acceptodds
Under review as a conference paper at ICLR 2027

Sampling the Shell: From Exponential to Constant Sample Complexity in Negative-Sampling Anomaly Detection

Abstract

Negative-sampling anomaly detection reduces unsupervised fault detection to binary classification: observed telemetry is labeled normal, synthetic negatives are drawn from a reference probability distribution, a binary classifier is trained, and Integrated Gradients converts its alarms into per-sensor symptoms for downstream root-cause analysis with language models. The recipe is deployed in **intelligent diagnostics**, yet no theory has said when, and at what dimension, it can work. We supply one, grounded in high-dimensional probability : concentration of measure, Le Cam's minimax theory, Markov semigroups, Poincar\'e inequalities, and Langevin dynamics. *Limits:* manufactured negatives are ancillary to learning decision boundaries; no negative-generation scheme adds boundary information, so positive samples are necessary at boundary resolution . And uniform sampling operates only inside a dimension window with closed-form optimum obeying a occupancy law, having certified budgets that are linear in inside the window and exponential outside it. *Constructions:* shell-concentrated, importance-weighted samplers restore a dimension-free band-hit rate, an exponential-to-constant reduction in negative sample complexity, realized by data-seeded Langevin dynamics and dimension-reducing autoencoders with dimension-free mixing and finite-sample certificates. We present an algorithmic framework as a ladder of eight detectors, from the MADI baseline to certified latent pipelines, each licensed by a theorem. Experiments confirm the predictions: on synthetic suites (–, one to five operating modes) detectors attain ROC-AUC of on box-structured modes at every dimension while the baseline method decays toward chance. On the multivariate track of the TSB-AD benchmark, under its official evaluation protocol and over fifteen series spanning the benchmark families, the strongest variants exceed the PCA, LSTMAD, and OmniAnomaly incumbents and match the top-ranked CNN (median AUC vs ). Finally, on real commercial flight-test data, certified and latent detectors reach , and AUC on fan-blade-out and compressor-stall events and up to on vibration, against - for the deployed baseline.

Primary area:
learning theory
Keywords:
anomaly detection, negative sampling, learning theory, minimax lower bounds, sample complexity, concentration of measure, Langevin dynamics, Poincaré inequality, time-series anomaly detection, explainability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hZ2KOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes negative-sampling anomaly detection in high dimensions. It establishes bounds on uniform negatives reaching the vicinity of a normal region, argues that generated negatives cannot add information beyond the positive sample, and characterizes a dimension-dependent operating window with a limiting 1/e occupancy rule. It proposes importance-weighted shell/envelope sampling, exact sampling for product envelopes, and a box-based boundary selector with conditional finite-sample guarantees. Experiments include controlled synthetic geometry, the 180-series TSB-AD-M benchmark, and flight telemetry with injected failures. The training-free CANONICAL method reaches 0.359 VUS-PR on TSB-AD-M, below xLSTMAD (0.392) but above the reported classical/deep benchmark baselines.

Critical strengths
  1. The paper offers a principled account of high-dimensional negative-sampling failure and connects boundary-hit rates, minimax information limits, and sampler design. This is a substantive conceptual contribution beyond a heuristic detector.

  2. The evaluation tests the central geometric predictions directly and includes broad time-series benchmarking with paired uncertainty estimates. The authors explicitly report that xLSTMAD remains ahead on VUS-PR, improving the credibility of the empirical account.

Critical weaknesses
  1. The strongest certificate assumes well-separated, axis-aligned box modes, density flatness, correct partitioning, and sufficient samples. Several deployed variants use trimmed boxes or latent transformations outside that theorem. In the D=128 sweep, the stated sufficient budgets are far above the actual experimental budgets; AUC 1.000 is therefore not a demonstration of the formal certificate in operation. This limits the claimed theory-to-practice connection.

  2. The main empirical gain appears tied to replacing a learned head with a geometric box estimator rather than to improved negative generation alone: envelope negatives dramatically improve band-hit efficiency, but the neural detector remains close to the uniform-negative head. The incremental value of the weighted negative-based selector over a positive-only box detector is not isolated.

Questions for authors
  1. Can the authors provide a matched-budget ablation of positive-only box scoring, box scoring plus envelope sampling, and the full median-of-means selector, reporting where the negatives change predictions? This would materially affect the algorithmic novelty assessment.

