acceptodds
Under review as a conference paper at ICLR 2027

Drift-admissibility via Optimal Transport

Abstract

We introduce structure-aware optimal transport diagnostics for assessing whether empirical time-series dynamics are compatible with a prescribed predictable-drift class. Our approach translates a pathwise property of stochastic processes into a finite-dimensional projection problem over empirical transition laws, encoding admissibility as row-wise moment constraints on local couplings and quantifying deviations via a Kullback–Leibler projection gap . This yields both a principled diagnostic and a scalable Bregman–Dykstra algorithm that retains the efficiency of Sinkhorn-type methods while incorporating structural constraints. Calibrated against a Brownian reference null, separates process classes that are indistinguishable by second-order statistics alone—including fractional Brownian motion and BSS processes sharing identical Hurst indices—without asymptotic resampling or continuous-time recovery. This positions the method as a general tool for structure-aware model validation of sequential data.

Primary area:
learning on time series and dynamical systems
Keywords:
optimal transport, conditional moments, time series dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cWuROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an optimal transport diagnostic for assessing whether discretely observed time series are compatible with a prescribed bounded predictable drift class. Within rolling windows, it compares reference and moment constrained solutions of a marginally relaxed entropic transport problem through a KL divergence, with drift bounds calibrated using Brownian motion. The authors develop and analyze a Bregman–Dykstra solver combining Sinkhorn scaling with row wise exponential tilting. Experiments examine the gap and its temporal variability on simulated processes with different roughness, memory, and jump structures, assess sensitivity to the drift envelope and sampling resolution, and benchmark numerical accuracy and runtime against general purpose convex solvers.

Critical strengths

The constrained transport formulation has a tractable computational structure. The paper characterizes row wise KL projections under moment constraints through one-dimensional exponential tilting and incorporates them into a Bregman–Dykstra solver with Sinkhorn scaling. This provides an efficient approach to solving the constrained transport problem, supported by numerical comparisons with a general purpose optimizer.

Critical weaknesses
  1. The projection gap is not clearly linked to the stated conditional drift property. Definition 1 constrains the conditional mean of the increment, but the implemented moment operator uses g(z,y)=y-z, although y is already defined as the increment (Sec. 3, Algorithm 1). Also, the transport problems use separate empirical predictor and increment marginals rather than their observed pairings. Rearranging these pairings leaves the gap unchanged despite potentially changing the conditional transition law. These issues level down the interpretation of the gap as a diagnostic of predictable drift admissibility.
  2. The main empirical evidence does not establish sensitivity beyond second order structure. The fBM–BSS comparison matches Hurst indices but not full covariance structure (Table 1, Fig. 1). Both simulated process families are centered Gaussian, so their distributional differences are determined by second order properties. The observed gap separation therefore does not establish sensitivity to conditional memory beyond second order statistics, as claimed in the abstract and conclusion.
Questions for authors
  1. Definition 1 bounds the conditional mean of the increment, but Sec. 3 and Algorithm 1 use g(z,y)=y-z, where y is already the increment. What conditional quantity is actually constrained? If the zero mean jump processes in Table 1 are martingales, how should their gap spikes in Fig. 4 be interpreted as drift admissibility violations?
  2. Do the observed (Z_s,Y_s) pairings enter the computation of Gλ,kG_{\lambda,k}, and which predictor is used in Figs. 1–4? If the gap depends only on the separate empirical marginals, how does it identify the observed conditional transition structure?
  3. Since fBM and BSS are centered Gaussian processes, matching their Hurst indices does not control their full covariance structure. What evidence supports the claimed sensitivity beyond second order statistics, rather than differences in covariance?
  4. Lemma A1 characterizes a KL projection with fixed row masses, whereas Eq. (4) relaxes the marginal constraints. How does this result apply to Algorithm 1, and what guarantees convergence to the constrained minimizer of Eq. (5)?
  5. Proposition 1 derives monotonicity of the projection gap from monotonicity of the constrained minimum of Jε,λJ_{\varepsilon,\lambda}. Since the gap is defined separately as KL(Tkproj∥Tkref)\mathrm{KL}(T_k^{\mathrm{proj}}\Vert T_k^{\mathrm{ref}}), what justifies this implication in the presence of marginal KL penalties?
2 comments
AuthorsOpenReview ↗

Reply to reviewer cWuR

We are grateful for a careful and expert report; its questions shaped the revision.

