Abstract
Clarke’s inverse function theorem for Lipschitz mappings states that a bi-Lipschitz mapping f is locally invertible about a point \(x_0\) if the generalized Jacobian \(\partial f(x_0)\) does not contain singular matrices. It is shown that under these assumptions the generalized Jacobian of the inverse mapping at \(f(x_0)\) is the convex hull of the set of matrices that can be obtained as limits of sequences \(J_f(x_k)^{-1}\) with f differentiable in \(x_k\) and \(x_k\) converging to \(x_0\). This identity holds as well if f is assumed to be locally bi-Lipschitz at \(x_0\).
Similar content being viewed by others
1 Introduction
F. H. Clarke introduced in [1, 2] the concepts of generalized gradients and generalized Jacobians as a tool to investigate Lipschitz continuous real-valued functions and proved a version of the inverse function theorem. To state this result precisely, we denote the Jacobian matrix of a mapping \(f:\mathbb {R}^n\rightarrow \mathbb {R}^m\), which is differentiable in \(x_0\), by \(J_f(x_0)\). By Rademacher’s theorem, every Lipschitz mapping is \(\lambda ^n\)-a.e. differentiable and the norm of the Jacobian matrix is bounded by the Lipschitz constant. The convex hull of a set \(A\subseteq \mathbb {R}^n\) is denoted by \({{\,\textrm{conv}\,}}(A)\). Throughout this note, the Euclidean norm of a vector \(x\in \mathbb {R}^n\) is written as \(\Vert x\Vert \) and the induced matrix norm is denoted by the same symbol.
Definition 1.1
For \(m, n\in \mathbb {N}\) and \(f: \mathbb {R}^n\rightarrow \mathbb {R}^m\) locally Lipschitz continuous, the set of points in which f is not differentiable is denoted by \(N_f\). For any \(x_0\in \mathbb {R}^n\) define
and the Clarke generalized Jacobian of f at \(x_0\) by \(\partial f(x_0):= {{\,\textrm{conv}\,}}(\mathcal {J}_f(x_0))\).
Since bounded sets in \(\mathbb {R}^{m\times n}\) are precompact, this set is not empty. Clarke’s inverse function theorem gives a sufficient condition for local invertibility of a Lipschitz continuous mapping \(f: \mathbb {R}^d\rightarrow \mathbb {R}^d\) at some point \(x_0\in \mathbb {R}^d\) by requiring that every matrix in the Clarke generalized Jacobian \(\partial f(x_0)\) be invertible.
Theorem 1.1
([2, Theorem 1]) Assume that \(x_0\in \mathbb {R}^d\) and that \(f: \mathbb {R}^d\rightarrow \mathbb {R}^d\) is Lipschitz continuous in a neighborhood of \(x_0\) with any matrix in \(\partial f(x_0)\) being invertible. Then there exist neighborhoods \(U\subseteq \mathbb {R}^d\) of \(x_0\), \(V\subseteq \mathbb {R}^d\) of \(f(x_0)\) and a Lipschitz continuous mapping \(g: V\rightarrow \mathbb {R}^d\), such that
-
(a)
\(g(f(u))= u\) for every \(u\in U\),
-
(b)
\(f(g(v))= v\) for every \(v\in V\).
The fundamental assertion of the theorem is the existence of an inverse mapping which is Lipschitz continuous. Clarke did not include a statement about the generalized Jacobian of the inverse mapping, and it was recently shown in [3] that the assumptions of Theorem 1.1 imply the relations
It was left open whether equality holds in (1), an assertion, which gives (2) as an immediate consequence by taking the convex hull on both sides. In the next section it is shown that equality holds in (1) even if only the existence of a Lipschitz continuous local inverse mapping g is assumed instead of the invertibility of any matrix in \(\partial f(x_0)\). A crucial observation is that in fact \(f(N_f)= N_g\) and \(g(N_g)= N_f\). Under these weaker assumptions, a formula for the Clarke generalized Jacobian is, to the best of our knowledge, not available in the literature.