  2. Which benchmark and flight-test runs satisfy the precise hypotheses and budgets of Theorem 4.5, and what evidence supports approximate validity of the box/flatness/mode-partition assumptions in the others?

  3. How does CANONICAL degrade under misspecified K, overlapping modes, and correlated nominal data, with the same selection protocol?

1 comment
AuthorsOpenReview ↗

The certificate now runs where budgets are payable (193/200 vs the 95% guarantee); the ablation is exact: benchmark predictions match the positive-only box, conceded; negatives fire where evidence thins; robustness swept.

Thank you for pressing on exactly the right two questions: does the certificate ever actually run, and do the negatives earn their place? We answered both with experiments (D, E, F; scripts included). Symbols: D is dimension, M = 4,000 positives, N = 8,000 synthetic negatives, t̂ is the padding the selector grants (zero means the tight min/max box stands).

  1. Does the certificate ever run? As submitted, no, and we say so plainly. No benchmark or flight run meets both Theorem 4.5's hypotheses and its budgets; that sentence moves from an appendix remark into the main text. The discussion period closes the gap from both sides. Experiment D runs the complete procedure where the budgets are payable, at D = 8 and 16, with every hypothesis exact and every budget at its formula ceiling (M_k = 1,832 and 8,037 positives per mode; N_k = 557,354 draws, which does not grow with D; about a second per run). The full certificate event held in 193/200 and 100/100 runs against the guaranteed 95%. The seven failures are the best part: all arrived through the box-accuracy clause, at 3.5%, inside that clause's 5% allocation. A guarantee that fails at its budgeted rate, through its budgeted clause, is operating, and its positive budget is nearly tight. Experiment E measures where the conservatism lives. The theorem asks 6,501 draws per block; about 50 suffice at D = 8. That 129× gap splits into worst-case envelope-geometry constants (27 to 43×, measured directly) and the Chebyshev step (3 to 5×: on the collar the weights are nearly constant, so the statistic is nearly a coin flip). The block count is not slack; one removable D/m factor on the positive side costs 16× at D = 128. End to end, about 116× at D = 8, matching your 60 to 150×. The revision adds this table and reframes the budgets as structural results rather than sizing rules. Outside the theorem, per-run diagnostics take over: clustering quality, box-plus-collar disjointness (it flags every overlapping run below), edge mass (0.3 to 1.8%), and the selector's own firing pattern. The lens keeps the certificate under one explicit assumption (flatness becomes a copula bound; proposition added); encoder latents keep only the resolution clause. A per-rung warrant table records this.
  2. What do the negatives add? Your ablation is exact. t̂ enters the score once, G(x) = Σ_d(dist_d(x) − t̂)₊, so arm one (positive-only box) is our detector with t̂ = 0, and whenever the selector returns zero the arms are bit-identical. Arm two collapses as well: the envelope feeds nothing but the selector. The whole ablation therefore reduces to: when is t̂ nonzero? On the paper's benchmarks, never, and we concede this plainly: 65/65 archived zeros, maximum score difference 0.0, and the main text now says the deployed detector and a positive-only box predict identically there. In multimode geometries the negatives work: three modes sharing N = 8,000 leave each mode's evidence thin, the selector fires 20/20 with collar-capped t̂ ≈ 0.004 to 0.007, predictions genuinely diverge, and AUC stays 1.000 because fault depths exceed the padding. Under stress it is load-bearing: starved budgets fire 100% at N ≤ 1,600; a 4× envelope-scale error fires at full budget; with everything correct, 0 false alarms in 300 runs against the allowed 5%. And even at t̂ = 0 the negatives buy the level (Lemma 4.1: the sampling ratio fixes the operating point), the receipt (an error-controlled verification no positive-only method can produce, by Prop. 3.3), and feasibility (at matched N = 557,354 and D = 64, uniform negatives verify 0 of 8 bands; the envelope verifies 8 of 8). The introduction now locates the novelty accordingly: not boxes, not better negatives for learned heads (Cor. 3.5 forbids that), but the certified calibration-and-verification layer around a simple core.
  3. Robustness (Experiment F; protocol fixed; 20 seeds). Wrong mode count is one-sided: too few is costly (AUC 0.831 at K = 2, 0.667 at K = 1; merged boxes swallow the corridor between modes), too many is nearly free (1.000 up to K = 8 against a true 3). The rule writes itself: over-partition. Overlapping modes never hurt: AUC 1.000 from well separated to 50% overlap, because the union of boxes still covers the support; separation is a warranty boundary, not a performance cliff. Correlation splits cleanly: amplitude faults are caught at 1.000 even at correlation 0.9 (box walls are marginal objects), while structural faults, which break dependence yet stay in range, sit at 0.500 everywhere. Axis-aligned geometry is categorically blind to dependence; that is exactly the gap the lens and latent rungs exist for.

