Pith. sign in

REVIEW 2 major objections 5 minor 42 references

A modified drift-plus-penalty policy can provision probabilistic throughput and delay guarantees in shared multi-infrastructure networks, with an optimality gap that shrinks as the service frame grows.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

The MDP policy linearizes probabilistic QoS constraints and provably achieves near-optimal utility and stability in multi-infrastructure-sharing networks, with error vanishing in the frame size.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The paper's core linearization idea is neat, but Theorem 3's proof has an algebraic gap that invalidates the main O(1/sqrt(T_s)) guarantees as written. the 2 major comments →

arxiv 2509.01694 v2 pith:OWUHMX2L submitted 2025-09-01 cs.NI cs.SYeess.SY

A QoS Framework for Service Provision in Multi-Infrastructure-Sharing Networks

Quang Minh Nguyen , Eytan Modiano This is my paper
classification cs.NI cs.SYeess.SY
keywords Quality-of-ServiceInfrastructure-sharing networksResource allocationDrift-plus-penaltyProbabilistic QoSStability regionLyapunov optimizationNetwork utility maximization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 asks whether a network orchestrator can give clients hard, tunable Quality-of-Service promises—throughput delivered with a stated probability within a stated time window—when several infrastructure providers share their links. The authors propose a three-parameter service model (frame size, guaranteed delivery ratio, and reliability), then show that the probabilistic constraint can be safely replaced by linear upper-confidence bounds without losing much feasibility. On this linearized model they run a Modified Drift-plus-Penalty (MDP) policy that stabilizes queues and maximizes utility. The central result is that the MDP policy supports an arrival-rate region arbitrarily close to the true one, and its utility gap vanishes as the frame length grows. The paper also proves that randomized scheduling is sometimes necessary to meet such probabilistic SLAs, something deterministic schedulers cannot do.

Core claim

The paper establishes that probabilistic QoS constraints of the form "deliver at least gamma fraction of service with probability at least q within each frame" can be enforced by a linearized feasible domain, and that a drift-plus-penalty algorithm operating on that domain is near-optimal. Specifically, under a Slater condition and a mild condition on the frame length, the MDP policy achieves mean rate stability and a utility gap of O(1/sqrt(Ts)) compared to the optimal policy, where Ts is the frame size. The linearized domain sits within the true feasible region, and the true region sits within the linearized domain after shrinking the guaranteed throughput by O(1/sqrt(Ts)); hence the stabi

What carries the argument

The key object is the linearized feasible domain eD(gamma,q), which replaces the intractable Poisson-binomial probabilistic constraint with a set of linear upper-confidence-bound (UCB) constraints. The UCB term B_c^i(Gamma,p) is written as a maximum over subsets of service opportunities and then, via duality, as a polynomial-size polyhedron. The MDP policy is the standard drift-plus-penalty rule applied to this polyhedron: each frame it maximizes the queue-weighted service rate plus V times the utility, over linear constraints. The approximation tightness, stability region characterization, and vanishing gap all rest on concentration properties of the Poisson-binomial distribution and on Lya

Load-bearing premise

The guarantees require a strictly feasible interior schedule (Slater condition) with positive slack zeta; when the QoS contract sits very close to the boundary of feasibility, the stated 1/sqrt(Ts) gap can carry a very large constant.

