Pith. sign in

REVIEW 4 major objections 4 minor 22 references

Symmetrization for high dimensional dependent random variables

T0 review · 4 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 This paper proves a symmetrization inequality for dependent, high-dimensional random vectors: the expected maximum of a sample mean is bounded by the same functional of a block-multiplied sum, up to a remainder that vanishes when a…

desk verdict Genuinely new symmetrization idea for dependent high-dimensional data, but the proof of Proposition 2.1 contains a load-bearing CDF error that breaks the main inequality as written. read the letter →

arxiv 2506.00547 v1 pith:7IGBZQUH submitted 2025-05-31 math.PR math.STstat.TH

Jonathan B. Hill This is my paper
classification math.PRmath.STstat.TH MSC 60F1060F25
keywords symmetrizationmaximalinequalityhigh-dimensionalrandomvectorsdependenceGaussianapproximationblockbootstrapmomentboundsphysical
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

Symmetrization is a standard tool for independent data: it bounds the expected maximum of a sample mean by the same functional applied to a randomly re-weighted version, often by introducing an independent copy of the data and Rademacher multipliers. This paper asks whether that comparison can survive dependence and high dimensionality, with the dimension $p$ much larger than the sample size $n$. Its answer is that the classic sandwich inequality, which bounds the symmetrized quantity above and below by the original mean, holds for arbitrary $\mathbb{R}^p$-valued dependent sequences once the multipliers are assigned block-wise and a remainder term is added. The remainder records the cost of blocking and of a truncation approximation, and it is controlled by high-dimensional Gaussian approximation distances that make the remainder vanish when they are small. The payoff is a Nemirovski-style maximal $L_q$ moment bound and a route to carrying independent-data maximal inequalities into weakly dependent, high-dimensional settings.

What carries the argument

Block-wise multiplier substitution: split the sample into $N_n = n/b_n$ consecutive blocks, attach independent multipliers $\varepsilon_l$ to each block, and compare the maximum of the original normalized sums with the maximum of the block-multiplied sums. The comparison is mediated by two objects: the Gaussian approximation Kolmogorov distances $\rho_n := \sup_{z\geq 0}|P(\max_i|X_n(i)|\leq z) - P(\max_i|\mathcal{X}_n(i)|\leq z)|$ and its blocked analogue $\rho^*_n$, together with a truncation level $U_n$ used to split the expectation into a bounded part and a tail part. The remainder formulas show that all dependence costs are captured by these distances; the argument itself is a change-of-variables and telescoping-block computation, not a distributional equality.

What would settle it

Take a dependent $\mathbb{R}^p$-valued process that satisfies the paper's tail and moment conditions, for example a mixing or physical-dependent process, and compute or simulate the Kolmogorov distances $\rho_n$ and $\rho^*_n$; if any such process has $\rho_n + \rho^*_n$ bounded away from zero, or decaying slower than the rate required for $R'_{n,1} \to 0$ at the chosen $U_n$, then the asymptotic symmetrization claim of Proposition 2.2 fails for that process even though its finite-sample inequality is true.

Watch

Extended reading notes

Core claim

The central claim is Proposition 2.2: for dependent $\mathbb{R}^p$-valued random variables with mean zero, any non-decreasing convex $\psi$ with $\psi(0)=0$ satisfies $$E\psi(\max_i|\bar{x}_{i,n}|) \leq \tfrac{1}{2}E\psi\!\left(2\max_i\left|\tfrac{1}{n}\sum_{l=1}^{N_n}\varepsilon_l S_{n,l}(i)\right|\right)+R'_{n,1}+R'_{n,2} \leq \tfrac{1}{2}E\psi(2\max_i|\bar{x}_{i,n}|)+2(R'_{n,1}+R'_{n,2}),$$ where $\varepsilon_l$ are independent block multipliers, $S_{n,l}(i)$ are the block sums, $R'_{n,1}$ is a Gaussian-approximation remainder built from Kolmogorov distances $\rho_n$ and $\rho^*_n$, and $R'_{n,2}$ is a truncation remainder. The paper's point is that the inequalities themselves require almost no structure on the dependence; the dependence enters only through the remainders. When the Gaussian approximation distances converge to zero at suitable rates and the truncation tail vanishes, the classic symmetrization and desymmetrization sandwich holds asymptotically, and under independence with block size one the result reduces exactly to the classical form.