Closing. The certificate now runs where its budgets are payable and fails exactly as budgeted; the negatives' contribution is isolated, conceded where absent and measured where present. If this resolves your concerns, we would be grateful if you would consider strengthening your assessment.

Reviewer_8dCBOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper analyzes negative-sampling anomaly detection [1] under a box model of normal operation. It argues three things: uniform negatives land near the normal region with probability at most (ρ+2ε)^D; no negative generator can lower the positive-sample requirement M ≳ ρ/ε; and uniform sampling works only inside a closed-form dimension window whose optimum sits at 1/e occupancy. It then proposes importance-weighted "envelope" samplers with a dimension-free band-hit rate, an "(ε, α)-certificate" for per-mode min/max boxes, and a ladder of eight detectors. These are evaluated on a synthetic dimension sweep, synthetic suites, the 180-series TSB-AD-M benchmark [2] (certified boxes reach VUS-PR 0.359, above the classical/deep baselines and behind xLSTMAD), and a seeded-fault flight test with 63 failure flights.

References

  • [1] J. Sipple. Interpretable, multidimensional, multimodal anomaly detection with negative sampling for detection of device failure. ICML 2020, PMLR 119:9016–9025.
  • [2] Q. Liu and J. Paparrizos. The elephant in the room: Towards a reliable time-series anomaly detection benchmark. NeurIPS 2024, Datasets and Benchmarks Track.
Critical strengths
  • The closed-form dimension window (Theorem 3.6) is the paper's one genuinely new result: the band mass is unimodal in D with maximizer D⋆ = ln(a/b)/(a−b), and the optimal occupancy tends to 1/e as ε → 0. It quantifies what earlier work described only qualitatively [3, 4]; the derivation is correct, but it is elementary and holds only within the box model.

References

  • [3] D. M. J. Tax and R. P. W. Duin. Uniform object generation for optimizing one-class classifiers. JMLR 2:155–173, 2001.
  • [4] K. Hempstalk, E. Frank, and I. H. Witten. One-class classification by combining density and class probability estimation. ECML PKDD 2008, LNCS 5211, pp. 505–519.
Critical weaknesses
  • The selector (Alg. 1) tests the non-negative statistic w(v)·1{v ∈ band} and never evaluates the positives, so it certifies zero inflation for essentially any box. In my re-implementation of the notebook code it did so in 30 of 30 runs, including for a box [0.3, 0.7]^D that excludes at least 99.6% of the normal mass. The best rungs therefore reduce to per-cluster limit checking, and wherever negatives do matter the envelope sampler never helps: the gap is ≤ 0.02 in Fig. 2a, AUC is 0.711 vs. 0.710 on TSB-AD-M, and 0.598 vs. 0.736 on the flights. No experiment supports the "exponential-to-constant" claim.

  • Uncited work already says that uniform artificial negatives become infeasible as D grows and must be generated near the target class with a density correction [3, 4], and negative generation designed for high dimensions or concentrated on shells and boundaries is established [5–7]. What remains is elementary (maximizing Dρ^D, plus a classical 1/M two-point bound combined with data processing [8, 9]), and it is misapplied:

    • Corollary 3.5 is used to "explain" empirical ties, although Eq. (2) is non-trivial only for M < (ρ+ε)/ε ≤ 81, while the experiments use M = 4,000.
    • The C-BCLA code runs ULA on a classifier-logit energy, not on Theorem 4.3's envelope.
    • Theorem 4.5's budgets at D = 128 are 60–150× those used.
  • Per the supplementary notebook:

    • The TSB-AD-M robustness layer and score smoother were designed over at least four runs using test metrics on 15 preview series that are themselves part of the 180 evaluated.
    • The rungs load Optuna-tuned "dial books" whose tuning data are never stated.
    • The sentinel gate and smoothing window read test-segment features, despite the "train-prefix-decided" claim.

