Stochastic Gradient Push

Updated 27 April 2026

Stochastic Gradient Push is a decentralized algorithm for optimizing models across directed networks, often used in distributed and privacy-preserving learning.
The method combines local stochastic gradient updates with the Push-Sum protocol, enabling unbiased parameter averaging even in challenging communication conditions.
SGP is robust to network adversities like message loss and asynchronous updates, ideal for applications such as federated optimization and deep learning.

Stochastic Gradient Push (SGP) is a class of decentralized algorithms for distributed stochastic optimization and learning over directed, possibly time-varying, and adverse networks. SGP employs the Push-Sum protocol to achieve unbiased averaging of local model parameters or gradients when standard distributed averaging or consensus methods—relying on symmetric, doubly-stochastic mixing—are inapplicable or inefficient due to asymmetric communication. SGP and its robust and variance-reduced variants serve as foundational protocols for a broad range of applications, including decentralized deep learning, federated optimization, and privacy-preserving inference.

1. Network Model, Problem Formulation, and Assumptions

SGP addresses distributed convex and non-convex optimization problems of the form

$\min_{z\in\mathbb{R}^d} F(z)=\sum_{i=1}^n f_i(z)$

where each node $i$ holds a private function $f_i$ , typically the local loss over a data subset. Communication occurs on a (potentially time-varying) directed graph $\mathcal{G}=(V,E)$ . Original formulations often permit harsh network adversities: asynchronous updates, arbitrary message delays ( $L_{\mathrm{del}}$ ), packet drops ( $L_f$ ), and lack of coordination. Only strong connectivity of $\mathcal{G}$ is typically assumed (Spiridonoff et al., 2018).

For analysis and rate optimality, common objective function assumptions include μ-strong convexity and L-Lipschitz gradients (Spiridonoff et al., 2018, Lin et al., 2023):

Each $f_i$ is μ_i-strongly convex ( $\mu = \sum_i \mu_i > 0$ )
Each $f_i$ is L_i-smooth

Stochasticity is modeled by noisy gradients: $i$ 0 with $i$ 1 zero-mean, possibly bounded. In privacy-preserving extensions, $i$ 2 may aggregate true stochasticity and DP-induced Gaussian noise (Zhu et al., 15 Dec 2025, Zhu et al., 2024).

2. Push-Sum and Stochastic Gradient Push Algorithmic Principle

The essence of SGP combines the Push-Sum gossip mechanism and local stochastic gradient steps to ensure unbiased aggregation over directed graphs.

Each node $i$ 3 maintains:

$i$ 4: biased model (accumulator)
$i$ 5: Push-Sum weight
$i$ 6: de-biased estimate

A general SGP iteration for node $i$ 7 at time $i$ 8 consists of:

Stochastic gradient step:

$i$ 9

where $f_i$ 0 may be a raw or variance-reduced stochastic gradient.

Push-Sum mixing:

$f_i$ 1

$f_i$ 2

where $f_i$ 3 are the (column-stochastic) mixing weights compliant with the communication graph.

De-biasing:

$f_i$ 4

This protocol is robust to lossy, delayed, and asynchronous communications, and only requires local knowledge of out-degree or link structure (Spiridonoff et al., 2018, Nedic et al., 2014).

3. Convergence Theory: Robustness and Optimality

SGP achieves rates competitive with centralized and symmetric decentralized SGD when optimized for stepsize, even in environments with message loss, directed communication, and asynchrony.

Strongly convex case: Robust Asynchronous SGP (RASGP) with stepsize $f_i$ 5 attains

$f_i$ 6

where $f_i$ 7 bounds inter-update intervals (Spiridonoff et al., 2018). All network parameters ( $f_i$ 8, delays, drops) appear exclusively in lower-order terms—establishing network-independent O(1/k) convergence matching centralized SGD (modulo asynchrony-induced slowdown).

General convex/nonconvex case: For smooth but possibly nonconvex $f_i$ 9, SGP with constant stepsize $\mathcal{G}=(V,E)$ 0 satisfies

$\mathcal{G}=(V,E)$ 1

matching the rate and linear speedup of centralized data-parallel SGD (Assran et al., 2018, Zhu et al., 15 Dec 2025).

Variance reduction: Methods such as Push-LSVRG-UP and PrivSGP-VR employ SAGA- or SVRG-type gradient correction tables, yielding (under strong convexity) linear convergence to the exact optimum (Hu et al., 2023) or optimal nonconvex-stationarity bounds under DP constraints (Zhu et al., 2024).
Lyapunov analysis: Explicit construction of quadratic Lyapunov functions based on the absolute probability sequence (APS) for the Push-Sum chain facilitates explicit $\mathcal{G}=(V,E)$ 2 and $\mathcal{G}=(V,E)$ 3 rates for SGP and Subgradient-Push analogues, closing the theoretical gap between centralized and decentralized methods (Lin et al., 2023).

4. Design, Communication, and Scalability

SGP accommodates column-stochastic mixing matrices, enabling sparse and directed graph topologies, which are not accessible to doubly-stochastic (symmetric) decentralized SGD.

Network design: Optimization of the mixing matrix to minimize convergence time (especially over wireless broadcast networks) leads to objective functions depending on graph-theoretic quantities, such as diameter ( $\mathcal{G}=(V,E)$ 4), max degree, and minimal nonzero mixing weight ( $\mathcal{G}=(V,E)$ 5), e.g.,

$\mathcal{G}=(V,E)$ 6

where $\mathcal{G}=(V,E)$ 7 are the out- and in-degrees of the activated graph (Nguyen et al., 16 Apr 2026).

