{"id":"f1145b2e-c5ef-4cac-834d-d6bbb098fc68","arxiv_id":"2608.10470","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proves that HSIC, a closed-form joint dependence statistic, is equivalent to the conditional MMD fairness integral up to a spectral tail, and uses it as a fast fair-representation regularizer.","lead":"Fair representation learning usually enforces independence from a continuous sensitive variable by averaging many local estimates, one per sensitive value. This paper shows that one joint dependence statistic, HSIC, targets the same fairness condition, can be computed directly from paired samples, and is about 36 times faster per epoch than a leading conditional method in their experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HSIC-to-conditional-target equivalence rests on a spectral-tail assumption the paper itself concedes is vacuous for Gaussian kernels on continuous S; without a non-vacuous regime, the claimed O(n^{-1/2}) statistical-efficiency gain estimates a different functional.","rationale":"The paper's disintegration identity (Theorem 3.2) and Corollary 3.4 are mathematically correct, and the zero-level equivalence of Proposition 3.1 is real: both formulations vanish exactly when Z⊥S. The experiments are reported honestly, with self-flagged limitations in Appendix H and Section D. The problem is in the nonzero regime, which is exactly where the advertised statistical-efficiency gain lives. A V-statistic converging at n^{-1/2} is only a gain over the conditional estimator's n^{-2/5} if it is estimating the same target or a quantity that controls it. Theorem 4.5 is the only population-level control, and it has the form lambda_m^{-1} HSIC + rho_m^2. For Gaussian kernels on continuous S, the kernel integral operator has no spectral gap and the paper's Appendix H.3 concedes the population bound is vacuous. There is no argument that learned representations satisfy the smoothness condition making rho_m^2 small, and a high-frequency dependence Z=f(S) is a concrete case where HSIC is small but the conditional MMD integral is not. This matches the reader's weakest assumption exactly; I would keep the CONDITIONAL verdict, with the condition being that the authors either demonstrate a non-vacuous spectral regime or revise the abstract and Section 6.2 to avoid claiming the n^{-1/2} rate transfers to the conditional-integral fairness target. The concern is not about internal inconsistency in the proofs but about a load-bearing external premise that the paper itself flags as unsupported.","tokens_in":37020,"tokens_out":6937,"duration_ms":69546,"concrete_test":"Run a synthetic experiment with S~Unif(0,1), Z_M = sin(2π M S) + ε, ε~N(0,0.01), M∈{1,5,20,100}. For each M, compute the biased HSIC with Gaussian kernels at median-heuristic bandwidth and estimate E_S[MMD^2(P_{Z|S},P_Z)] by averaging the empirical MMD between {Z_i: S_i in a narrow bin around s} and the global {Z_i} over a dense grid of s. If HSIC decreases with M while the conditional-MMD integral remains roughly constant, the residual rho_m^2 in Theorem 4.5 is large for continuous Gaussian S, confirming that the population equivalence is vacuous; the abstract's rate claim should then be reworded to state that HSIC directly penalizes joint dependence, not that it provably enforces the conditional-integral target at the faster rate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that FRHSIC reaches the conditional-integral fairness target (GDP/EIPM) at the O(n^{-1/2}) HSIC rate. The bridge is Theorem 4.5: E_S[MMD^2(P_{Z|S},P_Z)] <= lambda_m^{-1} HSIC(Z,S) + rho_m^2. This is useful only if the residual rho_m^2, the spectral tail of Delta(s)=mu_{Z|S=s}-mu_Z under the sensitive-kernel integral operator, is small. Appendix H.3 explicitly states that for Gaussian kernels on continuous S the population-level bound becomes vacuous because the analogous constant collapses to zero. All real datasets in Section 6 use continuous S (age or racial composition). Consequently, a small empirical HSIC is not guaranteed to imply a small conditional MMD integral or a small GDP gap; the O(n^{-1/2}) versus O(n^{-2/5}) comparison in the abstract and Section 6.2 may be comparing errors for different functionals. A concrete failure mode is Z=f(S) with f a high-frequency sinusoid: with a Gaussian S-kernel, HSIC decays with the frequency