Pith. sign in

REVIEW 4 minor

Diffusing weighted particles by heat flow yields a geometry-aware effective sample size that ordinary ESS cannot see.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 23:13 UTC pith:MROX4JOZ

load-bearing objection Clean, usable multiscale geometric ESS from heat-kernel Rényi-2; theory is solid within its stated compact-manifold scope, experiments are illustrative not inferential.

arxiv 2607.06696 v3 pith:MROX4JOZ submitted 2026-07-07 stat.ML cs.LGstat.ME

Heat-Kernel Entropy Profiles and Geometric Effective Sample Size for Weighted Measures on Manifolds

Kisung You , Boram Cho This is my paper
classification stat.ML cs.LGstat.ME MSC 62R3058J3562G07
keywords heat-kernel entropy profilesgeometric effective sample sizeweighted empirical measuresRényi entropycompact Riemannian manifoldsspherical harmonicsimportance sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Ordinary effective sample size only looks at weight labels and treats every particle as fully distinct, even when two particles sit on top of each other or form tight clusters. This paper replaces that label-only count with a multiscale profile obtained by letting the weighted atoms diffuse under the manifold heat equation and measuring how nonuniform the smoothed density remains at each scale. For the order-two Rényi entropy the profile collapses to a simple quadratic form of pairwise heat-kernel overlaps, so it can be computed from a Gram matrix. Normalizing those overlaps by the self-overlap of a single heat kernel produces a geometric effective sample size that automatically merges nearby or duplicate atoms while recovering ordinary ESS when particles are well separated. On spheres the same unlogged profile further decomposes into spherical-harmonic energies that recover classical mean-direction and Bingham summaries and expose higher-order structure that first-moment diagnostics miss.

Core claim

On any compact boundaryless manifold the order-two heat-kernel Rényi profile of a weighted empirical measure is exactly the inverse of the sum of pairwise heat-kernel overlaps, is monotone under diffusion, and yields a geometric ESS that is invariant to exact atom splitting, equals ordinary ESS for well-separated particles, and admits explicit small-time volume/curvature expansions and large-time spectral expansions.

What carries the argument

The pairwise heat-kernel identity: the squared L2 norm of the heat-smoothed density equals the quadratic form w^T K_{2t} w of the heat-kernel Gram matrix, which simultaneously defines both the entropy profile and the normalized geometric ESS.

Load-bearing premise

Everything is proved only for compact connected manifolds without boundary, and statistical rates hold only away from the zero-time singularity where heat kernels blow up.

What would settle it

On the sphere, take equal-weight exact duplicates at 30 well-separated locations; if the geometric ESS at a fine angular scale of 15 degrees fails to recover approximately 30, or if an antipodal two-mode cloud produces a flat profile indistinguishable from uniformity, the central claim is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 4 minor

Summary. The paper introduces heat-kernel Rényi-2 entropy profiles for weighted empirical measures on compact, connected, boundaryless Riemannian manifolds. Diffusing the atoms by the intrinsic heat semigroup and measuring nonuniformity of the smoothed density yields a multiscale effective-volume curve U2(t) that is exactly the reciprocal of the quadratic form of pairwise heat-kernel overlaps (Theorem 1). Normalizing those overlaps by local self-overlaps produces a geometric effective sample size gESS_w(t) that is merge-invariant, lies between 1 and ordinary ESS, recovers ordinary ESS for well-separated atoms, and collapses to 1 at large scales (Theorem 3). Small-time parametrix expansions give volume/curvature asymptotics and a two-atom merging law (Theorem 4, Proposition 5); large-time spectral expansions recover the lowest eigenmode energies, which on spheres specialize to mean-resultant and Bingham-type quantities (Theorem 6, Corollary 7). Stability under W1 and rates for deterministic and bounded-ratio self-normalized importance weights are proved on scale intervals bounded away from zero (Theorem 9). Sphere experiments and a public code repository illustrate that the profiles detect antipodal, girdle, multimodal and duplicate structure missed by weight-only ESS and first-moment summaries.

