{"id":"0bfa0ab6-f24b-4a2a-84ef-31821eafe4ea","arxiv_id":"2506.07760","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Adaptive intervention for quickest causal change point detection is proposed, but the KL divergence formula underpinning the claimed first-order optimality contains an error.","lead":"This paper proposes an online algorithm that detects changes in linear causal systems by actively choosing which variables to intervene on, using a transformation that concentrates a change signal into a single dimension. The authors claim the intervention strategy is asymptotically delay-optimal, but a mathematical error in the key KL divergence formula invalidates the stated optimality theorem.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (11) adds a spurious +1/2 to the KL divergence; because Theorems 1–3 use this quantity as the rate constant, the claimed first-order optimality is not established.","rationale":"The single most load-bearing point is the KL divergence in Eq. (11), because it is the denominator in every optimality statement. I independently recomputed the KL from Proposition 1's post-change distribution: the covariance has determinant 1 and the trace correction exactly cancels the '+1/2' constant, so Eq. (11) is off by an additive half-nat. Since c_j is fixed, this is not an asymptotically negligible error in γ, and the stated rate constant cannot be right. The reader's rationale identified the same equation, though their formal 'weakest_assumption' field focused on Assumption 1; hence partial agreement. I found no separate fatal flaw in the algorithmic idea, and the simulations look internally plausible, but the main theoretical guarantee is invalid as written. A corrected paper might be salvageable by replacing I with the true KL and re-proving the theorems, but the present version does not support the headline optimality claim. Therefore I would keep the reader's REJECT verdict.","tokens_in":45406,"tokens_out":12200,"duration_ms":150541,"concrete_test":"Recompute Eq. (11) in the toy model p=2, A=0, μ=(1,0)^T, Σ=I, with change [A]_{2,1}=Δ and intervention on node 1 with value c_1. Proposition 1 gives Y_2^{do(1)} ~ N(Δ c_1 / σ_2, 1), so the true KL is Δ^2 c_1^2 / (2σ_2^2), while Eq. (11) returns that value plus 1/2. For Δ=0.5, c_1=2, σ_2=1, this is 0.5 vs 1.0. Then simulate Multi-AI with b=log(100): the paper's claimed asymptotic delay is log(100)/1 ≈ 4.6, but the corrected lower bound log(100)/0.5 ≈ 9.2 is the minimum achievable at ARL=100; Monte Carlo will show the EDD is near 9.2, confirming the rate constant must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that Max-AI and Multi-AI achieve first-order optimality with detection delay log γ / I^{Δ,[k,j]}_{j,c_j}. The definition of I in Eq. (10)–(11) is the KL divergence between the post-change and pre-change centralized distributions. For the post-change distribution in Proposition 1, Y ~ N(μ, Σ) with μ = σ_k^{-1} Δ (μ_j^{do} + Σ_{l∈anc(j)} [B^{do}]_{j,l} μ_l^{do}) E_k, and Σ satisfying det Σ = 1 and tr Σ = p−1 + σ_k^{-2}Δ^2 Σ_{l∈anc(j)∪j} ([B^{do}]_{j,l})^2 σ_l^2. The Gaussian KL formula D(N(μ,Σ)||N(0,I)) = 1/2(tr Σ − (p−1) + ||μ||^2 − log det Σ) gives exactly the first two terms of Eq. (11), with no '+1/2'. The stated Eq. (11) therefore overstates every KL number by a constant 1/2. This error propagates directly into Theorem 1's lower bound and Theorems 2–3's upper bounds, all of which use I^{Δ,[k,j]}_{j,c_j} as the denominator. Since c_j is a fixed parameter for fixed problem data, the +1/2 does not vanish as γ→∞, so the claimed rate constant log γ / I is not achieved; the correct bound would involve I−1/2. Section 7.4's simplified expression I = 1/2(1 + c_j Δ^2/σ_k^2) also shows the spurious constant and is dimensionally inconsistent. Thus the central optimality guarantee is unsupported as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies online change point detection in linear causal structural equation models when the observer can intervene on nodes. It introduces a centralization transform that concentrates a single-edge change in the structural matrix into one coordinate, proposes an intervention-value rule (Algorithm 1), and presents two detection schemes, Multi-AI and Max-AI, with adaptive intervention policies that balance exploration and exploitation. The central theoretical claim is first-order asymptotic optimality: the worst-case detection delay satisfies J ~ log γ / I^{Δ,[k,j]}_{j,c_j} as γ→∞ (Theorems 2 and 3), with a matching lower bound in Theorem 