Pith. sign in

REVIEW 3 major objections 5 minor 56 references

OLALa: Online Learned Adaptive Lattice Codes for Heterogeneous Federated Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read OLALa shows per-client time-varying lattice quantizers preserve FedAvg's O(1/T) convergence and beat fixed quantizers in FL experiments.

desk verdict OLALa is a sensible adaptive-lattice FL scheme with solid-looking experiments, but its convergence theorem does not cover the implemented algorithm—there is a scaling mismatch in the proof and a deliberate violation of the no-overload assumption. read the letter →

arxiv 2506.20297 v1 pith:PIMARESG submitted 2025-06-25 eess.SP cs.LG

classification eess.SPcs.LG
keywords federatedlearninglatticequantizationonlineadaptivesubtractiveditheringmodelcompressionheterogeneousdatacommunication-efficienttraining
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

The paper proposes OLALa, a way for each client in federated learning to adapt its own lattice quantizer round by round instead of using a fixed one. It proves that when the quantizers are never overloaded, FedAvg with these per-client time-varying dithered lattices still converges at O(1/T) (Theorem III.2), and that properly adapted lattices can tighten the bound. The method learns the generator matrix with a small fixed-input network and an exact differentiable reformulation, sending only a compact set of quantizer metadata. Experiments on MNIST and CIFAR-10 show consistent accuracy gains over fixed and statically learned lattices across quantization rates.

What carries the argument

The load-bearing object is the dithered lattice quantizer Q_SDQ with generator matrix G_t^u and support radius gamma; under the no-overload assumption, its distortion is white noise uniformly distributed over the basic lattice cell with variance $sigma^{2}$_SDQ(L_gamma(G)). The enabler for online learning is the exact reformulation Q_L(G)(x) = G * argmin_{l in I_L(G)} ||x - G l||, whose linear dependence on G gives an approximate gradient for training the generator, combined with a DNN augmentation (fixed input, output reshaped to G) that stabilizes learning and keeps the communicated metadata compact. The scale factor zeta controls the trade-off between avoiding overloading and reducing distortion, and the two candidate losses (empirical MSE and task-based learning objective) drive the adaptation.

What would settle it

Run OLALa on MNIST-CNN with overloading thresholds of 1%, 5%, and 10% and record E[F(w_T)] - F(w*) against T; if the empirical convergence rate visibly slows below O(1/T) or the quantization error becomes input-dependent once overloading exceeds a small fraction, the assumption behind Theorem III.2 fails.

Watch

Extended reading notes

Core claim

The central claim is that adaptive, client-specific lattice quantizers are not only feasible in federated learning but beneficial: the convergence rate of FedAvg is preserved at O(1/T) even when every user changes its lattice every round, and the non-asymptotic convergence bound is minimized by letting lattices vary per user and per round (Theorem III.3). This generalizes the fixed-lattice FL convergence result [19] to time-varying, user-dependent generator matrices by bounding the aggregated quantization noise with the average SDQ distortion (Theorem III.1). The method then realizes this adaptivity by learning the generator matrix through a small DNN with a fixed input, using an exact reformulation of lattice quantization that supports gradient-based online training, and by scaling updates with a factor zeta to keep overloading small. Empirically, OLALa consistently outperforms fixed and statically learned lattice quantizers in both final accuracy and convergence speed.

Load-bearing premise

The analysis assumes every client's lattice is never overloaded in any round (AS3), yet OLALa deliberately permits small overloading by scaling updates; if overloading becomes non-negligible, the white-noise distortion model and the O(1/T) convergence guarantee are no longer proven.

Editorial extensions

If this is right

  • Federated training keeps its O(1/T) convergence rate even when every user switches its lattice every round, so personalizing compression does not cost asymptotic performance.
  • The non-asymptotic convergence bound contains the average SDQ distortion, so any scheme that lowers per-user lattice distortion tightens the finite-time guarantee.
  • Because the generator matrix is rate-independent and small (L by L), communicating per-round quantizer metadata is negligible relative to the model updates themselves.
  • Permitting small, controlled overloading (around 0.5% to 1%) gives better accuracy than strictly avoiding it, which is why OLALa scales updates by zeta rather than forcing no overload.

