Pith. sign in

REVIEW 3 major objections 4 minor 28 references

Transform-Resampled Double Bootstrap Percentile with Applications in System Reliability Assessment

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 A transformed-resampling double bootstrap yields O(n^{-1})-accurate system reliability confidence limits at a fraction of the usual computational cost.

desk verdict A genuinely useful computational trick for double-bootstrap reliability bounds, with numerical evidence that outruns its proof; the O(n^{-1}) coverage claim needs work before it is quoted. read the letter →

arxiv 2506.04573 v1 pith:N5ZLXZQT submitted 2025-06-05 stat.ME

Junpeng Gong , Xu He , Zhaohui Li This is my paper
classification stat.ME MSC 62N0562F4062E2062G09
keywords systemreliabilityassessmentdoublebootstrappercentilelog-location-scalefamilymoment-basedestimationcoverageaccuracybend-backconfidencelimittransformedresamples
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

This paper proposes a way to build lower confidence limits for the reliability of a coherent system from component-level failure data, using a double bootstrap that is both fast and second-order accurate. The key move is to generate bootstrap samples of component reliability by transforming auxiliary statistics (the mean and standard deviation of standardized log-lifetimes) rather than by resampling raw lifetime data and re-estimating parameters. This works because, for log-location-scale lifetime distributions, the moment-based reliability estimator depends on the unknown parameters only through the true reliability, so one set of transformed resamples can serve for all first-layer bootstrap replicates. The paper proves that the resulting limit stays in [0,1] with coverage 1-$\alpha$+O($n^{{-1}}$), matching the conventional double bootstrap, while reducing computation from O(snBC) to O(s(B+C)+sBC).

What carries the argument

The central object is the pivotal identity of Theorem 3: for log-location-scale lifetimes with standardized CDF F_i, the moment-based reliability estimator satisfies \hat r_i(t) \sim 1 - F_i[ \{ $F_i^{{-1}}$(1-r_i(t)) - \bar Z_{n_i} \} \kappa_{i2}/M_{n_i} + \kappa_{i1} ], where \bar Z and M are the sample mean and sample standard deviation of n_i standardized log-lifetimes. This identity converts bootstrap resampling into a pure transformation problem: each bootstrap replicate of component reliability is obtained by drawing the auxiliary statistics from their known distributions (e.g., t and chi-square for log-normal, gamma for exponential, or i.i.d. draws from F_i in the general case) and applying the formula. The second layer reuses one set of auxiliary statistics across all first-layer replicates, which is what cuts the cost, while the prepivoting argument of the double bootstrap (via Beran's result) supplies the O($n^{{-1}}$) coverage improvement.

What would settle it

Simulate a coherent system (say a series system of three Weibull components) with the DBPT algorithm, but draw component lifetimes from a distribution outside the log-location-scale family, such as a gamma or generalized-gamma distribution with the same moments; if the empirical coverage of the 90% lower limit does not approach the nominal level at the O($n^{{-1}}$) rate as n grows (e.g., coverage error stays at O($n^{{-1/2}}$) or systematically deviates), the pivotal-transformation premise is refuted. A more direct check is to verify numerically whether the moment estimator's Studentized statistic admits a valid Edgeworth expansion under the stated moment conditions.

Watch

Extended reading notes

Core claim

The central claim is Theorem 4: for a coherent system whose component lifetimes follow continuous log-location-scale distributions with finite fourth moments, the double bootstrap percentile with transformed resamples (Algorithm 2) produces a lower confidence limit R_DBPT(t) within [0,1] whose coverage is P(R(t) >= R_DBPT(t)) = 1 - $\alpha$ + O($n^{{-1}}$), the same order as the conventional double bootstrap percentile. The discovery that makes this possible is the pivotal representation of the moment-based estimator: \hat{r}_i(t) depends on the data only through the sample mean and variance of standardized log-lifetimes, with the unknown parameters entering only through the true component reliability r_i(t) (Theorem 3). This lets the second bootstrap layer be generated by drawing the auxiliary statistics (\bar{Z}, M) and plugging them into the known transformation, instead of resampling raw component data and re-running estimation. The paper also claims that in simulations the method matches or beats the delta method, the R-WCF expansion, and the ordinary bootstrap percentile in coverage, while almost never producing limits outside [0,1] and rarely showing bend-back, and that it handles Type-II censored data by imputation.

