pith. sign in

arxiv: 2604.19231 · v1 · submitted 2026-04-21 · 💻 cs.IT · eess.SP· math.IT

Reliable Remote Inference from Unreliable Components: Joint Communication and Computation Limits

Zhenyu Liu , Yi Ma , Rahim Tafazolli This is my paper

Pith reviewed 2026-05-10 01:43 UTC · model grok-4.3

classification 💻 cs.IT eess.SPmath.IT
keywords remote inferenceunreliable computationjoint communication and computationmin-cut converseinformation bottleneckshard separationcommitted receiver closure
0
0 comments X

The pith

Unreliable receivers force extra min-cuts on inference tasks

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

This paper examines remote inference when both the communication link and the receiver's processing elements are noisy and subject to a finite redundancy budget. It establishes that under a committed receiver model without explicit bypasses, task-relevant information must traverse a collection of vulnerable primitives and is therefore subject to serial cuts inside the computation graph in addition to the external channel limit. The central result is a receiver-internal compute min-cut converse that captures these extra first-order bottlenecks. The twofold loss seen in symmetric two-stage hard-separation architectures is shown to arise from the hard-separation choice combined with the committed closure rather than from unreliability itself. When downstream modules retain soft visibility to raw channel outputs, the bound collapses back to the single communication bottleneck.

Core claim

Under the committed/no-bypass receiver closure, task-relevant information can affect the final estimate only by passing through a budgeted collection of vulnerable primitives modeled as memoryless noisy channels. This produces a baseline supply-demand converse: the information needed for a target distortion cannot exceed the smaller of the total supply from the communication channel and the total supply from the vulnerable compute budget. Committed intermediate interfaces create additional first-order serial cuts and receiver-internal computation-graph cuts, captured by a general receiver-internal compute min-cut converse. The twofold loss in the symmetric two-stage hard-separation special c

What carries the argument

The receiver-internal compute min-cut converse, which identifies the serial cuts created by committed intermediate interfaces inside the computation graph under the committed/no-bypass closure.

If this is right

  • Hard-separation under committed interfaces incurs an extra first-order information tax that is absent when soft visibility to raw channel outputs is retained.
  • The twofold loss in symmetric two-stage hard-separation is induced by the closure and can be avoided by preserving soft information paths.
  • Under a stronger protected-support closure with reliable decoder and control support, both task-direct and serial hard-separation constructions achieve the information limits.
  • In the fully noisy-logic regime only a conservative depth-dependent converse is obtained, and matched achievability remains open.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Hardware interfaces that allow explicit bypasses or soft visibility around vulnerable primitives could eliminate the additional serial cuts identified by the min-cut bound.
  • System designers may benefit from prioritizing reliable control and soft-output paths when building inference hardware from noisy components.
  • Relaxing the memoryless-channel model for primitives could yield tighter or qualitatively different min-cut characterizations for real devices.

Load-bearing premise

The receiver operates under a committed/no-bypass closure in which all information must pass through vulnerable primitives without explicit protected bypasses, together with the modeling of each primitive as a memoryless noisy channel.

What would settle it

A concrete receiver construction or simulation under committed interfaces that achieves lower distortion than the receiver-internal compute min-cut bound predicts, or that removes the twofold loss in symmetric two-stage hard-separation without relaxing the closure.

Figures

Figures reproduced from arXiv: 2604.19231 by Rahim Tafazolli, Yi Ma, Zhenyu Liu.

Figure 1
Figure 1. Figure 1: Protected-side-information extension of the committed/no-bypass receiver model. The protected side variable [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Two baseline receiver organizations under the same vulnerable-compute budget [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Toy computation graphs for Example IV.10. Under a fixed total interface budget [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic “equivalent supply” picture for hard-separation with unequal stage reliabilities. The optimal split in Corollary IV.16 equalizes the active [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Unified “architecture phase diagram” in the supply plane. Each curve separates the channel-limited and compute-limited regions for one interface [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Finite-blocklength benchmark for scalar Gaussian remote estimation under the normal approximation (Remark A.7). The vulnerable-compute task-direct [PITH_FULL_IMAGE:figures/full_fig_p032_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Square-root illustration in normalized excess distortion for scalar Gaussian remote estimation in the compute-limited regime. The task-direct curve [PITH_FULL_IMAGE:figures/full_fig_p033_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Diagonal vector Gaussian example for the [PITH_FULL_IMAGE:figures/full_fig_p044_8.png] view at source ↗
read the original abstract

