Iteration complexity of the Difference-of-Convex Algorithm for unconstrained optimization: a simple proof

Abstract: We propose a simple proof of the worst-case iteration complexity for the Difference of Convex functions Algorithm (DCA) for unconstrained minimization, showing that the global rate of convergence of the norm of the objective function's gradients at the iterates converge to zero like o(1/k). A small example is also provided indicating that the rate cannot be improved.

• The paper provides a simplified proof of the worst-case iteration complexity for the Difference-of-Convex Algorithm achieving an $o(1/k)$ convergence rate directly without relying on semi-definite programming frameworks
• The study enumerates an explicit construction showing the rate matched except for an arbitrarily close gap, ensuring the tightness of the derived bound
• The authors conclude and demonstrate the absence of worst-case complexity advantages in DCA's iteration structure for structured decompositions of DC functions, indicating a need for strategies to avoid potential pitfalls when exploiting this decomposition

Overview

This short note by Gratton and Toint revisits the worst-case iteration complexity of the Difference-of-Convex Algorithm (DCA) for unconstrained minimization of DC functions, $f(x) = g(x) - h(x)$ with $g$ convex and strongly convex and $h$ convex with Lipschitz-continuous gradient (2601.15970). Prior analyses of this question relied on performance estimation techniques in the spirit of Drori and Teboulle's semidefinite-programming-based framework, which involves relaxing an infinite-dimensional optimization problem over an infinite set of constraints. The contribution here is methodological: both the upper bound on gradient norm decay and a matching lower bound are derived through direct, elementary arguments that require no SDP machinery, at the cost of some loss of generality in the assumptions.

The algorithm and its monotonicity

The DCA iteration is

$$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$

so each step solves a convex subproblem (the objective inherits the convexity of $g$). Because $h$ is convex, $g(x) - h(x_k) - \nabla h(x_k)^T(x - x_k)$ overestimates $f$, which immediately yields monotone decrease $f(x_{k+1}) \le f(x_k)$. A key structural identity follows from the first-order optimality condition of the subproblem: $\nabla g(x_{k+1}) = \nabla h(x_k)$. This single relation drives the entire analysis.

Upper bound: an $g$0 rate

Combining strong convexity of $g$1 with convexity of $g$2 around the stationarity identity gives, for all $g$3,

$g$4

and summing telescopes to $g$5. Meanwhile, the Lipschitz continuity of $g$6 together with $g$7 yields $g$8. Chaining these inequalities produces the average-square-gradient bound

$g$9

since $h$0 converges by monotonicity and boundedness below. This recovers the upper bound of Abbade et al.'s performance-estimation analysis but strengthens it from $h$1 to $h$2 — a small but genuine improvement obtained essentially for free from the direct argument. For an $h$3-stationarity termination test $h$4, this guarantees termination in at most $h$5 iterations. The proof also recovers known results on the iterates themselves, notably the summability of squared successive displacements previously established in the DC literature.

Lower bound: tightness via an explicit construction

The paper then shows the rate cannot be improved: for any $h$6, there exist admissible $h$7 and $h$8 such that DCA started at $h$9 satisfies

$$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$0

which is arbitrarily close to the $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$1 behavior implied by the upper bound as $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$2. The construction is explicit and interpretable: take $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$3 univariate, prescribe iterates $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$4 (which diverge to $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$5), and build $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$6 as a piecewise quadratic interpolant satisfying $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$7, extended linearly for $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$8. The construction verifies that $$x_{k+1} = \arg\min_{x} \; g(x) - \nabla h(x_k)^T (x - x_k),$$9 is convex with unit-Lipschitz gradient and that $g$0 is bounded below (by roughly $g$1), so all hypotheses of the upper-bound theorem hold. The DCA subproblem then reproduces exactly the prescribed slow sequence, and $g$2. The example extends verbatim to any dimension by making $g$3 depend on a single coordinate, and it is independent of any particular termination rule — unlike several slow-convergence examples in the optimization literature.

Discussion: what the bounds imply about exploiting DC structure

The central conclusion is negative for worst-case theory: DCA applied to a structured DC decomposition achieves no better iteration complexity than steepest descent applied directly to $g$4 while ignoring the DC structure. Exploiting the difference-of-convex decomposition per se therefore confers no worst-case benefit under this algorithm. More pointedly, the lower-bound example shows structure exploitation can be detrimental: the constructed $g$5 is itself convex (its Hessian is strictly positive on each interval $g$6, though it is not strongly convex), yet DCA performs worse on it than optimal methods designed for general convex objectives would. The authors are careful to note the standard caveat that worst-case bounds do not preclude good practical performance, just as they do not for Newton-type variants whose worst case matches steepest descent.

On evaluation complexity, the note observes only that it is bounded below by iteration complexity, since each inner minimization requires at least one oracle call; the cost of solving the convex subproblems to sufficient accuracy is not analyzed.

Limitations and open questions

The analysis carries assumptions worth flagging. The upper bound requires $g$7 to be strongly convex, not merely convex, and $g$8 to be bounded below; whether the $g$9 rate persists under weaker conditions is not addressed. On the lower-bound side, the construction genuinely depends on $h$0: for $h$1 the constructed $h$2 ceases to be bounded below unless an $h$3-termination rule is imposed, so exact matching of the $h$4 constant regime remains open. Finally, the equivalence claimed between DCA and steepest descent concerns iteration counts only; total oracle-evaluation complexity, including the inner convex solves, is left open beyond the trivial one-evaluation-per-iteration lower bound.

Conclusion

The paper delivers a concise, self-contained proof that DCA's worst-case gradient-norm convergence rate is $h$5 on average, matched up to arbitrarily small gaps by an explicit one-dimensional example, thereby reproducing and slightly sharpening prior performance-estimation results without their technical overhead. Its substantive message is that DC decomposability, when exploited through the standard DCA, yields no worst-case iteration-complexity advantage over structure-agnostic first-order methods — and can even be counterproductive on functions that are themselves convex.