Abstract
This chapter investigates the behavior of cascaded dynamical systems through the lens of the Koopman operator and, in particular, its so-called principal eigenfunctions. It is shown that there exist perturbation functions for the initial conditions of each component system that make the orbits for the cascaded system and the decoupled component systems have zero asymptotic relative error. This in turn implies that the evolutions are asymptotically equivalent. By analyzing the exact form of the initial condition perturbation functions, the maximum error between the trajectories of the decoupled systems and the cascaded system can be bounded. More colloquially, these results say that cascaded compositions of stable systems are stable. It is also shown that the process of wiring the component systems together in a cascade structure preserves the principal eigenvalues and these principal eigenvalues are preserved between topologically conjugate systems. Thus, the analysis of cascaded systems is reduced to the determination of the principal eigenvalues and eigenfunctions of each component’s Koopman operator and the form of the perturbation functions.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
We use \(\circ t\) instead of t to remind ourselves that these are compositions of nonlinear operators.
- 2.
In (5.5) and later, we are using the notation that if we have a collection of maps \(f_i : X \rightarrow X_i\), \((i=1,\ldots ,n)\), the vector-valued map \(\mathbf {f} : X \rightarrow X_1\times \cdots \times X_n\) defined by \(\mathbf {f}(\mathbf {x}) := (f_1(\mathbf {x}),\ldots ,f_n(\mathbf {x}))\) can be written as \(\mathbf {f}(\mathbf {x}) \equiv (f_1,\ldots , f_n)(\mathbf {x})\).
- 3.
Note that \(\mathbf {w}_{i,s} := (\mathbf {e}_{i,s}^*\mathbf {V}_i^{-1})^{*} = (\mathbf {V}_{i}^{*})^{-1} \mathbf {e}_{i,s}\) is the sth dual basis vector in system i; that is \(\langle {\mathbf {v}_{i,t}},{\mathbf {w}_{i,s}}\rangle _{\mathbb {C}^{d_i}} = \mathbf {w}_{i,s}^* \mathbf {v}_{i,t} = \delta _{s,t}\), where \(\mathbf {v}_{i,t}\) is the tth eigenvector of \(\mathbf {L}_i\).
- 4.
We take the empty sum \(\sum _{j=1}^{0}\) to be 0.
- 5.
We choose seven layers here merely for display purposes—graphs for the behavior of the six downstream component systems are easily displayed in a 2-by-3 table.
References
Banaszuk, A., Fonoberov, V.A., Frewen, T.A., Kobilarov, M., Mathew, G., Mezić, I., Pinto, A., Sahai, T., Sane, H., Speranzon, A., Surana, A.: Scalable approach to uncertainty quantification and robust design of interconnected dynamical systems. Annu. Rev. Control 35(1), 77–98 (2011)
Budisic, M., Mohr, R., Mezić, I.: Applied koopmanism. Chaos 22(4), 047,510 (2012)
Callier, F., Chan, W., Desoer, C.: Input-output stability theory of interconnected systems using decomposition techniques. IEEE Trans. Circuits Syst. 23(12), 714–729 (1976)
Heersink, B., Warren, M.A., Hoffmann, H.: Dynamic mode decomposition for interconnected control systems (2017). arXiv.org
Lan, Y., Mezić, I.: On the architecture of cell regulation networks. BMC Syst. Biol. 5(1), 37 (2011)
