Distributed Nonparametric Estimation: from Sparse to Dense Samples per Terminal

Deheng Yuan; Tao Guo; Zhongyi Huang

REVIEW 3 major objections 4 minor 26 references

Distributed Nonparametric Estimation: from Sparse to Dense Samples per Terminal

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper proves that the minimax L2 error of distributed nonparametric estimation with $m$ terminals, $n$ samples each, and $l$ bits per terminal is $N_{\mathrm{ess}}^{-2r/(2r+1)}$ up to logarithmic factors, for every sparse-to-dense…

desk verdict A substantial all-regime minimax characterization that deserves a serious referee, contingent on the authors' external parametric-protocol lemma. read the letter →

arxiv 2501.07879 v1 pith:UL2TJ7YA submitted 2025-01-14 cs.LG cs.ITmath.ITmath.STstat.TH

Deheng Yuan , Tao Guo , Zhongyi Huang This is my paper · ORCID

classification cs.LGcs.ITmath.ITmath.STstat.TH MSC 62G0562C2094A1760C05

keywords distributedestimationnonparametricminimaxratescommunicationconstraintsstrongdataprocessinginequalitywaveletphasetransitionseffectivesamplesize

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks: if each of $m$ terminals holds $n$ i.i.d. samples from a distribution parameterized by an unknown Sobolev-smooth function, and each terminal may send only $l$ bits, what is the fastest possible $L^2$ error at the decoder? The answer, up to logarithmic factors, is the same as centralized estimation from $N_{\mathrm{ess}}$ samples, where $N_{\mathrm{ess}}$ is an explicit function of $m,n,l,r$ given by equation (2). This collapses the entire sparse-to-dense spectrum into one effective sample size and identifies phase transitions where the role of the communication budget changes from exponential to polynomial. The result matters because it closes the gap between earlier work that handled either dense samples or exactly one sample per terminal, and because it yields minimax rates for density estimation and several regression models as corollaries. The proof uses a layered protocol in which a wavelet outer layer converts nonparametric estimation into parametric distribution estimation, and an information-theoretic lower bound built on a tensorized strong data processing inequality with balls-and-bins bounds.

What carries the argument

The load-bearing object is the effective sample size $N_{\mathrm{ess}}$, whose explicit formula determines the rate in every regime. For the upper bound, the machinery is a two-layer protocol: the outer layer uses a Daubechies wavelet basis with $S$ vanishing moments to approximate the Sobolev function by $K=2^H$ coefficients; each encoder maps each sample to a truncated, quantized wavelet coefficient, producing a parametric alphabet $\mathcal{W}=\{1,\dots,K\}\times\{0,1\}^{2S+2}$; the inner layer then runs a parametric distribution-estimation protocol on the induced distribution $p_{\mathcal{W}}$, and the decoder reconstructs the wavelet coefficients. For the lower bound, the machinery is a finite wavelet sieve $\mathcal{F}(k,C_0,\epsilon)$ over Rademacher perturbations, a terminal-wise likelihood ratio $L_{s,z_k}(x^n)=p^n_{z_k\odot e_s}(x^n)/p^n_{z_k}(x^n)$, and a generalized tensorization of the strong data processing inequality whose strong data processing constant is controlled by the balls-and-bins occupancy $V_s$, the number of samples falling in bin $s$.

What would settle it

For density estimation with $r=1$, take $m=n^4$, $l=1$, and let $n$ grow. The theorem predicts $R(n^4,n,1,1)\asymp (2n^5)^{-1/2}\mathrm{Poly}(\log n)$. Simulate the layered protocol on this instance and measure the average squared $L^2$ error; if the error decays polynomially faster or slower than $n^{-5/2}$ times logarithmic factors, the Case 1 boundary analysis is wrong.

