Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Multidimensional specific relative entropy between continuous martingales

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read For multidimensional martingales, specific relative entropy has a sharp quadratic-variation lower bound, and the bound is the convex lower semicontinuous envelope of the entropy.

desk verdict Solid multidimensional extension of Gantert's inequality with a real but fixable gap in the convex-envelope theorem. read the letter →

arxiv 2411.11408 v1 pith:V4TSTATS submitted 2024-11-18 math.PR q-fin.MF

Julio Backhoff , Edoardo Kimani Bellotto This is my paper · ORCID
classification math.PRq-fin.MF MSC 60G4460J6594A17
keywords specificrelativeentropymartingalesWienermeasurequadraticvariationGantert'sinequalityconvexlowersemicontinuousenvelopeBlack-Scholesmodeltensorization
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

Continuous-time martingale laws are typically mutually singular, so ordinary relative entropy between them is infinite; the specific relative entropy rescales relative entropies on dyadic time grids and lets the leading rate emerge. This paper carries that construction over to $\mathbb{R}^l$-valued martingales, defining $h_l(Q|P)=\liminf_{n\to\infty}2^{-n}H(Q|P)|F_{2^n}$. The main result is a multidimensional version of Gantert's inequality: for a square-integrable martingale measure $Q$ with absolutely continuous quadratic covariation $d\langle X\rangle_t/dt = \Sigma_t^Q$ and $X_0=0$, $h_l(Q|B^l) \geq E_Q[\int_0^1 F_l(\Sigma_t^Q)dt]$, where $F_l(\Sigma)=\frac12(\operatorname{tr}\Sigma - l - \log\det\Sigma)$. The paper then proves this lower bound is not arbitrary: it is exactly the convex lower semicontinuous envelope of the specific relative entropy, the largest convex lower semicontinuous minorant, which is new even in dimension one. Closed-form formulas for multidimensional Black-Scholes models show where the inequality is an equality.

What carries the argument

The engine is the matrix function $F_l(\Sigma)=\frac12(\operatorname{tr}\Sigma - l - \log\det\Sigma)$, which is convex, nonnegative, and vanishes only at the identity, together with the dyadic scaling limit $h_l(Q|P)=\liminf_{n\to\infty}2^{-n}H(Q|P)|F_{2^n}$ that makes sense of entropy between singular martingale laws. The convex-envelope theorem is carried by two further devices: a pathwise quadratic covariation density that defines the density of $\langle X\rangle$ simultaneously for all continuous martingale measures and makes the functional affine, and the approximating class $\mathcal{M}_l^2(N)$ of martingales whose diffusion matrices are predictable at the previous grid points, for which the specific relative entropy is computed exactly as $E[\int_0^1 F_l(\sigma_t^2)dt]$. With those pieces, the convex conjugate of $h_l$ is shown to coincide with that of the quadratic-variation functional $L_l$.

What would settle it

Take $Q$ to be the law of $X_t = x_0 + B_t$ with $x_0 \neq 0$ and $B$ a standard $\mathbb{R}^l$ Brownian motion, so $X_0 = x_0$ while $B^l$ starts at $0$. Then $E_Q[\int_0^1 F_l(I)dt] = 0$, but the relative entropy on every grid $F_n$ contains the term $H(\delta_{x_0}|\delta_0) = \infty$, hence $h_l(Q|B^l) = \infty$; this directly contradicts the claimed equality in Theorem 1.2 unless a matching starting law is imposed.

Watch

Extended reading notes

Core claim

The central claim is that the map $Q \mapsto E_Q[\int_0^1 F_l(\Sigma_t^Q)dt]$ is the convex lower semicontinuous envelope of $Q \mapsto h_l(Q|B^l)$. In plainer terms, among all convex lower semicontinuous functionals of a martingale measure that never exceed the specific relative entropy to Wiener measure, the quadratic-variation functional is the largest one, so the multivariate Gantert inequality cannot be improved within that class. The proof proceeds by introducing an approximating family of martingale measures whose diffusion coefficients are simple functions of the past, for which the inequality is an equality, and then showing that every martingale measure with finite cost can be approximated by such simple measures; a universal pathwise quadratic covariation density makes the functional affine and weakly lower semicontinuous. The same machinery yields equality for Gaussian martingales and for multidimensional Black-Scholes models, for which closed-form specific entropies are derived.