Lasota, A., Mackey, M.C.: Chaos, Fractals, and Noise: Stochastic Aspects of Dynamics. Applied Mathematical Sciences, vol. 97, 2nd edn. Springer, Berlin (1994)
Mauroy, A., Mezić, I.: Global stability analysis using the eigenfunctions of the Koopman operator. IEEE Trans. Autom. Control 61(11), 3356–3369 (2016)
Mesbahi, A., Haeri, M.: Conditions on decomposing linear systems with more than one matrix to block triangular or diagonal form. IEEE Trans. Autom. Control 60(1), 233–239 (2015)
Mezić, I.: Coupled nonlinear dynamical systems: asymptotic behavior and uncertainty propagation. In: 43rd IEEE Conference on Decision and Control, pp. 1778–1783. Atlantis, Paradise Island, Bahamas (2004)
Mezić, I.: Spectral properties of dynamical systems, model reduction and decompositions. Nonlinear Dyn. 41, 309–325 (2005)
Mezić, I., Banaszuk, A.: Comparison of systems with complex behavior: spectral methods. In: 39th IEEE Conference on Decision and Control, pp. 1224–1231. UCSB, Sydney, Australia (2000)
Mezić, I., Banaszuk, A.: Comparison of systems with complex behavior. Phys. D: Nonlinear Phenom. 197(1), 101–133 (2004)
Michel, A.N.: On the status of stability of interconnected systems. IEEE Trans. Autom. Control 28(6), 639–653 (1983)
Mohr, R., Mezić, I.: Construction of eigenfunctions for scalar-type operators via laplace averages with connections to the Koopman operator, 1–25 (2014). arXiv.org
Mohr, R., Mezić, I.: Koopman principal eigenfunctions and linearization of diffeomorphisms (2016). arXiv.org
Pichai, V., Sezer, M.E., Siljak, D.D.: A graph-theoretic algorithm for hierarchical decomposition of dynamic-systems with applications to estimation and control. IEEE Trans. Syst. Man Cybern. 13(2), 197–207 (1983)
Ryan, R.A.: Introduction to Tensor Products of Banach Spaces. Springer Monographs in Mathematics. Springer, London (2002)
Shen-Orr, S.S., Milo, R., Mangan, S., Alon, U.: Network motifs in the transcriptional regulation network of Escherichia coli. Nature Genet. 31(1), 64–68 (2002)
Acknowledgements
This research was partially funded under a subcontract from HRL Laboratories, LLC under DARPA contract N66001-16-C-4053 and additionally funded by the DARPA Contract HR0011-16-C-0116.
The views expressed are those of the authors and do not reflect the official policy or position of the Department of Defense or the U.S. Government. Distribution Statement “A”: Approved for Public Release, Distribution Unlimited.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
5.6 Appendix
5.6 Appendix
Here, we collect all the technical proofs of the above results. It can be skipped on a first reading.
1.1 5.6.1 Proof of Theorem 5.2: 0 Asymptotic Relative Error—Linear, Chained Cascades
The first lemma gives the general solution for the ith level of the chained linear cascade system.
Lemma 5.1
For all \(i=1,\ldots , n\) and \(t \ge 0\), denote by \(\mathbf {x}_i(t)\) the solution \(\varPi _i \circ \mathsf {Lin}^{\circ t}(\mathbf {x}_1,\ldots ,\mathbf {x}_n)\) of the ith level of (5.31). For \(i\ge 2\), the general solution satisfies
Proof
