{"id":"9983a148-cc36-4c1c-aef7-e2d0156bf91c","arxiv_id":"2411.12936","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For the power-of-two supermarket model, an approximate least-squares estimator based on the mean-field ODE is consistent and asymptotically normal as the number of servers and observation times grow.","lead":"The paper builds a least-squares estimator for the arrival and service rates of the power-of-two supermarket model from aggregate queue-length observations, and proves it is consistent and asymptotically normal. It opens statistical inference for discrete mean-field queueing systems, which matter for load balancing in large data centers and cloud systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (2.7)'s service term breaks mass conservation (Σ_j F_j(x) = -ν x_0), so the estimator targets a misspecified mean-field limit.","rationale":"The paper aims to estimate (λ, ν) of the power-of-two supermarket model by fitting the mean-field ODE to empirical-measure observations. For that program to work, the ODE must be the actual law-of-large-numbers limit of the empirical measure. Summing the displayed F in (2.7) over all coordinates gives -ν x_0, so the vector field points out of the probability simplex whenever the fraction of empty servers is positive. Since the process starts with x_0 > 0 in any realistic queueing system, the claimed unique solution in C([0,T],S) cannot exist. This is an internal inconsistency, not a disagreement with an external convention: the same error appears in the definitions of V_j and in the normal equations used to define the estimator. The Reader's weakest-assumption concern about identifiability condition (3.7) is secondary; even under (3.7), the estimator solves the wrong least-squares problem. I do not see how the simulations detect this, because they compare only the estimator's moments to the true parameters, not the fitted ODE trajectory to the simulated empirical measure. The error is fixable in principle, but fixing it changes the estimator, the sufficient conditions, and the numerical study, so the manuscript as written should not be accepted.","tokens_in":22421,"tokens_out":22136,"duration_ms":226963,"concrete_test":"Evaluate the claimed generator at x = e_0 = (1,0,0,...) in (2.7): U_0 = -1 and V_0 = -1, so F_0 = -(λ+ν); also U_1 = 1 and V_1 = 0, so F_1 = λ. Thus Σ_j F_j(x) = -ν ≠ 0, showing F does not preserve the simplex S. Independently re-derive F_0 from queueing balance: arrivals out of state 0 occur at rate λ(2x_0 - x_0^2) and service completions from state 1 add at rate ν x_1, giving F_0 = λ(x_0^2 - 2x_0) + ν x_1, not λ(x_0^2 - 2x_0) + ν(x_1 - x_0). This settles that (2.7) is missing the +ν x_0 term; correcting it requires revising V_0, Lemma 3.1, and the estimator.","verdict_should_be":"REJECT","load_bearing_attack":"The vector field in (2.2)/(2.7) is not a valid generator for the empirical measure. For any probability vector x, the arrival part satisfies Σ_j U_j(x) = 0, but the service part satisfies Σ_j (x_{j+1}-x_j) = -x_0. Hence Σ_j F_j(x) = -ν x_0, which is nonzero whenever x_0>0. Consequently the ODE (2.1) cannot stay in the simplex S, so the recalled LLN (Theorem 2.1) is inconsistent with the displayed F. The root of the error is V_j in (3.2): a service completion at a queue of length 1 increases ϱ_0, while no service can occur at length 0, so the correct generator has +ν x_1 at coordinate 0, not ν(x_1 - x_0). This error propagates into the normal equations (3.4)-(3.5), the definition of b1, b2, and the LSE (3.10). Even if (3.7) holds and all convergence assumptions are granted, the estimator is built from the wrong U, V pair and need not converge to the true (λ, ν) of the supermarket model. This is more fundamental than the proof gaps identified by the Reader: the inference target itself is misspecified unless the equations are corrected.