{"id":"26568bec-06d2-48aa-95c1-19bf63ecb403","arxiv_id":"2505.18510","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TSS stratifies the input distribution into nested tail sets with geometrically decreasing probabilities, yielding failure-probability estimates with exponentially decaying bias and variance that does not grow as the failure probability shrinks.","lead":"This paper introduces Tail Stratified Sampling (TSS), a Monte Carlo method that estimates rare failure probabilities by concentrating simulation samples in the low-probability tails of the input distribution. If the method works as claimed, it gives engineers a simpler, variance-efficient alternative to subset simulation for many reliability problems, though its performance in high dimensions depends on how well the relevant tails can be identified.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"High-dimensional efficiency of TSS rests on hand-crafted A0; the paper supplies no general construction, so 'identifiable tails' is an assumption rather than a demonstrated capability.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: TSS efficiency depends on a high-probability, failure-free null stratum A0, and the paper's general design-point construction fails in high dimensions. The paper is honest about this limitation, and the central claim is explicitly conditional on failures lying in identifiable tails. That condition is not met by any general procedure in the paper; the only high-dimensional success uses a hand-crafted A0. This is a real limitation on the scope of the central claim, but it is not an internal inconsistency. The reader's CONDITIONAL verdict already reflects this, so no change is needed. Secondary issues—the dimensionally inconsistent Eq. (50) and the use of the same data in Algorithm 2 to define strata and estimate failure probabilities—are also real but do not change the overall verdict; they reinforce that the paper should be accepted only after these technical points are cleaned up and the high-dimensional scope is stated more carefully.","tokens_in":32374,"tokens_out":9945,"duration_ms":96115,"concrete_test":"Run the Section 5.3 stranded-wire problem with N=1000 and A0 constructed automatically, for example by fitting a one-class density-quantile threshold on an independent pilot sample such that no pilot failures fall in A0, without using the known Amin expression. Repeat 100 trials and compare the mean PF and CoV against the reported TSS-ML and SuS results in Figure 11. If the automatic-A0 version does not reproduce the order-of-magnitude CoV reduction over SuS, the efficiency claim is contingent on problem-specific tail knowledge rather than on TSS itself.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 makes the efficiency condition explicit: TSS is only provably efficient when the conditional failure probabilities P(Fi) are large, which requires a null stratum A0 with substantial probability and no failures. Section 3.1 offers the design-point construction as the general way to obtain A0, but Sections 5.1 and 5.2 show that in 1000-D problems this A0 has vanishingly small probability, so TSS has larger CoV than subset simulation. The only successful high-dimensional example, Section 5.3, uses a problem-specific A0 (AT(x) > Amin) derived from physical insight, not from a general algorithm. Thus the central claim of orders-of-magnitude variance reduction is established only for low-dimensional problems and for special cases where the user already knows how to isolate the failure-relevant tail. The variance formula in Eq. (26) is correct for fixed strata, but its practical precondition—an identifiable, high-probability A0—is exactly what the paper cannot provide in high dimensions without external problem knowledge.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Tail Stratified Sampling (TSS), a stratified Monte Carlo estimator for rare-event failure probabilities. The sample space is partitioned into nested regions defined by level sets of the input density, with an auxiliary ordering for plateau regions; a user-specified safe set A0 may be removed, and the remaining probability mass is subdivided so that the stratum probabilities follow P(Ai)=p0^{i-1}(1-p0)P(A*). The resulting estimator is a weighted sum of stratum failure indicators, with bias bounded by p0^m P(A*) and variance given in closed form. The paper derives special constructions for Gaussian and uniform inputs, discusses proportional, equal, and Neyman sample allocations, and compares TSS with subset simulation on seven 2-D benchmark problems, a 7-D buckling problem, and 1000-D beam, oscillator, and cable examples. The authors are explicit that efficiency depends on identifying