Significance. If the results hold, the work supplies a geometrically intrinsic, multiscale replacement for ordinary ESS that is computable from heat-kernel Gram matrices and comes with clean small- and large-time interpretations. The pairwise identity, merge-invariance, and spectral recovery of classical directional moments are non-trivial and useful for importance sampling, particle methods and spherical representation learning. Strengths include fully written appendix proofs that rely only on standard heat-kernel tools, an explicit algorithm, reproducible synthetic constructions, and a public code link. The deliberate restriction to compact boundaryless manifolds and to scales bounded away from zero is stated transparently and does not undermine the central claims within that scope.

minor comments (4)
  1. In the abstract and introduction the phrase “order-two Rényi entropy” is used interchangeably with the Rényi-2 divergence from uniformity; a single clarifying sentence would avoid any ambiguity for readers who expect the absolute entropy rather than the divergence.
  2. Figure 1 caption and the experimental section both introduce the angular scale α=√(8t); stating the convention once in the main text (near Proposition 5) would make the figures self-contained.
  3. The discussion correctly notes that gESS need not be monotone on non-homogeneous manifolds; a short remark or numerical illustration on a non-homogeneous example would make that caveat more concrete.
  4. A few typographical inconsistencies appear (e.g., “R´ enyi”, “Garc´ ıa-Portugu´ es”); a light copy-edit pass would clean them.

Circularity Check

0 steps flagged

No significant circularity: definitions of heat-kernel Rényi-2 profiles and gESS are followed by independent derivations of their properties from standard heat-kernel analysis.

full rationale

The paper defines the order-two heat-kernel entropy profile and geometric ESS from the heat semigroup and pairwise overlaps (Theorem 1, Definition 2), then derives monotonicity, merge-invariance, small-time volume/curvature expansions, large-time spectral expansions, and W1/stability rates on compact boundaryless manifolds (Theorems 3, 4, 6, 9) via bilinearity, semigroup identities, Minakshisundaram–Pleijel expansions, and spectral representations that are proved in the appendix. Spherical specialization recovers classical B1/B2 moment energies by direct orthonormal-basis calculation (Corollary 7), not by redefinition. Experiments are descriptive profile demonstrations on synthetic and embedding data; no free parameters are fitted and then re-presented as predictions. Citations are to standard external literature (heat kernels, ESS, directional statistics, MMD). The construction is self-contained against its stated assumptions; no step reduces a claimed prediction to its own input by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central claims rest on standard Riemannian heat-kernel theory plus the modeling choice that the objects of interest are finite weighted atomic measures on compact boundaryless manifolds. No free parameters are fitted to produce the main theorems; experimental concentrations and truncation levels are simulation choices only. The two invented entities are definitional summaries rather than postulated physical objects.

free parameters (3)
  • heat-scale grid {t_k} / angular resolution α=√(8t)
    Chosen by the user for each profile evaluation; not fitted to data but required to produce any numerical curve.
  • spherical-harmonic truncation L_max (default 140)
    Numerical cutoff used in all S2 experiments; adequacy is checked in Appendix Figure 5 but remains a free computational parameter.
  • vMF concentration κ and girdle noise sd in synthetic examples
    Hand-chosen simulation parameters (κ=140 main, κ=100 replications) that control how strongly the stress-test configurations separate; not estimated from real data.
axioms (4)
  • domain assumption M is a compact, connected, m-dimensional Riemannian manifold without boundary, equipped with normalized volume measure ν.
    Stated at the opening of Section 3 and used for every theorem; excludes manifolds with boundary and non-compact spaces.
  • standard math The heat kernel k_t exists, is smooth for t>0, satisfies the semigroup property, and admits the Minakshisundaram–Pleijel diagonal expansion and Gaussian off-diagonal bounds.
    Invoked throughout Theorems 1, 4, 5 and the proofs in Appendix A; standard results from Rosenberg/Grigoryan.
  • domain assumption For statistical rates, the heat-kernel class on any interval [t0,T] with t0>0 is uniformly bounded and Lipschitz (Assumption 8).
    Required for the uniform-in-t consistency of Theorem 9; constants deteriorate as t0↓0.
  • domain assumption Self-normalized importance weights arise from a bounded density ratio ρ≤M.
    Explicit hypothesis of the final clause of Theorem 9; heavy-tailed weights are left for future work.
invented entities (2)
  • heat-kernel entropy profile (D2,P(t), U2,P(t)) no independent evidence
    purpose: Multiscale summary of nonuniformity of a weighted measure under intrinsic heat flow.
    Defined in Section 3 from the order-two Rényi divergence of the heat-smoothed density; independent evidence is the pairwise formula and the asymptotic theorems, not external measurement.
  • geometric effective sample size gESS_w(t) no independent evidence
    purpose: Scale-dependent effective number of geometrically distinguishable atoms obtained by normalizing heat overlaps by local self-overlaps.
    Definition 2; recovers ordinary ESS at fine scales for distinct atoms and equals 1 at infinite scale; no external physical referent.

