Abstract
A class of differential Riccati equations (DREs) is considered for which the evolution of any solution can be identified with the propagation of a value function of a corresponding optimal control problem arising in \({\mathscr {L}_2}\)-gain analysis. By exploiting the semigroup properties inherited from the attendant dynamic programming principle, a max-plus primal space fundamental solution semigroup of max-plus linear max-plus integral operators is developed that encapsulates all such value function propagations. Using this semigroup, a one-parameter fundamental solution semigroup of matrices is constructed for the aforementioned class of DREs. It is demonstrated that this semigroup can be used to compute particular solutions of these DREs, and to characterize finite escape times (should they exist) in a simple way.
Similar content being viewed by others
References
Anderson B, Moore J (1971) Linear optimal control. Prentice-Hall, Englewood Cliffs
Doyle J, Glover K, Khargonekar P, Francis B (1989) State space solutions to standard \({\fancyscript {H}_{2}}\) and \({\fancyscript {H}_\infty }\)-control problems. IEEE Trans Autom Control 34(8):831–847
Petersen I, Anderson B, Jonckheere E (1991) A first-principles solution to the non-singular \({\fancyscript {H}_\infty }\) control problem. Int J Robust & Nonlinear Control 1:171–185
Green M, Limebeer D (1995) Linear robust control, ser. Information and systems sciences. Prentice-Hall
Davison E, Maki M (1973) The numerical solution of the matrix Riccati differential equation. IEEE Trans Autom Control 18:71–73
Lawson J, Lim Y (2006) The symplectic semigroup and Riccati differential equations. J. Dyn Control Syst 12(1):49–77
McEneaney W (2008) A new fundamental solution for differential Riccati equations arising in control. Automatica 44:920–936
Dower P, McEneaney W (2015) A max-plus dual space fundamental solution for a class of operator differential Riccati equations. SIAM J. Control & Optim 53(2):969–1002
Zhang H, Dower P (2015) Max-plus fundamental solution semigroups for a class of difference Riccati equations. Automatica 52:103–110
Zhang H, Dower P (2015) A max-plus based fundamental solution for a class of discrete time linear regulator problems. Linear Algebra Appl 471:693–729
Fleming W, McEneaney W (2000) A max-plus-based algorithm for a Hamilton–Jacobi–Bellman equation of nonlinear filtering. SIAM J Control Optim 38(3):683–710
Kolokoltsov V, Maslov V (1997) Idempotent analysis and applications. Kluwer Publishing House, Dordrecht
Litvinov G, Maslov V, Shpiz G (2001) Idempotent functional analysis: an algebraic approach. Math Notes 69(5):696–729
Akian M, Gaubert S, Kolokoltsov VN (2005) Set coverings and invertibility of functional galois connections. In: Litvinov GL, Maslov VP (eds) Idempotent mathematics and mathematical physics, ser. contemporary mathematics. American Mathematical Society, Providence, pp 19–51
Cohen G, Gaubert S, Quadrat J-P (2004) Duality and separation theorems in idempotent semimodules. Linear Algebra & Appl 379:395–422
McEneaney W (2006) Max-plus methods for nonlinear control and estimation, ser. systems & control: foundations & applications. Birkhauser, Basel
Sasagawa T (1982) On the finite escape phenomena for matrix Riccati equations. IEEE Trans Autom Control 27:977–979
Kilicaslan S, Banks S (2012) Existence of solutions of Riccati differential equations. J Dyn Syst Meas Control 134:031001
Moreau J-J (1970) Inf-convolution, sous-additivité, convexité des fonctions numériques. J Math Pures Appl 9(49):109–154
Rockafellar R (1974) Conjugate duality and optimization. CBMS-NSF regional conference series in applied mathematics, vol. 16, SIAM, Philadelphia
Rockafellar R, Wets R (1997) Variational analysis. Springer, Berlin
Dower P, McEneaney W, Zhang H (2015) “Max-plus fundamental solution semigroups for optimal control problems,” In: Proceeding SIAM conference on control theory and its applications (Paris, France), pp. 368–375
Bensoussan A, Prato GD, Delfour M, Mitter S (2007) Representation and control of infinite dimensional systems, 2nd edn. Birkhaüser, Basel
Author information
Authors and Affiliations
Corresponding author
Additional information
Research partially supported by the Australian Research Council and AFOSR/AOARD Grants DP120101549 and FA2386-12-1-4084.
