# Equal arrival moments do not identify mean waiting **Finding.** In the specified single-server model at 50% utilization, three iid arrival laws with the same mean and variance give mean waits of 1.000000, 0.958912 and 0.699539 ms. The common Kingman VUT estimate is 1 ms. At 80%, the two discrete laws lie on opposite sides of the common 4 ms estimate. This is a reproducible worked counterexample using classical GI/M/1 theory, not a new theorem, hardware measurement or simulation. ## Model and estimand One work-conserving, nonpreemptive FCFS server has iid exponential service times with mean 1 ms. Independent iid positive interarrival intervals are A = Y/rho ms. There is unlimited waiting space, no vacation, abandonment, rejection or feedback. The estimand is steady-state mean waiting before service for an arriving request, including requests with zero wait. It excludes service time. Y has one of these laws: - Exponential with mean 1. - 80/20: Y = 1/2 with probability 4/5, or 3 with probability 1/5. - 90/10: Y = 2/3 with probability 9/10, or 4 with probability 1/10. - Deterministic Y = 1 is a separate SCV-zero baseline, not part of the matched-SCV comparison. For both discrete laws, direct rational summation gives E[Y] = 1 and E[Y^2] = 2, hence Var(Y) = 1, equal to the exponential law. Scaling gives E[A] = 1/rho ms and Var(A) = 1/rho^2 ms^2. All three arrival SCVs equal one. At rho = 0.5, the two discrete interval pairs are (1,6) ms and (4/3,8) ms. At rho = 0.8, they are (5/8,15/4) ms and (5/6,5) ms. Each new interval is sampled independently; these are not repeating schedules. ## Classical solution and what is being added Let L_A(s) = E[exp(-sA)]. The unique interior root r in (0,1) solves r = L_A(mu(1-r)). There is also the uninformative boundary root r = 1. For this FCFS model, P(Q>t) = r exp[-mu(1-r)t], t >= 0, and P(Q=0)=1-r. Integrating the tail gives E[Q] = r/[mu(1-r)]. MIT HW4 problem 1 states the root and distribution; Li and Niu (1992), section 2.2, equations 23, 24 and 29 state the ordinary GI/M/1 special case. Their D-policy extension is not used here. With exponential mean-one Y, the interior root is exactly rho: the M/M/1 baseline. With a discrete Y, L_Y(z) is the probability-weighted sum of exp(-zY). Equal first two moments do not make these transforms equal. The contribution here is the explicit matched pair, its reproducible numerical comparison with the existing Hesela VUT example, and checked numerical enclosures. The common VUT expression is ((c_a^2+c_s^2)/2)*rho/(1-rho)*E[S]. It is exact for the exponential-arrival baseline under these assumptions, but is an approximation for the other laws. ## Results and interpretation The JSON has 16 evaluations: four laws at rho = 0.2, 0.5, 0.8 and 0.95. The article shows the six matched-SCV rows at rho = 0.5 and 0.8. Six decimal places are display rounding, not a claim about physical precision. At 50%, using the common 1 ms estimate overstates the 90/10 mean by about 42.95%, but the 80/20 mean by about 4.28%. At 80%, it understates the 80/20 mean by about 1.45% and overstates the 90/10 mean by about 4.66%. These are errors relative to the exact means of these declared models, not a universal error envelope for Kingman's approximation. The SCV-zero deterministic case has its own different VUT input; do not describe all 16 rows as moment-matched. The chosen examples do not prove a ranking for all arrival laws. No finite arrival trace or empirical population was sampled; there is no sample size, censoring, bootstrap, seed or statistical confidence interval. For storage work, matching IOPS and interarrival SCV is insufficient to establish equivalent queueing workloads even before introducing serial dependence. Retain the arrival law and timestamps where available, plus the service and scheduling assumptions. Real devices have parallelism, non-exponential service, caching and correlated work; this model is not a predictor of their mean or p99. ## Numerical certificate Decimal arithmetic at precision 80 proposes an interior root bracket. It is not trusted as the certificate. Each endpoint is then converted to an exact Fraction and the residual signs are checked using rational arithmetic. For x >= 0, the Taylor polynomial of exp(-x) ending at an odd power is a lower bound; the preceding even polynomial is an upper bound (Taylor remainder sign). The code continues until their difference is below 10^-45 and n > x. Weighted sums give lower/upper transform bounds. The candidate is accepted only if the residual is provably positive on the left and negative on the right. Strict convexity of L_A(mu(1-r))-r, its value at zero and derivative at one establish that the interior crossing is unique for these stable laws. The waiting map E[S]*r/(1-r) is strictly increasing. Mapping both root endpoints gives the published rational waiting interval. The residual bounds in the JSON are conservatively rounded outward to a rational grid of 10^-50 for compactness. Tests require positive/negative certified signs and waiting-interval width below 10^-18 ms for every published case. These bounds cover numerical evaluation, not model misspecification, sampling or display rounding. The solver intentionally supports only the four named laws, exact inputs, 0.01 <= rho <= 0.999 and positive mean service. It fails if an interior bracket or rational sign certificate cannot be established. It is not a generic queue solver. ## Reproduce and audit Python 3.10+ standard library only; tested with the version in CHECKPOINT.md. ~~~sh python3 -B model.py --check python3 -B -m unittest discover -s . -p test_model.py python3 -B -O -m unittest discover -s . -p test_model.py python3 -B review_test.py python3 -B -O review_test.py ~~~ Without --check, model.py regenerates results.json. In the source repository, package.py copies the six public research files and checks that the copies match. The JSON records the model source SHA-256. No third-party data are bundled. Tests include exact moment identities, analytic M/M/1 roots, the independent closed-form MIT hyperexponential example (5-sqrt(3.4))/4, sign certificates, load monotonicity, time scaling, distinct waiting intervals, invalid inputs, A separate review implementation iterates the transform from zero without importing the model solver. It checks all 16 published roots, means and error denominators, and rejects deliberately corrupted results. freshness and one near-saturation sanity check. The latter is not an asymptotic proof. ## Sources and access limits 1. MIT OpenCourseWare, 15.072, HW4 Solutions, 24 April 2006, problems 1-2, PDF pp. 1-3. Root/distribution and an independent solved numerical baseline. https://ocw.mit.edu/courses/15-072j-queues-theory-and-applications-spring-2006/993e13a9105287dd88f66d9f3c8fd2f3_sol4.pdf 2. Jingwen Li and Shun-Chen Niu, The Waiting-Time Distribution for the GI/G/1 Queue under the D-Policy, Probability in the Engineering and Informational Sciences 6 (1992), 287-308. Author manuscript revised January 1992; section 2.2, printed pp. 7-8 (PDF pp. 9-10), equations 23, 24, 29. Only the ordinary GI/M/1 formulas are applied. DOI not independently verified. https://personal.utdallas.edu/~scniu/documents/gig1-dp.pdf 3. Mor Harchol-Balter and Ziv Scully, The most common queueing theory questions asked by computer systems practitioners, section 1. VUT convention and SCVs. https://www.cs.cmu.edu/~harchol/Papers/TeaPACS21.pdf The 1969 Izzet Sahin report, Technical Memorandum 134, was also located: https://commons.case.edu/wsom-ops-reports/367/ Its repository describes root computation for Erlang and deterministic arrivals. Only metadata/abstract were accessible (PDF returned 403), so it is a historical lead, not evidence for this calculation or an inspected algorithm. No access restriction was bypassed. Source URLs, hashes and access scopes are recorded in sources.json. Copyrighted source PDFs are not redistributed.