Power Enhancement of Permutation-Augmented Partial-Correlation Tests via Fixed-Row Permutations

Changliang Zou; Guanghui Wang; Tianyi Wang; Zhaojun Wang

REVIEW 3 major objections 4 minor 14 references

Power Enhancement of Permutation-Augmented Partial-Correlation Tests via Fixed-Row Permutations

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Permutation tests regain power by fixing selected rows, preserving the 2α error bound

desk verdict Useful fix for PALMRT's power collapse, but the power theory in Proposition 1 has a real quantile-direction error; the method is valid and promising, the stated lower bound is not. read the letter →

arxiv 2506.02906 v1 pith:SWKVOFQJ submitted 2025-06-03 stat.ME math.STstat.TH

Tianyi Wang , Guanghui Wang , Zhaojun Wang , Changliang Zou This is my paper

classification stat.MEmath.STstat.TH MSC 62G1062G0962J05

keywords CollinearityExchangeablenoisesGroupinvariancetestPartial-correlationPermutationmethodsFixed-rowpermutationsPowerenhancementConformalp-value

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

The paper claims that the power collapse of permutation-augmented partial-correlation tests, caused by the permuted design aligning with the covariate of interest, can be cured by permuting only a design-selected subset of rows. It proves that sampling from the resulting restricted-permutation subgroup preserves the finite-sample worst-case Type I error bound pr(pval ≤ α | H0) < 2α for any fixed design and exchangeable noise. The fixed rows are chosen greedily to maximize the α-quantile of a squared residual norm, which acts as a lower bound on power. Simulations show the refined test keeps nominal size and gains substantial power over unrestricted permutations, especially in high-collinearity designs.

What carries the argument

The central object is the restricted permutation subgroup S_{n|R}, the set of permutation matrices whose action fixes every row in R; since it is a subgroup, uniform sampling from it preserves exchangeability and the conformal argument behind PALMRT. Row choice is driven by two constructs: the squared residual norm D_R = ||(I - H_{P_{π|R}X, Z, P_{π|R}Z})X||$_2^{2}$, whose α-quantile Proposition 1 ties to a lower bound on power, and the rank criterion μ_R(Z) = E{rank(Z - P_{π|R}Z)}, whose monotonicity guarantees that adding rows in Stage I shrinks the danger of X lying in C(Z, P_{π|R}Z).

What would settle it

Construct a design where the greedy Stage II row set has a strictly larger α-quantile of D_R than the unrestricted set yet lower empirical rejection probability at a fixed β; finding such a design would show the quantile proxy does not order power. Concretely, simulate the paired design with p=40 and check whether the rows returned by Algorithm 1 satisfy pr(X ∈ C(Z, P_{π|R}Z)) < 0.05; if not, the shield stage failed to deliver its safety condition.

Watch

Extended reading notes

Core claim

For any fixed design matrix (X,Z) and arbitrary exchangeable noise, replacing the full permutation group by any subgroup that fixes a set R of rows yields a valid conformal p-value with pr(pval ≤ α | H0) < 2α for all α>0. The new contribution is a design-dependent choice of R that reduces the probability that X falls into the column space C(P_{π|R}X, Z, P_{π|R}Z), which is exactly the event that makes paired F-statistics unable to reject. A two-stage greedy algorithm first builds a shield to keep X outside C(Z, P_{π|R}Z), then augments rows to maximize the α-quantile of the squared residual norm D_R, giving a provable lower bound on power. In the paired-design example where the original test has zero size and zero power, the row-selected version recovers both.

Load-bearing premise

The greedy algorithm assumes that the α-quantile of the squared residual norm is a reliable proxy for power; the proof gives a lower bound that grows with this quantile but does not establish that actual power is monotone in the quantile for every design.

Editorial extensions

If this is right

Any user of PALMRT can replace full permutations with fixed-row permutations and keep the same worst-case 2α Type I error guarantee.
In high-collinearity designs such as the paired design, the procedure removes the zero-size and zero-power failure mode of unrestricted PALMRT.
In low-collinearity designs, fixing rows does not hurt power, so the refinement is safe to apply uniformly.
The power lower bound in Proposition 1 means the greedy choice of R comes with a certificate: the selected set is at least as good, in the bound, as any smaller set found along the greedy path.
Because the argument only needs a subgroup, the same fixed-row idea extends to any paired statistic satisfying the transferability condition.