Appendix A: Proofs
Appendix A: Proofs
1.1 A.1 Proof of Theorem 1
Since \(Q_t\in {\mathbb {R}}^{2n\times 2n}\) satisfies DRE (22), (23) for all \(t\in {\mathbb {R}}_{\ge 0}\), see (28), it may be represented by a corresponding symplectic fundamental solution of the form (5), denoted here by \(\widehat{\varSigma }_t\in {\mathbb {R}}^{4n\times 4n}\). In order to apply (5), define \({\widehat{{\mathcal {H}}}},{\varDelta }\in {\mathbb {R}}^{4n\times 4n}\) by
where \({\hat{A}}\in {\mathbb {R}}^{2n\times 2n}\), \({\hat{B}}\in {\mathbb {R}}^{2n\times m}\), \(\hat{C}\in {\mathbb {R}}^{p\times 2n}\) are as per (24). Note by inspection that \({\varDelta }= {\varDelta }' = {\varDelta }^{-1}\). By substitution, a straightforward calculation yields that
where \({\mathcal {H}}\in {\mathbb {R}}^{2n\times 2n}\) is as per (5). Hence, the symplectic fundamental solution \(\widehat{\varSigma }_t\) for DRE (22), (23) is, again by (5),
for all \(t\in {\mathbb {R}}_{\ge 0}\), where the notation of (3) has been applied. Hence, the particular solution \(Q_t\) of DRE (22), (23) is given in terms of the symplectic fundamental solution (5), with respect to \(\widehat{\varSigma }_t\), by
for all \(t\in [0,t^*(Q_0))\equiv {\mathbb {R}}_{\ge 0}\), see (28), in which
and \(\mu (M)\) is defined by (10). For any fixed \(t\in {\mathbb {R}}_{\ge 0}\), note in particular that
in which \(({\varSigma }_t^{11} + {\varSigma }_t^{12} M)^{-1}\) is well defined as \(t^*(M)=+\infty \) by Assumption 1 and (6), (28). That is, \(\widehat{X}_t^{-1}\) is well defined. Its substitution in (62), along with \(\widehat{Y}_t\) from (63), yields \(Q_t = \widehat{Y}_t \widehat{X}_t^{-1} \doteq {\varXi }({\varSigma }_t)\), where \({\varXi }:{\mathbb {R}}^{2n\times 2n}\rightarrow {\mathbb {R}}^{2n\times 2n}\) is defined by
using the notation of (3), with
As M is invertible by Assumption 1, it may be verified directly that \({\varXi }\) of (64) is invertible, with \({\varXi }^{-1}:{\mathbb {R}}^{2n\times 2n}\rightarrow {\mathbb {R}}^{2n\times 2n}\) given by
where
That is, (29) holds. \(\square \)
1.2 A.2 Proof of Lemma 1
Fix \(M_0\in {\mathbb {S}}_{\ge 0}^{n\times n}\) as the stabilizing solution of ARE (30) indicated in the lemma statement. Let \(t^*(M_0)\in {\mathbb {R}}_{>0}^+\) denote the maximal horizon of existence (6) of the DRE
As \(M_0\) is the stabilizing solution of ARE (30), note that \(R_t \doteq M_0\) is the unique solution of this DRE for all \(t\in {\mathbb {R}}_{\ge 0}\). Consequently, \(t^*(M_0) = +\infty \). Choose any invertible \(M\in {\mathbb {S}}^{n\times n}\) such that (31) holds, and note that such a choice is always possible. Recalling (25), let \(Q_t^{11}\in {\mathbb {S}}^{n\times n}\), \(t\in [0,t^*(M))\) denote the unique solution of DRE (25) initialized with \(Q_0^{11} = M\). As DREs (25) and (66) are identical, Proposition 1 (see below) and (31) imply that solutions \(Q_t^{11}\) and \(R_t\) satisfy the monotonicity property