pith-pipeline@v1.1.0-grok45 · 20939 in / 2939 out tokens · 39186 ms · 2026-07-10T23:13:42.639371+00:00 · methodology

0 comments
read the original abstract

Weighted empirical measures on compact manifolds appear in importance sampling, particle approximations, posterior summaries, quadrature, and representation learning. Ordinary effective sample size and related weight summaries ignore the geometry of the support. We introduce heat-kernel entropy profiles to measure nonuniformity after intrinsic diffusion at a range of scales. For order-two R\'enyi entropy, pairwise heat-kernel overlaps give an exact profile and a geometric effective sample size. This effective sample size discounts nearby or duplicate particles. It approaches ordinary effective sample size as overlaps between distinct particles vanish. On compact boundaryless manifolds, we establish profile monotonicity, gESS scale limits, deterministic-weight consistency, and a bounded-ratio result for self-normalized importance sampling. On spheres, the unlogged profile decomposes into spherical-harmonic energies. The first terms are squared mean-resultant and traceless-second-moment energies, which give vMF- and Bingham-type scalar summaries. Experiments identify antipodal, girdle, multimodal, and duplicate-particle structures that weight-only and first-moment summaries miss.

Figures

Figures reproduced from arXiv: 2607.06696 by Boram Cho, Kisung You.

Figure 1
Figure 1. Figure 1: Preview of the proposed profiles on represen [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The reported synthetic values correspond to the con￾structions described in Appendix B. All quantities are computed by the routine in Section 4. The only sphere-specific step is evaluating the heat Gram ma￾trix. The heat kernel is kt(x, y) = X∞ ℓ=0 (2ℓ + 1)e −ℓ(ℓ+1)tPℓ(x ⊤y), (14) where Pℓ is the Legendre polynomial. For a fixed weighted cloud, we compute the harmonic energies Bℓ = (2ℓ + 1)X i,j wiwjPℓ(x ⊤… view at source ↗
Figure 2
Figure 2. Figure 2: Mollweide projections of the five synthetic [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two-regime validation for a unimodal vMF cloud on [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Replicated synthetic profiles on S 2 . Each curve is the median over 30 replications, and the shaded region is the interquartile range. Left: geometry-aware effective sample size, which counts distinguishable atoms or clusters. Right: effective occupied volume fraction, which increases as heat flow removes nonuniformity. The plot checks that the qualitative profile shapes are stable across replications. sa… view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity of the S 2 heat-profile computation to harmonic truncation. The plot shows relative error in Abw(t) against a longer-series reference as a function of angular resolution. Smaller angular scales require more spherical-harmonic terms because high frequencies have not yet been damped by heat flow. Errors below 10−16 are clipped for log-scale visualization. 101 102 Angular Scale α (Degrees) 0 10 20… view at source ↗
Figure 6
Figure 6. Figure 6: Self-normalized importance-sampling stress test. Particles are drawn from the uniform proposal on [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Full profiles for the normalized-embedding example from Table 2. Left: geometry-aware effective [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.