Preprints of the 16th IFAC Symposium on System Identification The International Federation of Automatic Control Brussels, Belgium. July 11-13, 2012 Minimax state estimation for linear stationary differential-algebraic equations ? Sergiy Zhuk ∗ ∗ IBM Research, Dublin, Ireland (Tel: 35 31 826 94 51; e-mail: [email protected]). Abstract: This paper presents a generalization of the minimax state estimation approach for singular linear Differential-Algebraic Equations (DAE) with uncertain but bounded input and observation’s noise. We apply generalized Kalman Duality principle to DAE in order to represent the minimax estimate as a solution of a dual control problem for adjoint DAE. The latter is then solved converting the adjoint DAE into ODE by means of a projection algorithm. Finally, we represent the minimax estimate in the form of a linear recursive filter. Keywords: Minimax; Robust estimation; Descriptor systems; Optimization under uncertainties. 1. INTRODUCTION This paper presents a generalization of the minimax state estimation approach to linear Differential-Algebraic Equation (DAE) in the form d(F x) = Ax(t) + f (t), F x(t0 ) = x0 , (1) dt where F, A ∈ Rm×n . State estimation theory is one of the main tools in mathematical modeling. It is used, in particular, for parameter identification in oceanography [Heitz et al., 2010] or operational forecasts in meteorology and air pollution [Wu et al., 2008]. Mathematically, geophysical models are represented by systems of Partial Differential Equations (PDE) that make state estimation for those models very expensive from the computational point of view. In practice, different model reduction techniques are applied in order to get a reasonable computational time. For instance, the classical Galerkin projection method allows one to project the state of an infinite dimensional state equation, described by a PDE, onto a finite dimensional manifold and focus on the dynamics of the projection coefficients only. In the closed form this dynamics may be represented by the DAE (1): the main idea is to model the impact of the unresolved part of the state vector onto projection coefficients introducing the model error f and restricting it to belong to a certain subspace [Mallet and Zhuk, 2011]. Apart from the model reduction DAEs are applied in robotics [Schiehlen, 2000]. Pros and cons of using DAEs for modeling were discussed by Müller [2000]. The common way of deriving a state estimate for DAE is to 1) convert the DAE into an Ordinary Differential Equation (ODE) transforming the matrix pencil F − λA to Weierstrass canonical form [Gantmacher, 1960] and 2) apply the classical Kalman Duality (KD) principle to the resulting ODE. For the detailed description of 1), we refer the reader to [Darouach et al., 1997], [Gerdin ? This work was carried out during the author’s tenure of an ERCIM ”Alain Bensoussan” Fellowship Programme. This Programme is supported by the Marie Curie Co-funding of Regional, National and International Programmes (COFUND) of the European Commission. © IFAC, 2012. All rights reserved. et al., 2007] and citations there. We note that applying the matrix pencils theory to DAEs one restricts coefficients F and A (det (F − λA) 6= 0 for some λ) and might need to differentiate f . Although smoothness of f may be appropriate for control problems, it turns out to be a very restrictive assumption in the context of state estimation where f is the model error which is often represented by a measurable function with bounded L2 -norm. In what follows we derive a minimax state estimate for DAE (1) without these limitations. The main contribution of this paper is an optimal recursive state estimation algorithm for stationary DAEs in the form (1). The algorithm is applied to DAE directly that allows to avoid the restrictions imposed by the matrix pencils theory. To achieve this we transform the state estimation problem for DAE (1) to the dual control problem by means of generalized Kalman Duality (KD) principle proposed by Zhuk [2012] and the latter is then solved using a projection method. For the case of ellipsoidal G (see formula (7)) generalized KD states that the optimal Rt estimate û(y) = t01 ûT ydt of the linear function `T F x(t1 ) may be found as a solution of quadratic control problem for adjoint DAE: d(F 0 z) = −A0 z(t) + H 0 u(t), F 0 z(t1 ) = F 0 ` . (2) dt If F = I then this statement reduces to the classical KD principle (see [Åström, 2006]). In the general case, the set of all ` ∈ Rm such that (2) has a solution — this set will be referred to as a minimax observable subspace L (t1 ) — describes all possible directions in the state space of (1) that have the following property: if ` ∈ L (t1 ) then the worst-case estimation error (see formula (9)) is finite. In other words, the estimate û(y) with finite worst-case estimation error may be constructed for the projection of F x(t1 ) on L (t1 ) only. It is worth noting that for all ` ∈ L (t1 ) the estimate of `F x(t1 ) corresponds, in fact, to the solution of (1) which has the smallest eucledian norm and is, therefore, unique (see formula (5) and the following discussion). Therefore, there is no need to restrict F and A in order to force unique solvability for (1). Instead, it 143 16th IFAC Symposium on System Identification Brussels, Belgium. July 11-13, 2012 is sufficient to take ` from L (t1 ) and so constraints are moved from F, A onto `. We present a projection algorithm allowing one to convert (2) into equivalent ODE. It efficiently computes the minimax observable subspace and determines algebraic constraints for u imposed by the structure of (2). The algorithm splits DAE (2) into differential and algebraic parts (see (13)-(14)) and constructs an ODE (see (22)) such that all its solutions belong to the linear manifold defined by the algebraic part of DAE. This ODE is obtained applying a finite number of linear transformations (see (15)- (18)) to the differential part of DAE. We stress that in the general case the number of required transformations would be infinite if the DAE was non-stationary. An alternative way of constructing the set of all solutions for (2) is based on the theory of singular matrix pencils. We stress that this approach constrains u and its derivatives as well (see [Gantmacher, 1960, p.336,form.(79)]) that, in turn, complicates the procedure of computing the optimal control. Transforming the cost function of the dual control problem we find an optimal estimate û using Pontryagin maximum principle. We note that necessary optimality conditions for linear quadratic control problem with stationary regular DAE as a constraint were derived in Bender and Laub [1987]. The general case of a singular non-stationary DAE was considered by Zhuk [2012] where Tikhonov regularization is used to derive optimality conditions. Although the approach of this paper is applicable to the class of DAEs with constant coefficients only, it gives exact necessary and sufficient optimality conditions in contrast to the general approach presented in [Zhuk, 2012]. Finally, we derive the minimax estimate in the form of the linear recursive filter. We refer the reader to [Milanese and Tempo, 1985], [Chernousko, 1994] and [Kurzhanski and Vályi, 1997] for the basic information on the minimax framework. Minimax estimates for linear singular DAEs with discrete time were obtained by Zhuk [2010]. Let us note that our approach is applicable for the case of unbounded input f : indeed, one can always introduce an extended state x̃ = (x, f )0 , a new matrix F1 = F 0 and add a dummy bounded input f1 in (1). In this regard let us mention H∞ framework [Başar and Bernhard, 1995] which allows one to deal with unbounded f . Connections between minimax and H∞ frameworks were revealed in [Baras and Kurzhanski, 1995]. This paper is organized as follows. At the beginning of section 2 we describe the formal problem statement and definitions, and discuss ill-posedness of DAE. Then we present the dual control problem (Proposition 2) and projection algorithm (Proposition 3). At the end of the section the state estimation algorithm is derived (Theorem 4). In section 3 we consider a simple ODE with unbounded inputs and describe its minimax observable subspace and derive the minimax estimate. Section 4 contains conclusions. All proofs and technical statements are located in Appendix. Notation: Rn denotes the n-dimensional Euclidean space; L2 (t0 , t1 , Rm ) denotes a space of square-integrable functions with values in Rm (in what follows we will often write L2 (t0 , t1 ) referring L2 (t0 , t1 , Rk ) where the dimension k will be defined by the context); H1 (t0 , t1 , Rm ) denotes a space of absolutely continuous functions with L2 (t0 , t1 )- derivative; the prime 0 denotes the operation of taking the adjoint: L0 denotes adjoint operator, F 0 denotes the transposed matrix; R(L), N (L) denote the range and null-space of the operator L, PR (F )(PN (F )) denotes the orthogonal projection matrix onto the range (null-space) of the matrix F ; c(G, ·) denotes the support function of a set G; δ(G, ·) denotes the indicator of G: δ(G, f ) = 0 if f ∈ G and +∞ otherwise; xT y denotes the inner product of vectors x, y ∈ Rn , kxk2 := xT x; S > 0 means xT Sx > 0 for all x ∈ Rn ; F + denotes the pseudoinverse matrix; In denotes n × n-identity matrix, 0n×m denotes n × m-zero matrix, I0 := 0; 2. LINEAR MINIMAX ESTIMATION FOR DAES Consider a pair of systems d(F x) = Ax(t) + f (t) , dt y(t) = Hx(t) + η(t) , m×n F x(t0 ) = x0 , (3) (4) p×n where F, A ∈ R ,H∈R , t0 , t1 ∈ R and x(t) ∈ Rn , f (t) ∈ Rm , y(t) ∈ Rp , η(t) ∈ Rp represent the state, input, observation and observation’s noise respectively. In what follows we assume that initial condition x0 , input f and observation’s noise η are unknown and belong to the given bounding set G . Our aim is to estimate the state F x(t1 ) of (3) given observations y(t), t ∈ (t0 , t1 ). Let us briefly describe the main points of our strategy. In the case F = I the classical state estimation procedure (see [Åström, 2006]) suggests to look for an estimate of a linear function `(x) := `T F x(t1 ) in the class of linear Rt functions u(y) := t01 uT (t)y(t)dt, provided x solves (3) and the output y is given in the form (4). One computes an estimate û which is optimal with respect to some criteria σ̃. Assuming that G is bounded, σ̃ may be defined (see [Nakonechny, 1978] for details) as a worst-case estimation error σ̃(t1 , `, u) := sup (`(x) − u(y))2 (x0 ,f,η)∈G where x is a unique solution of ODE (3) corresponding to data x0 , f . In this case, the optimal estimate û solves the equation σ̃(t1 , `, û) = inf u σ̃(`, u). In general case F ∈ Rm×n (3) is solvable not for all x0 and f , and so one needs to assume that there is at least one x0 , f, η in G such that (3) has a solution. Further, if (3) is solvable for a given x0 , f , it may have non-unique solution so σ̃(t1 , `, u) is not well-defined. In fact, any solution of (3) corresponding to the data x0 , f admits the following representation: x = x0 + x1 , x0 ∈ X0 , x1 ⊥ X0 , (5) where X0 contains all x such that x solves (3) with x0 = 0 and f = 0. We stress that the data x0 , f determines x1 uniquely and so we may define the worst-case error as follows: σ(t1 , `, u) := sup sup (`(x) − u(y))2 , x∈X0 x0 ,f,η where an additional sup is taken over the set X0 . This allows one to eliminate the part of the solution x belonging to X0 . Notice that, in contrast with ODEs, σ(t1 , `, u) may be infinite for some `, u as X0 is a linear subspace. This observation suggests to introduce a so called minimax observable subspace L (t1 ). 