Reading between the lines

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

The fixed-row idea should transfer to other permutation-augmented tests, such as residual-based tests, wherever permuted design columns can align with the covariate of interest.
In structured designs, the shield and augmentation steps could be replaced by closed-form row sets, reducing the Monte Carlo cost of row selection.
A practical extension would be an automatic choice of the budget T and the number of restricted permutations B, balancing the power lower bound against computation.

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

Quick take: the fixed-row permutation trick is a genuinely useful repair for PALMRT's power collapse, and Theorem 1 is correct. But Proposition 1, the result that supposedly justifies the greedy row-selection rule, has a quantile-direction error in its proof. The power-enhancement claim currently rests on simulations, not on the stated lower bound.

What's new: the design-driven, greedy choice of fixed rows is a sensible application of the subgroup-permutation idea to permutation-augmented partial-correlation tests. The paper correctly observes that unrestricted PALMRT can lose power when permutation-augmented columns are collinear with X, and that fixing rows breaks the collinearity while preserving exchangeability. Theorem 1's argument is a direct subgroup extension of Guan's conformal p-value bound, and it holds for any design-dependent R. The paired-design simulations are dramatic: PALMRT drops to 0% size/power, and the fixed-row version recovers nominal size and good power. Honestly, the core method is valid and likely useful.

Where it's soft: the proof of Proposition 1 (Appendix A.4) doesn't deliver the claimed lower bound. The event {max_Pπ g(Pπ,ε) ≤ β² F^{-1}_{D_R}(α−δ)} implies only that pr(g ≥ β²D_R) ≤ F_{D_R}(F^{-1}(α−δ)) ≥ α−δ. That's the wrong direction: it doesn't control the proportion of permutations below α−δ. The proposition statement even writes α+O(...) where the proof uses α−O(...). So the greedy objective 'maximize F^{-1}_{D_R}(α)' is not backed by a proven power bound. It's a heuristic with intuitive appeal and decent simulation support, but the paper oversells it as a lower-bound-maximizer. The Stage I shield is also under-justified: monotonicity of μ_R(Z) doesn't guarantee that greedy additions reach pr(X∈C(Z,P_{π|R}Z)) < α, or that the chosen R is safe. Simulations report medians without error bars; the mixture design gains are less crisp than the paired gains.

None of this touches Theorem 1. The Type I error guarantee is intact for any R. So this is a sound method with a broken theoretical advertisement.

Who it's for: people working on permutation tests, conformal inference, or regression diagnostics. It deserves a serious referee. I would recommend major revision: fix or honestly qualify Proposition 1, tighten the Stage I analysis, and report simulation variability. The central idea is worth publishing, but not as-is.

Referee Report

3 major / 4 minor

Summary. The paper proposes a modification of PALMRT for testing H0: beta = 0 in the fixed-design linear model Y = X beta + Z theta + epsilon with exchangeable noise. Instead of sampling uniformly from the full symmetric group, the authors fix a design-dependent subset R of rows and permute only the remaining rows, sampling uniformly from the subgroup S_{n|R}. They prove (Theorem 1) that the conformal p-value retains the worst-case guarantee pr(pval <= alpha) < 2 alpha for any design-dependent R, by a subgroup-invariance argument. They propose a two-stage greedy algorithm: Stage I builds a 'shield' by adding rows that minimize mu_R(Z) until pr(X in C(Z, P_{pi|R} Z)) < alpha; Stage II adds rows that maximize the alpha-quantile of the squared residual norm D_R = ||(I - H_{P_{pi|R} X, Z, P_{pi|R} Z}) X||^2. Proposition 1 is claimed to give a power lower bound increasing in this quantile, and Proposition 2 gives monotonicity properties used to justify the greedy steps. Simulations compare the row-selected PALMRT with original PALMRT and CPT across Gaussian, paired, and mixture designs with four noise distributions. The reported findings are that the row-selection restores nontrivial size and power in the paired design where original PALMRT has zero power, and gives power gains in the mixture design while matching PALMRT in the Gaussian design.

