Pith. sign in

REVIEW 4 major objections 3 minor 26 references

Expected regret equals the covariance between costs and decisions, and its steepest-descent direction is the contrarian policy -(c-c̄), so regret minimization reduces to subtracting cost covariance from the decision rule.

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 14:05 UTC pith:QPGH3VCO

load-bearing objection The paper's central optimization result is wrong: tr(AΣ_c) is linear in A and unbounded below, so A*=0 is not a minimizer, and the convergence results collapse. the 4 major comments →

arxiv 2607.18866 v2 pith:QPGH3VCO submitted 2026-07-21 econ.EM cs.LGstat.ML

Optimizing Regret

Irene Aldridge This is my paper
classification econ.EM cs.LGstat.ML
keywords regret minimizationcovariance regret identityGâteaux derivativelinear decision policiesminimum-variance portfoliostochastic optimizationdecision-focused learningfinite-sample bounds
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 tries to turn regret minimization into calculus. It starts from a companion identity, expected regret equals the covariance between costs and decisions, and differentiates that covariance functional with respect to the policy. The Gâteaux derivative is again a covariance, and it is independent of the current policy; hence the steepest-descent direction is always the contrarian policy -(c-c̄). For linear decision rules π̂(c)=Ac+b, regret becomes tr(AΣ_c), whose gradient is the cost covariance matrix Σ_c and whose Hessian is zero, so the optimum lies on the boundary of the feasible policy set—in the unconstrained problem, at A=0, the minimum-variance policy. If the paper is right, regret minimization is a zero-curvature problem with geometric convergence, and minimizing regret versus maximizing alpha are just descent and ascent on one gradient.

Core claim

The central discovery is that the covariance regret functional C[π̂]=Cov(c,π̂(c)) is closed under differentiation. For any direction φ, DC[π̂](φ)=Cov(c,φ(c)), so the steepest-descent direction is φ*(c)=-(c-c̄), a contrarian/mean-reversion policy, regardless of the current policy. In the linear-policy class, C[A,b]=tr(AΣ_c), giving matrix gradient Σ_c, zero Hessian, and a unique unconstrained minimizer A=0—the minimum-variance portfolio with zero regret. The paper derives KKT conditions for constrained versions, recovers the minimum-variance portfolio and a regret-return frontier, proves a sign duality under which alpha maximization is the same gradient with the opposite sign, and gives finit

What carries the argument

The covariance functional C[π̂]=Cov(c,π̂(c)) and its Gâteaux derivative — a direction-wise derivative of that functional. The derivative theorem states that the first-order change in regret from perturbing the policy by φ is Cov(c,φ(c)), which makes c-c̄ the gradient. In the linear class, this specializes to the matrix gradient Σ_c; because the functional is linear in the policy matrix A, its Hessian vanishes, so every unconstrained optimum sits on the boundary of the policy set and gradient descent contracts at a rate set by the condition number of Σ_c.

Load-bearing premise

The whole derivative theory rests on the imported identity E[R(c)] = Cov(c, π̂(c)), asserted to hold whenever E[π̂(c)] = π*; that condition is not shown to be sufficient for realized optimal decisions, and if the identity fails, the descent direction, matrix gradient, and algorithms all lose their foundation.

What would settle it

Take c equally likely to be -1 or +1, and a policy π̂(c)=a c, so E[π̂]=0. The covariance Cov(c,π̂(c)) is a. Against the pointwise optimal decision π*(c)=-c with action set [-1,1], the true expected regret is a+1, which differs from a even though E[π̂]=E[π*]. Checking whether this violates the paper's stated condition would settle whether the imported identity, and therefore the derivative theory, has a valid domain.

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

If this is right

  • Unconstrained regret minimization over linear policies ends at A=0, the minimum-variance-style policy; any nonzero cost sensitivity adds regret in proportion to tr(AΣ_c).
  • Gradient descent on regret needs only observed cost vectors: the update A_{k+1} = Π_Z(A_k - ηΣ̂_c) does not require decision outputs or outcome labels.
  • A tilt away from the minimum-variance portfolio by ΔA costs tr(ΔAΣ_c), so the least regret is incurred along directions where cost covariance is smallest.
  • Regret minimization and alpha maximization are the same update with opposite signs: subtract ηΣ_c to minimize regret, add ηΣ_c to maximize alpha.
  • Under sub-Gaussian (light-tailed) costs, excess regret ε is reached in about κ ln(C_0/ε) gradient steps, with sample complexity O(σ^4 d log(d/δ)/ξ^2) for gradient estimation.