Classical information theory typically assumes reliable receiver-side processing. We study remote inference when communication is noisy and the receiver itself is built from unreliable components under a finite redundancy budget. Under a committed/no-bypass receiver closure, task-relevant information can affect the final estimate only by passing through a budgeted collection of vulnerable primitives unless an explicit protected bypass is modeled. Modeling each vulnerable primitive as a memoryless noisy channel yields a baseline supply--demand converse: the task-relevant information needed to attain a target distortion cannot exceed the smaller of the total information supplied by the communication channel and the total information supplied by the vulnerable compute budget. Our main converse shows that committed intermediate interfaces create additional first-order serial cuts and receiver-internal computation-graph cuts, captured in general by a receiver-internal compute min-cut converse. In particular, the twofold loss in the symmetric two-stage hard-separation special case is not inherent to unreliable receiver computation but induced by hard-separation under the committed/no-bypass closure. This extra first-order tax is therefore closure-dependent rather than universal. On the converse side, if downstream modules retain soft visibility to the raw channel output, the converse reduces to the single-bottleneck supply, up to any explicitly reserved soft-path budget. Under a separate stronger protected-support closure with reliable decoder and control support, we establish achievability results for task-direct and serial hard-separation constructions. For the fully noisy-logic regime, we obtain only a conservative depth-dependent converse, and matched achievability remains open.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper studies remote inference under noisy communication and unreliable receiver computation with a finite redundancy budget. Under the committed/no-bypass receiver closure, it derives a baseline supply-demand converse together with a receiver-internal compute min-cut converse that accounts for additional first-order serial cuts created by committed intermediate interfaces. It shows that the twofold loss observed in the symmetric two-stage hard-separation case is induced by the hard-separation assumption under this closure rather than being inherent to unreliable computation, contrasts this with the single-bottleneck supply that holds under soft-visibility or protected-support closures, establishes achievability for task-direct and serial hard-separation constructions under the stronger protected-support closure, and provides only a conservative depth-dependent converse for the fully noisy-logic regime with achievability left open.

Significance. If the converses and achievability results hold, the work supplies a principled min-cut framework for analyzing joint communication-computation limits when receiver primitives are modeled as memoryless noisy channels. The explicit dependence of the bounds on receiver closure assumptions (committed/no-bypass versus protected-support) is a useful distinction that clarifies when unreliable computation imposes extra first-order costs beyond the communication bottleneck. This extends classical information-theoretic tools to computation graphs with noise and could inform the design of reliable inference pipelines built from unreliable hardware components.

major comments (1)
  1. [Main converse] The main converse (receiver-internal compute min-cut): the manuscript must supply the explicit definition of the computation-graph cuts and the precise way they produce the additional first-order serial cuts; without this derivation the claim that the twofold loss is closure-dependent rather than universal cannot be verified as load-bearing.
minor comments (2)
  1. The abstract is dense with multiple technical terms (committed/no-bypass closure, protected-support closure, fully noisy-logic regime); a short table or bullet list summarizing the four modeling regimes and their corresponding bounds would improve readability.
  2. Notation for the vulnerable primitives (each modeled as a memoryless noisy channel) should be introduced once with a consistent symbol and reused; currently the description shifts between 'vulnerable primitive' and 'noisy channel' without a single defining equation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address the major comment below and indicate the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Main converse] The main converse (receiver-internal compute min-cut): the manuscript must supply the explicit definition of the computation-graph cuts and the precise way they produce the additional first-order serial cuts; without this derivation the claim that the twofold loss is closure-dependent rather than universal cannot be verified as load-bearing.

    Authors: We agree that an explicit definition of the computation-graph cuts and a precise derivation of the additional first-order serial cuts are required to make the main converse and the closure-dependence claim fully verifiable. In the revised manuscript we will add a dedicated subsection (new Section III-C) that formally defines the receiver-internal computation-graph cuts under the committed/no-bypass closure, states the min-cut expression, and derives step-by-step how committed intermediate interfaces create the serial cuts. The subsection will also apply the construction to the symmetric two-stage hard-separation case to illustrate why the twofold loss is induced by the hard-separation assumption rather than being universal. This addition will directly address the referee's concern. revision: yes

Circularity Check

0 steps flagged

No significant circularity; bounds derived directly from stated models

full rationale

The paper's converses are obtained by applying standard min-cut arguments to an explicitly defined system model consisting of memoryless noisy channels for each vulnerable primitive together with the committed/no-bypass receiver closure. The supply-demand bound and the receiver-internal compute min-cut are direct consequences of these modeling choices rather than reductions to fitted parameters or self-referential definitions. Different closures are treated as alternative modeling assumptions whose consequences are compared; the twofold loss in the hard-separation case is shown to follow from the chosen closure rather than being asserted as universal. No load-bearing self-citations, ansatzes smuggled via prior work, or renaming of known results appear in the derivation chain. Achievability is established only under stronger closures while the fully noisy-logic regime is left open, confirming that the central claims remain independent of the target bounds.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on modeling unreliable primitives as memoryless noisy channels and on two distinct receiver closures (committed/no-bypass and protected-support). No free parameters or invented physical entities are introduced.

