Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Uncovering the topology of an infinite-server queueing network from population data

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Snapshot counts alone reveal the routing structure of a queueing network.

desk verdict New moment-inversion result for topology recovery in infinite-server networks is real and mostly solid, but Theorem 1 hangs on an unproved invertibility condition that needs to be supplied. read the letter →

arxiv 2506.07057 v1 pith:HWBN2C3L submitted 2025-06-08 math.PR math.STstat.MEstat.TH

Hritika Gupta , Michel Mandjes , Liron Ravner , Jiesen Wang This is my paper
classification math.PRmath.STstat.MEstat.TH MSC 60K2562F1290B22
keywords infinite-serverqueuesroutingmatrixestimationmethodofmomentsPoissonsamplingnetworktopologyinferenceidentifiabilityconsistencycross-moments
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that the directed topology of a network of infinite-server queues can be recovered from nothing more than repeated snapshots of how many customers are at each station, taken at Poisson-distributed times. It proposes a method-of-moments estimator, built on explicit formulas for the mean population vector and for the cross-covariances between station populations one or two observation epochs apart, and proves that the estimator converges in probability to the true arrival rates and routing matrix as the number of snapshots grows, for any positive sampling rate. If true, this gives a statistical way to uncover who routes to whom, including link direction, without ever observing individual customers or their paths. The method also distinguishes networks that have the same stationary population distribution, because the lagged cross-covariances are asymmetric in the station indices.

What carries the argument

The machinery is the lag-one cross-moment matrix $\mathbb{E}[\boldsymbol{M}(0)\boldsymbol{M}(T_\beta)^\top]$ for an independent exponential time $T_\beta$ with rate $\beta$. Its ingredients are $P(\beta) = (I - \mathrm{diag}\{\mathcal{G}(\beta)\}Q)^{-1}(I - \mathrm{diag}\{\mathcal{G}(\beta)\})$, the probability that a fresh customer starting at one station is at another when the exponential clock rings, and its residual-service-time analogue $P^{(\mathrm{res})}(\beta)$; these enter through Proposition 1, and differentiating them in $\beta$ gives the Erlang-2 cross-moments used when service-time parameters or observation probabilities must also be estimated. The routing matrix is then isolated as a solution of a linear matrix equation, $Q = \Omega_1^{-1} \Omega_2$, and the estimator plugs empirical moments into this expression, with projection back onto the parameter space when finite-sample estimates fall outside it.

What would settle it

Find a two-station network and parameter values satisfying Assumptions 1 and 2 for which the determinant of $\Omega_1$ is zero while the moment system still has two distinct solutions $(\boldsymbol{\lambda}, Q)$, or compute $\det \Omega_1$ symbolically along a family of routing matrices and exhibit one that vanishes; either would give a concrete instance where the claimed identifiability fails.

Watch

Extended reading notes

Core claim

The central claim is that the moment equations $\mathbb{E}[M_i(0)] = \rho_i$ and $\mathbb{E}[M_j(0) M_i(T_\beta)] = [(\boldsymbol{\rho} \boldsymbol{\rho}^\top + \mathrm{diag}\,\boldsymbol{\rho})\, P^{(\mathrm{res})}(\beta) + \beta^{-1} \boldsymbol{\rho} \boldsymbol{\lambda}^\top P(\beta)]_{ji}$ determine $(\boldsymbol{\lambda}, Q)$ uniquely, given the service-time Laplace-Stieltjes transforms. The paper derives these equations by conditioning on an independent exponential clock, defines the estimator as the solution of the empirical versions of these $n^2 + n$ equations, and proves consistency by combining pointwise almost sure convergence of the empirical moments, a Lipschitz and Glivenko-Cantelli argument on the compact parameter space, and identifiability through the linear relation $Q = \Omega_1^{-1} \Omega_2$. The consistency theorem is stated for every sampling rate $\beta > 0$. A load-bearing step in the identifiability argument is the invertibility of $\Omega_1$, which the paper uses to define $Q$; Section 4.2 does not supply a proof of non-singularity over the full parameter space.

Load-bearing premise

The load-bearing premise is that the matrix $\Omega_1$ used to recover the routing matrix is invertible for every parameter vector in the allowed space; Section 4.2 defines $Q$ as $\Omega_1^{-1} \Omega_2$ but does not prove the inverse exists, and a singular $\Omega_1$ would leave the moment equations unable to pin down the topology.