Load-bearing premise

The advertised asymptotic symmetrization and the vanishing remainders require the high-dimensional Gaussian approximation distances $\rho_n$ and $\rho^*_n$ to shrink fast enough relative to the truncation level $U_n$ for the dependent process at hand, and the main text does not prove these Gaussian approximations, deferring them to supplemental material.

Editorial extensions

If this is right

  • For any dependent process admitting a high-dimensional Gaussian approximation with $\rho_n + \rho^*_n \to 0$, the maximal expectation $E\psi(\max_i|\bar{x}_{i,n}|)$ is asymptotically equivalent, up to constants, to the same functional of the block-multiplied sum, so symmetrization and desymmetrization both hold.
  • Theorem 3.1 delivers a Nemirovski-type bound of the form $E\max_i|\bar{x}_{i,n}|^q \leq 2^{q/2}c^q(\ln(2p)/n)^{q/2}E(\max_i(\tfrac{1}{n}\sum_l S_{n,l}(i)^2)^{q/2}) + \text{remainder}$, generalizing a standard independent-data moment inequality to physical dependence.
  • In mixing and physical-dependence settings, the supplemental verification of Gaussian approximations turns the generic inequality into explicit bounds with growth conditions on $p$ relative to $n$.
  • Under independence, choosing block size $b_n=1$ and Rademacher multipliers makes the remainders vanish and recovers the classical symmetrization inequality exactly.
  • The same convexity and scaling argument transfers the result to any Orlicz norm $\|\cdot\|_\psi$, so $L_q$ and exponential-type maximal bounds are covered simultaneously.

Reading between the lines

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

  • The real work of the paper is shifted to the Gaussian approximation distances; a reader applying Theorem 3.1 must first establish $\rho_n + \rho^*_n \to 0$ for their specific process, so the advertised freeing-up of independence is conditional on an external Gaussian approximation input.
  • The inequality suggests a natural block-multiplier bootstrap interpretation: since the original and block-multiplied maxima are interchangeable at the level of convex functionals, valid inference for functions of the maximum could be built from block multipliers whenever the Gaussian comparison holds, a route the paper does not pursue.
  • Because the two remainders move in opposite directions in the truncation level, the paper's analysis exposes an explicit trade-off: truncating more aggressively reduces the tail remainder but inflates the Gaussian-approximation remainder, so the best bound requires a $U_n$ tuned to the process's tail thickness.
  • Sharpening the remainder balance between $U_n$ and $\rho_n+\rho^*_n$, currently optimized only in the sub-exponential example, could lead to near-optimal growth conditions on $p$; the paper leaves cross-coordinate dependence, which would improve the $M_n$ bound, to future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper claims a generic symmetrization inequality for dependent, possibly high-dimensional random vectors: E psi(max_i |bar x_i,n|) is bounded above and below by 1/2 E psi(2 max_i |(1/n) sum_l epsilon_l S_{n,l}(i)|) plus remainder terms, where {epsilon_l} are block-wise independent multipliers and S_{n,l}(i) are block sums. The remainder terms are expressed through high-dimensional Gaussian approximation Kolmogorov distances rho_n and rho*_n and a truncation remainder. The proof strategy is to decompose the sample into expanding blocks, apply Gaussian comparison to the block-multiplied version, and then handle unboundedness by truncation. The paper further applies the result to obtain a maximal L_q-moment inequality in the style of Nemirovski (2000), with Gaussian approximation verification deferred to separate supplemental appendices.

Significance. If the main inequality were correctly established, it would be a genuinely useful finite-sample comparison: it would allow bounding E psi(max_i |bar x_i,n|) by a functional of a block-multiplied version, without requiring independence or Rademacher multipliers. The remainder structure is explicit and the paper identifies the exact quantities (rho_n, rho*_n, truncation level U_n) that drive the approximation. The paper also correctly emphasizes that the finite-sample comparison holds regardless of whether the Gaussian approximations are sharp. However, the central proof as written contains an invalid tail-probability replacement and an unjustified identification of a truncated integral with an untruncated expectation; these errors affect Proposition 2.1, Proposition 2.2, and Theorem 3.1. The asymptotic content additionally rests on Gaussian approximation results that are deferred to other documents, so the advertised convergence of the remainders is not verifiable from the manuscript itself.