for all \(t\in [0,t^*(M))\). By inspection, this provides an upper bound for \(Q_t^{11}\). In order to determine a lower bound, choose \(w_s = 0\) for all \(s\in [0,t]\) suboptimal in the definition (20) of \(S_t(x,0)\). Recalling (16), (17), (21),
for all \(x\in {\mathbb {R}}^n\), in which \(O_t\in {\mathbb {S}}^{n\times n}\) is well defined by
for all \(t\in {\mathbb {R}}_{\ge 0}\). Note that \(O_t\in {\mathbb {S}}^{n\times n}\) is finite for all \(t\in {\mathbb {R}}_{\ge 0}\), and provides a lower bound for \(Q_t^{11}\in {\mathbb {S}}^{n\times n}\). Hence, combining (67) and (68),
for all \(t\in [0,t^*(M))\). A simple contradiction argument subsequently implies that \(Q_t^{11}\in {\mathbb {S}}^{n\times n}\) is finite for all \(t\in {\mathbb {R}}_{\ge 0}\), so that \(t^*(M) = +\infty \). As \(M^{-1}\) exists by definition, it follows that Assumption 1 holds, as required. \(\square \)
Proposition 1
Given initializations \(P_0, {\widetilde{P}}_0\in {\mathbb {S}}^{n\times n}\) satisfying \(P_0 - {\widetilde{P}}_0\in {\mathbb {S}}_{\le 0}^{n\times n}\), and \(t^*\doteq \min (t^*(P_0), t^*({\widetilde{P}}_0))\), the respective unique solutions \(P_t, {\widetilde{P}}_t\in {\mathbb {S}}^{2n\times 2n}\) of DRE (1) defined for all \(t\in [0,t^*)\), satisfy
for all \(t\in [0,t^*)\).
Proof
Fix \(t\in [0,t^*)\). Let \({\mathcal {T}}:{\varDelta }_{0,t}\rightarrow {\mathbb {R}}^{n\times n}\) denote the evolution operator associated with the time-dependent ordinary differential equation (ODE)
defined for \(s\in [0,t]\), with \({\varDelta }_{0,t} \doteq \{ (r,s)\in {\mathbb {R}}_{\ge 0}^2 \, \bigl | \, 0\le r \le s \le t \}\). By definition, see, for example, [23, Proposition 3.6, p.138],
for all \((s,\sigma )\in {\varDelta }_{0,t}\). Define \(\pi :[0,t]\rightarrow {\mathbb {S}}^{n\times n}\) by
for all \(\sigma \in [0,t]\). Differentiating with respect to \(\sigma \),
for all \(\sigma \in [0,t]\), where
in which the equality with zero follows by virtue of the fact that \(P_\sigma \), \({\widetilde{P}}_\sigma \) both satisfy the DRE (1). Consequently, (72) implies that \(\dot{\pi }_\sigma = 0\) for all \(\sigma \in [0,t]\), so that integration with respect to \(\sigma \in [0,t]\) yields \(\pi _t = \pi _0\). Recalling (71), it follows immediately that
Recalling that \(P_0 - {\widetilde{P}}_0\in {\mathbb {S}}_{\le 0}^{n\times n}\), and noting that \(t\in [0,t^*)\) is arbitrary, yields the required assertion (69). \(\square \)
1.3 A.3 Proof of Lemma 2
Suppose that Assumption 1 holds. Fix \(x,y\in {\mathbb {R}}^n\), \(t\in {\mathbb {R}}_{>0}\). Note that \(G_t(x,y)\in {\mathbb {R}}^-\) by Theorem 2.
(Necessity) Suppose that \(G_t(x,y)\in {\mathbb {R}}\). Recalling the value function interpretation of \(G_t(x,y)\), if the dynamics (18) are not controllable from x to y in time t, it immediately follows by definitions (16), (36) that \(G_t(x,y) = -\infty \). Hence, the dynamics (18) must be controllable from x to y in time t. Necessity follows as \(x,y\in {\mathbb {R}}^n\) and \(t\in {\mathbb {R}}_{>0}\) are arbitrary.
