Betting on Moments: Legendre Jumper Martingales for Online Exchangeability Testing

Johan Hallberg Szabadv\'ary

REVIEW 2 major objections 4 minor 9 references

Betting on Moments: Legendre Jumper Martingales for Online Exchangeability Testing

T0 review · 2 major / 4 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Shifted Legendre polynomials turn the Simple Jumper into a scalable martingale that bets on variance, skewness and higher moments of conformal p-values.

desk verdict Clean, usable multi-moment extension of the Simple Jumper with a practical linear-cost variational form; math holds, empirics are narrow but honest. read the letter →

arxiv 2606.20859 v2 pith:2VLRNQZZ submitted 2026-06-18 stat.ML cs.LGmath.STstat.TH

Johan Hallberg Szabadváry This is my paper · ORCID

classification stat.MLcs.LGmath.STstat.TH MSC 62L1062G1060G42

keywords conformaltestmartingalesexchangeabilitytestingshiftedLegendrepolynomialsSimpleJumperdistributionshiftmean-fieldapproximationonlinemonitoring

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

Exchangeability—the idea that the future looks like the past—is routinely violated by distribution shift, yet many online tests only notice when the mean of conformal p-values moves. This paper shows that the familiar Simple Jumper can be rewritten with shifted Legendre polynomials of any degree, so a single betting function can target variance, skewness or kurtosis in isolation. Combining several degrees into a product betting function captures multi-moment shifts at once, but the joint state space grows exponentially and dilutes capital—the “jumping tax.” A mean-field (variational) factorisation runs one cheap sub-jumper per degree, forms a consensus parameter vector, and places a single product bet; the resulting process remains a valid conformal test martingale, costs linear time, and empirically matches the exact product’s capital growth. Averaging over a small grid of jumping rates further supplies a wealth floor and automatic adaptation to unknown shift timescales. On a real wine-quality classification stream the multi-degree versions accumulate substantially more evidence against exchangeability than any single-degree martingale, while staying quiet under pure permutation.

What carries the argument

The Variational Legendre Jumper: independent Simple Legendre sub-jumpers (one per degree) produce wealth-weighted consensus parameters; a single normalised product of the corresponding Legendre betting functions is then used as the global bet, guaranteeing a valid test martingale whose cost scales linearly in the number of degrees.

What would settle it

Run both the Product and Variational Legendre Jumpers with the same multi-degree set K on a controlled multi-moment shift (e.g., a known Beta or mixture sequence) and check whether the final log-wealth of the variational version falls more than a few percent below the exact product version; a large, systematic gap would falsify the “minimal power loss” claim.

Watch

Extended reading notes

Core claim

A family of conformal test martingales built from shifted Legendre polynomials—Simple, Product, Variational and Composite—extends the Simple Jumper from mean shifts to simultaneous higher-order moment deviations while preserving the martingale property and, in the variational case, reducing per-step cost from exponential to linear with negligible empirical power loss.

Load-bearing premise

That the mean-field consensus parameters keep essentially the same detection power as a fully joint product chain for the degrees people actually use, even though the paper leaves formal error bounds open and shows the claim on one real data set plus synthetic betas.

Editorial extensions

If this is right

Practitioners can monitor real-time conformal streams for variance or skewness collapse without waiting for a mean shift.
The Composite form (average over a grid of jumping rates) becomes the default online exchangeability test when the shift timescale is unknown.
Higher-degree Legendre bets remain valid on any conformal p-value sequence, so the same code can be dropped into existing online conformal pipelines.
The linear-cost variational construction makes it feasible to keep many polynomial degrees active simultaneously rather than committing to a single order in advance.

Reading between the lines

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

The same mean-field factorisation could be applied to other orthogonal polynomial families or to sleeping/waking jumper architectures already used in conformal testing.
A sequential model-selection rule that grows or shrinks the active degree set K on the fly would mirror classical data-driven smooth tests while staying inside the martingale framework.
If formal approximation bounds can be obtained via variational divergence techniques, the method would supply explicit power guarantees for |K|≥3 rather than relying solely on empirical match.

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.