Load-bearing premise

The load-bearing premise is that each component's lifetime follows a known log-location-scale distribution, so the transformed resamples faithfully reproduce the bootstrap distribution; if that family is misspecified, or the assumed Edgeworth regularity conditions fail, the O($n^{{-1}}$) coverage guarantee is not established.

Editorial extensions

If this is right

  • The DBPT lower confidence limit always lies in [0,1] and is monotonically decreasing in mission time t, so it avoids the 'falling outside' problem that plagues delta and Cornish-Fisher methods.
  • For sample sizes n <= 100, the recommended B=1000, C=500 settings give empirical coverage that tracks the nominal level even at true reliabilities above 0.999, where competing methods drift badly.
  • Computational cost drops from O(snBC) to O(s(B+C)+sBC); a 9-out-of-16 system example ran in 0.2 hours versus 78.8 hours for the conventional double bootstrap, with limits agreeing to two decimal places.
  • The same acceleration principle transfers to censored data when the data are first completed by imputation, preserving the coverage advantage.

Reading between the lines

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

  • Because the pivotal identity also holds for unequal component sample sizes without modification, the method should extend naturally to heterogeneous data collection schemes, a setting where higher-order asymptotic formulas are even harder to derive.
  • The same 'transform, don't resample' idea could be applied to other pivotal estimators in log-location-scale families, e.g., quantile estimation or accelerated life testing with covariates, whenever a moment-based statistic has a parameter-free representation.
  • The near-elimination (not elimination) of bend-back suggests a small residual nonmonotonicity that may be removable by a post-hoc isotonic correction, though the paper does not propose one.
  • If moment-based estimators are replaced by maximum likelihood, the pivotal property fails; so the method's efficiency gain is tied to accepting the slightly less efficient moment estimator in exchange for the closed-form bootstrap.
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

3 major / 4 minor

Summary. This paper proposes a computationally efficient double-bootstrap percentile method for system reliability assessment when component lifetimes follow a log-location-scale distribution. The key idea is to use moment-based estimators of component reliability and to exploit Theorem 3, which represents the sampling distribution of the moment estimator as a function of a single parameter r_i(t) and known auxiliary statistics (\bar Z, M). Algorithm 2 then generates first- and second-layer bootstrap reliability estimates by transforming independently generated auxiliary statistics, avoiding nested resampling of raw component data. The paper claims in Theorem 4 that the resulting lower confidence limit has coverage error O(n^{-1}), matching the conventional double bootstrap, at computational cost O(s(B+C)+sBC) instead of O(snBC). Numerical experiments on series, parallel, 2x2 series-parallel, and k-out-of-n systems, including censored data, show improved coverage and reduced bend-back behavior relative to the delta, R-WCF, and bootstrap percentile methods.

Significance. If the O(n^{-1}) claim is established, the paper makes a useful contribution: it provides an automated, non-analytic route to second-order accurate confidence limits for general coherent systems, while keeping lower confidence limits in the valid range and largely avoiding the bend-back problem. The transformation and reuse of auxiliary statistics is elegant, the use of moment estimators is well suited to the log-location-scale family, and the proof strategy via Newton-Geyer's lemma is appropriate. The paper also includes extensive simulations across system structures, sample sizes, and censoring scenarios, as well as a concrete runtime comparison. However, the central theoretical guarantee is not fully supported as written, and the theorem statement needs to be made precise before the claimed accuracy result can be accepted.

