{"id":"6b290b88-cc64-49dc-bda0-2a087ee144f8","arxiv_id":"2502.04530","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A moment-matching Erlang-mixture method approximates cumulative reward distributions in discrete-time Markov chains, enabling chance-constrained model checking on continuous reward spaces.","lead":"Probabilistic model checking usually verifies only the expected value of a system's reward. This paper proposes approximating the entire reward distribution of Markov-chain models with mixtures of Erlang distributions matched to high-order moments, so that engineers can verify chance constraints such as 'the reward exceeds a threshold with at most a given probability,' including for continuous reward domains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem IV.1's sup-norm approximation claim is false for atomic reward distributions; the proof establishes only weak convergence, so the advertised 'theoretically bounded error' is not delivered.","rationale":"The paper describes a useful empirical heuristic, but the central theoretical claim does not hold. The reader's REJECT is justified. My concern is more fundamental than the reader's finite-K point: even in the limit of infinite moments and components, no uniform sup-norm approximation of atomic CDFs by continuous Erlang mixtures is possible, so the 'theoretically bounded error' is not just unproved but false as stated. The concrete counterexample isolates the flaw in Theorem IV.1. The paper could potentially be salvaged by changing the error metric (e.g., Wasserstein or Levy-Prokhorov) and by adding a decision procedure with explicit conservative bounds, but as written it does not deliver robust model checking. Hence the verdict remains REJECT.","tokens_in":18809,"tokens_out":7569,"duration_ms":75650,"concrete_test":"Build the two-exit DTMC: s0 -> s1 with probability 0.5 and reward 1, s0 -> s2 with probability 0.5 and reward π, with s1,s2 absorbing. The true CDF has jumps at 1 and π. Apply the δ-discretization of Theorem IV.1 for an arbitrary δ>0. If the jump at 1 is exact (ceil(1/δ)δ=1), then the jump at π is shifted because π/δ cannot be integer; taking x=π gives F(π)=1 while F~_D(π)=0.5, error 0.5. If the jump at 1 is shifted, the error is already >=0.5 for x just above 1. Thus sup-norm error >=0.5 for all δ, contradicting the claimed bound for ε<0.5. This is an analytic check requiring no simulation; it settles whether Theorem IV.1 is correct as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract promises 'theoretically bounded error' for the Erlang-mixture approximation, and Section III-B builds a chance-constraint decision procedure on that bound. The supporting Theorem IV.1 claims sup_{r>=0} |F(r)-F~_D(r)| <= ε for a DPH approximation after reward discretization. The proof, however, only establishes weak convergence: after bounding P(|R~_T-R_T|>=ε) <= E[T]δ/ε, it concludes lim F~_D = F at continuity points. Weak convergence does not imply uniform sup-norm convergence. The claim is in fact false for atomic reward distributions, which are the paper's own discrete examples (Section II-B assigns each state a constant reward). For a DTMC with two absorbing exits giving reward 1 (prob 0.5) and π (prob 0.5), any single δ shifts at least one jump of the DPH CDF, so sup-norm distance is at least 0.5 for every δ. Since Erlang mixtures are continuous, a uniform sup-norm bound against an atomic CDF is impossible (any continuous CDF is at sup-norm distance at least 0.5 from a two-point atomic CDF). Moreover, no finite-K or finite-n error bound is derived; Section V-B concedes that the non-convex optimizer provides no global-optimality certificate and that accuracy is not monotone in K and n. Consequently the practical procedure (K=3, n=3, shapes 3^i) has no soundness guarantee, and the decision Pr(X<=r*)>=α can be wrong if the CDF error exceeds the margin.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for distributional probabilistic model checking of DTMCs with continuous or discrete rewards. The method computes the first K moments of the cumulative reward to absorption analytically via a moment-generating-function recurrence, then fits a mixture of Erlang distributions to those moments by solving a non-convex optimization problem with a maximum-entropy regularizer. The fitted CDF is used to decide chance-constrained properties of the form Pr(X ≤ r*) ≥ α, with Cantelli's inequality offered as an early sufficient test. The paper claims a theoretically bounded approximation error, supported by three theorems on discretization of rewards, Erlangization of discrete phase-type distributions, and denseness of Erlang mixtures.","tokens_in":19067,"tokens_out":4955,"duration_ms":50753,"significance":"If the bounded-error claim were correct, the paper would offer a useful alternative to histogram-based distributional model checking, especially for models with continuous rewards, and