Reading between the lines

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

  • Editorial inference: the same differentiable-lattice recipe could be dropped into other distributed optimization loops, such as sparsified or periodic-averaging FL, as long as the shared dither and approximate no-overload regime are maintained.
  • Editorial inference: because the theory only requires the distortion terms B_t to be bounded, one could try to learn lattices by directly minimizing the convergence bound rather than local MSE; the paper's task-based loss already moves in that direction.
  • Editorial inference: a testable extension is to let the dither distribution itself adapt along with the lattice; the authors list it as future work, and the white-noise property would then need to be re-derived.
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

3 major / 5 minor

Summary. The paper proposes OLALa, a federated learning compression scheme in which each client learns a lattice generator matrix online and transmits quantized model updates together with compact lattice metadata. The authors state convergence guarantees for FL with time-varying, client-specific lattice quantizers (Theorems III.1-III.3), argue that per-client adaptation can tighten the convergence bound, and support the method with experiments on MNIST and CIFAR-10 against fixed and statically learned lattice baselines. The central theoretical claim is that OLALa preserves the O(1/T) convergence rate of FedAvg.

Significance. The idea of replacing a single fixed lattice with online-learned, per-client lattices is natural, and the experimental comparison is useful: the code is provided, the baselines are external, and OLALa consistently improves accuracy over fixed alternatives, e.g., 93.00% versus 90.36% for Fixed-Hexagon at R=3 on MNIST-CNN (Table III). However, the theoretical analysis is not currently connected to the algorithm actually evaluated. The proof quantizes gradients, while the algorithm quantizes step-size-scaled updates, and the algorithm deliberately permits overloading while Assumption AS3 forbids it. These are load-bearing gaps, so the paper's central theoretical claim requires substantial revision before the guarantees can be accepted.

major comments (3)
  1. [Section III-B and Appendix C (Eq. (A.3), Lemma C.1)] The proof does not match Algorithm 1. Algorithm 1 and Section III-B take the local update as \tilde h_t^u = -\eta_t \nabla F_u(\tilde w_t, i_t^u) and the server aggregates Q^{SDQ}_{L}(\tilde h_t^u) in Eq. (11). However, \widehat{g}^{SDQ}_t in Eq. (A.3) is defined as (1/U) \sum_u Q^{SDQ}_{L}(\nabla F_u(\tilde w_t, i_t^u)), and Lemma C.1 uses \tilde w_{t+1} = \tilde w_t - \eta_t \widehat{g}^{SDQ}_t. This equality requires Q^{SDQ}_{L}(\eta_t v) = \eta_t Q^{SDQ}_{L}(v), which is false for a fixed dithered lattice quantizer because the nearest-neighbor mapping is not homogeneous under scaling of the input unless the lattice and dither are correspondingly rescaled. Until \eta_t is absorbed into the lattice and dither scaling (or the theorem is restated for the actually quantized, step-size-scaled update), Theorem III.2 does not establish convergence of Algorithm 1 as written.
  2. [Section III-C and Table II] The convergence analysis relies on AS3, which requires that the quantizer is never overloaded for every user and round. OLALa deliberately violates AS3: the text in Section III-C states that the design allows some level of overloading, Algorithm 2 uses the scaling parameter \zeta_t^u to keep overloading below 1%, and Table II evaluates overloading thresholds of 0%, 0.5%, 1%, 10%, and 50%, with 0% overloading reported as harmful. When overloading occurs, Theorem II.3's white-noise distortion model and the unbiasedness identity (A.5) fail, so the distortion bound (12) and the O(1/T) bound (14) do not apply to the scheme evaluated in Section IV. The manuscript itself notes in Section III-C that the overloaded regime has no tractable characterization, which confirms that the current proof cannot cover the implemented algorithm.
  3. [Theorem III.3 and Appendix D (Eq. (16), Eq. (D.5))] The displayed equality in Theorem III.3 is not a valid mathematical statement: an argmin over generator matrices cannot equal a scalar multiple of an argmin over unit-volume shape matrices. The proof instead yields min_G \sigma^2_{SDQ}(L_\gamma(G)) = \gamma^2 \min_A G(A)/(r_A(R))^2, with the minimizing generator given by (\gamma / r_{A_\star}(R)) A_\star when A_\star minimizes G(A)/(r_A(R))^2. As written, Eq. (16) conflates the minimum value with the minimizing matrix and should be corrected, with the dependence of r_A on A made explicit.
minor comments (5)
  1. [Section IV-B, Table II] The table caption says the '-1' heuristic allows at most 0.05% overloading, while the text states 0.3%; these numbers should be reconciled.
  2. [Section IV-C, Figs. 7-8] The captions of Figs. 7 and 8 state R=3, but the text says CIFAR-10 was evaluated at R=5 bits per weight; the captions and text should be harmonized.
  3. [Theorem III.2 and Appendix C] The parameter \nu is defined as \max\{8\kappa, 1\} in the theorem statement, while the proof chooses \nu+1 = \max\{8L/\mu, 1\}; Eq. (14) uses \nu+T-1 while the proof derives a denominator \nu+t. These definitions should be made consistent.
  4. [Appendix introduction] There is a typo, 'the low of total expectation', which should read 'the law of total expectation'.
  5. [Section III-B, final paragraph] 'systemically showcased' should be 'systematically showcased'.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the empirical contribution is benchmarked against external fixed-lattice baselines, and the convergence proof is an independent adaptation of [43]; the main weaknesses (eta-scaling commutation and AS3 overloading) are proof/correctness gaps, not input-output circularity.