while the conditional MMD integral stays bounded away from zero, placing all mass in rho_m^2. This is not a flaw in the disintegration algebra; it is an unsupported premise about the spectral tail, and the paper should either exhibit a non-vacuous regime or soften the equivalence/efficiency claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to enforce fairness with respect to a continuous sensitive attribute S by penalizing a single joint discrepancy d(P_{Z,S}, P_Z ⊗ P_S) rather than the conditional-integral criteria (GDP, EIPM, mutual information) that average a per-value discrepancy over S. The authors prove a disintegration identity (Theorem 3.2) and, under a decomposable witness class, recover the conditional-integral functional (Corollary 3.4). They instantiate the joint discrepancy with HSIC, prove a spectral equivalence between HSIC and the conditional MMD integral (Theorem 4.5), derive population and finite-sample GDP control (Corollaries 4.6 and Theorem 4.7), and give a uniform concentration bound for the empirical HSIC over encoder classes at an O(n^{-1/2}) rate (Theorem 5.1). The resulting algorithm, FRHSIC, is evaluated on synthetic data and five real datasets, reporting fairness–accuracy tradeoffs comparable to baselines and a large per-epoch speedup over FREM.","tokens_in":37356,"tokens_out":6936,"duration_ms":69405,"significance":"If the central equivalence were fully established, the paper would make a useful contribution: it gives a clean measure-theoretic connection between joint-distribution dependence measures and conditional-integral fairness criteria, provides a closed-form O(n^2) regularizer, supplies explicit spectral and concentration bounds, and releases code and experimental details. The disintegration identity and the algebraic derivations in Sections 3–4 are sound, and the empirical evaluation is more careful than typical in this literature, including matched operating points, five random splits, and a reproducibility checklist. However, the paper's headline statistical-efficiency claim rests on a spectral-tail assumption that the authors themselves concede is vacuous for Gaussian kernels on continuous S, which is the exact setting of every real-data experiment. The significance is therefore conditional on either a non-vacuous tail regime or a substantive reinterpretation of what the HSIC rate controls.","major_comments":[{"comment":"The upper bound E_S[MMD^2(P_{Z|S},P_Z)] ≤ λ_m^{-1} HSIC(Z,S) + ρ_m^2 is useful only when the spectral tail ρ_m^2 is small, but Appendix H.3 explicitly states that for Gaussian kernels on continuous S the population-level bound becomes vacuous because the analogous constant collapses to zero. Since all real datasets in Section 6.1 use continuous S (age or racial composition) with Gaussian kernels, the paper does not establish that a small empirical HSIC implies a small conditional MMD integral in its own operating regime. Consequently, the abstract and Section 6.2 compare the O(n^{-1/2}) rate of HSIC with the O(n^{-2/5}) rate of the conditional-route estimator while the two statistics may be converging to different population functionals. The authors should either prove a non-vacuous tail bound under explicit regularity conditions matching the experiments, or revise the equivalence and efficiency claims to refer only to the spectral-projected quantities actually controlled.","section":"Theorem 4.5 and Appendix H.3"},{"comment":"The decomposability condition used to recover the exact conditional-integral IPM is not satisfied by the HSIC witness class. Corollary 3.4 requires the class F to contain every measurable selector s ↦ g_s ∈ V, but for the tensor-product RKHS unit ball corresponding to HSIC this is generally false: for a Gaussian kernel on S, even a constant selector g(z) does not lie in F_S ⊗ F_Z because constant functions are not in the RKHS of a Gaussian kernel. Thus HSIC is not covered by the exact equality of Corollary 3.4, and the bridge from HSIC to the conditional MMD integral is entirely the spectral bound of Theorem 4.5, whose vacuity is the subject of the previous comment. The manuscript should state which natural witness classes are decomposable and explicitly acknowledge that the HSIC witness class is not among them.","section":"Definition 3.3 and Corollary 3.4"},{"comment":"The uniform concentration guarantee at