the analytical moment computation via MGFs (Eq. 2) is a genuine contribution. The Cantelli-based early-termination test is a sound and cheap sufficient condition, and the experimental evaluation covers several nontrivial benchmarks. However, the central theoretical guarantee is not established, and the practical decision procedure inherits no formal soundness from the provided theorems. The contribution is therefore currently an empirically evaluated heuristic, not a method with the advertised bounded-error verification semantics.","major_comments":[{"comment":"The theorem states a uniform sup-norm bound, sup_{r≥0} |F(r) − F_~D(r)| ≤ ε, but the proof only establishes weak convergence: the Markov-inequality argument bounds the probability that the discretized cumulative reward differs by at least ε, and the proof concludes with pointwise convergence at continuity points. Weak convergence does not imply uniform sup-norm convergence. For atomic reward distributions, which are the paper's own discrete examples (e.g., Section II-B assigns each state a constant reward), a δ-shift of the reward lattice moves jump locations, so the sup-norm distance between the original and discretized CDFs cannot be made arbitrarily small. Since an Erlang-mixture CDF is continuous, its sup-norm distance from an atomic CDF is at least half the largest atom mass (e.g., a two-point distribution with masses 0.5 and 0.5). Consequently the 'theoretically bounded error' claimed in the abstract and used in Section III-B for the chance-constraint decision procedure is not delivered by the stated theorems.","section":"Theorem IV.1 (Section IV)"},{"comment":"No finite-K, finite-n error bound is derived. The theorems are asymptotic (δ → 0, m → ∞, or denseness as the mixture grows), whereas the practical procedure fixes K=3, n=3 and restricts shapes to the heuristic a_i = 3^i. The paper itself reports in Section V-B that accuracy is not monotone in K and n because the non-convex optimizer provides no global-optimality certificate; Table II shows, for example, D_KS = 0.17 for (K=3,n=4) but 0.26 for (K=3,n=5), and 0.07 for (K=4,n=6) but 0.09 for (K=4,n=9). Therefore the decision Pr(X ≤ r*) ≥ α can be wrong when the fitted CDF error exceeds the property's margin, and the method as deployed has no soundness guarantee. The paper should either supply a finite bound for the practical configuration or explicitly reframe the procedure as a heuristic approximation without the bounded-error claim.","section":"Sections III-A and V-A (Algorithm 1, Table II)"},{"comment":"The construction introduces n_i = k_i − 1 intermediate states for each state with discretized reward k_iδ. For states with zero reward, k_i = 0 and n_i = −1, which is invalid. Since the reward function is only assumed to be non-negative and the paper's own examples include zero rewards (e.g., the landing state with reward 0.0), the DPH discretization proof is incomplete even before the sup-norm issue is considered.","section":"Theorem IV.1, DPH construction"}],"minor_comments":[{"comment":"The pseudocode for the moment computation is garbled: f and G are defined with e^{k r(x)} rather than the MGF parameter θ, and the update formula µ_k = r^k f + Σ_{i=0}^k C(k,i) G^{(k−i)} does not match the correct recurrence in Eq. (2). Please align the pseudocode with Eq. (2).","section":"Algorithm 1, lines 1–2"},{"comment":"The location shift loc = μ − σ can be negative when the standard deviation exceeds the mean, assigning positive probability to negative rewards even though the model rewards are non-negative. This also means the moments of the shifted fitted distribution are not the moments of the original reward variable; the paper should justify the shift or restrict it to loc ≥ 0.","section":"Section V, experimental settings"},{"comment":"The coefficient d_{n,n−1} in the moment-expansion formula E[X^n_m] − E[X^n] = E[X^{n−1}] d_{n,n−1}/m + o(1/m) is never defined. Please provide the definition or an explicit reference.","section":"Theorem IV.2"},{"comment":"The generalized Cantelli bound is stated for an arbitrary parameter b, but the condition on b (e.g., b ≥ 0 and how b relates to a) is not given. For n > 2 the optimal choice of b is nontrivial; please state the range of validity of the inequality and the intended choice of b.","section":"Section III-B, Eq. (9)"},{"comment":"There are several typos and inconsistencies: 'Chebysehv's' in Section II-B, 'prosed' in Section VI, 'LeadSync' versus 'LeaderSync' in Table I, and an inconsistent use of the subscript in F_{X_i}(x; a_i, λ_i) in Eq. (8) where the mixture components already share a common rate λ. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The core theoretical claim of bounded approximation error is false as stated for atomic reward distributions, and the practical configuration (K=3, n=3, heuristic shapes) has no derived error bound. This is not a local presentation issue; it undermines the paper's central contribution. A substantial reframing as a heuristic with empirical validation, plus new finite-error analysis, would be needed to make the manuscript publishable in its current scope. The reader's report and the skeptic's analysis align with my reading of the proof of Theorem IV.1."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper has a genuinely useful idea—compute exact moments of DTMC cumulative rewards via MGFs, fit an Erlang mixture, and use that to evaluate chance constraints—but the central theoretical claim does not hold up. Theorem IV.1 states a uniform sup-norm bound on the DPH approximation, yet the proof establishes only weak convergence, and weak convergence to a CDF with atoms cannot be uniform. Since their own discrete-reward examples produce atomic distributions, the claimed bound is simply false. The paper's abstract promises 'theoretically bounded error,' and that promise is not delivered.\n\nWhat is new: the combination of MGF moment recurrence (Eq. 2 is correct) with Erlang-mixture moment matching for model checking does extend distributional analysis to continuous rewards, which histogram methods like Elsayed-Aly et al. cannot handle cleanly. The Cantelli pre-test as an early termination check is also a nice practical touch. The experiments are honest: they compare against histograms on standard PRISM benchmarks, include continuous-reward examples, and report KS statistics and runtimes. And the authors are refreshingly upfront in Section V-B that accuracy is not monotonic in K and n because the optimizer is non-convex and lacks a global-optimality certificate.\n\nThe soft spots are exactly there. No finite-K or finite-n error bound is derived anywhere. The theorems are asymptotic denseness and weak-convergence results, not the bounded-error guarantee advertised. That means the decision procedure Pr(X ≤ r*) ≥ α is unsound when the fitted CDF's error exceeds the margin. The paper acknowledges this indirectly but does not fix it. The location shift and shape heuristic ai = 3^i are presented without formal justification.\n\nIs the paper worth your time? Yes, if you work on distributional model checking or risk-sensitive verification. The MGF moment computation and the empirical comparison are useful references. But I would not cite it as a sound method until the bound is repaired or the claim is weakened to a heuristic. A serious referee should engage with it, because the idea is novel and the flaw is fixable: either prove a uniform bound for the discrete DPH case (which will require handling atoms, maybe by allowing a small slack or using a different metric), or drop the 'bounded error' wording and reframe the contribution as an empirical approximation technique. As it stands, the paper is a solid empirical method sketch with a broken headline theorem.\n\nRecommendation: send to peer review, but expect major revision. The reviewer should not desk-reject it; there is real substance here.","headline":"The empirical moment-matching approach is worth a look, but the advertised bounded-error guarantee rests on a theorem whose proof only shows weak convergence and is false for atomic reward distributions.","tokens_in":19657,"tokens_out":2731,"would_cite":false,"duration_ms":29497,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J10","62E17","65C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"A moment-matched Erlang mixture makes chance constraints on Markov reward distributions decidable with bounded error.","keywords":["probabilistic model checking","distributional verification","Erlang mixture","moment matching","Markov reward process","chance constraints","moment generating function","continuous rewards"],"falsifier":"Take a small DTMC whose exact cumulative-reward distribution is computable in closed form, fit the paper's default mixture with $K=3$, $n=3$, and shapes $3^i$, and evaluate $\\Pr(X \\le r^*)$ at a threshold where the true CDF lies just above $\\alpha$ but the fitted CDF lies just below; if such a threshold exists, the practical decision procedure returns the wrong verdict, showing that the finite-$K$ fit is not sound without an instance-specific error certificate.","tokens_in":18539,"feed_emoji":"📊","tokens_out":10798,"duration_ms":105037,"temperature":0.7,"pith_summary":"This paper tries to move probabilistic model checking beyond expected values by verifying properties of the entire cumulative-reward distribution of a discrete-time Markov chain. The proposed method matches the first $K$ moments of the true reward distribution to a mixture of Erlang distributions, with the moments computed analytically by differentiating the moment generating function of the reward-to-absorption process. Once the mixture is fitted, chance requirements of the form $\\Pr(X \\le r^*) \\ge \\alpha$ are evaluated in closed form from the mixture CDF, and a moment-based inequality can sometimes