major comments (4)
  1. [Appendix A, eq. (A.1) and (A.2)] The proof of Proposition 2.1 uses the Kolmogorov distance bound |F_A(z) - F_B(z)| <= delta to conclude P(A > v) <= P(B <= v) + delta. This is false: the correct implication is P(A > v) <= P(B > v) + delta, since P(A > v) = 1 - F_A(v) and F_A(v) >= F_B(v) - delta. For a concrete failure, take A = B and delta = 0: the displayed inequality would claim P(A > v) <= P(A <= v), which fails whenever v is below the median. The same error appears in the 'reverse' inequality (A.2), where P(B > v) is bounded by P(A <= v) + delta. This is load-bearing: the first and second inequalities of Proposition 2.1, and hence all later results that invoke 'arguments in the proof of Proposition 2.1', are not established.
  2. [Appendix A, eq. (A.1); Section 2.2, display for E_{n,1}] Even after correcting the tail direction, the identification of the integral over [0, sqrt(n) U_n] with the full expectation E psi(max_i |(1/n) sum_l epsilon_l S_{n,l}(i)|) is not valid as stated. For a nonnegative random variable B with support in [0, U], one has integral_0^U P(B <= v) psi'(v/sqrt(n)) dv = sqrt(n)[psi(U/sqrt(n))P(B <= U) - E(psi(B/sqrt(n)) 1_{B <= U})], which is not equal to E psi(B/sqrt(n)). With the corrected tail P(B > v), the integral over [0, sqrt(n) U_n] equals E[psi(B/sqrt(n)) 1_{B <= sqrt(n) U_n}], not the full expectation, unless B <= sqrt(n) U_n almost surely. The paper does not state any boundedness condition on the multipliers epsilon_l in Proposition 2.1 or Proposition 2.2, so the block-multiplied statistic is generally unbounded even when the x_t are bounded. This invalidates the final equality in (A.1) and the corresponding step in the derivation of E_{n,1} in Section 2.2.
  3. [Section 2.2, proof of Proposition 2.2, second inequality] The second inequality of Proposition 2.2, namely 1/2 E psi(2 max_i |(1/n) sum_l epsilon_l S_{n,l}(i)|) <= 1/2 E psi(2 max_i |bar x_i,n|) + 2(R'_n,1 + R'_n,2), is not proved by the sentence 'This, along with a standard desymmetrization argument, proves the main result.' For unbounded x_t, the multiplier statistic B_n = max_i |(1/n) sum_l epsilon_l S_{n,l}(i)| is not a.s. bounded by U_n, so the Gaussian comparison applied on [0, sqrt(n) U_n] only controls the truncated expectation E[psi(2 B_n) 1_{B_n <= U_n}]. The contribution of the tail event {B_n > U_n} must be handled explicitly and bounded by the stated truncation remainder; the manuscript does not supply that argument. Since this second inequality is part of the advertised symmetrization/desymmetrization pair, the gap is load-bearing.
  4. [Section 1, Remark 2.7, and Corollary 3.2] The asymptotic claims that the remainders vanish, and in particular the o(rho_n + rho*_n) and o(1/g_n) statements in Corollary 3.2, depend on the high-dimensional Gaussian approximation bounds for rho_n and rho*_n. These bounds are not proved in the manuscript; they are deferred to Hill [2025b, Appendix B] and Hill [2024b, Appendix B], which are not included in the submission. The finite-sample inequalities are valid for arbitrary rho_n, rho*_n, but the advertised convergence of the remainder terms is a central part of the paper's contribution. The main text should either state the precise conditions and rates used for rho_n and rho*_n or include the relevant arguments, rather than relying on external supplemental documents.
minor comments (4)
  1. [Section 2, definition of Psi] The text says the class Psi consists of non-decreasing convex functions that are continuously differentiable on their support, but the displayed definition only states non-decreasingness and psi(0) = 0. The differentiability condition should be included in the set definition or stated separately.
  2. [Lemma 2.3(a)] The optimized choice lambda = ln(bar P_{U_n}^{-1} ln(p)) requires bar P_{U_n}^{-1} ln(p) > 1; otherwise the logarithm is non-positive and the bound is not meaningful. The statement should include the relevant condition on U_n and p.
  3. [Remark 2.3] The claim that under independence with b_n = 1 the result 'yields classic symmetrization' is imprecise: the displayed inequality has the non-classical factors 1/2 and 2 and is one-sided as a comparison, whereas the classical symmetrization identity is an equality of expectations under Rademacher multipliers. This should be clarified to avoid overstating the connection.
  4. [References] There are two references with nearly identical titles, Hill [2024b] and Hill [2025b], both described as 'Supplemental material for symmetrization for high dimensional dependent random variables'. The relationship between them should be clarified, and the manuscript should indicate which one contains the Gaussian approximation results used here.

