{"id":"bb9f4b61-f362-4953-8949-2c0b638d39ba","arxiv_id":"2505.00830","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Intersectional Divergence is a new regression fairness measure that combines group error differences with relevance-weighted target values, and it can be optimized as a loss to reduce unfairness.","lead":"This paper proposes Intersectional Divergence, a fairness metric for regression that compares prediction errors across combinations of protected attributes while weighting errors by how important the target value is to users. It also turns the metric into a trainable loss and reports improved fairness on four public datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The IDLoss convergence guarantee rests on a proof with false steps: Prop. A.3's example satisfies (not violates) convexity, and Thm. A.5's claim that each region-boundary crossing reduces IDLoss is contradicted by a two-group counterexample.","rationale":"I read the paper's central claim as having two load-bearing parts: (a) ID is a novel measure of intersectional, domain-imbalance-aware regression fairness, and (b) IDLoss satisfies convergence guarantees and is practically optimizable. Part (b) is where the argument is least secure. The relevance-function choice flagged by the reader is a genuine but standard parameterization inherited from SERA; a sensitivity analysis could close it. Appendix A, by contrast, contains statements that are demonstrably false. Proposition A.3 computes f(0.5x + 0.5y) = 0 and 0.5f(x) + 0.5f(y) = 0.05 and concludes this violates convexity, but the inequality direction is exactly the convexity inequality; with two groups IDLoss = max(e1, e2), the maximum of two convex functions, which is convex, so the example cannot work. Theorem A.5's Step 4 claim that each region-boundary crossing reduces IDLoss by at least delta > 0 is also false: switching alpha_min replaces the excluded error curve with the new minimum curve, and the net change can be positive, as the two-group example shows (IDLoss rises 25 to 26.01 across an argmin switch). Definition A.2's partition requires a single alpha_min for all t, leaving trajectories with threshold-varying argmin patterns unanalyzed, and no uniform delta is established. These are errors in a headline claim, not merely missing details.\n\nI partially agree with the reader: both concerns are real, but the convergence proof is more load-bearing because it is demonstrably invalid rather than empirically under-supported. The reader's phi concern remains valid; the NLSY narrative highlights the high-income tail of a symmetric boxplot relevance function, and without a phi sensitivity analysis the domain-imbalance novelty is not fully established.\n\nThe empirical evidence supports the practical claim that IDBoost reduces ID across all four datasets (Appendix B raw values show ID reductions versus XGB-SERA of roughly 4 to 15 percent on Communities, LSAC, NLSY, and COMPAS), so rejection is not warranted. The verdict stays CONDITIONAL, with conditions including a corrected convergence proof or a retracted guarantee, a phi sensitivity analysis, and the reader's evaluation fixes (no test-set selection for Agarwal; report raw metric magnitudes alongside ranks).","tokens_in":19799,"tokens_out":30175,"duration_ms":280767,"concrete_test":"Independently re-derive Theorem A.5 by testing the region-crossing step. With two protected groups and a single relevance threshold, set targets y = (0, 10) and errors e1(yhat) = yhat1^2, e2(yhat) = (yhat2 - 10)^2, so IDLoss = max(e1, e2). Walk the path yhat(t) = (5, 10 - t) from t = 0 to t = 6 and record IDLoss at the crossing yhat = (5, 5) (value 25) and immediately after the argmin switches, e.g. yhat = (5, 4.9) (value 26.01). If IDLoss increases across the boundary, the claim that every region-boundary crossing reduces IDLoss by at least delta > 0 is false and the Lojasiewicz argument in Theorem A.5 fails. Confirm with Proposition A.3: the displayed inequality 0 < 0.05 satisfies convexity rather than violating it, so that example does not establish non-convexity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and Section 4.2 claim IDLoss 'satisfies convergence guarantees' via a Lojasiewicz analysis described in the conclusion as 'the first rigorous analysis of convergence properties for intersectional fairness optimization in regression.' Appendix A does not establish this. Two concrete failures:\n\n(1) Proposition A.3's non-convexity demonstration is directionally wrong. With two groups, IDLoss = e1 + e2 - min(e1, e2) = max(e1, e2), the pointwise maximum of two convex squared-error terms, which is convex; no two-group example can demonstrate non-convexity. The displayed computation gives IDLoss(0.5 yA + 0.5 yB) = 0 < 0.05 = 0.5 IDLoss(yA) + 0.5 IDLoss(yB), exactly the convexity inequality f(lambda x + (1-lambda) y) <= lambda f(x) + (1-lambda) f(y); the text calls this a violation, inverting the definition.