axioms (2)
  • domain assumption Each vulnerable primitive behaves as a memoryless noisy channel.
    Stated in the abstract as the modeling choice that yields the baseline supply-demand converse.
  • domain assumption The receiver operates under a committed/no-bypass closure unless an explicit protected bypass is modeled.
    This closure is invoked to derive the additional first-order serial cuts and the min-cut bound.

pith-pipeline@v0.9.0 · 5575 in / 1389 out tokens · 28414 ms · 2026-05-10T01:43:42.409949+00:00 · methodology

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

38 extracted references · 38 canonical work pages

  1. [1]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,”Bell Syst. Tech. J., vol. 27, no. 3, pp. 379–423, 1948

    work page 1948

  2. [2]

    Near-threshold computing: Reclaiming Moore’s law through energy efficient integrated circuits,

    R. G. Dreslinski, M. Wieckowski, D. Blaauw, D. Sylvester, and T. Mudge, “Near-threshold computing: Reclaiming Moore’s law through energy efficient integrated circuits,”Proc. IEEE, vol. 98, no. 2, pp. 253–266, Feb. 2010

    work page 2010

  3. [3]

    Approximate computing: An emerging paradigm for energy-efficient design,

    J. Han and M. Orshansky, “Approximate computing: An emerging paradigm for energy-efficient design,” inProc. 18th IEEE Eur. Test Symp. (ETS), 2013, pp. 1–6

    work page 2013

  4. [4]

    Razor: A low-power pipeline based on circuit-level timing speculation,

    D. Ernst, N. S. Kim, S. Das, S. Pant, R. Rao, T. Pham, C. Ziesler, D. Blaauw, T. Austin, K. Flautner, and T. Mudge, “Razor: A low-power pipeline based on circuit-level timing speculation,” inProc. 36th Annu. IEEE/ACM Int. Symp. Microarchitecture (MICRO-36), 2003, pp. 7–18

    work page 2003

  5. [5]

    Comprehensive evaluation of supply voltage underscaling in FPGA on-chip memories,

    B. Salami, O. S. Unsal, and A. Cristal Kestelman, “Comprehensive evaluation of supply voltage underscaling in FPGA on-chip memories,” inProc. 51st Annu. IEEE/ACM Int. Symp. Microarchitecture (MICRO), 2018, pp. 724–736

    work page 2018

  6. [6]

    Impact of voltage scaling on soft errors susceptibility of multicore server CPUs,

    D. Agiakatsikas, G. Papadimitriou, V . Karakostas, D. Gizopoulos, M. Psarakis, C. B ´elanger-Champagne, and E. Blackmore, “Impact of voltage scaling on soft errors susceptibility of multicore server CPUs,” inProc. 56th Annu. IEEE/ACM Int. Symp. Microarchitecture (MICRO). ACM, 2023, pp. 957–971

    work page 2023

  7. [7]

    REMU: Memory-aware radiation emulation via dual addressing for in-orbit deep learning system,

    L. Xu, M. Wang, H. Qiu, J. Liu, Y . Li, and H. Li, “REMU: Memory-aware radiation emulation via dual addressing for in-orbit deep learning system,” inProc. 62nd ACM/IEEE Des. Autom. Conf. (DAC), 2025, pp. 1–7

    work page 2025

  8. [8]

    A case for application-aware space radiation tolerance in orbital computing.arXiv preprint arXiv:2407.11853, 2024

    M. Wang, H. Qiu, L. Xu, D. Wang, Y . Li, T. Zhang, J. Liu, and H. Li, “A case for application-aware space radiation tolerance in orbital computing,” arXiv:2407.11853, Jul. 2024. [Online]. Available: https://arxiv.org/abs/2407.11853

    work page arXiv 2024

  9. [9]

    SA VE: Software-implemented fault tolerance for model inference against GPU memory bit flips,

    W. Zheng, B. Xu, J. Gu, and H. Chen, “SA VE: Software-implemented fault tolerance for model inference against GPU memory bit flips,” inProc. USENIX Annu. Tech. Conf. (USENIX ATC), 2025, pp. 1585–1604. [Online]. Available: https://www.usenix.org/conference/atc25/presentation/zheng

    work page 2025

  10. [10]