144 16th IFAC Symposium on System Identification Brussels, Belgium. July 11-13, 2012 Definition 1. Given t1 < +∞, u ∈ L2 (t0 , t1 , Rp ) and ` ∈ Rm define a worst-case estimation error Z t1 σ(t1 , `, u) := sup (`T F x(t1 ) − uT (t)y(t)dt)2 x∈X0 ,(x0 ,f,η)∈G t0 Rt A function û(y) = t01 ûT (t)y(t)dt is called a minimax estimate in the direction ` (`-estimate) if inf u σ(t1 , `, u) = σ(t1 , `, û). The number σ̂(t1 , `) := σ(t1 , `, û) is called a minimax error in the direction ` at time-instant t1 (`error). The set L (t1 ) := {` ∈ Rm : σ̂(t1 , `) < +∞} is called a minimax observable subspace. We impose the following uncertainty description. Let Q0 , Q(t) ∈ Rm×m , Q0 = Q00 > 0, Q = Q0 > 0, R(t) ∈ Rp×p , R0 = R > 0; Q(t), R(t), R−1 (t) and Q−1 (t) are continuous functions of t on [t0 , t1 ]. Define Z t1 E(x0 , f, η) := xT0 Q0 x0 + f T Q(t)f + η T R(t)ηdt (6) t0 and let G = {(x0 , f, η) : E(x0 , f, η) ≤ 1} (7) It is easy to see that there is at least one (x0 , f, η) ∈ G such that (3) has a solution x. A geometrical representation of the `-estimate and error. Define a linear function Fu (x, η) := `T F x(t1 ) − R t1 T u (t)y(t)dt and operator t0 d(F x) (Lx)(t) = F x(t0 ), − Ax(t) , x ∈ D(L) dt (8) D(L) := {x ∈ L2 (t0 , t1 ) : F x ∈ H1 (t0 , t1 , Rm )} and note that (3) is equivalent to Lx(t) = (x0 , f (t)) and X0 = N (L). Since (x0 , f, η) ∈ G we can write that: 1 2 σ (t1 , `, u) = sup {Fu (x, η) : (Lx, η) ∈ G } Proposition 3. Take A01,2 ∈ Rn×n , A03,4 ∈ Rm×n consider DAE dp = A01 p + A02 q, p(t1 ) = w, dt 0 = A03 p + A04 q . Let us set by definition Ak+1 = (Ak1 − Ak2 (Ak4 )+ Ak3 )PN (I − PR (Ak4 ))Ak3 , 1 Ak+1 = Ak2 PN (Ak4 ), 2 Ak+1 = (I − PN (I − PR (Ak4 ))Ak3 )Ak+1 , 3 1k+1 k+1 k k A4 = (I − PN (I − PR (A4 ))A3 )A2 , Pis := usk=i PN (I − PR (Ak4 ))Ak3 , pk = PN (I − PR (Ak4 ))Ak3 pk+1 , pj = p∗ , q k = −(Ak4 )+ Ak3 pk + PN (Ak4 )q k+1 , q j = q ∗ . Then 1) there exists s ≤ n such that As3 = As4 = DAE (13)-(14) has a solution p, q iff w = P0s w; 0 0 0 0 (13) (14) (15) (16) (17) (18) (19) (20) (21) 0m×n ; 2) 3) if p, q solve (13)-(14) then p = p , q = q where p , q are defined from (20)-(21) with p∗ := p, q ∗ := q, k ∈ 0, s − 1 and dp = As1 p(t) + As2 q(t), p(t1 ) = w, q ∈ L2 (t0 , t1 ). (22) dt Im and set: Let us define I˜ = 0p×m 0 + 0 0 0 A1 = −(F ) A PR (F ) , A2 = −(F + )0 A0 PN (F 0 ) (F + )0 H 0 , A03 = −PN (F )A0 PR (F ), A04 = −PN (F )A0 PN (F 0 ) PN (F )H 0 s 0 −1 i 0 Q̃ = Q := PN 0(F ) I0p × us−1 −1 , β i=0 PN (A4 ) , 0 R γ := β s 0 Q̃β s , αs := − s−1 X −1 i + i s PN (Ai−1 4 )(A4 ) A3 Pi , A4 = 0, i=0 −1 0 + 0 D := (I − (PN (F 0 )Q−1 0 PN (F )) PN (F )Q0 ), ˜ 0s + P0s 0 I˜0 Q̃αs , ˜ 0s + αs 0 Q̃IP Q := αs 0 Q̃αs + P0s 0 I˜0 Q̃IP (9) x∈D(L),η so that `-error is a support function of the convex set X = {(x, η) : (Lx, η) ∈ G } (10) The support function of X describes the distance of the supporting hyperplane H (u) := {(x, η) : Fu (x, η) ≤ 1 σ 2 (t1 , `, u)} from the origin. Then, the `-estimate û defines a direction in a functional space L2 (t0 , t1 ) such that the distance of H (û) from the origin is minimal and so in this direction the maximal deviation of elements of X from the origin is minimal. Proposition 2. (dual control problem). Take ` ∈ Rm . The `-error σ̂(t1 , `) is finite iff there are z ∈ L2 (0, t1 , Rm ) and u ∈ L2 (0, t1 , Rp ) such that: d(F 0 z) = −A0 z(t) + H 0 u(t), F 0 z(t1 ) = F 0 ` . (11) dt If σ̂(t1 , `) < +∞ then `-estimate û is a unique solution of the following optimal control problem: Z t1 I (z, d, u) := (uT R−1 u + z T Q−1 z)dt(:= I1 (z, u)) and ˜ s , Q̃0 := D0 Q−1 D , Q0 := P s 0 Q̃0 P s . B s := αs + IP 0 0 0 0 The next theorem describes the minimax observable subspace, `-error and represents `-estimate in the form of a linear recursive filter on a subspace. Theorem 4. Assume that s ≤ n is defined as in Proposition 3 and define dK = (−As1 0 + B s 0 Q̃β s γ + As2 0 )K dt + K(−As1 + As2 γ + β s 0 Q̃B s ) − KAs2 γ + As2 0 K + + Q − B s 0 Q̃β s γ + β s 0 Q̃B s , K(t0 ) = Q0 , dx̂ = (−As1 0 + B s 0 Q̃β s γ + As2 0 − KAs2 γ + As2 0 )x̂ dt + (αs 0 + (KAs2 − B s 0 Q̃β s )γ + β s 0 ) y0 , x̂(t0 ) = 0 . Then û(y) = `T F F + x̂(t1 ) and σ̂(t1 , `) = `T F F + K(t1 )F F + ` , ∀` ∈ L (t1 ) , L (t1 ) = {` ∈ Rn : P0s (F + )0 F 0 ` = (F + )0 F 0 `} . (23) t0 + (F 0+ 3. EXAMPLE + 0 F z(t0 ) − d)T Q−1 F 0 z(t0 ) − d) → min 0 (F 0 d(F 0 z) = −A0 z + H 0 u, dt z,d,u F 0 d = 0, F 0 z(t1 ) = F 0 ` . (12) Next proposition presents the projection algorithm allowing to convert DAE to ODE. This example illustrates application of Theorem 4 to an ill-posed DAE which represents an ODE with unbounded inputs. As it was pointed out in Zhuk [2012] application of Pontryagin maximum principle to the dual control problem for this DAE would lead to the restriction of the minimax observable subspace. We apply a projection algorithm 145 16th IFAC Symposium on System Identification Brussels, Belgium. July 11-13, 2012 (Proposition 3) in order to preserve the structure h 1of0L0(t01i). 