the O(n^{-1/2}) rate is a load-bearing part of the train-to-population argument, but the proof in Appendix G.2 consists of a mapping onto Corollary 23 of Ni and Huo (2024) and explicitly says 'We do not reprove that result.' Because the cited work is the authors' own preprint and is not stated verbatim in the manuscript, the reader cannot verify the exact assumptions, constants, or scope of the imported theorem. The authors should either reproduce the proof in an appendix or state the imported result in full, including the precise definition of the two-sample statistic, the complexity measure, and the conditions under which Corollary 23 applies to the FRHSIC objective.","section":"Theorem 5.1 and Appendix G.2"},{"comment":"The synthetic convergence experiment measures the error of each estimator relative to its own population value, not the error relative to the fairness target E_S[MMD^2(P_{Z|S},P_Z)] or to the GDP gap. The fitted slopes of −0.46 for HSIC and −0.44 for EIPM therefore support the claim that empirical HSIC converges quickly as an estimator of population HSIC, but they do not directly support the claim that FRHSIC reaches the conditional-integral fairness target at the O(n^{-1/2}) rate. If the spectral-tail issue in the first comment is unresolved, the faster rate is a statement about a different functional; this distinction should be made explicit in the text.","section":"Section 6.2 and Figure 2"}],"minor_comments":[{"comment":"The phrase 'basel es' in the abstract appears to be a typo for 'baselines'; please correct it.","section":"Abstract"},{"comment":"The caption of Figure 6(c) refers to 'Thm. 3.4', which appears to be a typo; the intended reference is likely Theorem 4.7, the finite-sample GDP control bound.","section":"Appendix E.1, Figure 6"},{"comment":"The reproducibility checklist says the reported numbers 'agree with our earlier leakage-affected numbers to within standard deviations'; please clarify what 'leakage-affected' refers to, since the phrase is ambiguous and could be read as suggesting undisclosed earlier experiments.","section":"Appendix D.6"},{"comment":"The assumption ∥(I−P_m)Δ∥²_{L^2(P_S;F_Z)} ≤ ρ_m^2 is stated as a hypothesis of Theorem 4.5, but ρ_m^2 is the quantity whose smallness the theorem needs to be useful. Please add a remark explaining how one might verify or bound ρ_m^2 for specific kernels and target classes, and note explicitly the negative result reported in Appendix H.3.","section":"Section 4.1, Theorem 4.5"},{"comment":"The notation G({id_S}(S)) = 0 relies on a 'no-absolute-value convention' from Ni and Huo (2024); since this convention is not standard, please restate it in the appendix so the vanishing of the singleton-class complexity term is self-contained.","section":"Appendix G.2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid measure-theoretic core and a genuinely useful algorithmic package, but the headline equivalence to the conditional-integral fairness target is not supported in the Gaussian-kernel/continuous-S regime that the experiments use. The authors may be able to fix this by proving a non-vacuous tail bound under concrete spectral-decay assumptions, or by reframing the contribution as a smoothed/projected substitute rather than an equivalence. I would also encourage the editor to require that Theorem 5.1's imported proof be made available in the manuscript, since it is central and the cited source is the authors' own unpublished preprint."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the disintegration identity and the recovery of GDP/EIPM from a joint discrepancy are clean, correct, and genuinely new. The spectral equivalence theorem is a real attempt to quantify the link, and the empirical section is honestly reported with code and explicit limitations. But the headline statistical-efficiency claim is overstated, because the bridge between HSIC and the conditional MMD integral collapses to a vacuous bound exactly in the continuous-S Gaussian-kernel setting used in the experiments. This needs revision before publication, not rejection.