full rationale

The central empirical claim that OLALa outperforms fixed and statically learned lattice quantizers is tested against external baselines (hexagonal, A2, D2 lattices and non-compressed FL), so no fitted parameter is renamed as a prediction. The convergence analysis is self-contained in that Theorem III.2 is proved by the induction of [43, Thm. 1]; [19] is cited only as the fixed-lattice result being generalized and is not load-bearing. Self-citations [1], [32], [34] are background for the lattice-learning machinery and privacy/compression variants, not for the central derivation. The statement that adaptive lattices can tighten the convergence bound (Theorem III.3) is tautological: B_t in (15) contains sigma^2_SDQ of the chosen lattices, so choosing smaller distortion lowers the bound by construction. However, this is a conditional statement about the bound, not a prediction about test accuracy, and the experiments establish the latter independently. Two proof gaps are flagged but are not circularity: (i) Section III-B sets the update as -eta_t times the gradient, while Appendix A.3 defines bgSDQ_t using QSDQ of the un-scaled gradient, and Lemma C.1 writes the one-step recursion as if QSDQ(eta*grad) = eta*QSDQ(grad), which is false for a fixed dithered lattice; (ii) AS3 requires zero overloading, whereas OLALa deliberately permits about 0.5% overloading via the zeta scaling in Section III-C and Table II. These gaps undermine the theorem's applicability to Algorithm 1 as written, but they do not make the derivation equivalent to its own inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central analysis rests on four standard FL assumptions (AS1-AS4), but AS3 is the fragile one: it is explicitly violated by the OLALa algorithm, which permits controlled overloading. This disconnect means the stated convergence theorem does not certify the actual method. No new physical or mathematical entities are introduced.

free parameters (3)
  • Overloading threshold = 0.5% default; heuristic -1 uses 0.3% and 3x standard deviation
    Chosen by hand to balance distortion and clipping; Table II shows performance varies strongly with this value.
  • Scaling parameter zeta_t^u = Not fixed; set per user per round to achieve target overloading within radius gamma=1
    Acts as the effective support radius; it is adapted to the current update statistics, so it is a data-fitted parameter.
  • DNN fixed input s = Arbitrary vector
    Introduced as the fixed input to the deep prior; its values are not learned, but must be chosen.
assumptions (5)
  • domain assumption AS1: local datasets are i.i.d. within each user but heterogeneous across users
    Standard FL assumption; needed for unbiased stochastic gradients and the heterogeneity gap definition.
  • domain assumption AS2: stochastic gradient variance is bounded by sigma_u^2 per user
    Standard for FL convergence proofs; used in Theorem III.1 bound.
  • ad hoc to paper AS3: each lattice quantizer is never overloaded for every user and round
    Required for the SDQ white-noise distortion model (Theorem II.3). OLALa's actual implementation violates this by allowing up to 0.5% overloading, so the theory does not cover the proposed scheme as run.
  • domain assumption AS4: local objectives are L-smooth and mu-strongly convex
    Standard for FedAvg convergence; restricts applicability to convex problems, while experiments use deep networks which are non-convex.
  • ad hoc to paper Metadata (generator matrix and scaling) is transmitted at negligible cost
    The paper assumes 64-bit representation of G_t^u and zeta_t^u adds negligible overhead; reasonable for large models but an assumption about the communication budget.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OLALa: Online Learned Adaptive Lattice Codes for Heterogeneous Federated Learning." pith.science (2026). https://pith.science/paper/PIMARESG