a high-probability null stratum A0 with no failures, and they report both successful and unsuccessful high-dimensional applications.","tokens_in":32524,"tokens_out":9060,"duration_ms":79130,"significance":"If the estimator and its variance expressions are correct, TSS is a clean, interpretable variance-reduction scheme with several genuine strengths: the estimator follows directly from the law of total probability, the bias and variance formulas are explicit and falsifiable, the method uses only failure indicators and therefore applies to stochastic performance functions, and the 2-D benchmarks show large CoV reductions relative to subset simulation, including cases where SuS does not converge. The paper is also unusually candid about the central limitation, namely that efficiency requires a null stratum A0 of substantial probability in which no failures occur; no general construction of such a stratum is provided for high dimensions. The significance of the contribution is therefore real but narrower than the title and abstract suggest: the demonstrated order-of-magnitude gains are established for low-dimensional problems and for high-dimensional problems in which the user can supply problem-specific tail knowledge.","major_comments":[{"comment":"Equation (50) defines the stratum radii r_i by adding a probability term to a density value: r_i = phi_d^{-1}( phi_d(r0) + (1-p0^i)(1 - F_chi(...)) ). A density value and a probability cannot be added, so the argument is dimensionally inconsistent. The formula also violates nesting: for d=2, p0=0.1, and r0=3, the i=1 radius is approximately 2.28, which is smaller than r0, making A1 = {r0 <= r <= r1} empty and invalidating the subsequent strata. The correct condition should be F_chi(r_i) = 1 - p0^i (1 - F_chi(r0)), i.e., r_i = F_chi^{-1}(1 - p0^i(1 - F_chi(r0))). Because Eq. (50) is the concrete Gaussian specialization underlying the benchmark examples, it must be corrected and the affected experiments re-verified, or the text must state explicitly that the examples do not use Eq. (50).","section":"§2.5.1, Eq. (50)"},{"comment":"The abstract and contribution list present TSS as a generally efficient method and emphasize the absence of explicit dimension dependence. The load-bearing condition for that benefit is the existence of a null stratum A0 with substantial probability and no failures (Section 2.3). The only general construction offered, the design-point hypersphere of Section 3.1, is shown in Sections 5.1 and 5.2 to yield higher CoV than subset simulation in 1000-D examples. The one successful 1000-D example in Section 5.3 uses a hand-crafted, problem-specific A0 based on physical insight rather than a general algorithm. The claimed high-dimensional efficiency should therefore be presented as conditional on problem-specific tail identification; the abstract and conclusions should not imply a demonstrated general high-dimensional advantage unless a general construction of A0 is supplied.","section":"§5.1–5.3 and §2.3"},{"comment":"The bias and variance guarantees in Section 2.3 are derived for a fixed partition with known probabilities P(Ai). In Algorithm 2, the strata are constructed from the same N samples that are subsequently used to estimate P(Fi). Because the order statistics of fX(X) are generally dependent on the performance function g(X), it is not immediate that Eq. (23) and Eq. (26) remain valid for this adaptive empirical variant. The authors should provide an asymptotic justification or a numerical verification of the variance estimator's coverage; otherwise the claimed precise error quantification is not established for the general arbitrary-distribution implementation.","section":"§3.3, Algorithm 2"}],"minor_comments":[{"comment":"Line 9 contains an unresolved placeholder annotation '▷ ??'; this should be completed.","section":"Algorithm 1, line 9"},{"comment":"'Anther case of practical importance' should read 'Another case'.","section":"§2.5.2, first paragraph"},{"comment":"References [17] and [27] are the same article and should be merged; Section 7 contains the typo 'probem' for 'problem'.","section":"References"},{"comment":"The notation '= 0 [ ∵ P(F0):=0]' should be typeset as a proper sentence explaining that the term vanishes because P(F0)=0.","section":"Eq. (31)"},{"comment":"The caption contains stray glyph strings such as '1 22 2' and '12 23 34 45 5' that should be removed or replaced with the intended labels.