Pith. sign in

REVIEW 3 major objections 4 minor 12 references

A signed kernel statistic reads which way a distribution shifted

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 · deepseek-v4-flash

2026-08-01 10:43 UTC pith:KKM6T6HF

load-bearing objection The DKMD construction and its O(N log N) prefix-suffix algorithm are real, but the paper's key robustness claim is contradicted by its own Table 2 and the consistency proof is missing; it deserves a serious referee, not acceptance. the 3 major comments →

arxiv 2607.20119 v1 pith:KKM6T6HF submitted 2026-07-22 stat.ML cs.LG

Directional Kernel Mean Difference: A Fast Signed Statistic for Univariate Distribution Comparison

Shijie Zhong , Jiangfeng Fu This is my paper
classification stat.ML cs.LG MSC 62G1062G20
keywords Directional Kernel Mean Differencesigned two-sample statistickernel mean embeddingMMDstochastic dominanceRiemann estimatorprefix-suffix algorithmheavy-tailed robustness
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.

This paper introduces a signed univariate statistic, the Directional Kernel Mean Difference (DKMD), that answers not just whether two distributions differ but in which direction. Instead of squaring the RKHS distance as MMD does, DKMD integrates the difference of kernel mean embeddings against a fixed odd weight function such as tanh(z/τ). The paper proves three structural properties: antisymmetry, immunity to symmetric distributional differences, and a strictly positive sign under first-order stochastic dominance. It also contributes a data-driven Riemann estimator claimed to converge to the continuous integral, and an O(N log N) prefix–suffix algorithm that makes the statistic practical at million-sample scale. If the claims hold, DKMD offers a robust, direction-preserving summary of distributional shift with bounded influence from outliers.

Core claim

DKMD(P,Q) is defined as ∫ w(z)(μ_P(z) − μ_Q(z)) dz, where w is a bounded, monotone, odd weight and μ_P, μ_Q are Laplacian kernel mean embeddings. By Fubini's theorem, this reduces to E_{X~P}[ψ_w(X)] − E_{Y~Q}[ψ_w(Y)], with the witness function ψ_w(x) = ∫ w(z)k(x,z)dz. The statistic is deliberately not a metric: it can be negative and can vanish for different distributions when the embedding difference is even. For strictly increasing odd weights and strictly positive translation-invariant kernels, the sign of DKMD is guaranteed to match first-order stochastic dominance, and the empirical Riemann estimator is claimed to inherit these guarantees asymptotically.

What carries the argument

The central object is the pair (k, w): a translation-invariant characteristic kernel (Laplacian, k(x,z)=exp(−|x−z|/σ)) and a bounded monotone odd weight (tanh(z/τ)). Their convolution defines the witness function ψ_w, whose strict monotonicity carries the sign guarantee of Theorem 3.3 and whose boundedness |ψ_w(x)| ≤ 2σ gives outlier robustness. The prefix–suffix scanning recurrences on sorted pooled samples compute all discrete witness values in O(N) after sorting, reducing the kernel sum from quadratic to O(N log N) time and O(N) memory.

Load-bearing premise

The load-bearing premise is the unproved consistency of the data-driven Riemann estimator: as N grows and max Δz_k shrinks, the right-Riemann sum must converge to the Lebesgue integral, including in heavy tails — yet the paper's own k=5 Cauchy experiment reports a DKMD standard deviation of 10.45, suggesting the discrete estimator's variance is not controlled by the continuous bound |ψ_w(x)| ≤ 2σ, which would break the finite-sample sign and robustness guarantees.

What would settle it

Run the paper's heavy-tailed contamination scenario (P=N(0,1), Q=N(0.3,1), five Cauchy outliers) across many independent trials. If the empirical DKMD standard deviation stays far above the theoretical bound |ψ_w(x)| ≤ 2σ — as the reported 10.45 suggests — then the finite-sample estimator is not inheriting the continuous robustness guarantee, and the claim of sign stability under heavy tails would need revision.