@misc{pith2026250620297,
  author       = {Pith},
  title        = {Pith review of: OLALa: Online Learned Adaptive Lattice Codes for Heterogeneous Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PIMARESG}},
  note         = {Machine review of arXiv:2506.20297}
}
read the original abstract

Federated learning (FL) enables collaborative training across distributed clients without sharing raw data, often at the cost of substantial communication overhead induced by transmitting high-dimensional model updates. This overhead can be alleviated by having the clients quantize their model updates, with dithered lattice quantizers identified as an attractive scheme due to its structural simplicity and convergence-preserving properties. However, existing lattice-based FL schemes typically rely on a fixed quantization rule, which is suboptimal in heterogeneous and dynamic environments where the model updates distribution varies across users and training rounds. In this work, we propose Online Learned Adaptive Lattices (OLALa), a heterogeneous FL framework where each client can adjust its quantizer online using lightweight local computations. We first derive convergence guarantees for FL with non-fixed lattice quantizers and show that proper lattice adaptation can tighten the convergence bound. Then, we design an online learning algorithm that enables clients to tune their quantizers throughout the FL process while exchanging only a compact set of quantization parameters. Numerical experiments demonstrate that OLALa consistently improves learning performance under various quantization rates, outperforming conventional fixed-codebook and non-adaptive schemes.

Figures

Figures reproduced from arXiv: 2506.20297 by the authors.

Figure 1
Figure 1. Example: truncated lattices structures. •’s mark lattice points (codewords); concentric circles represent the radii (i.e., γ values). The left quantizer has the identity as its generator matrix, while the right has the hexagonal generator matrix scaled to have unit determinant. The number of lattice points in L is countable but infinite. Thus, to obtain a finite-bit representation, it is common to restrict L to incl… view at source ↗
Figure 2
Figure 2. Overview of OLALa at user u at time-step t, performing local model (left) and lattice (right) learning using input w˜ t to form the generator matrix Gθ u t (s). D. Discussion The proposed OLALa framework introduces a novel mech￾anism for incorporating adaptive lattice quantization in FL. Our key insight is the identification of the value in tailoring the lattice quantizer to the local update at each user and communi… view at source ↗
Figure 3
Figure 3. Evolution of OLALa’s lattice quantizer over training [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Accuracy vs. training rounds, MNIST, linear model, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Accuracy vs. training rounds, MNIST, MLP, [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 8
Figure 8. Figure 8: Accuracy vs. training rounds, CIFAR-10. CNN, last 30 [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 54 canonical work pages

  1. [1]

    Data-driven lattices for vector quantization,

    N. Lang, I. Assaf, O. Bokobza, and N. Shlezinger, “Data-driven lattices for vector quantization,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2024, pp. 8080–8084

  2. [2]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial Intelligence and Statistics. PMLR, 2017, pp. 1273–1282

  3. [3]

    A survey on federated learning: The journey from centralized to distributed on-site learning and beyond,

    S. AbdulRahman et al. , “A survey on federated learning: The journey from centralized to distributed on-site learning and beyond,” IEEE Internet Things J., vol. 8, no. 7, pp. 5476–5497, 2020

  4. [4]

    A survey on federated learning: challenges and applications,

    J. Wen et al., “A survey on federated learning: challenges and applications,” International Journal of Machine Learning and Cybernetics , vol. 14, no. 2, pp. 513–535, 2023

  5. [5]

    Advances and open problems in federated learning,

    P. Kairouz et al., “Advances and open problems in federated learning,” Foun- dations and trends® in machine learning, vol. 14, no. 1–2, pp. 1–210, 2021. 13

  6. [6]

    Federated learning: A signal processing perspective,

    T. Gafni, N. Shlezinger, K. Cohen, Y . C. Eldar, and H. V . Poor, “Federated learning: A signal processing perspective,” IEEE Signal Process. Mag. , vol. 39, no. 3, pp. 14–41, 2022

  7. [7]

    Communication-efficient federated learning,

    M. Chen, N. Shlezinger, H. V . Poor, Y . C. Eldar, and S. Cui, “Communication-efficient federated learning,” Proceedings of the National Academy of Sciences , vol. 118, no. 17, p. e2024789118, 2021