0 0 1 0 1 0 0 0 Let F = [ 0 1 0 0 ], A = −1 0 0 −1 , H = 0 0 0 1 . 0 1 0 0 dx ): Then (3)-(4) has the following form:(ẋ := dt ẋ1 = x3 + f1 , ẋ2 = −x1 − x4 + f2 , x1,2 (t0 ) = x1,2 0 , y1 = x1 + η1 , y2 = x4 + η2 , y3 = x2 + η3 . where x1,2 0 , f1,2 and η1,2,3 belong to (7) and E in (6) is defined with Q0 = I2 , Q = I2 , R = I3 . We see that x3,4 ∈ L2 (t0 , t1 ) are arbitrary functions. In this case, by Proposition 2 the `-estimate û of F x(t1 ) may be obtained minimizing the worst-case error Z t1 X Z t1 3 2 X 2 2 u2j dt (24) σ(t1 , `, u) = zi dt + kzi (t0 )k + i=1 t0 t0 j=1 over solutions to the adjoint DAE: ż1 − z2 − u1 = 0, z1 (t1 ) = `1 , −z1 ≡ 0 , (25) ż2 − u3 = 0, z2 (t1 ) = `2 , −z2 − u2 = 0 . The structure of (25) is simple and so we can explicitly compute û minimizing (25). This will allow to compare the corresponding `-estimate with the one derived from Theorem 4. (25) implies that û1,2 = −z2 . Hence, û3 is a unique solution to the following control problem Z t1 σ(t1 , `, u) = kz2 (t0 )k2 + 3z22 + u23 dt, Fig. 1. `-estimate ûε (y) = x̂2 (t) (Dotted), error σ̂(t, `) (DotDashed) and simulated trajectory x2 (t) (Solid), t ∈ [0, 1]. It is easy to see that the minimax estimate is a “central” point of the set of all possible realizations of the “true” state x2 . The estimate is “biased” due to the presence of unknown functions x2 , x3 . Nevertheless, the minimax error stays bounded at infinity that proves robustness of the `-estimate with respect to unknown and possibly unbounded disturbances in the state equation. t0 ż2 = u3 , z2 (t1 ) = `2 . The optimality condition takes the classical form: û3 = p where ż2 = p, z2 (t1 ) = `2 , ṗ = 3z2 , p(t0 ) = z2 (t0 ). Introducing k as a solution of the Riccati equation k̇ = 3− k 2 , k(0) = 1 we find that û3 = kz2 where ż2 = kz2 , z2 (t1 ) = `2 . Let x̂ be a solution to x̂˙ = −kx̂ − y1 − y2 + ky3 , x̂(t0 ) = 0 . (26) R t1 T Then it is easy to see that û(y) = t0 û ydt = `2 x̂(t1 ) and so x̂ represents a minimax state estimate. Now, let us use Proposition 3. We find that s = 2, L (t1 ) = {0} × R and 0 0 k1 k2 k̇1 k̇2 0 0 − k1 k2 = 0 3 0 1 k k k2 k4 , 2 4 k̇2 k̇4 (27) x̂˙ 1 = k2 y3 , x̂˙ 2 = −k2 x̂1 − k4 x̂2 − y1 − y2 + k4 y3 , 0 ) k2 (t0 ) with x̂1,2 (t0 ) = 0 and kk12 (t = 00 01 . From the (t0 ) k4 (t0 ) first equation of (27) we see that k1 = k2 = 0 so that x̂1 = 0 and the equation for x̂2 coincides with (26). In order to generate y1,2,3 we take t0 = 0, t1 = 1, x3 = cos(t) and x4 = sin(t), x1 (0) = 0.1, x2 (0) = −0.1, f1 = f2 = 0, η1 = −0.1, η2 = −0.2, η3 = 0.3. In Figure 1 the `-estimate û(y), and `-error are presented. As L(t) ≡ {0} × R we see that x1 is not observable in the minimax sense although y1 = x1 + η1 is observed. This can be explained as follows: the derivative x3 of x1 may be any element of L2 . It is not hard to compute RT that the expression for σ(t1 , `, u) contains t0 x3 z1 dt and so σ(t1 , `, u) < +∞ if and only if z1 (t) ≡ 0 for any t ∈ (t0 , t1 ). As z1 is absolutely continuous it follows that z1 (t1 ) = 0. Hence, if `1 = 0 then `-error is finite. If `1 6= 0 then the only candidate for û1 is the impulse control δ(t1 −t)`1 switching z1 from 0 to `1 at time-instant t1 . However, δ 6∈ L2 so that `-error is infinite. In other words, the algebraic structure of the adjoint DAE (25) can not “compensate” the unbounded derivative of x1 though it compensates the unbounded derivative of x2 and so the minimax observable subspace is non-trivial in that direction. It is curious to note that minimax error in the direction (0, 1) stays bounded with time as the solution of Riccati equation is bounded at infinity. 