Abstract
In a Hilbert framework, we introduce a new class of second-order dynamical systems that combine viscous and geometric damping but also a time rescaling process for nonsmooth convex minimization. A main feature of these systems is to produce trajectories that lie in the graph of the Fenchel subdifferential of the objective. Moreover, they do not incorporate any regularization or smoothing processes. This new class originates from some combination of a continuous Nesterov-like dynamic and the Minty representation of subdifferentials. These models are investigated through first-order reformulations that amount to dynamics involving three variables: two solution trajectories (including an auxiliary one) and another one associated with subgradients. We prove the weak convergence towards equilibria for the solution trajectories, as well as properties of fast convergence to zero for their velocities. Remarkable convergence rates (possibly of exponential-type) are also established for the function values. We additionally state notable properties of fast convergence to zero for the subgradients trajectory and for its velocity. Some numerical experiments are performed so as to illustrate the efficiency of our approach. The proposed models offer a new and well-adapted framework for discrete counterparts, especially for structured minimization problems. Inertial algorithms with a correction term are then suggested relative to this latter context.
Similar content being viewed by others
Data Availibility
We do not analyse or generate any datasets, because our work proceeds within a theoretical and mathematical approach. One can obtain the relevant materials from the references below.
References
Abbas, B., Attouch, H.: Dynamical systems and forward-backward algorithms associated with the sum of a convex subdifferential and a monotone cocoercive operator. Optimization 64, 2223–2252 (2015)
Abbas, B., Attouch, H., Svaiter, B.F.: Newton-like dynamics and forward-backward methods for structured monotone inclusions in Hilbert spaces. J. Optim. Theory Appl. 161(2), 331–360 (2014)
Alvarez, F., Attouch, H., Bolte, J., Redont, P.: A second-order gradient-like dissipative dynamical system with Hessian driven damping. Application to optimization and mechanics. J. Math. Pures Appl. 81(8), 747–779 (2002)
Apidopoulos, V., Aujol, J.-F., Dossal, Ch.: The differential inclusion modeling the FISTA algorithm and optimality of convergence rate in the case \(b \le 3\). SIAM J. Optim. 28(1), 551–574 (2018)
Attouch, H., Cabot, A.: Convergence of damped inertial dynamics governed by regularized maximally monotone operators. J. Differ. Equ. 264, 7138–7182 (2018)
Attouch, H., Cabot, A.: Convergence of a relaxed inertial forward-backward algorithm for structured monotone inclusions. Appl. Math. Optim. 80, 547–598 (2019)
Attouch, H., László, S.C.: Continuous Newton-like inertial dynamics for monotone inclusions. Set Valued Var. Anal. 29, 555–581 (2021)
Attouch, H., Peypouquet, J.: Convergence of inertial dynamics and proximal algorithms governed by maximally monotone operators. Math. Program. 174, 391–432 (2019). https://doi.org/10.1007/s10107-018-1252-x
Attouch, H., Svaiter, B.F.: A continuous dynamical Newton-like approach to solving monotone inclusions. SIAM J. Control Optim. 49(2), 574–598 (2011)
Attouch, H., Bolte, J., Redont, P.: Optimizing properties of an inertial dynamical system with geometric damping: Link with proximal methods. Control Cybern. 31, 643–657 (2002)
Attouch, H., Peypouquet, J., Redont, P.: Fast convex minimization via inertial dynamics with Hessian driven damping. J. Differ. Equ. 261, 5734–5783 (2016)
Attouch, H., Chbani, Z., Peypouquet, J., Redont, P.: Fast convergence of inertial dynamics and algorithms with asymptotic vanishing viscosity. Math. Program. 168(1–2), 123–175 (2018)