Watch

Extended reading notes

Core claim

The central discovery is a minimax characterization: under Assumptions 1, 2, and 3, the risk $R(m,n,l,r)$ is bounded above and below by $N_{\mathrm{ess}}^{-2r/(2r+1)}$ up to polynomial factors of $\log N$, where $N_{\mathrm{ess}}$ is the effective sample size defined in equation (2). Thus the communication-constrained problem is equivalent, in rate, to a centralized problem with $N_{\mathrm{ess}}$ samples. The result covers all five regimes of $(m,n,l)$ and shows that in the sparsest regime the rate decays exponentially in $l$, while in all other regimes it depends on $l$ only polynomially. It also gives the minimal $l$ needed to match the centralized rate: $n$ for $m>n^{2r}$, and $(mn)^{1/(2r+1)}$ for $m\le n^{2r}$.

Load-bearing premise

The upper-bound proof assumes that the inner parametric distribution-estimation protocol truly attains the three-regime error rates stated in Lemma 2, and only one of those three regimes is proved in this paper; if that oracle rate fails at a boundary, the matched upper bound collapses.

Editorial extensions

If this is right

For all parameter regimes covered by Assumptions 1 through 3, the minimax $L^2$ error is $N_{\mathrm{ess}}^{-2r/(2r+1)}$ up to logarithmic factors, so the communication-constrained problem is equivalent in rate to a centralized problem with $N_{\mathrm{ess}}$ samples.
In the sparsest regime $m\ge n^{2r+1}$ with $1\le l\le \frac{1}{2r+1}\log(m/n^{2r+1})$, the optimal rate decays exponentially in $l$; in all other regimes it depends on $l$ only polynomially.
To match the centralized rate, terminals need about $n$ bits each when $m>n^{2r}$, and about $(mn)^{1/(2r+1)}$ bits each when $m\le n^{2r}$.
The same theorem supplies minimax rates for nonparametric density estimation and for Gaussian, binary, Poisson, and heteroskedastic regression as direct corollaries.
Previously solved special cases, such as dense samples with $m<n^\gamma$ and one-sample-per-terminal density estimation, are recovered as boundary cases of the same effective-sample-size formula.

Reading between the lines

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

An extension the authors leave implicit: the same balls-and-bins tensorization should yield lower bounds for heterogeneous per-terminal sample sizes (different $n_i$ per terminal), with $\sum_i n_i$ replacing $mn$ in the occupancy bounds; the paper does not state this.
Another consequence of the same machinery: the layered protocol can be read as a design rule — pick the bit budget $l$ by inverting the rate formula for a target error — even though the paper presents the protocol only as an achievability proof.
A testable extension: run the protocol with unequal $n_i$ and check whether the rate is governed by the same effective-sample-size formula with $mn$ replaced by $\sum_i n_i$; the paper's fixed-$n$ analysis does not cover this.

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.

Desk Editor's Note

read the letter

The paper delivers what it promises: a unified minimax rate for distributed nonparametric estimation across all sparse-to-dense regimes, summarized by the effective sample size Ness. The formula (2)-(3) is new, and the phase transitions are clearly laid out. The layered protocol that reduces nonparametric estimation to parametric distribution estimation is a neat idea, and the lower bound machinery—strong data processing inequalities plus the balls-and-bins analysis—is genuinely worked out. The assumptions are explicit and verified for density, Gaussian, binary, Poisson, and heteroskedastic regression, which makes the corollaries credible. This is a serious theoretical contribution, not a repackaging of known results.

The soft spots are real but not fatal. The upper bound's inner layer leans on Lemma 2, which is cited to the authors' own prior preprint [8] rather than proved here; only item 1' is derived in Appendix A-C. If any of the rates in items 1-3 fail at the regime boundaries, the matched upper bound in Theorem 4 collapses. That is the load-bearing point a referee must check. I do not see circularity—the parametric problem is separate—but the dependency should be made fully visible, ideally by proving Lemma 2 in this paper or citing a published version.

Second, the plain-text rendering has exponent ambiguities: what should be 2^l appears as 2l, and the threshold (2^l - 1)n appears as (2l - 1)n. This looks like an OCR/formatting artifact rather than a mathematical error, but it genuinely obscures the boundary verifications in Cases 1 and 2. The authors should be asked to clean this up; a referee cannot independently check the case analysis as written.