1. The paper supports the claims with extensive simulations and two case studies.","tokens_in":45717,"tokens_out":21376,"duration_ms":262794,"significance":"The problem is well motivated, and the centralization idea is elegant: if the theory were correct, the paper would be a useful first step connecting causal intervention design with quickest change detection. Strengths include the single-dimension concentration property, the exploration-exploitation node-selection framework, and a broad simulation study covering graph size, sparsity, change magnitude, exogenous variance, and multiple-change scenarios, together with two real-data case studies. However, several load-bearing algebraic and conceptual issues in the theoretical sections prevent acceptance in the current form; the core rate constant and the intervention-value rule are not established as written.","major_comments":[{"comment":"The KL divergence in Eq. (11) is algebraically incorrect. For the post-change centralized distribution in Proposition 1, the determinant of the covariance matrix is 1, so the Gaussian KL formula gives D = 1/2[ σ_k^{-2}Δ^2( μ_j^{do} + Σ_{l∈anc(j)} B_{j,l} μ_l^{do} )^2 + σ_k^{-2}Δ^2 Σ_{l∈anc(j)∪j} (B_{j,l})^2 Σ_l ], with no '+1/2' and with the square on B in the second term. This quantity I is the rate constant in Theorems 1-3; the spurious +1/2 does not vanish as γ→∞, so the asserted first-order rate logγ/I is not the true rate. Section 7.4's expression I = 1/2(1 + c_j Δ^2/σ_k^2) is also inconsistent with both Eq. (11) and the corrected formula.","section":"Section 4.3, Eq. (11)"},{"comment":"The dominance condition is incompletely stated. For i∈anc(j)∪0, the inequality I_{j,c_j} − I_{i,c_i} > δ yields c_j² > ( μ_j^{do(i)} + Σ_{l∈anc(j)} B_{j,l} μ_l^{do(i)} )² + Σ_{l∈anc(j)} (B_{j,l})² Σ_l + σ_j² + 2δσ_k²/Δ². Eq. (42) and Algorithm 1 omit the μ_j and σ_j² terms. Consequently the computed intervention values C may violate Proposition 2, and Lemma 3 and Corollary 1, which are used in the optimality proofs, are unsupported as written.","section":"Section 4.3, Eq. (42) and Algorithm 1"},{"comment":"The claimed optimality is relative to intervention values fixed by Algorithm 1, not global optimality over intervention values. The lower bound is taken over detection methods (A,T) in Cγ while C is held fixed, and Section 6.4 itself notes that increasing c_j reduces the detection delay. Without an intervention-cost model the problem is ill-posed because c_j is unbounded; with such a model, Algorithm 1 has no optimality guarantee. The abstract and Section 6 should qualify the optimality claim as optimality among policies that use the proposed intervention-value rule.","section":"Section 6.1, Theorem 1, and Section 6.4"}],"minor_comments":[{"comment":"Eq. (12) defines aopt as arg min, but the surrounding text and the definition in (10) require arg max; this appears to be a typo.","section":"Section 4.3, Eq. (12)"},{"comment":"The marginal KL identity is missing a factor 1/2: the correct formula is I_{i,c_i}[k] = I_{i,c_i} − (1/2) log([Cov(Y^{t,do(i)})]_{k,k}) for i≠k. The error is in the proof of Proposition 2; it is conservative for inequality (16), but the proof as written is not correct.","section":"Appendix B, Eq. (43)"},{"comment":"The theorem statement says b = logγ, while Lemma 5 and the proof require b = logγ + logp to guarantee (A*, T^max)∈Cγ; the statement and proof should be reconciled.","section":"Theorem 3"},{"comment":"The displayed formula I = 1/2(1 + c_j Δ^2/σ_k^2) also omits the square on c_j that would follow from the corrected KL computation; it should be 1/2 c_j² Δ^2/σ_k² (up to the same correction as Eq. (11)).","section":"Section 7.4"},{"comment":"There are two items numbered Definition 1 (centralization and sequential intervention policy); the second should be renumbered.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The algebraic errors are load-bearing but appear fixable: correcting Eq. (11), the dominance condition in Algorithm 1, and the marginal KL identity, and then re-verifying the proofs, could make the main claims valid. The paper would need a careful rewrite of Sections 4.3 and 6 and a rerun of the simulations after Algorithm 1 is corrected. I do not see a novelty or attribution concern beyond the ordinary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague – quick read on Xu and Zhang.