    This post-processing (+0.032 VUS-PR on the preview, against a +0.046 margin over CNN) is given only to the proposed rungs, and the closest training-free baselines are missing [10–12]. Separately, the abstract's flight-test "lead" is not statistically resolved (ΔAUC +0.026, 95% CI [−0.030, +0.087]) and is confounded by different inputs (41 vs. 72 features).

References

  • [5] C. Désir, S. Bernard, C. Petitjean, and L. Heutte. One class random forests. Pattern Recognition 46(12):3490–3506, 2013.
  • [6] S. Goyal, A. Raghunathan, M. Jain, H. V. Simhadri, and P. Jain. DROCC: Deep robust one-class classification. ICML 2020, PMLR 119:3711–3721.
  • [7] L. Tao, X. Du, X. Zhu, and Y. Li. Non-parametric outlier synthesis. ICLR 2023.
  • [8] A. P. Korostelev and A. B. Tsybakov. Minimax Theory of Image Reconstruction. Lecture Notes in Statistics 82, Springer, 1993.
  • [9] A. B. Tsybakov. Introduction to Nonparametric Estimation. Springer Series in Statistics, Springer, 2009.
  • [10] M. Goldstein and A. Dengel. Histogram-based outlier score (HBOS): A fast unsupervised anomaly detection algorithm. KI-2012: Poster and Demo Track, pp. 59–63, 2012.
  • [11] Z. Li, Y. Zhao, N. Botta, C. Ionescu, and X. Hu. COPOD: Copula-based outlier detection. ICDM 2020, pp. 1118–1123.
  • [12] Z. Li, Y. Zhao, X. Hu, N. Botta, C. Ionescu, and G. H. Chen. ECOD: Unsupervised outlier detection using empirical cumulative distribution functions. IEEE TKDE 35(12):12181–12193, 2023.
Questions for authors
  • Please report the distribution of t̂ on TSB-AD-M and the flights, rerun CANONICAL with the envelope draws and selector removed (t̂ ≡ 0), and run the selector on deliberately wrong boxes. Also, under your own Lemma 4.1 risk at D ∈ {32, 128} with M = 4,000, isn't R(N̂^ε) − R(N̂) negative, since P(N∖N̂) ≈ 2D/M outweighs Leb(band), which scales like ρ^D? If so, the risk-optimal choice is to inflate N̂ rather than certify zero inflation.
  • What data and objective were the Optuna dial books tuned on? How do the results change on the 165 non-preview series, with every typing and smoothing decision taken from the training prefix and the same post-processing given to the baselines (including [10–12])? Why is StreamVAE, which your notebook lists among "the leaderboard's strongest non-pretrained methods", excluded from Table 1 as pretrained?
  • Can you show one matched-budget experiment where envelope negatives beat uniform negatives for a learned head (e.g., without the 99.5% weight clipping), and one experiment that satisfies all of Theorem 4.5's hypotheses? On real data, does performance actually peak near 1/e occupancy when d_z is swept, given that the experiments fix d_z (8, or 10 for CUBE CERTIFY) and per-coordinate occupancy is ≈1/(1+2δ) by construction?
1 comment
AuthorsOpenReview ↗

Your falsification test ran half the procedure and confirmed Proposition 3.3: coverage is carried by the positives. Four experiments: the sweep fires on demand, error rate measured, certificate verified in-hypothesis.

Thank you — this is the deepest reading the paper has received, and because you re-implemented the notebook, every disagreement below comes with an experiment (A–D, scripts included).