What would settle it

Construct a single-server, 101-client instance placed exactly on the boundary of the linearized feasible domain so that the Slater slack zeta is zero, then run the MDP policy for increasing frame sizes Ts and measure the actual utility gap; if the gap does not shrink as Ts grows, the vanishing-gap claim of Theorem 5 would be contradicted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the MDP guarantees hold, an orchestrator can offer SLAs with probabilistic throughput and delay bounds while still keeping all request queues stable, meaning both hard-deadline and soft-deadline applications can share the same infrastructure.
  • The vanishing gap with frame size means that choosing longer orchestration frames—which naturally aligns with coarser but more latency-tolerant applications—yields near-optimal utility without sacrificing feasibility.
  • Randomized scheduling is shown to be necessary in some feasible QoS scenarios; deterministic schedulers, which dominate current practice, may silently violate probabilistic SLAs even when a feasible randomized policy exists.
  • The explicit delay bound (Theorem 6) gives network operators a closed-form condition on gamma and q to meet a target average queueing delay, turning the QoS model into a practical SLA design tool.
  • Because the model reduces to classical NUM when q=0 and to probability-1 delivery-ratio guarantees as q approaches 1, it unifies existing throughput-stability formulations with deadline-based QoS models.
  • The same linearization recipe—using upper-confidence bounds on bounded independent service indicators—should extend to other service processes beyond the Poisson-binomial model, since the concentration argument only needs boundedness and independence.
  • In practice, the O(1/sqrt(Ts)) approximation error suggests an explicit trade-off: operators can set the frame length to balance latency strictness against conservatism in the guaranteed throughput gamma.
  • The necessity of randomized scheduling hints that mixed-integer or deterministic schedulers deployed in real infrastructure-sharing platforms may need a randomization layer to honor probabilistic SLAs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper studies QoS-aware resource provisioning in a shared multi-infrastructure network. It introduces a three-parameter SLA (frame length T_s, guaranteed delivery ratio γ, reliability q) and formulates a network utility maximization problem (P) with per-frame probabilistic delivery constraints and queue stability. To make the problem tractable, the probabilistic constraints are replaced by linearized upper-confidence-bound constraints, yielding a polyhedral approximate domain eD(γ,q) (Theorems 1–2). The main theoretical results are Theorem 3 (eD and D are O(1/sqrt(T_s))-close), Theorem 4 (polyhedral stability region), and Theorem 5/Corollary 2 (the MDP policy supports a near-true stability region and attains an O(1/sqrt(T_s)) utility gap). A numerical study for video streaming, environmental monitoring, and backup traffic illustrates the framework.

Significance. If the main theorems hold, this is a useful contribution: the closed-form formula for Γ in (18) avoids fitted parameters, the polyhedral reformulation makes the per-frame optimization convex, and the randomized-scheduling necessity example in Appendix E is a nice structural observation. The paper also makes falsifiable predictions about stability and optimality that are testable in simulation. However, the central approximation theorem currently has a load-bearing proof gap in Appendix A.3; until that step is repaired, the stability-region closeness and vanishing optimality gap of the MDP policy are not established as written.