What is constrained, and the jump processes in Fig. 4 (Q1, W1). The constrained quantity is the conditional mean of the increment given the predictor, as in Definition 1. ν^\hat\nu is the empirical measure of next states, so g(z,y)=y−zg(z, y) = y - z is a displacement, equal to the realised increment only on observed pairs; Sections 3–4 now say so. The reviewer is right that zero-mean compound Poisson processes are martingales. Their spikes in Fig. 4 (Fig. 3 of the revision) arise because a single jump widens the spread of levels in every window that contains it, relative to a fixed envelope; they are not drift violations, and the caption now says so. In checking this we also found that Table 1 misdescribed the "CP diffusion": it was simulated as a mean-reverting Ornstein–Uhlenbeck process with the same jumps (κ=5\kappa = 5), not as a martingale. Table 1 and the captions are corrected; its spikes have the same origin. Under the wild-bootstrap calibration the gap rejects the jump martingales at 0.04–0.07 (Appendix E).

Do the observed pairings enter, and which predictor (Q2)? In the submitted construction they did not; the general response and Appendix F describe what that gap measures instead. The coupling of observed transitions (Section 4.3) ties each increment to the state that preceded it, so permuting pairs changes the gap. The figures of Section 5 use the marginals of Xs−1X_{s-1} and XsX_s; Appendix E reports the level, lag and level-plus-lag predictors, all with size close to nominal.

Beyond second order (Q3, W2). We agree, and the point goes further than we had appreciated: the separation in Fig. 1 is explained largely by the wider spread of BSS levels over a window, a second-order property (Appendix F). The claim has been removed from the abstract, introduction and conclusion.

Lemma A1 and convergence (Q4). The experiments of Section 5 compute the row-wise gap, which involves no iteration: TrefT^{\mathrm{ref}} is computed once by relaxed Sinkhorn, and each active row is projected at its own fixed mass, exactly the setting of Lemma A1. The test of Section 4.3 uses the joint minimiser of the relaxed problem. Its dual is smooth and concave in the blocks (u,v,γ)(u, v, \gamma); the unbalanced Sinkhorn updates maximise it exactly in uu and vv, and each tilt maximises it exactly in γk\gamma_k subject to that multiplier's sign constraint. Written on the conditional mean, ∑j(gkj−dk) Tkj≤0\sum_j (g_{kj} - d_k)\, T_{kj} \le 0, the row constraint is a homogeneous halfspace, so the tilt is unchanged when the row is rescaled and Lemma A1 applies at whatever mass the current iterate carries. The scheme is block-coordinate ascent, equivalently Bregman–Dykstra (Bauschke & Lewis, 2000; Tseng, 2001), and converges to the unique minimiser. In revising this we corrected the activation test in Algorithm 1, which should be evaluated at γk=0\gamma_k = 0. We have also reworded Table 2: it compares the row-wise projection with MOSEK's joint optimum (objective within 2%, drifts within 0.003), so it shows how close the two are rather than validating a joint solver.

Proposition 1 with marginal penalties (Q5). The reviewer is right: the submitted argument established monotonicity of the constrained objective, not of the KL gap. With marginal penalties the statement is now made for ΔJ\Delta J and for the row-wise gap; these coincide with the KL gap, up to ε\varepsilon, only with hard marginals. For the row-wise gap we added the qualification that active rows remain feasible: our implementation removes an infeasible row, which then contributes only its mass, so the gap can fall as the box tightens. Assigning such rows their limiting divergence makes the statement unconditional.

I appreciate the authors' response and substantial revisions.

My main concern remains whether the proposed diagnostic meaningfully reflects the conditional drift property it aims to assess. The authors now acknowledge that the original marginal based gap measures level dispersion rather than conditional drift, and introduce a new observed transition coupling with wild bootstrap calibration. It remains unclear to me how far this new construction supports the original claim of assessing a general prescribed predictable drift class.

