betaFeedback

WorksMathematics

Linear Exponential Quadratic Gaussian Covariance Steering

Abstract

We formulate and analyze the linear exponential quadratic Gaussian (LEQG) covariance steering problem in continuous time over a given deadline (finite time horizon). The solution for this problem can be seen as a risk-sensitive Schrödinger bridge between Gaussian endpoints in the linear quadratic setting. Unlike the risk-neutral case, the LEQG covariance steering controller–still a linear state feedback–can no longer be written in closed form. We show that the optimal controller is parameterized by a symmetric matrix solving an algebraic equation that encodes the implicit dependence on the risk-sensitivity parameter. We explain how the structure of this optimal controller significantly generalizes the existing results for the risk-neutral case. Building on these results, for the matched noise and input channel case, we prove the existence-uniqueness of solution for the LEQG covariance steering problem in the neighborhood of the known risk-neutral optimal solution. We give an illustrative numerical example.

Introduction

For the standard integral quadratic cost

C:=∫01(xtTQtxt+utTRtut) dt,Qt⪰0,Rt≻0,(1)C := \int_{0}^{1} \left(x_t^{\mathsf T}Q_tx_t + u_t^{\mathsf T}R_tu_t\right)\,\mathrm{d}t,\quad Q_t \succeq0,\quad R_t \succ0, \tag*{(1)}

and fixed finite θ∈R\theta\in\mathbb{R}, we consider the linear exponential quadratic Gaussian (LEQG) covariance steering problem:

inf⁡γ∈Γ1θlog⁡E[exp⁡(θC)](2)\inf_{\gamma\in\Gamma}\frac{1}{\theta}\log\mathbb{E}\left[\exp(\theta C)\right] \tag*{(2)}

subject to

dxt=(Atxt+Btut) dt+Ft dwt,(3)\mathrm{d}x_t=(A_tx_t+B_tu_t)\,\mathrm{d}t+F_t\,\mathrm{d}w_t, \tag*{(3)}
x0∼ρ0:=N(0,Σ0),x1∼ρ1:=N(0,Σ1),(4)x_0\sim\rho_0:=\mathcal{N}(0,\Sigma_0),\qquad x_1\sim\rho_1:=\mathcal{N}(0,\Sigma_1), \tag*{(4)}

in continuous time t∈[0,1]t \in[0,1], where the state xt∈Rnx_t \in\mathbb{R}^n, the control input ut∈Rmu_t \in\mathbb{R}^m, and the standard Wiener process wt∈Rpw_t \in\mathbb{R}^p. Problem (2) seeks to design a Markovian feedback controller γ:[0,1]×Rn↦Rm\gamma:[0,1]\times\mathbb{R}^n\mapsto\mathbb{R}^m, i.e., ut=γ(t,xt)u_t=\gamma(t,x_t), such that the optimally controlled state is steered between the centered Gaussians ρ0,ρ1\rho_0,\rho_1 with given covariances Σ0,Σ1\Sigma_0,\Sigma_1 over the given time horizon [0,1][0,1].

The set of feasible control policies is

Γ:={γ:[0,1]×Rn↦Rm∣E∫01∥γ(t,xt)∥22dt<∞}.\Gamma:=\left\{\gamma:[0,1]\times\mathbb{R}^n\mapsto\mathbb{R}^m\mid\mathbb{E}\int_{0}^{1}\left\|\gamma(t,x_t)\right\|_2^2\mathrm{d}t<\infty\right\}.

where the expectation operator E[⋅]\mathbb{E}[\cdot] is taken w.r.t. the controlled state under the chosen policy γ\gamma. For a given γ∈Γ\gamma\in\Gamma, the E[⋅]\mathbb{E}[\cdot] in (2a) is w.r.t. the law of the controlled state.

There is no loss of generality in considering the endpoint Gaussians in (2c) to be centered because the solution for the nonzero-mean case is obtained by adding a time-varying drift to the optimal controller for problem (2). This time-varying drift depends on the solution of problem (2) but not the other way around; see [1], Remark 9 mutatis mutandis for θ≠0\theta\ne0. Throughout, we make the following assumptions.

A1. The trajectory tuple (At,Bt,Ft)(A_t,B_t,F_t) is bounded and continuous in t∈[0,1]t\in[0,1], and the pair (At,Bt)(A_t,B_t) is uniformly controllable over the time horizon [0,1][0,1].

A2. The matrix trajectories Qt⪰0Q_t\succeq0 and Rt≻0R_t\succ0 are bounded and continuous w.r.t. t∈[0,1]t\in[0,1].

A3. The given covariance matrices Σ0,Σ1≻0\Sigma_0,\Sigma_1\succ0.

Problem (2) can be seen as a risk-sensitive generalization of the risk-neutral linear quadratic Gaussian (LQG) covariance steering problem—the latter differs from (2) only in that the objective then is replaced by the average cost E[C]\mathbb{E}[C]. The constant θ∈R\theta\in\mathbb{R} is called the risk-sensitivity parameter. Notice that log⁡E[exp⁡(θC)]\log\mathbb{E}[\exp(\theta C)] is the logarithmic moment generation function (Log-MGF) of the random variable CC.

By L’Hôpital’s rule, lim⁡θ→01θlog⁡E[eθC]=E[C]\lim_{\theta\to0}\frac{1}{\theta}\log\mathbb{E}[e^{\theta C}]=\mathbb{E}[C], i.e., the limit θ→0\theta\to0 corresponds to the risk-neutral LQG objective. When θ>0\theta>0, the LEQG objective becomes risk-averse. When θ<0\theta<0, the objective becomes risk-seeking.

While the LQG problem minimizes the average cost E[C]\mathbb{E}[C], the LEQG objective is such that the realizations of the (random) cost CC that are above average, are penalized more than the realizations of CC that are below average. So minimizing 1θlog⁡E[eθC]\frac{1}{\theta}\log\mathbb{E}[e^{\theta C}] ensures that for θ>0\theta>0 large, a larger CC is more aggressively penalized. The motivation is that if the random variable CC has a large variance, then we prefer a more conservative controller that might be worse on average, but perform better in the worst case.

Related Works. In classical stochastic control, the solution of the LEQG problem with terminal cost is well-known via dynamic programming [2], via stochastic maximum principle [3], [4], and via completion-of-square and Radon-Nikodym derivative [5]. Notice, however, that problem (2) has no terminal cost and instead, has endpoint covariance constraints.

On the other hand, the LQG covariance steering problem (i.e., with objective E[C]\mathbb{E}[C]) was first studied in [1], [6], [7]. Specifically, the work in [1] derived the optimal controller for the case Ft=BtF_t=B_t, Qt≡0Q_t\equiv0. The developments in [6] considered different input and noise coefficient matrices (i.e., Ft≠BtF_t\ne B_t) and Qt≡0Q_t\equiv0. The work in [7] obtained the solution for Ft=BtF_t=B_t and Qt⪰0Q_t\succeq0. The purpose of our work is to initiate a generalization of these results for the risk-sensitive objective 1θlog⁡E[eθC]\frac{1}{\theta}\log\mathbb{E}[e^{\theta C}].