major comments (2)
  1. [Appendix A.3, Eq. (48)–(49); Theorem 3] The derivation of the second inequality in (48) is algebraically unsupported. From Lemma 5, Var[μ_i^c(t)] ≥ A T_s q_i^c with A=(1−r_max)|K|γ_i^c, so 0.7915/sqrt(Var) ≤ 0.7915/sqrt(A T_s q_i^c). The proof then replaces this by q_i^c/(sqrt(K_i^c) sqrt(T_s)). This requires 0.7915/sqrt(A T_s q_i^c) ≤ q_i^c/(sqrt(K_i^c) sqrt(T_s)). Substituting K_i^c=0.7952/(A (q_i^c)^3) from (21), the required condition becomes A (q_i^c)^3 ≥ 0.7058. Condition (21) only gives A (q_i^c)^3 > 0.7952/T_s, which is far weaker and can be arbitrarily small. In the paper's own simulation setup (T_s=300, r_max=0.9, |K|=10, γ=0.0204, q=0.99), A q^3 ≈ 0.0198, an order of magnitude below 0.7058. Thus Eq. (49), the second inclusion in (22), and all downstream results that invoke Theorem 3 (Corollary 1, Lemma 2, Theorem 5, Corollary 2) are not proved as written.
  2. [Corollary 2 and Theorem 5] The asymptotic statement is ambiguous when both T_s and T are allowed to grow. From (30), mean-rate stability requires sqrt(V T) + T_s sqrt(T) = o(T). If V=Θ(T^β), this holds for fixed T_s, but if T_s grows, e.g. T_s=Θ(sqrt(T)), then T_s sqrt(T) is not o(T). Similarly, the utility gap in (31) has the term T_s^2/V; with V=Θ(T^β) this is O(1/sqrt(T_s)) only under an additional joint growth condition that is not stated. The paper should either fix T_s and state the result as T→∞, or specify the relative growth of T_s and T needed for the 'vanishing with frame size' claim.
minor comments (5)
  1. [Abstract and Conclusion] Typographical errors: 'Dirft-plus-Penalty' in the abstract and 'gaurantees' in the conclusion should be corrected.
  2. [Notation, Eq. (1)] The definition [n]={0,...,n−1} is used inconsistently with expressions such as τ∈[T_s], which by the definition would be {0,...,T_s−1}, matching the intended range. Please align the notation throughout.
  3. [Appendix D.1, Lemma 12, Eq. (101)] The bound in Eq. (101) is written with the infinity norm ∥γ−γ′∥∞, while the preceding derivation and the final statement of Lemma 12 use the 1-norm ∥γ−γ′∥1. The norm should be made consistent or the constants adjusted.
  4. [Section 6.2 and Theorem 3] The text states that the simulation parameters satisfy condition (21) and Assumption 1. They do satisfy Ts>K_i^c, but they do not satisfy the implicit requirement A q^3 ≥ 0.7058 needed in the proof of (48). The simulation therefore cannot be used as evidence for Theorem 3 as currently proved; it should be revisited once the condition in Theorem 3 is repaired.
  5. [Section 6.3] The simulation chooses V=5T^{1/3}. This is a legitimate tuning choice, but the paper does not explain how it relates to the theoretical prescriptions V=Θ(min{T_s^{2.5},T^β}) in Corollary 2, since T_s^{2.5}≈1.56×10^6 for T_s=300 and the chosen V is much smaller.

Circularity Check

0 steps flagged

No significant circularity: the QoS guarantees are derived from closed-form UCB parameters and external approximation results, not fitted to the claimed predictions.

full rationale

I walked the derivation chain from the QoS model (Section 3), through the linearized feasible domain (Section 4), to the stability and optimality guarantees (Section 5), and found no step in which a prediction is equivalent to its inputs by construction. The UCB budget Gamma is set by the closed-form expression (18) to match the reliability q exactly; it is not estimated from data. Theorem 1 proves that the linearized constraint implies the probabilistic QoS constraint using a concentration inequality from Bertsimas and Sim [3]. Theorem 3 derives the two-sided inclusion eD(gamma,q) subset D(gamma,q) subset eD(gamma - K1/sqrt(Ts),q) from an external Gaussian-approximation bound for Poisson binomial variables [31], a variance lower bound proved from the QoS constraint itself, and a bound on the inverse normal CDF [5]. The Slater condition (Assumption 1) is stated as an explicit assumption, not derived from the target result, and the constants K1-K4 are explicit functions of the problem parameters, not fitted quantities. Lemma 2 and Theorem 5 then use standard Lyapunov drift arguments with these explicit bounds. The simulations in Section 6 are presented after the theoretical results and are used for validation, not as inputs. Self-citations to [19] and [29] are used as baseline comparisons or standard technical lemmas (e.g., the Skorokhod representation), not as load-bearing justification for the central claim. The skeptic's algebraic objection to inequality (48) is a correctness concern about whether Theorem 3 is valid under condition (21); it does not indicate circularity, because the proof does not assume the inclusion it claims to establish. A false theorem is not the same as a theorem whose conclusion is built into its assumptions. Therefore, I find no significant circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

No fitted constants appear in the theory. The only tunable hyper-parameter is V, standard in drift-plus-penalty. All QoS parameters (Ts, gamma, q) are SLA inputs, not fitted. No new physical or conceptual entities such as particles or forces are postulated.

