A Unifying Approximate Method of Multipliers
for Distributed Composite Optimization

We first propose a family of optimization algorithms that effectively approximate the Method of Multipliers [35] and serve as a cornerstone of developing distributed algorithms for solving problem (1).

Let each node $i\in\mathcal{V}$ keep a copy $x_{i}\in\mathbb{R}^{d}$ of the global decision variable $x\in\mathbb{R}^{d}$ in problem (1), and define

where $\mathbf{x}=(x_{1}^{T},\ldots,x_{N}^{T})^{T}\in\mathbb{R}^{Nd}$. Then, problem (1) can be equivalently transformed into

Next, let $\tilde{H}\in\mathbb{R}^{Nd\times Nd}$ be such that $\tilde{H}=\tilde{H}^{T}$, $\tilde{H}\succeq\mathbf{O}$, and $\operatorname{Null}(\tilde{H})=S$. It has been shown in [18] that the consensus constraint $\mathbf{x}\in S$ in problem (2) is equivalent to $\tilde{H}^{\frac{1}{2}}\mathbf{x}=\mathbf{0}$. Therefore, (2) is identical to

Problems (1), (2), and (3) have the same optimal value. Also, $x^{\star}\in\mathbb{R}^{d}$ is an optimum of (1) if and only if $\mathbf{x}^{\star}=((x^{\star})^{T},\ldots,(x^{\star})^{T})^{T}\in\mathbb{R}^{Nd}$ is an optimum of (2) and (3).

Consider applying the Method of Multipliers [35] to solve problem (3), which gives

Here, $\mathbf{x}^{k}\in\mathbb{R}^{Nd}$ is the primal variable at iteration $k\geq 0$, which is updated by minimizing an augmented Lagrangian function with the penalty $\frac{\rho}{2}\|\mathbf{x}\|_{\tilde{H}}^{2}$, $\rho>0$ on the consensus constraint $\tilde{H}^{\frac{1}{2}}\mathbf{x}=\mathbf{0}$. In addition, $\mathbf{v}^{k}\in\mathbb{R}^{Nd}$ is the dual variable, whose initial value $\mathbf{v}^{0}$ can be arbitrarily set.

Although the Method of Multipliers may be applied to solve (3) with properly selected parameters, it is not implementable in a distributed fashion over $\mathcal{G}$, even when the problem reduces to linear programming. To address this issue, our strategy is to first derive a paradigm of approximating the Method of Multipliers and then design its distributed realizations.

Our approximation approach is as follows: Starting from any $\mathbf{v}^{0}\in\mathbb{R}^{Nd}$, let

Compared to the Method of Multipliers, we adopt the same dual update (5) but construct a different primal update (4). In (4), we use a possibly time-varying surrogate function $u^{k}:\mathbb{R}^{Nd}\rightarrow\mathbb{R}$ to replace $f(\mathbf{x})$ in the primal update of the Method of Multipliers, whose conditions are imposed in Assumption 2 below. Additionally, to introduce more flexibility, we use a different weight matrix $H\in\mathbb{R}^{Nd\times Nd}$ to define the penalty term $\frac{\rho}{2}\|\mathbf{x}\|_{H}^{2}$, $\rho>0$. We suppose $H$ has the same properties as $\tilde{H}$, i.e., $H=H^{T}\succeq\mathbf{O}$ and $\operatorname{Null}(H)=S$.

The strong convexity condition in Assumption 2(b) guarantees that $\mathbf{x}^{k+1}$ in (4) is well-defined and uniquely exists. Assumption 2(d) is the key to making (4)–(5) solve problem (3). To explain this, note that (4) is equivalent to finding the unique $\mathbf{x}^{k+1}$ satisfying

Let $(\mathbf{x}^{\star},\mathbf{v}^{\star})$ be a primal-dual optimal solution pair of problem (3), which satisfies

If $(\mathbf{x}^{k},\mathbf{v}^{k})=(\mathbf{x}^{\star},\mathbf{v}^{\star})$, then $\mathbf{x}^{k+1}$ has to be $\mathbf{x}^{\star}$ because of Assumption 2(d), $H\mathbf{x}^{\star}=\mathbf{0}$, (7), and the uniqueness of $\mathbf{x}^{k+1}$ in (6). It follows from (5) and $\tilde{H}^{\frac{1}{2}}\mathbf{x}^{\star}=\mathbf{0}$ that $\mathbf{v}^{k+1}=\mathbf{v}^{\star}$. Therefore, $(\mathbf{x}^{\star},\mathbf{v}^{\star})$ is a fixed point of (4)–(5). The remaining conditions in Assumption 2 will be used for convergence analysis later in Section IV.