\n\n(2) Theorem A.5 Step 4 asserts each region-boundary crossing reduces IDLoss by at least delta > 0. Switching alpha_min changes the loss by the integral of (e_new_min(t) - e_old_min(t)) over t, which need not be negative. Counterexample with two groups and target values y = (0, 10): e1 = yhat1^2, e2 = (yhat2 - 10)^2. At yhat = (5, 5) both errors tie at 25 and IDLoss = 25; moving yhat2 to 4.9 raises e2 to 26.01, switches alpha_min to group 1, and IDLoss becomes 26.01 > 25. The crossing increased the loss.\n\nAlso, Definition A.2 partitions the prediction space by requiring the same alpha_min for every t in [0,1]; trajectories where the argmin pattern varies with t are not covered by the within-region Lojasiewicz argument, and the uniform delta needed for the finite-crossing claim is never established. The convergence guarantee is therefore unproven as written. This does not invalidate the empirical evidence that IDLoss reduces ID, but the headline theoretical claim needs a correct proof (e.g., via the Kurdyka-Lojasiewicz property of this piecewise-quadratic objective) or a scaled-back statement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Intersectional Divergence (ID), a new fairness measure for regression tasks that combines the SERA relevance-weighting framework with intersectional protected-attribute groups. ID is defined as the integrated, normalized difference between the best- and worst-performing intersectional groups across relevance thresholds. The paper also defines IDLoss, a surrogate loss for training, and claims that IDLoss is non-convex yet satisfies Lojasiewicz-based convergence guarantees and has piecewise Lipschitz continuous gradients. Experiments on four datasets compare an IDLoss-based boosting method (IDBoost) against several fairness-agnostic and fairness-aware baselines on MSE, SERA, BGL, statistical parity, and ID. The paper argues that ID reveals biases hidden by single-attribute average-error measures and that IDBoost improves fairness while retaining competitive predictive performance.","tokens_in":20162,"tokens_out":7174,"duration_ms":75677,"significance":"If the measure and its optimization are sound, ID addresses a genuine gap: no existing regression fairness measure combines intersectional protected attributes with domain-imbalance weighting. The paper has several concrete strengths: it builds on the externally validated SERA framework rather than inventing a relevance model from scratch; it provides visualizations and a reproducible implementation; and the experimental study spans four datasets with multiple protected attributes. The basic ID construction is intuitive and could be useful for auditing regression models. However, the advertised theoretical contributions are currently not established: the non-convexity proof is mathematically inverted, and the convergence proof contains a false step about region-boundary crossings. Because the abstract and conclusion explicitly promise 'convergence guarantees' and 'the first rigorous analysis of convergence properties for intersectional fairness optimization in regression,' these flaws are load-bearing. The empirical results are suggestive but would need to be paired with either a corrected theory or a substantially weakened set of claims.","major_comments":[{"comment":"The definition of ID is internally inconsistent: in Eq. (5), alpha_min is defined as argmin over alpha of SER_t^alpha / |D_alpha|, but Eq. (6) evaluates the ID integrand with SER_t^{alpha_min} / |D_t^{alpha_min}|. The denominator in the argmin does not match the denominator in the integral. If the intended normalization is by group size |D_alpha|, then Eq. (6) should divide by |D_alpha_min|, not |D_t^{alpha_min}|; if the intended normalization is by the threshold-dependent set |D_t^alpha|, then Eq. (5) should use |D_t^alpha|. This is not a cosmetic point: every reported ID value depends on which normalization is used, and the current text cannot be used to reproduce the numbers.","section":"Section 4, Eqs. (5)-(6)"},{"comment":"The proof of non-convexity is directionally wrong. The displayed computation shows IDLoss(0.5 y_A + 0.5 y_B) = 0 < 0.05 = 0.5 IDLoss(y_A) + 0.5 IDLoss(y_B), which is exactly the convexity inequality f(lambda x + (1-lambda) y) <= lambda f(x) + (1-lambda) f(y) for lambda = 0.5. The text calls this a violation, inverting the definition of convexity. Moreover, for two protected groups IDLoss = e_1 + e_2 - min(e_1, e_2) = max(e_1, e_2), the pointwise maximum of two convex squared-error terms, which is convex. Thus the claimed non-convexity is not demonstrated, and the motivation for the Lojasiewicz analysis in Section 4.2 and the conclusion is unsupported.","section":"Appendix A.2, Proposition A.3"},{"comment":"The assertion that 