  1. Your falsification test ran half the procedure — and the half it ran confirmed one of our theorems. The certificate (Def. 4.4) makes two promises, carried by two data sources. Coverage — the box sits inside the truth, each edge accurate to ε_box - is carried by the positives, through line 1's min/max box and Lemma C.8. Resolution — no outward padding t̂ is supported by evidence is carried by the negatives, and is the only thing lines 2–6 check. Your 30/30 ran lines 2–6 alone, on a box the full algorithm cannot produce: line 1's box contains every positive, so with data on [0.1, 0.9]^D your box [0.3, 0.7]^D is unreachable (probability ~10^(−9,633)). Experiment A reruns your geometry with line 1 restored: the true box is recovered to within 2×10⁻³ per edge in all 30 runs, and your box is rejected by the positives before a single negative is drawn — ≈3,980 of 4,000 violate it. The part we most want you to weigh: the silence of the negatives on that wrong box is not a flaw you found. It is Proposition 3.3 — no statistic built on negatives can detect a coverage failure — demonstrated empirically. Coverage was never the sweep's job; we prove nothing on that side can do it.
  2. The sweep's "always zero" is a test passing, not a test that cannot fail. The verdict is a median over 41 blocks and returns t̂ > 0 whenever a majority cannot verify a band. Experiment B produces that state on demand: starve the negative budget and it fires in 100% of runs (N ≤ 1,600), silent at the paper's N = 8,000; corrupt the sampler's length scale by 4× and it fires at full budget. With everything correct: 0 false triggers in 300 runs at the certified budget, against a guaranteed 5%. The zeros you observed are a working detector reporting that its failure mode is absent.
  3. The t̂ ≡ 0 ablation you asked for is exact — and we concede its result. t̂ enters scoring once, G(x) = Σ_d(dist_d(x) − t̂)₊, so whenever t̂ = 0 the deployed and ablated detectors are bit-identical. The archive shows t̂ = 0 on 65/65 series; the rerun gives max |Δscore| = 0.0; both harnesses now bank t̂ per series. Plainly: on these benchmarks the sweep's value is the α-controlled receipt that the tight boundary stands, not the padding itself. The padding path is real and activates when evidence thins (Experiment B).
  4. Your risk-optimality question improved the paper. The calculation is right: at D = 128 the uncovered mass (~6.2%) dwarfs a band's volume (~10⁻¹⁴), so the risk-optimal padding t* ≈ s/(M+1) is positive. But t* is of order ε_box - inside the slack the coverage clause already prices - and Prop. 3.3 is exactly why negatives cannot see it. Graded-distance scoring is why AUC stays 1.000 regardless. We add your observation as a remark, with an ε_box-padding variant.
  5. "The envelope never helps." On accuracy it cannot: the AUC ties (≤0.0004 over 180 series) are what Corollary 3.5 predicts for any learned head - demanding envelope-beats-uniform there asks us to refute our own theorem. The exponential-to-constant claim is about verification cost, now measured at matched budget: given the same 557,354 draws at D = 64, uniform negatives verify 0 of 8 certificate bands; the envelope verifies 8 of 8. Flights: conceded — confounded (41 vs 72 features); the abstract becomes "indicative" and a matched-feature arm is added.
  6. Theory and novelty. Eq. (2) is indeed vacuous at M = 4,000; the ties follow from the corollary's M-free kernel identity, and we reword. cbcla is deliberately the learned-energy rung Cor. 3.5 covers; a warrant table maps each rung to its backing result. Experiment D runs Theorem 4.5 with every hypothesis and budget met (D = 8/16; M_k = 1,832/8,037, N_k = 557,354): the certificate held in 193/200 and 100/100 runs against the 95% guarantee, all failures arriving through the coverage clause at 3.5% — inside its 5% allocation. The cited lines (Tax & Duin, Hempstalk, Désir, DROCC, NPOS) enter §1–2; none contain the closed-form window, the ancillarity bound, exact O(D) sampling, or the certificate.
  7. Hygiene. Dial books were tuned by Optuna on TSB-AD's official tuning split, disjoint from the 180 evaluation series — now documented. Preview exposure: conceded; the revision reports the 165-series split, adds HBOS/COPOD/ECOD under the identical stack and budget, with per-arm post-processing deltas.

Closing. The vacuity objection dissolves once the two clauses are kept apart: coverage lives with the positives, the sweep's failing state is reachable on demand, its error rate is measured at its guarantee, and the certificate holds where its budgets are payable. If this resolves your concerns, we would be grateful if you would reconsider the rating - your Part 4 question made the paper better.

Reviewer_zcmvOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the theoretical foundations and high-dimensional limits of negative-sampling anomaly detection (NSAD). The authors prove that synthetic negative samples are ancillary—they define the target density level but provide zero boundary information, which must come entirely from positive samples. Under a box model of normal data, the paper shows that uniform negative sampling only functions within a narrow dimension window governed by a 1/e≈36.8%1/e \approx 36.8\% occupancy law, beyond which uniform negatives miss the boundary exponentially fast. To overcome this, the authors introduce shell-concentrated and Langevin-based negative samplers with dimension-free hit rates, along with an end-to-end finite-sample certificate for axis-aligned box estimators. These components are organized into an eight-rung ladder of detectors and evaluated across synthetic suites, 180 multivariate series from TSB-AD-M, and a NASA flight-test telemetry dataset.