  8. [8]

    PAUSE: Privacy- aware active user selection for federated learning,

    O. Peleg, N. Lang, S. Rini, N. Shlezinger, and K. Cohen, “PAUSE: Privacy- aware active user selection for federated learning,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025

Show all 56 references
  1. [9]

    Adaptive gradient sparsification for efficient federated learning: An online learning approach,

    P. Han, S. Wang, and K. K. Leung, “Adaptive gradient sparsification for efficient federated learning: An online learning approach,” in IEEE Interna- tional Conference on Distributed Computing Systems , 2020, pp. 300–310

  2. [10]

    Sparse communication for distributed gradient descent,

    A. Aji and K. Heafield, “Sparse communication for distributed gradient descent,” in Conference on Empirical Methods in Natural Language Process- ing. Association for Computational Linguistics (ACL), 2017, pp. 440–445

  3. [11]

    The convergence of sparsified gradient methods,

    D. Alistarh et al. , “The convergence of sparsified gradient methods,” Advances in Neural Information Processing Systems , vol. 31, 2018

  4. [12]

    Machine learning at the wireless edge: Distributed stochastic gradient descent over-the-air,

    M. M. Amiri and D. G ¨und¨uz, “Machine learning at the wireless edge: Distributed stochastic gradient descent over-the-air,” IEEE Trans. Signal Process., vol. 68, pp. 2155–2169, 2020

  5. [13]

    Over-the-air federated learning from heterogeneous data,

    T. Sery, N. Shlezinger, K. Cohen, and Y . C. Eldar, “Over-the-air federated learning from heterogeneous data,” IEEE Trans. Signal Process. , vol. 69, pp. 3796–3811, 2021

  6. [14]

    Federated learning via over-the-air computation,

    K. Yang, T. Jiang, Y . Shi, and Z. Ding, “Federated learning via over-the-air computation,” IEEE Trans. Wireless Commun. , vol. 19, no. 3, pp. 2022–2035, 2020

  7. [15]

    Learned parameter compression for efficient and privacy-preserving federated learning,

    Y . Chen, L. Abrahamyan, H. Sahli, and N. Deligiannis, “Learned parameter compression for efficient and privacy-preserving federated learning,” IEEE Open Journal of the Communications Society , vol. 5, pp. 3506–3516, 2024

  8. [16]

    QSGD: Communication-efficient SGD via gradient quantization and encoding,

    D. Alistarh, D. Grubic, J. Li, R. Tomioka, and M. V ojnovic, “QSGD: Communication-efficient SGD via gradient quantization and encoding,” Advances in Neural Information Processing Systems , vol. 30, pp. 1709–1720, 2017

  9. [17]

    Fedpaq: A communication-efficient federated learning method with periodic averaging and quantization,

    A. Reisizadeh, A. Mokhtari, H. Hassani, A. Jadbabaie, and R. Pedarsani, “Fedpaq: A communication-efficient federated learning method with periodic averaging and quantization,” in International Conference on Artificial Intelligence and Statistics . PMLR, 2020, pp. 2021–2031

  10. [18]

    signSGD: Compressed optimisation for non-convex problems,

    J. Bernstein, Y .-X. Wang, K. Azizzadenesheli, and A. Anandkumar, “signSGD: Compressed optimisation for non-convex problems,” in Inter- national Conference on Machine Learning . PMLR, 2018, pp. 560–569

  11. [19]

    UVeQFed: Universal vector quantization for federated learning,

    N. Shlezinger, M. Chen, Y . C. Eldar, H. V . Poor, and S. Cui, “UVeQFed: Universal vector quantization for federated learning,” IEEE Trans. Signal Process., vol. 69, pp. 500–514, 2020

  12. [20]

    Natural compression for distributed deep learning,

    S. Horv ´oth et al., “Natural compression for distributed deep learning,” in Mathematical and Scientific Machine Learning. PMLR, 2022, pp. 129–141

  13. [21]

    Quantized federated learning under transmission delay and outage constraints,

    Y . Wang, Y . Xu, Q. Shi, and T.-H. Chang, “Quantized federated learning under transmission delay and outage constraints,” IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 323–341, 2021

  14. [22]