From a probabilistic viewpoint, problem (2) can be interpreted as a risk-sensitive variant of the Schrödinger bridge Chiran B. Cherian and Abhishek Halder are with the Department of Aerospace Engineering, Iowa State University, Ames, IA 50011, USA, {cbckbc,ahalder}@iastate.edu. Yasemin Isik is with the Department of Mathematics, Iowa State University, Ames, IA 50011, USA, [email protected].

This research was partially supported by NSF awards 2111688, 2450377. problem in the linear quadratic setting [8], [9]:

inf⁡(ρ,γ)∈P01×Γ1θlog⁡E[exp⁡(θC)](5)\inf_{(\rho,\gamma)\in\mathcal{P}_{01}\times\Gamma} \frac{1}{\theta}\log\mathbb{E}\left[\exp\left(\theta C\right)\right] \tag*{(5)}

subject to

∂tρ+∇x⋅(ρ(Atx+Btγ))=12⟨FtFtT,∇x2ρ⟩,(6)\partial_t\rho+\nabla_x\cdot\left(\rho\left(A_tx+B_t\gamma\right)\right)=\frac{1}{2}\left\langle F_tF_t^{\mathsf T},\nabla_x^2\rho\right\rangle, \tag*{(6)}

where P01\mathcal{P}_{01} is the collection of probability density function (PDF)-valued curves that are continuous in t∈[0,1]t\in[0,1], and connect the given centered Gaussian endpoints ρ0,ρ1\rho_0,\rho_1. In (3b), the symbol ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle denotes the Frobenius inner product, and ∇x⋅\nabla_x\cdot, ∇x2\nabla_x^2 denote the Euclidean divergence and Hessian, respectively. In particular, (3b) is the Fokker-Planck or Kolmogorov’s forward PDE describing the evolution of state PDF ρ(t,x)\rho(t,x) under the controlled dynamics. In this letter, we will not develop this connection (i.e., the large deviation viewpoint). Instead, we will focus on the structure and computation of the optimal solution.

Contributions. For the LEQG problem (2), we

  • deduce the optimal solution structure, and explain how this structure departs from the existing results known for the risk-neutral case (Sec. II),

  • establish existence-uniqueness of the optimal solution in the neighborhood of the risk-neutral optimal solution (Theorem 1 and Sec. III),

  • implement a numerical procedure that builds on the structural results deduced here, and illustrate the solution via a noisy double integrator example (Sec. IV).

Notations and acronyms. For natural number n>1n>1, we use Sn\mathbb{S}^n to denote the set of n×nn\times n real symmetric matrices, and S++n\mathbb{S}_{++}^n to denote its subset comprising of positive definite matrices. The notation λi\lambda_i denotes the iith eigenvalue, and ⊗\otimes denotes the Kronecker product. We use the acronyms IVP and BVP for initial and boundary value problem, respectively. We abbreviate “state transition matrix” as STM, and “with respect to” as w.r.t.

Main Results

Necessary Conditions for Optimality

Proposition 1. Consider problem (2) with fixed finite θ∈R\theta\in\mathbb{R} (equivalently problem (3)) and assumptions A1-A3. The optimal controller is

γopt(t,x)=−Rt−1BtTΠt∣θx,(7)\gamma_{\mathrm{opt}}(t,x)=-R_t^{-1}B_t^{\mathsf T}\Pi_{t\mid\theta}x, \tag*{(7)}

where (Ht∣θ,Πt∣θ)(H_{t\mid\theta},\Pi_{t\mid\theta}) jointly solve1 the coupled Riccati-Riccati matrix BVP:

−H˙t∣θ=AtTHt∣θ+Ht∣θAt+Ht∣θBtRt−1BtTHt∣θ−Qt−(Πt∣θ+Ht∣θ)(BtRt−1BtT−FtFtT)(Πt∣θ+Ht∣θ)−θΠt∣θFtFtTΠt∣θ,(8)-\dot{H}_{t\mid\theta}=A_t^{\mathsf T}H_{t\mid\theta}+H_{t\mid\theta}A_t+H_{t\mid\theta}B_tR_t^{-1}B_t^{\mathsf T}H_{t\mid\theta}-Q_t -\left(\Pi_{t\mid\theta}+H_{t\mid\theta}\right)\left(B_tR_t^{-1}B_t^{\mathsf T}-F_tF_t^{\mathsf T}\right)\left(\Pi_{t\mid\theta}+H_{t\mid\theta}\right) -\theta\Pi_{t\mid\theta}F_tF_t^{\mathsf T}\Pi_{t\mid\theta}, \tag*{(8)}
−Π˙t∣θ=AtTΠt∣θ+Πt∣θAt+Qt−Πt∣θ(BtRt−1BtT−θFtFtT)Πt∣θ,(9)-\dot{\Pi}_{t\mid\theta}=A_t^{\mathsf T}\Pi_{t\mid\theta}+\Pi_{t\mid\theta}A_t+Q_t-\Pi_{t\mid\theta}\left(B_tR_t^{-1}B_t^{\mathsf T}-\theta F_tF_t^{\mathsf T}\right)\Pi_{t\mid\theta}, \tag*{(9)}
Σ0−1=Π0∣θ+H0∣θ,Σ1−1=Π1∣θ+H1∣θ.(10)\Sigma_0^{-1}=\Pi_{0\mid\theta}+H_{0\mid\theta},\qquad\Sigma_1^{-1}=\Pi_{1\mid\theta}+H_{1\mid\theta}. \tag*{(10)}

Proof. That the optimal feedback policy is a linear state feedback follows from the standard linear-quadratic theory arguments [1], [2], [5], [7]. In particular, the necessary conditions of optimality for (2) yields an ODE BVP in the covariance state-costate pair (Σt∣θ,Πt∣θ)∈S++n×Sn(\Sigma_{t\mid\theta},\Pi_{t\mid\theta})\in\mathbb{S}_{++}^n\times\mathbb{S}^n, given by

Σ˙t∣θ=(At−BtRt−1BtTΠt∣θ)Σt∣θ+Σt∣θ(At−BtRt−1BtTΠt∣θ)T+FtFtT,(11)\dot{\Sigma}_{t\mid\theta}=\left(A_t-B_tR_t^{-1}B_t^{\mathsf T}\Pi_{t\mid\theta}\right)\Sigma_{t\mid\theta} +\Sigma_{t\mid\theta}\left(A_t-B_tR_t^{-1}B_t^{\mathsf T}\Pi_{t\mid\theta}\right)^{\mathsf T} +F_tF_t^{\mathsf T}, \tag*{(11)}
−Π˙t∣θ=AtTΠt∣θ+Πt∣θAt+Qt−Πt∣θ(BtRt−1BtT−θFtFtT)Πt∣θ,(12)-\dot{\Pi}_{t\mid\theta}=A_t^{\mathsf T}\Pi_{t\mid\theta}+\Pi_{t\mid\theta}A_t+Q_t -\Pi_{t\mid\theta}\left(B_tR_t^{-1}B_t^{\mathsf T}-\theta F_tF_t^{\mathsf T}\right)\Pi_{t\mid\theta}, \tag*{(12)}
Σ0∣θ=Σ0>0 given,Σ1∣θ=Σ1≽0 given.(13)\Sigma_{0\mid\theta}=\Sigma_0>0\ \text{given},\qquad\Sigma_{1\mid\theta}=\Sigma_1\succcurlyeq0\ \text{given}. \tag*{(13)}

The optimal controller (4) follows from the Pontryagin’s minimum principle.