Attouch, H., Chbani, Z., Riahi, H.: Fast proximal methods via time scaling of damped inertial gradient dynamics. SIAM J. Optim. 29(3), 2227–2256 (2019)
Attouch, H., Chbani, Z., Riahi, H.: Fast convex optimization via time scaling of damped inertial gradients dynamics. Pure Appl. Funct. Anal. 6(6), 1081–1117 (2021)
Attouch, H., Balhag, A., Chbani, Z., Riahi, H.: Fast convex optimization via inertial combining viscous and Hessian-driven damping with time rescaling dynamics. Evol. Equ. Control Theory 11(2), 487–514 (2022)
Attouch, H., Chbani, Z., Fadili, J., Riahi, H.: First-order optimization algorithms via inertial systems with Hessian driven damping. Math. Program. 193, 113–155 (2022). https://doi.org/10.1007/s10107-020-01591-1
Attouch, H., Chbani, Z., Fadili, J., Riahi, H.: Convergence of iterates for first-order optimization algorithms with inertia and Hessian driven damping. Optimization 72(5), 1199–1238 (2023). https://doi.org/10.1080/02331934.2021.2009828
Boţ, R.I., Hulett, D.A.: Second order splitting dynamics with vanishing damping for additively structured monotone inclusions. J. Dyn. Differ. Equ. (2022). https://doi.org/10.1007/s10884-022-10160-3
Boţ, R.I., Karapetyants, M.A.: A fast continuous time approach with time scaling for nonsmooth convex optimization. Adv. Cont. Disc. Models (2022). https://doi.org/10.1186/s13662-022-03744-2
Boţ, R.I., Csetnek, E., László, S.C.: On the strong convergence of continuous Newton-like inertial dynamics with Tikhonov regularization for monotone inclusions. J. Math. Anal. Appl. (2023). https://doi.org/10.13140/RG.2.2.20539.18729
Brezis, H.: Opérateurs maximaux monotones et semi-groupes de contractions dans les espaces de Hilbert. Math. Stud., vol. 5. North-Holland, Amsterdam (1973)
Brezis, H.: Function Analysis, Sobolev Spaces and Partial Differential Equations. Springer, New York (2010)
Cabot, A., Engler, H., Gadat, S.: On the long time behavior of second order differential equations with asymptotically small dissipation. Trans. Am. Math. Soc. 361, 5983–6017 (2009)
Cabot, A., Engler, H., Gadat, S.: Second order differential equations with asymptotically small dissipation and piecewise flat potentials. Electron. J. Differ. Equ. 17, 33–38 (2009)
Haraux, A.: Systémes dynamiques dissipatifs et applications, RMA17. Masson, Paris (1991)
Kim, D.: Accelerated proximal point method for maximally monotone operators (2019). Math. Program. 190, 57–87 (2021)
Labarre, F., Maingé, P.E.: First-order frameworks for continuous Newton-like dynamics governed by maximally monotone operators. Set Valued Var. Anal. 20(2), 425–451 (2022). https://doi.org/10.1007/s11228-021-00593-1
Luo, H.: Accelerated differential inclusion for convex optimization. Optimization 72(5), 1139–1170 (2023). https://doi.org/10.1080/02331934.2021.2002327
Maingé, P.E.: Accelerated proximal algorithms with a correction term for monotone inclusions. Appl. Math. Optim. 84(Suppl 2), 2027–2061 (2021)
Maingé, P.E., Weng-Law, A.: Fast continuous dynamics inside the graph of maximally monotone operators. Set Valued Var. Anal. (2023). https://doi.org/10.1007/s11228-023-00663-6
May, R.: Asymptotic for a second order evolution equation with convex potential and vanishing damping term. Turk. J. Math. 41(3), 681–785 (2015). https://doi.org/10.3906/mat-1512-28
Minty, G.J.: Monotone (nonlinear) operators in Hilbert spaces. Duke Math. J. 29, 341–346 (1962)
Opial, Z.: Weak convergence of the sequence of successive approximations for nonexpansive mappings. Bull. Am. Math. Soc. 73, 591–597 (1967)
Polyak, B.T.: Some methods of speeding up the convergence of iterative methods. USSR Comput. Math. Math. Phys. 4(5), 1–17 (1964)