Critical strengths
  1. Negative sampling has been deployed in industrial fault detection for years without a clear theoretical justification. Deriving the exact operable dimension window, the 1/e1/e occupancy law, and formalizing the ancillarity of negatives via Le Cam's two-point method provides genuine clarity on where and why uniform negative generation breaks down.
  2. The dimension sweep in Figure 2 provides a crisp sanity check for the theory. When replacing uniform negatives with shell negatives in an NS-NN, the sampling efficiency becomes dimension-free, yet the neural classifier still collapses past D≈16D \approx 16. This cleanly demonstrates that improving negative sampling reallocates variance but cannot inject missing boundary information.
  3. The authors report the full 180-series evaluation on TSB-AD-M with paired bootstrap confidence intervals, openly discussing where the method trails sequence models (such as xLSTMAD on range-aware metrics) and where box assumptions fail (such as rotated Gaussian structures collapsing to 0.50 AUC).
Critical weaknesses
  1. Much of the formal development (Theorems 3.6, 4.2, and 4.5) relies strictly on Assumption 3.1 and the multimodal box model (M), where normal operation is assumed to be an axis-aligned hypercube or disjoint boxes. The synthetic experiments in Section 6.2 show that as soon as data exhibits correlation (rotated Gaussians), the axis-aligned certified box drops to chance (0.50) on structural anomalies. While the authors propose latent autoencoder projections (B6) and quantile lenses (B5) to mitigate this, the finite-sample certificate of Theorem 4.5 does not formally hold in those learned, correlated latent spaces.
  2. In Appendix F (lines 1542–1544), the authors note that at D=128D = 128, the sufficient sample sizes required by Theorem 4.5 evaluate to M≥6×105M \ge 6 \times 10^5 positives and N≥5×105N \ge 5 \times 10^5 negatives. In practice, the experiments use only M=4,000M = 4,000 and N=8,000N = 8,000 and still achieve zero-inflation selection in 65 out of 65 runs. Because the theoretical sample complexity bounds are conservative by several orders of magnitude, their value as practical sizing rules for deployment remains limited.
  3. On TSB-AD-M, every evaluated model is technically a static point detector, but the pipeline applies a fairly intricate set of hand-crafted rules beforehand (Appendix H: tracking rolling event rates, gap ratios, first differences, piecewise-constant baselines, sentinel dithering) followed by a causal rolling maximum over an ACF period. In lines 1708–1710, removing the rolling smoother alone drops VUS-PR from 0.324 to 0.292. It is difficult to tell how much of the performance on TSB-AD-M comes from the negative-sampling mechanics versus this specialized time-series preprocessing and filtering stack.
Questions for authors
  1. Theorem 4.5 guarantees coverage and resolution for axis-aligned boxes under independent coordinate assumptions. When running CANONICAL on the latent representations of an autoencoder (AE_LATENT), the latent dimensions are inevitably correlated and rarely follow uniform or flat densities. Does any part of the finite-sample certificate survive in the latent space, or does it become a purely heuristic procedure there?
  2. At D=128D = 128, why is there an order-of-magnitude discrepancy between the required theoretical budget (M≈6×105M \approx 6 \times 10^5) and the empirical sample size (M=4,000M = 4,000)? Which bounding step introduces the bulk of this slack—the union bound across edges, the Chebyshev/Hoeffding steps in the median-of-means estimator, or the envelope tail bound?
  3. How does CANONICAL perform on TSB-AD-M if applied directly to standard normalized time series, without the rule-based channel-typing pass (the event-rate conversions, gap ratios, and setpoint tracking described in lines 1694–1705)? An ablation isolating the raw detector from this domain-specific preprocessing would help clarify its baseline capability.
  4. The introduction emphasizes that negative sampling is attractive in industrial diagnostics because Integrated Gradients can attribute alarms to individual sensors. However, the leading non-neural certified rungs (CANONICAL, LATENT_CANONICAL) do not train a classifier. How are per-sensor attributions generated for these box-based rungs in practice?
1 comment
AuthorsOpenReview ↗

Certificate scope mapped across raw, lens, and latent spaces; budget slack measured and itemized; equalized preprocessing ablations committed; box attributions shown to be exact.