Circularity Check

1 steps flagged · score 3.0 of 10

Main comparison inequality has independent content; asymptotic claims lean on self-cited Gaussian approximation appendices, and a separate proof error in Appendix A is a correctness issue, not circularity.

  1. self citation load bearing [Section 2, Remark 2.7; also Section 1 and Supplemental appendix Hill [2025b, Appendix B]]
    "In Hill [2025b, Appendix B] we prove ( ρn, ρ∗ n) → 0 with bounds on p under mixing and physical dependence, and a variety of tail conditions."

    The paper's advertised asymptotic symmetrization — the limiting two-sided inequality with Rn → 0 — depends on the Gaussian approximation Kolmogorov distances ρn and ρ*_n converging to zero. The proof of this convergence is not contained in the text; it is deferred to the author's own separate supplemental appendix (Hill [2025b]) and to Hill [2024b, Appendix B]. Thus the load-bearing premise for the main asymptotic claim is justified only by a self-citation of unpublished or separate work by the same author, without independent verification in the present paper.

full rationale

The core derivations in Section 2 (Propositions 2.1 and 2.2) establish a comparison between Eψ(max_i |bar x_{i,n}|) and the block-multiplier expectation, with an explicit remainder involving the Gaussian approximation distances ρn, ρ*_n and a truncation term. This comparison is not definitionally circular: the target expectation is not defined in terms of the bound's ingredients, and ρn, ρ*_n are independent quantities defined in (3). The application in Theorem 3.1 produces an upper bound on E max_i |bar x_{i,n}|^q through Hoeffding's inequality plus remainders; the truncation remainder R'_{n,2} includes max_i E|bar x_{i,n}|^q, a quantity related to but not identical to the LHS, so the inequality is implicit rather than a tautological reduction. The only circularity-adjacent feature is the self-cited supplemental appendices being used to verify ρn+ρ*_n → 0 in mixing and physical-dependence settings; that is load-bearing for the asymptotic claims but does not make the main inequality equivalent to an input. Separately, the proof of Proposition 2.1 in Appendix A (eq. A.1) appears to contain a technical error: it bounds the upper-tail probability P(max_i |√n bar x_{i,n}| > v) by P(max_i |Σ ε_l S_{n,l}(i)/√n| ≤ v) plus ρn+ρ*_n, whereas the Kolmogorov bound yields P(A>v) ≤ P(B>v)+δ, not P(A>v) ≤ P(B≤v)+δ. This is a correctness concern, not a circularity. Because the advertised remainder-based inequality is meant to hold for arbitrary ρn, ρ*_n, its truth does not reduce to the self-cited Gaussian approximation results. Overall circularity score reflects the moderate self-citation load-bearing for the asymptotic claim while acknowledging that the central comparison has independent content.

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

The paper introduces no new physical or mathematical entities beyond the block-wise multiplier and Gaussian comparison process, which are technical devices. The main result depends on user-chosen block size and truncation sequences, and on Gaussian approximation assumptions that are deferred to unpublished supplements.

free parameters (4)
  • block size b_n
    User-chosen sequence with b_n -> infinity and b_n = o(n); determines the blocks used for the multiplier. The proof requires N_n b_n = n.
  • truncation level U_n
    User-chosen sequence U_n -> infinity; balances the Gaussian approximation error and the truncation error in the unbounded case. The paper suggests an optimal U_n* minimizing the upper bound.
  • moment order r
    Assumed > 1 in Proposition 2.2 and Theorem 3.1; controls the truncation remainder via ||psi(2 max_i |x_{i,t}|)||_r. Not fitted, but arbitrary and influences constants.
  • tail exponent phi
    Any phi in (0, gamma) in the sub-exponential tail bound Lemma 2.3.c; a tuning parameter in the log-exp bound.
assumptions (4)
  • standard math Standard measure-theoretic probability and layer-cake representation for expectations of nonnegative random variables.
    Used to express E psi(max |bar x|) as an integral of tail probabilities in the proof of Proposition 2.1.
  • standard math Hoeffding's inequality for bounded independent random variables applied conditionally on the original sample.
    Used in Section 3 to bound the multiplier sum; cites Buhlmann and van de Geer [2011, Lemma 14.14].
  • domain assumption Existence of high-dimensional Gaussian approximations with Kolmogorov distances rho_n and rho*_n, and their convergence to zero for the asymptotic statements.
    The main inequality is meaningful only if rho_n + rho*_n is small; asymptotic symmetrization requires rho_n, rho*_n -> 0. Verification is deferred to supplements Hill [2025b] and Hill [2024b].
  • domain assumption Moment and tail conditions: E psi(max_i |bar x_{i,n}|) finite; ||psi(2 max_i |x_{i,t}|)||_r < infinity; sub-exponential tail bound (9) in examples.
    Needed for the truncation remainder and for the tail probability bounds in Lemma 2.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symmetrization for high dimensional dependent random variables." pith.science (2026). https://pith.science/paper/7IGBZQUH

