$	lower bdd. (of 6)	mean $\gamma_{f}$	clip effect
	$H(p_{A})$ alone		Residual (ours)
$0.000$	$4.00$	$6.76$	$0.00$	$0.01$	literal freeze
$0.062$	$5.00$	$6.77$	$0.44$	$0.46$	trickle
$0.070$	$5.00$	$6.77$	$3.56$	$3.61$	trickle
$0.100$ – $1.000$	$5.00$	$6.78$ – $6.82$	$4.00$	$\approx 4.07$ – $4.08$	trickle

Appendix

Table 12: Study (vi), floor sweep under the fair protocol (one

\mu

-init per candidate window, best-of-6 by lowest final objective, 9 seeds per floor,

E_{\text{TRUE}}{=}2

, noisy hard-task variant): lower-bounded-candidate count and mean converged

\gamma_{f}

, both objectives. Same lower-bounded/literal-starvation distinction as Table 8 (floor

=0

only).

	$H(p_{A})$ alone	Residual (ours)
Default (floor $=0.05$ , clean)	$0.960\pm 0.008$	$0.960\pm 0.010$
Hard variant (floor $=0$ , noisy)	$0.756\pm 0.016$	$0.805\pm 0.017$
Floor $=1$ (clean)	$0.960\pm 0.008$	$0.962\pm 0.007$

Appendix

Table 13: Study (vi), fair protocol (one init per window, best-of-6 by lowest final objective, 9 seeds). Same three floor/noise conditions as study (v)’s corresponding ablation (Appendix D.2 ).

gap	mean diff	95% CI	$p$	$d_{z}$
$0.00$	$-0.0029$	$[-0.0056,-0.0003]$	$0.042$	$-0.30$
$0.05$	$+0.0069$	$[+0.0035,+0.0103]$	$0.0003$	$+0.55$
$0.10$	$+0.0144$	$[+0.0108,+0.0180]$	$<10^{-4}$	$+1.08$
$0.20$	$+0.0358$	$[+0.0314,+0.0404]$	$<10^{-4}$	$+2.20$
$0.35$	$+0.0756$	$[+0.0704,+0.0810]$	$<10^{-4}$	$+3.88$
$0.50$	$+0.1414$	$[+0.1359,+0.1470]$	$<10^{-4}$	$+7.04$

Appendix

Table 14: Ordering payoff as a function of candidate interchangeability: invest in the chosen expert, then deploy it. Results are paired within run. 25 seeds per cell, floors pooled (

n{=}50

d_{z}

is the paired effect size relative to its across-seed variation. Values around 1 or larger are conventionally large.

Domain	Domain distance	Result
Bacillus (native)	native	$\sim$ 19–20% MSE improvement over mc_var ; 66/90 pairwise wins (5-seed re-validation); Table 4
Tetraselmis sp. (microalga)	diff. species, same modality	unanimous vs. uniform (3/3 seeds, 6/6 videos); only $+2.5\%$ , 4/6 vs. mc_var (1 seed)
DIC-C2DH-HeLa (mammalian)	diff. organism & modality	mixed vs. uniform (2/3, n.s., $p{=}0.36$ ); mc_var wins vs. sweetspot , $-5.2\%$ , 1/6 (1 seed)

Appendix

Table 15: Downstream case study: does the correctly-identified sweet-spot window improve a separately-developed video predictor?

Arm	W1	W2	W3	W4	W5
uniform	208.3	66.7	219.6	265.2	189.7
sweetspot ( $\mu^{\ast}{=}$ W3)	231.9	53.8	221.3	217.2	171.9
mc_var ( $=$ W2)	246.5	80.1	200.6	309.9	150.9

Appendix

Table 16: Study (vii) downstream results on the Bacillus held out synthetic subset (Tier A), reporting per window MSE for three arms.