Also, the original submission had significant gaps between its theoretical claims and supporting evidence, as well as inconsistencies in the formulation and interpretation. These raise concerns about the overall rigor of the work, especially for a theory paper.

I acknowledge the improvements, but these concerns remain. I will keep my assessment unchanged.

Reviewer_T7t6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an OT-based framework for measuring whether empirical time-series dynamics are compatible with a prescribed predictable-drift class. The method constructs an empirical transition coupling, projects it onto a set satisfying bounded conditional-drift constraints, and uses the KL distance between the original and projected couplings as a measure of structural incompatibility.

Critical strengths
  1. The paper studies an interesting and nontrivial problem: diagnosing structural compatibility of discretely observed stochastic processes.
  2. The formulation of predictable-drift constraints through constrained OT is conceptually novel and interesting.
  3. The theoretical development is relatively rigorous, including uniqueness, convergence, and monotonicity results.
Critical weaknesses
  1. The presentation lacks sufficient intuition for a general ML audience. The relationship between empirical drift estimation, OT coupling, admissible projection, and the resulting diagnostic could be explained much more clearly.
  2. The practical scope of the framework is somewhat unclear. The current formulation focuses on one-dimensional predictable first-moment constraints with a user-specified admissible drift class, and all experiments are synthetic. It would strengthen the paper to better discuss which real-world modeling problems naturally provide such constraints and how the approach scales to richer predictors or multivariate dynamics.
  3. see questions.
Questions for authors
  1. Since the method already constructs a nearest admissible coupling, it would be interesting to evaluate whether this projected coupling is itself a useful corrected estimator rather than using it only to define a diagnostic?
  2. If Yi=ΔiXY_i=\Delta_iX, why is the basic increment operator later written as (g(z,y)=y-z)? How should this be interpreted for a general predictor (z)? Some clean up of notations may be needed.
  3. Proposition 1 defines the projection gap as $ G_{\lambda,k} = \mathrm{KL}\left(T_k^{\mathrm{proj}} | T_k^{\mathrm{ref}}\right), $ while the argument for monotonicity relies on the constrained objective value increasing as the feasible set becomes smaller. What identity or Bregman/Pythagorean relation connects the objective-value difference to the stated KL projection gap?
  4. In terms of scope of the paper, the author should address the following: who actually needs this diagnostic, and in what settings is its structural assumption naturally available?
1 comment
AuthorsOpenReview ↗

Reply to reviewer T7t6

We thank the reviewer for a generous and constructive report, and for finding the formulation novel and the theory careful.

Intuition (W1). Section 3 now opens with a three-step overview: the reference coupling estimates the local drift at each predictor state by borrowing nearby transitions, much as a kernel smoother does; the projected coupling is the smallest change to it, in KL divergence, that brings every local drift within the box; the gap records how large that change had to be. A new figure (Appendix F) shows both couplings for one mean-reverting window, and Section 4 opens with a short roadmap. The rate-function identity in the general response gives the gap a plain reading: approximately, a mass-weighted sum of squared standardised excesses of the local means over the box.

Scope and users (W2, Q4). The diagnostic is natural wherever a drift tolerance is fixed by design rather than estimated. Section 6 discusses two examples. In reinforcement learning, a value function solves the Bellman equation for a policy exactly when its temporal-difference errors have zero conditional mean given the state, so a tolerance on the Bellman residual is a box on that conditional mean, with the state as a multivariate predictor. In derivative pricing, simulated transition kernels are routinely corrected to be martingales by minimum-relative-entropy reweighting (Avellaneda et al., 2001).

We are pursuing the first in separate work, which is not part of this submission, and mention it only briefly. In two small control problems where the Bellman error is known exactly, with an error planted in the critic, the per-state gaps on the coupling of observed transitions locate the states outside tolerance better than the best of several smoothers of the same temporal-difference errors once there are about 1,000 transitions or more (AUROC higher by 0.05–0.08; at 2,000 transitions, on all 30 datasets in each problem), and separate them from states that are merely noisy better still (by 0.10–0.12 at 2,000). With 500 transitions or fewer there is little or no advantage, and for a nearly correct critic transport is level with smoothing or worse. These results are preliminary, and the paper makes no claim on them; Section 6 states that all its experiments are synthetic.