Desk Editor's Note

read the letter

This is a solid methods note that does exactly what it claims. The new pieces are the Product Legendre Jumper (with the explicit Z(ε) normalisation for |K|≥3), the named jumping tax, and especially the Variational Legendre Jumper that factorises the product into independent sub-jumpers plus a consensus product bet. That last construction is the real algorithmic contribution: linear cost in the number of degrees while remaining a valid conformal test martingale by predictability of the consensus parameters.

What the paper does well is keep the foundations tight. Martingale validity follows immediately from the integral constraint and the fact that the betting functions are determined before the next p-value. Non-negativity of the standard grid for every degree is proved cleanly via |P_k|≤1. The cross-term coefficients for small K are derived correctly. The composite over jumping rates is the sensible default and inherits the wealth floor from the J=1 component. Code is already in online-cp, which is more than most notes deliver.

The soft spots are real but already flagged by the author. Formal power-loss bounds for the mean-field step are left open (Section 9), and the empirical case that VLJ tracks PLJ rests on one UCI stream (Wine Quality under four orderings) plus a few synthetic betas. That is enough to show the idea works and that single-degree SLJs of order 2 or 3 alone can miss the dominant location shift, but it is not a broad stress test. Free parameters (grid, J, K) are the usual ones; nothing is hidden. Citation pattern is appropriate: Vovk for the jumper infrastructure, Neyman for the historical Legendre motivation.

This is for people who already run conformal test martingales and want multi-moment power without exponential state spaces. It will not reorganise the field, but it is immediately usable and formally grounded. I would send it to peer review; a referee can push for more datasets and a start on the approximation bounds, but the core is ready.

Referee Report

2 major / 4 minor

Summary. The paper generalises the Simple Jumper conformal test martingale by replacing its linear betting function with shifted Legendre polynomials. It defines the Simple Legendre Jumper (arbitrary single degree k), the Product Legendre Jumper (product over a set K of degrees, with explicit normalisation Z(ε) for higher-order cross terms), and the Variational Legendre Jumper (mean-field factorisation into independent sub-jumpers whose wealth-weighted consensus parameters are plugged into the product bet). A Composite variant averages over a grid of jumping rates. Validity under exchangeability follows from the integral constraint and predictability; non-negativity of the standard grid E for all degrees is established via |P_k|≤1. Empirical trajectories on synthetic beta alternatives and the Wine Quality dataset (four orderings) show that multi-degree PLJ/VLJ accumulate more capital than single-degree SLJ under multi-moment shifts, that VLJ closely tracks PLJ at linear rather than exponential cost, and that the composite form supplies a wealth floor and automatic rate adaptation.

Significance. If the constructions hold, the work supplies a practical, distribution-free tool for online multi-moment exchangeability testing that removes the Simple Jumper’s restriction to location shifts while remaining computationally feasible. The explicit link to Neyman’s smooth test, the closed-form treatment of Z(ε), the proof that the standard grid is safe for every degree, and the open-source implementation in online-cp are concrete strengths. The variational reduction of the “jumping tax” is a useful engineering contribution even without formal power bounds, and the composite recommendation is immediately usable by practitioners who do not know the shift timescale.

major comments (2)

Section 9 and the empirical support for VLJ: the paper correctly flags that formal error bounds on the mean-field approximation remain open. The claim of “minimal power loss” (abstract, §5.1) rests on a single UCI stream (Wine Quality, four orderings) plus a handful of synthetic beta mixtures (Figs. 5–7). For |K|≥3 the approximation quality is therefore uncharacterised beyond these examples. Either additional controlled alternatives that isolate higher-order moments, or a quantitative statement of the observed relative log-capital gap, would be needed before VLJ can be presented as a drop-in default without caveat.
Section 8 / Table 3: only one real-world dataset and a fixed nonconformity measure (1-NN ratio) are used. The hierarchy PLJ/VLJ ≫ SLJ(k=1) ≫ SLJ(k=2,3) may be specific to the red/white wine change-point structure. At least one additional stream (or a synthetic multi-moment change-point with known ground-truth moments) would strengthen the claim that multi-degree betting is systematically superior under realistic distributional shift.

minor comments (4)