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an approximate least-squares estimator (LSE) for the arrival rate λ and service rate ν of the power-of-two supermarket model, based on discrete-time observations of the empirical measure of queue lengths over a fixed interval [0,T]. The estimator is obtained by matching the observed increments of the empirical measure to the mean-field ODE recalled in Theorem 2.1. The authors claim consistency (Theorem 3.1) and asymptotic normality (Theorem 3.2) as the number of servers N and the number of observations m tend to infinity, using the existing law of large numbers and central limit theorem for the supermarket model. Numerical experiments with simulated data are reported to support the theoretical results.","tokens_in":22658,"tokens_out":20666,"duration_ms":202758,"significance":"If correct, the paper would address a genuinely open problem: statistical inference for discrete-state mean-field queueing models, as opposed to the more developed inference theory for interacting diffusions. The construction is natural—regress empirical increments on the mean-field vector field—and the use of the external LLN/CLT results is appropriate and not circular. The paper also provides explicit proofs and numerical validation, which is a strength. However, the central estimator is built from a vector field that does not preserve total mass, and the consistency/normality proofs contain convergence arguments that are not justified by the cited results. The claimed theorems therefore do not hold for the estimator as defined in Section 3.2, and the paper needs substantial correction before its central claim can be accepted.","major_comments":[{"comment":"The service term in the mean-field vector field is incorrect at coordinate 0 and breaks mass conservation. For any probability vector x, V_j(x)=x_{j+1}-x_j gives Σ_{j=0}^∞ V_j(x) = -x_0, so Σ_{j=0}^∞ F_j(x) = -ν x_0 ≠ 0 when x_0>0. A service completion at a queue of length 1 increases the mass at state 0 by ν x_1, while no service completion can occur at length 0; hence the correct generator must have V_0(x)=x_1, not x_1-x_0. With that correction, Σ_j V_j(x)=0. As written, the ODE (2.1) cannot remain in the simplex S, contradicting Theorem 2.1. Because the LSE (3.10), the normal equations (3.4)-(3.5), and the definitions of a_12, a_22, b_1, b_2 all use the same erroneous V_j, the estimator targets the wrong mean-field limit. Consequently Theorem 3.1 cannot hold for the estimator as defined, and Theorem 3.2 inherits the same problem.","section":"Section 2.2, Eq. (2.7); Section 3.2, Eq. (3.2)"},{"comment":"The proof of Theorem 3.1 asserts that sup_{t∈[0,T]} Σ_{j=0}^∞ |ϱ^N_j(t)-ϱ_j(t)| → 0 in probability. This is a uniform-in-time l1 convergence statement, but Theorem 2.1 only provides convergence in distribution in D([0,T],S) in the Skorokhod topology. That topology does not by itself yield the uniform l1 bound used in the subsequent estimates for a^N,m_11, b^N,m_1, etc. A uniform tail estimate or an explicit statement of the metric on S and a continuity argument for the limit is needed; without it, the consistency proof has a gap.","section":"Appendix C, Eq. (C.1)"},{"comment":"The proof of Theorem 3.2 passes from convergence of Z^N in D([0,T],ℓ2) to the uniform convergence statement sup_t ∥Z^N(t)-Z(t)∥_2 → 0 in probability, and then uses quantities such as sup_t Σ_i |Z^N_i(t)| together with the weighted ℓ2 moment bound (D.2). Skorokhod convergence does not automatically imply the displayed uniform sup-norm convergence, and the sup of the l1 norm is not controlled by the second-moment bound used. These steps need a rigorous justification or a different argument for the convergence of the terms in (D.1).","section":"Appendix D, Eq. (D.5)"},{"comment":"The identifiability condition a_11 a_22 > (a_12)^2 in (3.7) is stated in terms of the unknown mean-field solution ϱ(·,θ*), and the sufficient conditions (3.8) and (3.9) also involve integrals of the unknown ϱ. If (3.7) fails, the matrix in (3.10) is singular and the estimator is undefined. The paper should discuss how this condition can be checked or avoided in practice; currently it is an assumption that cannot be verified from the data before estimation.","section":"Section 3.2, Lemma 3.1"}],"minor_comments":[{"comment":"The word 'mean-filed' in the abstract should be 'mean-field'.","section":"Abstract"},{"comment":"The statement 'as N, m, m√N → ∞' is ambiguous; based on the proof in Appendix D, the intended condition is m/√N → ∞, and the text should be corrected.","section":"Theorem 3.2"},{"comment":"In the row for (N,m)=(1000,10000), the first entry is written as '0.28, 0.74)' with a missing opening parenthesis.","section":"Table 1"},{"comment":"Reference [15] is cited for Ethier-Kurtz results used in (C.1) and (D.5), but the listed title 'The infinitely-many-alleles model with selection as a measure-valued diffusion' is not the standard Ethier-Kurtz text; the citation needs to be corrected and the page references checked.","section":"References"},{"comment":"The sums in (3.10) run over all j∈Z+, but any numerical implementation must truncate the state space; the paper does not discuss the effect of truncation on the estimator, which would be relevant for the simulation results in Section 4.