\n\nThe core idea is worth a look: they build an online change-detection scheme for linear causal models where the controller intervenes on one node each round, and they design intervention values so that, whatever edge changes, the origin node becomes the best node to intervene on. The centralization transformation (whitening by the pre-change intervention distribution) is elegant, and it gives a clean single-coordinate concentration of the change. That seems genuinely new relative to the offline causal CPD work they cite, and the simulation comparisons against no-intervention and random-intervention baselines show the adaptive intervention is doing real work. I'd want to see code, but the empirical story is plausible.\n\nThe problem is that the theory is not just shaky, it's built on a wrong KL formula. Equation (11) defines the KL between the centralized post-change and pre-change normals. For the Gaussian they derive in Proposition 1, the covariance has determinant 1, so the KL is 1/2 times the squared mean plus the trace of the variance part. That gives 1/2 σ_k^{-2} Δ^2[(μ_j + Σ_l [B]_{j,l} μ_l)^2 + Σ_l ([B]_{j,l})^2 σ_l^2]. Their Eq. (11) has two errors: a spurious +1/2, and the second term uses [B]_{j,l} σ_l without the square. The same error shows up in their simplified expression I = 1/2(1 + c_j Δ^2/σ_k^2) in Section 7.4, which is dimensionally inconsistent. This is load-bearing: Theorems 1–3 use this I as the rate constant in log γ / I. With the inflated I, the claimed lower bound is weaker than the true one, and the upper bound claims a delay faster than the true KL would allow. So the first-order optimality results are not established.\n\nThere's also a more conceptual soft spot. Theorem 1's lower bound is over methods that use the intervention values fixed by Algorithm 1, so the \"optimality\" is conditional on their own design choice. That is a legitimate relative optimality, but it isn't global optimality over intervention values, and the paper's language oversells it.\n\nBottom line: the algorithmic idea is worth publishing if the authors fix the KL algebra and either repair the optimality proofs or reframe the contribution as a heuristic with empirical support. I'd send this to a serious referee as is, because the novel idea and the well-executed centralization deserve scrutiny, but I'd be surprised if the theoretical claims survive contact with the correct formula.","headline":"The adaptive intervention idea is genuinely new and the simulations are compelling, but the central KL divergence formula is algebraically wrong, so the asymptotic optimality theorems do not hold as written.","tokens_in":46339,"tokens_out":5180,"would_cite":true,"duration_ms":57176,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62L10"],"pacs":[],"model":"deepseek-v4-flash","headline":"By intervening on the origin of a changed causal edge, this paper proves that a centralized CUSUM-type statistic attains the first-order optimal detection delay $\\log \\gamma / I^{\\Delta,[k,j]}_{j,c_j}$, matching the lower bound for any…","keywords":["linear causal model","sequential change point detection","adaptive intervention","centralization","Kullback-Leibler divergence","quickest detection","CUSUM","exploration-exploitation"],"falsifier":"Compute $I^{\\Delta,[k,j]}_{j,c_j}$ for the $p=6$, $d=2$, $\\Delta=0.1$ simulation of Section 7.3 and compare Max-AI's measured expected detection delay at ARL $=10^4$ with $\\log(10^4)/I^{\\Delta,[k,j]}_{j,c_j}$; if the ratio stays well above 1 as $\\gamma$ grows, Theorem 3's first-order optimality fails. A second check: with two simultaneous edge changes, if Max-AI still attains the single-change benchmark $\\log \\gamma / I^{\\Delta,[k,j]}_{j,c_j}$, then Proposition 2's concentration property is not the mechanism that carries the claimed lower bound.","tokens_in":45089,"feed_emoji":"🎯","tokens_out":11910,"duration_ms":106945,"temperature":0.7,"pith_summary":"The paper claims that in a linear causal model with a known graph, the quickest way to detect a change is not to passively observe but to intervene: set one node to a chosen value at each time step. Its centralization transform makes pre-change observations standard normal and concentrates a single-edge change into one coordinate, so the problem reduces to detecting a mean shift of size $\\Delta c_j / \\sigma_k$. A Kullback-Leibler-based rule for choosing intervention values guarantees that the origin of the changed edge is always the best node to intervene on, even though the algorithm does not know which edge changed. Two online monitoring schemes, Multi-AI and Max-AI, pair this transform with an exploration-exploitation policy