free parameters (1)
  • V (drift-plus-penalty weight) = 5*T^(1/3) in simulations
    Standard DP hyperparameter trading off utility and backlog; it is a user-chosen tuning parameter, not fitted to data, and the theory holds for any V>0.
axioms (5)
  • standard math Poisson binomial distribution and Gaussian approximation bound from [31, Theorem 3.3]
    Used to prove Theorem 3's closeness of linearized and original feasible domains.
  • standard math Concentration inequality for independent bounded random variables [3, Theorem 2]
    Used in Theorem 1 to translate the linearized UCB constraint into the probabilistic QoS guarantee.
  • standard math Exterior penalty function convergence [33]
    Used in Lemmas 8 and 14 to convert constrained optimization problems to unconstrained penalized forms.
  • domain assumption Assumption 1 (Slater condition with zeta>0)
    Needed for the optimality gap bounds in Lemma 2/3 and Theorem 5; not proven, stated as an assumption.
  • domain assumption Arrival second moment finite and system non-empty (delta_min > 0)
    Used for the delay bound in Theorem 6 and the stability guarantees in Theorem 5.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A QoS Framework for Service Provision in Multi-Infrastructure-Sharing Networks." pith.science (2026). https://pith.science/paper/OWUHMX2L

@misc{pith2026250901694,
  author       = {Pith},
  title        = {Pith review of: A QoS Framework for Service Provision in Multi-Infrastructure-Sharing Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWUHMX2L}},
  note         = {Machine review of arXiv:2509.01694}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We propose a framework for resource provisioning with QoS guarantees in shared infrastructure networks. Our novel framework provides tunable probabilistic service guarantees for throughput and delay. Key to our approach is a Modified Dirft-plus-Penalty (MDP) policy that ensures long-term stability while capturing short-term probabilistic service guarantees using linearized upper-confidence bounds. We characterize the feasible region of service guarantees and show that our MDP procedure achieves mean rate stability and an optimality gap that vanishes with the frame size over which service guarantees are provided. Finally, empirical simulations validate our theory and demonstrate the favorable performance of our algorithm in handling QoS in multi-infrastructure networks.

Figures

Figures reproduced from arXiv: 2509.01694 by Eytan Modiano, Quang Minh Nguyen.

Figure 1
Figure 1. Figure 1: Infrastructure sharing in NextG networks. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Sequence of events in our two time-scales model. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The MDP with our QoS model achieves better QoS than the DP without QoS in all applications. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: OTT service of type-3 job at client 20. The baseline without QoS over-provisions service for type-3 with no strict requirements within a time frame, i.e. 𝑞 3 𝑖 = 0 and 𝛾 3 𝑖 = 0. policies. We consider 𝛼-fair utility [2] where client 𝑖 receives util￾ity 𝑓 𝑐 𝑖 (𝑥) = 𝑤𝑐 𝑥 1−𝛼𝑐 1−𝛼𝑐 given the total service 𝑥 for type-𝑐. We set (𝑤1,𝑤2,𝑤3) = (0.106, 0.516, 0.7) and (𝛼1, 𝛼2, 𝛼3) = (0.5, 0.875, 0.75). Note that E[… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages · 3 internal anchors

  1. [1]

    Mai A Abdel-Malek, Muhammad M Sayed, and Mohamed Azab. 2022. UAV- based privacy-preserved trustworthy seamless service agility for NextG cellular networks.Sensors (Basel)22, 7 (April 2022), 2756

    work page 2022

  2. [2]

    Eitan Altman, Konstantin Avrachenkov, and Andrey Garnaev. 2008. Generalized 𝛼-fair resource allocation in wireless networks. In2008 47th IEEE Conference on Decision and Control. 2414–2419. doi:10.1109/CDC.2008.4738709

    work page arXiv 2008

  3. [3]

    Dimitris Bertsimas and Melvyn Sim. 2004. The Price of Robustness.Oper. Res.52 (2004), 35–53. https://api.semanticscholar.org/CorpusID:8946639

    work page 2004

  4. [4]

    2004.Convex Optimization

    Stephen Boyd and Lieven Vandenberghe. 2004.Convex Optimization. Cambridge University Press

    2004

  5. [5]