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The core idea is a fairly direct application of stratified sampling to density contours, and the novelty lies mostly in the detailed construction, the explicit bias/variance handling, and the honest benchmarking. The dimensional inconsistency in Eq. (50) looks like a typesetting or transcription error, but it is central enough that the paper cannot be accepted in its current form; the authors must either correct the formula and re-verify the examples or state clearly that the numerical experiments use a different radius construction. The paper fits the journal's scope and the low-dimensional results are strong; I would be willing to re-review a substantially revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real contribution, not a repackaging. TSS is stratified sampling on density level sets with geometric tail probabilities, and the bias/variance analysis in Eqs. (23) and (26) is correct and useful. The low-dimensional benchmarks are well chosen and the CoV gains over subset simulation are large. The paper is also unusually honest about where the method fails: the design-point construction of A0 collapses in high dimensions, and the one successful 1000-D example needs a problem-specific A0 from physical insight. That is not a hidden flaw; it is stated in Section 2.3 and demonstrated in Section 5.\n\nTwo soft spots are worth flagging. First, Eq. (50) in the normal special case adds a probability to a density: phi_d(r0) is a density value, and (1 - p0^i)(1 - F_chi(...)) is a probability, so the sum is dimensionally inconsistent. The correct construction should invert F_chi, not phi_d. As written, following Eq. (50) would give wrong stratum radii. This is likely a typo, but it is in a central formula and needs fixing.\n\nSecond, Algorithm 2 defines the empirical strata from the same sample used to estimate P(Fi). The paper does not prove unbiasedness of this adaptive post-stratification. It may be asymptotically fine, but given how carefully the paper treats bias elsewhere, the omission is noticeable. I would like to see either a proof of unbiasedness or an explicit caveat with a small simulation check.\n\nOn the stress-test concern: it is fair but the paper already concedes it. The central claim is conditional: TSS gives orders-of-magnitude variance reduction when failures lie in identifiable tails with a high-probability A0. For high dimensions, no general construction for A0 is given, and Sections 5.1–5.2 show the design point fails. That is a limitation, not a contradiction, and the authors say so themselves.\n\nThis deserves a serious referee. The method is novel, the general estimator is sound, and the experiments are extensive. I would accept with major revision: fix Eq. (50), clarify or prove the adaptive-strata property, and temper the abstract claim so it reflects the necessity of tail isolation. I would also bring it to our reading group, because the tail-definition discussion in Section 1.4 and the high-dimensional A0 problem are worth thinking through.","headline":"Solid stratified-sampling method with honest limits; the main fix needed is a wrong-looking formula in Eq. (50) and a missing unbiasedness argument for the adaptive strata.","tokens_in":33067,"tokens_out":5023,"would_cite":true,"duration_ms":44856,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62N05","65C05","62D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Tail Stratified Sampling estimates rare failure probabilities by concentrating samples in the tails, and its variance stays controlled as failure probabilities become extremely small.","keywords":["tail stratified sampling","rare event estimation","failure probability","stratified sampling","Monte Carlo variance reduction","subset simulation","reliability analysis","high-dimensional tails"],"falsifier":"Compute the conditional probabilities $P(F_i)$ in the 1000-dimensional cantilever or SDOF example using the design-point null stratum. If those conditional probabilities are all within a small factor of $P_F$, the paper's own condition for provably small variance is violated, and TSS should show roughly the same coefficient of variation as subset simulation rather than an order-of-magnitude improvement. A sharper test would construct a high-dimensional Gaussian failure surface spread across many coordinates, such as many small exceedances rather than one extreme direction, and check whether any density-threshold tail contains a large fraction of the failure mass; if not, no choice of $p_0$ and $m$ can make TSS's variance small.","tokens_in":32122,"feed_emoji":"🎯","tokens_out":9882,"duration_ms":71119,"temperature":0.7,"pith_summary":"Tail Stratified Sampling (TSS) is a Monte Carlo estimator for very small failure probabilities that spends nearly all of its samples in the tails of the input distribution, where failures actually happen. The paper's core claim is that stratifying the sample space by nested density thresholds