On scaling: the coupling of observed transitions accepts vector predictors through its kernel (Appendix E uses level and lag together), and the cost per window does not depend on the predictor's dimension, although the data required grow with it.

The projected coupling as an estimator (Q1). As an estimate of the drift it adds little: an active row is moved to the nearest box edge, so its mean is the reference mean clipped to the box. As an estimate of the conditional law it is more interesting: each projected row is the minimum-KL modification of the reference row that meets the constraint, which is the exponential-tilting estimator of a conditional distribution under a moment restriction (Kitamura & Stutzer, 1997), and a natural way to generate admissible scenarios. Section 6 now notes this use; we have not evaluated it as an estimator.

Notation (Q2). The text was inconsistent. In Sections 3–5, ν^\hat\nu is the empirical measure of next states XsX_s, not of increments, so g(z,y)=y−zg(z, y) = y - z is the displacement from state z to state y; it equals a realised increment only on observed pairs, which is the substance of the other reviewers' concerns. In the coupling of observed transitions the columns carry the increments themselves, and g(z,y)=yg(z, y) = y.

Proposition 1 (Q3). The identity is the Pythagorean identity for I-projections (Csiszár, 1975): with hard marginals it gives ε KL=ΔJ\varepsilon\,\mathrm{KL} = \Delta J exactly; with relaxed marginals the result is stated for ΔJ\Delta J and for the row-wise gap (general response).

Reviewer_GDCdOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a rolling-window diagnostic for compatibility with predictable-drift bounds. It measures the KL divergence between entropic transport plans fitted to empirical predictor and increment marginals, with and without row-wise moment constraints. The approach is potentially useful, but the diagnostic’s statistical interpretation and constraint specification.

Critical strengths
  1. Under hard marginal constraints, the drift bounds admit a tractable formulation as linear moment inequalities.

  2. The simulations provide useful evidence about the diagnostic’s behavior and numerical implementation.

Critical weaknesses
  1. Interpretation of the fitted coupling and choice of predictor.

Uniqueness of the entropic optimizer does not establish that its conditional means represent those of the observed process. The construction uses separate predictor and increment marginals, which do not identify their joint law. Could the authors explain what connects the selected coupling to the observed conditional dynamics, or clarify the scope of its interpretation under the chosen transport geometry?

The choice of ZiZ_i also needs guidance. A small predictor may omit relevant history, while a richer predictor increases data requirements. The paper does not establish when ZiZ_i adequately captures conditional drift given the full past, leaving its applicability under complex dependence unclear.

  1. Monotonicity argument.

Proposition 1 claims that tightening the drift bounds cannot make the KL gap in Equation (6) smaller. The proof establishes this only for the minimum of the full objective, which also includes penalties for changing the marginals. It does not establish the same result for the KL gap alone. Please provide that argument or revise the proposition.

  1. Comparison with existing diagnostics.

The experiments compare process families and validate the optimizer, but do not compare the diagnostic with conditional-mean or conditional-moment approaches. A focused comparison using the same predictors, windows, and drift bounds, with appropriate treatment of temporal dependence, would help clarify the additional value of the transport construction.

Questions for authors
  1. How should ZiZ_i be selected, and what connects the fitted coupling to conditional drift given the full past?

  2. What benefit does the transport construction provide over direct conditional-mean or conditional-moment diagnostics?

2 comments
AuthorsOpenReview ↗

Reply to reviewer CDCd

We thank the reviewer for a clear report and specific suggestions, which we have tried to follow.

What connects the coupling to the observed dynamics, and the choice of Z (W1, Q1). The reviewer's doubt is justified: uniqueness of the entropic optimiser says nothing about whether its conditional means are those of the process, and in the submitted construction they were not (general response, Appendix F). In the coupling of observed transitions (Section 4.3), each row averages the increments that actually followed nearby predictor states, so its mean is a kernel estimate of E[ΔX∣Z]\mathbb{E}[\Delta X \mid Z], and the wild bootstrap compares the gap with its distribution under zero conditional drift for the same series. The diagnostic tests admissibility with respect to the filtration generated by the chosen predictor, not the full past, and Section 6 now says so. We recommend the level, or a short vector of lags, when the drift class is Markov in the level, and a richer predictor only when the window is long enough to support it; Appendix E reports level, lag and both.

