8.3 Certainty-Equivalent Filters (CEFs)
We discuss in this section a class of estimators for the nonlinear system (8.1) which we refer to as “certainty-equivalent” worst-case estimators (see also [139]). The estimator is constructed on the assumption that the asymptotic value of ˆx equals x, and the gain matrices are designed so that they are not functions of the state vector x, but of ˆx and y only. We begin with the one degree-of-freedom (1-DOF) case, then we discuss the 2-DOF case. Accordingly, we propose the following class of estimators:
∑acef1 : {˙ˆx = f(ˆx)+g1(ˆx)ˆw⋆+ˆL(ˆx,y)(y−h2(ˆx)−k21(ˆx)w⋆)ˆz = h2(ˆx)˜z = y−h2(ˆx) |
(8.32) |
where ˆz=ˆy∈ℜm is the estimated output, ˜z∈ℜm is the new penalty variable, ˆw⋆ is the estimated worst-case system noise and ˆL:X×Y→Mn×m is the gain matrix for the filter.
Again, the problem is to find a suitable gain matrix ˆL (.,.) for the above filter, for estimating the state x(t) of the system (8.1) from available observations Yt≜{y(τ),τ≤t} t ∈ [t0, ∞) such that the L2-gain from the disturbance/noise signal w to the error output ˜z (to be defined later) is less than or equal to a desired number γ > 0, i.e.,
(8.33) |
for all w ∈ W, for all x0 ∈ O ⊂ X. Moreover, with w = 0, we also have limt→∞˜z(t)=0
The above problem can similarly be formulated by considering the cost functional
(8.34) |
as a differential game with the two contending players controlling L and w respectively. Again by making J ≤ 0, we see that the L2-gain constraint (8.33) is satisfied, and moreover, we seek for a saddle-point equilibrium solution, (ˆL⋆,w⋆), to the above game such that
(8.35) |
Before we proceed with the solution to the problem, we introduce the following notion of local zero-input observability.
Definition 8.3.1 For the nonlinear system Σa, we say that it is locally zero-input observable if for all states x1, x2 ∈ U ⊂ X and input w(.) ≡ 0
y(.,x1,w) = y(.,x2,w)⇒x1 = x2,
where y(., xi, w), i = 1, 2 is the output of the system with the initial condition x(t0) = xi. Moreover, the system is said to be zero-input observable, if it is locally zero-input observable at each x0 ∈ X or U = X.
Next, we first estimate ŵ⋆, and for this purpose, define the Hamiltonian function H : T∗X×W×Mn×m→ℜ for the estimator:
H(ˆx,w,ˆL,ˆVTˆx) = ˆVˆx(ˆx,y)[f(ˆx)+g1(ˆx)w+ˆL(ˆx,y)(y−h2(ˆx)−k21(ˆx)w)]+ˆVy˙y+ 12(‖˜z‖2−γ2‖w‖2) |
(8.36) |
for some smooth function ˆV:X×Y→ℜ and with the adjoint vector p=ˆVTˆx(ˆx,y).
Applying now the necessary condition for the worst-case noise
∂H∂w|w=ˆw⋆ = 0⇒ ˆw⋆=1γ2[gT1(ˆx)ˆVTˆx(ˆx)−kT21(ˆx)ˆL(ˆx,y)ˆVTˆx(ˆx,y)].
Then substituting ŵ⋆ into (8.36), we get
H(ˆx,ˆw⋆,ˆL,ˆVTˆx) = Vˆxf(ˆx)+ˆVy˙y+12γ2ˆVˆxg1(ˆx)gT1(ˆx)ˆVTˆx(ˆx)+ˆVˆxˆL(ˆx,y)(y−h2(ˆx))+ 12γ2ˆVˆxˆL(ˆx,y)ˆLT(ˆx,y)ˆVˆx+12(y−h2(ˆx))T(y−h2(ˆx)). |
(8.37) |
Completing the squares now for ˆL in the above expression for H(., ŵ⋆ , L, .), we get
H(ˆx,ˆw⋆,ˆL,ˆVTˆx) = Vˆxf(ˆx)+ˆVy˙y+12γ2ˆVˆxg1(ˆx)gT1(ˆx)ˆVTˆx−γ22(y−h2(ˆx))T(y−h2(ˆx))+ 12γ2‖ˆLT(ˆx,y)ˆVˆx+γ2(y−h2(ˆx))‖2+12(y−h2(ˆx))T(y−h2(ˆx)).