Editorial extensions

If this is right

  • For known service-time distributions, $n^2 + n$ equations from means and one-step cross-moments exactly determine the $n^2$ routing probabilities and $n$ external arrival rates, with no individual customer tracking needed.
  • The estimator is consistent for any sampling rate $\beta > 0$, so even sparse or slow Poisson snapshots suffice as the number of observations grows.
  • Because the cross-moments are not symmetric in the station indices, the method can orient edges and can tell apart networks that share the same stationary population distribution.
  • When service-time parameters or observation probabilities are unknown, adding Erlang-2 lag-two cross-moments supplies enough equations for consistency under the paper's regularity conditions.
  • In numerical experiments on line, circle, symmetric-circle, and clique topologies, the procedure recovers the routing structure, with empirical variances small relative to the estimated values.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: if the invertibility of $\Omega_1$ is established for general topologies, the same moment-identification scheme should carry over to networks with batch arrivals or phase-type service times, since only Laplace transforms of service times enter the formulas.
  • Editorial extension: the asymmetry of the cross-moments suggests a general principle for causal discovery from aggregate counts: two snapshots at a known random lag can orient directed interactions even when the marginal distribution at each node is the same.
  • Editorial extension: for finite samples, the need to project estimates back into the parameter space could bias small-sample estimates, so resampling-based confidence intervals are a natural next step that the paper does not develop.
  • Editorial extension: the model-free variant, which treats $\mathbb{E}[G_i]$, $\mathcal{G}_i(\beta)$, and $\dot{\mathcal{G}}_i(\beta)$ as free parameters, points toward a nonparametric misspecification test: if estimates from a parametric fit and from the model-free fit diverge, the assumed service-time family is wrong.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper studies statistical inference for a network of infinite-server queues observed only through the population vector at Poisson sampling epochs. The main object is a method-of-moments estimator for the routing matrix, external arrival rates, and, in extensions, service-time parameters and observation probabilities. The authors derive explicit expressions for first moments and cross-moments at exponential and Erlang inter-observation times, define an estimator from matching these moments, and prove consistency of the estimator when service-time distributions are known (Theorem 1). They then sketch extensions to parametric service-time estimation, a model-free version, and noisy observations, and support the method with numerical experiments on several topologies.

Significance. If the central consistency theorem is fully established, the paper would provide a practically attractive way to infer directed network structure from population snapshots alone, going beyond the stationary distribution and exploiting asymmetric cross-moments. The moment derivations in Sections 3 and 4.5 are careful and nontrivial, and the numerical experiments in Section 7 give concrete evidence that the method works on the tested topologies. The proof strategy, using a majorizing M/G/infinity process and Glivenko-Cantelli arguments, is appropriate. However, the main theorem currently rests on an unproved invertibility assertion, and the extended settings in Sections 5 and 6 are not supported by the same level of proof; these gaps are load-bearing for the paper's broader claims.