Significance. If the type I error guarantee (Theorem 1) and the power improvements in the simulations hold, this is a practically useful and conceptually clean contribution: it shows that the subgroup-invariance route to conformal validity can be used to avoid the collinearity collapse of permutation-augmented regression tests, and it connects to the recent 'representative subgroup' literature. The paper ships a theorem with a short proof, a clear algorithmic procedure, and simulations spanning heavy-tailed, discrete, and high-collinearity settings. The idea of fixing rows of the design to reduce covariate-permutation collinearity while preserving worst-case coverage is natural and likely to be of interest to the conformal-prediction and permutation-testing community. The main caveat is that the promised power rationale (Proposition 1) currently carries a quantile-direction gap in its proof, so the greedy objective is not yet theoretically justified as a lower-bound maximizer; the simulations are nevertheless extensive and the type I error claim is independent of that gap.

major comments (3)

[§2.3, Proposition 1 and Appendix A.4] The proof of Proposition 1 has a quantile-direction error that is load-bearing for the paper's central power rationale. The final step claims that the event {max_{P_pi} g <= beta^2 F^{-1}_{D_R}(alpha - delta)} gives the stated lower bound on power, but F_{D_R}(F^{-1}_{D_R}(alpha - delta)) >= alpha - delta, not <= alpha - delta, and for a fixed design D_R is discrete, so the quantile can be flat inside an atom. The proof needs an upper bound on pr(g >= beta^2 D_R), which the displayed event does not provide; it would need instead a condition such as g <= beta^2 times a value q with F_{D_R}(q) <= alpha - delta, i.e., a lower quantile rather than the standard upper quantile. Moreover, the proposition statement writes alpha + O(B^{-1/2}(log B)^{1/2}) while the proof uses alpha - O(B^{-1/2}(log B)^{1/2}), and the asymmetry matters inside atoms. As written, the derivation therefore does not establish that maximizing F^{-1}_{D_R}(alpha) maximizes a lower bound on power, and the greedy objective in Stage II currently rests on the simulations alone.
[§2.3, Stage I shield criterion and Proposition 2] The safety criterion used to terminate Stage I is not fully justified. Proposition 2(i) only shows that if pr(X in C(Z, P_{pi|R} Z)) >= alpha then F^{-1}_{D_R}(alpha) = 0; it does not show the converse, nor does it show that the greedy minimization of mu_R(Z) will reach a state with pr(X in C(Z, P_{pi|R} Z)) < alpha for a general design before R = [n]. The monotonicity of mu_R(Z) in Proposition 2(ii) is proven, but the link between reductions in mu_R(Z) and reductions in the safety probability is asserted rather than derived, and Remark 2 only states that both quantities can be estimated. Since Theorem 1 is unaffected, this is not a correctness flaw in the validity claim, but the algorithmic guarantee for Stage I and the statement that each added row 'strictly improves the collinearity shield' go beyond what Proposition 2 demonstrates.
[§3, Figure 2 and Table 1] The power comparison in the paired and mixture designs appears to calibrate the row-selected PALMRT at alpha = 5% and the original PALMRT at the same level, while the dashed curves calibrate both at alpha/2, but the numerical results are presented as median curves over only 50 design replicates with B = 2,000 permutations. Given the discreteness of the permutation p-value and the conservativeness of the 2-alpha guarantee, it would strengthen the power claim to report the distribution of the empirical rejection rates across replicates (or a Monte Carlo standard error) and to state explicitly which replicates select row sets that differ from the hand-constructed target set {2, ..., p+2} in the paired design. Without this, it is hard to judge how much of the reported gain is specific to the particular random draws of the design.

minor comments (4)