'each region boundary crossing reduces IDLoss by at least delta > 0' is false. Consider two groups with target values y = (0, 10) and predictions yhat = (5, 5): both groups have squared error 25, and IDLoss = 25 regardless of which group is excluded. If yhat_2 is changed to 4.9, group 1 becomes the unique minimum, so alpha_min switches to group 1; IDLoss becomes the error of group 2, which is 26.01, an increase over 25. The crossing increased the loss rather than decreasing it. Consequently, the 'finite region crossings' argument in Step 4 does not hold, and the proof does not establish convergence to a stationary point.","section":"Appendix A.3, Theorem A.5, Step 4"},{"comment":"The convergence argument also conflates a finite set of possible alpha_min values with a finite number of region transitions along a trajectory. A gradient-descent trajectory could in principle switch alpha_min many times, possibly infinitely often, if the loss does not decrease monotonically at each crossing (which, as shown by the counterexample in the previous comment, it need not). Definition A.2 additionally requires the argmin to be constant for all t in [0,1] simultaneously, but for a fixed prediction vector alpha_min(t) can vary with the relevance threshold; the proof does not analyze trajectories where the minimizing group changes at different t values at different times. A separate, correct finiteness or monotonicity argument is needed.","section":"Appendix A.3, Definition A.2 and Theorem A.5, Step 5"},{"comment":"The text states that IDBoost0.5 'is better than all XGB models in both fairness measures,' but Table 3 shows that XGB_Huber has a better average Statistical Parity rank (4.78) than IDBoost0.5 (6.26). IDBoost0.5 is better than all XGB models on the ID rank and on Delta BGL, but not on SP. The claim as written is contradicted by the paper's own table. In addition, the Agarwal baselines are selected per metric on the test set, an optimistic protocol that should be disclosed in the comparison and may inflate the relative standing of the proposed method.","section":"Section 5.4, Table 3"}],"minor_comments":[{"comment":"The notation in the manuscript is often unclear: Eq. (7) repeats Eq. (6) with a misrendered summation symbol, Eq. (12) uses a vertical-bar condition without a clear indicator function, and the derivative expressions should be written with explicit indicators 1(y_j in D_t^alpha) to be understood. These issues make the formal definitions harder to verify.","section":"Equations (6)-(7) and (12)-(13)"},{"comment":"The text says IDLoss is 'twice-differentiable,' but the presence of the indicator 1(y_j in D_t^alpha) and the dependence on alpha_min mean that the second derivative is at best piecewise and can be discontinuous at relevance thresholds where a sample enters or leaves D_t^alpha. The later Appendix correctly describes piecewise smoothness; Section 4.1 should be reworded to avoid the unqualified twice-differentiability claim.","section":"Section 4.1"},{"comment":"The manuscript contains several typos and editorial remnants: 'All measures disparate treatment' in the Figure 1 caption, 'Statistical Parity compare' in Section 2.1, placeholder conference formatting ('Conference acronym ’XX', '© 2018 ACM'), and inconsistent use of the reference numbering. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The relevance functions used for each dataset are described as interpolated from boxplot statistics, but no explicit sensitivity analysis is provided for the choice of relevance function. Since ID and IDLoss depend directly on phi(Y), a robustness check over alternative relevance functions would substantially strengthen the empirical claims.","section":"Section 5.1 and Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The core weakness is the theoretical appendix. Proposition A.3 is a simple mathematical error, and Theorem A.5 is not proven because Step 4 is false. The convergence guarantee is advertised in the abstract and conclusion, so this is not a peripheral issue. If the authors cannot provide a correct proof, they should remove the convergence-guarantee claims and reframe the contribution as an empirical measure and loss. The experimental study is otherwise reasonable, and the ID measure itself is worth pursuing once the normalization inconsistency in Eqs. (5)-(6) is fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the read on 2505.00830. The paper does one genuinely useful thing: it defines Intersectional Divergence (ID), a fairness measure for regression that combines SERA's relevance weighting with a max-min comparison across intersectional groups. That's a real gap, and the construction is simple enough to be usable. The IDLoss adaptation is also reasonable as a heuristic: exclude the current best group from the loss and push everyone else down. The empirical sections show the measure surfaces bias that MAE and BGL miss, and the IDBoost results are suggestive.