Pith. sign in

REVIEW 3 major objections 5 minor 34 references

For unbounded integrands on unbounded domains, RQMC-based self-normalized importance sampling achieves Lp error of order N^{-β+ε}, with β arbitrarily close to 1.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 22:23 UTC pith:UKPPWJXS

load-bearing objection Plain RQMC Lp theorem is solid and novel, but the SNIS corollaries mis-state the tail exponent, so the headline SNIS rate is not supported as written. the 3 major comments →

arxiv 2511.10599 v3 pith:UKPPWJXS submitted 2025-11-13 math.NA cs.NA

The L_p-error rate for randomized quasi-Monte Carlo self-normalized importance sampling of unbounded integrands

Jiarui Du , Zhijian He This is my paper
classification math.NA cs.NA MSC 65C0565D30
keywords randomized quasi-Monte Carloself-normalized importance samplingLp error rateunbounded integrandstransport mapsprojection operatorsub-Gaussian tailsBayesian computation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Self-normalized importance sampling (SNIS) is the standard way to estimate posterior expectations when the normalizing constant is unknown, but until now its error under randomized quasi-Monte Carlo (RQMC) had only been controlled for bounded integrands or bounded domains. This paper claims that for unbounded test functions on R^d, the Lp error (p≥1) of the RQMC-SNIS estimator is O(N^{-β+ε}), where β is a positive exponent determined by the growth rate of the product ωf and the sub-Gaussian tail exponent of the sampling distribution. When that product grows slowly—the 'QMC-friendly' case—β is arbitrarily close to 1, giving nearly O(1/N) convergence. That would beat Monte Carlo's O(N^{-1/2}) rate, so a Bayesian practitioner could get far more accurate posterior estimates from the same sample budget. The paper also proves a plain-RQMC Lp rate for a broad class of transport maps, which is the engine behind the SNIS result.

Core claim

The central claim is Theorem 3: under Assumptions 1–3 and a moment condition q(|f|^v)<∞, for RQMC point sets with star discrepancy O(N^{-1}(log N)^{d-1}), the Lp error satisfies {E|π_N(f)−π(f)|^p}^{1/p}=O(N^{-β+ε}) with β=1−pM_{ωf,τ}/α>0, provided p<min{v, α/M_{ωf,τ}}. Here M_{ωf,τ} is the exponential-quadratic growth rate of the composed integrand ωf after the transport map, and α is the sub-Gaussian tail rate of the proposal. The proof splits the estimation error into a 'good' event where the estimated normalizing constant is bounded away from zero and a rare 'bad' event, and it controls both by the plain-RQMC Lp bounds developed earlier. If the product ωf is QMC-friendly, M_{ωf,τ} is arbi

What carries the argument

At the center of the proof is a new radially symmetric smoothed projection P_r that agrees with the identity inside a ball of radius (1−δ)r, vanishes outside radius r, and interpolates through a C^d bump built from a Beta CDF; it is engineered so mixed partial derivatives of P_r are bounded by constants independent of dimension on the transition shell, and ||P_r(x)||≤||x||. Replacing the unbounded integrand f by f∘P_r makes the QMC discrepancy term tractable via the Koksma–Hlawka inequality, while the bias created by truncation is controlled by the sub-Gaussian tail assumption. The proof balances these two terms by taking the projection radius r proportional to sqrt(log N), which is what pro

Load-bearing premise

The load-bearing premise is that the sub-Gaussian tail exponent α used in Theorem 3 is the one that controls the plain-RQMC projection-error term; the proof of the general-transport step applies the plain-RQMC theorem to f∘τ with the base map G^{-1}, whose tail exponent need not equal that of the full transport map T=τ∘G^{-1}.

What would settle it

Run RQMC-SNIS on a one-dimensional integral with base map G^{-1}=Φ^{-1} (standard normal), proposal dilation τ(x)=cx with c>1, and a QMC-friendly f such as f(x)=x^2. Record the empirical Lp error exponent as N grows and compare it with Theorem 3's prediction β=1−pM_{ωf,τ}/α_q, where α_q=1/(2c^2), versus the alternative using the base exponent α_G=1/2. The two predictions diverge by a factor c^2 in the M term; whichever matches settles which tail exponent enters the rate.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • When ωf is QMC-friendly, the Lp rate is O(N^{-1+ε}) for any p<v, versus O(N^{-1/2}) for IID Monte Carlo.
  • Setting p=1 gives a bias rate and p=2 gives root mean square error, so one theorem covers both.
  • The plain-RQMC Lp bound applies to shifted, scaled, componentwise nonlinear, and composed transport maps (including common neural-network activation maps), not just the inverse-CDF normal map.
  • The derived skewness and kurtosis bounds imply the Student-t confidence interval for RQMC has coverage error nearly O(R^{-1/2}) after R independent replicates.
  • The rate degrades polynomially as the integrand growth M_{ωf,τ} approaches α/p, so the admissible moment order p is traded against integrand growth.