Therefore, choosing ˆL⋆ as
(8.38) |
minimizes H(., ., ., ., .) and gurantees that the saddle-point conditions (8.35) are satisfied by (ŵ⋆, ˆL⋆). Finally, substituting this value in (8.37) and setting
H(.,ˆw⋆,ˆL⋆,.,.)=0,
yields the HJIE
ˆVˆx(ˆx,y)f(ˆx)+ˆVy(ˆx,y)˙y+12γ2ˆVˆx(ˆx,y)g1(ˆx)gT(ˆx)ˆVTˆx(ˆx,y)− γ22(y−h2(ˆx))T(y−h2(ˆx))+ 12(y−h2(ˆx))T(y−h2(ˆx)) = 0, ˆV(0,0)=0, |
(8.39) |
or equivalently the HJIE
ˆVˆx(ˆx,y)f(ˆx)+ˆVy(ˆx,y)˙y+12γ2ˆVˆx(ˆx,y)g1(ˆx)gT1(ˆx)ˆVTˆx(ˆx,y)− 12γ2ˆVˆx(ˆx,y)ˆL(ˆx,y)ˆLT(ˆx,y)ˆVTˆx(ˆx,y) + 12(y−h2(ˆx))T(y−h2(ˆx)) = 0, ˆV(0,0)=0. |
(8.40) |
This is summarized in the following result.
Proposition 8.3.1 Consider the nonlinear system (8.1) and the N L H I F P for it. Suppose Assumption 8.1.1 holds, the plant Σa is locally asymptotically-stable about the equilibrium-point x = 0 and zero-input observable. Suppose further, there exists a C1 positive semidefinite function ˆV:ˆN׈Y→ℜ+ locally defined in a neighborhood ˆN׈Y⊂X×Y of the origin (ˆx, y) = (0, 0), and a matrix function ˆL:ˆN׈Y→Mn×m, satisfying the HJIE (8.39) or (8.40) together with the coupling condition (8.38). Then the filter ∑acef2 solves the N L H I F P for the system.
Proof: Let ˆV ≥ 0 be a C1 solution of the H J I E (8.39) or (8.40). Differentiating ˆV along a trajectory of (8.32) with w in place of ŵ⋆, it can be shown that
˙ˆV≤12(γ2‖ˆw‖2−‖˜z‖2).
Integrating this inequality from t = t0 to t = ∞, and since V ≥ 0 implies that the L2-gain condition ∞∫t0‖ˆz(t)‖2dt≤γ2∞∫t0‖w(t)‖2dt is satisfied. Moreover, setting w = 0 in the above inequality implies that ˙ˆV≤−12‖˜z‖2 and hence the estimator dynamics is stable. In addition,
˙ˆV=0∀t≥ts⇒˜z≡0⇒y=h2(ˆx)=ˆy.
By the zero-input observability of the system Σa, this implies that x = ˆx. □
Remark 8.3.1 Notice also that for γ = 1, a control Lyapunov-function ˆV [102, 170] for the system (8.1) solves the above HJIE (8.39).
8.3.1 2-DOF Certainty-Equivalent Filters
Next, we extend the design procedure in the previous section to the 2-DOF case. In this regard, we assume that the time derivative ˙y is available as an additional measurement information. This can be obtained or estimated using a differentiator or numerically. Notice also that, with y = h2(x) + k21(x)w and with w ∈ L2(0, T ) for some T sufficiently large, then since the space of continuous functions with compact support Cc is dense in L2, we may assume that the time derivatives ˙y, ÿ exist a.e. That is, we may approximate w by piece-wise C1 functions. However, this assumption would have been difficult to justify if we had assumed w to be some random process, e.g., a white noise process.