major comments (4)
  1. [§4.2, Proposition 3] The proposition defines Q(alpha^[0], alpha^[1]) = Omega_1(alpha^[0], alpha^[1])^{-1} Omega_2(alpha^[0], alpha^[1]) and states that there is a one-to-one relation between the moments and the parameters, but no proof is given that Omega_1 is nonsingular on the parameter space. This is not a minor gap: Proposition 4(iii) and hence Theorem 1 in §4.3 use exactly this uniqueness to conclude that the moment equations have only the true parameter as their solution. The text before Proposition 3 only says the equation is linear in Q and then asserts the inverse. The numerical experiments in Experiment 1 show invertibility for particular topologies, but they do not establish uniform invertibility over Theta, and the consistency theorem is stated for every admissible parameter. The authors should either prove invertibility under Assumptions 1–2, add an explicit assumption of nonsingularity of Omega_1 over Theta, or restrict Theorem 1 accordingly.
  2. [§5.1 and §5.2] The claimed extensions to unknown service-time distributions are not proved. In §5.1 the paper explicitly states that "deriving a counterpart to Proposition 3 may prove to be more challenging," which means the identifiability step, the essential ingredient of the consistency proof, is missing for the parametric variant. Section 5.2 similarly asserts that "one again obtains consistency" under Lipschitz assumptions, but it does not prove that the moment mapping is injective in the model-free parameter vector of dimension n^2+3n; having enough moment equations does not by itself imply a unique solution. Since the abstract and introduction present these variants as part of the contribution, the absence of identifiability proofs leaves the main claims for Sections 5.1 and 5.2 unsubstantiated.
  3. [§6] The section on unknown observation probabilities derives moment formulas in Lemma 4 and states that the estimation procedure "works essentially as before," but no consistency theorem is stated or proved for this setting. Given that the introduction highlights the estimation of observation probabilities as part of the main contribution, the paper should either provide a formal consistency result with the required identifiability conditions or clearly label Section 6 as heuristic. As written, the claim that the procedure estimates all model parameters including the observation probabilities is not backed by the paper's theoretical results.
  4. [§7.2] The efficient procedure replaces the model-based first moment rho by the empirical first-moment vector psi^[0] inside the cross-moment equations. This is a plug-in estimating procedure, not exactly the estimator defined by equations (13)–(14) in §4.1. Since Theorem 1 is proved for the latter estimator, the numerical procedure in Section 7.2 is not directly covered by Theorem 1. The authors should clarify the relationship, or prove consistency for the plug-in version, which should follow from continuity of the relevant maps but is not stated.
minor comments (5)
  1. [§3.3] Just before Proposition 2, the text says "Applying (10) in combination with Proposition 8"; this should refer to Proposition 1.
  2. [§7.2] In the displayed equation near the end of Section 7.2, the matrix diag{p} appears even though the subsection assumes p=1; this appears to be a leftover from the noisy-observation setting and should be removed or explained.
  3. [Figure 1 and §7.3] The caption of Figure 1 labels topology (d) as "2+3 Cliques," while the text in Experiment 1 calls it "3 + 2 Cliques"; the labels should be made consistent.
  4. [§4.3] The proof of Theorem 1 assumes the existence of a solution to the moment equations and does not address the possibility of multiple solutions or the need to select one; Remark 2 mentions truncation/projection for finite samples, but the theoretical statement would be cleaner if the estimator were defined as any measurable solution or via a well-defined minimization over Theta.
  5. [§2] The notation G_i is used both for the service-time random variable and for its cumulative distribution function; the distinction is mostly clear from context, but a note would help avoid confusion in the derivations of Sections 3 and 4.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the moment-based estimator follows from a self-contained derivation of stationary moments, and the cited results are standard external asymptotics tools.

full rationale

The paper's derivation chain is self-contained rather than circular. The central claims are (i) explicit moment formulas, (ii) a method-of-moments estimator defined by equating theoretical and empirical moments, and (iii) a consistency theorem built on identifiability of the moment map plus Glivenko--Cantelli convergence. The moment expressions in Propositions 1 and 2 are derived directly from the infinite-server queue dynamics, Lemma 1 and Lemma 2, and the memoryless/excess-life computations; they do not assume the parameters being estimated. Proposition 3 obtains (lambda, Q) by linear algebra from the moment map, and Proposition 4/THEOREM 1 then establish consistency using pointwise convergence and a Lipschitz condition. No fitted parameter is relabeled as a prediction: the empirical first and second moments are plugged in as sample analogues of the theoretical moments, which is standard moment estimation rather than circular reasoning. The cited [13] is prior work by two of the authors, but it is not load-bearing; the consistency proof instead relies on textbook results [1,16]. The only notable weakness is that Proposition 3 asserts Q = Omega_1^{-1} Omega_2 without proving Omega_1 is nonsingular on the whole parameter space; that is a regularity gap affecting the proof of identifiability, but it is not a circular step, because singularity concerns are not built into the definition of the moments or the estimator. The numerical experiments independently test the estimator against known parameters, further confirming the derivation is not circular. Therefore, on the seven enumerated circularity patterns, no step exhibits a self-definitional reduction, fitted-input-as-prediction, or self-citation chain that forces the conclusion.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central claim (consistency) rests on standard stationarity and Poisson-sampling assumptions, plus the unproven invertibility of Omega_1 and unproven regularity conditions for the model-free variants. No new physical entities are introduced.

