Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure

Mingyuan Zhang

Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure

T0 review · 0 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Exact convex calibration of the multi-label Jaccard loss requires exponentially many prediction coordinates.

desk verdict This paper's exponential lower bound on the exact convex calibration dimension of the Jaccard loss is new, well-proven, and worth refereeing carefully. read the letter →

arxiv 2608.13549 v1 pith:Q53YL4GU submitted 2026-08-13 cs.LG stat.ML

Mingyuan Zhang This is my paper · ORCID

classification cs.LGstat.ML

keywords multi-labelclassificationJaccardlossintersectionoverunionconvexcalibrationdimensionloss-matrixrankapproximateconsistencyMinHashrandomfeaturesF-measure

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper studies the per-instance Jaccard score (intersection over union) as a multi-label loss and asks how many real-valued prediction coordinates a convex surrogate must have to be exactly calibrated over all conditional label distributions. Its main theorem states that this convex calibration dimension is exponential in the number of labels: $2^{s-1} \leq \mathrm{CCdim}(L^{\mathrm{Jac}}) \leq 2^s - 1$ for $s$ labels, under the convention $\mathrm{Jac}(\varnothing,\varnothing)=1$. Along the way it proves that the Jaccard score, shifted-loss, and ordinary loss matrices have full rank $2^s$ and affine dimension $2^s - 1$. The paper also shows a sharp contrast: approximate prediction with any fixed additive regret tolerance is possible in polynomial dimension, via an $F_1$-to-Jaccard regret transfer and MinHash random-feature surrogates.

What carries the argument

The load-bearing construction is the factorially weighted witness distribution $q$ on the family $U = \{\{1\}\cup D : D \subseteq [s]\setminus\{1\}\}$, assigning weight proportional to $1/|D|!$ to outcome $\{1\}\cup D$. The factorial-balancing identity of Lemma 5.1 makes the expected Jaccard score of every report in $U$ equal to a common constant $\kappa$; mixing $q$ with the empty outcome makes the empty report tie as well, producing a set $A$ of $2^{s-1}+1$ tied Bayes-optimal reports. The active score submatrix indexed by $A$ is $\mathrm{diag}(1, S_{U,U})$, nonsingular by the MinHash Gram representation and Boolean Möbius inversion (which together prove the full Jaccard matrix is positive definite), so the relevant difference span is the whole orthogonal hyperplane and the two-sided feasible subspace is zero-dimensional; the lower-bound inequality then yields the exponential bound. The MinHash feature map $\Phi(A)$, whose inner product is the empirical collision probability of random permutations, carries the approximation results.

What would settle it

Evaluate the theorem for $s=2$, where it claims $2 \leq \mathrm{CCdim}(L^{\mathrm{Jac}}) \leq 3$ on a $4 \times 4$ loss matrix: an exhaustive computer search over convex surrogate-link pairs in one real dimension would either find an exactly calibrated one-dimensional surrogate, refuting the lower bound, or confirm the bound. A cheaper check is to test whether the witness $p$ defined in Eq. (42) lies in the relative interior of its trigger set; the external lower-bound inequality used in the proof may require an interior point, and if $p$ lies on a face the $2^{s-1}$ claim collapses.

Watch

Extended reading notes

Core claim

The central discovery is that exact convex calibration of the per-instance Jaccard loss is exponentially harder in terms of prediction dimension than its polynomial-sized output space might suggest. Every exactly calibrated convex surrogate must use at least $2^{s-1}$ real-valued coordinates, and an affine surrogate with $2^s - 1$ coordinates suffices, so $\mathrm{CCdim}(L^{\mathrm{Jac}}) = \Theta(2^s)$. The proof constructs a factorially weighted distribution over the $2^{s-1}+1$ outcomes consisting of the empty set and all sets containing a fixed core label; a combinatorial identity makes all corresponding Bayes-optimal reports tie exactly, the active score submatrix is nonsingular, and the two-sided feasible subspace is trivial, triggering a known lower-bound inequality from convex calibration theory. The same paper establishes that the Jaccard score and loss matrices have maximal rank and column-affine dimension, and gives two polynomial-dimensional approximation routes: an $F_1$-to-Jaccard transfer yielding asymptotic regret at most $3-2\sqrt{2}$, and MinHash square-loss surrogates that are $\alpha$-approximately consistent in dimension $O((s^2 + s\log(1/\rho))/\alpha^2)$ or $O((s + \log(1/\rho))/\alpha^2)$.