Where Pith is reading between the lines

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

  • One testable consequence left implicit: proposals with light tails (large α) and small linear-dilation constants C_τ should give the best observed Lp exponents; the formula suggests C_τ^2 enters the effective growth rate while α scales roughly inversely with C_τ^2, so shrinking the proposal's spread may not help as much as shrinking C_τ itself.
  • If the tail-exponent mismatch flagged below is real, the theorem's stated α should be replaced by the tail exponent of the base map G^{-1}(U), and experiments with a dilated Gaussian proposal (where the two exponents differ) would reveal which exponent actually governs convergence.
  • The projection radius r∝sqrt(log N) suggests finite-sample log-factor effects that the ε absorbs asymptotically; the paper's own high-dimensional experiments show these factors hurt for moderate N, so a dimension-robust variant (e.g., boundary damping) is a natural next test.
  • The framework only covers integrands with exponential-quadratic growth; an extension to heavier tails (polynomial growth of f) would require a different projection analysis, and the same machinery might be adapted by choosing a slower-growing truncation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Lp error rates (p≥1) for randomized quasi-Monte Carlo integration and for RQMC-based self-normalized importance sampling with unbounded integrands on unbounded domains. A new radial projection operator is introduced and used to prove a plain RQMC rate O(N^{-β+ε}) for functions with at most exp(M||x||^2) growth under a sub-Gaussian tail assumption on the sampling map (Theorem 1, Corollary 1). The paper then treats composite transport maps T=τ∘G^{-1} by an auxiliary composition-growth estimate (Theorem 2, Corollary 2) and applies the machinery to the SNIS estimators q_N(ωf) and q_N(ω), obtaining the headline rate for π_N(f) (Corollary 3, Theorem 3). Numerical experiments on a Bayesian inverse problem and Bayesian logistic regression are reported.

Significance. If sound, Theorem 3 would fill a genuine gap: prior RQMC-SNIS analyses for unbounded integrands on R^d were largely unavailable, and a near O(N^{-1}) Lp rate would improve on the classical O(N^{-1/2}) Monte Carlo rate. The paper's plain RQMC analysis—especially the dimension-independent radial projection operator and the transparent treatment of the projection, variation, and tail terms—is a useful contribution in its own right. The proofs are detailed and the numerical experiments probe the predicted rates. However, the central SNIS statements are currently not supported by the proofs because of the tail-exponent mismatch described in the major comments; the advertised rates and conditions need to be corrected before the contribution can be accepted.