and are proven to achieve first-order asymptotic optimality: their worst-case expected detection delay approaches $\\log \\gamma / I^{\\Delta,[k,j]}_{j,c_j}$ as the false-alarm constraint $\\gamma$ grows, matching the information-theoretic lower bound. This matters because intervention turns causal structure into an amplifier: small changes that passive detectors would miss can be detected with provably minimal delay.","feed_headline":"Adaptive intervention hits the causal-change detection speed limit","feed_subtitle":"Intervening on the changed edge's origin amplifies small causal shifts and provably minimizes worst-case delay.","key_machinery":"The central object is the centralization transform of Definition 1: after an intervention $\\mathrm{do}(X_i=c_i)$, the non-intervened coordinates are mapped to $Y = [\\Sigma^{\\mathrm{do}}]^{-1/2}(I - A^{\\mathrm{do}})(X^{\\mathrm{do}} - \\mu^{\\mathrm{do}}_x)$. Pre-change this is exactly standard normal; post-change, for an edge-weight perturbation in $[A]_{k,j}$, Proposition 1 shows the mean becomes a single nonzero coordinate at $k$ while the covariance stays close to identity, and intervening on $j$ makes the covariance exactly identity. Algorithm 1 then fixes intervention values $c_j$ so that the KL divergence from intervening on any other node is dominated by a gap $\\delta$ for every possible change origin, which is what makes the arg-max intervention rule of Corollary 1 recover the true origin $j$ without knowing which edge changed. The detection statistics are window-limited CUSUM: Multi-AI estimates full mean-covariance pairs over a window of length $O(p^2)$, while Max-AI estimates only per-coordinate means and variances over $O(p)$ and takes the maximum over $p$ separate CUSUM statistics, exploiting the one-dimensional concentration.","core_discovery":"The central claim, stated on the paper's own terms, is that a single-element change in a linear structural equation model can be detected at the quickest possible rate by actively intervening on the node where the changed edge originates. Under Assumption 1, the centralization transform of Definition 1 sends the post-change distribution of the centralized vector into a normal with mean concentrated on the target node $k$ and covariance a rank-one perturbation of the identity; intervening on the origin $j$ of the changed edge makes that covariance exactly identity and leaves only a scalar mean shift $\\Delta c_j / \\sigma_k$. Proposition 2 then establishes a strict dominance of KL divergences: for any change location $[k,j]$, the divergence $I^{\\Delta,[k,j]}_{j,c_j}$ exceeds every other intervention's divergence by at least $\\delta$, and all change information is confined to one dimension. Because of that dominance, the adaptive policies in (24) and (28) identify the optimal intervention node automatically, and Theorems 2 and 3 show the resulting window-limited CUSUM statistics are first-order optimal, i.e., $J(A,T) \\sim \\log \\gamma / I^{\\Delta,[k,j]}_{j,c_j}$ as $\\gamma \\to \\infty$, with Max-AI achieving this with an $O(p)$-length window and Multi-AI with an $O(p^2)$-length window but better tolerance for multiple changes.","pith_inferences":["Beyond the paper, the same concentration identity suggests a local version for nonlinear SEMs: around a chosen intervention value, linearization would give an approximate one-dimensional statistic, but no optimality guarantee would carry over.","A natural testable extension is budgeted intervention: Algorithm 1 currently ignores intervention cost, and Section 6.4 only sketches the trade-off, so an adaptive $\\delta$ or $c_j$ rule under a budget could be benchmarked against Theorem 1's bound.","When the graph itself is unknown, one could combine structure learning with this detector, but the concentration property of Proposition 2 would need re-derivation, since misspecified ancestry destroys the identity covariance of the centralized statistic.","The paper's one-change assumption is the real boundary: a practical deployment would run a multi-change diagnostic first, or choose Multi-AI, because Max-AI's optimality silently relies on a single changed edge."],"forward_implications":["Under Assumption 1, Max-AI and Multi-AI reach the asymptotic lower bound, so no detection method using these intervention values can detect the change faster in first-order terms.","The delay is inversely proportional to $I^{\\Delta,[k,j]}_{j,c_j}$, and since Algorithm 1 lets $c_j$ grow with the gap $\\delta$, users can trade larger