4. CONCLUSION In this paper we discuss the minimax state estimation approach for linear stationary DAEs with unknown but bounded initial condition, input and observation noise. Our approach is based on the generalized Kalman duality principle that allows to avoid constraints on DAE coefficients and input which are usually imposed in order to apply the theory of matrix pencils. The presented approach is relevant to high-dimensional numerical models that require state estimation algorithms in reduced form due to computational burden. In this case F might encapsulate the vectors which span a low dimensional reduced sub-space and the state of the resulting DAE describes the evolution of the projection coefficients. This reduction approach addresses stability issues of the classical Galerkin projection method (see [Mallet and Zhuk, 2010] for details). In perspective, we plan to apply the presented duality concept to stochastic DAEs. Another promising direction is an ensemble state estimation where one state is shared by different models and each model describes the dynamics of the same process reflecting various physical/chemical/numerical parameterizations for the same process (see [Garaud and Mallet, 2011] for an example in air quality modeling). 146 16th IFAC Symposium on System Identification Brussels, Belgium. July 11-13, 2012 REFERENCES A. Albert. Regression and the Moor-Penrose pseudoinverse. Acad. press, N.Y., 1972. K. Åström. Introduction to Stochastic Control. Dover, 2006. T. Başar and P. Bernhard. H∞ -optimal Control and Related Minimax Design Problems. Springer, 1995. J. S. Baras and A.B. Kurzhanski. Nonlinear filtering: The set-membership and the H∞ techniques. In Proc. 3rd IFAC Symp.Nonlinear Control Sys.Design. Pergamon, 1995. D. Bender and A. Laub. The linear-quadratic optimal regulator for descriptor systems. IEEE TAC, 32(8):672– 688, 1987. F. L. Chernousko. State Estimation for Dynamic Systems . Boca Raton, FL: CRC, 1994. M. Darouach, M. Boutayeb, and M. Zasadzinski. Kalman filtering for continuous descriptor systems. In ACC, pages 2108–2112, Albuquerque, June 1997. AACC. W. Fleming and R. Rishel. Deterministic and Stochastic Optimal control. Springer, 1975. F. Gantmacher. The theory of matrices. Chelsea Publish.Comp., N.-Y., 1960. Damien Garaud and Vivien Mallet. Automatic calibration of an ensemble for uncertainty estimation and probabilistic forecast: Application to air quality. J. Geophys. Res., 116, 2011. Markus Gerdin, Thomas B. Schön, Torkel Glad, Fredrik Gustafsson, and Lennart Ljung. On parameter and state estimation for linear differential-algebraic equations. Automatica J. IFAC, 43(3):416–425, 2007. D. Heitz, E. Mémin, and C. Schnörr. Variational fluid flow measurements from image sequences: synopsis and perspectives. Exp. Fluids, 48(3):369–393, 2010. A. Ioffe and V. Tikhomirov. Theory of extremal problems. North-Holland, Amsterdam, 1974. Alexander Kurzhanski and István Vályi. Ellipsoidal calculus for estimation and control. Systems & Control: Foundations & Applications. Birkhäuser Boston Inc., Boston, MA, 1997. ISBN 0-8176-3699-4. V. Mallet and S. Zhuk. Reduced minimax state estimation. Technical Report RR7500, INRIA, Paris-Rocquencourt, 2010. http://hal.archives-ouvertes.fr. V. Mallet and S. Zhuk. Reduced minimax filtering by means of differential-algebraic equations. In Proc. of 5th Int. Conf. on Physics and Control (PhysCon 2011). IPACS Electronic Library, 2011. Available at: lib.physcon.ru. Mario Milanese and Roberto Tempo. Optimal algorithms theory for robust estimation and prediction. IEEE Trans. Automat. Control, 30(8):730–738, 1985. P. Müller. Descriptor systems: pros and cons of system modelling by differential-algebraic equations. Mathematics and Computers in Simulation, 53(4-6):273–279, 2000. A. Nakonechny. A minimax estimate for functionals of the solutions of operator equations. Arch. Math. (Brno), 14 (1):55–59, 1978. ISSN 0044-8753. W. Schiehlen. Force coupling