(Sufficiency) Suppose that dynamics (18) are controllable. Proposition 2 (see below) implies that \(Q_t^{22} \in {\mathbb {S}}_{>M}^{n\times n}\), where \(Q_t^{22}\) is as per (27). Hence, \(S_t(x,\cdot )\in {\mathscr {S}_+^{{-M}}} = {\mathsf {dom}\, }({\mathcal {D}}_\varphi )\), so that \({\mathcal {D}}_\varphi S_t(x,\cdot )\in {\mathscr {S}_-^{{-M}}}\) is well defined. So, applying the semiconvex transform (11) to \(S_t(x,\cdot )\) yields
in which \((M-Q_t^{22})^{-1}\) is guaranteed to exist, so that \({\varLambda }_t\in {\mathbb {R}}^{2n\times 2n}\) exists and is finite by definition. Hence, applying the right-hand equality of (36) in Theorem 2,
which is finite, thereby demonstrating sufficiency. \(\square \)
Proposition 2
Under Assumption 1, controllability of the dynamics (18) implies that \(Q_t^{22} \in {\mathbb {S}}_{>M}^{n\times n}\) for all \(t\in {\mathbb {R}}_{>0}\).
Proof
With \(M\in {\mathbb {S}}^{n\times n}\) satisfying Assumption 1, recall that \(t^*(M) = +\infty \) as per (28). Fix any \(t\in {\mathbb {R}}_{>0}\). Consequently, the optimal dynamics associated with \(S_t(x,y)\) of (20), (21) are well defined by the time-dependent ODE
for all \(s\in [0,t]\). Let \({\mathcal {V}}_t:{\varDelta }_{0,t}\rightarrow {\mathbb {R}}^{n\times n}\) denote the evolution operator associated with (75), with \({\varDelta }_{0,t} \doteq \{ (r,s)\in {\mathbb {R}}_{\ge 0}^2 \, \bigl | \, 0\le r \le s \le t \}\). By definition, see, for example, [23, Proposition 3.6, p.138],
for all \((s,\sigma )\in {\varDelta }_{0,t}\). Define \({\mathcal {U}}_t:{\varDelta }_{0,t}\rightarrow {\mathbb {R}}^{n\times n}\) via (76) by
for all \((r,\tau )\in {\varDelta }_{0,t}\). By inspection of (76), (77),
That is, \({\mathcal {U}}_t:{\varDelta }_{0,t}\rightarrow {\mathbb {R}}^{n\times n}\) is the evolution operator for the dynamics associated with \((A + B B' Q_{s}^{11})'\), \(s\in [0,t]\). Comparing with (26), it immediately follows that \(Q_s^{12} = - {\mathcal {U}}_t(s,0) M\) for all \(s\in [0,t]\). Hence, (27) implies that
where \({\mathcal {C}}_t \doteq \int _0^t {\mathcal {V}}_t(t,t-s) B B' \, {{\mathcal {V}}_t(t,t-s)'}\, ds\in {\mathbb {S}}_{\ge 0}^{n\times n}\) is the controllability gramian for the pair \((A+B B' Q_{t-\cdot }^{11}, B)\) on [0, t], by definition of \({\mathcal {V}}_t\). However, recall that controllability is preserved under state feedback, see, for example, [1, p.48]. Hence, (A, B) completely controllable implies that \((A+B B' Q_{t-\cdot }^{11}, B)\) is completely controllable, which in turn implies that \({\mathcal {C}}_t\) is invertible for \(t\in {\mathbb {R}}_{>0}\). That is, \({\mathcal {C}}_t\in {\mathbb {S}}_{>0}^{n\times n}\) for all \(t\in {\mathbb {R}}_{>0}\). As M is invertible by Assumption 1, the assertion immediately follows by (79). \(\square \)
1.4 A.4 Proof of Theorem 3
Fix any \(t\in {\mathbb {R}}_{>0}\), \(x\in {\mathbb {R}}^n\). Applying Lemma 2, and in particular (73), (74), it follows that \(Q_t\in {\mathbb {R}}^{2n\times 2n}\), \({\varLambda }_t\in {\mathbb {S}}^{2n\times 2n}\) of (22), (73) are related via
with matrix operators \({\varPi }, {\varPi }^{-1}:{\mathbb {S}}^{2n\times 2n}\rightarrow {\mathbb {S}}^{2n\times 2n}\) defined using the notation of (3) by
In particular, (81) follows by inspection of (73), while (80) may be verified by demonstrating that \({\varPi }\circ {\varPi }^{-1}\) is the identity. \(\square \)
1.5 A.5 Proof of Theorem 4
Throughout, it is assumed that Assumptions 1 and 2 hold, with \(M\in {\mathbb {S}}^{n\times n}\) specified by the former, as per the theorem statement. Note in particular that \(t^*(M) = +\infty \), so that \(({\varSigma }_t^{11} + {\varSigma }_t^{12}\, M)^{-1}\) exists for all \(t\in {\mathbb {R}}_{\ge 0}\), where \({\varSigma }_t\) is the symplectic fundamental solution identified in (5). Consequently, \(Q_t\in {\mathbb {S}}^{2n\times 2n}\) is well defined as the unique solution of DRE (22), (23), for all \(t\in {\mathbb {R}}_{\ge 0}\) by Assumption 1, see Theorem 1 and its proof. Note that \(P_0\in {\mathbb {S}}_{>M}^{n\times n} = {\mathsf {dom}\, }({\varUpsilon })\) by hypothesis and (13).