Repeatedly using (5.31), we have
Replacing \(\mathbf {L}_{i}^{-k}\) with \(\mathbf {V}_{i} \varLambda _{i}^{-k} \mathbf {V}_{i}^{-1}\) in this final expression gives (5.55). \(\square \)
Lemma 5.2
Assume Condition 5.1 holds for (5.31) and each \(\mathbf {L}_i\) is diagonalized by \(\mathbf {L}_{i} = \mathbf {V}_{i} \varLambda _{i} \mathbf {V}_{i}^{-1}\). For any matrix \(\mathbf {B} \in \mathbb {C}^{d_i \times d_j}\), the following equality holds for any \(i,j \in \lbrace 1,\ldots , n \rbrace \) with \(i\ne j\):
where \(\tilde{\mathbf {B}} \in \mathbb {C}^{d_i \times d_j}\) is the matrix whose \((\ell ,m)\)th entry is given by
Proof
For any matrix \(\mathbf {M}\), we denote the \((\ell ,m)\)th entry as \([\mathbf {M}]_{\ell ,m}\). The \((\ell ,m)\)th entry of (5.56) is given by
Since \(\varLambda _i\) is diagonal, \([\varLambda _{j}^{k}]_{u,m} = 0\) for \(u\ne m\) and \([\varLambda _{j}^{k}]_{m,m} = \lambda _{j,m}^{k}\). This gives
Since \(\varLambda _{i}^{-k}\) is diagonal, we have
Summing from \(k=0,\ldots , t-1\), gives
Using (5.58), but with \(\mathbf {B}\) and k replaced by \(\tilde{\mathbf {B}}\) and t, respectively, we get
Therefore,
This is equivalent to (5.56). \(\square \)
Lemma 5.3
For each \(i=2,\ldots ,n\), the solution of (5.31) is
where
and the matrix \(\tilde{\mathbf {C}}_{i,j} \in \mathbb {C}^{d_{i} \times d_{j}}\) has elements
The perturbation functions \(\mathsf {pert}_{i} : \mathbb {C}^{d_1} \times \cdots \times \mathbb {C}^{d_i} \rightarrow \mathbb {C}^{d_i}\) are multilinear maps defined inductively by
Proof
We prove the result using induction. First note that the solution for \(\varPi _1\circ \mathsf {Lin}^{\circ t}(\mathbf {x}_1,\ldots ,\mathbf {x}_n)\) can be written as
Seed step: Consider \(\mathbf {x}_2(t) = \varPi _2 \circ \mathsf {Lin}^{\circ t}(\mathbf {x}_1,\ldots ,\mathbf {x}_n)\). By Lemma 5.1, Eq. (5.55), this is
where in the second line we have replaced \(\mathbf {x}_1(k)\) with (5.65) for \(t = k\). Using \(\mathbf {L}_{1}^{k} = \mathbf {V}_{1} \varLambda _{1}^{k}\mathbf {V}_{1}^{-1}\) in the second line gives
Lemma 5.2, (5.56), with \(\mathbf {B} \equiv \mathbf {V}_{2}^{-1} \mathbf {C}_{2,1} \mathbf {D}_{1,1} \mathbf {V}_{1}\) gives that
where the elements of \(\tilde{\mathbf {C}}_{2,1}\) are given as
Equation (5.69) is the same as (5.62) for \(i=2\). Using (5.68) in (5.67) gives
Since \(\mathbf {L}_{2}^{t-1} \mathbf {V}_{2}\varLambda _{2}^{-t} = \mathbf {L}_{2}^{-1}\mathbf {V}_{2}\) and \(\varLambda _{1}^{t}\mathbf {V}_{1}^{-1} = \mathbf {V}_{1}^{-1} \mathbf {L}_{1}^{t}\), we get
Defining \(\mathbf {D}_{2,1}\) as
and \(\mathsf {pert}_{2} : \mathbb {C}^{d_{1}}\times \mathbb {C}^{d_{2}} \rightarrow \mathbb {C}^{d_{2}}\) as
gives
since \(\mathbf {D}_{2,2} = \mathbf {I}_{d_2}\) by definition. Finally,
Using the change of variables \(j = 2 - s\), we have that
Equations (5.71)–(5.74) are equivalent to Eqs. (5.59), (5.61), and (5.64), for \(j=2\).
Induction step: Assume (5.59)–(5.64) hold for for all \(j \le i\) where \(i\in \lbrace 2,\ldots , n-1 \rbrace \). We show they hold for \(i+1\) as well.