Load-bearing premise

The exponential lower bound depends on an external theorem that says the convex calibration dimension is at least the number of supported outcomes minus the dimension of two-sided feasible directions minus one; the paper applies that theorem to its factorial tied-report witness without re-proving it or checking its regularity hypotheses at that boundary point.

Editorial extensions

If this is right

Every exactly calibrated convex surrogate for the per-instance Jaccard loss uses at least $2^{s-1}$ real-valued coordinates; in particular no polynomial-dimensional convex surrogate can be exactly calibrated over all conditional distributions.
The Jaccard score, shifted-loss, and ordinary loss matrices are nonsingular and the loss columns have affine dimension $2^s - 1$, so an affine surrogate in $2^s - 1$ dimensions achieves exact calibration.
An $F_1$-Bayes classifier has worst-case asymptotic Jaccard regret at most $3 - 2\sqrt{2} \approx 0.1716$, and any $F_1$-calibrated surrogate inherits this constant regret floor for Jaccard.
For any tolerance $\alpha > 0$ and confidence $\rho$, the MinHash square-loss surrogate is $\alpha$-approximately consistent in dimension $O((s^2 + s\log(1/\rho))/\alpha^2)$, and the signed variant in $O((s + \log(1/\rho))/\alpha^2)$.
Zero-regret prediction requires exponential dimension, while every fixed additive regret tolerance admits polynomial dimension, so the exponential lower bound is specific to exact calibration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The factor-of-two gap between the bounds suggests the true value of $\mathrm{CCdim}(L^{\mathrm{Jac}})$ may be $2^s - 1$, matching the affine-dimension upper bound; extending the factorial witness to a full-support distribution with zero lineality would close the gap.
The $O(s/\alpha^2)$ dimension of the signed MinHash surrogate matches the parametric rate for estimating an $s$-dimensional structure, hinting that the $\alpha$-approximate dimension may be $\Omega(s/\alpha^2)$ and that the exponential cost is an artifact of requiring exactly zero regret.
The factorial-balancing tie construction is a promising template for establishing exponential calibration dimension for other nondecomposable ratio losses, such as Dice or Tversky, whose loss matrices have similar combinatorial structure.
The MinHash surrogate's dimension bound is derived via uniform approximation of the entire score matrix; a data-dependent or Laplacian analysis could reduce the $\log(1/\rho)$ factor or the $4^s$ union-bound cost.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague — the headline is that this paper proves a clean exponential lower bound on the exact convex calibration dimension of the per-instance Jaccard loss: 2^{s-1} <= CCdim <= 2^s - 1. That's new and it's the main thing to know. The upper bound comes from affine dimension, while the lower bound uses a factorial witness distribution that forces 2^{s-1}+1 Bayes-optimal reports to tie, then shows the two-sided feasible subspace is trivial. I checked the balancing identity, the active-submatrix argument, and the mu = 0 computation; all hold. The paper also pins down the rank and affine dimension under the Jac(empty,empty)=1 convention via a MinHash Gram matrix plus Boolean Mobius inversion, which is a cleaner proof than the earlier infinite positive-definite decomposition. The two approximation results are genuinely useful context: an F1-to-Jaccard transfer with constant 3-2*sqrt(2), and MinHash square-loss surrogates with polynomial dimension for any fixed regret floor. The MinHash analysis is honest that the link still maximizes over 2^s reports, so this is a prediction-dimension theorem, not a tractable-inference one.