Minor issues are just that: typos like 'spase' and 'Gaussion' do not affect the math. The case enumeration in the proof of Theorem 1 omits some logarithmic boundary terms, which is standard but worth a footnote.

Overall the central argument holds up if Lemma 2 is correct. The paper is for researchers in communication-constrained estimation and federated learning theory. I would send it to a serious referee rather than desk reject, and I would ask for the Lemma 2 proof plus a version where the exponents are unambiguous.

Referee Report

3 major / 4 minor

Summary. The paper characterizes the minimax L2 rate of distributed nonparametric estimation when each of m terminals holds n i.i.d. samples and sends l bits, under Sobolev regularity of order r. The main result, Theorem 1, states that the rate is (Ness)^{-2r/(2r+1)} up to polylogarithmic factors, where Ness is an effective sample size depending on m, n, l, and r. The upper bound is obtained by a two-layer protocol: a wavelet outer layer converts the problem into parametric distribution estimation, and an inner layer invokes a protocol for that parametric problem. The lower bound uses a wavelet sieve, information-theoretic inequalities based on strong data processing inequalities, and balls-and-bins estimates. The paper also verifies the required assumptions for density estimation and for Gaussian, binary, Poisson, and heteroskedastic regression models, yielding rate results for those special cases.

Significance. If the proof is correct, this resolves a question left open by prior work that treated only dense samples or only n=1: it covers all relative sizes of m and n and identifies phase transitions in the dependence on the communication budget l. The construction is elegant in principle, and the lower-bound machinery, in particular the extension of strong data processing inequalities to irregular events, is potentially reusable. The paper is also commendably explicit about its assumptions and verifies them on several concrete models. However, the central upper bound depends on an unproved external lemma from the authors' own preprint, and there is a suspected gap in the proof of the main lower-bound lemma for the polynomial regimes; both points are load-bearing and need to be addressed before the characterization can be considered established.

major comments (3)

[§III-B, Lemma 2; §IV-B] Lemma 2 is the core of the upper bound: items 1)-3) are used in Cases 1-4 of the error analysis, yet only item 1') is proved in this manuscript; items 1)-3) are cited to the authors' own preprint [8]. The claimed minimax characterization is therefore not self-contained at its most load-bearing point. I ask the authors to include a proof of Lemma 2, or at least to state it with a complete proof in an appendix, and to make explicit which regimes depend on [8]. If the paper is meant to rely on a published version of [8], that should be stated clearly.
[§III-B, Lemma 2; Eq. (2); §IV-B] The notation '2l' in Lemma 2 items 2)-3), in equation (2), and in the error analyses of Section IV-B is ambiguous; it appears to mean 2^l (e.g., the error term k/(2lmn) in item 3 and the Case 1 verification rely on an exponential factor), but the text as written could be read as 2l. The distinction is load-bearing because it determines whether item 2 or item 3 applies at the boundary n<k and whether the phase transition in Case 1 is exponential or polynomial in l. Please replace '2l' by 2^\ell throughout and re-verify the inequalities justifying the use of items 2 and 3 in Cases 1 and 2.
[§V, Lemma 6; Appendix C-D] The proof of Lemma 6, the polynomial strong-data-processing bound used for Cases 2 and 4, contains a gap. After equation (44) the paper states that, under the conditional measure p_{Z^k}(·|E(x^n)=1), the variables L_{s,Z^k}(X^n|E=1) are independent and form a sub-Gaussian vector with parameter 2(δ1+δ2). This is not justified: conditioning on the global event E couples the disjoint bins, and the pointwise bound |L_s-1|≤2(δ1+δ2) does not by itself give the Euclidean-parameter sub-Gaussian property needed for the transportation lemma (Lemma 12). In the preceding Lemma 5 the independence of L_s across s is valid because these quantities depend on disjoint bins under the product measure, but that argument does not survive conditioning. Since Lemma 6 is essential for the lower bounds in Cases 2 and 4, the lower-bound proof is incomplete as written. Please provide a corrected argument that avoids conditioning on a global event or proves the required sub-Gaussian vector bound under the conditional measure.

minor comments (4)