Proposition 1 (W2). The reviewer is right; please see the general response for the corrected statement.

Comparison with conditional-mean and conditional-moment diagnostics (W3, Q2). We added the comparison (Appendix E), using the same paths, windows and box, with every statistic calibrated by the same wild bootstrap, which preserves each series' volatility path and so respects its temporal dependence under the null. No statistic dominates. Variance-ratio tests (Lo & MacKinlay, 1988) are strongest against roughness, a regression slope against periodic drift, and the mean-increment t-statistic against constant drift. Compared with a Nadaraya–Watson estimate of the conditional mean on the level predictor, the gap rejects more often against rough fBM with H=0.4H = 0.4 (0.36 against 0.23) and against the mean-reverting jump diffusion (0.33 against 0.19), and less often against constant drift (0.36 against 0.87).

As to what transport adds: to second order the row-wise gap is a self-normalised kernel conditional-moment statistic (general response). Beyond that, it contributes inequality constraints on rolling windows, self-normalisation by the local spread, an exact zero for rows within the box, the projected conditional laws, and a solver at Sinkhorn cost. As a test it is not more powerful than direct conditional-moment statistics; where we have seen it do better is in locating where a drift tolerance fails, given enough data (our reply to Reviewer T7t6). We have revised the text so that it claims no more than this.

Thanks for the revision and responding to my concerns. I have one follow-up on calibration under leverage, where Section 4.3 describes sign flipping as approximate. For the Heston model with (\xi=1) and (\rho=-0.7), Table 3 reports rejection rates of 4–6% at a nominal 5% level using a critical value pooled across replications. Have you checked whether the rejection rates are similar when you compute a separate bootstrap p-value for each simulated path using repeated sign flips, as described in Section 4.3?

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of PDF revision

We have uploaded a revised PDF. The principal changes are as follows:

  1. A coupling of the observed transitions, calibrated by a wild bootstrap (new Section 4.3, Appendix E). The observed pairing of predictor and increment enters the transport cost; the statistic is calibrated by sign flips that rebuild the path. Appendix E reports size under eight martingales and power under thirteen alternatives, for three predictors, alongside conditional-moment baselines calibrated the same way.
  2. What the original construction measures (Section 5, new Appendix F). Without the observed pairing, the mean gap is governed largely by how widely the levels spread within a window. Section 5 is now described in those terms, and the claims about memory are withdrawn.
  3. The gap and Proposition 1 (Section 4.1). The gap is now defined explicitly: row by row for the experiments of Section 5, and through the joint projection for the test of Section 4.3. A new remark identifies the row-wise gap as a sum of Cramér rate functions and corrects our earlier Sanov remark. Proposition 1 is restated in three parts, with a feasibility qualification for the row-wise gap.
  4. Notation (Sections 3–5). ν^\hat\nu is the empirical measure of next states; y−zy - z is a displacement, equal to the realised increment only on observed pairs.
  5. Algorithm 1 and convergence (Section 3.2, Appendices B and C.2). The row-wise gap and the joint scheme are given separately; the joint scheme is stated as block-coordinate dual ascent, with its activation test corrected.
  6. Calibration, solver comparison and Table 1 (Sections 4.2, 5). The Brownian envelope is described as a scale, not a calibrated null. Table 2 is described as comparing the row-wise projection with the joint optimum, not as validating a joint solver. Table 1 now describes the "CP diffusion" correctly: it is mean-reverting.
  7. Readability and scope. Figure 3 redrawn; a roadmap for Section 4; a figure of both couplings for one window (Appendix F); a notation table (Appendix J); Section 6 covers the predictor filtration, the synthetic experiments, and settings where a drift tolerance arises by design.
AuthorsOpenReview ↗

General response to all reviewers

We are grateful to all three reviewers. Read together, their reports ask two questions that we think are the right ones: what connects the fitted coupling to the observed conditional dynamics, and how large must a gap be before it counts as evidence? The revision is built around the answers.