Accordingly, consider the following class of filters:
Σacef2 :{‵˙x = f(‵x)+g1(‵x)‵w⋆+‵L1(‵x,y,˙y)(y−h2(‵x)−k21(‵x)‵w⋆)+ ‵L2(‵x,y,˙y)(˙y−Lf(‵x)h2(‵x))‵z =[ h2(‵x)Lf(‵x)h2(‵x)]˜z = [ y−h2(‵x)˙y−Lf(‵x)h2(‵x)]
where ‵z ∈ℜm is the estimated output of the filter, ˜z∈ℜs is the error or penalty variable, Lf h2 denotes the Lie-derivative of h2 along f [268], while ‵L1:X×TY→Mn×m, ‵L2:X×TY→Mn×m are the filter gains, and all the other variables and functions have their corresponding previous meanings and dimensions. As in the previous subsection, we can define the Hamiltonian ‵H:T⋆X×W×Mn×m×Mn×m→ℜ: of the system as
‵H(‵x,ˋw,'L1,‵L2,‵V‵xT)=‵V‵x(‵x,y,˙y){f(‵x)+g1(‵x)‵w+‵L1(‵x,y,˙y)(y−h2(‵x)−k21(‵x)‵w)+ ‵L2(‵x,y,˙y)(˙y−ℒf(‵x)h2(‵x))} +‵Vy˙y+‵V˙y¨y+12(‖˜z‖2+γ2‖‵w‖2) |
(8.41) |
for some smooth function ‵V : X × T Y → ℜ and by setting the adjoint vector ‵p=‵V‵xT. Then proceeding as before, we clearly have
‵w⋆=1γ2[gT1(‵x)‵VT‵x(‵x,y,˙y)−kT21‵L1(‵x,y,˙y)‵VT‵x(‵x,y,˙y)].
Substitute now ‵w⋆ into (8.41) to get
‵H(‵x,‵w,‵L1,‵L2,‵V‵xT)=‵V‵xf(‵x)+‵Vy˙y+‵V˙y¨y+12γ2‵V‵xg1(‵x)gT1(‵x)‵VT‵x+ 'Vˋx‵L1(‵x,y,˙y)(y−h2(‵x))+12γ2‵V‵x‵L1(‵x,y,˙y)‵LT1(‵x,y,˙y)‵VT‵x+ ‵V‵x‵L2(‵x,y,˙y)(˙y−ℒfh2(‵x))+12(˙y−ℒfh2(‵x))T(˙y−ℒfh2(‵x))+ 12(y−h2(‵x))T(y−h2(‵x))
and completing the squares for ‵L1 and ‵L2, we get
‵H(‵x,‵w⋆,‵L1,‵L2,‵V‵xT)=‵V‵xf(‵x)+‵Vy˙y+‵V˙y¨y+12γ2‵V‵xg1(‵x)gT1(‵x)‵VT‵x− γ22(y−h2(‵x))T(y−h2(‵x))+12γ2‖‵LT1(‵x,y,˙y)‵VT‵x+γ2(y−h2(‵x))‖2+ 12 ‖LT2(‵x,y,˙y)‵VT‵x+(˙y−ℒfh2(‵x))‖2−12 ‵V‵x L2(‵x,y,˙y)LT2(‵x,y,˙y)‵VT‵x+ 12(y−h2(‵x))T(y−h2(‵x))
Hence, setting
(8.42) |
(8.43) |
minimizes ‵H and gurantees that the saddle-point conditions (8.35) are satisfied. Finally, setting
‵H(‵x,‵w⋆,‵L⋆1,‵LT2,‵VT‵x)=0
yields the HJIE
‵V‵x(‵x,y,˙y)f(‵x)+‵Vy(‵x,y,˙y)˙y+‵V˙y(‵x,y,˙y)¨y + 12γ2‵V‵x(‵x,y,˙y)g1(‵x)gT1(‵x)‵VT‵x(‵x,y,˙y) −12(˙y−ℒfh2(‵x))T(˙y−ℒfh2(‵x))+(1−γ2)2(y−h2(‵x))T(y−h2(‵x)) = 0, ‵V(0,0) = 0. |
(8.44) |
Consequently, we have the following result.
Theorem 8.3.1 Consider the nonlinear system (8.1) and the N L H I F P for it. Suppose Assumption 8.1.1 holds, the plant Σa is locally asymptotically-stable about the equilibrium-point x = 0 and zero-input observable. Suppose further, there exists a C1 positive-semidefinite function ‵V:‵N×‵Y→ℜ+ locally defined in a neighborhood ‵N×‵Y⊂X×TY of the origin (‵x,y,˙y) = (0, 0, 0), and matrix functions 'L1:N×‵Y→Mn×m,‵L2:N×⌣Y→Mn×m, satisfying the HJIE (8.44) or equivalently the HJIE:
together with the coupling condition (8.42), (8.43). Then the filter ∑acef1 solves the H I N L F P for the system.