    Quantization,

    R. M. Gray and D. L. Neuhoff, “Quantization,” IEEE Trans. Inf. Theory , vol. 44, no. 6, pp. 2325–2383, 1998

  15. [23]

    An algorithm for vector quantizer design,

    Y . Linde, A. Buzo, and R. Gray, “An algorithm for vector quantizer design,” IEEE Trans. Commun. , vol. 28, no. 1, pp. 84–95, 1980

  16. [24]

    Learning vector quantization,

    T. Kohonen, “Learning vector quantization,” Self-organizing maps, pp. 245–261, 2001

  17. [25]

    Neural discrete representation learning,

    A. Van Den Oord and O. Vinyals, “Neural discrete representation learning,” Advances in Neural Information Processing Systems , vol. 30, 2017

  18. [26]

    Remote inference over dynamic links via adaptive rate deep task-oriented vector quantization,

    E. Fishel, M. Malka, S. Ginzach, and N. Shlezinger, “Remote inference over dynamic links via adaptive rate deep task-oriented vector quantization,” arXiv preprint arXiv:2501.02521 , 2025

  19. [27]

    Hardware-limited task-based quantization,

    N. Shlezinger, Y . C. Eldar, and M. R. Rodrigues, “Hardware-limited task-based quantization,” IEEE Trans. Signal Process. , vol. 67, no. 20, pp. 5223–5238, 2019

  20. [28]

    Lecture notes on information theory,

    Y . Polyanskiy and Y . Wu, “Lecture notes on information theory,” Lecture Notes for 6.441 (MIT), ECE563 (University of Illinois Urbana-Champaign), and STAT 664 (Yale), 2012-2017

  21. [29]

    Dithered quantizers,

    R. M. Gray and T. G. Stockham, “Dithered quantizers,” IEEE Trans. Inf. Theory, vol. 39, no. 3, pp. 805–812, 1993

  22. [30]

    Quantization and dither: A theoretical survey,

    S. P. Lipshitz, R. A. Wannamaker, and J. Vanderkooy, “Quantization and dither: A theoretical survey,” Journal of the audio engineering society , vol. 40, no. 5, pp. 355–375, 1992

  23. [31]

    On universal quantization by randomized uniform/lattice quantizers,

    R. Zamir and M. Feder, “On universal quantization by randomized uniform/lattice quantizers,” IEEE Trans. Inf. Theory , vol. 38, no. 2, pp. 428–436, 1992

  24. [32]

    Joint privacy enhancement and quantization in federated learning,

    N. Lang, E. Sofer, T. Shaked, and N. Shlezinger, “Joint privacy enhancement and quantization in federated learning,” IEEE Trans. Signal Process., vol. 71, pp. 295–310, 2023

  25. [33]

    Communication-efficient Laplace mechanism for differential privacy via random quantization,

    A. M. Shahmiri, C. W. Ling, and C. T. Li, “Communication-efficient Laplace mechanism for differential privacy via random quantization,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 4550–4554

  26. [34]

    Compressed private aggregation for scalable and robust federated learning over massive networks,

    N. Lang, N. Shlezinger, R. G. D’Oliveira, and S. E. Rouayheb, “Compressed private aggregation for scalable and robust federated learning over massive networks,” IEEE Trans. Mobile Comput. , 2025, early access

  27. [35]

    On lattice quantization noise,

    R. Zamir and M. Feder, “On lattice quantization noise,” IEEE Trans. Inf. Theory, vol. 42, no. 4, pp. 1152–1159, 1996

  28. [36]

    Asymptotically optimal block quantization,

    A. Gersho, “Asymptotically optimal block quantization,” IEEE Trans. Inf. Theory, vol. 25, no. 4, pp. 373–380, 1979

  29. [37]

    Better lattice quantizers constructed from complex integers,

    S. Lyu, Z. Wang, C. Ling, and H. Chen, “Better lattice quantizers constructed from complex integers,” IEEE Trans. Commun. , vol. 70, no. 12, pp. 7932–7940, 2022

  30. [38]

    On the best lattice quantizers,

    E. Agrell and B. Allen, “On the best lattice quantizers,” IEEE Trans. Inf. Theory, vol. 69, no. 12, pp. 7650–7658, 2023

  31. [39]

    Optimization of lattices for quantization,

    E. Agrell and T. Eriksson, “Optimization of lattices for quantization,” IEEE Trans. Inf. Theory , vol. 44, no. 5, pp. 1814–1828, 1998