Qu, X., Bian, W.: Fast inertial dynamic algorithm with smoothing method for nonsmooth convex optimization. Comput. Optim. Appl. 83, 287–317 (2022)
Shi, B., Du, S.S., Jordan, M.I., Su, W.J.: Understanding the acceleration phenomenon via high-resolution differential equations. Math. Program. 195, 79–148 (2022). https://doi.org/10.1007/s10107-021-01681-8
Sontaq, E.D.: Mathematical Control Theory, 2nd edn. Springer, New York (1998)
Su, W., Boyd, S., Candés, E.J.: A differential equation for modeling Nesterov’s accelerated gradient method: theory and insights. Neural Inf. Process. Syst. 27, 2510–2518 (2014)
Funding
The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors have no competing interests to declare that are relevant to the content of this article.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
1.1 Proof of Proposition 2.1
Let us prove that (i1) \(\Rightarrow \) (i2). Consider a solution \((x,\xi ) \in {\mathcal {A}}_c\times {\mathcal {A}}_c\) to (2.2). Set \(\zeta ( \cdot )=\sigma ( \cdot ) \xi ( \cdot )\) and suppose that \(x( \cdot )+ \zeta ( \cdot )\) is of class \(C^1\) and that \(\left( x( \cdot )+ \zeta ( \cdot ) \right) ^{(1)} \in {\mathcal {A}}_c\). Clearly, for \(t \ge 0\), as \(\dot{x}\), \({\dot{\zeta }}\) and \(\zeta \) are integrable on [0, t] (since x and \(\zeta \) belong to \({\mathcal {A}}_c\)), we can set as a well-defined quantity
Hence, \(z \in {\mathcal {A}}_c\), and by differentiating (1.9) we obtain
Therefore, by (2.2b) together with the above equality, we get
Moreover, recalling that \(\big (x( \cdot ) + \zeta ( \cdot ) \big )^{(1)} \in {\mathcal {A}}_c\) and \(z \in {\mathcal {A}}_c\), we get
\(\frac{d}{dt} \left( \big (x( \cdot ) + \zeta ( \cdot ) \big )^{(1)} + z( \cdot ) \right) (t) = \big (x( \cdot ) + \zeta ( \cdot ) \big )^{(2)}(t)+ \dot{z}(t)\), for a.e. \(t \ge 0\). Hence, we straightforwardly deduce
It follows immediately that
which, by the initial condition \(\big ( x( \cdot ) + \sigma ( \cdot )\xi ( \cdot )\big )^{(1)}(0) = q_0 \), yields
which readily implies that
Multiplying (1.15) by \(\beta (t)\) and adding the resulting equality to (1.10) give us
Now, since \(\theta ( \cdot )\) is positive, we introduce the function \(y( \cdot )\) defined for \(t \ge 0\) by
For simplification we also set \(u(t)=y(t)-x(t)\). Observe from (1.17) that we equivalently have
Thus, for \(t \ge 0\), (1.14) in light of the above equality entails
that is (2.4b). We now prove (2.4c). Differentiating (1.18), while noticing that \(\{x, \xi , u\} \subset {\mathcal {A}}_c\), readily implies, for a.e. \(t\ge 0\),
Moreover, using the definitions of \(\alpha ( \cdot )\), \(\beta ( \cdot )\) and \(b( \cdot )\) given by (1.3), namely \(\alpha (t) =-\frac{\dot{\theta }(t) }{\theta (t)} + \kappa -\theta (t)\), \(\beta (t) = -\frac{{\dot{\theta }}(t) }{\theta (t)} + \kappa +\omega (t) \) and \(b(t)= \omega (t) \left( \kappa +\frac{{\dot{\omega }} (t)}{\omega (t)}-\frac{\dot{\theta }(t)}{\theta (t)} \right) \), yields
Hence, by (1.16) and using (1.20), (1.18), \(\dot{u}=\dot{y} - \dot{x}\) and (1.21a), successively, we get, for a.e. \(t\ge 0\),
namely
In addition, by (1.19), while recalling that \(\{x,\xi \} \subset {\mathcal {A}}_c\), we obtain
Thus, combining (1.22) and (1.23), in light of \(\theta \ne 0\), yields
\(\dot{y}(t)+\kappa (u(t))=0\) for a.e. \(t\ge 0\),
that is (2.4c).
Finally, regarding the initial conditions, we have \((x(0),\xi (0))= (x_0,\xi _0)\) and \(\big (x( \cdot ) + \zeta ( \cdot )\big )(0)= q_0\) (according to (i1)) while (2.4b) at time \(t=0\) ensures that
\(\big ( x( \cdot ) + \zeta ( \cdot ) \big )^{(1)} (0) + \theta (0) \left( y(0)-x(0) \right) + \omega (0) \zeta (0) =0\).
Hence, we deduce that \(y(0) =x_0-\frac{1}{\theta (0)} \left( q_0 + \sigma (0) \omega (0) \xi _0 \right) \).
