pith. machine review for the scientific record. sign in

arxiv: 2602.03986 · v2 · submitted 2026-02-03 · 💻 cs.LG · cs.RO· cs.SY· eess.SY

Recognition: 2 theorem links

· Lean Theorem

eCP: Equivariant Conformal Prediction with pre-trained models

Nikolaos Bousias , Lars Lindemann , George Pappas
Authors on Pith no claims yet

Pith reviewed 2026-05-16 07:42 UTC · model grok-4.3

classification 💻 cs.LG cs.ROcs.SYeess.SY
keywords conformal predictionequivariant modelsgroup averagingnon-conformity scoresuncertainty quantificationtrajectory predictionsymmetry orbitsexchangeability
0
0 comments X

The pith

Group averaging of pre-trained predictors over known symmetries contracts non-conformity scores in increasing convex order.

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

Conformal prediction supplies distribution-free coverage guarantees but often produces large uncertainty regions in long-horizon settings. The paper shows that averaging a pre-trained model over the actions of a known symmetry group spreads non-conformity mass across orbits, so each point is evaluated as a representative of its symmetric equivalents. This operation produces non-conformity scores that are smaller in increasing convex order. The contraction tightens exponential tail bounds and yields narrower prediction sets on average, most noticeably at high confidence levels. Readers care because the improvement keeps the original finite-sample coverage while exploiting geometric structure already present in many robotics and vision tasks.

Core claim

Infusing conformal prediction with geometric symmetry via group-averaging of the pretrained predictor distributes non-conformity scores across orbits. The resulting scores are provably contracted in increasing convex order, which delivers improved exponential-tail bounds and sharper conformal sets in expectation, especially at high confidence levels, while the exchangeability assumption and coverage guarantees remain intact.

What carries the argument

Group-averaging the pretrained predictor over the known symmetry group actions to induce equivariant non-conformity scores across orbits.

If this is right

  • Non-conformity scores contract in increasing convex order.
  • Exponential tail bounds on the scores improve.
  • Conformal prediction sets become narrower in expectation.
  • The tightening is strongest at high confidence levels.
  • The method applies directly to long-horizon tasks such as pedestrian trajectory prediction.

Where Pith is reading between the lines

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

  • The same averaging step may reduce the calibration-set size needed to reach a target width in other symmetric domains.
  • Learned or approximate symmetry groups could extend the method beyond cases where the group is given exactly.
  • The convex-order contraction supplies a new quantitative link between equivariance and statistical efficiency that could be studied in isolation.
  • The approach is compatible with any base conformal score and therefore composes with existing post-hoc refinements.

Load-bearing premise

The symmetry group must be known in advance and the averaging step must preserve the exchangeability required for valid conformal coverage.

What would settle it

On a dataset with known symmetries, the group-averaged non-conformity scores fail to be smaller in increasing convex order than the unaveraged scores, or the empirical coverage after averaging falls below the nominal level.

Figures

Figures reproduced from arXiv: 2602.03986 by George Pappas, Lars Lindemann, Nikolaos Bousias.

Figure 1
Figure 1. Figure 1: Set size reduction via symmetrization of Conformal prediction in multi-step prediction in ETH-UCY and NBA Rebound/Score datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation studies on the group size and approximate equivariance. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Non-conformity score distributions for SocialVAE and Eq [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Calibration quantile (Qα=0.05) and empirical coverage (Cov95%) for 15-split conformal prediction under symmetry groups G ∈ {SO(2),C4,C8} across ETH-UCY and SDD datasets. eCP methods consistently lead to smaller calibration quantiles, while preserving the prescribed coverage rate. = Z G s  f(φg ′−1 (x)), ψg ′−1 (y)  dµG(g ′ ) = ΠG[s; f ](x,y), where we used dµG(g) = dµG(g ′ ) and the group homomor￾phism p… view at source ↗
read the original abstract