The paradigm described by (4)–(5) and obeying Assumption 2 is called Approximate Method of Multipliers (AMM). As there are numerous options of the surrogate function $u^{k}$, AMM unifies a wealth of optimization algorithms, including a variety of existing methods (cf. Section III) and many brand new algorithms. Moreover, since Assumption 2 allows $u^{k}$ to have a more favorable structure than $f$, AMM with appropriate $u^{k}$’s may induce a prominent reduction in computational cost compared to the Method of Multipliers. In the sequel, we will provide various options of $u^{k}$, which give rise to a series of distributed versions of AMM.

A number of existing works [36, 37, 38, 39, 40, 31, 32, 33, 34] also introduce surrogate functions to optimization methods, among which [31, 32, 33, 34] develop distributed algorithms and can address partially nonconvex problems on time-varying networks. The advantages of AMM as well as its differences from the algorithms in [31, 32, 33, 34] are highlighted as follows:

i) The algorithms carrying surrogate functions in [31, 32, 33, 34] are (primal) gradient-tracking-based methods. In contrast, AMM incorporates a surrogate function into a primal-dual framework, so that AMM is inherently different from the algorithms in [31, 32, 33, 34] and requires new analysis tools.

ii) The surrogate function conditions in Assumption 2 intersect with but still differ from those in [31, 32, 33, 34]. For instance, when $f$ is twice continuously differentiable, $u^{k}(\mathbf{x})=\langle\nabla f(\mathbf{x}^{k}),\mathbf{x}\rangle+\frac{1}{2}(\mathbf{x}-\mathbf{x}^{k})^{T}(\nabla^{2}f(\mathbf{x}^{k})+\epsilon I)(\mathbf{x}-\mathbf{x}^{k})$ with $\epsilon>0$ meets Assumption 2 but cannot be included by [31, 32, 33, 34].

iii) To enable distributed implementation, the surrogate functions in [31, 32, 33, 34] need to be fully separable in the sense that they can be written as the sum of $N$ functions with independent variables. In contrast, AMM allows $u^{k}$ to be densely coupled under proper design yet can still be executed in a distributed fashion (cf. Sections II-C and III-C4). This may lead to more diverse algorithm design.

iv) In [31, 32, 33, 34], it is required that the global nonsmooth objective function be accessible to every node, while problem (1) only assigns a local nonsmooth component $h_{i}$ to each node $i$, which is more applicable to scenarios concerning privacy and makes the analysis more challenging.

This subsection lays out the parameter designs of AMM for distributed implementations.

We first apply the following change of variable to AMM:

Then, AMM (4)–(5) can be rewritten as

Moreover, note that

where $S^{\bot}$ is the orthogonal complement of $S$ in (2). Hence, (8) requires $\mathbf{q}^{k}\in S^{\bot}$ $\forall k\geq 0$, which, due to (10), can be ensured simply by the following initialization:

Therefore, (9)–(11) is an equivalent form of AMM.

Next, partition the primal variable $\mathbf{x}^{k}$ and the dual variable $\mathbf{q}^{k}$ in (9)–(11) as $\mathbf{x}^{k}=((x_{1}^{k})^{T},\ldots,(x_{N}^{k})^{T})^{T}$ and $\mathbf{q}^{k}=((q_{1}^{k})^{T},\ldots,(q_{N}^{k})^{T})^{T}$. Suppose each node $i\in\mathcal{V}$ maintains $x_{i}^{k}\in\mathbb{R}^{d}$ and $q_{i}^{k}\in\mathbb{R}^{d}$. Clearly, the nodes manage to collectively meet (11) by setting, for instance, $q_{i}^{0}=\mathbf{0}$ $\forall i\in\mathcal{V}$. Below, we discuss the selections of $u^{k}$, $H$, and $\tilde{H}$ for the sake of distributed implementations of (9) and (10).

To this end, we let

and impose Assumption 3 on $P,\tilde{P}\in\mathbb{R}^{N\times N}$.

The nodes can jointly determine $P,\tilde{P}$ under Assumption 3 without any centralized coordination. For instance, we can let each node $i\in\mathcal{V}$ agree with every neighbor $j\in\mathcal{N}_{i}$ on $p_{ij}=p_{ji}<0$ and $\tilde{p}_{ij}=\tilde{p}_{ji}<0$, and set $p_{ii}=-\sum_{j\in\mathcal{N}_{i}}p_{ij}$ and $\tilde{p}_{ii}=-\sum_{j\in\mathcal{N}_{i}}\tilde{p}_{ij}$, which directly guarantee Assumption 3(a). Then, Assumption 3(c)(d) are satisfied effortlessly by means of Assumption 3(b) and the connectivity of $\mathcal{G}$. Typical examples of such $P$ and $\tilde{P}$ include the graph Laplacian matrix $L_{\mathcal{G}}\in\mathbb{R}^{N\times N}$ with $[L_{\mathcal{G}}]_{ij}=[L_{\mathcal{G}}]_{ji}=-1$ $\forall\{i,j\}\in\mathcal{E}$ and the Metropolis weight matrix $M_{\mathcal{G}}\in\mathbb{R}^{N\times N}$ with $[M_{\mathcal{G}}]_{ij}=[M_{\mathcal{G}}]_{ji}=-\frac{1}{\max\{|\mathcal{N}_{i}|,|\mathcal{N}_{j}|\}+1}$ $\forall\{i,j\}\in\mathcal{E}$. Also, the conditions on $H$ and $\tilde{H}$ in Section II-A hold due to Assumption 3.