Let us prove that (i2) \(\Rightarrow \) (i1). Consider a solution \((x,\xi ,y) \in {\mathcal {A}}_c\times {\mathcal {A}}_c\times C^1\) to (2.4). For simplification, we set again \(u( \cdot )=y( \cdot )-x( \cdot )\) and \(\zeta ( \cdot )=\sigma ( \cdot ) \xi ( \cdot )\). Clearly, by (2.4b), we have, for \(t \ge 0\),
This, by \((x,\xi ,y) \in {\mathcal {A}}_c\times {\mathcal {A}}_c\times C^1\) and by \(\{\omega ( \cdot )\), \(\theta ( \cdot )\), \(\sigma ( \cdot )\} \subset C^1([0, \infty ])\), entails that \(x( \cdot )+\zeta ( \cdot )\) is of class \(C^1\) and that \(\big ( x( \cdot ) + \sigma ( \cdot )\xi ( \cdot ) \big )^{(1)} \in {\mathcal {A}}_c\). Then, differentiating (1.24) gives us, for a.e \(t \ge 0\),
while we know from (2.4c) that \(\dot{y}(t)= -\kappa u(t) \). Consequently, we readily obtain, for a.e \(t \ge 0\),
Furthermore, for a.e. \(t \ge 0\), by (1.24) we readily have
\(u(t)=-\frac{1}{\theta (t)} \big ( ( x( \cdot )+ \zeta ( \cdot )) ^{(1)} + \omega (t) \zeta (t) \big )\), which, by (1.26), entails
Hence the expressions of \(\alpha ( \cdot ), \beta ( \cdot )\) and \(b( \cdot )\) defined in (1.3) amounts to (2.2b). In addition, from the initial conditions in (i2), we have \(~x(0)=x_0\), \(~\xi (0)=\xi _0\) and \(y(0) =x_0-\frac{1}{\theta (0)} \left( q_0 + \sigma (0) \omega (0) \xi _0 \right) \). This implies that \(y(0) =x(0)-\frac{1}{\theta (0)} \left( q_0 + \sigma (0) \omega (0) \xi (0) \right) \), while (2.4b) at time \(t=0\) yields
\(\big ( x( \cdot ) + \zeta ( \cdot ) \big )^{(1)} (0) + \theta (0) \left( y(0)-x(0) \right) + \omega (0) \zeta (0) =0\).
Hence, regarding the last two equalities, substituting the former in the latter gives us \(\big ( x( \cdot ) + \zeta ( \cdot )\big )^{(1)}(0) = q_0 \).
1.2 Proof of Proposition 2.2
1.2.1 The Yosida Regularization
Some useful properties of the Yosida regularization are recalled through the lemma below established in [27] (see also [8, 21, 22]).
Lemma 1.1
Let \(A: {\mathcal {H}}\rightrightarrows {\mathcal {H}}\) be a maximally monotone operator such that \(S:=A^{-1}(0) \ne \emptyset \). Let \(\gamma ,\delta > 0\) and \(x, y \in {\mathcal {H}}\). Then for \(z \in A^{-1}(\{0\})\), we have
Proof
The proof of (1.27) can be found in [8]. For proving (1.28), we simply observe that \(\begin{array}{l} A_{\gamma }x- A_{\delta }y= \frac{1}{ \delta } \left( \delta A_{\gamma }x- \delta A_{\delta }y\right) = \frac{1}{ \delta } \left( ( \delta -\gamma ) A_{\gamma }x + ( \gamma A_{\gamma }x -\delta A_{\delta }y\right) , \end{array}\) from which we get \(\begin{array}{l} \Vert A_{\gamma }x- A_{\delta }y \Vert \le \frac{1}{ \delta } \left( | \delta -\gamma | \times \Vert A_{\gamma }x \Vert + \Vert \gamma A_{\gamma }x -\delta A_{\delta }y\Vert \right) . \end{array}\)
Consequently, by \( \Vert A_{\gamma }x\Vert \le \frac{1}{\gamma } \Vert x-z\Vert \) and using (1.27), we obtain (1.28) \(\square \)
1.2.2 Main Proof of the Proposition
The proof follows the same lines as in [30] (see, also, [1, 2]), but it is developed through the following steps (s1)- (s3) with full details:
(s1) We begin by reformulating the (possibly) existing strong global solutions to (1.4)) (that are supposed to satisfy (2.3)) by means of the Minty representation of the maximal monotone operator \(\partial f\) (see [32]). Set \(J_{ \sigma ( \cdot )}^{ \partial f}=\big (I+\sigma ( \cdot ) \partial f\big )^{-1}\) and \((\partial f)_{\sigma ( \cdot ) }= \frac{1}{\sigma ( \cdot )} \big (I-J_{\sigma ( \cdot ) } ^{ \partial f}\big )\), namely the resolvent and the Yosida approximation of \(\partial f\) (with index \(\sigma ( \cdot )\)), respectively, which are well-known to be single-valued and everywhere defined. Associated with any strong global solution \((x( \cdot ),\xi ( \cdot ),y( \cdot ))\) to (1.4), we introduce the new unknown function
It is readily seen that \(v( \cdot )\) belongs to \({\mathcal {A}}_c\) (the set of absolutely continuous functions) and that
\(v(0)= x_0+ \sigma (0) \xi _0\).