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the incorrect service term at coordinate 0: as written, the estimator is not consistent for the true parameters. However, the approach is salvageable by correcting V_0 to x_1 and re-working the normal equations, proofs, and numerical experiments. The fact that the reported simulations converge to the true values suggests the numerical implementation may differ from the displayed formulas, and the authors should be asked to clarify this. The other proof gaps around (C.1) and (D.5) also need to be addressed. The citation to Ethier-Kurtz [15] appears mislabeled and should be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: the paper's inference scheme is built on a mean-field ODE that does not conserve mass. In equations (2.2) and (2.7), the service drift is ν Σ_j (x_{j+1} − x_j) e_j. Summing over j gives −ν x_0, not zero. In the supermarket model, a service completion at a length-1 queue moves mass into state 0 and no service occurs at length 0, so the coordinate-0 drift should be +ν x_1, not ν(x_1 − x_0). Because the displayed F leaks total mass, it cannot be the limit of the empirical measure, and the LSE in (3.10)—which uses V_j = x_{j+1} − x_j—targets the wrong object. The consistency and asymptotic normality theorems are therefore not about the model they claim.\n\nWhat is genuinely new: the problem is real. Statistical inference for discrete mean-field queueing models is nearly untouched, and the least-squares idea, regressing observed increments on the drift terms, is a sensible first move. The paper correctly identifies the gap and assembles the LLN and CLT from the literature in a plausible way. As a proposal, it points in the right direction.\n\nBeyond the sign error, the reader's technical concerns are legitimate but secondary. The uniform l1 convergence asserted in (C.1) does not follow from Skorokhod convergence; the infinite sums in the limiting distribution are never shown to converge; and the numerical section uses KS tests on residuals with estimated parameters, which is not a valid normality check, and never compares with the theoretical covariance. Those are fixable. The mass leak is not.\n\nMy recommendation: a serious editor should send this to referees, because the topic is worth refereeing and the error, while load-bearing, is localized and correctable. But the current version should not be accepted; the main theorems are false as stated. The authors need to fix the service term and redo the derivations. I would not cite it in its present form.","headline":"The paper's LSE is built on a service drift that leaks probability mass, so its main theorems do not describe the supermarket model.","tokens_in":23193,"tokens_out":8393,"would_cite":false,"duration_ms":75260,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K25","60K35","62F10","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A closed-form least-squares estimator recovers the arrival and service rates of a large power-of-two queueing network, with proven consistency and asymptotic normality.","keywords":["mean-field queueing","supermarket model","power of two choices","least squares estimation","consistency","asymptotic normality","empirical measure","parameter estimation"],"falsifier":"Take any true parameter pair $(\\lambda^*,\\nu^*)$ for the power-of-two model, solve the mean-field ODE (2.1) numerically on $[0,T]$, compute $a_{11},a_{12},a_{22}$, and verify the strict inequality $a_{11}a_{22}>(a_{12})^2$; then simulate the $N$-server process with large $N$ and $m$ satisfying $m/\\sqrt{N}\\to\\infty$. If for any such parameter pair the estimates fail to concentrate at $(\\lambda^*,\\nu^*)$, or the $\\sqrt{N}$-rescaled errors fail to match the Gaussian law of Theorem 3.2, the central claim falls.","tokens_in":22208,"feed_emoji":"📈","tokens_out":6030,"duration_ms":58571,"temperature":0.7,"pith_summary":"The paper asks whether the arrival rate $\\lambda$ and service rate $\\nu$ of a