Proof: The first part of the proof has already been established above, that ‵(w⋆,‵[L⋆1,‵L⋆2]) constitute a saddle-point solution for the game (8.34), (8.41). Therefore, we only need to show that the L2-gain requirement (8.33) is satisfied, and the filter provides asymptotic estimates when w = 0.
Let ‵V≥0 be a C1 solution of the HJIE (8.44), and differentiating it along a trajectory of the filter ∑acef1 with ‵w in place of ‵w⋆, and ‵L⋆1,‵L⋆2, we have
‵V = ‵V‵x(‵x,y,˙y)[f(‵x)+g1(‵x)‵w+‵L1(‵x,y)(y−h2(‵x)−k21(‵x)‵w)+ L2(‵x,y,˙y)(˙y−ℒfh2(‵x))] +‵Vy˙y+‵V˙y¨y ={‵V‵xf(‵x)+‵Vy˙y+‵V˙y¨y+12γ2‵V‵xg1(‵x)gT1(‵x)‵VT‵x+ (1−γ2)2(y−h2(‵x)T(y−h2(‵x) −12(˙y−ℒfh2(‵x))T(˙y−ℒfh2(‵x))}+ ‵V‵xg1(‵x)‵w−‵V‵x‵L⋆1(‵x,y)k21(‵x)‵w−12γ2‵V‵xg1(‵x)gT1(‵x)‵VT‵x− γ22(y−h2(‵x)T(y−h2(‵x)+‵V‵xL2(‵x,y,˙y)(˙y−ℒfh2(‵x))− 12(y−h2(‵x))T(y−h2(‵x))+12(˙y−ℒfh2(‵x))T(˙y−ℒfh2(‵x))
=−γ22‖w−1γ2gT1(‵x)‵V‵xT+1γ2kT21(‵x)‵L⋆T1(‵x,y)‵VT‵x‖2+γ22‖‵w‖2−12‖˜z‖2≤12(γ2‖‵w‖2−12‖˜z‖2).
Integrating the above inequality from t = t0 to t = ∞, we get that the L2-gain condition (8.33) is satisfied.
Similarly, setting ‵w⋆ = 0, we have .ˊV=−12‖˜z‖2. Therefore, the filter dynamics is stable.
Moreover, the condition
˙ˊV≡0∀t≥ts⇒˜z≡0⇒y=h2(ˆx), ˙y≡Lfh2(ˋx).
By the zero-input observability of the system Σa, this implies that `x ≡ x ∀t ≥ ts. □
We consider an example to illustrate and compare the performances of the 1-DOF and the 2-DOF.
Example 8.3.1 Consider the nonlinear system
˙x1 = −x31−x2˙x2 = −x1−x2 y = x1+w
where w = 5w0 + 5 sin(t) and w0 is a zero-mean Gaussian white-noise with unit variance. It can be checked that the system is locally zero-input observable and the functions ˆV(x)=12(ˆx21+ˆx22),ˋV(‵x)=12(‵x21+‵x22) solve the HJI-inequalities corresponding to (8.39), (8.44) for the 1-DOF and 2-DOF certainty-equivalent filters respectively with γ = 1. Subsequently, we calculate the gains of the filters as
ˆL(ˆx,y)=[ 1−y/ˆx2], ‵L1(‵x,y,˙y)= [ 1−y/ˆx2], ‵L2(‵x,y,˙y)= [ −‵x22−(˙y−‵x2)/‵x2]
and construct the filters ∑acef1,∑acef2 respectively. The results of the individual filter performance with the same initial conditions but unknown system initial conditions are shown in Figure 8.5. It is seen from the simulation that the 2-DOF filter has faster convergence than the 1-DOF filter, though the latter may have better steady-state performance. The estimation errors are also insensitive or robust against a significantly high persistent measurement noise.
8.4 Discrete-Time Nonlinear H∞-Filtering
In this section, we consider H∞-filtering problem for discrete-time nonlinear systems. The configuration for this problem is similar to the continuous-time problem shown in Figure 8.1 but with discrete-time inputs and output measurements instead, and is shown in Figure 8.7. We consider an affine causal discrete-time state-space system with zero input defined on X⊆ℜn in coordinates x = (x1,&, xn):
Σda : { xk+1 = f(xk)+g1(xk)wk; x(k0) =x0 zk = h1(xk) yk = h2(xk)+k21(xk)wk |
(8.46) |
where x ∈ X is the state vector, w ∈ W ⊂ ℓ 2[k0, ∞) is the disturbance or noise signal, which belongs to the set W ⊂ ℜr of admissible disturbances or noise signals, the output y ∈ ℜm is the measured output of the system, while z ∈ ℜs is the output to be estimated. The functions f : χ → χ, g1 : χ → Mn×r(χ ), h1 : χ → ℜ s, h2 : χ → ℜ m, and k12 : χ → Ms×p(χ ), k21 : χ → Mm×r(χ) are real C∞ functions of x. Furthermore, we assume without any loss of generality that the system (8.46) has a unique equilibrium-point at x = 0 such that f(0) = 0, h1(0) = h2(0) = 0.