[Appendix C-C] The displayed statement for Case 3 reads 'We show that R(m,n,l,r) ≼ (lm)^{-2r}', but since this is a lower bound it should be ≽.
[Appendix A-D, Lemma 8(2)] The event in Lemma 8(2) is misstated: the probability should be P[∃s: V_s ≥ cn/k], not 'max_{1≤s≤k} V_s ≥ cn/k, ∀s', which is not a well-formed event and does not match the proof by union bound.
[Remark 2] The displayed formula for Ness in Remark 2 has a corrupted exponent: '(2lm)^{2r+1/(2r+1)}' cannot be correct as written. This is likely related to the 2^l notation issue and should be corrected.
[Theorem 3] The word 'Gaussion' in the statement of Theorem 3 should be 'Gaussian'.

Circularity Check

1 steps flagged · score 2.0 of 10

No circular reduction; only a load-bearing self-citation to the authors' own parametric-distribution protocol.

self citation load bearing [Section III-B, Lemma 2; used in Section IV-B to prove Theorem 4]
"The cases in 1-3) are achieved by [8], see Theorem 1 therein for the proof. The proof of 1’) can be found in Appendix A-C."

Theorem 4's upper bound is built as a two-layer protocol: the outer layer produces a parametric distribution-estimation subproblem, and the inner layer invokes the protocol ASR(m,n,l,|W|) from Lemma 2. Lemma 2's items 1-3, which supply the inner-layer rates for all four upper-bound cases, are not proved in this paper but are cited to the authors' own prior preprint [8]. Thus the main upper bound is contingent on a load-bearing self-citation rather than on a self-contained derivation. This is not a definitional or fitted-input circularity: Lemma 2 is a separate parametric estimation statement with its own explicit conditions, and the paper proves the additional item 1' in Appendix A-C.

full rationale

The central claim, R(m,n,l,r) = (Ness)^(-2r/(2r+1)) up to logarithmic factors, is obtained by combining Theorem 4 (upper bound) and Theorem 5 (lower bound). The lower bound is proved in-paper via information-theoretic inequalities, strong data processing constants, and balls-and-bins arguments; it does not depend on the authors' prior work. The upper bound is a genuinely modular construction: Assumption 1 provides unbiased sub-exponential sample-wise estimators of wavelet coefficients, Lemma 1 gives the wavelet approximation error, and Lemma 3 converts the parametric distribution-estimation error into nonparametric error. The only externally imported ingredient is Lemma 2, whose items 1-3 are quoted from the authors' own arXiv preprint [8]. Under the review rule that a cited result is independent support when it is a parameter-free statement with assumptions that do not include the target result, Lemma 2 qualifies: it concerns a distinct parametric density-estimation problem with explicit regimes in k,n,m,l and does not assume the nonparametric rate under study. Therefore the paper does not commit a definitional or statistical circularity, and the self-citation is not a logical loop. The score of 2 reflects the residual concern that the full upper bound, and hence the claimed complete characterization, is not self-contained at its most load-bearing point without the authors' separate protocol.

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

The central claim depends on three statistical assumptions (Assumptions 1-3), a wavelet basis result, and a black-box parametric protocol from prior work. No free parameters are fitted, and no entities are invented.

assumptions (5)

domain assumption Lemma 2: there exists an interactive protocol ASR(m,n,l,k) for parametric distribution estimation with error rates in three regimes (k<=n, n<k<=(2^l-1)n, k>(2^l-1)n).
Used as the inner layer of the upper bound (Section IV-A). The proof of items 1-3 is cited to Theorem 1 of [8]; only item 1' is proved in Appendix A-C.
domain assumption Assumption 1: for every wavelet coefficient fHs there is an unbiased sub-exponential estimator fHatHs(X) with parameters (sqrt(c1 K), c2 sqrt(K)).
Required for the outer-layer quantization and error analysis in Section IV; verified for the five examples in Appendix D.
domain assumption Assumptions 2 and 3: on the lower-bound sieve, the distribution factorizes across bins with sub-exponential log-likelihood ratios with scale k^{-2r} and k^{-r}.
Required for the lower bound in Section V; verified for the examples in Appendix D.
standard math Lemma 1 (wavelet basis properties): orthonormal wavelet basis with bounded sup-norm, local support, and L2 approximation rate 2^{-2Hr}.
Used throughout the upper and lower bounds; proof cited to [21], [24], [25].
standard math Sobolev embedding Hr([0,1],L) is contained in L-infinity for r>1/2.
Used in Lemma 1 and in verifying the examples.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distributed Nonparametric Estimation: from Sparse to Dense Samples per Terminal." pith.science (2026). https://pith.science/paper/UL2TJ7YA