Conformal prediction, a post-hoc, distribution-free, finite-sample method of uncertainty quantification that offers formal coverage guarantees under the assumption of data exchangeability. Unfortunately, the resulting uncertainty regions can grow significantly in long horizon missions, rendering the statistical guarantees uninformative. To that end, we propose infusing CP with geometric information via group-averaging of the pretrained predictor to distribute the non-conformity mass across the orbits. Each sample now is treated as a representative of an orbit, thus uncertainty can be mitigated by other samples entangled to it via the orbit inducing elements of the symmetry group. Our approach provably yields contracted non-conformity scores in increasing convex order, implying improved exponential-tail bounds and sharper conformal prediction sets in expectation, especially at high confidence levels. We then propose an experimental design to test these theoretical claims in pedestrian trajectory prediction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes eCP, which augments standard conformal prediction by applying a fixed group-averaging transformation to the outputs of a pre-trained predictor. This distributes non-conformity mass across orbits induced by a known symmetry group, with the central claim that the resulting scores are contracted in increasing convex order. The contraction is asserted to yield improved exponential-tail bounds and smaller conformal sets in expectation while preserving marginal coverage under exchangeability. An experimental design is outlined to evaluate the approach on pedestrian trajectory prediction.

Significance. If the contraction in increasing convex order is established, the method supplies a symmetry-aware efficiency improvement to conformal prediction that leaves validity untouched. This is a targeted advance for domains with known group actions (e.g., SE(2) or SE(3) trajectories) and could tighten high-confidence sets without extra data or retraining. The explicit preservation of exchangeability under pointwise averaging is a clean technical feature.

major comments (2)
  1. [Theoretical section on contraction] Theoretical derivation of the contraction (likely the section containing the main theorem): the abstract asserts that group-averaged scores are contracted in increasing convex order, but the provided text does not contain the full step-by-step argument, the precise definition of the averaged non-conformity score, or verification that the stochastic order holds for arbitrary group actions. This derivation is load-bearing for the efficiency claims and must be supplied in detail.
  2. [Non-conformity score definition] Definition of the non-conformity score after averaging (Eq. defining the score): it is unclear whether the averaging operation is applied before or after the residual computation and whether the resulting score remains a function of the original data point alone; any dependence on other orbit members would need to be shown not to violate the exchangeability argument used for coverage.
minor comments (2)
  1. [Experimental design] The experimental design paragraph should specify concrete baselines (e.g., standard CP, equivariant CP variants), quantitative metrics for set size and coverage at multiple confidence levels, and how the symmetry group is chosen or validated.
  2. [Notation and preliminaries] Notation for the symmetry group G and orbit representatives should be introduced once and used consistently; currently the abstract mixes “orbit” and “symmetry group” without a clear preliminary definition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review. The comments highlight important points for strengthening the theoretical presentation. We address each major comment below and will revise the manuscript to supply the requested details while preserving the core claims.

read point-by-point responses

  1. Referee: [Theoretical section on contraction] Theoretical derivation of the contraction (likely the section containing the main theorem): the abstract asserts that group-averaged scores are contracted in increasing convex order, but the provided text does not contain the full step-by-step argument, the precise definition of the averaged non-conformity score, or verification that the stochastic order holds for arbitrary group actions. This derivation is load-bearing for the efficiency claims and must be supplied in detail.

    Authors: We agree that the submitted manuscript omitted a complete step-by-step derivation. In the revised version we will expand the theoretical section with a full proof. The averaged non-conformity score is defined by first applying the fixed group actions to the pre-trained predictor outputs to obtain the orbit of predictions, then averaging the resulting residuals. We will prove contraction in increasing convex order for finite groups (the setting used in the pedestrian trajectory experiments) by showing that the group average is a convex combination that reduces the spread of the score distribution, and we will state the precise conditions under which the stochastic order extends to compact groups. revision: yes

  2. Referee: [Non-conformity score definition] Definition of the non-conformity score after averaging (Eq. defining the score): it is unclear whether the averaging operation is applied before or after the residual computation and whether the resulting score remains a function of the original data point alone; any dependence on other orbit members would need to be shown not to violate the exchangeability argument used for coverage.

    Authors: We will clarify the definition explicitly in the revision. The group averaging is applied to the predictor outputs before the residual (non-conformity) is computed. For any single data point the orbit is generated deterministically by applying the known group elements to that point alone; the averaged score is therefore a function of the individual observation. Because the transformation is fixed, deterministic, and identical across all points, exchangeability of the original data implies exchangeability of the transformed scores, so the standard marginal coverage argument continues to hold. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained on standard CP and group averaging

full rationale

The central claim applies a fixed group-averaging transformation to a pretrained predictor to obtain new non-conformity scores. Exchangeability of the original data tuples is inherited directly by the transformed scores because the averaging operator is identical and independent for every point; the rank of the test score among calibration scores therefore remains uniform, preserving marginal coverage by the standard conformal argument. The additional statement that the new scores are contracted in increasing convex order follows from the properties of the averaging operator applied to the non-conformity function and is an efficiency result separate from validity. No equation reduces a prediction to a fitted parameter defined by the same paper, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled via self-citation. The construction therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method rests on the standard exchangeability assumption of conformal prediction plus the existence of a usable symmetry group for the pretrained model.