Notice that (6a)-(6b) comprise a one-way coupled system of Lyapunov-Riccati ODEs subject to boundary conditions (6c) only in covariance. We use the following change-of-variable Σt∣θ∈S++n→Ht∣θ∈Sn\Sigma_{t\mid\theta}\in\mathbb{S}_{++}^n\to H_{t\mid\theta}\in\mathbb{S}^n proposed in [1]:

Ht∣θ:=Σt∣θ−1−Πt∣θ,(14)H_{t\mid\theta}:=\Sigma_{t\mid\theta}^{-1}-\Pi_{t\mid\theta}, \tag*{(14)}

to rewrite the coupled Lyapunov-Riccati BVP (6) in indeterminate pair (Σt∣θ,Πt∣θ)(\Sigma_{t\mid\theta},\Pi_{t\mid\theta}) as the Riccati-Riccati BVP (5) in indeterminate pair (Ht∣θ,Πt∣θ)(H_{t\mid\theta},\Pi_{t\mid\theta}). ■

In the risk neutral case (θ=0\theta=0), the highlighted terms in (5a)-(5b) vanish. If in addition, the noise and input channels coincide, i.e.,

FtFtT=BtRt−1BtT,(15)F_tF_t^{\mathsf T}=B_tR_t^{-1}B_t^{\mathsf T}, \tag*{(15)}

then (5) reduces to the boundary value problem in [7], eq. (2). The system of equations (4)-(5) can be seen as a risk-sensitive generalization of the same in prior work [7].

To isolate the complexities due to the noise and input channel mismatch2 from that of risk-sensitivity, in the remaining of this letter, we consider the version of (5a)-(5b) together with (8), i.e.,

−H˙t∣θ=AtTHt∣θ+Ht∣θAt+Ht∣θBtRt−1BtTHt∣θ−Qt−θΠt∣θBtRt−1BtTΠt∣θ,(16)-\dot{H}_{t\mid\theta}=A_t^{\mathsf T}H_{t\mid\theta}+H_{t\mid\theta}A_t+H_{t\mid\theta}B_tR_t^{-1}B_t^{\mathsf T}H_{t\mid\theta}-Q_t -\theta\Pi_{t\mid\theta}B_tR_t^{-1}B_t^{\mathsf T}\Pi_{t\mid\theta}, \tag*{(16)}
−Π˙t∣θ=AtTΠt∣θ+Πt∣θAt+Qt+(θ−1)Πt∣θBtRt−1BtTΠt∣θ,(17)-\dot{\Pi}_{t\mid\theta}=A_t^{\mathsf T}\Pi_{t\mid\theta}+\Pi_{t\mid\theta}A_t+Q_t +(\theta-1)\Pi_{t\mid\theta}B_tR_t^{-1}B_t^{\mathsf T}\Pi_{t\mid\theta}, \tag*{(17)}

and the boundary conditions (5c). As earlier, the highlights in (9) emphasize the terms that are new compared to the risk-neutral setting. In Sec. II-B, we will explain how these terms give rise to new mathematical challenges not seen in the risk-neutral case.

Our main contribution is to establish the existence-uniqueness of the solution of the system of equations (9) and (5c), in the neighborhood of the risk-neutral (θ=0\theta=0) solution3. To state our main result (Theorem 1), we need to parameterize the solution for (9) and (5c) in terms of certain matrix Y0∣θ∈SnY_{0\mid\theta} \in\mathbb{S}^{n} and the given risk-sensitivity parameter θ∈R\theta\in\mathbb{R}. In Sec. II-B, we detail this (implicit) parametrization.

Parametric Solution Structure

Our starting point is to write the solutions of the Riccati ODEs (9a) and (9b) in terms of the solutions of the following linear Hamiltonian ODEs:

\begin{bmatrix} \dot{\hat{X}}_{t\mid\theta}\\ \dot{\hat{Y}}_{t\mid\theta} \end{bmatrix} = \begin{bmatrix} A_t & -B_tR_t^{-1}B_t^{\mathsf{T}}\\ -(Q_t+\theta\Pi_{t\mid\theta}B_tR_t^{-1}B_t^{\mathsf{T}}\Pi_{t\mid\theta}) & -A_t^{\mathsf{T}} \end{bmatrix} \begin{bmatrix} \hat{X}_{t\mid\theta}\\ \hat{Y}_{t\mid\theta} \end{bmatrix}, \qquad =: \hat{M}_{t\mid\theta} \tag*{(18)}\end{bmatrix}
\begin{bmatrix} \dot{X}_{t\mid\theta}\\ \dot{Y}_{t\mid\theta} \end{bmatrix} = \begin{bmatrix} A_t & (\theta-1)B_tR_t^{-1}B_t^{\mathsf{T}}\\ -Q_t & -A_t^{\mathsf{T}} \end{bmatrix} \begin{bmatrix} X_{t\mid\theta}\\ Y_{t\mid\theta} \end{bmatrix}, \qquad =: M_{t\mid\theta} \tag*{(19)}\end{bmatrix}

The solutions of (9a) and (9b) admit the well-known [11, p. 156] linear fractional representations

Ht∣θ=−(X^t∣θT)−1Y^t∣θT,Πt∣θ=Yt∣θXt∣θ−1,(20)H_{t\mid\theta}=-\left(\hat{X}_{t\mid\theta}^{\mathsf{T}}\right)^{-1}\hat{Y}_{t\mid\theta}^{\mathsf{T}}, \qquad \Pi_{t\mid\theta}=Y_{t\mid\theta}X_{t\mid\theta}^{-1}, \tag*{(20)}

provided that Xt∣θ,X^t∣θX_{t\mid\theta},\hat{X}_{t\mid\theta} are invertible ∀t∈[0,1],θ∈R\forall t\in[0,1],\theta\in\mathbb{R}.

Remark 1. Unlike the risk-neutral case in [7], M^t∣θ≠Mt∣θ\hat{M}_{t\mid\theta}\ne M_{t\mid\theta} for θ≠0\theta\ne0.