Explore similar work

Jun 9, 2026stat.ML

Annealed Entropic Allocation for Ranking and Selection

We propose annealed entropic allocation, an adaptive sampling policy based on an annealed, weighted soft-min formulation of static budget allocation. We replace the maximin large-deviation rate objective with a weighted log-sum-exp surrogate that blends challenger-specific pairwise scores through soft-min weights, avoiding hard switching when several challengers are nearly active. To capture tail behavior beyond the leading exponent, the surrogate incorporates saddlepoint prefactors from refined pairwise tail asymptotics. Because these corrections are subexponential, decreasing the annealing temperature with the budget preserves the same first-order target allocation. For the static problem, we prove uniform convergence to the hard minimum, concentration of soft-min weights on active challengers, and continuity of the induced target-allocation map under fixed weights. Experiments show that the proposed methods are consistently competitive: the no-saddlepoint ablation performs best in symmetric Gaussian and exponential slippage settings, while saddlepoint weighting can help in heterogeneous or asymmetric cases.

Xin Fei, Juergen Branke

Business School · The University of Edinburgh, Edinburgh, UK · Warwick Business School +1

Minimax Rate Token Budget Allocation Ranking

Ask PDF

Jun 20, 2026cs.LG

Ranking-and-Selection with Multiple Correct Answers and Non-Answerable Estimates

We study fixed-precision ranking-and-selection in structured settings where the answer may be non-unique and where noisy estimates may temporarily admit no valid answer at all. This phenomenon arises naturally in problems such as multi-fidelity ranking-and-selection and identifying a Condorcet winner from pairwise comparisons. To address this, we propose a unified framework based on answer-wise acceptance sets, restricted generalized likelihood ratio stopping, and an answer-pitfall decomposition that yields a max-max-min characteristic value and a common sampling principle. We introduce ENDS, a general procedure that combines estimation, nomination, pitfall detection, and cost-aware information-directed selection. We instantiate ENDS for various problems by deriving explicit formulas. Extensive numerical experiments show that this unified recipe performs well across a broad range of pure-exploration problems and offers a practical framework and proof-of-concept algorithmic recipe.

Qiaoqiao Wang, Wei You

Dept. of Industrial Engineering and Decision Analytics, The Hong Kong University of Science and Technology, Hong Kong, China

Selection Bias Ranking Finite-Sample Answer Stochastic Exploration Acceptance+1

Ask PDF

Apr 26, 2026cs.AI

Non-Stationarity Breaks Permutation Surrogates in Multi-Agent Reinforcement Learning: Diagnosis and Remedies

Reporting guidance for information-theoretic measures is rarely tested against ground truth. We test one guardrail in two multi-agent reinforcement learning games, a social dilemma and a coordination race, where directed influence between selected agent pairs is zero by construction, over 100 seeds. Omitting one precondition, exclusion of the non-stationary training transient, gives false-positive rates of 100.00% and 99.95%: agents annealing exploration independently, in runs that never met, are flagged as influencing one another. Excluding the transient reaches 3.0% in the social dilemma but 11.8% in the coordination game, which stationarity tests explain: 95.7% of social-dilemma series are stationary afterwards against 56.8% of coordination series. So the non-stationarity must be treated, and exclusion is neither the only way nor sufficient. What we recommend instead changes the null model rather than the data: permuting the source within blocks of training time reaches 5.25% and 5.50%, the only one of four constructions at the size of the test in both games, leaving series, statistic and estimand untouched. Titrating injected links of known strength in both games shows it is also the most sensitive of the three, detecting 89.0% in the coordination game where conditioning detects 61.0% on identical pairs, while the ablated test reports 100% with or without a link, so its apparent sensitivity is uninformative. The block count is not critical: every setting from 16 to 256 lands in the nominal region, and a partition derived from the stationarity test removes the parameter, though less sensitively. Code and data are released.

Nikolaos Al. Papadopoulos, Konstantinos E. Psannis

Department of Applied Informatics, University of Macedonia, Egnatia 156, 54636 Thessaloniki, Greece

Information-Theoretic Limits Entropy Intelligence Measures Decisions

Ask PDF