versus differential algebraic description of constrained multibody systems. Multibody System Dynamics, 4:317–340, 2000. Lin Wu, Vivien Mallet, Marc Bocquet, and Bruno Sportisse. A comparison study of data assimilation algorithms for ozone forecasts. J. Geophys. Res., 113 (D20310), 2008. S. Zhuk. Minimax state estimation for linear discrete-time differential-algebraic equations. Automatica J. IFAC, 46(11):1785–1789, 2010. doi: 10.1016/j.automatica.2010.06.040. S. Zhuk. Kalman duality principle for a class of illposed minimax control problems with linear differentialalgebraic constraints. Applied Mathematics & Optimisation, 2012. submitted. APPENDIX Proof of Proposition 2. Generalized Kalman duality for DAEs in the form (3) with zero initial condition was presented in [Zhuk, 2012, Th.2.4]. The proof given there is based on the operator interpretation of DAEs (see geometrical representation of `-estimate above). Applying the method of Zhuk [2012] (see the first part of the proof of Theorem 2.4) to the definition of operator L given by formula (8) it is easy to see that (11) has a solution and (see [Zhuk, 2012, frml.(2.10)]): 1 σ 2 (t1 , `, u) = min + (z0 ,v)∈N (L0 ) c(G , F 0 F 0 z(t0 ) − z0 , z − v, u) , where L0 is defined as follows: dF 0 v − A0 v, b ∈ D(L0 ), (L0 b)(t) = − dt D(L0 ) := {b = (z0 , v) : F 0 v ∈ H1 (t0 , t1 , Rm ), F 0 v(t1 ) = 0, v0 = F +0 (.1) (.2) F 0 v(t0 ) + d, F 0 d = 0} . Since (z0 , v) ∈ N (L ) and z solves (11), it follows that z+v solves (11), and so taking the min in (.1) with respect to (z0 , v) ∈ N (L0 ) is equivalent to minimize with respect to all possible d and z such that d : F 0 d = 0 and z solves (11) for the fixed u: 1 σ 2 (t1 , `, u) (.3) + = min c(G , F 0 F 0 z(t0 ) − d, z, u) 0 0 z solves (11) and F d=0 ˜ Noting that the support and min is attained at z̃, d. function of the ellipsoid G coincides with I , namely + c2 (G , F 0 F 0 z(t0 ) − d, z, u) = I (z, d, u) and recalling (.3) we obtain σ(t1 , `, u) = min I (z, d, u) (.4) 0 z solves (11) and F d=0 This completes the proof. Proof of Proposition 3. Consider (14). It is solvable iff (I − PR (A04 ))A03 p(t) ≡ 0, and its general solution has the following form (see [Albert, 1972]): q(t) = −(A04 )+ A03 p(t)+ PN (A04 )v(t), v ∈ L2 (t0 , t1 ). Using this we transform (13)(14) into the equivalent form (k = 0): dp = (Ak1 − Ak2 (Ak4 )+ Ak3 )p(t) + Ak2 PN (Ak4 )v(t), (.5) dt (I − PR (Ak4 ))Ak3 p(t) ≡ 0, p(t1 ) = w . (.6) Define r1 := dimN ((I − PR (A04 ))A03 ) and r0 := n. There are three different possibilities: r1 = r0 , r1 = 0 and 0 < r1 < r0 . Consider the case 0 < r1 < r0 . Define 147 16th IFAC Symposium on System Identification Brussels, Belgium. July 11-13, 2012 dpk = Ak1 pk (t) + Ak2 q k (t), dt 0 = Ak3 pk (t) + Ak4 q k (t) . k p (t1 ) = w, (.7) (.8) We claim that for k = 1 any solution of (.7)-(.8) corresponds to some solution of (13)-(14) through (20)-(21), provided P01 w = w. Let us prove this. Assume P01 w = w and p1 , q 1 solve (.7)-(.8). Let p0 , q 0 be defined by (20)(21) with p∗ = p1 , q ∗ = q 1 , k = 0. Then, by direct substitution one checks that p0 , q 1 solve (.5)-(.6) so p0 , q 0 solve (13)-(14). On the other hand, if p, v solve (.5)-(.6) then P01 w = w and p, v verify (.7) (this may be checked by direct substitution noting that PN ((I − PR (A04 ))A03 )p = p by (.6)) and (.8) (one checks this differentiating (.6)). Now, as above we transform (.7)-(.8) into (.5)-(.6) but with k = 1 and define r2 := dim N ((I −PR (A14 ))A13 ). Assuming 0 < r2 < r1 we continue constructing (.7)-(.8) and transforming it into (.5)-(.6). At some point for ri := dim N (I− i−1 PR (Ai−1 should be either ri = ri−1 or ri = 0. 