    Learning by mistakes in memristor networks

    Seok-Ho Chang, Pamela C. Cosman, and Laurence B. Milstein. 2011. Chernoff- Type Bounds for the Gaussian Error Function.IEEE Transactions on Communica- tions59, 11 (2011), 2939–2944. doi:10.1109/TCOMM.2011.072011.100049

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  6. [6]

    Giovanni Farias da Silva, Francisco Vilar Brasileiro, Raquel Vigolvino Lopes, Fábio Jorge Almeida Morais, Marcus Carvalho, and Daniel Turull. 2020. QoS-driven scheduling in the cloud.Journal of Internet Services and Applications11 (2020). https://api.semanticscholar.org/CorpusID:226301495

    work page 2020

  7. [7]

    Mohammad Hadi, Mohammad Reza Pakravan, and Erik Agrell. 2019. Dynamic Resource Allocation in Metro Elastic Optical Networks Using Lyapunov Drift Optimization.J. Opt. Commun. Netw.11, 6 (Jun 2019), 250–259. doi:10.1364/JOCN. 11.000250

    doi:10.1364/jocn 2019

  8. [8]

    Clifford K Ho, Alex Robinson, David R Miller, and Mary J Davis. 2005. Overview of sensors and needs for environmental monitoring.Sensors (Basel)5, 1 (Feb. 2005), 4–37

    work page 2005

  9. [9]

    I.-H. Hou, V. Borkar, and P. R. Kumar. 2009. A Theory of QoS for Wireless. In IEEE INFOCOM 2009. 486–494. doi:10.1109/INFCOM.2009.5061954

    work page arXiv 2009

  10. [10]

    Longbo Huang and Michael J. Neely. 2013. Utility Optimal Scheduling in Energy- Harvesting Networks.IEEE/ACM Transactions on Networking21, 4 (2013), 1117–

    work page 2013

  11. [11]

    Satya Krishna Joshi, K. B. Shashika Manosha, Marian Codreanu, and Matti Latva- aho. 2017. Dynamic Inter-Operator Spectrum Sharing via Lyapunov Optimization. IEEE Transactions on Wireless Communications16, 10 (2017), 6365–6381. doi:10. 1109/TWC.2017.2722999

    work page arXiv 2017

  12. [13]

    An Experimental Study of Low-Latency Video Streaming over 5G

    Imran Khan, Tuyen X. Tran, Matti Hiltunen, Theodore Karagioules, and Dimitrios Koutsonikolas. 2024. An Experimental Study of Low-Latency Video Streaming over 5G. arXiv:2403.00752

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [14]

    Paschos, Panayotis Mertikopoulos, and Ulaş C

    Mathieu Leconte, Georgios S. Paschos, Panayotis Mertikopoulos, and Ulaş C. Kozat. 2018. A Resource Allocation Framework for Network Slicing. InIEEE INFOCOM 2018 - IEEE Conference on Computer Communications. 2177–2185. doi:10. 1109/INFOCOM.2018.8486303

    arXiv 2018

  14. [15]

    Qianru Li, Zhehui Zhang, Yanbing Liu, Zhaowei Tan, Chunyi Peng, and Songwu Lu. 2023. CA++: Enhancing Carrier Aggregation Beyond 5G. InProceedings of the 29th Annual International Conference on Mobile Computing and Networking (Madrid, Spain)(ACM MobiCom ’23). Association for Computing Machinery, New York, NY, USA, Article 5, 14 pages. doi:10.1145/3570361.3592500

    arXiv 2023

  15. [16]

    Bai Liu and Eytan Modiano. 2022. Universal Policy Tracking: Scheduling for Wireless Networks with Delayed State Observation. In2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton). 1–8. doi:10. 1109/Allerton49937.2022.9929348

    arXiv 2022

  16. [17]

    Bai Liu, Quang Minh Nguyen, Qingkai Liang, and Eytan Modiano. 2024. Tracking Drift-Plus-Penalty: Utility Maximization for Partially Observable and Control- lable Networks.IEEE/ACM Transactions on Networking32, 2 (2024), 1064–1079. doi:10.1109/TNET.2023.3307684

    work page arXiv 2024

  17. [18]