Abstract vs. body: the abstract states “constant time per step”; Algorithm 4 and §5.1 correctly give O(|K|·g). Align the wording.
Figure 2 caption still reads f^{(1)}_ε while the plot is for k=2; likewise a few “decreses” / “comaratively” typos remain.
Section 5: the derivation of Z(ε) for |K|≥3 is clear for K={1,2,3,4}, but a short remark on how the lookup table is built for larger K (or a pointer to the package) would help reproducibility.
Notation: the same symbol E is used both for the ε-grid and for expectation; a brief disambiguation would avoid momentary confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; all constructions are valid CTMs by orthogonality of shifted Legendre polynomials plus predictability of consensus parameters, with empirical power comparisons measured on held-out orderings rather than fitted inputs.

full rationale

The derivation chain is self-contained and non-circular. Simple Legendre Jumper betting functions 1+εP̃k(p) integrate to 1 by the defining orthogonality ∫P̃k=0 (k≥1) and are non-negative on the standard grid E because |Pk|≤1 on [-1,1] (Section 7, Szegő). Product forms add the explicit finite normalisation Z(ε) whose closed-form Gaunt coefficients are pre-computed; the Markov mixture is the same as Vovk et al.’s Simple Jumper and therefore inherits the martingale property. Variational Legendre Jumper replaces the joint chain by independent marginal sub-jumpers whose wealth-weighted means ε̄k are fully determined before pn, so the product bet fK_ε̄(pn) remains a predictable valid betting function (Algorithm 4 and preceding paragraph). Composite averaging over rates is convexity. Empirical trajectories (Wine Quality orderings, synthetic betas) measure capital growth after the methods are fully specified; no parameter is fitted to the same quantity later called a “prediction.” Self-citations to Vovk supply only the background jumper infrastructure, not load-bearing uniqueness theorems that force the multi-moment claims. The open problem on mean-field error bounds (Section 9) is a limitation of power analysis, not a circularity in the validity argument.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The paper sits on standard conformal-test-martingale foundations (uniform p-values under exchangeability, Ville) plus classical Legendre orthogonality. Free choices are the discrete ε-grid, jumping rates, and which degrees enter K; the only substantive modelling invention is the mean-field consensus product used as a betting function. No new physical entities are postulated.

free parameters (3)

ε-grid E = {-1/2,-1/4,0,1/4,1/2}
Hand-chosen discrete betting fractions; paper notes dependence is not heavy and proves safety inside [-1,1], but the specific five-point grid is a free design choice carried from Vovk et al.
jumping rate J (and composite set {10^{-4},...,1})
Controls adaptation speed; fixed J=0.01 used in main table; composite averages five rates chosen by hand following Vovk et al.
degree set K
Which Legendre degrees are included (e.g. {1,2,3}) is chosen by the user; paper follows classical Neyman advice of small k but does not derive an optimal K.

assumptions (4)

domain assumption Under exchangeability, conformal p-values are i.i.d. Uniform[0,1] (Vovk et al., Theorem 11.1).
Load-bearing null used throughout Sections 1–2 to justify testing uniformity via CTMs.
standard math Ville’s inequality: P(∃n: S_n ≥ C) ≤ 1/C for non-negative martingales with S_0=1.
Supplies the false-alarm guarantee for all constructions.
standard math Shifted Legendre polynomials are orthogonal on [0,1] with ∫ P̃_k = 0 for k≥1 and |P_k|≤1 on [-1,1].
Used to enforce the integral betting constraint and non-negativity of the standard grid (Sections 3, 7).
ad hoc to paper Mean-field consensus parameters yield a predictable betting function whose product remains a valid CTM.
Validity of the martingale property is proved; the claim of near-full power is empirical and left theoretically open (Section 5.1, 9).

invented entities (2)

Variational Legendre Jumper (mean-field consensus product bet)
purpose: Replace the joint Markov chain over ε^K with independent sub-jumpers whose wealth-weighted means form a single product betting function, eliminating exponential state growth.
New algorithmic object introduced in Section 5.1; independent evidence is only the empirical trajectories and the formal martingale property, not an external falsifiable prediction.
Jumping tax
purpose: Name the capital dilution that occurs when jump mass is spread uniformly over an exponentially large product state space.
Conceptual label for a known mixture-cost phenomenon; useful but not an independent physical or statistical entity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Betting on Moments: Legendre Jumper Martingales for Online Exchangeability Testing." pith.science (2026). https://pith.science/paper/2VLRNQZZ

