implicit-func-relaxations

Convex relaxations are used in typical methods for global optimization and reachable set generation. Suppose that:

• we have an implicit function $\mathbf{x}$ defined in terms of a known residual function $\mathbf{f}$ as satisfying:

$$ \mathbf{f}(\mathbf{x}(\mathbf{p}),\mathbf{p}) = \mathbf{0}, \qquad\forall\mathbf{p}\in P, $$

• we can also generate convex/concave relaxations $\mathbf{f}^{\text{cv}}/\mathbf{f}^{\text{cc}}$ for $\mathbf{f}$ on $P$, perhaps using McCormick.jl, and
• we know a box $X\subset\mathbb{R}^{n_x}$ containing the range of $\mathbf{x}$.

This repository illustrates our new approach for constructing convex/concave relaxations for the unknown implicit function $\mathbf{x}$ in terms of known information, and contains our Julia code for all numerical examples in our accompanying article.

This implementation was developed by Huiyi Cao in Julia. This repository is tied to the accompanying article, and will not be updated except for bug fixes. If you make use of this code, please cite our article:

Huiyi Cao and Kamil A. Khan, General convex relaxations of implicit functions and inverse functions, Journal of Global Optimization, in press, 2023. doi:10.1007/s10898-023-01281-0

This work was supported by the McMaster Advanced Control Consortium (MACC), and by the Natural Sciences and Engineering Research Council of Canada (NSERC) under Grant RGPIN-2017-05944.

Dependencies

• McCormick.jl v0.13.1
• DataFrames.jl v1.3.4
• IntervalArithmetic.jl v0.20.6
• JuMP.jl v0.21.5
• Ipopt.jl v0.6.5
• NLsolve.jl v4.5.1
• Plots.jl v1.31.3
• LaTeXStrings v1.3.0

Method outline

With the setup described above, we describe convex/concave relaxations of the implicit function $\mathbf{x}$ as follows, for each component $i$ and each $\mathbf{p}\in P$:

$$ \begin{align*} x_i^{\mathrm{cv}}(\mathbf{p})&:=\inf_{\mathbf{z}\in X} z_i \quad\text{subject to}\quad\mathbf{f}^{\text{cv}}(\mathbf{z},\mathbf{p})\leq\mathbf{0}\leq\mathbf{f}^{\text{cc}}(\mathbf{z},\mathbf{p}), \\ x_i^{\mathrm{cc}}(\mathbf{p})&:=\sup_{\mathbf{z}\in X} z_i \quad\text{subject to}\quad\mathbf{f}^{\text{cv}}(\mathbf{z},\mathbf{p})\leq\mathbf{0}\leq\mathbf{f}^{\text{cc}}(\mathbf{z},\mathbf{p}). \end{align*} $$

When linearity or monotonicity can be exploited, these relaxations are evaluated particularly easily. This formulation is simple and versatile, and does not actually require existence or uniqueness of $\mathbf{x}$ on the presumed domain. Inverse functions and constraint-satisfaction problems may be relaxed analogously. For more details, please refer to the accompanying manuscript.