major comments (3)
  1. [§3 Corollary 2 and §4 Corollaries 3/Theorem 3] The statements place Assumption 2 on the full map T=τ∘G^{-1}, but the proof applies Corollary 1 to fτ (resp. ωf∘τ) with the base inversion map G^{-1}. The projection-error Lemma 3.5 is controlled by the tail of Y=G^{-1}(U), not by that of T(U); Assumption 3 gives only an upper bound on ∥τ∥ and no lower bound, so a contractive τ can make the tail of T much lighter than that of Y. Example: d=1, G=Φ, τ(y)=0.1y, f(x)=e^{102x^2}. Then T(U)∼N(0,10^{-2}) has α=50, fτ(y)=e^{1.02y^2} gives M_{f,τ}=1.02, and p=1 satisfies M_{f,τ}<α; yet π(f)=E[e^{1.02Y^2}]=∞. Hence the stated β and the condition p<α/M are not valid as written; α must be the tail exponent of Y (or a lower-bound/non-contraction condition on τ must be added).
  2. [Theorem 2, Eq. (15)] The formula for M_{f,τ} is asserted for every M_f∈R. The proof uses exp(M_f∥τ(x)∥²)≤exp(M_fCτ²(1+ε)∥x∥²), which is false when M_f<0 because multiplying by M_f reverses the inequality. Thus a negative value of (15) is not generally an admissible Assumption 1 rate for fτ. The correct statement should use max(M_f,0)Cτ²(1+ε)+d max(Mτ,0). This is local for Theorem 3 (which assumes M_f>0, M_{ω,τ}>0), but Theorem 2 and Corollary 2 need the amendment.
  3. [§4 Theorem 3] 'M_{ω,τ}>0 is arbitrarily small' is not a precise hypothesis. In the proof, after p and v are fixed, the constant s must satisfy (22) and s<α/M_{ω,τ}; the second inequality forces M_{ω,τ} to be smaller than a bound depending on p,v,ε. The theorem and proof should state explicit quantifiers (e.g., for every δ>0, assume 0<M_{ω,τ}<δ, with constants depending on p,v,ε,δ). As written, the Markov bound O(N^{-(1-ε)s}) in the first term of (21) is not justified uniformly.
minor comments (5)
  1. [Table 1] Softplus: log(1+e^x) grows linearly, so CT should be 1, not 0; as printed, Table 1's claim that softplus has M_{f,τ}=0 is incorrect (though Mτ=0 remains true).
  2. [Theorem 3 proof] In the first term of the proof, 'max_{1≤i≤N} f(x_i)' should be 'max_{1≤i≤N}|f(x_i)|' before applying Hölder with q(|f|^v).
  3. [Remark 6] Typo: M_{f,τ}=M_T M_L(1+ε) should read M_f M_L(1+ε).
  4. [Abstract] 'L1-error (bias)' is imprecise; the L1 error is a mean absolute error, not the bias.
  5. [§5.1] The statement 'M_q>0 is arbitrarily small' for t proposals should be quantified ('for any ε>0, M_q≤ε'), matching the epsilon-delta style used elsewhere.

Circularity Check

0 steps flagged

No substantive circularity: the Lp rates are derived from explicit growth/tail assumptions, and the self-citations are supporting lemmas, not recycled predictions.

full rationale

The central chain is Theorem 1/Corollary 1 (plain RQMC) from Assumptions 1, 2 and the projection operator, then Theorem 2 (composition growth rate), Corollary 3 (RQMC-IS), and Theorem 3 (RQMC-SNIS) via the [7] decomposition. Each rate beta = 1 - pM/alpha is calculated from the assumed growth rate and tail exponent; no fitted constant is later renamed as a prediction. The paper cites [20] for Lemma A.2 and the projection idea, and [13,14] for auxiliary bounds, but these are prior lemmas/framework, not a uniqueness claim or an assumption of the target rate, so they do not make the argument circular. Lemma 3.1 is justified by a self-citation, but it is an elementary integral inequality and is not load-bearing in a circular sense. The notable weakness is a condition gap rather than circularity: Corollaries 2/3 state Assumption 2 for T = tau o G^{-1}, while the proof path applies Corollary 1 to f_tau with base map G^{-1}, so the tail exponent that matters is that of G^{-1}(U); the printed alpha can be the wrong exponent. This is a correctness concern and does not change the circularity verdict.

Axiom & Free-Parameter Ledger

0 free parameters · 7 axioms · 1 invented entities

No data-fitted free parameters appear; the central claim is conditional on explicit growth and tail assumptions. The hand-chosen quantities, such as the projection smoothness δ and balancing radius θ*, are auxiliary and absorbed into the ε factor of the final rate. Invented content is limited to a novel projection operator, not an empirical or physical entity.