    Michael J. Neely. 2010.Stochastic Network Optimization with Application to Communication and Queueing Systems. Morgan and Claypool Publishers

    work page 2010

  18. [19]

    Neely, Eytan Modiano, and Chih-Ping Li

    Michael J. Neely, Eytan Modiano, and Chih-Ping Li. 2008. Fairness and Optimal Stochastic Control for Heterogeneous Networks.IEEE/ACM Transactions on Networking16, 2 (2008), 396–409. doi:10.1109/TNET.2007.900405

    work page arXiv 2008

  19. [20]

    Quang Minh Nguyen and Eytan Modiano. 2023. Learning to Schedule in Non- Stationary Wireless Networks With Unknown Statistics. InProceedings of the Twenty-Fourth International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing(Washington, DC, USA)(MobiHoc ’23). Association for Computing Machinery, New...

    arXiv 2023

  20. [21]

    Quang Minh Nguyen and Eytan Modiano. 2025. Optimal Control for Distributed Wireless SDN: Theory and Architecture.IEEE Transactions on Networking(2025), 1–17. doi:10.1109/TON.2025.3560605

    work page arXiv 2025

  21. [22]

    Quang Minh Nguyen and Eytan H. Modiano. 2024. Linear-Time Scheduling for Time-Varying Wireless Networks via Randomization. In2024 60th Annual Allerton Conference on Communication, Control, and Computing. 1–8. doi:10.1109/ Allerton63246.2024.10735302

    work page arXiv 2024

  22. [23]

    Shahir Rahman, Xinzhe Fu, Sastry Kompella, Joseph Macker, and Eytan H

    Quang Minh Nguyen, M. Shahir Rahman, Xinzhe Fu, Sastry Kompella, Joseph Macker, and Eytan H. Modiano. 2022. An Optimal Network Control Framework for Wireless SDN: From Theory to Implementation. InMILCOM 2022 - 2022 IEEE Military Communications Conference (MILCOM). 102–109. doi:10.1109/ MILCOM55135.2022.10017713

    work page arXiv 2022

  23. [24]

    Pedersen, Maciej Niparko, Jens Steiner, Jakub Oszmianski, Luke Mudolo, and Saeed R

    Klaus I. Pedersen, Maciej Niparko, Jens Steiner, Jakub Oszmianski, Luke Mudolo, and Saeed R. Khosravirad. 2016. System Level Analysis of Dynamic User-Centric Scheduling for a Flexible 5G Design. In2016 IEEE Global Communications Confer- ence (GLOBECOM). 1–6. doi:10.1109/GLOCOM.2016.7842312

    work page arXiv 2016

  24. [25]

    Qiaofeng Qin, Nakjung Choi, Muntasir Raihan Rahman, Marina Thottan, and Leandros Tassiulas. 2020. Network Slicing in Heterogeneous Software-defined RANs. InIEEE INFOCOM 2020 - IEEE Conference on Computer Communications. 2371–2380. doi:10.1109/INFOCOM41043.2020.9155532

    arXiv 2020

  25. [26]

    Rajkumar, Chen Lee, J.P

    R. Rajkumar, Chen Lee, J.P. Lehoczky, and D.P. Siewiorek. 1998. Practical solutions for QoS-based resource allocation problems. InProceedings 19th IEEE Real-Time Systems Symposium (Cat. No.98CB36279). 296–306. doi:10.1109/REAL.1998.739755

    work page arXiv 1998

  26. [27]

    Veselin Rakocevic, John Griffiths, and Graham Cope. 2001. Performance anal- ysis of bandwidth allocation schemes in multiservice IP networks using utility functions. InTeletraffic Engineering in the Internet Era, Jorge Moreira de Souza, Nelson L.S. da Fonseca, and Edmundo A. de Souza e Silva (Eds.). Teletraffic Science and Engineering, Vol. 4. Elsevier, 2...