The objective is to synthesize a causal filter, Fk, for estimating the state xk or a function of it, zk = h1(xk), from observations of yk up to time k ∈ Z+, i.e., from
Yk≜ {yi : i≤k},
such that the ℓ2-gain from w to ⌣z, the error (or penalty variable), of the interconnected system defined as
(8.47) |
where w := {wk}, ⌣z:={⌣zk}, is rendered less or equal to some given positive number γ > 0. Moreover, with wk ≡ 0, we have limk→∞ ⌣zk = 0.
For nonlinear systems, the above condition is interpreted as the ℓ2-gain constraint and is represented as
(8.48) |
for all wk ∈ W, and for all x0 ∈ O ⊂ χ .
The discrete-time nonlinear H∞ suboptimal filtering or estimation problem can be defined formally as follows.
Definition 8.4.1 (Discrete-Time Nonlinear ℋ∞ (Suboptimal) Filtering/Prediction Problem (DNLHIFP)). Given the plant Σda and a number γ⋆ > 0. Find a filter Fk : Y → χ such that
ˆxk+1 =Fk(Yk), k=k0,…,K
and the constraint (8.48) is satisfied for all γ ≥ γ⋆ , for all w ∈ W, and for all x0 ∈ O. Moreover, with wk ≡ 0, we have limk→∞ ⌣zk = 0.
Remark 8.4.1 The problem defined above is the finite-horizon filtering problem. We have the infinite-horizon problem if we let K → ∞.
To solve the above problem, we consider the following class of estimators:
Σdaf : { ˆxk+1 = f(ˆxk)+L(ˆxk,k)[y−h2(ˆxk)], ˆx(k0) =ˆx0 ˆzk = h1(ˆxk) |
(8.49) |
where ˆxk∈X is the estimated state, L(., .) ∈ Mn×m(X × Z) is the error-gain matrix which has to be determined, and ⌣z∈ℜs is the estimated output of the filter. We can now define the estimation error or penalty variable, ⌣z, which has to be controlled as:
⌣zk : zk −ˆzk = h1(xk)−h1(ˆxk).
Then we combine the plant (8.46) and estimator (8.49) dynamics to obtain the following augmented system:
⌣xk+1 = ⌣f(⌣xk)+⌣g(⌣xk)wk, ⌣x(k0) = (x0T ˆx0T)T ⌣zk = ⌣h(⌣xk)}, |
(8.50) |
where⌣xk = (xTk ˆxTk)T,
⌣f(⌣x) = ( f(xk)f(ˆxk)+L(ˆxk,k)(h2(xk)−h2(ˆxk))) , ⌣g(⌣x)=( g1(xk)L(ˆxk,k)k21(xk)) ,⌣h(⌣xk) = h1(xk) −h2(ˆxk).
The problem is then formulated as a two-player zero-sum differential game with the cost functional:
(8.51) |
where w := {wk}, L := L(xk, k). By making J ≤ 0, we see that the H∞ constraint ‖Σdaf ◦ Σa‖H∞ ≤ γ is satisfied. A saddle-point equilibrium solution to the above game is said to exist if we can find a pair (L⋆ , w⋆ ) such that
(8.52) |
Sufficient conditions for the solvability of the above game are well known [59]. These are also given as Theorem 2.2.2 which we recall here.
Theorem 8.4.1 For the two-person discrete-time nonzero-sum game (8.51)-(8.50), under memoryless perfect information structure, a pair of strategies (L⋆, w⋆) provides a feedback saddle-point solution if, and ony if, there exist a set of K − k0 functions V : X × Z → ℜ, such that the following recursive equations (or discrete-Hamilton-Jacobi-Isaac’s equations (DHJIE)) are satisfied for each k ∈ [k0, K]:
Equation (8.53) is also known as Isaac’s equation.