turns rare-event estimation into a sum of conditional probabilities with known geometric weights, so the bias can be made exponentially small and the variance does not grow as the failure probability $P_F$ shrinks. In seven two-dimensional benchmarks with failure probabilities from $10^{-2}$ to $10^{-9}$, TSS reports coefficients of variation one to two orders of magnitude lower than subset simulation at the same budget, and it succeeds in cases where subset simulation fails to converge. The paper also shows that this performance depends on identifying a high-probability safe region $A_0$ whose complement contains all failures; in high dimensions the standard design-point construction of $A_0$ has vanishing probability and TSS loses its advantage unless a problem-specific tail is defined.","feed_headline":"Tails-only sampling beats subset simulation on rare failures","feed_subtitle":"New stratified estimator keeps bias and variance under control when failure probabilities drop below one in a billion.","key_machinery":"The machinery is the density-threshold tail decomposition. A set $S_{u,v}$ contains points whose joint density is above a threshold $u$, with an auxiliary parameter $v\\in[0,1]$ resolving plateaus where the density is constant; the tail is the complement $T_{u,v}=S_{u,v}^c$. A nested sequence $S_{u_1,v_1}\\subset S_{u_2,v_2}\\subset\\cdots$ is chosen so that each tail retains a fixed fraction $p_0$ of the previous tail's probability, giving strata $A_i=S_{u_i,v_i}\\setminus S_{u_{i-1},v_{i-1}}$ with exactly known geometric probabilities $P(A_i)=p_0^{i-1}(1-p_0)P(A^*)$. Because the strata are defined by the scalar density value, the scheme is effectively one-dimensional and has no explicit dependence on the input dimension; the dimension enters only through the difficulty of finding a suitable null stratum $A_0$.","core_discovery":"The central claim of the paper is that a failure probability $P_F$ can be estimated accurately by the Tail Stratified Sampling estimator\n$$\\hat $P_F^{{TSS}}$ = (1-p_0)P(A^*) \\sum_{i=1}^{m} $p_0^{{i-1}}$ \\frac{1}{N_i}\\sum_{k=1}^{N_i} I_{\\{g(x)\\le 0\\}}(x_k\\mid A_i),$$\nwhere $A_1,\\dots,A_m$ are nested tail strata with probabilities $P(A_i)=p_0^{i-1}(1-p_0)P(A^*)$. The paper proves that the bias of this truncated estimator is at most $p_0^m P(A^*)$ and that its variance is\n$$\\mathrm{Var}[\\hat $P_F^{{TSS}}$] = [(1-p_0)P(A^*)]^2 \\sum_{i=1}^{m} \\frac{$p_0^{{2(i-1)}}$ P(F_i)(1-P(F_i))}{N_i},$$\nso the stratum weights vanish exponentially and the estimate is dominated by the high-probability strata where failures almost never occur. Provided failures are confined to a known safe-region complement $A^*$ and the conditional failure probabilities $P(F_i)$ are large compared with $P_F$, the variance does not depend on the magnitude of $P_F$; the method, as the paper puts it, can therefore work better for smaller failure probabilities that occur deeper in identifiable tails.","pith_inferences":["Read as a recipe, TSS says that for extremely rare events the best use of a fixed budget is to certify a large safe core and spend nearly every evaluation inside the certified tail, which inverts the usual importance-sampling logic of placing weight near a design point.","The paper's high-dimensional examples leave the construction of $A_0$ as an open step; a natural next test is whether data-driven dimension reduction can identify a high-probability safe region automatically in problems where the design point fails.","Because TSS consumes only the failure indicator, the same decomposition should transfer to rare-event estimation for stochastic simulators, queues, and other black-box systems, not just deterministic structural performance functions."],"forward_implications":["Choosing $p_0=0.1$ and $m=4$ gives a bias below $10^{-4} P(A^*)$, and the paper uses this as a default in all examples.","For the seven two-dimensional benchmark problems, TSS achieves coefficients of variation one to two orders of magnitude lower than subset simulation with the same total sample budget.","TSS uses only the failure indicator $I_{\\{g(x)\\le 0\\}}$, so it can estimate failure probabilities when the performance function is stochastic or unavailable; the paper demonstrates this on a cable with an unknown number of strands.","In high dimensions the design-point definition of $A_0$ yields a vanishingly small safe-region probability, and TSS then loses its efficiency; replacing it with a physically motivated safe region restores an order-of-magnitude coefficient-of-variation advantage on the 1001-dimensional cable problem.","The estimator's variance does not depend explicitly on $P_F$, so within its regime of validity TSS improves rather