Load-bearing premise

The load-bearing premise is that $Q$ and the reference Brownian motion begin at the same initial distribution; the formal statements do not consistently impose this, and without it $h_l(Q|B^l)$ can be infinite while $E_Q[\int_0^1 F_l(\Sigma_t^Q)dt]$ is finite, breaking the claimed convex-envelope identity.

Editorial extensions

If this is right

  • In any dimension, the specific relative entropy between a martingale measure and Wiener measure admits a computable lower bound $E_Q[\int_0^1 F_l(\Sigma_t^Q)dt]$, so entropy comparisons can be replaced by an integral of a convex matrix function.
  • The lower bound is the convex lower semicontinuous envelope of $h_l$, so convex variational problems whose cost is a function of the martingale law can equivalently use the quadratic-variation functional as their canonical relaxation.
  • Gantert's inequality is an equality for Gaussian martingales and for multidimensional Black-Scholes models, yielding the closed-form expressions (5) and (6) for specific relative entropy.
  • Tensorization holds: when the reference martingale has independent coordinates, $h_l$ is at least the sum of the coordinate specific entropies; when coordinates are independent and identically distributed, it is exactly $l$ times the one-dimensional value.
  • The dyadic-grid comparison works for any $p^n$ refinement with $p \geq 2$, so the scaling limit does not depend on the choice of base grid.

Reading between the lines

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

  • The paper leaves implicit that the affine, convex functional $E_Q[\int_0^1 F_l(\Sigma_t^Q)dt]$ can be optimized directly in martingale optimal transport problems, with $h_l$ recovered afterwards by convex relaxation.
  • A testable extension suggested by the proof machinery is that the same approximating-class argument should identify convex envelopes for other rescaled divergences between martingale laws, with envelope functionals built from the same matrix function $F_l$.
  • The closed-form Black-Scholes formulas indicate the natural formula for time-dependent coefficients $\Gamma_t$: replace $\Gamma\Gamma^T$ and the integrals by their time-dependent versions, an extension the paper notes as achievable without writing it down.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper defines a multidimensional version h_l(Q|P) of Gantert's specific relative entropy between continuous martingale laws on Wiener space, by taking the scaled liminf of finite-dimensional relative entropies. Its main results are: Theorem 2.4, the multidimensional Gantert inequality h_l(Q|B^l) ≥ E_Q∫_0^1 F_l(Σ^Q_t)dt for square-integrable martingales with X_0=0 and absolutely continuous quadratic covariation; Theorem 1.2, which claims that the functional L_l(Q)=E_Q∫_0^1 F_l(Σ_t)dt is the convex lower semicontinuous envelope of h_l(·|B^l) over the class M_2^l; and closed-form expressions for the specific relative entropy of multidimensional Black-Scholes models against Brownian motion and between two such models. The paper also contains tensorization properties of h_l and a number of auxiliary lemmas preparing the envelope theorem.

Significance. If the envelope theorem is correctly formulated, it is a substantial contribution: it identifies a computable affine functional as the best convex lower semicontinuous lower bound for the specific relative entropy, a result that the authors state is new even in dimension one. The multidimensional extension of Gantert's inequality is natural, and the closed-form Black-Scholes computations are concrete and consistent with the integrated quadratic-variation functional. The proofs are mostly elementary and transparent, and the lower bound is derived from first principles without fitted parameters. The main problem is a domain error in the statement and proof of Theorem 1.2: the initial law is not fixed, so the claimed envelope identity is false on the stated class. This is a local and fixable issue, but it is load-bearing for the paper's central claim.