Moreover, for \(t \ge 0\), by \(\xi (t) \in \partial f (x(t))\) we obtain \(v(t) \in x(t)+ \sigma (t) \partial f(x(t))\) and \( \xi (t)=\frac{1}{\sigma (t)}(v(t)-x(t))\), hence, by Minty’s representation we simply have
Differentiating (1.29), in light of (2.3b), gives us, for a.e. \(t \ge 0\),
\(\dot{v}(t)=\dot{x}(t)+ (\sigma ( \cdot ) \xi ( \cdot ))^{(1)}(t) = -\theta (t) (y(t)-x(t))-\omega (t) \sigma (t) \xi (t)\), hence, by (1.30), we obtain
Hence, from (2.3), we deduce that \((v( \cdot ),y( \cdot ))\) are implicitly given, for a.e. \(t \in [0,\infty )\), by
together with \(y(0)=y_0\) and \(v(0)=x_0+ \sigma (0) \xi _0\).
This shows us that any strong global solution \((x( \cdot ),\xi ( \cdot ),y( \cdot ))\) to (1.4) is entirely determined (thanks to the two formulas in (1.30)) by some (strong) solution \((v( \cdot ),y( \cdot ))\) to (1.31). So, for proving existence and uniqueness of a strong global solution to (1.4), we just state (as argued below) the existence and uniqueness of a (strong) global solution \((v( \cdot ),y( \cdot ))\) to (1.31), but also the existence of a strong global solution \((x( \cdot ),\xi ( \cdot ),y( \cdot ))\) to (1.4).
(s2) Existence, uniqueness and regularity of a (strong) global solution \((v( \cdot ),y( \cdot ))\) to (1.31). First, we show that (1.31) is relevant to the Cauchy–Lipschitz theorem. Indeed, (1.31) can be expressed as
where \(U( \cdot )=(v( \cdot ),y( \cdot ))\) and \(F(t,.): {\mathcal {H}}^2 \rightarrow {\mathcal {H}}^2\) is defined for any \(t \ge 0\) and \( ({\bar{v}}, {\bar{y}}) \in {\mathcal {H}}^2\) by \(F(t,({\bar{v}},{\bar{y}}))= \big ( \phi _1(t,({\bar{v}},{\bar{y}})), \phi _2(t,({\bar{v}},{\bar{y}})) \big )\), together with
In view of applying the global Cauchy–Lipschitz theorem, we establish two main properties on F(., .) through the following items (a) and (b):
(a) Given \(({\bar{v}},{\bar{y}}) \in {\mathcal {H}}^2\), we prove that \(F(., (\bar{v},{\bar{y}}))\) is continuous on \([0,\infty )\). Indeed, let \(z \in (\partial f)^{-1}(0)\) and \((t_1,t_2) \in [0,\infty )^2\). By Lemma 1.1 with \(A= \partial f\) and \(\sigma ( \cdot ) \ge \sigma _0 >0\) (from (1.5b)), we obtain \( \Vert (\partial f)_{\sigma (t_1) }{\bar{y}}- (\partial f)_{\sigma (t_2) }{\bar{y}} \Vert \le 3 \frac{ | \sigma (t_1) - \sigma (t_2) |}{\sigma _0^2}\Vert {\bar{y}}-z\Vert \).