Thus, we can apply the above theorem to derive sufficient conditions for the solvability of the D N L H I F P. To do that, we define the Hamiltonian function H : (X×X ) × W × ℳn×m × ℜ → ℜ associated with the cost functional (8.51) and the estimator dynamics as
H(⌣x,wk,L,V) = V(⌣f(⌣x)+⌣g(⌣x)wk(⌣x),k+1)−V(⌣x,k)+12‖⌣zk(⌣x)‖2−12γ2‖wk(⌣x)‖2, |
(8.54) |
where the adjoint variable has been set to p = V. Then we have the following result.
Theorem 8.4.2 Consider the nonlinear system (8.46) and the DNLHIF P for it. Suppose the function h1 is one-to-one (or injective) and the plant Σda is locally asymptotically-stable about the equilibrium-point x = 0. Further, suppose there exists a C1 (with respect to both arguments) positive-definite function V : N × N × Z → ℜ locally defined in a neighborhood N ⊂ χ of the origin ⌣x=0, and a matrix function L : N × Z → Mn×m satisfying the following DHJIE:
V(⌣x,k) = V(⌣f⋆(⌣xk)+⌣g⋆(⌣xk)w⋆k(⌣x),k+1)+12‖⌣zk(⌣x)‖2−12γ2‖w⋆k(⌣x)‖2, V(⌣x,K+1)=0 |
(8.55) |
k = k0,…, K, together with the side-conditions:
(8.56) |
(8.57) |
(8.58) |
(8.59) |
Then: (i) there exists a unique saddle-point equilibrium solution (w⋆, L⋆) for the game (8.51), (8.50) locally in N; and (ii) the filter Σdaf with the gain-matrix L(ˆxk,k) satisfying (8.57) solves the finite-horizon D N L H I F P for the system locally in N.
Proof: Assume there exists positive-definite solution V (., k) to the DHJIEs (8.55) in N ⊂ X for each k, and (i) consider the Hamiltonian function H(., ., ., .). Apply the necessary condition for optimality, i.e.,
∂TH∂w|w=w⋆ = ⌣gT(⌣x)∂TV(λ,k+1)∂λ|λ=⌣f(⌣x)+⌣g(⌣x)w⋆k−γ2w⋆k=0,
to get
w⋆k = 1γ2⌣gT(⌣x)∂TV(λ,k+1)∂λ|λ=⌣f(⌣x)+⌣g(⌣x)w⋆k:=α0(⌣x,w⋆k). |
(8.60) |
Thus, w⋆ is expressed implicitly. Moreover, since
∂TH∂w2 = ⌣gT(⌣x)∂TV(λ,k+1)∂λ2|λ=⌣f(⌣x)+⌣g(⌣x)w⋆k⌣g(⌣x)−γ2I
is nonsingular about (⌣x,w)=(0,0), the equation (8.60) has a unique solution α0(⌣x), α0(0) = 0 in the neighborhood N × W of (x, w) = (0, 0) by the Implicit-function Theorem [234].
Now substitute w⋆ in the expression for H(., ., ., .) (8.54), to get
H(⌣x,w⋆k,L,V) = V(⌣f(⌣x)+⌣g(⌣x)w⋆k(⌣x),k+1)−V(⌣x,k)+12‖⌣zk(⌣x)‖2−12γ2‖w⋆k(⌣x)‖2
and let
L⋆ = arg minL{H(⌣x,w⋆k,L,V)}.
Then, by Taylor’s theorem, we can expand H(., ., ., .) about (L⋆ , w⋆ ) [267] as
H(⌣x,w,L,V) =H(⌣x,w⋆,L,V)+12(w−w⋆)T∂2H∂w2(w,L⋆)(w−w⋆)+ 12Tr{[In⊗(L−L⋆)T]∂2H∂L2(w⋆,L)[Im⊗(L−L⋆)T]}+ O(‖w−w⋆‖3+‖L−L⋆‖3). |
(8.61) |
Thus, taking L⋆ as in (8.57) and w⋆ = α0(⌣x, w⋆) and if the conditions (8.58), (8.59) hold, we see that the saddle-point conditions
H(w,L⋆)≤H(w⋆,L⋆)≤H(w⋆,L)∀L∈Mn×m,∀w∈l2[k0,K], k∈[k0,K]
are locally satisfied. Moreover, substituting (w⋆ , L⋆) in (8.53) gives the DHJIE (8.55).