major comments (3)
  1. [Appendix A, proof of Theorem 4 (final paragraph)] The O(n^{-1}) coverage claim is not established because Beran's prepivoting theorem is invoked without verifying its hypotheses. The proof shows only that, for fixed n and fixed data, the Monte Carlo distribution \tilde G converges almost surely to G as B,C tend to infinity. To conclude C(x)=1-G(x)+O(n^{-1}), one needs the prepivot U(R) to have a valid Edgeworth expansion and the bootstrap estimator of its distribution to be second-order accurate. The assumptions stated in Theorem 4 (continuity and finite fourth moments) do not imply these conditions; the earlier argument in the proof of Theorem 1, which derives Edgeworth regularity from absolute continuity via the Riemann-Lebesgue lemma and finite fourth moments, is not sufficient and should be replaced by explicit regularity conditions, such as Cram\'er's condition and suitable smoothness of the polynomial map \psi and of the moment-based estimator.
  2. [Section 3.4, Theorem 4 and its proof] The order of limits is ambiguous. The theorem is stated as B,C,n tending to infinity jointly, but the proof first sends B,C to infinity at fixed n and then lets n tend to infinity. No rate conditions on B and C are given, so the O(n^{-1}) coverage error is not uniform in B,C. For the recommended values B=10^3 and C=500, the Monte Carlo error in estimating G^{-1}(\alpha) is not shown to be o(n^{-1}), and hence Eq. (9) is not a statement about the implemented algorithm. The proof also passes from pointwise convergence \tilde G(x) to G(x) to convergence of the quantile \hat u_{(k)} to G^{-1}(\alpha) without assuming that G is continuous and strictly increasing at \alpha; this should be stated as an additional condition or proved.
  3. [Section 3.4, Theorem 4 statement] The statement of Theorem 4 omits assumptions that make Algorithm 2 well defined. It assumes only that the component lifetime distributions are continuous with finite fourth moments, but Algorithm 2 and Eq. (8) require that each component follows a log-location-scale distribution with known standardized CDF F_i, or a known special case such as Weibull, log-normal, or exponential. Without these assumptions, the transformed resamples do not simulate the relevant bootstrap distribution, and the claimed coverage guarantee does not apply. The theorem should state these assumptions explicitly rather than leaving them implicit in the surrounding text.
minor comments (4)
  1. [Appendix A, proof of Theorem 3] In the proof of Theorem 3, \bar Z_{n_i} is written as a sum rather than an average, which conflicts with its definition in the theorem statement as the sample mean; please correct this notation.
  2. [Algorithm 2, line 24] The order statistic notation \hat u_{(k)} is introduced only in the proof; the algorithm should define it explicitly and should also handle the boundary case \hat u_{(k)}=0, where k'=0 would make R^*_{(0)} undefined.
  3. [Section 4.2] The censoring indicator \delta_i is assigned 0 for observed failures and 1 for censored observations, which is the opposite of the usual convention; please redefine it to avoid confusion.
  4. [Figure 6] The x-axis labels are run together (for example, '0.57210.6984'); add separators so the true reliability values are readable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DBPT coverage claim is derived from a proven pivotal representation plus external prepivoting theory, so no prediction is equivalent to its input by construction.

full rationale

The central claim (Theorem 4, Eq. 9) is not defined in terms of its own output. The construction rests on Theorem 3, whose single-parameter pivotal representation (Eq. 8) is proved in the appendix by standardizing log-location-scale samples; it is not imported from the cited Li et al. (2023, 2016) papers, which are credited only as background for moment estimation. Algorithm 2 generates first- and second-layer estimates through the exact functional form of Eq. 8, so the transformed resamples are a computational shortcut for the nested bootstrap rather than a fitted input dressed as a prediction. The O(n^{-1}) coverage statement is obtained by applying Beran's prepivoting result and Newton-Geyer's ergodic lemma, both external to the authors, to the empirical distribution of the adjusted levels u-hat_j; the paper's proof of almost-sure convergence of G-tilde(x) to G(x) is a standard Monte Carlo argument. Self-citations to Li et al. (2020) appear only as a numerical baseline, not as a premise. The acknowledged limitation that bend-back is not completely eliminated (Section 5) and the reviewer-identified gaps about Edgeworth regularity and the B, C, n order of limits are correctness concerns about whether Beran's external theorem applies; they do not make the derivation circular. No fitted constant, calibration, or author-specific uniqueness claim is load-bearing.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on three scientific assumptions: (1) the log-location-scale model with known F_i, (2) the regularity conditions for Edgeworth expansions of the moment estimator, and (3) the external prepivoting theory. No invented entities or fitted parameters appear; the only tuning constants are the resample sizes B and C.