Soft spots: the lower bound imports the Ramaswamy-Agarwal inequality CCdim(L) >= ||p||_0 - mu - 1, and the paper does not re-prove it. I don't see evidence that the theorem fails for this trigger set; the stress-test's local check of mu is sound. But a referee should verify the regularity conditions against the original source. Minor: the factor-of-two gap remains, and the exact value of CCdim is open; the authors say so. The F1-transfer result depends on the published Zhang-Ramaswamy-Agarwal surrogate, which is fine. The paper also cites a companion F1 paper (Zhang 2026) for the quadratic CCdim of F1; that seems appropriate, but a reviewer may want to confirm that result is real or in press.

The empty-set convention deserves attention: the exponential conclusion survives under Jac(empty,empty)=0, but the lower bound shifts to 2^{s-1}-1, so the headline constant depends on a convention choice. That's not a flaw, but it should be stated loudly, and it is in the appendix.

Verdict: this is a real result with complete proofs, a clear related-work section, and no red flags in the citation pattern. The external dependency is the only thing I'd want a careful referee to check. I'd bring it to reading group and would cite it. Send it to peer review.

Referee Report

0 major / 4 minor

Summary. This paper studies the convex calibration dimension (CCdim) of the per-instance multi-label Jaccard loss. It proves that, under the convention Jac(∅,∅)=1, the Jaccard score matrix, the shifted-loss matrix, and the ordinary loss matrix are all nonsingular and have affine dimension 2^s−1. The main result is Theorem 5.2: 2^{s−1} ≤ CCdim(L^{Jac}) ≤ 2^s−1, so exact convex calibration requires Θ(2^s) prediction coordinates. The lower bound is obtained from a factorially weighted distribution over sets containing a core label, mixed with the empty set, and a feasible-subspace argument that gives a trivial lineality space. The paper also provides two polynomial-dimensional approximation guarantees: an F1-to-Jaccard regret transfer with asymptotic regret floor 3−2√2, and MinHash-based square-loss surrogates that achieve any fixed additive regret floor α in dimension O((s+log(1/ρ))/α^2). The paper thus establishes an exactness–dimension tradeoff.

Significance. If the main theorem is correct, it is a significant contribution to the theory of convex calibration. It shows that the instance-wise Jaccard loss, despite having a simple closed-form score, is substantially harder to calibrate than the F1 measure (whose calibration dimension is Θ(s^2) in related work): no convex surrogate can be exactly calibrated unless it uses exponentially many coordinates. The paper also gives constructive, distribution-free approximate surrogates with explicit regret bounds and dimension guarantees, which is practically relevant. The proofs are detailed and mostly self-contained; the two external inputs—the Ramaswamy–Agarwal lower bound and the Zhang et al. F1 surrogate—are cited precisely and appear to be used correctly. The finite MinHash Gram representation and the factorial balancing identity are elegant tools that may be of independent interest.

minor comments (4)

[Appendix C, Eq. (46)] The derivation of the lineality space is compressed into a single sentence ('It follows directly from the simplex constraints and the active trigger inequalities'). Since this is the crux of the exponential lower bound, I recommend spelling out the three conditions: two-sided feasibility forces v_{A^c}=0; the normalization constraint gives 1_A^T v_A=0; and preserving the active ties requires v_A ⊥ E. Then using E=p_A^⊥ yields v_A=0. This would make the proof easier to verify.
[Section 3 and references] The paper relies on the unpublished preprint Zhang (2026) for the F1 calibration dimension and as a comparison baseline. The authors should indicate its preprint status (and ideally update the reference if it has been published) and confirm that its empty-set convention aligns with the convention F(∅,∅)=1 used here.
[Section 6.2, Theorems 6.3 and 6.4] The link rules in (25) and (31) are defined via argmax with 'any fixed rule to break ties.' The proofs are valid for any deterministic tie-breaking rule, but it would be clearer to state explicitly that the regret bounds hold uniformly over all tie-breaking rules, so that no hidden dependence on the tie-breaking rule is introduced.
[Throughout] There are several typographical issues with superscripts and spacing: for example, '2 s −1' should be '2^s − 1' in the abstract and elsewhere, and 'CCdim(L Jac)' should be 'CCdim(L^{Jac})'. A careful proofreading pass would improve readability.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the exponential lower bound rests on a self-contained witness construction, and the only imported inequality is external and independent.