Then, the continuity of \(\sigma ( \cdot )\) on \([0,\infty )\) yields that the mappings \(t \rightarrow (\partial f)_{\sigma (t) }{\bar{y}}\) and \(t \rightarrow J_{\sigma (t)}^{\partial f}{\bar{v}} \) (given by \(J_{\sigma (t)}^{\partial f}{\bar{v}}:={\bar{v}}-\sigma (t) (\partial f)_{\sigma (t)}{\bar{v}}\)) are also continuous on \([0,\infty )\). So, in light of the definition of \(\phi _1(.,.)\) and \(\phi _2(.,.)\), together with the continuity of \(\{ \theta ( \cdot ), \omega ( \cdot ) \}\), we infer that \(F(.,(\bar{v},{\bar{y}}))\) is continuous on \([0,\infty )\) (as are \(\phi _1(.,.)\) and \(\phi _2(.,.)\)).
(b) Given \(t \ge 0\), we prove that F(t, .) is \(\iota (t)\)-Lipschitz continuous on \({\mathcal {H}}^2\), for some continuous mapping \(\iota : [0,\infty ) \rightarrow [0,\infty )\). Indeed, for \((v_i,y_i)\in {\mathcal {H}}^2\) (for \(i=1,2\)), while noticing that \(J_{\sigma (t)}^{\partial f}\) and \(\frac{1}{2}\sigma (t) (\partial f)_{\sigma (t)}\) are nonexpansive on \({\mathcal {H}}\), by (1.33) we get
while an easy computation gives us
It follows from the previous arguments that F(t, .) satisfies
hence F(t., .) is \( \iota (t)\)-Lipschitz continuous on \({\mathcal {H}}^2\) along with \( \iota ( \cdot )=\theta ( \cdot )+2 \omega ( \cdot )+ \kappa \) which is continuous (by the continuity of \(\theta ( \cdot )\) and \(\omega ( \cdot )\)).
Thus, for any given \((x_0,y_0,\xi _0)\in {\mathcal {H}}^3\), applying the global Cauchy–Lipschitz theorem yields existence and uniqueness of a global classical solution \((v( \cdot ),y( \cdot ))\) to (1.31) (namely, \(y( \cdot )\) and \(v( \cdot )\) are of class \(C^1\)) such that \(y(0)=y_0\) and \(v(0)=x_0+ \sigma (0) \xi _0\). Furthermore, the previous arguments (a) and (b) guarantee existence and uniqueness of a strong global solution \((v( \cdot ),y( \cdot ))\) to the same problem (1.31), by invoking the version of the Cauchy–Lipschitz theorem involving absolutely continuous trajectories, see for example [25, Proposition 6.2.1.], [37, Theorem 54].
(s3) Existence of a strong global solution \((x( \cdot ),\xi ( \cdot ),y( \cdot ))\) to (1.4). Let \((x_0,\xi _0,y_0) \in {\mathcal {H}}^3\) be such that \( \xi _0 \in \partial f (x_0)\). Given a global classical solution \((v( \cdot ),y( \cdot ))\) to (1.31) such that \(y(0)=y_0\) and \(v(0)=x_0+ \sigma (0) \xi _0\), we consider the functions \(x( \cdot )\) and \(\xi ( \cdot )\) defined by (1.30), and we show through the following items (s3-a)–(s3-b) that \((x( \cdot ),\xi ( \cdot ),y( \cdot ))\) is a strong global solution to (1.4):
(s3-a) Let us prove the absolute continuity of \(x( \cdot )\) and \(\xi ( \cdot )\) on any bounded subset of \([0, \infty )\). As \(v( \cdot )\) is of class \(C^1\) on \([0, \infty )\), we immediately see that \(v( \cdot )\) is absolutely continuous from the characterization (i1) of Definition 2.1. Hence, given \(\epsilon >0\) and finitely many intervals \(I_k=(a_k,b_k)\) such that \(I_k \cap I_j = \emptyset \) (for \(k \ne j\)), by using Definition 2.1-(i3) we know for some \(\eta >0\) that \(\sum _k |b_k-a_k| \le \eta \) implies that \(\sum _k \Vert v(b_k)-v(a_k)\Vert \le \min (\epsilon , \frac{2 \epsilon }{\sigma _0}).\) So, invoking the non-expansiveness of \(J_{\sigma ( \cdot )}^{ \partial f}\) and \( \frac{1}{2} \sigma ( \cdot ) (\partial f)_{\sigma ( \cdot )}\) while recalling that \(\sigma ( \cdot ) \ge \sigma _0 >0\) entails that
\(\sum _k \Vert J_{\sigma (t)}^{ \partial f} v(b_k)- J_{\sigma (t)}^{\partial f} v (a_k)\Vert \le \sum _k \Vert v(b_k)- v (a_k)\Vert \le \epsilon \)
and that
\(\sum _k \Vert (\partial f)_{\sigma (t)} v(b_k)-(\partial f)_{\sigma (t)} v(a_k)\Vert \le \sum _k \frac{2}{\sigma (t)}\Vert v(b_k)-v(a_k)\Vert \le \frac{2}{\sigma _0} \sum _k \Vert v(b_k)-v(a_k)\Vert \le \epsilon \).