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

If this is right

  • The sign of DKMD tells which of the two distributions is stochastically larger, information that squared MMD discards.
  • Pure scale or other symmetric variations leave the statistic near zero, so it filters out directionally meaningless differences.
  • The bounded witness function limits the influence of heavy-tailed outliers that can arbitrarily flip the sign of the mean difference.
  • The O(N log N) algorithm makes kernel-based distribution comparison practical on datasets with millions of points.
  • At large sample sizes the empirical Riemann sum is claimed to recover the continuous integral and thereby preserve the three theoretical guarantees.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural next step the paper leaves open is calibrating a one-sided hypothesis test with permutation or bootstrap thresholds; the paper notes the null distribution remains an open question.
  • The signed framework could be extended to multivariate data by averaging DKMD over random projections or slices, a direction the authors explicitly flag as future work.
  • In drift-detection pipelines, DKMD's sign could serve directly as an alert direction (e.g., latency increased rather than merely changed), which most existing two-sample tests cannot provide.
  • One testable extension would be comparing DKMD's sign against the one-sided Kolmogorov–Smirnov statistic and the signed Wasserstein distance to map where the new statistic adds information beyond first-order trends.

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

3 major / 4 minor

Summary. The paper proposes the Directional Kernel Mean Difference (DKMD), a signed statistic for comparing univariate distributions, defined as the integral of the difference of kernel mean embeddings against an odd weighting function. It proves three structural properties: antisymmetry (Thm 3.1), symmetric immunity (Thm 3.2), and directional monotonicity under first-order stochastic dominance (Thm 3.3). A data-driven Riemann estimator is introduced in §3.3, and an O(N log N) prefix–suffix algorithm is given in §4. §5 reports synthetic experiments on location shifts, scale changes, heavy-tailed contamination, monotonicity, and hyperparameter sensitivity. The central claims are that the Riemann estimator is asymptotically consistent with the continuous integral and that the empirical statistic inherits the bounded influence and robustness of the continuous formulation.

Significance. The theoretical properties of the continuous DKMD are elementary but valid under the stated conditions, and the fast O(N log N) algorithm is a useful computational contribution for univariate comparisons. If the empirical estimator were consistent and robust, the paper would provide a practical signed kernel statistic with attractive structural guarantees. However, the consistency assertion is unproved and the paper's own experiments contradict the robustness claims. The empirical estimator is not controlled by the bounded continuous witness function, so the headline advantage over the mean difference is unsupported.

major comments (3)
  1. [§3.3, Remark 2] The consistency claim is asserted without proof. The condition max Δz_k → 0 is not established and is generally false for heavy-tailed distributions: for an unbounded-support distribution such as the Cauchy, the largest spacing among N iid samples grows without bound (linearly in N). Consequently, the right-Riemann sum in Eq. (8) does not converge to the Lebesgue integral, and the estimator does not inherit the continuous bound |ψ_w(x)| ≤ 2σ. This is load-bearing because the paper's central robustness claim rests on this consistency.
  2. [§5.2, Table 2] Table 2 directly contradicts the claimed bounded influence. With k=5 Cauchy outliers, DKMD has standard deviation 10.45, versus 0.0506 at k=0, and the mean jumps from -0.2268 to -0.1557. The text asserts 'bounded standard deviation' and 'numerically harmless' outliers, but the reported values show a variance explosion comparable in scale to the mean-difference failure. This internal inconsistency undermines the core empirical contribution.
  3. [§3.4] The robustness argument confuses the continuous witness ψ_w with the discrete witness ψ̂ in Eq. (10). The boundedness of w does not limit the influence of extreme observations because ψ̂ includes the random spacing Δz_k, which is unbounded. A single tail outlier contributes ≈ Δz_out to ψ̂ at its own location, shifting the statistic by O(Δz_out/n). Thus the claimed 'bounded influence guaranteed by our theoretical framework' does not apply to the actual estimator.
minor comments (4)
  1. [§3.3] Eq. (8) uses z_k as both partition and evaluation points, but the partition is random. The proof of consistency must handle both the shrinking of internal spacings and the truncation outside [z_1, z_N]; neither is mentioned. Even for light-tailed distributions the tail truncation term must be shown to vanish.
  2. [§3.3] The notation Δz_k is defined as z_k − z_{k−1}, but the sum starts at k=2. In Eq. (11), Δz_1 is set to 0; this should be stated explicitly in the main text, not only in Algorithm 1.
  3. [§5.1, Table 1] The reported DKMD means for location shifts are negative (e.g., −0.7583 at μ=1.0) while the text says 'negative sign explicitly indicates that Q is shifted toward larger values.' This is consistent with the definition, but the sign convention is not stated explicitly; clarify that DKMD(P,Q) < 0 corresponds to Q being to the right of P.
  4. [§6] The claim that 'No existing kernel two-sample statistic satisfies these three properties simultaneously' is not substantiated with a literature survey or a specific comparison. Consider softening or providing evidence.