@misc{pith2026250107879,
  author       = {Pith},
  title        = {Pith review of: Distributed Nonparametric Estimation: from Sparse to Dense Samples per Terminal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UL2TJ7YA}},
  note         = {Machine review of arXiv:2501.07879}
}

read the original abstract

Consider the communication-constrained problem of nonparametric function estimation, in which each distributed terminal holds multiple i.i.d. samples. Under certain regularity assumptions, we characterize the minimax optimal rates for all regimes, and identify phase transitions of the optimal rates as the samples per terminal vary from sparse to dense. This fully solves the problem left open by previous works, whose scopes are limited to regimes with either dense samples or a single sample per terminal. To achieve the optimal rates, we design a layered estimation protocol by exploiting protocols for the parametric density estimation problem. We show the optimality of the protocol using information-theoretic methods and strong data processing inequalities, and incorporating the classic balls and bins model. The optimal rates are immediate for various special cases such as density estimation, Gaussian, binary, Poisson and heteroskedastic regression models.

Figures

Figures reproduced from arXiv: 2501.07879 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

[8]

Adaptive Refinement Protocols for Distributed Distribution Estimation under $\ell^p$-Losses

D. Y uan, T. Guo, and Z. Huang, “Adaptive reﬁnement protoc ols for distributed distribution estimation under ℓp-losses,” 2024. [Online]. Available: https://arxiv.org/abs/2410.06884 2, 6, 8, 9, 1 0

work page Pith review arXiv 2024
[1]

Communication-efﬁcient learning of deep networks fr om decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Ar cas, “Communication-efﬁcient learning of deep networks fr om decentralized data,” in International Conference on Artiﬁcial Intelligence and St atistics, vol. 54, Fort Lauderdale, FL, USA, Apr. 2017, pp. 1273–1282. 1

work page 2017
[2]

Federated l earning: Challenges, methods, and future directions,

T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated l earning: Challenges, methods, and future directions,” IEEE Signal Processing Magazine, vol. 37, no. 3, pp. 50–60, May 2020. 1

work page 2020
[3]

Advances and open problems in federa ted learning,

P . Kairouz, et al., “Advances and open problems in federa ted learning,” F oundations and Trends in Machine Learning , vol. 14, no. 1–2, pp. 1–210, Jun. 2021. 1

work page 2021
[4]

Distributed nonparametric estim ation under communication constraints,

A. Zaman and B. Szabó, “Distributed nonparametric estim ation under communication constraints,” 2022. [Online]. A vailable: https://arxiv.org/abs/2204.10373 1, 2, 5

arXiv 2022
[5]

Adaptive distributed method s under communication constraints,

B. Szabó and H. van Zanten, “Adaptive distributed method s under communication constraints,” The Annals of Statistics , vol. 48, no. 4, pp. 2347–2380, Aug. 2020. 2, 5

work page 2020
[6]

Lower bounds for learn ing distributions under communication constraints via ﬁsh er information,

L. P . Barnes, Y . Han, and A. Ozgur, “Lower bounds for learn ing distributions under communication constraints via ﬁsh er information,” Journal of Machine Learning Research , vol. 21, no. 236, pp. 1–30, Feb. 2020. 2

work page 2020
[7]

Opti mal rates for nonparametric density estimation under commu nication constraints,

J. Acharya, C. L. Canonne, A. V . Singh, and H. Tyagi, “Opti mal rates for nonparametric density estimation under commu nication constraints,” IEEE Transactions on Information Theory , vol. 70, no. 3, pp. 1939–1961, Mar. 2024. 2, 5, 10

work page 1939

Show all 26 references

[9]