free parameters (5)
  • lambda_i (external arrival rates) = e.g., lambda=(3,2,4,3,4) in Experiment 1
    n unknown arrival rates estimated from the n stationary mean moment equations (13); target of inference, not hidden degrees of freedom.
  • q_ij (routing probabilities) = e.g., line topology q_{i,i+1}=0.5
    n^2 unknown routing probabilities estimated from the n^2 lag-one cross-moment equations (14); target of inference.
  • mu_i or eta_i (service-time parameters) = e.g., mu=(2,3,3,4,3) in Experiment 1
    In Section 5, n additional parameters enter when service-time distributions are in a parametric family; estimated using lag-two cross-moments. In the model-free version, each station contributes three functionals: E[G_i], G_i(beta), and G'_i(beta).
  • p_i (observation probabilities) = not specified in numerical experiments
    In Section 6, n observation probabilities are estimated from extended moments; the paper does not report numerical values for them.
  • model-free functionals E[G_i], G_i(beta), G'_i(beta) = e.g., 0.3330 and 0.2003 for E[G] in Experiment 4
    The model-free approach treats these three numbers per station as parameters, replacing the service-time distribution by three point functionals.
assumptions (7)
  • domain assumption The network is strongly stationary from time 0; the queue-length process is in equilibrium at all observation times.
    Stated in Section 2; used to justify the moment expressions E[M_i(t)] = rho_i and to apply ergodic theorems.
  • domain assumption Observation times follow an independent Poisson process with known rate beta.
    Sections 2 and 3; the exponential inter-observation times are the basis for the LST expressions P(beta) and the cross-moment formulas.
  • domain assumption Assumption 1(i): every station has positive incoming flow; Assumption 1(ii): uniform epsilon lower bound on the probability of an exit path.
    Section 4.1; used to ensure the effective arrival rate vector is well defined, to bound E[H(theta)] uniformly, and to prove Lipschitz continuity and Glivenko-Cantelli convergence.
  • domain assumption Assumption 2: E[G_j] is positive and finite for all j.
    Section 4.1; excludes degenerate zero-service-time nodes and ensures finite moments.
  • domain assumption In the service-time estimation setting, q_ii = 0 for all i.
    Remark 1 and Section 5.1; imposed to avoid the non-identifiability of q_ii and an exponential service rate mu_i, which enter only through the product (1-q_ii) mu_i.
  • ad hoc to paper In the model-free approach, E[G_i], G_i(beta), and G'_i(beta) are Lipschitz in theta and E[G_i] >= g_- > 0.
    Section 5.2; the paper states these regularity conditions are needed for consistency, but no proof is given and no check on natural parametric families is provided.
  • ad hoc to paper The matrix Omega_1(alpha^[0], alpha^[1]) is invertible over Theta.
    Section 4.2, Proposition 3; the closed-form inversion Q = Omega_1^{-1} Omega_2 requires non-singularity, which the paper does not prove.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncovering the topology of an infinite-server queueing network from population data." pith.science (2026). https://pith.science/paper/HWBN2C3L

@misc{pith2026250607057,
  author       = {Pith},
  title        = {Pith review of: Uncovering the topology of an infinite-server queueing network from population data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HWBN2C3L}},
  note         = {Machine review of arXiv:2506.07057}
}
read the original abstract

This paper studies statistical inference in a network of infinite-server queues, with the aim of estimating the underlying parameters (routing matrix, arrival rates, parameters pertaining to the service times) using observations of the network population vector at Poisson time points. We propose a method-of-moments estimator and establish its consistency. The method relies on deriving the covariance structure of different nodes at different sampling epochs. Numerical experiments demonstrate that the method yields accurate estimates, even in settings with a large number of parameters. Two model variants are considered: one that assumes a known parametric form for the service-time distributions, and a model-free version that does not require such assumptions.

Figures

Figures reproduced from arXiv: 2506.07057 by the authors.

Figure 1
Figure 1. Four topologies considered in Experiment 1. Experiment 1. In the first experiment, we consider an instance in which we wish to estimate the vector of external arrival rates 𝝀 and the transition rate matrix 𝑄; the service times are exponentially distributed with known rates 𝝁, and the observation probabilities 𝒑 are known to equal 1. It means [PITH_FULL_IMAGE:figures/full_fig_p021_1.png] view at source ↗
Figure 2
Figure 2. Histograms pertaining to estimates of 𝜆1, based on 𝑅 = 1000 experi￾ments, each with 𝑚 = 250 000 observations. the value of the corresponding routing probability is below 0.01, and is dotted if it is below 0.02. The figures show that, when increasing the number of observations 𝑚, the algorithm increasingly well succeeds in identifying the structure of the underlying routing matrix. Experiment 3. In this experiment we… view at source ↗
Figure 3
Figure 3. An example with five stations and a cyclic network. the model-free approach we need to estimate (E[𝐺1], E[𝐺2]), (𝒢1(𝛽),𝒢2(𝛽)), (𝒢¤ 1 (𝛽),𝒢¤ 2(𝛽)). In this experiment we take 𝒑 = 1, while the external arrival rates are 𝝀 = (3, 4) ⊤ and the transition probabilities 𝑞1,2 = 𝑞2,1 = 0.5, 𝑞1,1 = 𝑞2,2 = 0; see [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Colormap of |𝑄 − 𝑄b| I II [PITH_FULL_IMAGE:figures/full_fig_p027_4.png]
Figure 5
Figure 5. Figure 5: Two-station network used to test the model-free approach. We consider two cases: in the former the data is generated such that the service times are ex￾ponentially distributed with rates 𝝁, whereas in the latter the service times follow and Erlang-2 distribution with r…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Parameter inference for partially observed branching processes