2 Main Result
Theorem 2.1
Assume that \(U,V\subseteq \mathbb {R}^d\) are open and that \(f: U\rightarrow \mathbb {R}^d\) and \(g: V\rightarrow \mathbb {R}^d\) are Lipschitz continuous with \(g\circ f= \textrm{Id}_U\) and \(f\circ g= \textrm{Id}_V\). Then \(f(N_f)= N_g\), \(g(N_g)= N_f\) and for any \(x_0\in U\) the set \(\mathcal {J}_f(x_0)\) is invertible with
In particular
Proof
Let \(L>0\) be a Lipschitz constant for g and fix any \(x\in U{\setminus } N_f\). Then \(J_f(x)\) is injective, since for every \(h\in \mathbb {R}^d\) and \(t\in \mathbb {R}{\setminus } \{0\}\) with \(x+th\in U\)
This lower bound implies injectivity of \(J_f(x)\) since
Once \(J_f(x)\) is shown to be injective, according to [7, III, Theorem 3.3], this matrix is invertible. By [3, Lemma 1], or more generally by [4, Corollary 3.1], g is differentiable at f(x) with \(J_g(f(x))= J_f(x)^{-1}\). In particular, \(f(U{\setminus } N_f)\subseteq V{\setminus } N_g\) and consequently
With the same argument for g, we have \(N_f\subseteq g(N_g)\) and
which implies \(f(N_f)= N_g\). Analogously we obtain \(g(N_g)= N_f\). Fix now \(x_0\in U\), \(M\in \mathcal {J}_f(x_0)\) and a sequence \(x: \mathbb {N}\rightarrow U{\setminus } N_f\) converging to \(x_0\) with \(J_f(x_\ell )\) converging to M for \(\ell \rightarrow \infty \). Then M is injective, since by (4) for any \(h\in \mathbb {R}^d\) the uniform lower bound
is satisfied. Hence by [7, III, Theorem 3.3], M is invertible. By the continuity of the matrix inversion on the set of invertible matrices, \((J_g(f(x_\ell )))_{\ell \in \mathbb {N}}= (J_f(x_\ell ))^{-1}\) converges to \(M^{-1}\in \mathcal {J}_g(f(x_0))\) for \(\ell \rightarrow \infty \). Consequently \(\mathcal {J}_f(x_0)^{-1}\subseteq \mathcal {J}_g(f(x_0))\). By the same argument for g we have that the set \(\mathcal {J}_g(f(x_0))\) is invertible with \(\mathcal {J}_g(f(x_0))^{-1}\subseteq \mathcal {J}_f(g(f(x_0)))\). \(\square \)
For a convenience of a reader we present a special case of the example constructed in [6, Example 2.2] and [5, Example 3.9], which provides a piecewise linear bi-Lipschitz mapping from \(\mathbb {R}^2\) to \(\mathbb {R}^2\), to which Theorem 1.1 cannot be applied at the origin due to singularity of the Clarke generalized Jacobian. Nevertheless the inverse function and its Clarke generalized Jacobian can be given explicitly, which verifies formula (3) from Theorem 2.1.
Example 2.1
Setting \(i=\sqrt{-1}\), we use polar coordinates in the plane with \(x=r \exp (i\phi )\), \(\phi \in [0,2\pi )\), \(r>0\). Define
and \(a_j=\exp (i\alpha _j)\), \(b_j=\exp (i\beta _j)\), \(j=1,\ldots ,7\). For \(j=1,\ldots ,6\), define the matrix \(M_j\) by \(M_j a_j=b_j\) and \(M_{j}a_{j+1}=b_{j+1}\), which leads to
and
Since \(\alpha _{j+1}-\alpha _j\), \(\beta _{j+1}-\beta _j\in (0,\pi )\), all the matrices \(M_j\) are invertible with positive determinant. The mapping
is bi-Lipschitz with the inverse mapping
The mappings f and g are \(\lambda _2\)-a.e. differentiable with
the range of the derivative consists for f and g of the six matrices \(M_j\) and \(M_j^{-1}\), respectively. Moreover, we have \(N_g= f(N_f)\), \(N_f= g(N_g)\) and
as asserted in Theorem 2.1. Since \(0=\frac{1}{2}(\textrm{Id}+(-\textrm{Id}))\in \partial f(0) \cap \partial g(0)\), the assumptions in Clarke’s inverse function theorem are not satisfied for f or g. However, since the inverse mapping is known to exist, Theorem 2.1 is applicable and this example nicely illustrates formula (3).
3 Conclusions
We presented a formula for the Clarke generalized Jacobian of the inverse \(f^{-1}\) of a bi-Lipschitz map \(f:\mathbb {R}^d\rightarrow \mathbb {R}^d\) which relates the Clarke generalized Jacobian of \(f^{-1}\) to the Clarke generalized Jacobian of f. In particular, if the assumptions of Clarke’s inverse function theorem hold, then our result provides a formula for the subgradient of the inverse map. Of course, the most important part in Clarke’s theorem is the existence of the inverse map and it is a challenging open problem to identify sufficient conditions for the existence of an inverse map which is Lipschitz continuous if the Clarke generalized Jacobian of a Lipschitz function contains singular matrices.
Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.
References
Clarke, F.H.: Generalized gradients and applications. Trans. Am. Math. Soc. 205, 247–262 (1975)
Clarke, F.H.: On the inverse function theorem. Pac. J. Math. 64, 97–102 (1976)
Fabian, M., Hiriart-Urruty, J.-B., Pauwels, E.: On the generalized Jacobian of the inverse of a Lipschitzian mapping. Set-Valued Var. Anal. 30, 1443–1451 (2022)
Fabián, M.J.: Concerning a geometrical characterization of Fréchet differentiability. Časopis pro pěstování matematiky 104(1), 45–64 (1979)
Fusek, P., Klatte, D., Kummer, B.: Examples and counterexamples in Lipschitz analysis. Control. Cybern. 31, 471–492 (2002)
Kummer, B.: Lipschitzian inverse functions, directional derivatives, and applications in \(C^{1,1}\) optimization. J. Optim. Theory Appl. 70, 561–581 (1991)
Lang, S.: Linear Algebra. Springer, New York (1987)
Acknowledgements
The authors gratefully acknowledge funding by the German Research Foundation (DFG) within project 441211072/441468770 “Variational modeling of pressure-dependent plasticity - a paradigm for model reduction via relaxation” within the Priority Program 2256 “Variational Methods for Predicting Complex Phenomena in Engineering Structures and Materials.” They thank the anonymous referee for providing them with the references [3] and [4] which resulted in shortening our argument.
Funding
Open Access funding enabled and organized by Projekt DEAL.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Aram Arutyunov.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Behr, F., Dolzmann, G. A Note on Clarke’s Generalized Jacobian for the Inverse of Bi-Lipschitz Maps. J Optim Theory Appl 200, 852–857 (2024). https://doi.org/10.1007/s10957-023-02333-x
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10957-023-02333-x