In teractive inference under information constraints,

J. Acharya, C. L. Canonne, Y . Liu, Z. Sun, and H. Tyagi, “In teractive inference under information constraints,” IEEE Transactions on Information Theory , vol. 68, no. 1, pp. 502–516, Jan. 2022. 2, 20

2022
[10]

Uniﬁed l ower bounds for interactive high-dimensional estimation u nder information constraints,

J. Acharya, C. L. Canonne, Z. Sun, and H. Tyagi, “Uniﬁed l ower bounds for interactive high-dimensional estimation u nder information constraints,” in International Conference on Neural Information Processin g Systems , vol. 36, New Orleans, LA, US, Dec. 2023, pp. 51 133–51 16...

2023
[11]

Distributed nonparametric reg ression under communication constraints,

Y . Zhu and J. Lafferty, “Distributed nonparametric reg ression under communication constraints,” in International Conference on Machine Learning, vol. 80, Stockholm, Sweden, Jul. 2018, pp. 6009–6017. 2, 3, 5

2018
[12]

Distributed function estim ation: Adaptation using minimal communication,

B. Szabó and H. van Zanten, “Distributed function estim ation: Adaptation using minimal communication,” Mathematical Statistics and Learning, vol. 5, no. 3, pp. 159–199, Dec. 2022. 2, 3, 5

2022
[13]

Distributed nonparametric functi on estimation: Optimal rate of convergence and cost of adapt ation,

T. T. Cai and H. Wei, “Distributed nonparametric functi on estimation: Optimal rate of convergence and cost of adapt ation,” The Annals of Statistics, vol. 50, no. 2, pp. 698–725, Apr. 2022. 2, 3, 5, 23

2022
[14]

Local d ifferential privacy: Elbow effect in optimal density estim ation and adaptation over Besov ellipsoids,

C. Butucea, A. Dubois, M. Kroll, and A. Saumard, “Local d ifferential privacy: Elbow effect in optimal density estim ation and adaptation over Besov ellipsoids,” Bernoulli, vol. 26, no. 3, pp. 1727–1764, Aug. 2020. 2

2020
[15]

On density estimation at a ﬁxed point under lo cal differential privacy,

M. Kroll, “On density estimation at a ﬁxed point under lo cal differential privacy,” Electronic Journal of Statistics , vol. 15, no. 1, pp. 1783–1813, Jan. 2021. 2

2021
[16]

Density estimation under local differential privacy and Hellinger loss,

M. Sart, “Density estimation under local differential privacy and Hellinger loss,” Bernoulli, vol. 29, no. 3, pp. 2318–2341, Aug. 2023. 2

2023
[17]

About the co st of central privacy in density estimation,

C. Lalanne, A. Garivier, and R. Gribonval, “About the co st of central privacy in density estimation,” Transactions on Machine Learning Research, Aug. 2023. 2

2023
[18]

Optimal fe derated learning for nonparametric regression with hetero geneous distributed differential privacy constraints,

T. T. Cai, A. Chakraborty, and L. Vuursteen, “Optimal fe derated learning for nonparametric regression with hetero geneous distributed differential privacy constraints,” 2024. [Online]. Avail able: https://arxiv.org/abs/2406.06755 2

2024 arXiv
[19]

Communication complexity of two-party nonpar ametric global density estimation,

J. Liu, “Communication complexity of two-party nonpar ametric global density estimation,” in Annual Conference on Information Sciences and Systems , Princeton, NJ, USA, Mar. 2022, pp. 292–297. 2

2022
[20]

A few interactions improve distributed nonparame tric estimation, optimally,

——, “A few interactions improve distributed nonparame tric estimation, optimally,” IEEE Transactions on Information Theory , vol. 69, no. 12, pp. 7867–7886, Dec. 2023. 2

2023
[21]

Giné and R

E. Giné and R. Nickl, Mathematical F oundations of Inﬁnite-Dimensional Statist ical Models . New Y ork: Cambridge University Press,
[22]