\n\nWhere it falls down is the theory. The abstract and conclusion claim IDLoss \"satisfies convergence guarantees\" via a Łojasiewicz analysis. Appendix A does not establish that. Proposition A.3's attempted non-convexity example actually satisfies convexity: they show IDLoss(0.5 yA + 0.5 yB) = 0 < 0.05 = 0.5 IDLoss(yA) + 0.5 IDLoss(yB), which is exactly the convexity inequality. So the paper's own demonstration inverts the definition. More seriously, Theorem A.5's Step 4 asserts each region-boundary crossing reduces the loss by at least δ > 0. That's unsupported, and a simple two-group example (errors tied at 25, then one group's error rises to 26.01, switching the argmin) shows the loss can increase at a crossing. Definition A.2 also partitions the space by requiring the same argmin for all t in [0,1], so trajectories where the best group varies with t aren't covered. The \"finite crossings\" argument needs a uniform δ that is never established. Either the theorem needs a correct proof (the piecewise-quadratic structure suggests a Kurdyka-Łojasiewicz argument might work) or the claim should be scaled back to \"empirically effective loss with piecewise smooth structure.\"\n\nOther soft spots: the Agarwal baseline is selected on the test set, which is a methodological error, though it favors the baseline, so the main qualitative results survive. Results are mostly average ranks; Appendix B has raw values, but the magnitude of fairness gains is hard to eyeball. And the whole approach leans on the user-supplied relevance function φ; no sensitivity analysis, so we don't know how robust the conclusions are to that choice.\n\nBottom line: the measure is a legitimate contribution to fair regression, and the paper deserves a serious referee. But the convergence guarantee as written is not proven, and the experimental protocol needs cleanup. I'd recommend major revision.","headline":"A genuinely useful fairness measure for regression, undermined by a convergence proof that is incorrect as written.","tokens_in":20828,"tokens_out":2568,"would_cite":true,"duration_ms":24659,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Single-attribute fairness audits miss regression bias; Intersectional Divergence measures every protected group at every relevance level, with a provably convergent loss version.","keywords":["fairness","intersectionality","regression","imbalanced domains","relevance function","fair loss function","SERA","protected attributes"],"falsifier":"Two concrete tests. First, recompute ID and rerun the IDBoost comparison on the same four datasets under two or more defensible relevance functions — the default boxplot interpolation, a wider or narrower whisker rule, and an expert-supplied weighting — and check whether the ranking of models by ID, and the fairness improvements claimed for IDBoost0.5, survive; if the 'fairest model' verdict flips, ID is measuring the chosen weighting, not the model. Second, build a dataset with two protected groups whose normalized squared errors are exactly tied across a range of relevance thresholds, making $\\alpha_{\\min}$ non-unique, and run gradient descent on IDLoss in that tie regime; if it fails to converge or cycles between regions, the finite-region-transition argument behind the convergence guarantee breaks at its boundary case.","tokens_in":19530,"feed_emoji":"⚖️","tokens_out":19834,"duration_ms":171119,"temperature":0.7,"pith_summary":"The paper claims that fairness in regression cannot be trusted when measured as an average error over a single protected attribute: real bias hides at the intersection of several protected attributes, and inside target ranges the user cares about most. It proposes Intersectional Divergence (ID) as the first regression fairness measure that combines the intersectionality of protected attributes with importance weighting over the target range: the integral, over relevance thresholds, of the gap between the normalized squared-error curves of the best- and worst-predicted intersectional subgroups. The paper further claims that the same construction converts into a loss, IDLoss, that is non-convex yet provably converges to stationary points via the Łojasiewicz inequality, and it demonstrates on four datasets that a dual-boosting learner trained with IDLoss and SERA (IDBoost) improves intersectional fairness measures while keeping predictive performance competitive. If these claims hold, an auditor gains a two-dimensional picture of model unfairness — which subgroup suffers, and at which prediction values — that single-attribute average-error metrics cannot show, plus a training objective that directly targets that unfairness.","feed_headline":"Regression fairness metric catches bias that average error hides","feed_subtitle":"Compares