full rationale

The central claim, Theorem 5.2, is not circular. The lower bound is derived in the paper: a factorially weighted distribution q is constructed in Equation (39), Lemma 5.1 shows all reports in U tie under q, Equation (42) mixes in the empty outcome to make the empty report tie, and Equation (43) identifies the Bayes-optimal set A. Steps 3 and 4 of Appendix C then compute the active loss-difference space E and show that the two-sided feasible subspace is trivial, giving mu = 0; only the final substitution into inequality (5) uses the external general theorem of Ramaswamy and Agarwal (2016). That theorem is imported as an established result, not re-derived, but it does not contain the Jaccard-specific construction and is not a fitted input. The self-citations are not circular: Zhang et al. (2020) is a published, parameter-free F1 surrogate whose transfer is applied only for the constant-floor guarantee in Section 6.1, with Proposition 6.1 proved in the present paper; Zhang (2026) is only a comparison. The MinHash approximation results in Section 6.2 are self-contained. No quantity is defined in terms of the claimed prediction dimension, and no fitted parameter is renamed as a prediction. Therefore no circular step was identified.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central exponential lower bound is self-contained apart from the Ramaswamy-Agarwal dimension inequalities, which are standard external tools in this literature. Two further building blocks are imported: the published F1 surrogate (co-authored by the current author) for the constant-floor result, and standard probabilistic and analytic inequalities. No fitted parameters, hand-tuned constants dependent on data, or new postulated entities are introduced.

assumptions (5)

domain assumption Ramaswamy-Agarwal theorem: CCdim(L) <= affdim(L) and CCdim(L) >= ||p||_0 - mu - 1 for p in a trigger set.
The upper and lower bounds in Theorems 4.2 and 5.2 rely on these external inequalities (Eqs. (4) and (5), Section 2). The paper does not re-prove them.
domain assumption Existence of an (s^2+1)-dimensional convex calibrated surrogate and polynomial-time link for the multi-label F1 measure (Zhang et al., 2020).
Section 6.1's constant-floor Jaccard approximation inherits its F1 surrogate from this cited work; the construction and its proof are referenced, not reproduced.
standard math MinHash collision identity: for a uniform random permutation pi, Pr(m(A)=m(B)) = Jac(A,B), extended to empty sets with a special symbol.
Used for the Gram representation in Lemma 4.1 and for the feature maps in Section 6.2; the paper proves the nonempty case and extends it by definition to the empty set.
standard math Standard concentration and analytic facts: Hoeffding's inequality, Jensen's inequality, Cauchy-Schwarz, convexity of g(t)=t/(2-t), and concavity of H.
Used throughout the approximation proofs in Appendix D without proof.
standard math Boolean Möbius inversion on the subset lattice.
Used in Lemma 4.1 to conclude that vanishing subset sums imply the zero vector.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure." pith.science (2026). https://pith.science/paper/Q53YL4GU

@misc{pith2026260813549,
  author       = {Pith},
  title        = {Pith review of: Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q53YL4GU}},
  note         = {Machine review of arXiv:2608.13549}
}