free parameters (2)
  • B (first-layer resamples) = 1000
    User-set resampling count recommended as B=10^3 in Section 3.4; chosen by simulation heuristics, not estimated from data.
  • C (second-layer resamples) = 500
    User-set resampling count recommended as C=500; chosen by simulation heuristics.
assumptions (5)
  • domain assumption Component lifetimes follow a log-location-scale distribution with known standardized CDF F_i
    Stated in Section 3.3; eq. (5) and (8) rely on this for the pivotal property of the moment estimator.
  • domain assumption The system is coherent with independent components and known structure function psi
    Section 2.1 defines the system model; Lemma 1 uses the polynomial representation of R(t) in r_i.
  • standard math The prepivoting identity C(x)=1-G(x)+O(n^-1) from Beran (1987) holds for the moment-based estimator
    Invoked in the proof of Theorem 4 to derive the O(n^-1) coverage; the required Edgeworth regularity conditions are assumed (absolute continuity, finite fourth moments) but not explicitly verified.
  • standard math Lemma 2 of Newton and Geyer (1994) applies to the Monte Carlo approximation of uhat_j
    Used in the proof of Theorem 4 to justify almost sure convergence of the second-layer Monte Carlo estimates.
  • domain assumption The standardized CDF F_i is known exactly for each component
    Required by Algorithm 1 and equations (8) to generate auxiliary statistics; misspecification would invalidate the transformation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transform-Resampled Double Bootstrap Percentile with Applications in System Reliability Assessment." pith.science (2026). https://pith.science/paper/N5ZLXZQT

@misc{pith2026250604573,
  author       = {Pith},
  title        = {Pith review of: Transform-Resampled Double Bootstrap Percentile with Applications in System Reliability Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N5ZLXZQT}},
  note         = {Machine review of arXiv:2506.04573}
}
read the original abstract

System reliability assessment(SRA) is a challenging task due to the limited experimental data and the complex nature of the system structures. Despite a long history dating back to \cite{buehler1957confidence}, exact methods have only been applied to SRA for simple systems. High-order asymptotic methods, such as the Cornish-Fisher expansion, have become popular for balancing computational efficiency with improved accuracy when data are limited, but frequently encounter the "bend-back" problem in high-reliability scenarios and require complex analytical computations. To overcome these limitations, we propose a novel method for SRA by modifying the double bootstrap framework, termed the double bootstrap percentile with transformed resamples. In particular, we design a nested resampling process for log-location-scale lifetime models, eliminating the computational burden caused by the iterative resampling process involved in the conventional double bootstrap. We prove that the proposed method maintains the high-order convergence property, thus providing a highly accurate yet computationally efficient confidence limit for system reliability. Moreover, the proposed procedure is straightforward to implement, involving only a simple resampling operation and efficient moment estimation steps. Numerical studies further demonstrate that our approach outperforms the state-of-the-art SRA methods and, at the same time, is much less susceptible to the bend-back issue.

Figures

Figures reproduced from arXiv: 2506.04573 by the authors.

Figure 1
Figure 1. System diagrams of a s-series (top left), s-parallel (top-right) and 2 × 2 series￾parallel (bottom) system. at t, and the system structure function ψ can be determined by incorporating mechanical domain knowledge and engineering features. For instance, some system structures can be represented by the system diagram shown in [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. The real reliability (black solid line), ML estima [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the conventional double bootstra [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The empirical coverage probability (top) and the 9 [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: The empirical coverage probability (top) and the 9 [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]
Figure 6
Figure 6. Figure 6: The empirical coverage rate versus the true reliab [PITH_FULL_IMAGE:figures/full_fig_p027_6.png]
Figure 7
Figure 7. Figure 7: The empirical coverage probability (left) and the [PITH_FULL_IMAGE:figures/full_fig_p029_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages

  1. [1]

    Beran, R. (1987). Prepivoting to reduce level error of confidence sets. Biometrika , 74(3):457--468

    work page 1987

  2. [2]

    and Presnell, B

    Booth, J. and Presnell, B. (1998). Allocation of Monte C arlo resources for the iterated bootstrap. Journal of Computational and Graphical Statistics , 7(1):92--112

    work page 1998

  3. [3]

    Booth, J. G. and Hall, P. (1994). Monte C arlo approximation and the iterated bootstrap. Biometrika , 81(2):331--340

    work page 1994

  4. [4]

    Buehler, R. J. (1957). Confidence intervals for the product of two binomial parameters. Journal of the American Statistical Association , 52(280):482--493

    work page 1957

  5. [5]

    Byun, J.-E., Noh, H.-M., and Song, J. (2017). Reliability growth analysis of k-out-of-n systems using matrix-based system reliability method. Reliability Engineering & System Safety , 165:410--421

    work page 2017

  6. [6]

    Cai, X., Tian, Y., Xu, H., and Wang, J. (2017). WCF approach of reliability assessment for solid state power controller with accelerate degradation data. Communications in Statistics-Simulation and Computation , 46(1):458--468

    work page 2017

  7. [7]

    and Hall, P

    Chang, J. and Hall, P. (2015). Double-bootstrap methods that use a single double-bootstrap simulation . Biometrika , 102(1):203--214

    work page 2015

  8. [8]