    Radshield: Software radiation protection for commodity hardware in space,

    H. Wang, S. Myint, V . Verma, Y . Winetraub, J. Yang, and A. Cidon, “Radshield: Software radiation protection for commodity hardware in space,” in Proc. 31st ACM Int. Conf. Architectural Support Program. Lang. Oper. Syst. (ASPLOS). New York, NY , USA: Association for Computing Machinery, 2026, pp. 219–233. [Online]. Available: https://doi.org/10.1145/3760...

    work page doi:10.1145/3760250.3762218 2026

  11. [11]

    Probabilistic logics and the synthesis of reliable organisms from unreliable components,

    J. von Neumann, “Probabilistic logics and the synthesis of reliable organisms from unreliable components,” inAutomata Studies, ser. Annals of Mathematics Studies, C. E. Shannon and J. McCarthy, Eds. Princeton, NJ, USA: Princeton University Press, 1956, no. 34, pp. 43–98

    work page 1956

  12. [12]

    Lower bounds for the complexity of reliable boolean circuits with noisy gates,

    P. G ´acs and A. G ´al, “Lower bounds for the complexity of reliable boolean circuits with noisy gates,”IEEE Trans. Inf. Theory, vol. 40, no. 2, pp. 579–583, Mar. 1994

    work page 1994

  13. [13]

    Signal propagation and noisy circuits,

    W. S. Evans and L. J. Schulman, “Signal propagation and noisy circuits,”IEEE Trans. Inf. Theory, vol. 45, no. 7, pp. 2367–2373, Nov. 1999

    work page 1999

  14. [14]

    On the capacity of noisy computations,

    F. Simon, “On the capacity of noisy computations,” inProc. 2011 IEEE Inf. Theory Workshop (ITW). IEEE, Oct. 2011, pp. 185–189

    work page 2011

  15. [15]

    Information theory and noisy computation,

    W. S. Evans, “Information theory and noisy computation,” Ph.D. dissertation, University of California at Berkeley, Berkeley, CA, USA, Nov. 1994

    work page 1994

  16. [16]

    Performance of LDPC codes under faulty iterative decoding,

    L. R. Varshney, “Performance of LDPC codes under faulty iterative decoding,”IEEE Trans. Inf. Theory, vol. 57, no. 7, pp. 4427–4444, Jul. 2011

    work page 2011

  17. [17]

    Belief propagation algorithms on noisy hardware,

    C.-H. Huang, Y . Li, and L. Dolecek, “Belief propagation algorithms on noisy hardware,”IEEE Trans. Commun., vol. 63, no. 1, pp. 11–24, Jan. 2015

    work page 2015

  18. [18]

    T. M. Cover and J. A. Thomas,Elements of Information Theory, 2nd ed. Hoboken, NJ, USA: Wiley-Interscience, 2006

    work page 2006

  19. [19]

    Berger,Rate Distortion Theory: A Mathematical Basis for Data Compression

    T. Berger,Rate Distortion Theory: A Mathematical Basis for Data Compression. Englewood Cliffs, NJ, USA: Prentice-Hall, 1971

    work page 1971

  20. [20]

    Indirect rate distortion problems,

    H. S. Witsenhausen, “Indirect rate distortion problems,”IEEE Trans. Inf. Theory, vol. 26, no. 5, pp. 518–521, Sep. 1980

    work page 1980

  21. [21]

    Beyond transmitting bits: Context, semantics, and task-oriented communications,

    D. G ¨und¨uz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, semantics, and task-oriented communications,”IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 5–41, 2023

    work page 2023

  22. [22]

    Deep joint source-channel coding for semantic communications,

    J. Xu, T.-Y . Tung, B. Ai, W. Chen, Y . Sun, and D. G ¨und¨uz, “Deep joint source-channel coding for semantic communications,”IEEE Commun. Mag., vol. 61, no. 11, pp. 42–48, 2023

    work page 2023

  23. [23]

    Information transmission with additional noise,

    R. Dobrushin and B. Tsybakov, “Information transmission with additional noise,”IRE Trans. Inf. Theory, vol. 8, no. 5, pp. 293–304, 1962

    work page 1962

  24. [24]

    Source encoding in the presence of random disturbance,

    D. Sakrison, “Source encoding in the presence of random disturbance,”IEEE Trans. Inf. Theory, vol. 14, no. 1, pp. 165–167, 1968

    work page 1968

  25. [25]