Where Pith is reading between the lines

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

  • Editorial extension: if the covariance identity survives scrutiny, the same differentiation argument should generalize to any bilinear objective Cov(data, policy), yielding contrarian updates in nonlinear policy classes through the Hessian formula in Theorem 4.2.
  • Editorial extension: the bias–regret decomposition in Section 10 suggests a model-selection rule—choose the fitted model with the smallest induced decision sensitivity tr(A_m Σ_c)—which predicts underfitting whenever extra model complexity raises decision sensitivity. This is testable by comparing holdout regret across polynomial degrees.
  • Editorial extension: the zero-curvature structure suggests that generic online-convex-optimization rates are unnecessarily pessimistic for this functional; if correct, exact geometric convergence is available whenever the cost covariance is estimable and positive definite.

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

4 major / 3 minor

Summary. The paper claims that expected regret equals Cov(c, π̂(c)) (from the author's prior work [Ald26]), and builds a derivative theory of this covariance functional. For linear policies π̂(c)=Ac+b, it derives the gradient Σ_c, asserts a zero Hessian, identifies A=0 as the unique unconstrained minimizer, and develops constrained KKT solutions, a sign-gradient duality, projected gradient algorithms with linear convergence, and finite-sample bounds. A numerical shortest-path experiment and a finance application are used to illustrate the theory.

Significance. If valid, the paper would provide a very simple, input-only gradient method for regret minimization and a closed-form relation between regret and cost covariance. The claimed sign-gradient duality between return maximization and regret minimization is an elegant observation, and the idea of estimating the gradient from cost observations alone is practically appealing. However, the central optimization claims are internally inconsistent: the covariance functional is linear in A and unbounded below in the unconstrained class, A=0 is not a minimizer, the Hessian statements contradict each other, and the constrained KKT solutions do not satisfy their own stationarity equations. Because these errors infect the main theorems, algorithms, and bounds, the paper's substantive contributions are not established.

major comments (4)
  1. [§3.1, Eq. (2)] The paper's central identity E[R] = Cov(c, π̂(c)) is imported from [Ald26] without proof, and the stated sufficient condition E[π̂(c)] = π* does not imply it. For a realized optimal decision π*(c), E[c^T π*(c)] is not generally equal to c̄^T E[π*(c)], so E[R] and Cov(c, π̂) can differ. Since all subsequent derivative and algorithm results rely on this equality, the foundational claim is unsupported.
  2. [§3.3, Corollary 3.3 and §4, Theorem 4.1] These two results are inconsistent. Corollary 3.3 states that the Hessian of C with respect to vec(A) is Σ_c ⊗ I_n, while Theorem 4.1 states that all second derivatives are zero. For C[A,b]=tr(AΣ_c), the second derivative is indeed zero; Corollary 3.3's Hessian is incorrect. This contradiction is load-bearing because the zero-curvature claim is used later to justify linear convergence in §2, §8, and §9.
  3. [§3.3, §4, §8, §9] C[A,b]=tr(AΣ_c) is linear in A. On the unconstrained linear-policy class, if Σ_c ≻ 0, the infimum is −∞ (take A=−tI and let t→∞), and A=0 is not a stationary point since ∇_A C=Σ_c ≠ 0. Theorem 4.1's boundary conclusion does not apply to the unconstrained problem, and A*=0 is not a minimizer. Consequently, the projected gradient update A_{k+1}=A_k−ηΣ̂_c diverges, and Theorems 8.1, 9.2, and 9.3, which all assert convergence to A*=0, are invalid.
  4. [§5.2, Eqs. (12)–(14), Corollaries 5.1 and 5.2] For κ=0, stationarity gives ∇_A L = Σ_c + (μ1−ν)c̄^T = 0 and ∇_b L = μ1−ν = 0, jointly implying Σ_c=0, which contradicts Σ_c ≻ 0. Thus the proposed A*=0, b*=Σ_c^{-1}1/(1^TΣ_c^{-1}1) is not a KKT point. For Corollary 5.2, substituting A*=κΣ_c^{-1} r̄ c̄^T into tr(AΣ_c) gives κ c̄^T r̄, not the stated κ* c̄^TΣ_c^{-1} r̄ · tr(Σ_c) in Eq. (14).
minor comments (3)
  1. [§10, Figures 1–2] The text says the experiments use 10,000 simulated samples, while the figure captions say 100,000 samples. Please reconcile.
  2. [§10, Eq. (22)] The generating model c_{ij}(X) is not fully specified: X_{0,i}, deg, ϵ_i, and p are undefined, and the meaning of 'deg' in the simulation is unclear.
  3. [Notation throughout] The symbol C is used both for the feasible cost set and for the covariance functional, which is confusing near Eq. (2) and §5.