@misc{pith2026250600547,
  author       = {Pith},
  title        = {Pith review of: Symmetrization for high dimensional dependent random variables},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IGBZQUH}},
  note         = {Machine review of arXiv:2506.00547}
}
abstract

We establish a generic symmetrization property for dependent random variables $\{x_{t}\}_{t=1}^{n}$ on $\mathbb{R}^{p}$, where $p$ $>>$ $n$ is allowed. We link $\mathbb{E}\psi (\max_{1\leq i\leq p}|1/n\sum_{t=1}^{n}(x_{i,t}$ $-$ $\mathbb{E}x_{i,t})|)$ to $\mathbb{E}\psi (\max_{1\leq i\leq p}|1/n$ $\sum_{t=1}^{n}\eta _{t}(x_{i,t}$ $-$ $\mathbb{E}% x_{i,t})|)$ for non-decreasing convex $\psi $ $:$ $[0,\infty )$ $\rightarrow $ $\mathbb{R}$, where $\{\eta _{t}\}_{t=1}^{n}$ are block-wise independent random variables, with a remainder term based on high dimensional Gaussian approximations that need not hold at a high level. Conventional usage of $% \eta _{t}(x_{i,t}$ $-$ $\tilde{x}_{i,t})$ with $\{\tilde{x}% _{i,t}\}_{t=1}^{n} $ an independent copy of $\{x_{i,t}\}_{t=1}^{n}$, and Rademacher $\eta _{t}$, is not required in a generic environment, although we may trivially replace $\mathbb{E}x_{i,t}$ with $\tilde{x}_{i,t}$. In the latter case with Rademacher $\eta _{t}$ our result reduces to classic symmetrization under independence. We bound and therefore verify the Gaussian approximations in mixing and physical dependence settings, thus bounding $\mathbb{E}\psi (\max_{1\leq i\leq p}|1/n\sum_{t=1}^{n}(x_{i,t}$ $-$ $\mathbb{E}x_{i,t})|)$; and apply the main result to a generic % Nemirovski (2000)-like $\mathcal{L}_{q}$-maximal moment bound for $\mathbb{E}\max_{1\leq i\leq p}|1/n\sum_{t=1}^{n}(x_{i,t}$ $-$ $\mathbb{E}x_{i,t})|^{q}$, $q$ $\geq $ $1$.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    High-dimensional methods and inference on structural and treatment effects

    Belloni, A., Chernozhukov, V., Hansen, C., 2014. High-dimensional methods and inference on structural and treatment effects. J. Econom. Perspect. 28, 29–50

    work page 2014

  2. [2]

    On hoeffding’s inequalities

    Bentkus, V., 2004. On hoeffding’s inequalities. Ann. Probab. 32, 1650–1673

    work page 2004

  3. [3]

    An extension of the hoeffding inequality to unbounded random vari- ables

    Bentkus, V., 2008. An extension of the hoeffding inequality to unbounded random vari- ables. Lith. Math. J. 48, 137–157

    work page 2008

  4. [4]

    Statistics for High-Dimensional Data

    Buhlmann, P., van de Geer, S., 2011. Statistics for High-Dimensional Data. Springer, Berlin

    work page 2011

  5. [5]

    Gaussian approximations and multi- plier bootstrap for maxima of sums of high-dimensional random vectors

    Chernozhukov, V., Chetverikov, D., Kato, K., 2013. Gaussian approximations and multi- plier bootstrap for maxima of sums of high-dimensional random vectors. Ann. Statist. 41, 2786–2819

    work page 2013

  6. [6]

    Inference on causal and structural parameters using many moment inequalities

    Chernozhukov, V., Chetverikov, D., Kato, K., 2019. Inference on causal and structural parameters using many moment inequalities. Rev. Econ. Stud. 86, 1867–1900

    work page 2019

  7. [7]