Consequently, the mappings \(x( \cdot )= J_{\sigma ( \cdot )}^{\partial f }v( \cdot )\) and \(\xi ( \cdot ) = (\partial f)_{\sigma ( \cdot )} v( \cdot )\) also comply with characterization (i3) of Definition 2.1, which proves that \(x( \cdot )\) and \(\xi ( \cdot )\) are absolutely continuous on \([0, \infty )\).
(s3-b) Let us show that the triplet \((x( \cdot ),y( \cdot ),\xi ( \cdot ))\) satisfies system (2.3). Indeed, by \(x( \cdot )= J_{\sigma ( \cdot )}^{\partial f }v( \cdot )\) (from (1.30)), and \(\sigma ( \cdot ) \xi ( \cdot )=v( \cdot )-x( \cdot )\), we readily deduce that \(\sigma ( \cdot ) \xi ( \cdot ) \in \left( \sigma ( \cdot ) \partial f \right) (x( \cdot ))\) (because \(v( \cdot ) \in x( \cdot ) + {\sigma ( \cdot )}\partial f (x( \cdot ))\)), which by the positivity of \(\sigma ( \cdot )\) proves (2.3a). Moreover, in (1.31), substituting \(v( \cdot )\), \( J_{\sigma ( \cdot )}^{\partial f }v( \cdot )\) and \({\sigma ( \cdot )} (\partial f)_{\sigma ( \cdot )}v( \cdot )\), by \(x( \cdot ) + \sigma ( \cdot ) \xi ( \cdot )\), \(x( \cdot )\) and \(\sigma ( \cdot ) \xi ( \cdot )\), respectively, gives us immediately (2.3b) and (2.3c). In addition, regarding the initial conditions we obtain \(x(0)=J_{\sigma (0)}^{\partial f } v(0)=x_0\) (since \(v(0)=x_0+\sigma (0) \xi _0\) and \( \xi _0 \in \partial f (x_0)\)), \(y(0)=y_0\) and \(\sigma (0) \xi (0)=v(0)-x(0)=\sigma (0) \xi _0\).
Consequently, by items (s3-a)– (s3-b), we get the existence of a strong global solution to (1.4) \(\square \)
1.3 Proof of Lemma 3.2
(See [30, Lemma 4.1]). Given \(t\in [ 0,\infty )\) and \(h\in (0,\infty )\), we have \(\xi (t)\in \partial fx(t)\) and \(\xi (t+h)\in \partial f x(t+h)\) hence, by monotonicity of \(\partial f\), we simply have
Clearly, assuming that \(x( \cdot )\) and \(\xi ( \cdot )\) are absolutely continuous on \([0,\infty )\), yields that, for a.e. \(t \in [0,\infty )\), and as \(h \rightarrow 0^+\), we have
Thus, letting h tend to \(0^+\) in (1.35), implies \(\langle \dot{\xi }(t),\dot{x}(t)\rangle \ge 0\), that is the desired result \(\square \)
1.4 Proof of Proposition 4.3
By Remark 3.1, we know under condition (1.15) that \(\theta ( \cdot )\) is well-defined and positive on \([0,\infty )\). Moreover, by \(\nu ( \cdot ) \in C^2\) and \(\theta ( \cdot )=\frac{\kappa \nu ( \cdot ) -\dot{\nu }( \cdot )}{\nu ( \cdot )+ e_*}\) (hence \( \theta ( \cdot )=\kappa -\frac{ \dot{\nu }( \cdot )+\kappa e_*}{\nu ( \cdot ) + e_*}\)), we can see that \(\theta ( \cdot ) \in C^1([0,\infty ))\) and (omitting the variable t) we readily get
Let us recall (from (1.3)) that \(\alpha := \frac{\dot{\theta }}{\theta } + \kappa - \theta \). Consequently, by the previous arguments we obtain
\(\alpha := \frac{1}{\theta } (\theta ( \kappa - \theta ) - {\dot{\theta }}) = \frac{1}{\kappa \nu - {\dot{\nu }}} \left( (\dot{\nu } +\kappa e_*) \frac{\kappa \nu - 2 {\dot{\nu }}}{\nu + e_*} + \ddot{\nu }\right) \).