Circularity Check

0 steps flagged

No significant circularity: the statistic's structural properties follow directly from its definition, and the main load-bearing gap (unproved consistency of the data-driven Riemann estimator) is an omitted proof rather than circular reasoning.

full rationale

DKMD is defined directly as the integral of a fixed odd weight times the kernel-embedding difference (Section 3.1, Eq. 5-6). Theorems 3.1-3.3 (antisymmetry, symmetric immunity, directional monotonicity) are immediate analytic consequences of oddness, evenness, and the convolution structure of the witness function; they do not assume the conclusion. No parameters are fitted to data: sigma and tau are user-chosen, and the empirical estimator (Eq. 8) is a literal right-Riemann sum over the pooled empirical partition, with Eq. 9 being an algebraic rearrangement of the same finite sums. The consistency claim in Remark 2 is asserted rather than proved, and Table 2 (k=5: DKMD std = 10.45) shows the discrete estimator is not controlled by the continuous bound |psi_w| <= 2sigma; this is a genuine correctness/robustness gap and an omitted proof, which I flag per the reviewing rule, but it is not circular because the estimator is not fitted to the target statistic nor justified by a self-citation chain. The reference list contains no self-citations, and no uniqueness theorem or ansatz is imported from the authors' prior work. The Gaussian experiments for monotonicity and symmetric immunity instantiate the proved theorems rather than circularly generating them. Overall, there is no circular reduction of the paper's central claims to its inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central construction is self-contained except for standard RKHS embedding facts and calculus. The main uncharged premisses are the unproved consistency of the right-Riemann estimator and the unstated symmetry condition for scale immunity; σ and τ are user-chosen free parameters.

free parameters (2)
  • kernel bandwidth σ = 1 in main benchmarks; swept in §5.4
    Hand-chosen; controls the spatial reach of the Laplacian kernel and directly affects the magnitude of DKMD and its estimator.
  • weight scale τ = 1 in main benchmarks; swept in §5.4
    Hand-chosen; controls the saturation of tanh(w), setting the trade-off between directional sensitivity and tail robustness.
axioms (5)
  • standard math The kernel mean embedding is injective for characteristic kernels, so P ≠ Q implies μ_P ≠ μ_Q.
    Invoked in §2.2 to justify comparing distributions through their embeddings; a standard result from the cited RKHS literature.
  • standard math Fubini's theorem applies: w is bounded and the kernel is integrable, so the order of integration can be interchanged in §3.1.
    Used to derive the witness function ψ_w(x) and the expectation-difference form of DKMD.
  • standard math For Theorem 3.3, ψ'_w(x) = ∫ w'(x+u)κ(u)du > 0 for strictly increasing w and strictly positive translation-invariant kernel.
    Assumes differentiation under the integral sign and strict positivity of the kernel profile; necessary for directional monotonicity under stochastic dominance.
  • ad hoc to paper The data-driven Riemann estimator is consistent: max Δz_k → 0 and tail integrability ensure the right-Riemann sum converges to the Lebesgue integral.
    Remark 2 asserts this without proof. It is required for the estimator to 'strictly preserve' the theoretical guarantees, and it is not guaranteed for the heavy-tailed distributions used in §5.2.
  • domain assumption Pure scale changes produce an even embedding difference Δ(z).
    Theorem 3.2's symmetric immunity is conditional on Δ being even; the paper presents it as immunity to 'pure scale changes' without stating the extra symmetry/centering condition on the two distributions.

pith-pipeline@v1.3.0-alltime-deepseek · 8565 in / 19513 out tokens · 191248 ms · 2026-08-01T10:43:19.245848+00:00 · methodology

0 comments
read the original abstract

We introduce the Directional Kernel Mean Difference (DKMD), a signed statistic for univariate distribution comparison that preserves the direction of distributional shifts. Unlike the squared Maximum Mean Discrepancy (MMD), which discards directional information by squaring the RKHS distance, DKMD integrates the difference of kernel mean embeddings against a fixed odd weighting function. This construction yields three structural properties: antisymmetry, immunity to symmetric distributional differences, and directional monotonicity under stochastic dominance. We derive a data-driven Riemann estimator that ensures asymptotic consistency with the continuous formulation, strictly preserving the theoretical guarantees of the signed statistic in empirical evaluations. To overcome the quadratic computational cost of kernel methods, we develop an $O(N \log N)$ prefix--suffix scanning algorithm that exploits the total order of the real line while requiring only $O(N)$ memory. Experiments on synthetic benchmarks demonstrate that DKMD correctly isolates directional shifts from symmetric perturbations, remains robust to heavy-tailed outliers that can flip the sign of the mean difference, and scales to millions of samples in seconds.