axioms (7)
  • domain assumption Assumption 1: test and weight functions have derivatives up to order d with |D^a f(x)| ≲ exp(M||x||^2).
    Defines the admissible function class; all rate bounds depend on M. Invoked in Theorem 1 and Lemmas 3.5-3.7.
  • domain assumption Assumption 2: the relevant random variable has sub-Gaussian tail P(||X||>t) ≲ t^η exp(-αt^2).
    Controls the projection bias E|f∘P_r(X)-f(X)|^p. Needed for Gaussian and t proposals. The tail-parameter assignment is the paper's weakest point.
  • domain assumption Assumption 3 / Propositions 1-2: transformation τ has linear growth and derivative growth exp(M_τ||x||^2).
    Used to propagate growth rates to composed integrands fτ and to cover neural-network-like transport maps.
  • domain assumption RQMC point sets satisfy D*_N = O(N^{-1}(log N)^{d-1}) almost surely (Eq. 6).
    Standard property of scrambled low-discrepancy sequences; needed for the Koksma-Hlawka bound on the projected integrand.
  • domain assumption Theorem 3 assumes q(|f|^v) < ∞ for some v > 1 and 1 ≤ p < min{v, α/M_{ωf,τ}}.
    Moment condition controls the maximum of |f| on the bad event in the SNIS decomposition.
  • domain assumption Theorem 3 assumes M_{ω,τ} > 0 is arbitrarily small (QMC-friendly condition).
    This is the load-bearing condition that yields β ≈ 1; without it the rate degrades. It is stated explicitly but is strong.
  • standard math Background analytic tools: Koksma-Hlawka inequality, Faa di Bruno formula, Hölder/Markov inequalities.
    Standard results used throughout Sections 3 and 4 without proof.
invented entities (1)
  • Radial projection operator P_r (Eq. 11) no independent evidence
    purpose: Smoothed truncation of unbounded integrands that keeps Hardy-Krause variation finite without a dimension-dependent operator-norm factor.
    A new mathematical construction used as the main proof tool. It is not a physical entity and has no falsifiable handle outside the paper, but it is internally well-defined and verified in Lemmas 3.2-3.4.

pith-pipeline@v1.3.0-alltime-deepseek · 24877 in / 22386 out tokens · 195986 ms · 2026-08-03T22:23:58.854740+00:00 · methodology

0 comments
read the original abstract

Self-normalized importance sampling (SNIS) is a fundamental tool in Bayesian inference when the posterior distribution involves an unknown normalizing constant. In many applications, both the test function of interest and the underlying state space are unbounded, making direct $L_1$-error (mean absolute error) and $L_2$-error (root mean square error) estimates challenging for SNIS under randomized quasi-Monte Carlo (RQMC) sampling. In this work, we derive the $L_p$-error rate $(p\ge1)$ for RQMC-based SNIS (RQMC-SNIS) estimators with unbounded integrands on unbounded domains. A key step in our analysis is to first establish the $L_p$-error rate for plain RQMC integration. Our results allow for a broader class of transport maps used to generate samples from RQMC points. Under mild function boundary growth conditions, we further establish the \(L_p\)-error rate of order \(\mathcal{O}(N^{-\beta + \epsilon})\) for RQMC-SNIS estimators, where $\epsilon>0$ is arbitrarily small, $N$ is the sample size, and \(\beta \in (0,1]\) depends on the boundary growth rate of the resulting integrand. Numerical experiments validate the theoretical results.

Figures

Figures reproduced from arXiv: 2511.10599 by Jiarui Du, Zhijian He.

Figure 2
Figure 2. Figure 2: The Lp-error for different proposals with different κ and d = 5. 5.2 Bayesian logistic regression We consider a Bayesian logistic regression model following [14]. Let X ∈ R m×d denote the obser￾vation matrix of predictor variables with Xi ∈ R 1×d the i-th row, β = (β1, β2, . . . , βd) T ∈ R d×1 denote the regression parameters, and Y = (Y1, Y2, . . . , Ym) denote the observation with Yi ∈ {0, 1}. The logis… view at source ↗
Figure 3
Figure 3. Figure 3: The Lp-error for different proposals with κ = 1 and d = 30. and the covariance Σ⋆ is given by Σ⋆ = [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The Lp-error for different proposals with Pima dataset. a simple Bayesian inverse problem, examining the impact of various parameters, and find that the experimental outcomes align well with the theory. In addition, we conduct experiments on Bayesian logistic regression, where the results suggest that the linear t proposals constructed using µ and Σ determined by the LapIS method, appear to provide superio… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 6 linked inside Pith

  1. [1]

    Im- portance sampling: Intrinsic dimension and computational cost

    Sergios Agapiou, Omiros Papaspiliopoulos, Daniel Sanz-Alonso, an d Andrew M Stuart. Im- portance sampling: Intrinsic dimension and computational cost. Statistical Science , pages 405–431, 2017

    2017

  2. [2]

    Transformations and Hardy–Kraus e variation

    Kinjal Basu and Art B Owen. Transformations and Hardy–Kraus e variation. SIAM Journal on Numerical Analysis , 54(3):1946–1966, 2016

    1946

  3. [3]