@misc{pith2026260620859,
  author       = {Pith},
  title        = {Pith review of: Betting on Moments: Legendre Jumper Martingales for Online Exchangeability Testing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2VLRNQZZ}},
  note         = {Machine review of arXiv:2606.20859}
}

read the original abstract

A fundamental assumption in statistics and machine learning is that ``the future looks like the past,'' formalized as exchangeability: the joint data distribution is order-invariant. In practice, this assumption is often violated due to distribution shifts over time. Early detection of exchangeability violations is crucial to prevent performance degradation and enable timely interventions like model retraining. Conformal test martingales offer a flexible, distribution-free framework for sequential exchangeability testing with guaranteed false-alarm rate control by betting against the uniformity of conformal p-values. While alternatives such as plug-in martingales and mixture-based strategies exist, computationally efficient baselines like the Simple Jumper are limited to detecting mean location shifts. We propose a family of conformal test martingales based on shifted Legendre polynomials that extend the Simple Jumper to higher-order moments. The Simple Legendre Jumper replaces linear betting functions with polynomials of arbitrary degree, enabling rapid detection of variance, skewness, and other higher-order deviations. The Product Legendre Jumper combines multiple polynomial degrees into a single betting function but suffers from exponential state-space growth, termed the jumping tax. To resolve this, we introduce the Variational Legendre Jumper, which employs a mean-field approximation to reduce complexity to constant time per step with minimal power loss, providing an expressive, scalable framework for real-time distribution shift monitoring.

Figures

Figures reproduced from arXiv: 2606.20859 by the authors.

**Figure 2.** Betting functions of the Simple Legendre Jumper for [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Trajectories of two Simple Legendre Jumper martingales of degrees [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Some betting functions of the Product Legendre Jumper for [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Trajectories of Product Legendre Martingales with degrees [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Trajectories of the Product Legendre Jumper and the Variational Legendre Jumper, both with [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Trajectories of the Product Legendre Jumper and the Variational Legendre Jumper, both with [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

**Figure 8.** Figure 8: Log-martingale trajectories (log10 Mn) for seven Legendre Jumper configurations on the Wine Quality dataset under four orderings. Under exchangeability (top-left), all martingales remain near zero. Under distributional shift, PLJ and VLJ consistently dominate SLJ(k=1),…

**Figure 9.** Figure 9: Fixed-J versus Composite Legendre Jumpers on the Wine Quality dataset. Solid lines denote composite variants (averaging over five jumping rates); dashed lines denote fixed J = 0.01. The composite provides a wealth floor under exchangeability (top-left) and substantiall…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 1 canonical work pages

[1]

doi: 10.1080/01621459.2017. 1285773. URLhttps://doi.org/10.1080/01621459.2017.1285773. Henrik Boström.An Investigation of Conformal Test Martingales, pp. 118–143. Springer Nature Switzerland, Cham,

doi:10.1080/01621459.2017 2017
[2]

doi: 10.1007/978-3-032-15120-9_7

ISBN 978-3-032-15120-9. doi: 10.1007/978-3-032-15120-9_7. URLhttps://doi.org/10. 1007/978-3-032-15120-9_7. Paulo Cortez, António Cerdeira, Fernando Almeida, Telmo Matos, and José Reis. Modeling wine preferences by data mining from physicochemical properties.Decision Support Systems, 47(4):547–553,

doi:10.1007/978-3-032-15120-9_7
[3]

Bruno De Finetti

doi: 10.1016/j.dss.2009.05.016. Bruno De Finetti. La prévision: ses lois logiques, ses sources subjectives. InAnnales de l’institut Henri Poincaré, volume 7, pp. 1–68,

doi:10.1016/j.dss.2009.05.016 2009
[4]

URLhttp://www

ISSN 00029947, 10886850. URLhttp://www. jstor.org/stable/1992999. Teresa Ledwina. Data-driven version of neyman’s smooth test of fit.Journal of the American Statistical Association, 89(427):1000–1005,

arXiv
[5]

John CW Rayner and DJ Best.Smooth tests of goodness of fit

URLhttps://doi.org/10.1080/14786440009463897. John CW Rayner and DJ Best.Smooth tests of goodness of fit. Oxford University Press,

doi:10.1080/14786440009463897
[6]