axioms (2)
  • domain assumption Data exchangeability assumption required for conformal coverage guarantees
    Invoked implicitly as the foundation for all conformal prediction results.
  • domain assumption Symmetry group is known and the predictor can be averaged over its actions
    Central premise that enables distribution of non-conformity mass across orbits.

pith-pipeline@v0.9.0 · 5451 in / 1262 out tokens · 28884 ms · 2026-05-16T07:42:41.597254+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

  1. [1]

    V ovk, A

    V . V ovk, A. Gammerman, and G. Shafer,Algorithmic Learning in a Random World. Springer, 2005

    work page 2005

  2. [2]

    A tutorial on conformal prediction,

    G. Shafer and V . V ovk, “A tutorial on conformal prediction,”Journal of Machine Learning Research, vol. 9, pp. 371–421, 2008

    work page 2008

  3. [3]

    Distribution-free predictive inference for regression,

    J. Lei, M. G’Sell, A. Rinaldo, R. J. Tibshirani, and L. Wasserman, “Distribution-free predictive inference for regression,”Journal of the American Statistical Association, vol. 113, no. 523, pp. 1094–1111, 2018

    work page 2018

  4. [4]

    A gentle introduction to conformal prediction and distribution-free uncertainty quantification,

    A. N. Angelopoulos and S. Bates, “A gentle introduction to conformal prediction and distribution-free uncertainty quantification,”Founda- tions and Trends in Machine Learning, vol. 16, no. 4, pp. 494–682, 2021

    work page 2021

  5. [5]

    Conformalized quantile regression,

    Y . Romano, E. Patterson, and E. J. Cand `es, “Conformalized quantile regression,” inAdvances in Neural Information Processing Systems (NeurIPS), 2019

    work page 2019

  6. [6]

    Conformal prediction under covariate shift,

    R. J. Tibshirani, R. F. Barber, E. J. Cand `es, and A. Ramdas, “Conformal prediction under covariate shift,” inAdvances in Neural Information Processing Systems, 2019

    work page 2019

  7. [7]

    Adaptive conformal inference under distribution shift,

    A. Gibbs and E. J. Cand `es, “Adaptive conformal inference under distribution shift,” inAdvances in Neural Information Processing Systems, 2021

    work page 2021

  8. [8]

    Copula conformal prediction for multi-step time series prediction,

    S. H. Sun and R. Yu, “Copula conformal prediction for multi-step time series prediction,” inThe Twelfth International Conference on Learning Representations, 2023

    work page 2023

  9. [9]

    Multi-modal conformal prediction regions by optimizing convex shape templates,

    R. Tumu, M. Cleaveland, R. Mangharam, G. Pappas, and L. Lin- demann, “Multi-modal conformal prediction regions by optimizing convex shape templates,” inProceedings of the 6th Annual Learn- ing for Dynamics &; Control Conference(A. Abate, M. Cannon, K. Margellos, and A. Papachristodoulou, eds.), vol. 242 ofProceedings of Machine Learning Research, pp. 1343...

    work page 2024

  10. [10]

    Safe planning in dynamic environments using conformal prediction,

    L. Lindemann, M. Cleaveland, G. Shim, and G. J. Pappas, “Safe planning in dynamic environments using conformal prediction,”IEEE Robotics and Automation Letters, vol. 8, no. 8, pp. 5116–5123, 2023

    work page 2023

  11. [11]

    Conformal prediction for uncertainty-aware planning with dif- fusion dynamics model,

    J. Sun, Y . Jiang, J. Qiu, P. Nobel, M. J. Kochenderfer, and M. Schwa- ger, “Conformal prediction for uncertainty-aware planning with dif- fusion dynamics model,”Advances in Neural Information Processing Systems, vol. 36, pp. 80324–80337, 2023

    work page 2023

  12. [12]