every protected-group combination at every relevance level, and its loss version provably converges.","key_machinery":"The load-bearing object is the relevance-weighted error curve of each intersectional subgroup: for a combination $\\alpha$ of protected-attribute values, $SER^\\alpha_t / |D^\\alpha_t|$ is the mean squared error over cases in that subgroup whose target relevance $\\phi(y) \\geq t$. ID is the integral over $t \\in [0,1]$ of the difference between the maximum and minimum of these curves, so it measures the largest error gap between any two protected groups at whatever target range matters; the ideal score is 0. The relevance function $\\phi: Y \\to [0,1]$ — supplied by a domain expert or interpolated from boxplot statistics, with extremes weighted 1 and the median near 0 — imports domain preferences into the measure, and is also the input that makes ID adjustable to different application stakes. IDLoss is the same integral with the current best subgroup $\\alpha_{\\min}$ excluded from the sum, which lowers every group's error toward the best group; its non-convexity comes from $\\alpha_{\\min}$ switching during optimization, and the convergence argument applies the Łojasiewicz inequality region by region, exploiting the finite number of possible $\\alpha_{\\min}$ identities. The experimental engine is IDBoost, which averages two boosted ensembles — one weighted by IDLoss and one by SERA — with a user-chosen performance/fairness weight.","core_discovery":"The central claim, stated the way the authors would put it, is that regression fairness is a two-dimensional quantity: it varies across the subgroups formed by every combination of protected attributes, and it varies across how relevant the predicted target value is to the user. ID formalizes this as the area between the highest and lowest normalized squared-error curves ($SER^\\alpha_t/|D^\\alpha_t|$) taken over all relevance thresholds $t$, so the score reports the worst gap between any two protected groups at any target range. The same construction yields IDLoss: instead of measuring the gap, optimize by summing every subgroup's error curve except the current best one, pulling the underperforming groups up toward the best rather than dragging the best group down. The paper's theoretical analysis claims that although $\\alpha_{\\min}$ can switch during training, making the loss non-convex, IDLoss satisfies the Łojasiewicz inequality within each region of constant $\\alpha_{\\min}$, the regions are finite, and gradient descent therefore converges to a stationary point with piecewise Lipschitz gradients. Empirically, ID surfaces biases invisible to single-attribute measures — in COMPAS the race disparity is near zero overall but large within each sex — and the authors report that IDBoost0.5 achieves the best average ranks among the fair-regression baselines compared on both ID and $\\Delta BGL$ while staying competitive in SERA.","pith_inferences":["Beyond the paper: the same area-between-curves construction could be applied to classification by swapping squared error for 0/1 or log loss, yielding an intersectional divergence that also handles class imbalance — a direction the paper does not pursue.","Beyond the paper: because ID reports a worst-case gap rather than an average over groups, it is sensitive to noise in the smallest subgroup; a bootstrap confidence band around the ID curve would tell auditors whether a reported divergence is real or sampling artifact.","Beyond the paper: ID's max-versus-min structure is a worst-case measure, so it naturally extends to a fairness budget — train with a constraint that the ID gap stay below a user-chosen threshold, turning the audit measure into a tunable fairness specification rather than a fixed objective weight.","Beyond the paper: a practical deployment recipe suggested by the construction is to audit each model under several relevance functions and report all verdicts — if the fairness ranking changes across them, the disagreement itself reveals how preference-dependent the model's fairness is."],"forward_implications":["Auditors can score any regression model with binary protected attributes by ID and read off, from a single curve, which intersectional subgroup is worst-predicted and at which relevance levels — information single-attribute MAE comparisons cannot provide.","Any algorithm with a differentiable loss can incorporate IDLoss, and the paper's convergence proof means fairness-aware training does not require ad-hoc constraint tuning or convexity assumptions.","The reported experiments imply that single-attribute fairness audits can both hide large bias (NLSY: race disparity 48.6% among men versus 3.3% among women) and misread overall near-zero disparity as fairness (COMPAS: 1.0% overall, but +13.9% for women and −1.4% for men).","Across the four datasets, IDBoost0.5 