major comments (3)
  1. [Theorem 1.2 (statement)] Theorem 1.2 is not true on the stated domain M_2^l. Definition 2.2 imposes no condition on X_0, while B^l is the standard Wiener measure with X_0=0. For any Q with Q(X_0≠0)>0, the restrictions Q|F^{2^n} and B^l|F^{2^n} are singular at time zero by Lemma 2.2, so H(Q|B^l)|F^{2^n}=∞ for every n and hence h_l(Q|B^l)=∞; nevertheless L_l(Q)=E_Q∫_0^1 F_l(Σ_t)dt is finite and may even vanish. For instance, if Q is the law of x+B_t for x≠0 and B a standard Brownian motion, then L_l(Q)=0. Taking the bounded continuous test function f(ω)=φ(ω(0)) with φ(0)=0 and φ(x)=1 gives h_l^{**}(Q)≥E_Q[f]-h_l^*(f)≥1, while L_l(Q)=0, contradicting the claimed identity L_l=h_l^{**}. The statement must restrict Q to laws with X_0=0, or must replace B^l by Brownian motion with the same starting law as Q, as the remark after Theorem 1.1 already indicates.
  2. [Lemma 3.3] Lemma 3.3 asserts h_l(Q|B^l)=E∫_0^1 F_l(σ_t^2)dt for every Q∈M_2^l(N), but the construction of M_2^l(N) permits an arbitrary initial value M_0. The proof's first displayed identity, H(Q|B^l)|F^{2^n}=E[∑_{k=0}^{2^n-1} A_k(...)], omits the initial-time term H(L(M_0)|δ_0) that Lemma 2.2 places in the finite-dimensional relative entropy; this term is +∞ whenever L(M_0)≠δ_0. Consequently equality (15) is false in general and holds only when M_0=0, or when the reference Brownian motion is started at the same law. Since this equality is precisely the bridge from h_l to L_l in the proof of Theorem 1.2, the proof of the envelope theorem collapses for measures with nonzero or random initial value.
  3. [Lemma 3.4 / Proof of Theorem 1.2] The approximation argument does not rescue the missing initial-law assumption. In Step 1 of Lemma 3.4 the process is defined as Z_t=X_t+εB_t, so the initial value is unchanged; the approximating sequence Q_n∈M_2^l(n) therefore inherits the arbitrary starting law of Q. The proof of Theorem 1.2 then applies Lemma 3.3 to these Q_n, which is not legitimate unless X_0=0 has been imposed. The fix is to add X_0=0 to the domain in Definition 2.2 and throughout Section 3, or to work throughout with a reference Brownian motion having the same starting law; this is a local domain correction rather than a reworking of the technical argument.
minor comments (5)
  1. [Abstract] The abstract spells the name as "Ganter's inequality"; it should be "Gantert's inequality".
  2. [Title/header] The title on page 1 appears as "MUL TIDIMENSIONAL SPECIFIC RELA TIVE ENTROPY" with spurious spaces; this should be corrected.
  3. [Lemma 3.3, proof] In the displayed conditional covariance, the integral is written as ∫_{k2^{-n}}^{(k-1)2^{-n}} σ_u^2 du; the upper limit should be (k+1)2^{-n}, matching the conditioning sigma-field and the subsequent Jensen bound.
  4. [Lemma 4.3, proof] In the final displayed derivation of h_l(L(M^Γ)|B^l), the term "e−1/2" appears to be a typesetting corruption of "-l/2"; as printed the line does not match formula (17).
  5. [MSC classification] The Mathematics Subject Classification is left as "Primary XXX; Secondary XXX" and should be filled in.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the lower bound, the equality on the auxiliary class, and the convex-envelope argument are self-contained; author-overlap citations are not load-bearing.

full rationale

No circular step is present. Theorem 1.1 (Gantert's inequality in R^l) is proved from first principles by discretizing the quadratic covariation and comparing Gaussian conditional laws; no fitted parameter or assumed conclusion enters. Lemma 3.3 establishes equality h_l(Q|B^l)=E[∫F_l(σ^2)dt] on the auxiliary class M2_l(N); it uses Theorem 2.4 only for the lower half of the equality, while the upper half is a separate convexity/Jensen argument from finite-dimensional relative entropies. Theorem 1.2 then applies the Fenchel–Moreau mechanism: Lemma 3.4 approximates general Q in M2_l by Q_n in M2_l(n) with L_l(Q_n)→L_l(Q), so the Legendre transform of h_l dominates that of L_l, and the reverse inequality is exactly the already-proved minorant property L_l≤h_l. No equation is equal to its conclusion by construction, and no parameter is fitted and then renamed as a prediction. Author-overlap references [4,6,7,8,21] appear mainly as context and applications; [5] is used once in Lemma 3.1 through a published theorem on lower semicontinuity, not as the premise of the envelope identity, so there is no self-citation chain forcing the result. One correctness caveat, orthogonal to circularity, is that Theorem 1.2 and Lemma 3.3 as written do not restrict Q to a common starting law with B^l (X0=0), so the stated identities can fail for Q with nonzero or random initial law; this is a domain error in the statement, not a circular reduction in the derivation.