For 0≤s<t≤10\le s<t\le1, let Φ^s→t∣θ,Φs→t∣θ∈R2n×2n\hat{\Phi}_{s\to t\mid\theta},\Phi_{s\to t\mid\theta}\in\mathbb{R}^{2n\times2n} denote the STMs from time ss to tt, associated with the coefficient matrices M^t∣θ,Mt∣θ\hat{M}_{t\mid\theta},M_{t\mid\theta}, respectively, i.e.,

∂∂tΦ^s→t∣θ=M^t∣θΦ^s→t∣θ,Φ^s→s∣θ=I,(21)\frac{\partial}{\partial t}\hat{\Phi}_{s\to t\mid\theta}=\hat{M}_{t\mid\theta}\hat{\Phi}_{s\to t\mid\theta}, \qquad \hat{\Phi}_{s\to s\mid\theta}=I, \tag*{(21)}
∂∂tΦs→t∣θ=Mt∣θΦs→t∣θ,Φs→s∣θ=I.(22)\frac{\partial}{\partial t}\Phi_{s\to t\mid\theta}=M_{t\mid\theta}\Phi_{s\to t\mid\theta}, \qquad \Phi_{s\to s\mid\theta}=I. \tag*{(22)}

Expressing the corresponding STMs in terms of their n×nn\times n blocks, for any t∈[0,1]t\in[0,1], we write

[X^t∣θY^t∣θ]=[Φ^0→t∣θ11Φ^0→t∣θ12Φ^0→t∣θ21Φ^0→t∣θ22]⏟Φ^0→t∣θ[X^0∣θY^0∣θ],[Xt∣θYt∣θ]=[Φ0→t∣θ11Φ0→t∣θ12Φ0→t∣θ21Φ0→t∣θ22]⏟Φ0→t∣θ[X0∣θY0∣θ].\begin{bmatrix} \hat{X}_{t\mid\theta}\\ \hat{Y}_{t\mid\theta} \end{bmatrix} = \underbrace{\begin{bmatrix} \hat{\Phi}^{11}_{0\to t\mid\theta} & \hat{\Phi}^{12}_{0\to t\mid\theta}\\ \hat{\Phi}^{21}_{0\to t\mid\theta} & \hat{\Phi}^{22}_{0\to t\mid\theta} \end{bmatrix}}_{\hat{\Phi}_{0\to t\mid\theta}} \begin{bmatrix} \hat{X}_{0\mid\theta}\\ \hat{Y}_{0\mid\theta} \end{bmatrix}, \qquad \begin{bmatrix} X_{t\mid\theta}\\ Y_{t\mid\theta} \end{bmatrix} = \underbrace{\begin{bmatrix} \Phi^{11}_{0\to t\mid\theta} & \Phi^{12}_{0\to t\mid\theta}\\ \Phi^{21}_{0\to t\mid\theta} & \Phi^{22}_{0\to t\mid\theta} \end{bmatrix}}_{\Phi_{0\to t\mid\theta}} \begin{bmatrix} X_{0\mid\theta}\\ Y_{0\mid\theta} \end{bmatrix}.

As in the proof of [7], without loss of generality, we can set X^0∣θ=X0∣θ=I\hat{X}_{0\mid\theta}=X_{0\mid\theta}=I since their initial values can be absorbed into Y^0∣θ,Y0∣θ\hat{Y}_{0\mid\theta},Y_{0\mid\theta} keeping H0∣θ,Π0∣θ∈SnH_{0\mid\theta},\Pi_{0\mid\theta}\in\mathbb{S}^{n} invariant. Then from (11), both Y^0∣θ,Y0∣θ\hat{Y}_{0\mid\theta},Y_{0\mid\theta} are symmetric. In particular, combining (7) and (11), we obtain

Y^0∣θ=Y0∣θ−Σ0−1.(23)\hat{Y}_{0\mid\theta}=Y_{0\mid\theta}-\Sigma_0^{-1}. \tag*{(23)}

Furthermore,

Xt∣θ=Φ0→t∣θ11+Φ0→t∣θ12Y0∣θ,(24)X_{t\mid\theta}=\Phi^{11}_{0\to t\mid\theta}+\Phi^{12}_{0\to t\mid\theta}Y_{0\mid\theta}, \tag*{(24)}
Yt∣θ=Φ0→t∣θ21+Φ0→t∣θ22Y0∣θ,(25)Y_{t\mid\theta}=\Phi^{21}_{0\to t\mid\theta}+\Phi^{22}_{0\to t\mid\theta}Y_{0\mid\theta}, \tag*{(25)}

and hence from (11), the matrix Πt∣θ∈Sn\Pi_{t\mid\theta}\in\mathbb{S}^{n} is determined by the choice of Y0∣θ∈SnY_{0\mid\theta}\in\mathbb{S}^{n}.

A notable departure from the risk-neutral case is that the matrix M^t∣θ\hat{M}_{t\mid\theta} in (10a) depends explicitly on Πt∣θ∈Sn\Pi_{t\mid\theta}\in\mathbb{S}^{n}, and therefore on Y0∣θY_{0\mid\theta}. Hence, the blocks of Φ^0→t∣θ\hat{\Phi}_{0\to t\mid\theta} should be viewed as functions of both θ\theta and Y0∣θY_{0\mid\theta}. In contrast, the blocks of Φ0→t∣θ\Phi_{0\to t\mid\theta} are functions of θ\theta, but do not depend on Y0∣θY_{0\mid\theta}.

Particularizing the above observations for t=1t=1, we find

H1∣θ=−(X^1∣θT)−1Y^1∣θT=−((Φ^0→1∣θ11)T+Y^0∣θ(Φ^0→1∣θ12)T)−1((Φ^0→1∣θ21)T+Y^0∣θ(Φ^0→1∣θ22)T),(26)H_{1\mid\theta} = -\left(\hat{X}_{1\mid\theta}^{\mathsf{T}}\right)^{-1}\hat{Y}_{1\mid\theta}^{\mathsf{T}} = -\left(\left(\hat{\Phi}^{11}_{0\to1\mid\theta}\right)^{\mathsf{T}}+\hat{Y}_{0\mid\theta}\left(\hat{\Phi}^{12}_{0\to1\mid\theta}\right)^{\mathsf{T}}\right)^{-1} \left(\left(\hat{\Phi}^{21}_{0\to1\mid\theta}\right)^{\mathsf{T}}+\hat{Y}_{0\mid\theta}\left(\hat{\Phi}^{22}_{0\to1\mid\theta}\right)^{\mathsf{T}}\right), \tag*{(26)}
Π1∣θ=Y1∣θX1∣θ−1=(Φ0→1∣θ21+Φ0→1∣θ22Y0∣θ)(Φ0→1∣θ11+Φ0→1∣θ12Y0∣θ)−1.(27)\Pi_{1\mid\theta}=Y_{1\mid\theta}X_{1\mid\theta}^{-1} = \left(\Phi^{21}_{0\to1\mid\theta}+\Phi^{22}_{0\to1\mid\theta}Y_{0\mid\theta}\right) \left(\Phi^{11}_{0\to1\mid\theta}+\Phi^{12}_{0\to1\mid\theta}Y_{0\mid\theta}\right)^{-1}. \tag*{(27)}

Since (13) allows us to write Y^0∣θ\hat{Y}_{0\mid\theta} in terms of Y0∣θY_{0\mid\theta}, we view (15a)-(15b) as functions of θ\theta and Y0∣θY_{0\mid\theta}. Therefore, using (7), we arrive at an implicit algebraic equation in indeterminate Y0∣θY_{0\mid\theta} given by

G(θ,Y0∣θ):=Π1∣θ+H1∣θ−Σ1−1=0.(28)G(\theta,Y_{0\mid\theta}):=\Pi_{1\mid\theta}+H_{1\mid\theta}-\Sigma_1^{-1}=0. \tag*{(28)}

If a matrix Y0∣θ∈SnY_{0\mid\theta}\in\mathbb{S}^{n} solving (16) exists and is unique in a suitable sense, then it determines Πt∣θ\Pi_{t\mid\theta} (via (11)), and thus the optimal control (4). However, an explicit solution of (16) in the form Y0∣θ=Y0∣θ(θ)Y_{0\mid\theta}=Y_{0\mid\theta}(\theta) is not available in general.