The proof proceeds by demonstrating a sequence of implications concerning the following claims, posed with respect to arbitrary fixed \(t\in {\mathbb {R}}_{>0}\) and \(P_0\in {\mathbb {S}}_{>M}^{n\times n}\):
- 1):
-
\(t\in (0,t^*(P_0))\);
- 2):
-
\({\varUpsilon }(P_0) + Q_s^{22}\in {\mathbb {S}}_{<0}^{n\times n}\) for all \(s\in (0,t]\);
- 3):
-
\({\varUpsilon }(P_0) + Q_t^{22}\in {\mathbb {S}}_{<0}^{n\times n}\);
- 4):
-
\(P_0 + {\varLambda }_t^{22}\in {\mathbb {S}}_{<0}^{n\times n}\); and
- 5):
In particular, it is shown that (1) \(\Leftrightarrow \) (2) \(\Leftrightarrow \) (3) \(\Leftrightarrow \) (4) \(\Rightarrow \) (5).
2) \(\Rightarrow \) 1): Suppose that \({\varUpsilon }(P_0) + Q_s^{22} \in {\mathbb {S}}_{<0}^{n\times n}\) for all \(s\in (0,t]\). Applying (13) and Theorem 1,
where it may be noted that the inverses on the right-hand side are guaranteed to exist. By hypothesis, the left-hand side is invertible, so that a matrix \(K_s\in {\mathbb {R}}^{n\times n}\) is well defined for an arbitrary \(s\in (0,t]\) by
However, the Woodbury lemma implies that
That is, \({{\varSigma }_s^{11}} + {\varSigma }_s^{12} P_0\in {\mathbb {S}}^{n\times n}\) is invertible. Recalling (6), and that \(s\in (0,t]\) is arbitrary, immediately implies that 1) holds.
1) \(\Rightarrow \) 2): Fix an arbitrary \(t\in (0,t^*(P_0))\). Analogously to the proof of Theorem 1, let \({\widetilde{Q}}_s\in {\mathbb {S}}^{2n\times 2n}\) denote the unique solution of DRE (22) subject to the initialization
defined, via (10), for all \(s\in [0, t^*({\widetilde{Q}}_0))\), where \(t^*({\widetilde{Q}}_0)\in {\mathbb {R}}_{>0}\) is the corresponding maximal horizon of existence (6). Analogously to the argument yielding (28), observe that \(t^*(\widetilde{Q}_0) = t^*(P_0)\), so that \(t\in (0,t^*({\widetilde{Q}}_0))\). An application of the symplectic fundamental solution (4), (5), (61) yields
for all \(s\in [0,t]\), in which
for all \(s\in [0,t]\). In particular,
in which \(({\varSigma }_s^{11} + {\varSigma }_s^{12} P_0)^{-1}\) is well defined for all \(s\in [0,t]\), as \(t\in (0,t^*(P_0))\), see (6). Consequently, recalling (84),
is well defined for all \(s\in [0,t]\). Recalling (23), (83), as \({\widetilde{Q}}_0 = \mu (P_0) \ge \mu (M) = Q_0\), Proposition 1 applied to (22) implies that \({\widetilde{Q}}_s - Q_s\in {\mathbb {S}}_{\ge 0}^{2n\times 2n}\), so that in particular
for all \(s\in [0,t]\). Fix any \(s\in (0,t]\). Rearranging (85) and applying (86) and Proposition 2,
Theorem 1 and (64) imply via the notation of (3) that