    Group equivariant convolutional net- works,

    T. S. Cohen and M. Welling, “Group equivariant convolutional net- works,” inProceedings of the 33rd International Conference on Machine Learning, pp. 2990–2999, 2016

    work page 2016

  13. [13]

    Convolutional networks for spherical signals,

    T. S. Cohen, M. Geiger, J. K ¨ohler, and M. Welling, “Convolutional networks for spherical signals,” inProceedings of the 36th Interna- tional Conference on Machine Learning, pp. 1321–1330, 2019

    work page 2019

  14. [14]

    Symmetries-enhanced multi-agent reinforcement learning,

    N. Bousias, S. Pertigkiozoglou, K. Daniilidis, and G. Pappas, “Symmetries-enhanced multi-agent reinforcement learning,” inPro- ceedings of the 7th Annual Learning for Dynamics & Control Conference(N. Ozay, L. Balzano, D. Panagou, and A. Abate, eds.), vol. 283 ofProceedings of Machine Learning Research, pp. 999–1011, PMLR, 04–06 Jun 2025

    work page 2025

  15. [15]

    Deep equivariant multi- agent control barrier functions,

    N. Bousias, L. Lindemann, and G. Pappas, “Deep equivariant multi- agent control barrier functions,” 2025

    work page 2025

  16. [16]

    Geometric deep learning: Going beyond euclidean data,

    M. M. Bronstein, J. Bruna, Y . LeCun, A. Szlam, and P. Vandergheynst, “Geometric deep learning: Going beyond euclidean data,”IEEE Signal Processing Magazine, vol. 34, no. 4, pp. 18–42, 2017

    work page 2017

  17. [17]

    Data augmentation vs. equivariant networks: A theory of generalization on dynamics forecasting,

    R. Wang, R. Walters, and R. Yu, “Data augmentation vs. equivariant networks: A theory of generalization on dynamics forecasting,” 2022

    work page 2022

  18. [18]

    Symmpi: Predictive inference for data with group symmetries,

    E. Dobriban and M. Yu, “Symmpi: Predictive inference for data with group symmetries,” 2024

    work page 2024

  19. [19]

    Predictive inference with group symme- tries,

    J. Pillow and A. Rakhlin, “Predictive inference with group symme- tries,”arXiv preprint arXiv:2203.xxxx, 2022

    work page 2022

  20. [20]

    Distribution-free predictive inference with exchangeability and in- variance,

    S. Bates, E. J. Cand `es, J. Lei, Y . Romano, and R. J. Tibshirani, “Distribution-free predictive inference with exchangeability and in- variance,”Journal of the American Statistical Association, vol. 118, no. 541, pp. 1–15, 2023

    work page 2023

  21. [21]

    Cp 2: Leveraging geometry for conformal prediction via canonicalization,

    P. A. van der Linden, A. Timans, and E. J. Bekkers, “Cp 2: Leveraging geometry for conformal prediction via canonicalization,” 2025

    work page 2025

  22. [22]

    Conformal prediction: A gentle introduction,

    A. N. Angelopoulos and S. Bates, “Conformal prediction: A gentle introduction,”Foundations and Trends in Machine Learning, vol. 16, no. 5, pp. 494–591, 2023

    work page 2023

  23. [23]

    You'll never walk alone: Modeling social behavior for multi-target tracking,

    S. Pellegrini, A. Ess, K. Schindler, and L. V . Gool, “You'll never walk alone: Modeling social behavior for multi-target tracking,” in Proceedings of the IEEE International Conference on Computer Vision, pp. 261–268, 2009

    work page 2009

  24. [24]

    Learning social etiquette: Human trajectory understanding in crowded scenes,

    A. Robicquet, A. Sadeghian, A. Alahi, and S. Savarese, “Learning social etiquette: Human trajectory understanding in crowded scenes,” inProceedings of the European Conference on Computer Vision, pp. 549–565, 2016

    work page 2016

  25. [25]

    Socialvae: Human trajectory prediction using timewise latents,

    P. Xu, J.-B. Hayet, and I. Karamouzas, “Socialvae: Human trajectory prediction using timewise latents,” inEuropean Conference on Com- puter Vision, pp. 511–528, Springer, 2022

    work page 2022

  26. [26]

    Tutr: A transformer-based approach for trajectory prediction,

    H. Zhao, J. Chen, Y . Mao, and Q. Zhang, “Tutr: A transformer-based approach for trajectory prediction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13133– 13142, 2021

    work page 2021