intervention values for shorter detection delay.","Multi-AI is the safer default when multiple edges may change: its optimality does not rely on single-dimension concentration, and Section 7.6 shows it beats Max-AI in multi-change scenarios.","No-intervention and random-intervention counterparts (MAX-NI, MULTI-NI, MAX-RI, MULTI-RI) have strictly smaller KL denominators and therefore cannot reach the lower bound, isolating adaptive intervention as the source of the speedup.","The method transfers to real decision settings where a node's value can be set directly, as demonstrated in the ecological and psychological case studies."],"supporting_citations":[{"why":"Supplies the information-theoretic lower-bound framework used in Theorem 1 for quick detection of parameter changes.","marker":"Lai (1998)"},{"why":"Provides the quickest-change-with-controlled-sensing setting and the Proposition 1 (Lemma 9 here) that the adaptive intervention policies extend to continuous post-change parameter spaces.","marker":"Fellouris and Veeravalli (2022)"},{"why":"The CUSUM-with-estimated-parameters template on which the Multi-AI and Max-AI detection statistics are built.","marker":"Lorden and Pollak (2008)"},{"why":"Introduces the window-limited CUSUM construction that motivates the finite time windows used by both monitoring schemes.","marker":"Xie et al. (2023)"},{"why":"Supplies the optional stopping theorem used in proving the false-alarm (ARL) guarantees in Lemmas 1 and 5.","marker":"Tartakovsky et al. (2014)"}],"fun_headline_variants":["Intervening on the cause speeds up change detection","Adaptive interventions make causal change detection optimal","Proven fastest detection by intervening on the source","Pick the right node to amplify causal shifts quickly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire concentration and optimality argument rests on Assumption 1: the post-change model differs from the pre-change model in exactly one element of exactly one component, with the change magnitude $|\\Delta|$ between known bounds. If this fails, for instance if two edges change at once, the single-dimension concentration property of Proposition 2 breaks and Max-AI's first-order optimality is no longer guaranteed, a degradation the paper's own multi-change simulation confirms.","fun_headline_variants_meta":{"raw":{"variants":["Intervening on the cause speeds up change detection","Adaptive interventions make causal change detection optimal","Proven fastest detection by intervening on the source","Pick the right node to amplify causal shifts quickly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1197,"prompt_tokens":941,"completion_tokens":256,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":197}},"tokens_in":557,"tokens_out":256,"duration_ms":3421,"temperature":1.0,"reasoning_tokens":197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:30:07.914563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $I^{\\Delta,[k,j]}_{j,c_j}$ for the $p=6$, $d=2$, $\\Delta=0.1$ simulation of Section 7.3 and compare Max-AI's measured expected detection delay at ARL $=10^4$ with $\\log(10^4)/I^{\\Delta,[k,j]}_{j,c_j}$; if the ratio stays well above 1 as $\\gamma$ grows, Theorem 3's first-order optimality fails. A second check: with two simultaneous edge changes, if Max-AI still attains the single-change benchmark $\\log \\gamma / I^{\\Delta,[k,j]}_{j,c_j}$, then Proposition 2's concentration property is not the mechanism that carries the claimed lower bound.","supporting_citations":[{"cited_title":"Information bounds and quick detection of parameter changes in stochastic systems","cited_arxiv_id":null,"evidence_quote":"Supplies the information-theoretic lower-bound framework used in Theorem 1 for quick detection of parameter changes."},{"cited_title":"Sequential change-point detection procedures that are nearly optimal and computationally simple","cited_arxiv_id":null,"evidence_quote":"The CUSUM-with-estimated-parameters template on which the Multi-AI and Max-AI detection statistics are built."},{"cited_title":"Window-limited cusum for sequential change detection","cited_arxiv_id":null,"evidence_quote":"Introduces the window-limited CUSUM construction that motivates the finite time windows used by both monitoring schemes."},{"cited_title":"Sequential analysis: Hypothesis testing and changepoint detection","cited_arxiv_id":null,"evidence_quote":"Supplies the optional stopping theorem used in proving the false-alarm (ARL) guarantees in Lemmas 1 and 5."}],"review_version":1}