    The Berry-Esseen bo und for Student’s statistic

    Vidmantas Bentkus and Friedrich G¨ otze. The Berry-Esseen bo und for Student’s statistic. The Annals of Probability , 24(1):491–503, 1996

    1996

  4. [4]

    Enhanced convergence rates of Adaptive Importance Sampling with recycling schemes via quasi-Mont e Carlo methods

    Jianlong Chen, Jiarui Du, Xiaoqun Wang, and Zhijian He. Enhanced convergence rates of Adaptive Importance Sampling with recycling schemes via quasi-Mont e Carlo methods. arXiv preprint arXiv:2505.05037, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Constantine and Thomas H

    Gregory M. Constantine and Thomas H. Savits. A Multivariate Faa di Bruno Formula with Applications. Transactions of the American Mathematical Society , 348:503–520, 1996

    1996

  6. [6]

    Randomization of number theoretic methods for multiple integration

    Roy Cranley and Thomas NL Patterson. Randomization of number theoretic methods for multiple integration. SIAM Journal on Numerical Analysis , 13(6):904–914, 1976

    1976

  7. [7]

    On importance sampling and independent Metropolis-Hastings with an unbounded we ight function

    George Deligiannidis, Pierre E Jacob, El Mahdi Khribch, and Guany ang Wang. On importance sampling and independent Metropolis-Hastings with an unbounded we ight function. arXiv preprint arXiv:2411.09514, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    High-dimensional inte gration: the quasi-Monte Carlo way

    Josef Dick, Frances Y Kuo, and Ian H Sloan. High-dimensional inte gration: the quasi-Monte Carlo way. Acta Numerica, 22:133–288, 2013

    2013

  9. [9]

    Digital Nets and Sequences: Discrepancy Theory and Quasi-Monte Carlo Integration

    Josef Dick and Friedrich Pillichshammer. Digital Nets and Sequences: Discrepancy Theory and Quasi-Monte Carlo Integration . 2010

    2010

  10. [10]

    Optimal constants in the Marcinkiewicz–Zygmu nd inequalities

    Dietmar Ferger. Optimal constants in the Marcinkiewicz–Zygmu nd inequalities. Statistics & Probability Letters, 84:96–101, 2014

    2014

  11. [11]

    Sequential quasi Monte Car lo

    Mathieu Gerber and Nicolas Chopin. Sequential quasi Monte Car lo. Journal of the Royal Statistical Society Series B: Statistical Methodology , 77(3):509–579, 2015

    2015

  12. [12]

    Monte Carlo Methods in Financial Engineering , volume 53

    Paul Glasserman. Monte Carlo Methods in Financial Engineering , volume 53. 2004

    2004

  13. [13]

    Quasi-Monte Carlo and importance sampling methods for Bayesian inverse problems

    Zhijian He, Hejin Wang, and Xiaoqun Wang. Quasi-Monte Carlo and importance sampling methods for Bayesian inverse problems. arXiv preprint arXiv:2403.11374 , 2024. 26

    Pith/arXiv arXiv 2024

  14. [14]

    On the error rate of importance sampling with randomized quasi-Monte Carlo

    Zhijian He, Zhan Zheng, and Xiaoqun Wang. On the error rate of importance sampling with randomized quasi-Monte Carlo. SIAM Journal on Numerical Analysis , 61(2):515–538, 2023

    2023

  15. [15]

    Remark on algorithm 659: Imple menting Sobol’s quasirandom sequence generator

    Stephen Joe and Frances Y Kuo. Remark on algorithm 659: Imple menting Sobol’s quasirandom sequence generator. ACM Trans. Math. Softw. (TOMS) , 29(1):49–57, 2003

    2003

  16. [16]

    Transport quasi-monte carlo

    Sifan Liu. Transport quasi-monte carlo. arXiv preprint arXiv:2412.16416 , 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Quasi-Monte Carlo methods with applications in fi nance

    Pierre L’Ecuyer. Quasi-Monte Carlo methods with applications in fi nance. Finance and Stochastics, 13:307–349, 2009

    2009

  18. [18]

    Confidence intervals for randomized quasi-Monte Carlo estimators

    Pierre l’Ecuyer, Marvin K Nakayama, Art B Owen, and Bruno Tuffin . Confidence intervals for randomized quasi-Monte Carlo estimators. In 2023 Winter Simulation Conference (WSC) , pages 445–456. IEEE, 2023