4 ))A3 i−1 Assume ri = 0. Then N ((I − PR (Ai−1 4 ))A3 ) = {0} i−1 i−1 and, thus, PN (I − PR (A4 ))A3 = 0n so by (15),(17): Ak1 ≡ 0n×n and Ak3 ≡ 0m×n for k ≥ i. We also have Ai2 = i−1 i+1 i i Ai−1 = Ai2 PN (Ai4 ) = 0n 2 PN (A4 ), A4 = A2 so that A2 i+1 and thus A4 = 0m×n by (18). We see that the statement 1) of the proposition holds for s = i + 1. If ri = ri−1 i−1 n then N ((I − PR (Ai−1 so that PN (I − 4 ))A3 ) = R i−1 i−1 PR (A4 ))A3 = In and thus Ai3 = 0m×n , Ai4 = 0m×n by (17)-(18) so 1) holds with s = i. Assume 1) is verified for k = s. By construction, for any solution p, q of (13)-(14) we have that p = p0 and q = q 0 where p0 , q 0 are defined by (20) and (21) with p∗ = ps and q ∗ = q s , provided ps , q s solve (.7)-(.8) and P0s w = w. Since As3 = As4 = 0m×n it follows that (22) is equivalent to (.7)-(.8) for k = s and so statement 3) is verified. By 3) w = p(t1 ) = p0 (t1 ) = P0s p(t1 ) = P0s w so equality w = P0s w follows from the solvability of (13)-(14). On the other hand, if P0s w = w and p, q solve (22) then p0 , q 0 defined by (20) and (21) with p∗ = p and q ∗ = q solve (13)(14) by construction. This completes the proof. Proof of Theorem 4. Define p := PR (F )z, q̃ := PN (F 0 )z, q := (q̃, u)0 and let A01,2,3,4 be defined as above. We note that F 0 z = F 0 p and so (11) is equivalent to DAE (13)-(14). By statement 2) of Proposition 3 DAE (13)-(14) is solvable iff P0s (F + )0 F ` = (F + )0 F `. This and Proposition 2 prove the second line in (23). As we saw above (11) is equivalent to DAE (13)-(14). Let us represent I in (12) as a function of p and q. Recalling 0 that p(t0 ) = F + F 0 z(t0 ) we write: −1 I (z, d, u) = kQ0 2 (p(t0 ) − d)k2 + I1 (p + q̃, 0 I q) . Define I˜1 (z, u) := mind:F 0 d=0 I (z, d, u). It is easy to com−1 pute that mind:F 0 d=0 kQ0 2 (p(t0 ) − d)k2 = pT (t0 )Q̃0 p(t0 ) . and so I˜1 (z, u) = pT (t0 )Q̃0 p(t0 ) + I1 (p + q̃, 0 I q) . Z t1 I2 (p, q) = p (t0 )Q0 p(t0 ) + pT Qpdt t0 Z t1 (q T γq + 2q T β s 0 Q̃B s p)dt . + T (.9) t0 Since q uniquely defines p through (22), it follows that min 0 I (z, d, u) = min I2 (p, q) (.10) z,u solve (11),F d=0 q Note that I2 is convex and smooth in q. Thus, by [Ioffe and Tikhomirov, 1974, p.81,Prop.1] q̂ is a minimizer for I2 iff the Gateaux derivative of I2 at q̂ vanishes: d I2 (pτ , q̂ + τ v) = 0, where pτ denotes the solution dτ τ =0 of (22) with q = q̂ + τ v. Define dp̂ = As1 p̂ + As2 q̂, p̂(t1 ) = F 0+ F 0 `, dt (.11) dŵ = −As1 0 ŵ + Qp̂ + B s 0 Q̃β s q̂, ŵ(t0 ) = Q0 p̂(t0 ) . dt It is not difficult to check that Z t1 d =2 v T γ q̂dt I2 (pτ , q̂ + τ v) dτ t0 τ =0 Z t1 v T β s 0 Q̃B s p̂ − As2 0 ŵ dt . +2 t0 So the Gateaux derivative of I2 at q̂ vanishes iff q̂ solves the linear algebraic equation: γ q̂ = −β s 0 Q̃B s p̂ + As2 0 ŵ . (.12) We note that, detγ = 0 in the general case. So (.12) is solvable iff the vector on the right hand side of (.12) belongs to the range R(γ) of matrix γ. As Q̃ is a symmetric positive definite matrix, then R(γ) = R(β s 0 ). It is clear that −β s 0 Q̃B s p̂ ∈ R(β s 0 ). By definition, β s 0 = PN (As−1 )× 4 0 0 · · · × PN (A04 ) and As2 0 = PN (As−1 ) × · · · × P (A )A N 4 2 4 s0 s0 so A2 ŵ ∈ R(β ). Thus, the minimizer for I2 admits representation: q̂ = γ + As2 0 ŵ − β s 0 Q̃B s p̂ + v, (.13) where γv = 0. Let us plug (.13) into (.11). Then γv = 0 implies As2 v = 0. Thus p̂, ŵ do not depend on v and so does I2 . Therefore, we can take v = 0 in (.13). Now we split (.11) using the standard argument of the linear regulator problem (see [Fleming and Rishel, 1975, p.23,Exmpl.2.3,p.88]): plugging (.13) with v = 0 into (.11) we find that ŵ = K p̂ and q̂ = γ + As2 0 K − β s 0 Q̃B s p̂ where K solves Riccati equation formulated in the theorem’s statement. Thus, q̂ is in the feed-back form. In order to compute `-estimate û which solves (11), we recall that u = 0 I q and q = αs p + β s q so that û = 0 I (αs p̂ + β s q̂). Using this representation and definition of x̂ we find integrating by parts that: û(y) = R t1 T 0 q̂ y dt = `T F F + x̂(t1 ). Let us compute `-error. We t0 find plugging (.11) into (.9) and using (.4), (.10) that σ̂(t1 , `) = I2 (p̂, q̂) = p̂T (t1 )F F + ` = `T F F + K(t1 )F F + `. This completes the proof. By statement 3) of Proposition 3 any solution p, q of (13)(14) may be represented as p = P0s p and q = αs p + β s q where p, q solve (22). Using this it is straightforward to check that I˜1 (z, u) = I2 (p, q) where: 148