Circularity Check

3 steps flagged

Regret-optimal conclusions are the covariance identity from the author's own prior work restated; the 'universal descent direction' reduces to the covariance functional's own gradient, and the MVP optimum is a known result imported under a new label.

specific steps
  1. self definitional [Section 3.2, Corollary 3.2]
    "The steepest-descent direction for C in L2(P ) is always ϕ∗(c) = −(c−c¯), regardless of ˆπ; ... Economically, ϕ∗(c) = −(c−c¯) is a contrarian/mean-reversion strategy ... a first-principles derivation of why contrarian policies minimize regret and momentum policies maximize alpha."

    C is defined in Eq. (2) as Cov(c, π̂(c)), and the regret interpretation comes only from [Ald26]. Corollary 3.2 minimizes Cov(c, ϕ) over directions ϕ; by Cauchy-Schwarz the minimizer is −(c−c̄). So the 'universal descent direction' is exactly the negative of the centered cost vector that appears in the covariance inner product defining C. The claim that contrarian policies minimize regret is therefore the input identity E[R]=Cov(c,π̂) restated in optimization form; no new information is added.

  2. self citation load bearing [Section 3.1, after Eq. (2)]
    "By [Ald26], C[ˆπ] = E[R] whenever E[ˆπ(c)] = π∗."

    This sentence is the only bridge between the covariance functional and the paper's subject ('regret'). It is attributed solely to the author's own prior arXiv paper, not proved or independently verified here. Every later result about regret — the descent direction, A*=0, MVP, finite-sample bounds — is a statement about this C. If the cited identity is unverified, the entire regret interpretation is forced by a self-citation chain: the paper's central premise reduces to its own previous work.

  3. renaming known result [Section 5.2, Corollary 5.1]
    "For Σc ≻ 0, the budget-constrained regret minimum over linear policies is A∗ = 0, b∗ = Σ−1 c 1/(1⊤Σ−1 c 1), the minimum-variance portfolio."

    From Eq. (5), C[A,b]=tr(AΣc) is independent of b. The proposed b∗ solves a variance-minimization problem (min b^TΣc b subject to 1^T b=1), not the paper's regret objective. The known Markowitz minimum-variance portfolio is imported and relabeled as the 'regret minimum' without being derived from C. This is the paper's own framework renaming a known result rather than obtaining it from 'regret equals covariance'.

full rationale

The Gâteaux derivative computation in Theorem 3.1 is mathematically self-contained as a statement about covariance functionals; I do not flag that calculation itself as circular. The circularity is at the level of interpretation and premises. The paper's advertised 'first-principles' conclusions about regret all flow from the identity E[R]=Cov(c,π̂), which is taken from the author's own [Ald26] and not proved or independently supported here. Corollary 3.2 then shows that the steepest descent direction for this covariance functional is −(c−c̄); since regret is asserted to be that same covariance, the 'contrarian policies minimize regret' prediction is the input identity rewritten in variational form, not an independent result. Similarly, Corollary 5.1 labels the classic minimum-variance portfolio as the regret-optimal solution even though C[A,b]=tr(AΣc) does not determine b; the MVP is a known result from a different objective, imported and renamed within the new 'covariance regret' vocabulary. These steps make the central claim partially circular: the derivative theory is real mathematics, but the regret-specific content reduces to a self-cited identity and to a relabeled known portfolio result. Correctness issues, such as the unboundedness of the linear objective, are outside this circularity pass.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. The 'asset-pair gaps' Δ_ij are just entries of the cost covariance matrix. The main burden falls on the unproven covariance regret identity from the author's own earlier work, plus standard distributional and model-class assumptions. Several of the stated auxiliary assumptions (e.g., A=0 as unique minimizer) are not merely unproven but incorrect.