In the risk-neutral case (θ=0\theta=0), M^t∣0=Mt∣0\hat{M}_{t\mid0}=M_{t\mid0}, so the corresponding STMs match4{}^{4}, i.e., Φ^s→t∣0=Φs→t∣0\hat{\Phi}_{s\to t\mid0}=\Phi_{s\to t\mid0}. This STM matching plays a crucial role in the solution derived in [7]. In that case, combining the terminal boundary condition Σ1−1=Π1∣0+H1∣0\Sigma_1^{-1}=\Pi_{1\mid0}+H_{1\mid0} with (15), and using (13), the symmetric identities in [7, Lemma 3] turn (16) into a quadratic equation in unknown Y0∣0Y_{0\mid0}. In the proof for [7], the two roots (Y0∣0+,Y0∣0−)(Y^+_{0\mid0},Y^-_{0\mid0}) of this quadratic equation found via completion-of-square, were

Y0∣0±=12Σ0−1−(Φ0→112)−1Φ0→111±Σ0−12(14I+Σ012(Φ0→112)−1)−1Σ1(Φ0→112)−TΣ0−12.(29)Y^\pm_{0\mid0} = \frac{1}{2}\Sigma_0^{-1} - \left(\Phi^{12}_{0\to1}\right)^{-1}\Phi^{11}_{0\to1} \pm \Sigma_0^{-\frac{1}{2}} \left( \frac{1}{4}I+\Sigma_0^{\frac{1}{2}} \left(\Phi^{12}_{0\to1}\right)^{-1} \right)^{-1} \Sigma_1 \left(\Phi^{12}_{0\to1}\right)^{-\mathsf{T}} \Sigma_0^{-\frac{1}{2}}. \tag*{(29)}

In that work, it was shown that only one of these roots, namely Y0∣0−Y^-_{0\mid0}, ensures invertibility4 of Xt∣0,X^t∣0 ∀t∈[0,1]X_{t\mid0},\hat{X}_{t\mid0}\ \forall t\in[0,1]. By (11), this in turn ensures the unique admissible solution of (9) for θ=0\theta=0.

In our θ≠0\theta\ne0 case, per Remark 1, Φ^s→t∣θ≠Φs→t∣θ\hat{\Phi}_{s\to t\mid\theta}\ne\Phi_{s\to t\mid\theta}, and the mixed STM terms no longer satisfy the same symmetric identities. Consequently, we have to contend with the implicit algebraic equation (16) in Y0∣θY_{0|\theta}. Now we are ready to state our main result.

Theorem 1. Consider problem (2) with fixed finite θ∈R\theta\in\mathbb{R}, assumptions A1-A3, and the matched channel condition (8). Let Y0∣0−Y_{0|0}^{-} denote the unique admissible solution of (16) in the risk-neutral LQG covariance steering problem. Then there exist open sets I⊂R\mathcal{I} \subset\mathbb{R}, U⊂Sn\mathcal{U} \subset\mathbb{S}^{n} such that (0,Y0∣0−)∈I×U(0,Y_{0|0}^{-}) \in\mathcal{I} \times\mathcal{U}, and a unique C1C^{1} mapping h:I→Uh:\mathcal{I}\to\mathcal{U} such that h(0)=Y0∣0−h(0)=Y_{0|0}^{-} and G(θ,h(θ))=0 ∀θ∈IG(\theta,h(\theta))=0\ \forall\theta\in\mathcal{I}. In other words, the LEQG problem (2) has a local C1C^{1} solution branch6 through Y0∣0−Y_{0|0}^{-}.

In Section III that follows, we prove Theorem 1 using several auxiliary results (Proposition 2, Lemma 1, Proposition 3) derived therein.

Proofs and Auxiliary Results

Our results in this Section are organized and interconnected as in the block diagram below.

Proposition 2. Consider GG in (16) as a nonlinear map G(θ,Y0∣θ):I×U↦VG(\theta,Y_{0|\theta}):\mathcal{I}\times\mathcal{U}\mapsto\mathcal{V} where I×U\mathcal{I}\times\mathcal{U} and V\mathcal{V} are bounded open subsets of R×Sn\mathbb{R}\times\mathbb{S}^{n} and Sn\mathbb{S}^{n}, respectively. Then G∈C1(I×U;V)G\in C^{1}(\mathcal{I}\times\mathcal{U};\mathcal{V}).

Proof. We establish the desired regularity of GG w.r.t. the tuple (θ,Y0∣θ)(\theta,Y_{0|\theta}) by following the dependence structure of the quantities entering the definition of GG: (θ,Y0∣θ)↦Πt∣θ↦Ht∣θ↦G(θ,Y0∣θ)(\theta,Y_{0|\theta})\mapsto\Pi_{t|\theta}\mapsto H_{t|\theta}\mapsto G(\theta,Y_{0|\theta}).

All matrix spaces involved are finite-dimensional and may therefore be identified with Euclidean spaces.

Consider the Riccati ODE (9b) in indeterminate Πt∣θ\Pi_{t|\theta} with initial condition Π0∣θ=Y0∣θ\Pi_{0|\theta}=Y_{0|\theta}, where θ\theta serves as a parameter. The corresponding vector field is continuous in tt, and C1C^{1} w.r.t. (Πt∣θ,θ)(\Pi_{t|\theta},\theta) since it comprises of continuous and bounded in time AtA_t, BtB_t, QtQ_t (Assumptions A1-A2), together with finite sums and products involving Πt∣θ\Pi_{t|\theta} and θ\theta. Therefore, by the smoothness theorem for parameter-dependent ODE IVP [12, Chapter V, Theorem 3.1], Πt∣θ\Pi_{t|\theta} is jointly C1C^{1} w.r.t. (θ,Y0∣θ)(\theta,Y_{0|\theta}).

Next, consider the Riccati ODE (9a) in indeterminate Ht∣θH_{t|\theta} with initial condition (see (7)) H0∣θ=Σ0−1−Y0∣θH_{0|\theta}=\Sigma_{0}^{-1}-Y_{0|\theta} (which is affine, thus C1C^{1} in Y0∣θY_{0|\theta}). Since Πt∣θ\Pi_{t|\theta} is jointly C1C^{1} w.r.t. (θ,Y0∣θ)(\theta,Y_{0|\theta}), the vector field of (9a) is continuous in tt, and C1C^{1} w.r.t. (Ht∣θ,θ,Y0∣θ)(H_{t|\theta},\theta,Y_{0|\theta}). By the same smoothness theorem [12, Chapter V, Theorem 3.1], Ht∣θH_{t|\theta} is jointly C1C^{1} w.r.t. (θ,Y0∣θ)(\theta,Y_{0|\theta}).