large supermarket-style queueing network can be estimated from discrete-time snapshots of the aggregate empirical measure, without observing individual servers. It constructs an approximate least-squares estimator from the mean-field ODE limit and proves that the estimator is consistent and asymptotically normal as both the number of servers $N$ and the number of observation times $m$ grow. This is presented as the first statistical inference scheme for a discrete-state mean-field queueing model, where the fluctuations are governed by a Gaussian process rather than by Brownian diffusion. If the main theorems are right, operators of large load-balancing networks can recover the fundamental arrival and service rates from coarse aggregate observations.","feed_headline":"Estimator recovers queueing rates from aggregate observations","feed_subtitle":"A least-squares scheme on the mean-field limit is consistent and Gaussian as servers and observations grow.","key_machinery":"The machinery is the linear-in-parameter decomposition of the mean-field drift, $F(x)=\\lambda\\sum_j U_j(x)e_j+\\nu\\sum_j V_j(x)e_j$, with $U_j$ and $V_j$ given by (3.1)-(3.2). Because $\\lambda$ and $\\nu$ enter linearly, the minimizer of the quadratic criterion $G(\\lambda,\\nu)$ has the closed form (3.6) in terms of integrals of $U_j$ and $V_j$ along the limiting trajectory; replacing those integrals by Riemann sums over the observed empirical measure produces the estimator (3.10). The argument also rests on the law of large numbers for the empirical measure (Theorem 2.1), the central limit theorem whose limiting process $Z$ solves the SDE (2.5), and the identifiability condition $a_{11}a_{22}>(a_{12})^2$, which keeps the $2\\times2$ normal matrix invertible.","core_discovery":"The central claim is that, for the power-of-two supermarket model, the unknown rates $\\theta^*=(\\lambda^*,\\nu^*)$ are identifiable from the finite dataset $D_{N,m}=\\{\\varrho^N(t_k):1\\le k\\le m\\}$ and can be recovered by the closed-form least-squares formula (3.10). Theorem 3.1 states that, under the strict identifiability condition $a_{11}a_{22}>(a_{12})^2$, the estimator converges in probability to $\\theta^*$ as $N,m\\to\\infty$. Theorem 3.2 states that $\\sqrt{N}(\\hat\\theta-\\theta^*)$ converges in distribution to a Gaussian vector as $N,m\\to\\infty$ with $m/\\sqrt{N}\\to\\infty$, and the limiting law is a linear combination of the Gaussian fluctuation process $Z$ arising from the central limit theorem for the empirical measure. The proof structure carries all randomness through $Z$, so the estimator's asymptotic distribution inherits Gaussianity from Proposition 2.1.","pith_inferences":["Beyond the paper, the Gaussian limit suggests a practical monitoring device: sliding-window estimates of $\\lambda$ and $\\nu$ from aggregate queue-length histograms could detect changes in load or service capacity, with standard errors read off the limiting covariance.","Because the identifiability condition is stated in terms of the unknown mean-field trajectory, a data-driven implementation would likely need to check the sample Gram matrix for near-singularity; poor conditioning would show up as unstable estimates rather than honest uncertainty.","The requirement $m/\\sqrt{N}\\to\\infty$ means the observation grid must be fine relative to the square root of the network size, which quantifies how much discretization bias the fluctuation analysis tolerates.","The authors' numerical results show skewness and kurtosis close to Gaussian values at moderate $N$ and $m$, suggesting the asymptotic regime may be reached earlier in practice than the theorem's joint limit requires."],"forward_implications":["The estimator (3.10) is a closed-form function of aggregate queue-length snapshots, so it can be applied when tracking every server individually is impractical.","As both the system size and the number of observation epochs grow, the estimates concentrate at the true arrival and service rates, as stated in Theorem 3.1.","The $\\sqrt{N}$-scaled estimation error converges to a Gaussian distribution, so the authors' expressions for its mean and covariance can in principle yield confidence intervals for $\\lambda$ and $\\nu$.","The authors note that the same approximate-LSE scheme is conceivable for mean-field models whose limiting equation is linear in the unknown parameters, such as