decide the requirement before the mixture is even fitted. If correct, the method gives engineers a smooth approximation of reward distributions in both continuous and discrete reward spaces, including heavy tails and multimodal behavior that histograms miss.","feed_headline":"Reward distributions replace expectations in Markov model checking","feed_subtitle":"Matching the first moments of a reward process to an Erlang mixture decides chance constraints in closed form.","key_machinery":"The machinery is a finite Erlang mixture $f_{\\mathrm{approx}}(x)=\\sum_{i=1}^n \\omega_i f_{\\mathrm{Erlang}}(x; a_i, \\lambda)$ with a common rate parameter $\\lambda$: a weighted sum of gamma-type densities with integer shapes, a family that can approximate any positive random variable to arbitrary accuracy. The argument runs through three pieces: an MGF first-step analysis that yields the $k$-th moments of cumulative reward to absorption; a truncated Stieltjes moment problem on $[0,+\\infty)$ whose objective trades moment discrepancies against differential entropy; and a fixed-shape heuristic ($a_i = 3^i$) that turns the mixed-integer optimization into a non-convex quadratic problem. The CDF of the fitted mixture is a finite sum of Erlang CDF terms, so the chance constraint $\\Pr(X \\le r^*) \\ge \\alpha$ is evaluated as a closed-form expression, with a one-sided moment bound available as an early-termination test.","core_discovery":"On the paper's own terms, the central discovery is that the cumulative reward to absorption in a DTMC can be represented, for verification purposes, by a finite mixture of Erlang distributions with a common rate parameter, where the mixture is chosen to match the first $K$ analytically computed moments of the true reward process. The moment generating function of the total reward satisfies a first-step linear system, so the moments are exact without simulation. Theorems IV.1 through IV.3 establish that this representation is dense: any positive reward distribution can be approached arbitrarily closely by Erlang mixtures, and the discretized-markov and Erlangization chain converges in distribution and in moments. The resulting density is the maximum-entropy density among those matching the given moments, and its CDF gives a direct closed-form decision procedure for chance constraints. On discrete-reward benchmarks the paper reports agreement with simulated reward distributions at least as close as histogram-based distributional model checking, and on continuous-reward models it provides a smooth distributional approximation without discretization.","pith_inferences":["Editorial inference: because the paper's experiments show that accuracy is not monotonic in $K$ and $n$ under the non-convex optimizer, a practical deployment should inspect the fitted CDF's margin against the threshold $r^*$ before trusting borderline chance-constraint verdicts.","Editorial inference: the same moment-matched Erlang mixture could serve as a distributional critic in reinforcement learning, replacing histogram and quantile critics with a smooth density that supports analytic gradients with respect to policy parameters.","Editorial inference: a testable extension would compare the fitted mixture against exact or simulated CDFs at the specific quantile used by the constraint, rather than only a global worst-case distance, since a small global discrepancy can still hide a local error large enough to flip a verdict."],"forward_implications":["Chance constraints of the form $\\Pr(X \\le r^*) \\ge \\alpha$ can be evaluated directly from the fitted mixture CDF, without simulations, for both continuous and integer reward spaces.","A one-sided moment inequality computed from the first $K$ moments can decide some requirements instantly; only when that test is inconclusive does the system need to fit the mixture.","Because moments are computed by solving $K$ linear systems of size equal to the state space, the distributional analysis carries polynomial cost in the number of states, independent of the reward range.","The method captures skewness, kurtosis, heavy tails, and multiple modalities that expected-value verification ignores, and it does not require binning or discretization of continuous rewards.","On discrete-reward benchmarks the paper reports agreement with empirical reward distributions comparable to histogram-based distributional model checking, while on continuous-reward models it handles cases the histogram approach cannot analyze cleanly."],"supporting_citations":[{"why":"Supplies the denseness result that mixtures of Erlang distributions approximate any positive random variable to arbitrary accuracy, the foundation for using the mixture family.","marker":"[42]"},{"why":"Frames the approximation as a truncated Stieltjes moment problem on the semi-infinite interval and supplies the kernel-density solution idea the optimization adapts.","marker":"[18]"},{"why":"Provides