Exponential Decay of Sensitivity in Graph-Structured Nonlinear Programs

Abstract

We study solution sensitivity for nonlinear programs (NLPs) whose structure is induced by a graph G = (V, E). These graph-structured NLPs arise in many applications such as dynamic optimization, stochastic optimization, optimization with partial differential equations, and network optimization. We show that the sensitivity of the primal-dual solution at node i ∈ V against a data perturbation at node j ∈ V is bounded by Υρ d G (i,j) for constants Υ > 0 and ρ ∈ (0, 1) and where d G (i, j) is the distance between i and j on G. In other words, the sensitivity of the solution decays exponentially with the distance to the perturbation point. This result, which we call exponential decay of sensitivity (EDS), holds under fairly standard assumptions used in classical NLP sensitivity theory: the strong second-order sufficiency condition and the linear independence constraint qualification. We also present conditions under which the constants (Υ, ρ) remain uniformly bounded; this allows us to characterize behavior for NLPs defined over subgraphs of infinite graphs (e.g., as those arising in problems with unbounded domains). Our results provide new insights on how perturbations propagate through the NLP graph and on how the problem formulation influences such propagation. Specifically, we provide empirical evidence that positive objective curvature and constraint flexibility tend to dampen propagation. The developments are illustrated with numerical examples.

A Fast Temporal Decomposition Procedure for Long-horizon Nonlinear Dynamic Programming

“…In addition, FOTD can be applied on graph-structured problems, seeing that a similar exponential decay of sensitivity for graph-structured problems was established in [37]. Finally, a reasonable conjecture for the improved performance of the Schwarz scheme over ADMM is the lack of information exchange between subproblems in ADMM.…”

Condensed interior-point methods: porting reduced-space approaches on GPU hardware

“…To improve the performance of reduction algorithms, we believe the most important item is to alleviate the dependence on the number of control variables. We plan to explore a way to accelerate the reduction by exploiting the exponentially decaying structure of the reduced Hessian [34].…”

A Unifying Modeling Abstraction for Infinite-Dimensional Optimization

“…This modeling feature can be useful in dynamic optimization and optimal control problems in which it is often desirable to place more/less emphasis on initial or final conditions. For instance, in infinite-horizon problems, wp¨q obtained from the probability density function of an exponential density function behaves as a discount factor [48,32].…”

“…For instance, it has been recently shown that a graph abstraction unifies a wide range of optimization problems such as discrete-time dynamic optimization (graph is a line), network optimization (graph is the network itself), and multistage stochastic optimization (graph is a tree) [30]. This unifying abstraction has helped identify new theoretical properties and decomposition algorithms [31,32]; this has been enabled, in part, via transferring techniques and concepts across disciplines. The limited availability of unifying modeling tools ultimately limits our ability to experiment with techniques that appear in different disciplines and limits our ability to identify new modeling abstractions to tackle emerging applications.…”