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

There are no fitted constants in the paper, and the only new object is the multidimensional specific relative entropy, which is a direct extension of Gantert's existing definition rather than an invented entity. The main unstated burden is the fixed-starting-law convention, which is necessary for the convex envelope theorem and for Lemma 3.3 to hold as stated.

assumptions (5)
  • standard math The relative entropy chain rule (Lemma A.2) and the Gaussian entropy formula (Eq. 7) are correct.
    Used throughout Section 4 for the Black-Scholes closed-form computations and in Lemma 2.2 for discretized relative entropies.
  • standard math The function F_l(X)=1/2(tr(X)-l-log det(X)) is convex, nonnegative, and vanishes only at X=I.
    Crucial for Jensen-type steps in Proposition 2.3 and Theorem 2.4, and for the convexity of L_l in Lemma 3.1.
  • standard math Karandikar's pathwise quadratic covariation construction provides a canonical density matrix that does not depend on the reference measure.
    Invoked in Lemma 3.2 and Remark 8 to identify the affine functional L_l and to define the universal density matrix.
  • domain assumption The measures Q considered are laws of square-integrable continuous martingales with absolutely continuous quadratic covariation.
    This is the class M_l^{2,abs} on which Gantert's inequality is stated; the central theorem is only claimed for this domain.
  • domain assumption The reference Brownian measure and the martingale measure Q share the same starting law whenever h_l is finite.
    This is needed for Lemma 3.3 and Theorem 1.2, but it is not stated in the definitions. As written, h_l with respect to standard Brownian is infinite for Q with a different initial law, so this assumption is load-bearing and currently implicit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multidimensional specific relative entropy between continuous martingales." pith.science (2026). https://pith.science/paper/V4TSTATS

@misc{pith2026241111408,
  author       = {Pith},
  title        = {Pith review of: Multidimensional specific relative entropy between continuous martingales},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V4TSTATS}},
  note         = {Machine review of arXiv:2411.11408}
}
read the original abstract

In continuous time, the laws of martingales tend to be singular to each other. Notably, N. Gantert introduced the concept of specific relative entropy between real-valued continuous martingales, defined as a scaling limit of finite-dimensional relative entropies, and showed that this quantity is non-trivial despite the aforementioned mutual singularity of martingale laws. Our main mathematical contribution is to extend this object, originally restricted to one-dimensional martingales, to multiple dimensions. Among other results, we establish that Gantert's inequality, bounding the specific relative entropy with respect to Wiener measure from below by an explicit functional of the quadratic variation, essentially carries over to higher dimensions. We also prove that this lower bound is tight, in the sense that it is the convex lower semicontinuous envelope of the specific relative entropy. This is a novel result even in dimension one. Finally we establish closed-form expressions for the specific relative entropy in simple multidimensional examples.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

  1. [1]

    D. Aldous. notes, https://www.stat.berkeley.edu/~aldous/Research/OP/ent-MG.pdf

    work page

  2. [2]

    D. Aldous. What is the max-entropy win-probability mart ingale? https://www.stat.berkeley.edu/~aldous/Research/OP/maxentmg.html

    work page

  3. [3]

    Avellaneda, C

    M. Avellaneda, C. Friedman, R. Holmes, and D. Samperi. Ca librating volatility surfaces via relative-entropy minimization. Applied Mathematical Finance , 4(1):37–64, 1997

    work page 1997

  4. [4]

    Backhoff-Veraguas and M

    J. Backhoff-Veraguas and M. Beiglb¨ ock. The most exciting game. Electronic Communications in Probability, 29:1–12, 2024

    work page 2024

  5. [5]

    Backhoff-Veraguas and G

    J. Backhoff-Veraguas and G. Pammer. Applications of weak transport theory. Bernoulli, 28(1):370–394, 2022

    2022

  6. [6]