    work page doi:10.1016/s1388-3437(01)80125-6 2001

  27. [28]

    Samdanis, F

    K. Samdanis, F. G. Mir, D. Kutscher, and T. Taleb. 2013. Service Boost: Towards on- demand QoS enhancements for OTT apps in LTE. In2013 21st IEEE International Conference on Network Protocols (ICNP). 1–6. doi:10.1109/ICNP.2013.6733663

    work page arXiv 2013

  28. [29]

    Abhishek Sinha and Eytan Modiano. 2018. Optimal Control for Generalized Network-Flow Problems.IEEE/ACM Transactions on Networking26, 1 (2018), 506–519. doi:10.1109/TNET.2017.2783846

    arXiv 2018

  29. [30]

    Stine and Carlos E

    John A. Stine and Carlos E. Caicedo Bastidas. 2014. Service level agreements with spectrum consumption models. In2014 IEEE International Symposium on Dynamic Spectrum Access Networks (DYSPAN). 206–214. doi:10.1109/DySPAN.2014.6817797

    work page arXiv 2014

  30. [31]

    Wenpin Tang and Fengmin Tang. 2019. The Poisson binomial distribution – Old & New. arXiv:1908.10024 [math.PR]

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  31. [32]

    Min Ye Thu, Wunna Htun, Yan Lin Aung, Pyone Ei Ei Shwe, and Nay Min Tun. 2018. Smart Air Quality Monitoring System with LoRaWAN. In2018 IEEE International Conference on Internet of Things and Intelligence System (IOTAIS). 10–15. doi:10.1109/IOTAIS.2018.8600904

    work page arXiv 2018

  32. [33]

    Klaus Truemper. 1975. Note on Finite Convergence of Exterior Penalty Functions. Management Science21, 5 (1975), 600–606. http://www.jstor.org/stable/2630043

    work page arXiv 1975

  33. [34]

    Hua Wang, Claudio Rosa, and Klaus Pedersen. 2010. Performance of Uplink Carrier Aggregation in LTE-Advanced Systems. In2010 IEEE 72nd Vehicular Technology Conference - Fall. 1–5. doi:10.1109/VETECF.2010.5594543

    work page arXiv 2010

  34. [35]

    Qin Xin and Jie Xiang. 2009. Joint QoS-aware admission control, channel as- signment, and power allocation for cognitive radio cellular networks. In2009 IEEE 6th International Conference on Mobile Adhoc and Sensor Systems. 294–303. doi:10.1109/MOBHOC.2009.5336984

    work page arXiv 2009

  35. [36]

    Shah Zeb, Muhammad Ahmad Rathore, Syed Ali Hassan, Saleem Raza, Kapal Dev, and Giancarlo Fortino. 2023. Toward AI-Enabled NextG Networks with Edge Intelligence-Assisted Microservice Orchestration.IEEE Wireless Communications 30, 3 (2023), 148–156. doi:10.1109/MWC.015.2200461 A QoS Framework for Service Provision in Multi-Infrastructure-Sharing Networks Mo...

    work page doi:10.1109/mwc.015.2200461 2023

  36. [38]

    be the set and index respectively that achieves the maximum in (15), and consider(𝑘𝑟,𝜏𝑟)=argmin (𝑘,𝜏)∈𝑆 𝑐∗ 𝑡𝑖∪(𝑘∗ 0,𝜏∗

    work page

  37. [39]