abstract

The per-instance Jaccard score, or intersection over union (IoU), is standard in multi-label classification and binary segmentation. With $s$ labels, its loss matrix has $2^s$ outcomes and reports. Under the convention $\mathrm{Jac}(\varnothing,\varnothing)=1$, we prove that the Jaccard score, shifted-loss, and ordinary loss matrices are nonsingular and that the loss columns have affine dimension $2^s-1$. The proof combines a finite MinHash Gram representation with Boolean M\"obius inversion. For exact calibration, we prove $2^{s-1} \leq \mathrm{CCdim}(L^{\mathrm{Jac}}) \leq 2^s-1$. The lower bound uses a factorially weighted distribution with $2^{s-1}+1$ supported outcomes and Bayes-optimal reports. Consequently, every exactly calibrated convex surrogate requires exponentially many prediction coordinates. We also give two polynomial-dimensional approximation guarantees with explicit regret transfers. A new $F_1$-to-Jaccard transfer turns an existing $(s^2+1)$-dimensional $F_1$ surrogate into a polynomial-time rule with asymptotic Jaccard regret at most $3-2\sqrt{2}$. For any $\alpha>0$ and $0<\rho<1$, a MinHash square-loss surrogate attains Jaccard-regret floor $\alpha$ uniformly over arbitrary conditional label distributions. With probability at least $1-\rho$, the direct construction has dimension $O((s^2+s\log(1/\rho))/\alpha^2)$, while a signed variant has dimension $O((s+\log(1/\rho))/\alpha^2)$. Thus zero-regret calibration requires exponential dimension, whereas every fixed additive regret tolerance admits polynomial prediction dimension.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

[1]

Calibrated surrogate maximization of linear-fractional utility in binary classification

Han Bao and Masashi Sugiyama. Calibrated surrogate maximization of linear-fractional utility in binary classification. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, volume 108 of PMLR, pages 2337--2347, 2020

work page 2020
[2]

Blaschko

Maxim Berman, Amal Rannen Triki, and Matthew B. Blaschko. The Lov\'asz--Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4413--4421, 2018

work page 2018
[3]

A proof for the positive definiteness of the Jaccard index matrix

Mathieu Bouchard, Anne-Laure Jousselme, and Pierre-Emmanuel Dor\' e . A proof for the positive definiteness of the Jaccard index matrix. International Journal of Approximate Reasoning, 54(5):615--626, 2013

work page 2013
[4]

Andrei Z. Broder. On the resemblance and containment of documents. In Compression and Complexity of Sequences 1997, pages 21--29. IEEE, 1997

work page 1997
[5]

Finding the Jaccard median

Flavio Chierichetti, Ravi Kumar, Sandeep Pandey, and Sergei Vassilvitskii. Finding the Jaccard median. In Proceedings of the Twenty-First Annual ACM--SIAM Symposium on Discrete Algorithms, pages 293--311, 2010

work page 2010
[6]

RankSEG : A consistent ranking-based framework for segmentation

Ben Dai and Chunlin Li. RankSEG : A consistent ranking-based framework for segmentation. Journal of Machine Learning Research, 24(224):1--50, 2023

work page 2023
[7]

On label dependence and loss minimization in multi-label classification

Krzysztof Dembczy\' n ski, Willem Waegeman, Weiwei Cheng, and Eyke H\"ullermeier. On label dependence and loss minimization in multi-label classification. Machine Learning, 88(1--2):5--45, 2012

work page 2012
[8]

Finocchiaro, Rafael Frongillo, and Enrique B

Jessica J. Finocchiaro, Rafael Frongillo, and Enrique B. Nueve. The structured abstain problem and the Lov\'asz hinge. In Proceedings of the 35th Conference on Learning Theory, volume 178 of PMLR, pages 3718--3740, 2022

work page 2022

Show all 17 references

[9]

John C. Gower. A general coefficient of similarity and some of its properties. Biometrics, 27(4):857--871, 1971

1971
[10]

Koyejo, Nagarajan Natarajan, Pradeep K

Oluwasanmi O. Koyejo, Nagarajan Natarajan, Pradeep K. Ravikumar, and Inderjit S. Dhillon. Consistent multilabel classification. In Advances in Neural Information Processing Systems 28, pages 3321--3329, 2015

2015
[11]