free parameters (3)
  • η (gradient step size) = 1/||Σ_c|| or 1/(||Σ_c||+ξ)
    Chosen by hand and used in Algorithm 8.1 and convergence bounds; affects the claimed rate but is not fitted to data.
  • κ* (return-constraint multiplier) = chosen so that the return constraint binds
    Introduced in Corollary 5.2 to parameterize the alleged efficient frontier; effectively fitted to the target return μ0.
  • C (universal constant in sample complexity)
    Appears in Theorem 9.3's bound; not specified numerically, so the bound is not quantitatively usable.
axioms (5)
  • domain assumption Expected regret equals the covariance between costs and decisions: E[R(c)] = Cov(c, π̂(c)) whenever E[π̂(c)] = π*
    Taken from [Ald26] and used throughout (§3.1). Not proven here and the stated condition is insufficient in general.
  • domain assumption Cost vectors are sub-Gaussian
    Assumed in §9 to derive sample complexity and convergence bounds.
  • domain assumption The policy class is linear: π̂(c) = Ac + b
    Used for the main gradient, Hessian, and convergence results (§3.3, §4, §8).
  • ad hoc to paper A = 0 is the unique unconstrained minimizer of the covariance regret
    Claimed in §4 and §5 but false for an unconstrained linear objective tr(AΣ_c), which is unbounded below when Σ_c ≻ 0.
  • domain assumption The optimal decision rule z*(·) is locally linear in the cost model
    Used in §10.1 to justify the bias–regret decomposition and the comparison of polynomial fits.

pith-pipeline@v1.3.0-alltime-deepseek · 6772 in / 15839 out tokens · 129178 ms · 2026-08-01T14:05:12.331694+00:00 · methodology

0 comments
read the original abstract

Building on the identity that expected regret equals the covariance between costs and decisions, this paper develops a derivative theory of the covariance regret functional. We derive the G\^ateaux derivative, showing that the universal steepest-descent direction is the contrarian policy $-(c-\bar c)$, while ascent yields momentum. For linear policies $\hat\pi(c)=Ac+b$, the gradient is the cost covariance matrix $\Sigma_c$, with a zero Hessian implying boundary-optimal solutions such as the minimum-variance portfolio. We extend to constrained optimization, sign-gradient duality between regret minimization and alpha maximization, finite-sample convergence bounds paralleling Thompson Sampling, and gradient-descent algorithms requiring only input observations.

Figures

Figures reproduced from arXiv: 2607.18866 by Irene Aldridge.

Figure 1
Figure 1. Figure 1: Average performance of different loss models over 100,000 samples in the shortest-path [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average performance of different loss models over 100,000 samples in the shortest-path [PITH_FULL_IMAGE:figures/full_fig_p010_2.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

26 extracted references · 1 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2605.14019 [econ.EM] , year =

    Aldridge, Irene , title =. arXiv preprint arXiv:2605.14019 [econ.EM] , year =

    Pith/arXiv arXiv

  2. [2]

    Proceedings of the 16th International Conference on Artificial Intelligence and Statistics (AISTATS) , series =

    Agrawal, Shipra and Goyal, Navin , title =. Proceedings of the 16th International Conference on Artificial Intelligence and Statistics (AISTATS) , series =. 2013 , url =

    2013

  3. [3]

    Prediction, Learning, and Games , publisher =

    Cesa-Bianchi, Nicol\`. Prediction, Learning, and Games , publisher =. 2006 , doi =

    2006

  4. [4]

    , title =

    Cover, Thomas M. , title =. Mathematical Finance , volume =. 1991 , doi =

    1991

  5. [5]