    normalized

    ˆ𝑝𝑐 𝑘𝑖(𝑡,𝜏) . We also define the random variable 𝜂𝑐 𝑘𝑖(𝑡,𝜏)= −𝑦𝑐 𝑘𝑖(𝑡,𝜏)+𝑟𝑘𝑖𝑝𝑐 𝑘𝑖(𝑡,𝜏) ˆ𝑝𝑐 𝑘𝑖(𝑡,𝜏) ∈ [−1, 1], which is the “normalized" version of𝑧𝑐 𝑘𝑖(𝑡,𝜏)=−𝑦 𝑐 𝑘𝑖(𝑡,𝜏) to be used later in the proof. We define the weights: 𝑤𝑐 𝑘𝑖(𝑡,𝜏)=    1,if(𝑘,𝜏)∈𝑆 𝑐∗ 𝑡𝑖 ˆ𝑝𝑐 𝑘𝑖(𝑡,𝜏) ˆ𝑝𝑐 𝑘𝑟𝑖(𝑡,𝜏𝑟),if(𝑘,𝜏)∈𝑁(𝑖)×[𝑇 𝑠]\𝑆 𝑐∗ 𝑡𝑖 . Now, we have: 𝑃 1 |K|𝑇𝑠 𝜇𝑐 𝑖(𝑡)≤𝛾 𝑐 𝑖 ...

    work page

  38. [40]

    Then, we can increase the objective value in the maximization problem (15) by exchanging(𝑘∗ 0,𝜏∗

    (otherwise,𝑤𝑐 𝑘𝑖(𝑡,𝜏) = 1) and hence (𝑘𝑟,𝜏𝑟) ∈𝑆𝑐∗ 𝑡𝑖 . Then, we can increase the objective value in the maximization problem (15) by exchanging(𝑘∗ 0,𝜏∗

    work page

  39. [41]

    In both cases, we obtain a new solution other than𝑆𝑐∗ 𝑡𝑖 ∪(𝑘∗ 0,𝜏∗

    with(𝑘𝑟,𝜏𝑟). In both cases, we obtain a new solution other than𝑆𝑐∗ 𝑡𝑖 ∪(𝑘∗ 0,𝜏∗

    work page

  40. [42]

    that increases the objective value of (15), which is a contradiction. Given that 𝑤𝑐 𝑘𝑖(𝑡,𝜏) ∈ [0, 1] for all(𝑘,𝜏) ∈𝑁(𝑖)×[𝑇 𝑠] and 𝜂𝑐 𝑘𝑖(𝑡,𝜏) ’s are independent and symmetrically distributed in[−1, 1], we use [3, Theorem 2] to obtain that: 𝑃 ∑︁ (𝑘,𝜏)∈𝑁(𝑖)×[𝑇 𝑠] 𝜂𝑐 𝑘𝑖(𝑡,𝜏)𝑤 𝑐 𝑘𝑖(𝑡,𝜏)≥Γ 𝑐 𝑖 ≤𝑒− (Γ𝑐 𝑖)2 2|K|𝑇𝑠.(41) Finally, combining (40) and (41), we conclud...

    work page

  41. [43]

    Since strong duality holds for any feasible linear optimization problem,𝐵𝑐 𝑖(Γ𝑐 𝑖,p(𝑡))admits the following dual form of (42): 𝐵𝑐 𝑖(Γ𝑐 𝑖,p(𝑡))=min ∑︁ 𝑘∈K 𝑇𝑠−1∑︁ 𝜏=0 𝑣𝑐 𝑖,𝑘𝜏+Γ 𝑐 𝑖𝑠𝑐 𝑖 (43) such that𝑠 𝑐 𝑖+𝑣𝑐 𝑘𝑖𝜏≥𝑟 𝑘𝑖𝑝𝑐 𝑘𝑖(𝑡,𝜏), ∀𝑘∈𝑁(𝑖),𝜏∈[𝑇 𝑠] 𝑠𝑐 𝑖+𝑣𝑐 𝑘𝑖𝜏≥1−𝑟 𝑘𝑖𝑝𝑐 𝑘𝑖(𝑡,𝜏), ∀𝑘∈𝑁(𝑖),𝜏∈[𝑇 𝑠] 𝑣𝑐 𝑘𝑖𝜏≥0,∀𝑘∈𝑁(𝑖),𝜏∈[𝑇 𝑠] 𝑠𝑐 𝑖≥0. Substituting (43) into the linearize...

    work page 2025

  42. [1130]

    doi:10.1109/TNET.2012.2230336

    work page arXiv 2012

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.