    Bounds on the Expectation of a Convex Functions

    Edmundson, H.P., 1956. Bounds on the Expectation of a Convex Functions. Technical Report 982. Rand Corp.. Santa Monica. 16

    work page 1956

  8. [8]

    Statistical challenges with high dimensionality: Feature selection in knowledge discovery, in: Sanz-Sole, M., Soria, J., Varona, J.L., Verdera, J

    Fan, J., Li, R., 2006. Statistical challenges with high dimensionality: Feature selection in knowledge discovery, in: Sanz-Sole, M., Soria, J., Varona, J.L., Verdera, J. (Eds.), Proceedings of the International Congress of Mathematicians, European Mathematical

    work page 2006

Show all 22 references
  1. [9]

    Sparse high-dimensional models in economics

    Fan, J., Lv, J., Qi, .L., 2011. Sparse high-dimensional models in economics. Annu. Rev. Economics 3, 291–317

    2011

  2. [10]

    Strong laws for dependent heterogeneous processes

    Hansen, B.E., 1991. Strong laws for dependent heterogeneous processes. Econometric Theory 7, 213–221

    1991

  3. [11]

    Erratum: Strong laws for dependent heterogeneous processes

    Hansen, B.E., 1992. Erratum: Strong laws for dependent heterogeneous processes. Econo- metric Theory 8, 421–422

    1992

  4. [12]

    symmetrization for high dimensional depen- dent random variables

    Hill, J.B., 2025b. Supplemental material for “symmetrization for high dimensional depen- dent random variables”. Dept. of Economics, University of North Carolina - Chapel Hill. K¨ unsch, H.R., 1989. The jackknife and the bootstrap for general stationary observations. Ann. Stat...

    1989

  5. [13]

    Bootstrap procedures under some non-i.i.d

    Liu, R.Y., 1988. Bootstrap procedures under some non-i.i.d. models. Ann. Statist. 16, 1696–1708

    1988

  6. [14]

    Bounds on the expectation of a convex function of a multivariate random variable

    Madansky, A., 1959. Bounds on the expectation of a convex function of a multivariate random variable. Ann. Math. Statist. 30, 743–746

    1959

  7. [15]

    Bernstein inequality and moderate deviations for weakly dependent sequences

    Merlevede, F., Peligrad, M., Rio, E., 2011. Bernstein inequality and moderate deviations for weakly dependent sequences. Probab. Theory Rel. 151, 435–474. Volume 5

    2011

  8. [16]

    Topics in nonparametric statistics, in: Emery, M., Nemirovski, A., Voiculescu, D., Bernard, P

    Nemirovski, A.S., 2000. Topics in nonparametric statistics, in: Emery, M., Nemirovski, A., Voiculescu, D., Bernard, P. (Eds.), Lectures on Probability Theory and Statistics: Ecole d’Ete de Probabilites de Saint-Flour XXVIII - 1998. Springer, New York. volume 1738, pp. 87–285

    2000

  9. [17]

    The stationary bootstrap

    Politis, D.N., Romano, J.P., 1994. The stationary bootstrap. J. Amer. Statis. Assoc. 89, 1303–1313

    1994

  10. [18]

    Convergence of Stochastic Processes

    Pollard, D., 1984. Convergence of Stochastic Processes. Springer Verlag, New York

    1984

  11. [19]

    A bootstrap-assisted spectral test of white noise under unknown depen- dence

    Shao, X., 2011. A bootstrap-assisted spectral test of white noise under unknown depen- dence. Journal of Econometrics 162, 213–224. van der Vaart, A., Wellner, J., 1996. Weak Convergence and Empirical Processes. Springer, New York

    2011

  12. [20]

    Nonlinear system theory: Another look at dependence

    Wu, W.B., 2005. Nonlinear system theory: Another look at dependence. Proc. Natl. Acad. Sci. 102, 14150–14154

    2005

  13. [21]

    On linear processes with dependent innovations

    Wu, W.B., Min, M., 2005. On linear processes with dependent innovations. Stochastic Process. Appl. 115, 939–958. 17

    2005

  14. [22]

    Gaussian approximation for high dimensional vector under physical dependence

    Zhang, X., Cheng, G., 2018. Gaussian approximation for high dimensional vector under physical dependence. Bernoulli 24, 2640–2675. 18

    2018

Pith tools

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