Sharp analysis of learning with discrete losses

Alex Nowak, Francis Bach, and Alessandro Rudi. Sharp analysis of learning with discrete losses. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, volume 89 of PMLR, pages 1920--1929, 2019

1920
[12]

Ramaswamy and Shivani Agarwal

Harish G. Ramaswamy and Shivani Agarwal. Classification calibration dimension for general multiclass losses. In Advances in Neural Information Processing Systems 25, pages 2087--2095, 2012

2012
[13]

Ramaswamy and Shivani Agarwal

Harish G. Ramaswamy and Shivani Agarwal. Convex calibration dimension for multiclass loss matrices. Journal of Machine Learning Research, 17(14):1--45, 2016

2016
[14]

On the Bayes-optimality of F-measure maximizers

Willem Waegeman, Krzysztof Dembczy\' n ski, Arkadiusz Jachnik, Weiwei Cheng, and Eyke H\"ullermeier. On the Bayes-optimality of F-measure maximizers. Journal of Machine Learning Research, 15(103):3513--3568, 2014

2014
[15]

Learning submodular losses with the Lov\'asz hinge

Jiaqian Yu and Matthew Blaschko. Learning submodular losses with the Lov\'asz hinge. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of PMLR, pages 1623--1631, 2015

2015
[16]

Ramaswamy, and Shivani Agarwal

Mingyuan Zhang, Harish G. Ramaswamy, and Shivani Agarwal. Convex calibrated surrogates for the multi-label F-measure. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of PMLR, pages 11246--11255, 2020

2020
[17]

Exact rank and convex calibration dimension lower bounds for the multi-label F_1 loss

Mingyuan Zhang. Exact rank and convex calibration dimension lower bounds for the multi-label F_1 loss. arXiv preprint arXiv:2608.08399 https://arxiv.org/abs/2608.08399, 2026

2026 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Calibrated surrogate maximization of linear-fractional utility in binary classification

Han Bao and Masashi Sugiyama. Calibrated surrogate maximization of linear-fractional utility in binary classification. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, volume 108 of PMLR, pages 2337--2347, 2020

work page 2020

[2] [2]

Blaschko

Maxim Berman, Amal Rannen Triki, and Matthew B. Blaschko. The Lov\'asz--Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4413--4421, 2018

work page 2018

[3] [3]

A proof for the positive definiteness of the Jaccard index matrix

Mathieu Bouchard, Anne-Laure Jousselme, and Pierre-Emmanuel Dor\' e . A proof for the positive definiteness of the Jaccard index matrix. International Journal of Approximate Reasoning, 54(5):615--626, 2013

work page 2013

[4] [4]

Andrei Z. Broder. On the resemblance and containment of documents. In Compression and Complexity of Sequences 1997, pages 21--29. IEEE, 1997

work page 1997

[5] [5]

Finding the Jaccard median

Flavio Chierichetti, Ravi Kumar, Sandeep Pandey, and Sergei Vassilvitskii. Finding the Jaccard median. In Proceedings of the Twenty-First Annual ACM--SIAM Symposium on Discrete Algorithms, pages 293--311, 2010

work page 2010

[6] [6]

RankSEG : A consistent ranking-based framework for segmentation

Ben Dai and Chunlin Li. RankSEG : A consistent ranking-based framework for segmentation. Journal of Machine Learning Research, 24(224):1--50, 2023

work page 2023

[7] [7]

On label dependence and loss minimization in multi-label classification

Krzysztof Dembczy\' n ski, Willem Waegeman, Weiwei Cheng, and Eyke H\"ullermeier. On label dependence and loss minimization in multi-label classification. Machine Learning, 88(1--2):5--45, 2012

work page 2012

[8] [8]

Finocchiaro, Rafael Frongillo, and Enrique B

Jessica J. Finocchiaro, Rafael Frongillo, and Enrique B. Nueve. The structured abstain problem and the Lov\'asz hinge. In Proceedings of the 35th Conference on Learning Theory, volume 178 of PMLR, pages 3718--3740, 2022

work page 2022