    Davison, A. C. and Hinkley, D. V. (1997). Bootstrap methods and their application . Number 1. Cambridge university press

    1997

Show all 28 references
  1. [9]

    DiCiccio, T. J. and Efron, B. (1996). Bootstrap confidence intervals. Statistical science , 11(3):189--228

    1996

  2. [10]

    Du, S., Li, Z., Yu, D., Li, D., and Hu, Q. (2020). Exact confidence limit for complex system reliability based on component test data. Quality Technology & Quantitative Management , 17(1):75--88

    2020

  3. [11]

    and Tibshirani, R

    Efron, B. and Tibshirani, R. J. (1994). An introduction to the bootstrap . Chapman and Hall/CRC

    1994

  4. [12]

    Hall, P. (1986). On the bootstrap and confidence intervals. The Annals of Statistics , 14(4):1431--1452

    1986

  5. [13]

    Henley, E. J. and Kumamoto, H. (1981). Reliability Engineering and Risk Assessment . Prentice Hall

    1981

  6. [14]

    and Meeker, W

    Hong, Y. and Meeker, W. Q. (2014). Confidence interval procedures for system reliability and applications to competing risks models. Lifetime data analysis , 20(2):161--184

    2014

  7. [15]

    Q., and Escobar, L

    Hong, Y., Meeker, W. Q., and Escobar, L. A. (2008). Avoiding problems with normal approximation confidence intervals for probabilities. Technometrics , 50(1):64--68

    2008

  8. [16]

    Li, M., Zhang, W., Hu, Q., Guo, H., and Liu, J. (2016). Design and risk evaluation of reliability demonstration test for hierarchical systems with multilevel information aggregation. IEEE Transactions on Reliability , 66(1):135--147

    2016

  9. [17]

    Li, Z., Ng, H. K. T., Yu, D., and Hu, Q. (2023). Optimal acceptance sampling testing plan with pivotal quantity for log-location-scale distributions. IEEE Transactions on Reliability , 72(4):1453--1465

    2023

  10. [18]

    Li, Z., Yu, D., Liu, J., and Hu, Q. (2020). Higher-order normal approximation approach for highly reliable system assessment. IISE Transactions , 52(5):555--567

    2020

  11. [19]

    E., Glen, A

    Marks, C. E., Glen, A. G., Robinson, M. W., and Leemis, L. M. (2014). Applying bootstrap methods to system reliability. The American Statistician , 68(3):174--182

    2014

  12. [20]

    Martin, M. A. (1990). On bootstrap iteration for coverage correction in confidence intervals. Journal of the American Statistical Association , 85(412):1105--1118

    1990

  13. [21]

    Q., Escobar, L

    Meeker, W. Q., Escobar, L. A., and Pascual, F. G. (2022). Statistical methods for reliability data . John Wiley & Sons

    2022

  14. [22]

    Newton, M. A. and Geyer, C. J. (1994). Bootstrap recycling: A Monte C arlo alternative to the nested bootstrap. Journal of the American Statistical Association , 89(427):905--912

    1994

  15. [23]

    and Weissfeld, L

    Schneider, H. and Weissfeld, L. (1986). Inference based on type II censored samples. Biometrics , 42(3):531--536

    1986

  16. [24]

    Shao, J. (2008). Mathematical statistics . Springer Science & Business Media

    2008

  17. [25]

    Winterbottom, A. (1980). Asymptotic expansions to improve large sample confidence intervals for system reliability. Biometrika , 67(2):351--357

    1980

  18. [26]

    Xiao, X., Hu, Q., Yu, D., and Xie, M. (2014). Study of an imputation algorithm for the analysis of interval-censored data. Journal of Statistical Computation and Simulation , 84(3):477--490

    2014

  19. [27]

    and Ding, J

    Xiong, W. and Ding, J. (2017). Confidence intervals for a population proportion using group testing in the presence of misclassification. Communications in Statistics-Simulation and Computation , 46(1):174--183

    2017

  20. [28]

    Yu, D., Li, X., Jiang, N., and Yang, J. (2007). Some statistical inference problems and research progresses on reliability analysis of complex systems. Journal of System Science and Mathematics Science , 27(1):68--81

    2007

Pith tools

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