M. J. Wainwright, High-Dimensional Statistics: A Non-Asymptotic Viewpoint . Cambridge University Press, 2019. 4, 14

2019
[23]

Dis tributed estimation with multiple samples per user: Sharp r ates and phase transition,

J. Acharya, C. Canonne, Y . Liu, Z. Sun, and H. Tyagi, “Dis tributed estimation with multiple samples per user: Sharp r ates and phase transition,” in International Conference on Neural Information Processin g Systems , vol. 34, Dec. 2021, pp. 18 920–18 931. 6

2021
[24]

Daubechies, Ten Lectures on W avelets

I. Daubechies, Ten Lectures on W avelets. Philadelphia: Society for Industrial and Applied Mathema tics, 1992. 7 37

1992
[25]

Sobolev, Besov and Nikolskii fractional spa ces: Imbeddings and comparisons for vector valued spaces on an interval,

J. Simon, “Sobolev, Besov and Nikolskii fractional spa ces: Imbeddings and comparisons for vector valued spaces on an interval,” Annali di Matematica Pura ed Applicata , vol. 157, pp. 117–148, Dec. 1990. 7

1990
[26]

Inference unde r information constraints II: Communication constraints a nd shared randomness,

J. Acharya, C. L. Canonne, and H. Tyagi, “Inference unde r information constraints II: Communication constraints a nd shared randomness,” IEEE Transactions on Information Theory , vol. 66, no. 12, pp. 7856–7877, Dec. 2020. 10

2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [8]

Adaptive Refinement Protocols for Distributed Distribution Estimation under $\ell^p$-Losses

D. Y uan, T. Guo, and Z. Huang, “Adaptive reﬁnement protoc ols for distributed distribution estimation under ℓp-losses,” 2024. [Online]. Available: https://arxiv.org/abs/2410.06884 2, 6, 8, 9, 1 0

work page Pith review arXiv 2024

[2] [1]

Communication-efﬁcient learning of deep networks fr om decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Ar cas, “Communication-efﬁcient learning of deep networks fr om decentralized data,” in International Conference on Artiﬁcial Intelligence and St atistics, vol. 54, Fort Lauderdale, FL, USA, Apr. 2017, pp. 1273–1282. 1

work page 2017

[3] [2]

Federated l earning: Challenges, methods, and future directions,

T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated l earning: Challenges, methods, and future directions,” IEEE Signal Processing Magazine, vol. 37, no. 3, pp. 50–60, May 2020. 1

work page 2020

[4] [3]

Advances and open problems in federa ted learning,

P . Kairouz, et al., “Advances and open problems in federa ted learning,” F oundations and Trends in Machine Learning , vol. 14, no. 1–2, pp. 1–210, Jun. 2021. 1

work page 2021

[5] [4]

Distributed nonparametric estim ation under communication constraints,

A. Zaman and B. Szabó, “Distributed nonparametric estim ation under communication constraints,” 2022. [Online]. A vailable: https://arxiv.org/abs/2204.10373 1, 2, 5

arXiv 2022

[6] [5]

Adaptive distributed method s under communication constraints,

B. Szabó and H. van Zanten, “Adaptive distributed method s under communication constraints,” The Annals of Statistics , vol. 48, no. 4, pp. 2347–2380, Aug. 2020. 2, 5

work page 2020

[7] [6]

Lower bounds for learn ing distributions under communication constraints via ﬁsh er information,

L. P . Barnes, Y . Han, and A. Ozgur, “Lower bounds for learn ing distributions under communication constraints via ﬁsh er information,” Journal of Machine Learning Research , vol. 21, no. 236, pp. 1–30, Feb. 2020. 2

work page 2020

[8] [7]

Opti mal rates for nonparametric density estimation under commu nication constraints,

J. Acharya, C. L. Canonne, A. V . Singh, and H. Tyagi, “Opti mal rates for nonparametric density estimation under commu nication constraints,” IEEE Transactions on Information Theory , vol. 70, no. 3, pp. 1939–1961, Mar. 2024. 2, 5, 10

work page 1939