Write \(\mathbf {x}_{i+1}(t) = \varPi _{i+1}\mathsf {Lin}^{\circ t}(\mathbf {x}_1,\ldots ,\mathbf {x}_n)\). By Lemma 5.1, Eq. (5.55), the solution is
By the induction hypothesis,
which gives that \(\mathbf {x}_{i+1}(t)\) is (after interchanging the finite sums)
Since \(\mathbf {L}_{j}\) is diagonalizable, we substitute \(\mathbf {V}_{j} \varLambda _{j}^{k} \mathbf {V}_{j}^{-1}\) for \(\mathbf {L}_{j}^{k}\) in the above equation to get
By Lemma 5.2, Eq. (5.56), with \(\mathbf {B} \equiv \mathbf {V}_{i+1}^{-1} \mathbf {C}_{i+1,i} \mathbf {D}_{i,j} \mathbf {V}_{j}\), we have
where for \(j \in \lbrace 1,\ldots , i \rbrace \) the matrix \(\tilde{\mathbf {C}}_{i+1,j} \in \mathbb {C}^{d_{i+1}\times d_{j}} \) has elements
Equation (5.78) is (5.62) for \(i+1\). Plugging (5.77) into (5.76) gives
Since \(\mathbf {L}_{i+1}^{t-1} \mathbf {V}_{i+1} \varLambda _{i+1}^{-t} = \mathbf {L}_{i+1}^{-1} \mathbf {V}_{i+1}\) and \(\varLambda _{j}^{t} \mathbf {V}_{j}^{-1} = \mathbf {V}_{j}^{-1} \mathbf {L}_{j}^{t}\), then
For \(j=1,\ldots , i\), define
as in Eq. (5.61) and \(\mathsf {pert}_{i+1} : \mathbb {C}^{d_1}\times \cdots \times \mathbb {C}^{d_{i+1}} \rightarrow \mathbb {C}^{d_{i+1}}\) as
as in Eq. (5.64) with the substitution \(i \mapsto i+1\). Substituting these definitions into the expression for the solution \(\mathbf {x}_{i+1}(t)\) and defining \(\mathbf {D}_{i+1,i+1} = \mathbf {I}_{d_{i+1}}\), we have
Comparing with (5.59) with the substitution \(i\mapsto i+1\), we see that the induction is complete. This completes the proof. \(\square \)
Corollary 5.3
Assume that Condition 5.1 holds for (5.31). Then for all \(i\in \lbrace 2,\ldots ,n \rbrace \) and \(t \in \mathbb {N}\),
where \(\mathbf {D}_{i,j}\) and \(\mathsf {pert}_{j}\) are given by (5.61) and (5.64). Furthermore,
Proof
Inequality (5.80) follows directly from Lemma 5.3, Eq. (5.59) and the fact that \(\mathbf {D}_{i,i} = \mathbf {I}_{d_{i}}\). Equation (5.81) follows from the condition 5.1, equation (5.34).
1.2 5.6.2 Proof of Theorem 5.3: Perturbation of Principal Eigenfunctions—Nominal, Linear System
We now prove Theorem 5.3. It is a straightforward application of Theorem 5.2.
Proof
(Proof of Theorem 5.3) We first show that for \(i \ge 1\) and \(t \ge 0\) that (5.45) holds.
By definition,
Therefore,
By Theorem 5.2, Eq. (5.36), for all \(t \ge 0\) and \(i\ge 1\),
This estimate along with (5.82) gives (5.45).
By Theorem 5.2, Eq. (5.43), for all \(i\in \lbrace 1,\ldots , n \rbrace \) and any \(\varepsilon > 0\),
for all t large enough. This is equivalent to (5.46). \(\square \)
The Generalized Laplace Analysis theorem uses Laplace averages of the Koopman operator to project a function onto an eigenspace. This is is the proof of Corollary 5.2. The following definition and theorem are taken from [14].
Definition 5.5
(Dominating point spectrum) For \(r > 0\), let \(\mathbb D_r\) be the open disc of radius r centered at 0 in the complex plane and let \(\sigma (U; \mathbb D_r) = \sigma (U) \cap \mathbb D_r\). If there exists an \(R > 0\) such that \(\sigma (U)\setminus \mathbb D_R\) is not empty and for every \(r > R\), we have
-
1.
if \(\sigma (U; \mathbb D_r) \cap \sigma _p(U) \ne \emptyset \), then the peripheral spectrum of \(\sigma (U; \mathbb D_r)\) is not empty, and
-
2.
the set \(\sigma (U) \setminus \mathbb D_r\) consists only of eigenvalues (i.e., \(\sigma (U) \setminus \mathbb D_r \subset \sigma _p(U)\)).
Theorem 5.6
Let \(\sigma (U)\) have a dominating point spectrum and assume that the point spectrum is concentrated on isolated circles in the complex plane. Let \(\lambda \) be an eigenvalue of U. The projection \(P_\lambda \) onto the \(N(\lambda I - U)\), the \(\lambda \)-eigenspace of U, can be computed as
where the limit exists in the strong operator topology and where \(\Omega = \lbrace \mu \in \sigma _p(U) : |\mu | > |\lambda | \rbrace \).