\n\nWhat is new and good: Theorem 3.2 (disintegration of the joint-vs-product IPM) and Corollary 3.4 (recovery of the conditional-integral functional on decomposable witness classes) are correct and give a useful unifying perspective. The spectral equivalence in Theorem 4.5 is a genuine attempt to make the relationship quantitative. The experiments are competitive, the ablation studies are thoughtful, and the paper consistently flags its own limitations (FREM anchor subsampling, no paired significance tests, the vacuous population bound in H.3). That level of honesty is rare.\n\nSoft spots, in proportion: The load-bearing issue is the spectral tail in Theorem 4.5. For Gaussian kernels on continuous S, the population-level bound becomes vacuous because the constant collapses to zero (Appendix H.3). All real datasets use continuous S. So a small empirical HSIC does not guarantee a small conditional MMD integral or a small GDP gap in the regime where the method operates. The stress-test example (Z = f(S) with high-frequency sinusoidal f) is a concrete failure mode, and the paper does not offer a non-vacuous alternative. Consequently, the O(n^{-1/2}) versus O(n^{-2/5}) comparison in the abstract and Section 6.2 compares estimation errors for two different functionals. That is a fair criticism, not a nitpick.\n\nSecond, the uniform concentration theorem (5.1) is imported from the authors' prior work without a self-contained proof. This is acceptable practice if the cited result is solid, but it makes the guarantee depend on an unreviewed dependency.\n\nThird, the experiments use 5 random splits with no paired significance testing, and FREM's anchor subsampling is an approximation. The paper admits these, but they limit the strength of the empirical claims, especially on datasets where methods are within noise of each other.\n\nWho this is for: researchers working on fair representation learning with continuous sensitive attributes, and anyone interested in the interplay between joint and conditional dependence measures. The disintegration identity alone is worth citing.\n\nRecommendation: send to peer review. The theory is largely sound and the empirical study is competent. The authors should be asked to either exhibit a non-vacuous spectral regime for continuous S or substantially soften the equivalence/efficiency claims. With that change, this would be a solid contribution.","headline":"Core theory is sound and worth engaging, but the headline O(n^{-1/2}) efficiency gain rests on a spectral-tail condition the paper itself admits is vacuous for Gaussian kernels on continuous S.","tokens_in":857,"tokens_out":991,"would_cite":true,"duration_ms":31034,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","62G05","62G20","46E22"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that fair representation learning with a continuous sensitive attribute can be enforced by a single joint-dependence statistic, the Hilbert–Schmidt independence criterion (HSIC), which matches the conditional-integral…","keywords":["fair representation learning","continuous sensitive attributes","Hilbert–Schmidt independence criterion","kernel methods","demographic parity","integral probability metrics","spectral analysis","uniform concentration"],"falsifier":"A concrete check: generate paired samples where the representation depends on $S$ chiefly through high-frequency eigenfunctions of the sensitive kernel with small eigenvalues, compute both HSIC and the conditional MMD integral, and verify whether small HSIC implies small conditional MMD. If the two quantities diverge substantially on such data, the spectral-tail assumption behind Theorem 4.5 fails.","tokens_in":36757,"feed_emoji":"⚖️","tokens_out":5480,"duration_ms":47944,"temperature":0.7,"pith_summary":"The paper tries to show that fairness with a continuous sensitive attribute $S$ can be enforced without estimating the conditional law of the representation given each sensitive value. Its route is a single discrepancy between the joint law of $(Z,S)$ and the product of marginals, instantiated as HSIC. The paper proves a disintegration identity that ties this joint discrepancy to the conditional-integral functional used by generalized demographic parity and expectation-of-IPM criteria, and shows that the empirical HSIC regularizer concentrates uniformly at the $O(n^{-1/2})$ rate. A sympathetic reader would care because the conditional route requires nonparametric smoothing with the slower $O(n^{-2/5})$ rate, while the joint route is a closed-form $O(n^2)$ statistic that reaches the same target.","feed_headline":"HSIC reaches conditional-fairness