Figures

Figures reproduced from arXiv: 2607.20119 by Jiangfeng Fu, Shijie Zhong.

Figure 1
Figure 1. Figure 1: Comparison of DKMD and MMD2 under location and scale shifts. DKMD effectively detects location shifts while maintaining immunity to scale (symmetric) changes [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Robustness to heavy-tailed Cauchy outliers. The classical mean difference suffers from unbounded variance [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Monotonic directional activation of DKMD across continuous location shifts (µ). 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity map of DKMD over a grid of kernel bandwidths (σ) and weight scales (τ ) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

12 extracted references · 2 linked inside Pith

  1. [1]

    A kernel two-sample test.The journal of machine learning research, 13(1):723–773, 2012

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.The journal of machine learning research, 13(1):723–773, 2012

    2012

  2. [2]

    Learning with kernels: support vector machines, regularization, optimization, and beyond

    Bernhard Scholkopf and Alex Smola. Learning with kernels: support vector machines, regularization, optimization, and beyond. InAdaptive computation and machine learning series, 2001

    2001

  3. [3]

    Sriperumbudur, Arthur Gretton, Kenji Fukumizu, Bernhard Scholkopf, and Gert R

    Bharath K. Sriperumbudur, Arthur Gretton, Kenji Fukumizu, Bernhard Scholkopf, and Gert R. G. Lanckriet. Hilbert space embeddings and metrics on probability measures.J. Mach. Learn. Res., 11:1517–1561, 2009

    2009

  4. [4]

    Kernel choice and classifiability for rkhs embeddings of probability distributions.Advances in neural information processing systems, 22, 2009

    Kenji Fukumizu, Arthur Gretton, Gert Lanckriet, Bernhard Schölkopf, and Bharath K Sriperumbudur. Kernel choice and classifiability for rkhs embeddings of probability distributions.Advances in neural information processing systems, 22, 2009

    2009

  5. [5]

    Fienberg

    George Casella, Ingram Olkin, and Stephen E. Fienberg. Testing statistical hypotheses. 2005

    2005

  6. [6]

    On the relation between universality, characteristic kernels and rkhs embedding of measures

    Bharath Sriperumbudur, Kenji Fukumizu, and Gert Lanckriet. On the relation between universality, characteristic kernels and rkhs embedding of measures. In Yee Whye Teh and Mike Titterington, editors,Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 ofProceedings of Machine Learning Research, pages 7...

    2010

  7. [7]

    Székely and Maria L

    Gábor J. Székely and Maria L. Rizzo. Energy statistics: A class of statistics based on distances.Journal of Statistical Planning and Inference, 143:1249–1272, 2013

    2013

  8. [8]

    Sejdinovic, Bharath K

    D. Sejdinovic, Bharath K. Sriperumbudur, Arthur Gretton, and Kenji Fukumizu. Equivalence of distance-based and rkhs-based statistics in hypothesis testing.ArXiv, abs/1207.6076, 2012. 10 APREPRINT- JULY23, 2026

    Pith/arXiv arXiv 2012

  9. [9]

    Peters, Dominik Janzing, and Bernhard Scholkopf

    Kun Zhang, J. Peters, Dominik Janzing, and Bernhard Scholkopf. Kernel-based conditional independence test and application in causal discovery.ArXiv, abs/1202.3775, 2011

    Pith/arXiv arXiv 2011

  10. [10]

    Fast two-sample testing with analytic representations of probability measures.Advances in Neural Information Processing Systems, 28, 2015

    Kacper P Chwialkowski, Aaditya Ramdas, Dino Sejdinovic, and Arthur Gretton. Fast two-sample testing with analytic representations of probability measures.Advances in Neural Information Processing Systems, 28, 2015

    2015

  11. [11]

    Chwialkowski, and Arthur Gretton

    Wittawat Jitkrittum, Zoltán Szabó, Kacper P. Chwialkowski, and Arthur Gretton. Interpretable distribution features with maximum testing power. InNeural Information Processing Systems, 2016

    2016

  12. [12]

    Sliced wasserstein kernels for probability distributions

    Soheil Kolouri, Yang Zou, and Gustavo Kunde Rohde. Sliced wasserstein kernels for probability distributions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5258–5267, 2015. 11

    2016