Empirical findings demonstrate that SGP with optimized communication topology can yield up to 45% reduction in wall-clock convergence time compared to traditional D-PSGD and other mixing designs, without compromising model quality (Nguyen et al., 16 Apr 2026).

Communication complexity: Each iteration involves sending $\mathcal{G}=(V,E)$ 8-dimensional vectors and a scalar weight. Compression techniques (e.g., sparse or quantized communication) with error-feedback allow SGP to further reduce per-iteration bandwidth while retaining convergence guarantees, even under differential privacy (Zhu et al., 15 Dec 2025).

5. Differential Privacy and Adaptive Noise

SGP algorithms admit principled integration of differentially private mechanisms via locally-injected Gaussian noise to gradients or updates.

DP guarantees: Both PrivSGP-VR and DP-CSGP establish that, with per-step Gaussian noise variance

$\mathcal{G}=(V,E)$ 9

per node (where $L_{\mathrm{del}}$ 0 is the local sample size, $L_{\mathrm{del}}$ 1 the decision variable dimension), the entire node transcript is $L_{\mathrm{del}}$ 2-DP (Zhu et al., 15 Dec 2025, Zhu et al., 2024). The resultant utility bound for nonconvex objectives is

$L_{\mathrm{del}}$ 3

matching centralized DP-SGD and exceeding prior decentralized approaches by a $L_{\mathrm{del}}$ 4 factor.

Variance reduction and adaptivity: Extensions such as ADP-VRSGP employ stepwise noise decay and matched dynamic step-sizes plus progressive/running-averaged gradient fusion to reduce the negative impact of early-stage heavy noise. The result is accelerated convergence under a predetermined privacy budget and improved test accuracy under DP constraints (Wu et al., 23 Oct 2025).

Plain SGP achieves only sublinear convergence due to stochastic gradient variance. Recent variants integrate variance reduction, gradient tracking, and adaptive consensus mechanisms:

Push-LSVRG-UP: Incorporates loopless SVRG and probabilistic triggering to attain linear convergence for strongly convex finite-sum objectives over unbalanced directed graphs. It maintains low storage overhead (no full history caches as in SAGA) and is effective even with uncoordinated local update schedules (Hu et al., 2023).
Lyapunov methodology: Quadratic Lyapunov tools based on the APS yield unified frameworks for analyzing convergence of standard, variance-reduced, and heterogeneous SGP-like protocols (Lin et al., 2023).
Robust asynchronous schemes: RASGP achieves network-independent $L_{\mathrm{del}}$ 5 convergence under extreme asynchrony, message loss, and delays. The step-size schedule and computation of missing-step compensations are designed so that the resulting iterate sequence closely tracks centralized SGD (Spiridonoff et al., 2018).

7. Empirical Studies and Practical Implications

Extensive empirical studies confirm SGP's practical effectiveness in large-scale decentralized machine learning:

Deep learning workloads: Distributed training of ResNet-50 on ImageNet and Transformer models on WMT’16 En—De using SGP achieves similar or better convergence and final accuracy than AllReduce-SGD or D-PSGD. On commodity interconnects, SGP accelerates time-to-target-accuracy by factors of up to 3×–5× due to reduced communication bottlenecks and robust progress in the presence of stragglers (Assran et al., 2018).
Wireless federation and topology design: SGP under optimized broadcast-aware topology design outperforms state-of-the-art decentralized SGD schemes in wall-clock convergence on challenging graph topologies (e.g., Windmill, Roofnet), with up to 45% fewer communication rounds required (Nguyen et al., 16 Apr 2026).
Differential privacy and compressed communication: DP-CSGP matches the accuracy of exact, uncompressed decentralized DP baselines while achieving up to 4×–10× reduction in actual communication cost (Zhu et al., 15 Dec 2025).

A plausible implication is that SGP's flexibility in mixing design and noise integration, coupled with robust theoretical guarantees, renders it the preferred protocol for decentralized optimization over non-ideal and adversarial networks, particularly when privacy and communication cost are critical.

Key papers referenced:

(Spiridonoff et al., 2018) Robust Asynchronous Stochastic Gradient-Push: Asymptotically Optimal and Network-Independent Performance for Strongly Convex Functions
(Lin et al., 2023) An Analysis Tool for Push-Sum Based Distributed Optimization
(Nguyen et al., 16 Apr 2026) Optimizing Stochastic Gradient Push under Broadcast Communications
(Assran et al., 2018) Stochastic Gradient Push for Distributed Deep Learning
(Wu et al., 23 Oct 2025) ADP-VRSGP: Decentralized Learning with Adaptive Differential Privacy via Variance-Reduced Stochastic Gradient Push
(Hu et al., 2023) Push-LSVRG-UP: Distributed Stochastic Optimization over Unbalanced Directed Networks with Uncoordinated Triggered Probabilities
(Zhu et al., 2024) PrivSGP-VR: Differentially Private Variance-Reduced Stochastic Gradient Push with Tight Utility Bounds
(Nedic et al., 2014) Stochastic Gradient-Push for Strongly Convex Functions on Time-Varying Directed Graphs
(Zhu et al., 15 Dec 2025) DP-CSGP: Differentially Private Stochastic Gradient Push with Compressed Communication