target at root-n speed","feed_subtitle":"A closed-form O(n^2) statistic enforces independence from a continuous sensitive attribute, skipping per-value smoothing.","key_machinery":"The central object is HSIC as the squared MMD between the joint law of $(Z,S)$ and the product of its marginals in a tensor-product RKHS, estimated by the closed-form biased V-statistic $\\widehat{\\mathrm{HSIC}}_n = n^{-2}\\operatorname{tr}(\\widetilde{K}\\widetilde{L})$. The argument is carried by three mechanisms: the disintegration identity that expresses the joint discrepancy as an $S$-average of conditional contrasts, the decomposable witness class condition under which that average equals the conditional-integral functional, and the spectral analysis of the sensitive-attribute kernel integral operator $T_S$ whose tail $\\rho_m^2$ controls the gap between HSIC and the conditional MMD integral. The empirical counterpart replaces the operator spectrum with the eigenvalues of the centered sensitive Gram matrix.","core_discovery":"The central claim is that the representation-level fairness target $Z \\perp S$ can be measured by $\\mathrm{HSIC}(Z,S)$, the squared maximum mean discrepancy between the joint law $P_{Z,S}$ and the product $P_Z \\otimes P_S$, and that this quantity is equivalent to the conditional-integral functional $\\mathbb{E}_S[d(P_{Z|S}, P_Z)]$ that existing continuous-sensitive criteria use. The paper establishes a disintegration identity (Theorem 3.2) rewriting the joint-vs-product IPM as an $S$-averaged conditional contrast, and shows that on decomposable witness classes this equals the conditional-integral IPM (Corollary 3.4). It then proves that HSIC is equivalent to the conditional MMD integral up to an explicit spectral tail of the sensitive-attribute kernel (Theorem 4.5), and that the empirical HSIC regularizer concentrates uniformly over the encoder class at rate $O(n^{-1/2})$ (Theorem 5.1). The resulting algorithm, FRHSIC, attains fairness–accuracy tradeoffs comparable to conditional-route baselines while training about 36 times faster per epoch at $n=20{,}000$.","pith_inferences":["Editorial inference: if the spectral tail of the sensitive-attribute kernel does not decay quickly in practice, small HSIC may not guarantee small conditional MMD or GDP gap; this could be tested by measuring both quantities on real data across kernel bandwidths.","Editorial inference: the root-$n$ rate applies to estimating the joint dependence statistic itself, not directly to the fairness metric; practitioners should verify that the downstream GDP gap also improves at the claimed rate.","Editorial inference: the structural identity of Theorem 3.2 would extend to other joint dependence measures such as distance covariance, but those would not inherit the closed-form estimator or root-$n$ rate unless they are also V-statistics.","Editorial inference: a testable extension is to replace the reference distribution $P_Z \\otimes P_S$ with $P_Z \\otimes \\mathrm{Unif}(S)$ to enforce minority-protection fairness, a modeling choice the paper notes does not change the structural identity."],"forward_implications":["A single HSIC penalty enforces the same representation-level independence target as GDP and EIPM, without constructing conditional laws at each sensitive value.","The empirical HSIC regularizer converges at the root-$n$ rate, so its statistical efficiency is better than the nonparametric $O(n^{-2/5})$ rate of conditional-route estimators.","Small HSIC controls the conditional MMD integral and the demographic-parity gap of RKHS heads, up to an explicit spectral tail and a spectral factor in the finite-sample bound.","As a minibatch regularizer, FRHSIC reaches fairness–accuracy tradeoffs comparable to the conditional-route baselines while reducing per-epoch training time by a large factor.","The framework extends to equal opportunity by restricting the statistic to the positive-outcome subset and to multiple continuous sensitive attributes through a product kernel."],"supporting_citations":[{"why":"Defines HSIC as the RKHS dependence measure and establishes the characteristic-kernel condition under which it vanishes exactly under independence.","marker":"(Gretton et al., 2005)"},{"why":"Supplies the disintegration theorem used to derive the joint-vs-product IPM identity and the regular conditional distributions