  32. [40]

    Optimization and identification of lattice quantizers,

    E. Agrell, D. Pook-Kolb, and B. Allen, “Optimization and identification of lattice quantizers,” IEEE Trans. Inf. Theory , 2025, early access

  33. [41]

    Discriminative and generative learning for linear estimation of random signals [lecture notes],

    N. Shlezinger and T. Routtenberg, “Discriminative and generative learning for linear estimation of random signals [lecture notes],” IEEE Signal Process. Mag., vol. 40, no. 6, pp. 75–82, 2023

  34. [42]

    Heterogeneous federated learning: State-of-the-art and research challenges,

    M. Ye, X. Fang, B. Du, P. C. Yuen, and D. Tao, “Heterogeneous federated learning: State-of-the-art and research challenges,” ACM Computing Surveys, vol. 56, no. 3, pp. 1–44, 2023

  35. [43]

    On the convergence of fedavg on non-iid data,

    X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the convergence of fedavg on non-iid data,” in International conference on learning representations (ICLR), 2020

  36. [44]

    Model-based deep learning: On the intersection of deep learning and optimization,

    N. Shlezinger, Y . C. Eldar, and S. P. Boyd, “Model-based deep learning: On the intersection of deep learning and optimization,” IEEE Access , vol. 10, pp. 115 384–115 398, 2022

  37. [45]

    Deep image prior,

    D. Ulyanov, A. Vedaldi, and V . Lempitsky, “Deep image prior,” in IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 9446–9454

  38. [46]

    Local SGD converges fast and communicates little,

    S. U. Stich, “Local SGD converges fast and communicates little,” in International Conference on Learning Representations , 2019

  39. [47]

    Vector quantization,

    R. Gray, “Vector quantization,” IEEE ASSP Mag. , vol. 1, no. 2, pp. 4–29, 1984

  40. [48]

    Closest point search in lattices,

    E. Agrell, T. Eriksson, A. Vardy, and K. Zeger, “Closest point search in lattices,” IEEE Trans. Inf. Theory , vol. 48, no. 8, pp. 2201–2214, 2002

  41. [49]

    The communication-aware clustered federated learning problem,

    N. Shlezinger, S. Rini, and Y . C. Eldar, “The communication-aware clustered federated learning problem,” in IEEE International Symposium on Information Theory (ISIT) , 2020, pp. 2610–2615

  42. [50]

    Communication-efficient algorithms for statistical optimization,

    Y . Zhang, M. J. Wainwright, and J. C. Duchi, “Communication-efficient algorithms for statistical optimization,” Advances in Neural Information Processing Systems, vol. 25, 2012

  43. [51]

    Stragglers-aware low-latency synchronous federated learning via layer-wise model updates,

    N. Lang, A. Cohen, and N. Shlezinger, “Stragglers-aware low-latency synchronous federated learning via layer-wise model updates,” IEEE Trans. Commun., vol. 73, no. 5, pp. 3333–3346, 2025

  44. [52]

    Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks,

    M. Huh, B. Cheung, P. Agrawal, and P. Isola, “Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks,” in International Conference on Machine Learning . PMLR, 2023, pp. 14 096–14 113

  45. [53]

    Deep task-based analog-to-digital conversion,

    N. Shlezinger, A. Amar, B. Luijten, R. J. Van Sloun, and Y . C. Eldar, “Deep task-based analog-to-digital conversion,” IEEE Trans. Signal Process., vol. 70, pp. 6021–6034, 2022

  46. [54]

    Learning task-based trainable neuromorphic ADCs via power-aware distillation,

    T. V ol, L. Danial, and N. Shlezinger, “Learning task-based trainable neuromorphic ADCs via power-aware distillation,” IEEE Trans. Signal Process., vol. 73, pp. 1246–1261, 2025

  47. [55]

    Entropy-constrained vector quantization,

    P. A. Chou, T. Lookabaugh, and R. M. Gray, “Entropy-constrained vector quantization,” IEEE Trans. Acoust., Speech, Signal Process. , vol. 37, no. 1, pp. 31–42, 1989

  48. [56]

    A catalogue of lattices

    G. Nebe and N. Sloane, “A catalogue of lattices.” [Online]. Available: https://www.math.rwth-aachen.de/homes/Gabriele.Nebe/LATTICES/

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.