the Erlangization theorem converting discrete phase-type rewards into continuous phase-type variables that converge in distribution and moments.","marker":"[23]"},{"why":"Establishes the component count floor(K/2)+1 needed for an Erlang mixture to match K moments, guiding the choice of the mixture size.","marker":"[28]"},{"why":"Defines the histogram-based distributional model-checking baseline that the paper compares against on discrete reward spaces.","marker":"[12]"},{"why":"Supplies the chance-constrained form Pr(X less than or equal to r*) >= alpha used to express the robust properties being verified.","marker":"[14]"},{"why":"Gives the one-sided moment inequality used as a fast sufficient test and early-termination check before fitting the mixture.","marker":"[19]"},{"why":"Supplies the maximum-entropy principle used to select the least-biased density among all distributions matching the given moments.","marker":"[36]"}],"fun_headline_variants":["Moment matching turns reward distributions into verifiable checks","Erlang moments close the gap in chance-constrained checking","Distributional checks for Markov rewards via moment matching","Move beyond expectation in model checking with Erlang mixtures","Chance constraints solved by density approximation in DTMCs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a finite Erlang mixture fitted to a finite number of moments (in the paper's default, three moments and three components with shapes $3^i$) is close enough to the true reward CDF that the verdict on $\\Pr(X \\le r^*) \\ge \\alpha$ is correct, since the proved guarantees are denseness and weak convergence rather than finite-$K$ error bounds.","fun_headline_variants_meta":{"raw":{"variants":["Moment matching turns reward distributions into verifiable checks","Erlang moments close the gap in chance-constrained checking","Distributional checks for Markov rewards via moment matching","Move beyond expectation in model checking with Erlang mixtures","Chance constraints solved by density approximation in DTMCs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000945,"raw_usage":{"total_tokens":4028,"prompt_tokens":932,"completion_tokens":3096,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":3034}},"tokens_in":548,"tokens_out":3096,"duration_ms":19551,"temperature":1.0,"reasoning_tokens":3034,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T22:24:22.574072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small DTMC whose exact cumulative-reward distribution is computable in closed form, fit the paper's default mixture with $K=3$, $n=3$, and shapes $3^i$, and evaluate $\\Pr(X \\le r^*)$ at a threshold where the true CDF lies just above $\\alpha$ but the fitted CDF lies just below; if such a threshold exists, the practical decision procedure returns the wrong verdict, showing that the finite-$K$ fit is not sound without an instance-specific error certificate.","supporting_citations":[{"cited_title":"Stochastic Models: An Algorithmic Approach","cited_arxiv_id":null,"evidence_quote":"Supplies the denseness result that mixtures of Erlang distributions approximate any positive random variable to arbitrary accuracy, the foundation for using the mixture family."},{"cited_title":"The truncated stieltjes moment problem solved by using kernel density functions","cited_arxiv_id":null,"evidence_quote":"Frames the approximation as a truncated Stieltjes moment problem on the semi-infinite interval and supplies the kernel-density solution idea the optimization adapts."},{"cited_title":"Continuous approximations of discrete phase-type distributions and their applications to reliability models","cited_arxiv_id":null,"evidence_quote":"Provides the Erlangization theorem converting discrete phase-type rewards into continuous phase-type variables that converge in distribution and moments."},{"cited_title":"A., AND TAAFFE , M","cited_arxiv_id":null,"evidence_quote":"Establishes the component count floor(K/2)+1 needed for an Erlang mixture to match K moments, guiding the choice of the mixture size."},{"cited_title":"Distributional proba- bilistic model checking","cited_arxiv_id":null,"evidence_quote":"Defines the histogram-based distributional model-checking baseline that the paper compares against on discrete reward spaces."},{"cited_title":"S., M AJUMDAR , R., P RABHU , V","cited_arxiv_id":null,"evidence_quote":"Supplies the chance-constrained form Pr(X less than or equal to r*) >= alpha used to express the robust properties being verified."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the one-sided moment inequality used as a fast sufficient test and early-termination check before fitting the mixture."},{"cited_title":"R., AND PAPANICOLAOU , N","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-entropy principle used to select the least-biased density among all distributions matching the given moments."}],"review_version":1}