1.3 5.6.3 Proof of Theorem 5.4: Asymptotic Equivalence for Nonlinear, Chained Cascades
We now prove Theorem 5.4. It is a straightforward application of Theorem 5.2 and the fact that the topological conjugacy is a homeomorphism.
Proof
(Proof of Theorem 5.4) Fix \(\varepsilon > 0\) and \(\mathbf {Y} = (\mathbf {y}_1,\ldots , \mathbf {y}_n) \in \mathbb {C}^{d_1}\times \cdots \times \mathbb {C}^{d_n}\). Define \(\mathbf {X} = \tau ^{-1}(\mathbf {Y})\). Denote by \(\overline{B_{i}}\) the closed unit ball of radius centered at the origin in \(\mathbb {C}^{d_i}\). Condition 5.1 and Theorem 5.2, Eq. (5.43) guarantee that \(\mathsf {Lin}^{\circ t}(\mathbf {X})\) and \(\mathsf {Nom}^{\circ t}(\mathsf {pert}(\mathbf {X}))\) are in the compact set \(\overline{B_{1}} \times \cdots \times \overline{B_{n}}\) for all t large enough.
Since \(\tau \) is continuous, it is uniformly continuous on \(\overline{B_{1}} \times \cdots \times \overline{B_{r}}\). Let \(\delta > 0\) be such that if \(\mathbf {X}, \mathbf {X}' \in \overline{B_1} \times \cdots \times \overline{B_n}\) and \(\Vert \mathbf {X} - \mathbf {X}'\Vert _{\times } < \delta \), then \(\Vert \tau (\mathbf {X}) - \tau (\mathbf {X}')\Vert _{\times } < \varepsilon \).
By Corollary 5.1, there is a \(T \in \mathbb {N}\) such that \(t \ge T\) implies
The uniform continuity of \(\tau \) implies that
Now, since \(\tau \) is a topological conjugacy, \(\mathsf {Lin}^{\circ t} = \tau ^{-1}\circ \mathsf {NonLin}^{\circ t} \circ \tau \). Plugging this into (5.87) gives, for all \(t \ge T\),
Therefore,
This completes the proof. \(\square \)
1.4 5.6.4 Proof of Theorem 5.5: Perturbation of Principal Eigenfunctions—Nominal, Nonlinear Cascades
To save space in the following proof, we will write \(\phi _{(0,\ldots ,0,s_i,0,\ldots ,0)}\) as \(\phi _{s_i \mathbf {e}_{n,i}}\), where \(\mathbf {e}_{n,i}\) is the ith canonical basis vector of length n.
Proof
(Proof of Theorem 5.5) Fix \(\mathbf {Y} = (\mathbf {y}_1,\ldots , \mathbf {y}_n) \in \mathbb {C}^{d_1}\times \cdots \times \mathbb {C}^{d_n}\) and let \(\mathbf {X} = \tau ^{-1}(\mathbf {Y})\). The topological conjugacy satisfies
Using this relation, we get
On the other hand,
Combining these two expression, we have
Theorem 5.3, Eq. (5.46), implies
\(\square \)
Rights and permissions
Copyright information
© 2020 Springer Nature Switzerland AG
About this chapter
Cite this chapter
Mohr, R., Mezić, I. (2020). Koopman Spectrum and Stability of Cascaded Dynamical Systems. In: Mauroy, A., Mezić, I., Susuki, Y. (eds) The Koopman Operator in Systems and Control. Lecture Notes in Control and Information Sciences, vol 484. Springer, Cham. https://doi.org/10.1007/978-3-030-35713-9_5
Download citation
DOI: https://doi.org/10.1007/978-3-030-35713-9_5
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-35712-2
Online ISBN: 978-3-030-35713-9
eBook Packages: Intelligent Technologies and RoboticsIntelligent Technologies and Robotics (R0)