[Throughout] The notation S_{n|R,k->j} is used in Appendices A.3 and A.5 but not defined before first use; a one-sentence definition would improve readability.
[§2.2, Theorem 1] The theorem states pr(pval <= alpha | H0) < 2 alpha for all alpha > 0; the bound should presumably be restricted to alpha <= 1, and the convention for the p-value in Eq. (2) should be stated consistently with Guan (2024) regarding ties, since the displayed formula includes a tie term that is easy to overlook in the proof of Theorem 1.
[§3, Table 1] The size values for the Gaussian design (about 2%) are well below the nominal 5% for both PALMRT variants, which is consistent with conservativeness but should be commented on explicitly, as the text says the methods 'control the nominal level' without noting the degree of conservativeness.
[§3.2, Paired design] The claim that CPT 'achieves moderate power' and 'excels for weak signals under multinomial noises' is based on the medians in Figure 2, but the panel for multinomial noise in the paired design is hard to read at the lowest beta values; enlarging the small-beta region or adding a table with numerical values at selected beta would clarify the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the validity result follows from subgroup invariance and the row-selection objective is design-based, not fitted to the response.

full rationale

The paper's Type I error guarantee (Theorem 1) is derived from the subgroup property of S_{n|R} for any design-dependent R, together with the transferability Condition 1 imported from Guan (2024); no parameter is fitted to the response and no quantity in the theorem is defined in terms of the p-value it seeks to bound. The power-oriented row selection is based on an explicit lower-bound calculation (Proposition 1) that expresses power in terms of a design-only quantile F^{-1}_{D_R}(α; X, Z); although the proof's use of α−O(B^{-1/2}(log B)^{1/2}) versus the statement's α+O(...) is a correctness concern, it is not a circular reduction—the bound is not calibrated to observed responses or to simulation outcomes. The row set R = R(X, Z) depends only on the design and α, not on Y, so the power gains reported in simulations are not fitted re-statements of the test output. The Wen et al. (2025) self-citation is a prior residual permutation test used as an external baseline ingredient, not as justification for the present paper's central claims. No load-bearing step reduces, by the paper's own equations or by self-citation, to its own inputs.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. It relies on the fixed design, exchangeable noise, and Guan's transferability condition; the rank condition n > 2p is an implicit assumption inherited from PALMRT.

assumptions (4)

domain assumption Noise vector epsilon is exchangeable under any fixed design.
Assumed throughout and stated in the abstract and Section 1; needed for the subgroup-based conformal argument.
domain assumption The paired statistics satisfy Condition 1 (transferability).
Inherited from Guan (2024); Theorem 1 and the power bound rely on it (Section 2.2).
domain assumption X is not in the column space of Z.
Stated in Section 1.2 so that beta is identifiable; without it the testing problem is degenerate.
domain assumption n is large enough relative to p for the augmented residual projections to be non-degenerate (e.g., n > 2p in PALMRT).
Not explicitly stated in Theorem 1 but required for nontrivial D_R; simulations use n=100 with p=15 or 40.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Power Enhancement of Permutation-Augmented Partial-Correlation Tests via Fixed-Row Permutations." pith.science (2026). https://pith.science/paper/SWKVOFQJ

@misc{pith2026250602906,
  author       = {Pith},
  title        = {Pith review of: Power Enhancement of Permutation-Augmented Partial-Correlation Tests via Fixed-Row Permutations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWKVOFQJ}},
  note         = {Machine review of arXiv:2506.02906}
}

read the original abstract

Permutation-based partial-correlation tests guarantee finite-sample Type I error control under any fixed design and exchangeable noise, yet their power can collapse when the permutation-augmented design aligns too closely with the covariate of interest. We remedy this by fixing a design-driven subset of rows and permuting only the remainder. The fixed rows are chosen by a greedy algorithm that maximizes a lower bound on power. This strategy reduces covariate-permutation collinearity while preserving worst-case Type I error control. Simulations confirm that this refinement maintains nominal size and delivers substantial power gains over original unrestricted permutations, especially in high-collinearity regimes.

Figures

Figures reproduced from arXiv: 2506.02906 by the authors.