Recall that \({\varSigma }_s\in {\mathsf {dom}\, }({\varXi })\) (i.e., the inverse involved is guaranteed to exist) by Assumption 1, as \(s\in (0,t^*(M)) \equiv {\mathbb {R}}_{>0}\). Furthermore, as \(s\in (0,t^*(P_0))\), definition (6) implies that \({\varSigma }_s^{11} + {\varSigma }_s^{12} P_0\) is invertible. Hence, a matrix \(L_s\in {\mathbb {S}}^{n\times n}\) is well defined by
where the second equality follows by adding and subtracting \({\varSigma }_s^{12} M\) within the inverse. Applying (87), and the fact that \(P_0\in {\mathbb {S}}_{>M}^{n\times n}\), note that \(L_s\in {\mathbb {S}}_{<0}^{n\times n}\) by definition. The Woodbury lemma subsequently implies that
where the second equality follows as per (82). Consequently, as \(M\in {\mathbb {S}}^{n\times n}\) is invertible and \(L_s\in {\mathbb {S}}_{<0}^{n\times n}\),
As \(s\in (0,t]\) is arbitrary, claim 2) immediately follows.
2) \(\Rightarrow \) 3): By hypothesis, \({\varUpsilon }(P_0) + Q_s^{22}\in {\mathbb {S}}_{<0}^{n\times n}\) for all \(s\in (0,t]\). Selecting \(s=t\) yields claim 3) as required.
3) \(\Rightarrow \) 2): By hypothesis, \({\varUpsilon }(P_0) + Q_t^{22}\in {\mathbb {S}}_{<0}^{n\times n}\). Furthermore, \({\varUpsilon }(P_0)\in {\mathbb {S}}_{<-M}^{n\times n}\) by (13), (14). Hence, \(Q_t^{22}\in {\mathbb {S}}^{n\times n}\), so that \((Q_\sigma ^{12})' B B' Q_\sigma ^{12}\) must be integrable with respect to \(\sigma \in [0,t]\) by definition (27). In particular,
for any fixed \(s\in (0,t]\). Hence, \(Q_s^{22} - Q_t^{22}\in {\mathbb {S}}_{\le 0}^{n\times n}\), so that
Recalling that \(s\in (0,t]\) is arbitrary yields claim 2) as required.
3) \(\Rightarrow \) 4): Recalling (13) and Theorem 3, see (42), (80),
Recalling that \({\varUpsilon }(P_0) + Q_t^{22}\in {\mathbb {S}}_{<0}^{n\times n}\) by hypothesis,
That is, claim 4) holds.
4) \(\Rightarrow \) 3): Note that (89) holds as per the 3) \(\Rightarrow \) 4) case above. By hypothesis, \(P_0 + {\varLambda }_t^{22}\in {\mathbb {S}}_{<0}^{n\times n}\). Hence, the string of equivalences (90) implies that 3) holds.
4) \(\Rightarrow \) 5): Recalling (15), (35) and (42), the value function \(W_t\) of (15), (19) satisfies
for all \(x\in {\mathbb {R}}^n\). By hypothesis, \(P_0 + {\varLambda }_t^{22}\in {\mathbb {S}}_{<0}^{n\times n}\), so that \((P_0 + {\varLambda }_t^{22})^{-1}\) exists. Hence, the above max-plus integration explicitly evaluates as
As \(x\in {\mathbb {R}}^n\) is arbitrary, (44) follows immediately. In addition, as 4) \(\Leftrightarrow \) 1), it immediately follows that
That is, (45) holds. \(\square \)
Rights and permissions
About this article
Cite this article
Dower, P.M., Zhang, H. A max-plus primal space fundamental solution for a class of differential Riccati equations. Math. Control Signals Syst. 29, 15 (2017). https://doi.org/10.1007/s00498-017-0200-2
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00498-017-0200-2