Since evaluation at the fixed terminal time t=1t=1 preserves C1C^{1}-regularity, both Π1∣θ,H1∣θ\Pi_{1|\theta},H_{1|\theta} are C1C^{1} in (θ,Y0∣θ)(\theta,Y_{0|\theta}). The matrix Σ1−1\Sigma_{1}^{-1} is constant. Therefore, G∈C1(I×U;V)G\in C^{1}(\mathcal{I}\times\mathcal{U};\mathcal{V}). ■

Lemma 1. At θ=0\theta=0, the partial Jacobian

DY0∣0G=X1∣0−T⊗X1∣0−T−X^1∣0−T⊗X^1∣0−T.(30)D_{Y_{0|0}}G=X_{1|0}^{-\mathsf{T}}\otimes X_{1|0}^{-\mathsf{T}}-\hat{X}_{1|0}^{-\mathsf{T}}\otimes\hat{X}_{1|0}^{-\mathsf{T}}. \tag*{(30)}

Proof. From (16), we have

dG(0,Y0∣0)=dΠ1∣0+dH1∣0.(31)\mathrm{d}G(0,Y_{0|0})=\mathrm{d}\Pi_{1|0}+\mathrm{d}H_{1|0}. \tag*{(31)}

Using (15b),

dΠ1∣0=d(Y1∣0X1∣0−1)=(dY1∣0)X1∣0−1−Y1∣0X1∣0−1(dX1∣0)X1∣0−1=(Φ0→1∣022dY0∣0−Π1∣0(Φ0→1∣012dY0∣0))X1∣0−1=(Φ0→1∣022−Π1∣0Φ0→1∣012)(dY0∣0)X1∣0−1.(32)\begin{aligned} \mathrm{d}\Pi_{1|0} &=\mathrm{d}\left(Y_{1|0}X_{1|0}^{-1}\right)\\ &=(\mathrm{d}Y_{1|0})X_{1|0}^{-1}-Y_{1|0}X_{1|0}^{-1}(\mathrm{d}X_{1|0})X_{1|0}^{-1}\\ &=\left(\Phi_{0\to1|0}^{22}\mathrm{d}Y_{0|0}-\Pi_{1|0}\left(\Phi_{0\to1|0}^{12}\mathrm{d}Y_{0|0}\right)\right)X_{1|0}^{-1}\\ &=\left(\Phi_{0\to1|0}^{22}-\Pi_{1|0}\Phi_{0\to1|0}^{12}\right)(\mathrm{d}Y_{0|0})X_{1|0}^{-1}. \tag*{(32)} \end{aligned}

Since Π1∣0∈Sn\Pi_{1|0}\in\mathbb{S}^{n}, we have Π1∣0=Π1∣0T=X1∣0−TY1∣0T\Pi_{1|0}=\Pi_{1|0}^{\mathsf{T}}=X_{1|0}^{-\mathsf{T}}Y_{1|0}^{\mathsf{T}}. Let A:=Φ0→1∣022−Π1∣0Φ0→1∣012=Φ0→1∣022−X1∣0−TY1∣0TΦ0→1∣012A:=\Phi_{0\to1|0}^{22}-\Pi_{1|0}\Phi_{0\to1|0}^{12}=\Phi_{0\to1|0}^{22}-X_{1|0}^{-\mathsf{T}}Y_{1|0}^{\mathsf{T}}\Phi_{0\to1|0}^{12}. Left-multiplying AA by X1∣0T=(Φ0→1∣011)T+Y0∣0(Φ0→1∣012)TX_{1|0}^{\mathsf{T}}=(\Phi_{0\to1|0}^{11})^{\mathsf{T}}+Y_{0|0}(\Phi_{0\to1|0}^{12})^{\mathsf{T}}, we obtain

X1∣0TA=(Φ0→1∣011)TΦ0→1∣022+Y0∣0(Φ0→1∣012)TΦ0→1∣022−(Φ0→1∣021)TΦ0→1∣012−Y0∣0(Φ0→1∣012)TΦ0→1∣012=(Φ0→1∣011)TΦ0→1∣022−(Φ0→1∣021)TΦ0→1∣012=I,\begin{aligned} X_{1|0}^{\mathsf{T}}A &=(\Phi_{0\to1|0}^{11})^{\mathsf{T}}\Phi_{0\to1|0}^{22} +Y_{0|0}(\Phi_{0\to1|0}^{12})^{\mathsf{T}}\Phi_{0\to1|0}^{22}\\ &\quad-(\Phi_{0\to1|0}^{21})^{\mathsf{T}}\Phi_{0\to1|0}^{12} -Y_{0|0}(\Phi_{0\to1|0}^{12})^{\mathsf{T}}\Phi_{0\to1|0}^{12}\\ &=(\Phi_{0\to1|0}^{11})^{\mathsf{T}}\Phi_{0\to1|0}^{22} -(\Phi_{0\to1|0}^{21})^{\mathsf{T}}\Phi_{0\to1|0}^{12}=I, \end{aligned}

where the second equality is due to the symplectic identity [7, eq. (7b)], and the third is due to the symplectic identity [7, eq. (7a)]. Thus A=X1∣0−TA=X_{1|0}^{-\mathsf{T}}, which simplifies (21) to

dΠ1∣0=X1∣0−T(dY0∣0)X1∣0−1.(33)\mathrm{d}\Pi_{1|0}=X_{1|0}^{-\mathsf{T}}(\mathrm{d}Y_{0|0})X_{1|0}^{-1}. \tag*{(33)}

Starting from (15a), similar computation7 yields

dH1∣0=−X^1∣0−T(dY0∣0)X^1∣0−1.(34)\mathrm{d}H_{1|0}=-\hat{X}_{1|0}^{-\mathsf{T}}(\mathrm{d}Y_{0|0})\hat{X}_{1|0}^{-1}. \tag*{(34)}

Substituting dΠ1∣0\mathrm{d}\Pi_{1|0} and dH1∣0\mathrm{d}H_{1|0} from (21) and (22) back into (19), we obtain

dG(0,Y0∣0)=X1∣0−T(dY0∣0)X1∣0−1−X^1∣0−T(dY0∣0)X^1∣0−1.(35)\mathrm{d}G(0,Y_{0|0})=X_{1|0}^{-\mathsf{T}}(\mathrm{d}Y_{0|0})X_{1|0}^{-1}-\hat{X}_{1|0}^{-\mathsf{T}}(\mathrm{d}Y_{0|0})\hat{X}_{1|0}^{-1}. \tag*{(35)}

Applying the vectorization (vec) operator to both sides of (23), and using the identity vec⁡(PQR)=(RT⊗P)vec⁡(Q)\operatorname{vec}(PQR)=(R^{\mathsf{T}}\otimes P)\operatorname{vec}(Q), we get

vec⁡ dG(0,Y0∣0)=[(X1∣0−T⊗X1∣0−T)−(X^1∣0−T⊗X^1∣0−T)]vec⁡ dY0∣0.(36)\operatorname{vec}\,\mathrm{d}G(0,Y_{0|0}) =\left[\left(X_{1|0}^{-\mathsf{T}}\otimes X_{1|0}^{-\mathsf{T}}\right)-\left(\hat{X}_{1|0}^{-\mathsf{T}}\otimes\hat{X}_{1|0}^{-\mathsf{T}}\right)\right]\operatorname{vec}\,\mathrm{d}Y_{0|0}. \tag*{(36)}