throughout.","marker":"(Kallenberg, 2002)"},{"why":"Provides the kernel-smoothing estimation route for conditional expectations whose slower nonparametric rate the paper contrasts with HSIC.","marker":"(Nadaraya, 1964; Watson, 1964)"},{"why":"Introduces generalized demographic parity, the conditional-integral criterion whose target FRHSIC is shown to match.","marker":"(Jiang et al., 2022)"},{"why":"Introduces the expectation-of-IPM criterion and the FREM estimator, the principal conditional-route baseline and its $O(n^{-2/5})$ rate.","marker":"(Kong et al., 2025)"},{"why":"Provides the uniform concentration inequality for kernel two-sample statistics that Theorem 5.1 adapts to obtain the $O(n^{-1/2})$ train-to-population bound.","marker":"(Ni and Huo, 2024)"},{"why":"Establishes adversarial fair representation learning and the transfer of fairness to downstream heads, which the cross-head stability experiment evaluates.","marker":"(Madras et al., 2018)"},{"why":"Demonstrates the earlier use of HSIC as a fairness penalty in kernel regression, which the paper extends with disintegration and spectral theory.","marker":"(Pérez-Suay et al., 2017)"}],"fun_headline_variants":["HSIC matches conditional fairness at root-n speed","Joint HSIC enforces fairness without per-value smoothing","Fairness via entire joint law: HSIC skips conditional estimates","FRHSIC: root-n fair representations, 36x faster training","Disintegration identity turns joint fairness into conditional target"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The equivalence between HSIC and the conditional MMD integral holds only up to a spectral tail of the sensitive-attribute kernel, and for Gaussian kernels on continuous $S$ the paper's Appendix H.3 notes that the population-level bound can become vacuous because the analogous constant collapses to zero.","fun_headline_variants_meta":{"raw":{"variants":["HSIC matches conditional fairness at root-n speed","Joint HSIC enforces fairness without per-value smoothing","Fairness via entire joint law: HSIC skips conditional estimates","FRHSIC: root-n fair representations, 36x faster training","Disintegration identity turns joint fairness into conditional target"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000643,"raw_usage":{"total_tokens":3037,"prompt_tokens":1108,"completion_tokens":1929,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":724,"completion_tokens_details":{"reasoning_tokens":1848}},"tokens_in":724,"tokens_out":1929,"duration_ms":12291,"temperature":1.0,"reasoning_tokens":1848,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:20:11.349030+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: generate paired samples where the representation depends on $S$ chiefly through high-frequency eigenfunctions of the sensitive kernel with small eigenvalues, compute both HSIC and the conditional MMD integral, and verify whether small HSIC implies small conditional MMD. If the two quantities diverge substantially on such data, the spectral-tail assumption behind Theorem 4.5 fails.","supporting_citations":[{"cited_title":"Measuring statistical dependence with Hilbert-Schmidt norms","cited_arxiv_id":null,"evidence_quote":"Defines HSIC as the RKHS dependence measure and establishes the characteristic-kernel condition under which it vanishes exactly under independence."},{"cited_title":"Foundations of Modern Probability","cited_arxiv_id":null,"evidence_quote":"Supplies the disintegration theorem used to derive the joint-vs-product IPM identity and the regular conditional distributions throughout."},{"cited_title":"Generalized demographic parity for group fairness","cited_arxiv_id":null,"evidence_quote":"Introduces generalized demographic parity, the conditional-integral criterion whose target FRHSIC is shown to match."},{"cited_title":"Fair representation learning for continuous sensitive attributes using expectation of integral probability metrics","cited_arxiv_id":null,"evidence_quote":"Introduces the expectation-of-IPM criterion and the FREM estimator, the principal conditional-route baseline and its $O(n^{-2/5})$ rate."},{"cited_title":"Fair kernel learning","cited_arxiv_id":null,"evidence_quote":"Demonstrates the earlier use of HSIC as a fairness penalty in kernel regression, which the paper extends with disintegration and spectral theory."}],"review_version":1}