than degrades as the target probability becomes smaller."],"supporting_citations":[{"why":"Supplies the subset simulation benchmark and the SDOF oscillator test problem used throughout the comparison.","marker":"[8]"},{"why":"Supplies the two-dimensional benchmark performance functions with complex and disjoint failure domains.","marker":"[16]"},{"why":"Provides the geometric argument that the design point becomes statistically insignificant in high dimensions, motivating the need for an alternative null stratum.","marker":"[18]"},{"why":"Supplies the classical stratified sampling variance formula and Neyman allocation on which TSS builds.","marker":"[21]"},{"why":"Provides the transformation that lets arbitrary non-Gaussian inputs be handled through independent standard normal variables.","marker":"[30]"},{"why":"Describes the related beta-sphere sampling method for standard normals that TSS extends by removing the safe stratum and exploiting tail variance decay.","marker":"[31]"},{"why":"Supplies the stable buckling problem used as the seven-dimensional numerical case study.","marker":"[39]"}],"fun_headline_variants":["Tail-only sampling makes rare failures computable","Sample the tails, skip the bulk: rare failure wins","Tail Stratified Sampling: variance tamed for tiny odds","Billion-to-one failure? Just sample the tails"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The efficiency guarantee collapses unless one can identify a high-probability safe region $A_0$ where failure is impossible, so that every tail stratum has a conditional failure probability much larger than the overall failure probability.","fun_headline_variants_meta":{"raw":{"variants":["Tail-only sampling makes rare failures computable","Sample the tails, skip the bulk: rare failure wins","Tail Stratified Sampling: variance tamed for tiny odds","Billion-to-one failure? Just sample the tails"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000308,"raw_usage":{"total_tokens":1840,"prompt_tokens":1102,"completion_tokens":738,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":675}},"tokens_in":718,"tokens_out":738,"duration_ms":6489,"temperature":1.0,"reasoning_tokens":675,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:30:14.557911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the conditional probabilities $P(F_i)$ in the 1000-dimensional cantilever or SDOF example using the design-point null stratum. If those conditional probabilities are all within a small factor of $P_F$, the paper's own condition for provably small variance is violated, and TSS should show roughly the same coefficient of variation as subset simulation rather than an order-of-magnitude improvement. A sharper test would construct a high-dimensional Gaussian failure surface spread across many coordinates, such as many small exceedances rather than one extreme direction, and check whether any density-threshold tail contains a large fraction of the failure mass; if not, no choice of $p_0$ and $m$ can make TSS's variance small.","supporting_citations":[{"cited_title":"Reliability analysis of discrete-state performance functions via adaptive sequential sam- pling with detection of failure surfaces","cited_arxiv_id":null,"evidence_quote":"Supplies the two-dimensional benchmark performance functions with complex and disjoint failure domains."},{"cited_title":"Geometric insight into the challenges of solving high-dimensional reliability problems","cited_arxiv_id":null,"evidence_quote":"Provides the geometric argument that the design point becomes statistically insignificant in high dimensions, motivating the need for an alternative null stratum."},{"cited_title":"The art of simulation","cited_arxiv_id":null,"evidence_quote":"Supplies the classical stratified sampling variance formula and Neyman allocation on which TSS builds."},{"cited_title":"An innovating analysis of the nataf transformation from the copula viewpoint","cited_arxiv_id":null,"evidence_quote":"Provides the transformation that lets arbitrary non-Gaussian inputs be handled through independent standard normal variables."},{"cited_title":"A stratified beta-sphere sam- pling method combined with important sampling and active learning for rare event analysis","cited_arxiv_id":null,"evidence_quote":"Describes the related beta-sphere sampling method for standard normals that TSS extends by removing the safe stratum and exploiting tail variance decay."},{"cited_title":"Modified replica exchange-based mcmc algorithm for estimation of structural reliability based on particle splitting method","cited_arxiv_id":null,"evidence_quote":"Supplies the stable buckling problem used as the seven-dimensional numerical case study."}],"review_version":1}