    Transmission of noisy information to a noisy receiver with minimum distortion,

    J. Wolf and J. Ziv, “Transmission of noisy information to a noisy receiver with minimum distortion,”IEEE Trans. Inf. Theory, vol. 16, no. 4, pp. 406–411, 1970

    work page 1970

  26. [26]

    Lossy joint source–channel coding in the finite blocklength regime,

    V . Kostina and S. Verd ´u, “Lossy joint source–channel coding in the finite blocklength regime,”IEEE Trans. Inf. Theory, vol. 59, no. 5, pp. 2545–2575, May 2013

    work page 2013

  27. [27]

    A systematic methodology for evaluating the error resilience of GPGPU applications,

    B. Fang, K. Pattabiraman, M. Ripeanu, and S. Gurumurthi, “A systematic methodology for evaluating the error resilience of GPGPU applications,”IEEE Trans. Parallel Distrib. Syst., vol. 27, no. 12, pp. 3397–3411, 2016

    work page 2016

  28. [28]

    GPU-Qin: A methodology for evaluating the error resilience of GPGPU applications,

    ——, “GPU-Qin: A methodology for evaluating the error resilience of GPGPU applications,” inProc. IEEE Int. Symp. Perform. Anal. Syst. Softw. (ISPASS), 2014, pp. 221–230

    work page 2014

  29. [29]

    SASSIFI: An architecture-level fault injection tool for GPU application resilience evaluation,

    S. K. S. Hari, T. Tsai, M. Stephenson, S. W. Keckler, and J. S. Emer, “SASSIFI: An architecture-level fault injection tool for GPU application resilience evaluation,” inProc. IEEE Int. Symp. Perform. Anal. Syst. Softw. (ISPASS), 2017, pp. 249–258

    work page 2017

  30. [30]

    NVBitFI: Dynamic fault injection for GPUs,

    T. Tsai, S. K. S. Hari, M. B. Sullivan, O. Villa, and S. W. Keckler, “NVBitFI: Dynamic fault injection for GPUs,” inProc. 51st Annu. IEEE/IFIP Int. Conf. Dependable Syst. Netw. (DSN), 2021, pp. 284–291

    work page 2021

  31. [31]

    Channel coding rate in the finite blocklength regime,

    Y . Polyanskiy, H. V . Poor, and S. Verd´u, “Channel coding rate in the finite blocklength regime,”IEEE Trans. Inf. Theory, vol. 56, no. 5, pp. 2307–2359, May 2010

    work page 2010

  32. [32]

    R. G. Gallager,Information Theory and Reliable Communication. New York, NY , USA: Wiley, 1968

    work page 1968

  33. [33]

    Fixed-length lossy compression in the finite blocklength regime,

    V . Kostina and S. Verd ´u, “Fixed-length lossy compression in the finite blocklength regime,”IEEE Trans. Inf. Theory, vol. 58, no. 6, pp. 3309–3338, Jun. 2012

    work page 2012

  34. [34]

    Deep joint source-channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source-channel coding for wireless image transmission,”IEEE Trans. on Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, Sep. 2019

    work page 2019

  35. [35]

    Nonlinear transform source-channel coding for semantic communications,

    J. Dai, S. Wang, K. Tan, Z. Si, X. Qin, K. Niu, and P. Zhang, “Nonlinear transform source-channel coding for semantic communications,”IEEE J. Sel. Areas Commun., vol. 40, no. 8, pp. 2300–2316, Aug. 2022

    work page 2022

  36. [36]

    Theoretical limitations on the transmission of data from analog sources,

    T. Goblick, “Theoretical limitations on the transmission of data from analog sources,”IEEE Trans. Inf. Theory, vol. 11, no. 4, pp. 558–567, 1965

    work page 1965

  37. [37]

    Yes, One-Bit-Flip matters! Universal DNN model inference depletion with runtime code fault injection,

    S. Li, X. Wang, M. Xue, H. Zhu, Z. Zhang, Y . Gao, W. Wu, and X. S. Shen, “Yes, One-Bit-Flip matters! Universal DNN model inference depletion with runtime code fault injection,” inProc. 33rd USENIX Secur. Symp. (USENIX Security). Philadelphia, PA: USENIX Association, Aug. 2024, pp. 1315–1330. [Online]. Available: https://www.usenix.org/conference/usenixse...

    work page 2024

  38. [38]

    Deep neural network approximation theory,

    D. Elbr ¨achter, D. Perekrestenko, P. Grohs, and H. B ¨olcskei, “Deep neural network approximation theory,”IEEE Trans. Inf. Theory, vol. 67, no. 5, pp. 2581–2623, May 2021

    work page 2021