From (24), using the Jacobian identification rule [13, p. 199], we arrive at (18). ■

Remark 2. A tedious computation shows that at θ≠0\theta\ne0,

DY0∣θG=X1∣θ−T⊗X1∣θ−T−X^1∣θ−T⊗X^1∣θ−T+θ∫01(Us∣θ⊗Vs∣θ+Vs∣θ⊗Us∣θ) ds.(37)\begin{aligned} D_{Y_{0|\theta}}G &=X_{1|\theta}^{-\mathsf{T}}\otimes X_{1|\theta}^{-\mathsf{T}} -\hat{X}_{1|\theta}^{-\mathsf{T}}\otimes\hat{X}_{1|\theta}^{-\mathsf{T}}\\ &\quad+\theta\int_{0}^{1}\left(U_{s|\theta}\otimes V_{s|\theta} +V_{s|\theta}\otimes U_{s|\theta}\right)\,\mathrm{d}s. \tag*{(37)} \end{aligned}

where

Us∣θ:=X^1∣θ−⊤X^s∣θ⊤Xs∣θ−⊤,(38)U_{s\mid\theta}:=\widehat{X}_{1\mid\theta}^{-\top}\widehat{X}_{s\mid\theta}^{\top}X_{s\mid\theta}^{-\top}, \tag*{(38)}
Vs∣θ:=X^1∣θ−⊤X^s∣θ⊤Πs∣θBsRs−1Bs⊤Xs∣θ−⊤.(39)V_{s\mid\theta}:=\widehat{X}_{1\mid\theta}^{-\top}\widehat{X}_{s\mid\theta}^{\top}\Pi_{s\mid\theta}B_sR_s^{-1}B_s^{\top}X_{s\mid\theta}^{-\top}. \tag*{(39)}

While (25) generalizes (18), we will not use this result in the proofs that follow.

Proposition 3. The partial Jacobian DY0∣0GD_{Y_{0\mid0}}G evaluated at Y0∣0−Y_{0\mid0}^{-} in (17), is nonsingular.

Proof. For convenience, let

Ψ01:=Σ0−12(14I+Σ012(Φ0→112)−1Σ1(Φ0→112)−⊤Σ012)12Σ0−12,(40)\Psi_{01}:=\Sigma_0^{-\frac{1}{2}}\left(\frac{1}{4}I+\Sigma_0^{\frac{1}{2}}(\Phi_{0\to1}^{12})^{-1}\Sigma_1(\Phi_{0\to1}^{12})^{-\top}\Sigma_0^{\frac{1}{2}}\right)^{\frac{1}{2}}\Sigma_0^{-\frac{1}{2}}, \tag*{(40)}
S0:=12Σ0−1.(41)S_0:=\frac{1}{2}\Sigma_0^{-1}. \tag*{(41)}

As explained in Sec. II-B, X^0∣θ=X0∣θ=I ∀θ∈[0,1]\widehat{X}_{0\mid\theta}=X_{0\mid\theta}=I\ \forall\theta\in[0,1], so the matrices X^1∣0=Φ0→1∣011+Φ0→1∣012Y^0∣0−\widehat{X}_{1\mid0}=\Phi_{0\to1\mid0}^{11}+\Phi_{0\to1\mid0}^{12}\widehat{Y}_{0\mid0}^{-} and X1∣0=Φ0→1∣011+Φ0→1∣012Y0∣0−X_{1\mid0}=\Phi_{0\to1\mid0}^{11}+\Phi_{0\to1\mid0}^{12}Y_{0\mid0}^{-} simplify to

X^1∣0=X1∣0−Φ0→1∣012Σ0−1=−Φ0→1∣012(Ψ01+S0),(42)\widehat{X}_{1\mid0}=X_{1\mid0}-\Phi_{0\to1\mid0}^{12}\Sigma_0^{-1}=-\Phi_{0\to1\mid0}^{12}(\Psi_{01}+S_0), \tag*{(42)}
X1∣0=−Φ0→1∣012(Ψ01−S0),(43)X_{1\mid0}=-\Phi_{0\to1\mid0}^{12}(\Psi_{01}-S_0), \tag*{(43)}

where we have used (13) and (17).

Per Assumption A3, both Ψ01,S0\Psi_{01},S_0 in (27) are strictly positive definite. In particular, the positive definiteness of Ψ01\Psi_{01} follows from that

  • a positive definite matrix has unique (principal) positive definite square root,

  • Φ0→112\Phi_{0\to1}^{12} is invertible [7],

  • the matrix Σ012(Φ0→112)−1Σ1(Φ0→112)−⊤Σ012≻0\Sigma_0^{\frac{1}{2}}(\Phi_{0\to1}^{12})^{-1}\Sigma_1(\Phi_{0\to1}^{12})^{-\top}\Sigma_0^{\frac{1}{2}}\succ0 since it is congruent [14] to Σ1≻0\Sigma_1\succ0 via congruence transform defined by invertible matrix Σ012(Φ0→112)−1\Sigma_0^{\frac{1}{2}}(\Phi_{0\to1}^{12})^{-1}.

Hence, Ψ01+S0≻0\Psi_{01}+S_0\succ0. Furthermore,

λi(14I+Σ012(Φ0→112)−1Σ1(Φ0→112)−⊤Σ012)>14∀i∈[n]\lambda_i\left(\frac{1}{4}I+\Sigma_0^{\frac{1}{2}}(\Phi_{0\to1}^{12})^{-1}\Sigma_1(\Phi_{0\to1}^{12})^{-\top}\Sigma_0^{\frac{1}{2}}\right)>\frac{1}{4}\quad\forall i\in[n]

implies that

Ψ01−S0=Σ0−12((14I+Σ012(Φ0→112)−1Σ1(Φ0→112)−⊤Σ012)12−12I)Σ0−12≻0.(44)\Psi_{01}-S_0=\Sigma_0^{-\frac{1}{2}}\left(\left(\frac{1}{4}I+\Sigma_0^{\frac{1}{2}}(\Phi_{0\to1}^{12})^{-1}\Sigma_1(\Phi_{0\to1}^{12})^{-\top}\Sigma_0^{\frac{1}{2}}\right)^{\frac{1}{2}}-\frac{1}{2}I\right)\Sigma_0^{-\frac{1}{2}}\succ0. \tag*{(44)}

From (28), taking the inverse transposes, we get

X^1∣0−⊤=−(Φ0→1∣012)−⊤(Ψ01+S0)−1,(45)\widehat{X}_{1\mid0}^{-\top}=-(\Phi_{0\to1\mid0}^{12})^{-\top}(\Psi_{01}+S_0)^{-1}, \tag*{(45)}
X1∣0−⊤=−(Φ0→1∣012)−⊤(Ψ01−S0)−1.(46)X_{1\mid0}^{-\top}=-(\Phi_{0\to1\mid0}^{12})^{-\top}(\Psi_{01}-S_0)^{-1}. \tag*{(46)}

Substituting (30) in (18) from Lemma 1, and using Kronecker product properties, we obtain

DY0∣0G∣Y0∣0=Y0∣0−=((Φ0→112)−⊤⊗(Φ0→112)−⊤)Ω,(47)\left.D_{Y_{0\mid0}}G\right|_{Y_{0\mid0}=Y_{0\mid0}^{-}}=((\Phi_{0\to1}^{12})^{-\top}\otimes(\Phi_{0\to1}^{12})^{-\top})\Omega, \tag*{(47)}

where

Ω:=((Ψ01−S0)−1⊗(Ψ01−S0)−1)−((Ψ01+S0)−1⊗(Ψ01+S0)−1).(48)\Omega:=((\Psi_{01}-S_0)^{-1}\otimes(\Psi_{01}-S_0)^{-1})-((\Psi_{01}+S_0)^{-1}\otimes(\Psi_{01}+S_0)^{-1}). \tag*{(48)}

Since Φ0→112\Phi_{0\to1}^{12} is invertible, (31) is nonsingular if and only if det⁡Ω≠0\det\Omega\ne0.