So, as \(t \rightarrow \infty \), by \(\nu (t) \rightarrow +\infty \), \(\dot{\nu }(t) \rightarrow l \in [0,\infty )\) and \(\ddot{\nu }(t) \rightarrow 0\) (from (4.40)), we immediately obtain that \(\alpha (t) \sim \frac{ l +\kappa e_*}{\nu (t) + e_*}\). We also recall (from (1.3)) that \(\beta :=-\frac{\dot{\theta }}{{\theta }}+ \kappa +\omega \), hence by the definition of \(\alpha \) we equivalently have \(\beta =\alpha + \theta + \omega \). Moreover, as \(t \rightarrow \infty \), by \(\omega :=\big (\kappa -\frac{\dot{\nu }}{\nu } \big )\vartheta -\frac{\delta }{\nu +e_*}\) (from (1.5c)), by \(\nu (t) \rightarrow \infty \), \({\dot{\nu }}(t) \rightarrow l\) and \(\vartheta (t) \rightarrow \vartheta _\infty \), we readily deduce that \(\omega (t) \rightarrow \kappa \vartheta _\infty \). Then, as \(t \rightarrow \infty \), by the latter formulation of \(\beta \), and remembering that (as \(t\rightarrow \infty \)) \(\theta (t) \rightarrow \kappa \), \(\alpha (t) \rightarrow 0\) and that \(\omega (t) \rightarrow \kappa \vartheta _\infty \), we get \(\beta (t) \rightarrow \kappa (1+\vartheta _\infty )\). Again from (1.3) we simply have
\(b(t):=\omega (t) \left( \kappa +\frac{{\dot{\omega }}(t)}{ \omega (t)}- \frac{\dot{\theta }(t)}{{\theta (t)}} \right) = \omega (t) \left( \kappa - \frac{\dot{\theta }(t)}{{\theta (t)}} \right) + {\dot{\omega }}(t)\).
In addition, we obviously see from its expression that \(\omega ( \cdot )\) is of class \(C^1\), and we have
It is also classically deduced from the convergence of \(\vartheta \) and the Lipschitz continuity of \({\dot{\vartheta }}\) that \({\dot{\vartheta (t)}} \rightarrow 0\) (as \(t \rightarrow \infty \)). This, in light of condition (4.40) and \(\lim _{t \rightarrow \infty } \vartheta (t)= \vartheta _\infty \) entails that \({\dot{\omega }}(t) \rightarrow \kappa \vartheta _\infty \) (as \(t \rightarrow \infty \)). Consequently, as \(t \rightarrow \infty \), by the previous formulation of b together with \(\frac{\dot{\theta }(t)}{{\theta }(t)} \rightarrow 0\), \(\omega (t) \rightarrow \kappa \vartheta _\infty \) and \({\dot{\omega }}(t) \rightarrow 0\), we deduce that \( b(t) \rightarrow \kappa ^2 \vartheta _\infty \)
\(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Maingé, PE., Weng-Law, A. Fast Continuous Dynamics Inside the Graph of Subdifferentials of Nonsmooth Convex Functions. Appl Math Optim 89, 1 (2024). https://doi.org/10.1007/s00245-023-10055-9
Accepted:
Published:
DOI: https://doi.org/10.1007/s00245-023-10055-9
Keywords
- Nonsmooth minimization
- Differential equations
- Dissipative dynamical systems
- Nonsmooth convex minimization
- Damped inertial dynamics
- Yosida approximation
- Coupled systems
- Nesterov acceleration