    2023

  19. [19]

    Random Number Generation and Quasi-Monte Carlo Methods

    Harald Niederreiter. Random Number Generation and Quasi-Monte Carlo Methods . 1992

    1992

  20. [20]

    Achieving high converg ence rates by quasi- Monte Carlo and importance sampling for unbounded integrands

    Du Ouyang, Xiaoqun Wang, and Zhijian He. Achieving high converg ence rates by quasi- Monte Carlo and importance sampling for unbounded integrands. SIAM Journal on Numerical Analysis, 62(5):2393–2414, 2024

    2024

  21. [21]

    Randomly permuted ( t,m,s )-nets and ( t,s )-sequences

    Art B Owen. Randomly permuted ( t,m,s )-nets and ( t,s )-sequences. In Monte Carlo and Quasi-Monte Carlo Methods in Scientific Computing , pages 299–317. Springer, 1995

    1995

  22. [22]

    Multidimensional variation for quasi-Monte Carlo

    Art B Owen. Multidimensional variation for quasi-Monte Carlo. In Contemporary Multivari- ate Analysis And Design Of Experiments: In Celebration of Pr ofessor Kai-Tai Fang’s 65th Birthday, pages 49–74. 2005

    2005

  23. [23]

    Halton sequences avoid the origin

    Art B Owen. Halton sequences avoid the origin. SIAM review , 48(3):487–503, 2006

    2006

  24. [24]

    Practical Quasi-Monte Carlo Integration

    Art B Owen. Practical Quasi-Monte Carlo Integration . 2023

    2023

  25. [25]

    Error estimation for quasi-Monte Carlo

    Art B Owen. Error estimation for quasi-Monte Carlo. arXiv preprint arXiv:2501.00150 , 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Coverage errors for Student’s t confidence inter vals comparable to those in Hall (1988)

    Art B Owen. Coverage errors for Student’s t confidence inter vals comparable to those in Hall (1988). arXiv preprint arXiv:2501.07645 , 2025

    Pith/arXiv arXiv 1988

  27. [27]

    Quasi-Monte Carlo integra tion over Rs with boundary- damping importance sampling

    Zexin Pan, Du Ouyang, and Zhijian He. Quasi-Monte Carlo integra tion over Rs with boundary- damping importance sampling. arXiv preprint arXiv:2509.07509 , 2025

    arXiv 2025

  28. [28]

    Skewness of a randomized quasi-Mont e Carlo estimate

    Zexin Pan and Art B Owen. Skewness of a randomized quasi-Mont e Carlo estimate. Journal of Complexity , page 101956, 2025

    2025

  29. [29]

    Inverse problems and data assimila- tion, volume 107

    Daniel Sanz-Alonso, Andrew Stuart, and Armeen Taeb. Inverse problems and data assimila- tion, volume 107. Cambridge University Press, 2023

    2023

  30. [30]

    Activation functions in neural networks

    Sagar Sharma, Simone Sharma, and Anidhya Athaiya. Activation functions in neural networks. Towards Data Sci , 6(12):310–316, 2017

    2017

  31. [31]

    High-Dimensional Probability: An Introduction with Appli cations in Data Science

    Roman Vershynin. High-Dimensional Probability: An Introduction with Appli cations in Data Science. 2018. 27

    2018

  32. [32]

    Efficient comput ation of option prices and Greeks by quasi–Monte Carlo method with smoothing and dimension re duction

    Chengfeng Weng, Xiaoqun Wang, and Zhijian He. Efficient comput ation of option prices and Greeks by quasi–Monte Carlo method with smoothing and dimension re duction. SIAM Journal on Scientific Computing , 39(2):B298–B322, 2017

    2017

  33. [33]

    Jesse Bennett Windle. Forecasting high-dimensional, time-varying variance-co variance matri- ces with high-frequency data and sampling P´ olya-Gamma ran dom variates for posterior dis- tributions derived from logistic likelihoods . PhD thesis, The University of Texas at Austin, 2013

    2013

  34. [34]

    Efficient importan ce sampling in quasi-Monte Carlo methods for computational finance

    Chaojun Zhang, Xiaoqun Wang, and Zhijian He. Efficient importan ce sampling in quasi-Monte Carlo methods for computational finance. SIAM Journal on Scientific Computing , 43(1):B1– B29, 2021. 28

    2021