Source-linked AI summary

Distributed Nonconvex Constrained Optimization over Time-Varying Digraphs

Gesualdo Scutari, Ying Sun

arXiv:1809.01106v1math.OCcs.DCcs.MA

TL;DR

The paper targets distributed constrained optimization for smooth, possibly nonconvex aggregate costs combined with DC regularization over time-varying directed graphs. It proposes a successive-convex-approximation framework with perturbed push-sum gradient tracking, proving sublinear rates for constant step-sizes and asymptotic stationarity for diminishing step-sizes. Numerical results compare favorably with current schemes.

  • Problem

    The gap is the lack of distributed schemes for the composite nonconvex constrained problem over general time-varying directed graphs, including nonsmooth regularization and potentially unbounded gradients.

  • Method

    The framework combines successive convex approximation with perturbed push-sum consensus to track the aggregate smooth gradient over directed or undirected time-varying graphs.

  • Results

    The method converges to stationary solutions; constant step-sizes provide a sublinear convergence rate, and numerical results compare favorably with state-of-the-art algorithms.

  • Takeaways & Limitations

    The framework enlarges the class of constrained and composite convex or nonconvex problems to which distributed algorithms have convergence guarantees.

  • Takeaways & Limitations

    The analysis assumes strongly convex surrogate functions, although compact feasible sets allow this to be weakened to convexity; inexact subproblems require additional assumptions on G+.

Abstract

from arXiv · show

This paper considers nonconvex distributed constrained optimization over networks, modeled as directed (possibly time-varying) graphs. We introduce the first algorithmic framework for the minimization of the sum of a smooth nonconvex (nonseparable) function--the agent's sum-utility--plus a Difference-of-Convex (DC) function (with nonsmooth convex part). This general formulation arises in many applications, from statistical machine learning to engineering. The proposed distributed method combines successive convex approximation techniques with a judiciously designed perturbed push-sum consensus mechanism that aims to track locally the gradient of the (smooth part of the) sum-utility. Sublinear convergence rate is proved when a fixed step-size (possibly different among the agents) is employed whereas asymptotic convergence to stationary solutions is proved using a diminishing step-size. Numerical results show that our algorithms compare favorably with current schemes on both convex and nonconvex problems.

1 Introduction

The paper addresses distributed optimization of a possibly nonconvex, nonseparable composite objective over constrained, time-varying directed networks. It introduces a provably convergent framework combining successive convex approximation with perturbed push-sum gradient tracking.

  • Problem setting: The problem combines smooth possibly nonconvex agent costs, a DC regularizer with nonsmooth convex part, and convex constraints.Each agent knows only its own cost function, while agents cooperate through local communications.
  • Motivation: The setting covers applications including network information processing, telecommunications, multi-agent control, machine learning, and data-intensive analytics.Examples include nonlinear least squares, dictionary learning, component analysis, low-rank approximation, and matrix completion.
  • Motivation: Distributed computation is motivated by scattered data whose volume and spatial or temporal disparity can make centralized processing infeasible or inefficient.Time-varying directed links naturally model mobile wireless networks and asymmetric communication conditions.
  • Contributions: Constant step-sizes yield sublinear convergence, while diminishing step-sizes yield convergence to stationary solutions.The framework also permits different constant step-sizes among agents.
  • Contributions: The analysis addresses nonconvexity, nonsmoothness, constraints, general directed time-varying graphs, and unbounded gradients, extending the scope of distributed methods.The paper reports preliminary numerical results comparing favorably with state-of-the-art algorithms.
  • Contributions: The proposed framework is designed for arbitrary time-varying directed or undirected graphs, preserves feasibility, and requires only column-stochastic weights on directed graphs.It uses successive convex approximation and perturbed push-sum consensus to track the gradient of the smooth aggregate objective.

2 Problem Setup and Motivating Examples

The paper formulates constrained distributed optimization with smooth, possibly nonconvex agent costs and a DC regularizer over time-varying directed networks. It motivates the formulation through sparse statistical-learning problems, including nonconvex sparse regression and sparse PCA.