(ii) Reconsider equation (8.61). Since the conditions (8.58) and (8.59) are satisfied about ⌣x = 0, by the Inverse-function Theorem [234], there exists a neighborhood U ⊂ N × N of ⌣x = 0 for which they are also satisfied. Consequently, we immediately have the important inequality
H(⌣x,w,L⋆,V) ≤H(⌣x,w⋆,L⋆,V) = 0 ∀⌣x∈U
(8.62) |
Summing now from k = k0 to k = K, we get the dissipation inequality [183]:
(8.63) |
Thus, the system has ℓ2-gain from w to ⌣z less or equal to γ. □
8.4.1 Infinite-Horizon Discrete-Time Nonlinear H∞-Filtering
In this subsection, we discuss the infinite-horizon discrete-time filtering problem, in which case we let K → ∞. Since we are interested in finding a time-invariant gain for the filter, we seek a time-independent function V:˜NטN→ℜ locally defined in a neighborhood ˜NטN⊂χ×χ of (x,ˆx)=(0,0) such that the following stationary DHJIE:
˜V(⌣f⋆(⌣x)+⌣g⋆(⌣x)w⋆(⌣x))−˜V(⌣x)+12‖⌣z(⌣x)‖2−12γ2‖w⋆(⌣x)‖2=0, ˜V(0)=0, x,ˆx∈ˆN |
(8.64) |
is satisfied together with the side-conditions:
(8.65) |
(8.66) |
(8.67) |
(8.68) |
where
⌣f⋆=⌣f(⌣x)|˜L=˜L⋆ , ⌣g⋆(⌣x)=⌣g(⌣x)|˜L=˜L⋆,
˜H(⌣x,wk,˜L,˜V) = ˜V(⌣f(⌣x)+⌣g(⌣x)w)−˜V(⌣x)+12‖⌣zk‖2−12γ2‖wk‖2, |
(8.69) |
and ˜w⋆,˜L⋆ are the asymptotic values of w⋆k,Lk⋆ respectively. Again here, since the estimation is carried over an infinite-horizon, it is necessary to ensure that the augmented system (8.50) is stable with w = 0. However, in this case, we can relax the requirement of asymptotic-stability for the original system (8.46) with a milder requirement of detectability which we define next.
Definition 8.4.2 The pair {f, h} is said to be locally zero-state detectable if there exists a neighborhood O of x = 0 such that, if xk is a trajectory of xk+1 = f(xk) satisfying x(k0) ∈ O, then h(xk) is defined for all k ≥ k0 and h(xk) ≡ 0 for all k ≥ ks, implies limk→∞ xk = 0.
A filter is also required to be stable, so that trajectories do not blow-up for example in an open-loop system. Thus, we define the “admissibility” of a filter as follows.
Definition 8.4.3 A filter F is admissible if it is (internally) asymptotically (or exponentially) stable for any given initial condition x(k0) of the plant Σda, and with w = 0
limx→∞⌣zk=0.
The following proposition can now be proven along the same lines as Theorem 8.4.2.
Proposition 8.4.1 Consider the nonlinear system (8.46) and the infinite-horizon D N L H I F P for it. Suppose the function h1 is one-to-one (or injective) and the plant Σda is zero-state detectable. Further, suppose there exist a C1 positive-definite function ˜V:˜NטN→ℜ locally defined in a neighborhood where ⌣xk = (xTk ˆxTk)T,˜N⊂χ of the origin ⌣x=0,˜V(0)=0 and a matrix function ˜L:˜N→Mn×m satisfying the DHJIEs (8.64) together with (8.65)-(8.69). Then: (i) there exists locally a unique saddle-point solution (˜w⋆,˜L⋆) for the game; and (ii) the filter Σdaf with the gain matrix ˜L(ˆx)=˜L⋆(ˆx) satisfying (8.66) solves the infinite-horizon D N L H I F P locally in ˜N.
Proof: (Sketch). We only prove that the filter Σdaf is admissible, as the rest of the proof is similar to the finite-horizon case. Using similar manipulations as in the proof of Theorem 8.4.1, it can be shown that with w = 0,
˜V(⌣xk+1)−˜V(⌣xk)≤ −12‖⌣zk‖2.