With Assumption 3, (10) can be accomplished by letting each node update as

This is a local operation, as each node $i\in\mathcal{V}$ only needs to acquire the primal variables associated with its neighbors.

It remains to design the surrogate function $u^{k}$ so that (9) can be realized in a distributed way. To this end, let $\psi_{i}^{k}:\mathbb{R}^{d}\rightarrow\mathbb{R}$ $\forall i\in\mathcal{V}$ $\forall k\geq 0$ be a set of functions under Assumption 4.

To meet Assumption 4(d), it suffices to let $\psi_{i}^{k}$ $\forall i\in\mathcal{V}$ $\forall k\geq 0$ be any strongly convex functions whose convexity parameters are larger than or equal to $\rho\lambda_{\max}(H)>0$, and one readily available upper bound on $\lambda_{\max}(H)$ is $\max_{i\in\mathcal{V},\;j\in\mathcal{N}_{i}}|p_{ij}|N$.

Now define

and let $D_{\phi^{k}}(\mathbf{x},\mathbf{x}^{k})$ be the Bregman divergence associated with $\phi^{k}$ at $\mathbf{x}$ and $\mathbf{x}^{k}$, i.e.,

Then, we set $u^{k}$ as

With Assumption 4, (15) is sufficient to ensure Assumption 2. To see this, note from Assumption 4(a)(d) that $u^{k}(\mathbf{x})$ in (15) is twice continuously differentiable and convex, i.e., Assumption 2(a) holds. Also note from (14) and (15) that

This, along with Assumption 4 (b)(c), guarantees Assumption 2(b)(c)(d).

To see how (15) results in a distributed implementation of (9), note from (16) that (9) is equivalent to

for some $g^{k+1}\in\partial h(\mathbf{x}^{k+1})$. Then, using (13) and the structure of $H$ given in Assumption 3, this can be written as

where $g_{i}^{k+1}\in\partial h_{i}(x_{i}^{k+1})$. In other words, (9) can be achieved by letting each node $i\in\mathcal{V}$ solve the following strongly convex optimization problem:

which can be locally carried out by node $i$.

Algorithms described by (11), (17), and (12) under Assumptions 3 and 4 constitute a set of distributed realizations of AMM, referred to as Distributed Approximate Method of Multipliers (DAMM), which can be implemented by the nodes in $\mathcal{G}$ via exchanging information with their neighbors only. Algorithm 1 describes how each node acts in DAMM.

Finally, we provide two examples of DAMM with two particular choices of $\psi_{i}^{k}$.

In this subsection, we focus on the smooth convex optimization problem $\min_{x\in\mathbb{R}^{d}}\sum_{i\in\mathcal{V}}f_{i}(x)$, i.e., (1) with $h_{i}(x)\equiv 0$ $\forall i\in\mathcal{V}$. For this special case of (1), we provide additional designs of the surrogate function $u^{k}$ in AMM, leading to a couple of variations of DAMM.

Here, we let $u^{k}$ still be in the form of (15), but no longer require $\phi^{k}+\frac{\rho}{2}\|\cdot\|_{H}^{2}$ be a separable function as in (13). Instead, we construct $\phi^{k}$ based upon another function $\gamma^{k}:\mathbb{R}^{Nd}\rightarrow\mathbb{R}$ under Assumption 5.

From Assumption 5(b)(c), each $(\gamma^{k})^{\star}$ is $(1/\bar{\gamma})$-strongly convex and $(1/\underline{\gamma})$-smooth [41], so that Assumption 5(d) holds as long as $I_{Nd}/\bar{\gamma}-\rho H\succeq\mathbf{O}$. Now we set

Unlike DAMM, $u^{k}$ given by (15) and (18) under Assumption 5 does not necessarily guarantee that $u^{k}(\cdot)+\frac{\rho}{2}\|\cdot\|_{H}^{2}$ is separable. Below, we show that such $u^{k}$ $\forall k\geq 0$ also satisfy Assumption 2, leading to another subclass of AMM.