  • Problem formulation: The problem uses a closed convex feasible set, smooth possibly nonconvex agent costs, and a DC regularizer with a convex possibly nonsmooth part.
  • Problem formulation: The regularizer G can promote extra solution structure, including sparsity through convex or nonconvex surrogates of the ℓ0 function.
  • Distributed information: Each agent knows only its own cost function and the shared regularizer, rather than the other agents’ cost functions.
  • Network model: Agents communicate over directed, possibly time-varying graphs, with information exchanged through time-dependent in- and out-neighborhoods.
  • Motivating examples: The motivating applications include nonconvex sparse linear regression and sparse PCA, both modeled as instances of the proposed problem class.

3 Preliminaries: The perturbed condensed push-sum algorithm

This section develops perturbed condensed push-sum as a constrained consensus mechanism for directed, time-varying graphs. Its auxiliary variables yield row-stochastic equivalent weights, supporting feasibility preservation, average tracking, and convergence analysis.

  • Algorithmic role: The framework combines local successive convex approximation with constrained consensus and gradient-average tracking over directed graphs.
  • Algorithmic role: Perturbed condensed push-sum is introduced as a reformulation better suited to integrating push-sum consensus with constrained optimization.
  • Protocol structure: The auxiliary variables φ_i dynamically rebuild row stochasticity in the equivalent weight matrix, a condition used to lock consensus.
  • Convergence analysis: Under column-stochastic weight matrices and B-strong connectivity, Proposition 1 establishes convergence properties for arbitrary perturbation sequences satisfying the stated conditions.
  • Special cases: With zero perturbations, the protocol reduces to condensed push-sum and yields geometric convergence to weighted average consensus.
  • Special cases: Because the equivalent weights are row stochastic, initializing all agent values in K preserves feasibility of every iterate.
  • Special cases: The same protocol can track averages of time-varying signals when agents generate signal samples and use the prescribed perturbations.
  • Convergence analysis: The convergence proof reduces consensus analysis to bounding products of row-stochastic equivalent matrices relative to a weight-averaging matrix.

4 Algorithmic Design

SONATA combines local successive convex approximation with perturbed push-sum consensus and gradient tracking for constrained composite optimization over time-varying directed graphs. Under stated step-size conditions, it achieves consensus at stationary solutions and provides asymptotic convergence and O(1/ϵ) complexity guarantees.

  • Gradient tracking: Each agent maintains a local copy of the global variable and an auxiliary variable for tracking the average gradient information unavailable locally.The auxiliary variable is updated using only local information through perturbed condensed push-sum.
  • Local successive convex approximation: SONATA has each agent solve a local convexification of the composite problem using a strongly convex approximation of its smooth cost and a linearized concave DC component.The local update uses a step-size α_n ∈ (0, 1).
  • Surrogate assumptions: The surrogate preserves first-order properties, is uniformly strongly convex in its optimization argument, and has a uniformly Lipschitz partial gradient.These conditions support simple convex local approximations of the agents’ smooth costs.
  • Information mixing: Perturbed push-sum updates mix optimization variables and auxiliary tracking variables using communication with in-neighbors and out-neighbors on directed, possibly time-varying graphs.The mechanism is designed to make local variables asymptotically consensual while tracking gradient averages.
  • Distributed implementation: The algorithm is distributed: local optimization uses locally available variables, while consensus and tracking exchange information only among graph neighbors.SONATA combines the local optimization step with two information-mixing steps.
  • Convergence and complexity: Under the convergence assumptions, local iterates approach a stationary solution, and with a constant step-size the first iteration attaining M(x_n) ≤ ϵ satisfies T_ϵ = O(1/ϵ).The asymptotic result gives lim_{n→∞} M(x_n) = 0; the results also extend with minor modifications to different constant agent step-sizes.

5 SONATA and special cases

SONATA provides a unifying framework that subsumes several distributed optimization algorithms as special cases under restricted problem and network settings.