URLhttp://www.jstor.org/stable/2958889

ISSN 00905364, 21688966. URLhttp://www.jstor.org/stable/2958889. Gábor Szegő.Orthogonal polynomials, volume

arXiv
[7]

Vladimir Vovk, Ilia Nouretdinov, and Alexander Gammerman

URLhttps: //arxiv.org/abs/2512.22162. Vladimir Vovk, Ilia Nouretdinov, and Alexander Gammerman. Testing exchangeability on-line.Proceedings of the Twentieth International Conference on Machine Learning (ICML-2003), pp. 768–775,

arXiv 2003
[8]

SpringerInternationalPublishing, 12022

VladimirVovk, AlexanderGammerman, andGlennShafer.Algorithmic Learning in a Random World, Second Edition. SpringerInternationalPublishing, 12022. ISBN9783031066498. doi: 10.1007/978-3-031-06649-8/ COVER. Martin J. Wainwright and Michael I. Jordan. Graphical models, exponential families, and variational in- ference.Foundations and Trends in Machine Learning...

work page doi:10.1007/978-3-031-06649-8/

Show all 9 references

[9]

doi: 10.1561/2200000001

ISSN 1935-8237. doi: 10.1561/2200000001. URLhttps://doi.org/10.1561/2200000001. 18

1935 doi

Pith tools

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

[1] [1]

doi: 10.1080/01621459.2017. 1285773. URLhttps://doi.org/10.1080/01621459.2017.1285773. Henrik Boström.An Investigation of Conformal Test Martingales, pp. 118–143. Springer Nature Switzerland, Cham,

doi:10.1080/01621459.2017 2017

[2] [2]

doi: 10.1007/978-3-032-15120-9_7

ISBN 978-3-032-15120-9. doi: 10.1007/978-3-032-15120-9_7. URLhttps://doi.org/10. 1007/978-3-032-15120-9_7. Paulo Cortez, António Cerdeira, Fernando Almeida, Telmo Matos, and José Reis. Modeling wine preferences by data mining from physicochemical properties.Decision Support Systems, 47(4):547–553,

doi:10.1007/978-3-032-15120-9_7

[3] [3]

Bruno De Finetti

doi: 10.1016/j.dss.2009.05.016. Bruno De Finetti. La prévision: ses lois logiques, ses sources subjectives. InAnnales de l’institut Henri Poincaré, volume 7, pp. 1–68,

doi:10.1016/j.dss.2009.05.016 2009

[4] [4]

URLhttp://www

ISSN 00029947, 10886850. URLhttp://www. jstor.org/stable/1992999. Teresa Ledwina. Data-driven version of neyman’s smooth test of fit.Journal of the American Statistical Association, 89(427):1000–1005,

arXiv

[5] [5]

John CW Rayner and DJ Best.Smooth tests of goodness of fit

URLhttps://doi.org/10.1080/14786440009463897. John CW Rayner and DJ Best.Smooth tests of goodness of fit. Oxford University Press,

doi:10.1080/14786440009463897

[6] [6]

URLhttp://www.jstor.org/stable/2958889

ISSN 00905364, 21688966. URLhttp://www.jstor.org/stable/2958889. Gábor Szegő.Orthogonal polynomials, volume

arXiv

[7] [7]

Vladimir Vovk, Ilia Nouretdinov, and Alexander Gammerman

URLhttps: //arxiv.org/abs/2512.22162. Vladimir Vovk, Ilia Nouretdinov, and Alexander Gammerman. Testing exchangeability on-line.Proceedings of the Twentieth International Conference on Machine Learning (ICML-2003), pp. 768–775,

arXiv 2003

[8] [8]

SpringerInternationalPublishing, 12022

VladimirVovk, AlexanderGammerman, andGlennShafer.Algorithmic Learning in a Random World, Second Edition. SpringerInternationalPublishing, 12022. ISBN9783031066498. doi: 10.1007/978-3-031-06649-8/ COVER. Martin J. Wainwright and Michael I. Jordan. Graphical models, exponential families, and variational in- ference.Foundations and Trends in Machine Learning...

work page doi:10.1007/978-3-031-06649-8/