To do so, first notice that the strong convexity and smoothness of $(\gamma^{k})^{\star}$ guarantee Assumption 2(b)(c). Also note from (16) that Assumption 2(d) is assured. In addition, due to Assumption 5(d), $\phi^{k}$ is convex and, thus, so is $u^{k}$. To show the twice continuous differentiability of $u^{k}$ in Assumption 2(a), consider the fact from [42] that due to Assumption 5(b)(c), $\nabla\gamma^{k}$ is invertible and its inverse function is $(\nabla\gamma^{k})^{-1}=\nabla(\gamma^{k})^{\star}$. This, along with (18), implies that

From the inverse function theorem [43], $(\nabla\gamma^{k})^{-1}$ is continuously differentiable, so that $\phi^{k}$ and therefore $u^{k}$ given by (15) and (18) are twice continuously differentiable. We then conclude that Assumption 2 holds.

Equipped with (15), (18), and Assumption 5, we start to derive additional distributed realizations of AMM (9)–(11) for minimizing $\sum_{i\in\mathcal{V}}f_{i}(x)$. As $h_{i}(x)\equiv 0$ $\forall i\in\mathcal{V}$, (9) is equivalent to

This, together with (19), gives

Like DAMM, we let each node $i\in\mathcal{V}$ maintain the $i$th $d$-dimensional blocks of $\mathbf{x}^{k}$ and $\mathbf{q}^{k}$, i.e., $x_{i}^{k}$ and $q_{i}^{k}$, where the update of $q_{i}^{k}$ is the same as (12). According to (21), the update of $x_{i}^{k}$ is given by $x_{i}^{k+1}=\nabla_{i}\gamma^{k}(\nabla\phi^{k}(\mathbf{x}^{k})-\nabla f(\mathbf{x}^{k})-\mathbf{q}^{k})$. From Assumption 5(e), this can be computed by node $i$ provided that it has access to $\nabla_{j}\phi^{k}(\mathbf{x}^{k})-\nabla f_{j}(x_{j}^{k})-q_{j}^{k}$ $\forall j\in\mathcal{N}_{i}\cup\{i\}$, where $\nabla_{j}\phi^{k}(\mathbf{x}^{k})\in\mathbb{R}^{d}$ denotes the $j$th block of $\nabla\phi^{k}(\mathbf{x}^{k})$. Therefore, the remaining question is whether each node $i\in\mathcal{V}$ is able to locally evaluate $\nabla_{i}\phi^{k}(\mathbf{x}^{k})$. In fact, this can be enabled by the following two concrete ways of constructing $\gamma^{k}$.

Way #1: Let $\gamma^{k}=\gamma$ $\forall k\geq 0$ for some $\gamma:\mathbb{R}^{Nd}\rightarrow\mathbb{R}$ satisfying Assumption 5. Then, according to (18), $\phi^{k}$ $\forall k\geq 0$ are fixed to some $\phi:\mathbb{R}^{Nd}\rightarrow\mathbb{R}$. We introduce an auxiliary variable $\mathbf{y}^{k}=((y_{1}^{k})^{T},\ldots,(y_{N}^{k})^{T})^{T}$ such that $\mathbf{y}^{k}=\nabla\phi(\mathbf{x}^{k})$ $\forall k\geq 0$. From (20), $\mathbf{y}^{k}$ satisfies the recursive relation

Due to the structure of $H$ in Assumption 3, each node $i\in\mathcal{V}$ is able to locally update $y_{i}^{k+1}$ as above. Nevertheless, the initialization $\mathbf{y}^{0}=\nabla\phi(\mathbf{x}^{0})=(\nabla\gamma)^{-1}(\mathbf{x}^{0})-\rho H\mathbf{x}^{0}$ cannot be achieved without centralized coordination, given that $\mathbf{x}^{0}$ is arbitrarily chosen. We may overcome this by imposing a restriction on $\mathbf{x}^{0}$ as follows: Let each node $i\in\mathcal{V}$ pick any $\tilde{z}_{i}\in\mathbb{R}^{d}$, and force $x_{i}^{0}=\nabla_{i}\gamma(\tilde{\mathbf{z}})$ with $\tilde{\mathbf{z}}=(\tilde{z}_{1}^{T},\ldots,\tilde{z}_{N}^{T})^{T}$, so that $\tilde{\mathbf{z}}=(\nabla\gamma)^{-1}(\mathbf{x}^{0})$. Then, by letting each $y_{i}^{0}=\tilde{z}_{i}-\rho\sum_{j\in\mathcal{N}_{i}\cup\{i\}}p_{ij}x_{j}^{0}$, we obtain $\mathbf{y}^{0}=\nabla\phi(\mathbf{x}^{0})$. Due to Assumption 5(e), such initialization is decentralized, which only requires each node $i$ to share $\tilde{z}_{i}$ and $x_{i}^{0}$ with its neighbors.

Incorporating the above into (21) yields another distributed version of AMM, which is called Distributed Approximate Method of Multipliers for Smooth problems with Constant update functions (DAMM-SC) and is specified in Algorithm 2.