load-balancing cloud storage models, provided the limiting equation remains tractable.","The linear-in-parameter structure is essential: for nonparametric estimation of interaction kernels the optimization moves to function space, and the paper's approach does not directly carry over."],"supporting_citations":[{"why":"Introduced the supermarket model and its convergence to a deterministic mean-field ODE, which is the limit used to build the estimator.","marker":"[35]"},{"why":"Introduced the power-of-two-choices model, the special case on which the inference scheme is developed.","marker":"[31]"},{"why":"Supplies the law of large numbers for the empirical measure, restated as Theorem 2.1 and used to justify replacing the true trajectory by the prelimit process.","marker":"[20]"},{"why":"Supplies the central limit theorem and the Gaussian fluctuation process $Z$ (Theorem 2.2), plus the well-posedness and moment estimates used in the proofs of consistency and asymptotic normality.","marker":"[6]"},{"why":"Provides the convergence criteria and Skorokhod representation results used in the proofs of the consistency and asymptotic-normality theorems.","marker":"[15]"},{"why":"Provides the continuous mapping theorem used to finish the proof of Theorem 3.2.","marker":"[34]"}],"fun_headline_variants":["Least-squares estimator for supermarket queueing is consistent and Gaussian","Power-of-two queueing rates recovered by least squares from discrete samples","Mean-field inference for discrete jump processes via least squares","Supermarket model parameters identified from discrete observations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the strict identifiability condition $a_{11}a_{22}>(a_{12})^2$, which is stated in terms of the unknown true mean-field trajectory; if it fails, the normal matrix in the estimator is singular and the whole scheme is undefined.","fun_headline_variants_meta":{"raw":{"variants":["Least-squares estimator for supermarket queueing is consistent and Gaussian","Power-of-two queueing rates recovered by least squares from discrete samples","Mean-field inference for discrete jump processes via least squares","Supermarket model parameters identified from discrete observations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000577,"raw_usage":{"total_tokens":2729,"prompt_tokens":963,"completion_tokens":1766,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":1700}},"tokens_in":579,"tokens_out":1766,"duration_ms":14215,"temperature":1.0,"reasoning_tokens":1700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:02:30.656861+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any true parameter pair $(\\lambda^*,\\nu^*)$ for the power-of-two model, solve the mean-field ODE (2.1) numerically on $[0,T]$, compute $a_{11},a_{12},a_{22}$, and verify the strict inequality $a_{11}a_{22}>(a_{12})^2$; then simulate the $N$-server process with large $N$ and $m$ satisfying $m/\\sqrt{N}\\to\\infty$. If for any such parameter pair the estimates fail to concentrate at $(\\lambda^*,\\nu^*)$, or the $\\sqrt{N}$-rescaled errors fail to match the Gaussian law of Theorem 3.2, the central claim falls.","supporting_citations":[{"cited_title":"Vvedenskaya, R.L","cited_arxiv_id":null,"evidence_quote":"Introduced the supermarket model and its convergence to a deterministic mean-field ODE, which is the limit used to build the estimator."},{"cited_title":"Mitzenmacher","cited_arxiv_id":null,"evidence_quote":"Introduced the power-of-two-choices model, the special case on which the inference scheme is developed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the law of large numbers for the empirical measure, restated as Theorem 2.1 and used to justify replacing the true trajectory by the prelimit process."},{"cited_title":"Budhiraja and E","cited_arxiv_id":null,"evidence_quote":"Supplies the central limit theorem and the Gaussian fluctuation process $Z$ (Theorem 2.2), plus the well-posedness and moment estimates used in the proofs of consistency and asymptotic normality."},{"cited_title":"Ethier and T.G","cited_arxiv_id":null,"evidence_quote":"Provides the convergence criteria and Skorokhod representation results used in the proofs of the consistency and asymptotic-normality theorems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the continuous mapping theorem used to finish the proof of Theorem 3.2."}],"review_version":1}