Thank you for a careful, constructive reading. Your account of Figure 2 — better negatives "reallocate variance but cannot inject missing boundary information" — is exactly the paper's central message. We take your weaknesses and questions in order; Experiments D and E were run during this discussion, and scripts accompany the revision.

  1. Scope, and what survives in learned spaces (W1, Q1). One clarification first: Theorem 4.5 assumes flatness — the density stays within a factor κ_P of uniform on each box mode — not independent coordinates. The real restriction is the box-shaped support. The certificate makes two promises: coverage (the box sits inside the truth with accurate edges), from the positives, and resolution (no outward padding is supported by the evidence), from the negatives. They degrade differently. (i) Raw sensors: the full certificate holds, and Experiment D now runs it with every hypothesis and budget met at D = 8 and 16 — the certificate event held in 193/200 and 100/100 runs against the guaranteed 95%. (ii) Quantile lens: the certificate survives, with one substitution (a proposition in the revision). The lens is a frozen, monotone, per-coordinate map, so it carries boxes to boxes and commutes with the min/max — coverage transfers verbatim. And since the lens makes every marginal uniform, flatness becomes a bound on the copula density, κ̃_P: the remaining exposure is dependence strength. A rotation is precisely the dependence the lens cannot flatten — hence the 0.50 rotated-Gaussian result, consistent with the stated scope. (iii) Autoencoder latents: we agree, with one refinement: the promises do not fail together. Resolution survives in any representation, because it concerns only objects we construct (latent boxes, the envelope, exact draws). Coverage needs flatness of the latent law, which is neither guaranteed nor verifiable; "heuristic" is the accurate word there. A per-rung warrant table in the revision states which rungs carry which guarantee.

  2. Budget conservatism (W2, Q2). Agreed: these are not sizing rules at D = 128. Their value is structural — polynomial budgets with a dimension-free negative side, against the exponential cost of uniform sampling. Two results bracket the gap. Experiment D shows the guarantee exercised where payable, failing only through the coverage clause at 3.5% against its 5% allocation — so the positive budget is nearly tight at D = 8. Experiment E answers "which step": the slack concentrates in the per-block draw requirement, 6,501 certified versus about 50 sufficient (≈129×), which splits into worst-case envelope-geometry constants (27–43× at D = 8, measured directly) and the Chebyshev step (3–5×: on the collar the weights are nearly constant, so the statistic is nearly Bernoulli). The block count K_B = 52 is not slack; e^{1/2} ≈ 1.65 is modest; and the positive side carries one removable D/m factor (edges proved to ρ₋/(4D²) where ρ₋/(4Dm) suffices) — 16× at D = 128. End to end, ≈116× at D = 8, consistent with your 60–150×. The revision adds the decomposition table and reframes the budgets as structural.

  3. Preprocessing versus detector (W3, Q3). The typing pass is representation repair: event channels become rates, counters become increments, setpoints become baseline deviations. It is train-prefix-decided, label-free, and applied identically to every rung, so internal comparisons are unaffected. Your question is the external one, and there we agree the attribution is not separable today (our own smoother ablation: 0.324 → 0.292). We commit to: (a) the factorized ablation you ask for — raw, then +typing, +guards, +smoother — on all 180 series; (b) the converse test: HBOS, COPOD, and ECOD given the identical stack and an equal tuning budget — if the stack explains the result, stacked-HBOS will match stacked-CANONICAL; (c) per-arm post-processing deltas in one table.

  4. Attribution (Q4). For the box rungs it is exact. The score is G(x) = Σ_d (dist_d − t̂)₊ — a sum over sensors of each sensor's exceedance past the certified limit — so the attribution to sensor d is its own term. Completeness holds by definition; Integrated Gradients from any in-box baseline returns exactly these terms (the slope is 1 outside the limit, 0 inside); and the units are physical: "this sensor is this far past its certified limit." The lens pulls attributions back one-to-one to named sensors; the autoencoder rungs use IG through the frozen encoder. A revision subsection states this.

Closing. Your concerns are, we believe, the fixable kind you marked — and several (the slack table, lens proposition, attribution subsection) made the paper better. If these answers resolve them, we would be grateful if you would consider strengthening your assessment; the Question 2 decomposition can be expanded to any depth on request.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.