Therefore, the augmented system is locally stable. Further, the condition that ˜V(⌣xk+1)≡˜V(⌣xk)∀k≥kc, for some kc ≥ k0, implies that ⌣zk≡0∀k≥kc. Moreover, the zero-state detectability of the system (8.46) implies the zero-state detectability of the system (8.50) since h1 is injective. Thus, by virtue of this, we have limk→∞ xk = 0, and by LaSalle’s invariance-principle, this implies asymptotic-stability of the system (8.50). Hence, the filter Σdaf is admissible. □
8.4.2 Approximate and Explicit Solution
In this subsection, we discuss how the D N L H I F P can be solved approximately to obtain explicit solutions [126]. We consider the infinite-horizon problem, but the approach can also be used for the finite-horizon problem. For simplicity, we make the following assumption on the system matrices.
Assumption 8.4.1 The system matrices are such that
k21(x)gT1(x) = 0 ∀x∈X,k21(x)kT21(x) = I ∀x∈X.
Now, consider the infinite-horizon Hamiltonian function ˜H(., ., ., .) defined by (8.69). Expanding it in Taylor-series about up to first-order1 and denoting this expansion by (., ., ., .) and L by , we get:
(8.70) |
where are the row-vectors of the partial-derivatives of with respect to x, respectively,
and
Then, applying the necessary condition for optimality, we get
(8.71) |
Now substitute in (8.70) to obtain
Completing the squares for in the above expression for (., ., ., .), we get
Thus, taking as
(8.72) |
minimizes (., ., ., .) and renders the saddle-point condition
satisfied.
Substitute now as given by (8.72) in the expression for (., ., ., .) and complete the squares in w to obtain:
Thus, substituting w = ŵ⋆ as given in (8.71), we see that the second saddle-point condition
is also satisfied. Hence, the pair constitutes a unique saddle-point solution to the game corresponding to the Hamiltonian (., ., ., .). Finally, substituting in the DHJIE (8.53), we get the following DHJIE:
(8.73) |
With the above analysis, we have the following theorem.
Theorem 8.4.3 Consider the nonlinear system (8.46) and the infinite-horizon D N L H I F P for this system. Suppose the function h1 is one-to-one (or injective) and the plant Σda is zero-state detectable. Suppose further there exists a C1 positive-definite function locally defined in a neighborhood of the origin , anda matrix function the DHJIE (8.73) together with the side-condition (8.72). Then:
(i) there exists locally in a unique saddle-point solution for the dynamic game corresponding to (8.51), (8.50);
(ii) the filter Σdaf with the gain matrix satisfying (8.72) solves the infinite-horizon D N L H I F P for the system locally in
Proof: Part (i) has already been shown above. For part (ii), consider the time variation of (a solution to the DHJIE (8.73)) along a trajectory of the augmented system (8.50) with , i.e.,
where use has been made of the Taylor-series approximation, equation (8.72) and the DHJIE (8.73). Finally, the above inequality clearly implies the infinitesimal dissipation-inequality:
Therefore, the system (8.50) has locally ℓ2-gain from w to less or equal to γ. The remaining arguments are the same as in the proof of Proposition 8.4.1. □
We now specialize the result of the above theorem to the linear-time-invariant (LTI) system:
(8.74) |
where all the variables have their previous meanings, and A ∈ ℜ n×n, B1 ∈ ℜn×r, C1 ∈ ℜs×n, C2 ∈ ℜm×n and D21 ∈ ℜm×r are constant real matrices. We have the following corollary to Theorem 8.4.3.
Corollary 8.4.1 Consider the discrete-LTI system Σdl defined by (8.74) and the D N L H I F P for it. Suppose C1 is full column rank and A is Hurwitz. Suppose further, there exists a positive-definite real symmetric solution to the discrete algebraic-Riccati
equation (DARE):
(8.75) |
together with the coupling condition:
(8.76) |
Then:
(i) there exists a unique saddle-point solution for the game given by
(ii) the filter F defined by
with the gain matrix satisfying (8.76) solves the infinite-horizon D N L H I F P for the system.
Proof: Take:
and apply the result of the theorem. □
Next we consider an example.
Example 8.4.1 We consider the following scalar example:
where wk = w0 + 0.1sin(2πk), and w0 is zero-mean Gaussian white noise. We compute the solution of the DHJIE (8.73) using the iterative scheme:
(8.77) |
starting with the initial guess and initial filter-gain l0 = 1, after one iteration, we get and . Then we proceed to compute the filter-gain using (8.72). The result of the simulation is shown in Figure 8.7.