**Figure 1.** Comparison of PALMRT’s performance with and without row-selection in a paired design with standard Gaussian noise (Y = 4X + ϵ) for varying p. The left panel illustrates the probability of collinearity, while the right panel shows the test power. The significance level is set to 5%. row indices R is fixed (see Section 2.1). These rows are selected greedily to maximize the angle between X and the space C(Pπi|RX, Z, Pπ… view at source ↗

**Figure 2.** Empirical power curves versus signal strength at the nominal level α = 5% across designs and noise distributions. PALMRTs with (red) and without (green) row-selection are each calibrated at α = 5% (solid lines) and α/2 = 2.5% (dashed lines), while CPT (blue, solid lines) is shown at α = 5%. To assess power we generate responses from Y = Xβ + ϵ, varying the signal strength β. 11 [PITH_FULL_IMAGE:figures/full_fig_p01… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 12 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := #2 '...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in " " * FUNCTION format....
[3]

F., Cand\`es, E

Barber, R. F., Cand\`es, E. J., Ramdas, A., and Tibshirani, R. J. (2021), Predictive inference with the jackknife+, Ann. Statist., 49, 486--507

work page 2021
[4]

(2024), A conformal test of linear models via permutation-augmented regressions, Ann

Guan, L. (2024), A conformal test of linear models via permutation-augmented regressions, Ann. Statist., 52, 2059--2080

work page 2024
[5]

Hartigan, J. A. (1970), Exact confidence intervals in regression problems with independent symmetric errors, Ann. Math. Statist., 41, 1992--1998

work page 1970
[6]

and Goeman, J

Hemerik, J. and Goeman, J. (2018), Exact testing with random permutations, TEST, 27, 811--825

work page 2018
[7]

Koning, N. W. (2024), More power by using fewer permutations, Biometrika, 111, 1405--1412

work page 2024
[8]

Koning, N. W. and Hemerik, J. (2024), More efficient exact group invariance testing: using a representative subgroup, Biometrika, 111, 441--458

work page 2024

Show all 14 references

[9]

and Bickel, P

Lei, L. and Bickel, P. J. (2021 a ), An assumption-free exact test for fixed-design linear models with exchangeable errors, Biometrika, 108, 397--412

2021
[10]

An Assumption-Free Exact Test For Fixed-Design Linear Models With Exchangeable Errors

--- (2021 b ), Supplementary Material of "An Assumption-Free Exact Test For Fixed-Design Linear Models With Exchangeable Errors", Biometrika, 108

2021
[11]

(2015), Group bound: confidence intervals for groups of variables in sparse high dimensional regression without assumptions on the design, J

Meinshausen, N. (2015), Group bound: confidence intervals for groups of variables in sparse high dimensional regression without assumptions on the design, J. R. Stat. Soc. Ser. B. Stat. Methodol., 77, 923--945

2015
[12]

F., Cand\`es, E

Ramdas, A., Barber, R. F., Cand\`es, E. J., and Tibshirani, R. J. (2023), Permutation tests using arbitrary permutation distributions, Sankhya A, 85, 1156--1177

2023
[13]

(2018), Cross-conformal predictive distributions, in conformal and probabilistic prediction and applications, PMLR, pp

Vovk, V., Nouretdinov, I., Manokhin, V., and Gammerman, A. (2018), Cross-conformal predictive distributions, in conformal and probabilistic prediction and applications, PMLR, pp. 37--51

2018
[14]

(2025), Residual permutation test for regression coefficient testing, Ann

Wen, K., Wang, T., and Wang, Y. (2025), Residual permutation test for regression coefficient testing, Ann. Statist., 53, 724--748

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := #2 '...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in " " * FUNCTION format....

[3] [3]

F., Cand\`es, E

Barber, R. F., Cand\`es, E. J., Ramdas, A., and Tibshirani, R. J. (2021), Predictive inference with the jackknife+, Ann. Statist., 49, 486--507

work page 2021

[4] [4]

(2024), A conformal test of linear models via permutation-augmented regressions, Ann

Guan, L. (2024), A conformal test of linear models via permutation-augmented regressions, Ann. Statist., 52, 2059--2080

work page 2024

[5] [5]

Hartigan, J. A. (1970), Exact confidence intervals in regression problems with independent symmetric errors, Ann. Math. Statist., 41, 1992--1998

work page 1970

[6] [6]

and Goeman, J

Hemerik, J. and Goeman, J. (2018), Exact testing with random permutations, TEST, 27, 811--825

work page 2018

[7] [7]

Koning, N. W. (2024), More power by using fewer permutations, Biometrika, 111, 1405--1412

work page 2024

[8] [8]

Koning, N. W. and Hemerik, J. (2024), More efficient exact group invariance testing: using a representative subgroup, Biometrika, 111, 441--458

work page 2024