ranks best among the fairness-aware baselines on both ID and $\\Delta BGL$ while remaining competitive on SERA, supporting the paper's claim that intersectional, relevance-aware fairness does not require sacrificing predictive performance.","Because ID normalizes per subgroup, it remains defined when intersectional groups have very unequal sizes; the authors note the exponential runtime growth in the number of protected attributes and report a curve-simplification procedure that cuts training time by over 30% with little change in SERA or ID."],"supporting_citations":[{"why":"Defines SERA and the relevance function that ID is built from; ID is an area between two SER-type curves.","marker":"[41]"},{"why":"Proves SERA is twice-differentiable and usable as a loss, the differentiability template IDLoss extends.","marker":"[44]"},{"why":"The inequality the convergence proof applies inside each region where the best group stays fixed.","marker":"[31]"},{"why":"Cited for the convergence-condition framework behind the abstract's claim that optimizing IDLoss converges.","marker":"[27]"},{"why":"Defines Bounded Group Loss and Statistical Parity, the single-attribute baselines ID is compared against and reported to improve on.","marker":"[2]"},{"why":"Supplies the motivating evidence that race-by-gender error disparities hide from single-attribute audits.","marker":"[6]"},{"why":"Grounds the construction of subgroups from all combinations of protected attributes in intersectionality theory.","marker":"[16]"},{"why":"The tree-boosting algorithm that IDBoost is implemented on.","marker":"[11]"},{"why":"Supplies the NLSY feature setup and a nonconvex fair-regression baseline for comparison.","marker":"[28]"}],"fun_headline_variants":["Regression fairness: beyond single attributes","Intersectional fairness metric for regression catches hidden bias","Fairness for regression: measure all protected group combos","New metric reveals intersectional bias in regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything ID reports is relative to the relevance function $\\phi(Y)$ that decides which target values matter, and the paper gives no sensitivity analysis over it: if a defensible alternative choice of $\\phi$ (a different boxplot rule, or an expert weighting) changes which model is judged fairest, then ID's verdict is an artifact of that weighting rather than a property of the model.","fun_headline_variants_meta":{"raw":{"variants":["Regression fairness: beyond single attributes","Intersectional fairness metric for regression catches hidden bias","Fairness for regression: measure all protected group combos","New metric reveals intersectional bias in regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000896,"raw_usage":{"total_tokens":3888,"prompt_tokens":1001,"completion_tokens":2887,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":2829}},"tokens_in":617,"tokens_out":2887,"duration_ms":20313,"temperature":1.0,"reasoning_tokens":2829,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:34:02.007925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Two concrete tests. First, recompute ID and rerun the IDBoost comparison on the same four datasets under two or more defensible relevance functions — the default boxplot interpolation, a wider or narrower whisker rule, and an expert-supplied weighting — and check whether the ranking of models by ID, and the fairness improvements claimed for IDBoost0.5, survive; if the 'fairest model' verdict flips, ID is measuring the chosen weighting, not the model. Second, build a dataset with two protected groups whose normalized squared errors are exactly tied across a range of relevance thresholds, making $\\alpha_{\\min}$ non-unique, and run gradient descent on IDLoss in that tie regime; if it fails to converge or cycles between regions, the finite-region-transition argument behind the convergence guarantee breaks at its boundary case.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines SERA and the relevance function that ID is built from; ID is an area between two SER-type curves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves SERA is twice-differentiable and usable as a loss, the differentiability template IDLoss extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The inequality the convergence proof applies inside each region where the best group stays fixed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Bounded Group Loss and Statistical Parity, the single-attribute baselines ID is compared against and reported to improve on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Grounds the construction of subgroups from all combinations of protected attributes in intersectionality theory."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NLSY feature setup and a nonconvex fair-regression baseline for comparison."}],"review_version":1}