  • SONATA-L: The linearized SONATA-L variant results from first-order surrogates with proximal terms and can use ATC or CAA mixing protocols.
  • NEXT-L: With double-stochastic mixing, the ATC-based SONATA-L updates coincide with the earlier NEXT algorithm.
  • Algorithms in [30,33,49,51] are all special cases of SONATA and NEXT.
  • Related algorithms: Aug-DGM and the algorithm in [33] coincide with ATC-NEXT-L under the stated equal-step-size or networking conditions.
  • Related algorithms: DIGing coincides with CAA-NEXT-L, while push-DIGing coincides with ATC-SONATA-L in the cited settings.
  • Related algorithms: ADD-OPT coincides with CAA-SONATA-L after the stated variable transformation.

6 Convergence Proof of SONATA

The convergence proof analyzes SONATA through weighted averages, consensus disagreement, tracking errors, and a Lyapunov-like function under the paper’s standing assumptions.

  • Proof strategy: The proof studies a Lyapunov function along the weighted average, consensus disagreement, and tracking-error dynamics.
  • Solution map: The error-free local solution map is Lipschitz continuous, and its fixed points coincide with d-stationary solutions of Problem (P).
  • Error dynamics: The analysis establishes that consensus and tracking disagreements asymptotically vanish under the relevant conditions.
  • Error dynamics: A new analysis couples consensus and tracking errors because constant step-sizes or unbounded subgradients prevent directly inferring summability from consensus-error dynamics alone.
  • Lyapunov analysis: The proof constructs a Lyapunov-like function whose descent over blocks of iterations yields summability of errors and convergence of all error sequences.
  • Step-sizes: The convergence argument treats constant and diminishing step-sizes separately, with constant step-sizes requiring sufficiently small values.

7 Numerical results

Numerical experiments evaluate SONATA on sparse regression and distributed PCA over time-varying digraphs, with results favoring SONATA against the tested baselines.

  • Sparse regression: Sparse regression experiments use 30 agents, time-varying digraphs, 100 Monte Carlo trials, and a nonconvex logarithmic sparsity regularizer.
  • Sparse regression: Both SONATA variants are much faster than distributed gradient, and SONATA-PL converges faster than SONATA-L under the same tuning.
  • Distributed PCA: The distributed PCA experiments compare SONATA with gradient projection on synthetic and gene-expression data using stationarity, consensus, and NMSE measures.
  • Synthetic PCA: On synthetic PCA data, SONATA efficiently finds a stationary point and the leading eigenvector, whereas the gradient algorithm progresses slowly and fails to achieve a small NMSE.
  • Gene-expression PCA: On the gene-expression data set, SONATA shows behavior similar to the synthetic experiment and converges quite fast despite the massive variable dimension.

A Proof of Lemma 3

The appendix proof develops bounds for products of time-varying mixing matrices and their disagreement from weighted consensus behavior.

  • Mixing bounds: Lemma 15 provides stochastic-vector and matrix-product properties used in proving the mixing result.
  • Relation to prior work: The appendix states that Lemma 15 follows arguments related to prior results while requiring only the paper’s weaker graph condition.
  • Proof construction: The proof analyzes each coordinate of an arbitrary stacked vector and bounds the deviation of matrix products from the weighted consensus operator.

B Proof of Lemma 11

The proof derives the stated results by recognizing both SONATA updates as instances of the perturbed condensed push-sum algorithm and applying Proposition 1.

  • B Proof of Lemma 11: The SONATA x-update is identified as a special case of the perturbed condensed push-sum algorithm, with a specified perturbation.This correspondence allows Proposition 1 to be applied directly to obtain (69).
  • B Proof of Lemma 11: The y-update is likewise recognized as a special case of the same algorithm, enabling the analogous argument for (70).The passage indicates that the perturbation is defined through the inverse diagonal scaling and gradient-related term shown there.
  • B Proof of Lemma 11: The argument concludes after these two update correspondences and their associated applications of Proposition 1.The proof text explicitly marks completion at the end.
Loading 1809.01106v1…