    Foundations and Trends in Optimization , volume =

    Hazan, Elad , title =. Foundations and Trends in Optimization , volume =. 2016 , doi =

    2016

  6. [6]

    Machine Learning , volume =

    Hazan, Elad and Agarwal, Amit and Kale, Satyen , title =. Machine Learning , volume =. 2007 , doi =

    2007

  7. [7]

    Econometrica , volume =

    Hart, Sergiu and Mas-Colell, Andreu , title =. Econometrica , volume =. 2000 , doi =

    2000

  8. [8]

    and Schapire, Robert E

    Helmbold, David P. and Schapire, Robert E. and Singer, Yoram and Warmuth, Manfred K. , title =. Mathematical Finance , volume =. 1998 , doi =

    1998

  9. [9]

    SIAM Journal on Optimization , volume =

    Nemirovski, Arkadi , title =. SIAM Journal on Optimization , volume =. 2004 , doi =

    2004

  10. [10]

    SIAM Journal on Optimization , volume =

    Nesterov, Yurii , title =. SIAM Journal on Optimization , volume =. 2005 , doi =

    2005

  11. [11]

    Proceedings of the 28th International Conference on Algorithmic Learning Theory (ALT) , series =

    Orseau, Laurent and Lattimore, Tor and Legg, Shane , title =. Proceedings of the 28th International Conference on Algorithmic Learning Theory (ALT) , series =. 2017 , url =

    2017

  12. [12]

    Shalev-Shwartz, Shai , title =

  13. [13]

    Foundations and Trends in Machine Learning , volume =

    Shalev-Shwartz, Shai , title =. Foundations and Trends in Machine Learning , volume =. 2012 , doi =

    2012

  14. [14]

    arXiv preprint arXiv:2510.04407 [cs.GT] , year =

    Zhang, Brian Hu and Anagnostides, Ioannis and Sandholm, Tuomas , title =. arXiv preprint arXiv:2510.04407 [cs.GT] , year =

    arXiv

  15. [15]

    Proceedings of the 20th International Conference on Machine Learning (ICML) , pages =

    Zinkevich, Martin , title =. Proceedings of the 20th International Conference on Machine Learning (ICML) , pages =. 2003 , url =

    2003

  16. [16]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Zinkevich, Martin and Bowling, Michael and Johanson, Michael and Piccione, Carmelo , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =. 2007 , url =

    2007

  17. [17]

    , title =

    Chen, Hao and Didisheim, Antoine and Somoza, Luis A. , title =. NBER Working Paper No.\ 34965 , year =

  18. [18]

    arXiv preprint arXiv:2304.07619 [q-fin.ST] , year =

    Lopez-Lira, Alejandro and Tang, Yuehua , title =. arXiv preprint arXiv:2304.07619 [q-fin.ST] , year =

    arXiv

  19. [19]

    SSRN Working Paper No.\ 5217505 , year =

    Lopez-Lira, Alejandro and Tang, Yuehua and Zhu, Mao , title =. SSRN Working Paper No.\ 5217505 , year =

  20. [20]

    Journal of Finance , volume =

    Markowitz, Harry , title =. Journal of Finance , volume =. 1952 , doi =

    1952

  21. [21]

    and Barto, Andrew G

    Sutton, Richard S. and Barto, Andrew G. , title =

  22. [22]

    Tyrrell and Wets, Roger J.-B

    Rockafellar, R. Tyrrell and Wets, Roger J.-B. , title =. 2009 , doi =

    2009

  23. [23]

    Journal of Multivariate Analysis , volume =

    Ledoit, Olivier and Wolf, Michael , title =. Journal of Multivariate Analysis , volume =. 2004 , doi =

    2004

  24. [24]

    , title =

    Tropp, Joel A. , title =. Foundations and Trends in Machine Learning , volume =. 2015 , doi =

    2015

  25. [25]

    and Grigas, Paul , title =

    Elmachtoub, Adam N. and Grigas, Paul , title =. Management Science , volume =

  26. [26]

    2024 , url=

    Decision-Focused Learning with Directional Gradients , author=. 2024 , url=

    2024