    Backhoff-Veraguas and C

    J. Backhoff-Veraguas and C. Unterberger. On the specific r elative entropy between martingale diffusions on the line. Electronic Communications in Probability , 28(none):1 – 12, 2023

    work page 2023

  7. [7]

    Backhoff-Veraguas, Z

    J. Backhoff-Veraguas, Z. W ang, and X. Zhang. The Aldous ma rtingale in arbitrary dimen- sions. Work-in-progress, 2024

    work page 2024

  8. [8]

    Backhoff-Veraguas and X

    J. Backhoff-Veraguas and X. Zhang. Specific W asserstein d ivergence between continuous mar- tingales. arXiv preprint arXiv:2404.19672 , 2024

    arXiv 2024

Show all 22 references
  1. [9]

    Benamou, G

    J.-D. Benamou, G. Chazareix, and G. Loeper. From entropi c transport to martingale trans- port, and applications to model calibration. arXiv preprint arXiv:2406.11537 , 2024

    2024 arXiv

  2. [10]

    Cattiaux and N

    P. Cattiaux and N. Gozlan. Deviations bounds and condit ional principles for thin sets. Sto- chastic processes and their applications , 117(2):221–250, 2007. 24 J. BACKHOFF AND E.K. BELLOTTO

    2007

  3. [11]

    F. Chen, G. Conforti, Z. Ren, and X. W ang. Convergence of Sinkhorn’s algorithm for entropic martingale optimal transport problem. arXiv preprint arXiv:2407.14186 , 2024

    2024

  4. [12]

    Cohen and Y

    A. Cohen and Y. Dolinsky. A scaling limit for utility ind ifference prices in the discretised Bachelier model. Finance Stoch., 26(2):335–358, 2022

    2022

  5. [13]

    De March

    H. De March. Entropic approximation for multi-dimensi onal martingale optimal transport. arXiv preprint arXiv:1812.11104 , 2018

    2018 arXiv

  6. [14]

    De March and P

    H. De March and P. Henry-Labordere. Building arbitrage -free implied volatility: Sinkhorn’s algorithm and variants. arXiv preprint arXiv:1902.04456 , 2019

    1902 arXiv

  7. [15]

    F¨ ollmer

    H. F¨ ollmer. Doob decomposition, Dirichlet processes , and entropies on Wiener space. In Dirichlet forms and related topics , volume 394 of Springer Proc. Math. Stat. , pages 119–141. Springer, Singapore, [2022] ©2022

    2022

  8. [16]

    F¨ ollmer

    H. F¨ ollmer. Optimal couplings on Wiener space and an ex tension of Talagrand’s trans- port inequality. In Stochastic analysis, filtering, and stochastic optimizati on, pages 147–175. Springer, Cham, [2022] ©2022

    2022

  9. [17]

    N. Gantert. Einige grosse Abweichungen der Brownschen Bewegung , volume 224 of Bonner Mathematische Schriften [Bonn Mathematical Publications ]. Universit¨ at Bonn, Mathema- tisches Institut, Bonn, 1991. Dissertation, Rheinische Fr iedrich-Wilhelms-Universit¨ at Bonn, Bonn, 1991

    1991

  10. [18]

    G. Guo, D. Possama ¨ ı, and C. Reisinger. Randomness and e arly termination: what makes a game exciting? ArXiv e-prints , 2306.07133, 2023

    2023 arXiv

  11. [19]

    Henry-Labordere

    P. Henry-Labordere. From (martingale) Schr¨ odinger b ridges to a new class of stochastic volatility model. Available at SSRN 3353270 , 2019

    2019

  12. [20]

    R. L. Karandikar. On the quadratic variation process of a continuous martingale. Illinois journal of Mathematics , 27(2):178–181, 1983

    1983

  13. [21]

    Kimani Bellotto

    E. Kimani Bellotto. Specific relative entropy between c ontinuous martingales. University of Vienna (Masterarbeit), 2024

    2024

  14. [22]

    Nutz and J

    M. Nutz and J. Wiesel. On the martingale Schr¨ odinger br idge between two distributions. arXiv preprint arXiv:2401.05209 , 2024. Appendix A. Elementary Properties of the Entropy We collect here a few elementary properties of the relative entropy . We start by recalling the we...

    2024 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.