What the submitted construction measures. Reviewers GDCd and cWuR are right. The submitted coupling joined the empirical predictor and next-state marginals, so permuting the observed pairs leaves the gap unchanged, and the row means are not estimates of E[ΔX∣Z]\mathbb{E}[\Delta X \mid Z]. We checked what that gap does respond to (Appendix F). A curve of mean gap against the standard deviation of the levels in each window, fitted on zero-drift Brownian motions alone, predicts the mean gap of fBM, BSS, Ornstein–Uhlenbeck and drifting Brownian paths: the correlation is 0.96 on the log scale, over 20 paths per process. A Brownian motion with σ=4\sigma = 4 scores higher than every rough process (0.144 against at most 0.122); a strongly mean-reverting OU path scores below Brownian motion (0.006 against 0.013); rescaled to Brownian increment variance, rough fBM falls to 0.002. The fBM–BSS separation in Fig. 1 therefore reflects the wider spread of BSS levels over a window, a second-order property. Section 5 now says so; its results describe departures from a fixed envelope, not drift.

The remedy: a coupling of the observed transitions with a calibrated test (Section 4.3). Rows index predictor states and columns index the observed transitions, each carrying its own increment; the kernel compares predictor states, so each row borrows the increments that actually followed nearby states. Its mean is a Nadaraya–Watson estimate of E[ΔX∣Z=Zs]\mathbb{E}[\Delta X \mid Z = Z_s], and permuting the pairs changes the gap. With a zero box the statistic is invariant to rescaling the series, so the level spread that drives the submitted construction cannot drive it. We calibrate it by a wild bootstrap: Rademacher signs, path rebuilt from the re-signed increments, predictors and statistic recomputed. Sign flips preserve absolute increments, so the bootstrap inherits each series' volatility clustering and jumps. The flip is exact when each increment is conditionally symmetric and depends on the past only through absolute increments (as in GARCH with symmetric innovations); under leverage or level-dependent volatility it is an approximation, and the Heston rows of Appendix E show how well it holds.

With a zero box, N=400N = 400 and 200 replications, rejection rates at the nominal 5% level lie between 0.02 and 0.10 across eight martingales and three predictors; the largest, 0.10 (Brownian motion, lag predictor), is about three Monte Carlo standard errors above nominal. On power we do not overstate the case: rough fBM (H=0.3H = 0.3) is rejected in 0.90–0.99 of series, but constant drift (μ=5\mu = 5) only in 0.36–0.61, where the mean-increment t-statistic reaches 0.98. It is an omnibus statistic.

What the gap is (new remark, Section 4.1). Projecting a row of mass mkm_k at fixed mass onto its nearest box edge dkd_k costs mkΛk∗(dk)m_k \Lambda^*_k(d_k), where Λk∗\Lambda^*_k is the Cramér rate function of that row's displacement distribution. The row-wise gap is therefore a sum of local large-deviation rates and, to second order, half the mass-weighted sum of squared standardised excesses of the local means over the box. This places it within exponential tilting for conditional moment restrictions (Kitamura & Stutzer, 1997; Kitamura, Tripathi & Ahn, 2004). The transport formulation adds inequality constraints on rolling windows of one path, self-normalisation by the local spread, an exact zero for rows inside the box, a solver at Sinkhorn cost, and the projected conditional laws. The identity also corrects our Sanov remark, which had the direction reversed.

Proposition 1. (i) The objective gap ΔJ\Delta J is monotone under nested boxes for every λ\lambda. (ii) With hard marginals, the Pythagorean identity for I-projections gives ε KL(Tjoint ∥ Tref)=ΔJ\varepsilon\,\mathrm{KL}(T^{\mathrm{joint}} \,\|\, T^{\mathrm{ref}}) = \Delta J, where TjointT^{\mathrm{joint}} is the joint minimiser (TprojT^{\mathrm{proj}} in the submission). (iii) The row-wise gap is monotone for every λ\lambda provided active rows remain feasible: a tighter box needs a larger tilt, and a row's divergence increases with the tilt.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.