    math.ST 2026-07 accept novelty 6.0 of 10

    Method-of-moments estimators for the parameters of a partially observed age-dependent branching process are identifiable and asymptotically normal when only aggregate population size is recorded.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    D. Andrews. Generic uniform convergence. Econometric Theory, 8(2):241–257, 1992

    work page 1992

  2. [2]

    Asanjarani, Y

    A. Asanjarani, Y. Nazarathy, and P. Taylor. A survey of parameter and state estimation in queues. Queueing Systems, 97:39–80, 2021

    work page 2021

  3. [3]

    Asmussen

    S. Asmussen. Applied Probability and Queues. Springer, 2003

    work page 2003

  4. [4]

    Bingham and S

    N. Bingham and S. Pitts. Nonparametric estimation for the M/G/ ∞ queue. Annals of the Institute of Statistical Mathematics, 51:71–97, 1999

    work page 1999

  5. [5]

    Blanghaps, Y

    N. Blanghaps, Y. Nov, and G. Weiss. Sojourn time estimation in an M/G/ ∞ queue with partial information. Journal of Applied Probability, 50:1044–1056, 2013

    work page 2013

  6. [6]

    Brillinger

    D. Brillinger. Cross-spectral analysis of processes with stationary increments including the stationary G/G/ ∞ queue. Annals of Probability, 2:815–827, 1974

    work page 1974

  7. [7]

    M. Brown. An M/G/ estimation problem. Annals of Mathematical Statistics, 41:651–654, 1970. UNCOVERING THE TOPOLOGY OF A QUEUEING NETWORK FROM POPULATION DATA 29

    work page 1970

  8. [8]

    Castro, M

    R. Castro, M. Coates, G. Liang, R. Nowak, and B. Yu. Network tomography: Recent developments. Statistical Science, 19:499–517, 2004

    work page 2004

Show all 18 references
  1. [9]

    Goldenshluger

    A. Goldenshluger. Nonparametric estimation of the service time distribution in the M/G/ ∞ queue. Advances in Applied Probability, 48:1117–1138, 2016

    2016

  2. [10]

    Goldenshluger

    A. Goldenshluger. The M/G/∞ estimation problem revisited. Bernoulli, 24:2531–2568, 2018

    2018

  3. [11]

    Hansen and S

    M. Hansen and S. Pitts. Nonparametric inference from the M/G/1 workload. Bernoulli, 12:737–759, 2006

    2006

  4. [12]

    J. Pearl. Causality: Models, Reasoning and Inference. Cambridge University Press, 2nd edition, 2009

    2009

  5. [13]

    Ravner, O

    L. Ravner, O. Boxma, and M. Mandjes. Estimating the input of a L ´evy-driven queue by Poisson sampling of the workload process. Bernoulli, 25:3734–3761, 2019

    2019

  6. [14]

    Schweer and C

    S. Schweer and C. Wichelhaus. Nonparametric estimation of the service time distribution in discrete-time queueing networks. Stochastic Processes and their Applications, 130:4643–4666, 2020

    2020

  7. [15]

    Sutton and M

    C. Sutton and M. Jordan. Bayesian inference for queueing networks and modeling of internet services. Annals of Applied Statistics, 5:254–282, 2011

    2011

  8. [16]

    van der Vaart

    A. van der Vaart. Asymptotic Statistics. Cambridge University Press, 2000

    2000

  9. [17]

    W. Whitt. Stochastic-Process Limits: An Introduction to Stochastic-Process Limits and Their Application to Queues. Springer, 2002

    2002

  10. [18]

    Wichelhaus and L

    C. Wichelhaus and L. Langrock. Nonparametric inference for stochastic feedforward networks based on cross- spectral analysis of point processes. Electronic Journal of Statistics, 6:1670–1714, 2012

    2012

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.