Since S0≻0S_0\succ0 and Ψ01\Psi_{01} are real symmetric, by the simultaneous diagonalizability theorem [14], [15], there exists nonsingular PP such that

P⊤S0P=I,P⊤Ψ01P=diag⁡(ψ1,…,ψn),(49)P^{\top}S_0P=I,\qquad P^{\top}\Psi_{01}P=\operatorname{diag}(\psi_1,\ldots,\psi_n), \tag*{(49)}

where ψi>0 ∀i∈[n]\psi_i>0\ \forall i\in[n] denote the iith eigenvalue of Ψ01≻0\Psi_{01}\succ0. Since Ψ01−S0≻0\Psi_{01}-S_0\succ0 (see (29)), the simultaneous diagonalizability (33) implies the stronger: ψi>1 ∀i∈{1,…,n}\psi_i>1\ \forall i\in\{1,\ldots,n\}. Then the eigenvalues of (Ψ01−S0)−1(\Psi_{01}-S_0)^{-1} and (Ψ01+S0)−1(\Psi_{01}+S_0)^{-1} are 1ψi−1\frac{1}{\psi_i-1} and 1ψi+1\frac{1}{\psi_i+1}, respectively. So the spectrum of Ω\Omega in (32) comprises of

1(ψi−1)(ψj−1)−1(ψi+1)(ψj+1)=2(ψi+ψj)(ψi2−1)(ψj2−1)∀(i,j)∈[n]×[n],\frac{1}{(\psi_i-1)(\psi_j-1)}-\frac{1}{(\psi_i+1)(\psi_j+1)} =\frac{2(\psi_i+\psi_j)}{(\psi_i^2-1)(\psi_j^2-1)} \qquad\forall(i,j)\in[n]\times[n],

which are positive because ψi>1 ∀i∈[n]\psi_i>1\ \forall i\in[n]. This proves Ω≻0\Omega\succ0, and in particular det⁡Ω≠0\det\Omega\ne0. Hence (31) is nonsingular. ■

Proof of Theorem 1. Specializing Proposition 2 at Y0∣0=Y0∣0−Y_{0\mid0}=Y_{0\mid0}^{-}, the mapping GG is C1C^1 in a neighborhood of (0,Y0∣0−)(0,Y_{0\mid0}^{-}). Moreover, since Y0∣0−Y_{0\mid0}^{-} is the unique admissible risk-neutral solution, from (16), we have that G(0,Y0∣0−)=0G(0,Y_{0\mid0}^{-})=0. By Proposition 3, the partial Jacobian DY0∣0GD_{Y_{0\mid0}}G evaluated at Y0∣0−Y_{0\mid0}^{-}, is nonsingular. Hence, the implicit function theorem [16] applies at (0,Y0∣0−)(0,Y_{0\mid0}^{-}). Therefore, there exist an open interval I⊂R\mathcal{I}\subset\mathbb{R} containing θ=0\theta=0, a neighborhood U⊂Sn\mathcal{U}\subset\mathbb{S}_n containing Y0∣0−Y_{0\mid0}^{-}, and a unique C1C^1 mapping h:I→Uh:\mathcal{I}\to\mathcal{U} such that h(0)=Y0∣0−h(0)=Y_{0\mid0}^{-} and G(θ,h(θ))=0 ∀θ∈IG(\theta,h(\theta))=0\ \forall\theta\in\mathcal{I}. This completes the proof. ■

Numerical Example

As an illustrative example, we implemented the solution of the LEQG covariance steering over a noisy double integrator, i.e., an instance of problem (2) with

(At,Bt)≡([0100],[01]),Bt≡Ft, Qt=Rt=I ∀t∈[0,1],(A_t,B_t)\equiv\left(\begin{bmatrix}0&1\\0&0\end{bmatrix},\begin{bmatrix}0\\1\end{bmatrix}\right),\quad B_t\equiv F_t,\ Q_t=R_t=I\ \forall t\in[0,1],

and randomly generated endpoint covariances

Σ0=[3.3356−0.0796−0.07962.0147],Σ1=[6.9230−2.0741−2.07413.7050].\Sigma_0=\begin{bmatrix}3.3356&-0.0796\\-0.0796&2.0147\end{bmatrix},\qquad \Sigma_1=\begin{bmatrix}6.9230&-2.0741\\-2.0741&3.7050\end{bmatrix}.

These problem data satisfy assumptions A1-A3.

With these problem data, we used Newton’s method with backtracking line search to solve the implicit nonlinear equation (16) in unknown Y0∣θY_{0\mid\theta}. Starting with Y0∣0Y_{0\mid0} in (17) as an initial guess for Y0∣θY_{0\mid\theta}, we solved the IVP defined by (9b) and Π0∣θ=Y0∣θ\Pi_{0\mid\theta}=Y_{0\mid\theta}, followed by the IVP defined by (9a) and H0∣θ=Σ0−1−Π0∣θH_{0\mid\theta}=\Sigma_0^{-1}-\Pi_{0\mid\theta}. Due to the one-way coupling among (9a)-(9b), the latter IVP solution uses the trajectory Πt∣θ\Pi_{t\mid\theta} from the former IVP. For the Newton update, we used the Jacobian (37). Figure 1 shows the steering of the optimally controlled state covariance Σt∣θ\Sigma_{t\mid\theta} for θ=−1,0,1\theta= -1,0,1.

Block diagram showing the relationships among Proposition 2, Lemma 1, Proposition 3, and Theorem 1

Figure 1.

Figure 1. LEQG steering of the state covariance Σt∣θ\Sigma_{t\mid\theta}, shown as respective 2σ2\sigma ellipses for t∈[0,1]t \in[0,1], θ∈{−1,0,1}\theta\in\{-1,0,1\}, and state domain [−6,6]×[−4,4][-6,6] \times[-4,4]. All ellipses are centered at the origin.

As θ\theta increases from negative to positive, the optimal solution transitions from risk-seeking to risk-averse. Consequently, log⁡det⁡Σt∣θ\log\det\Sigma_{t\mid\theta} for any fixed tt, is a decreasing function of θ\theta. Figure 2 shows this hedging against uncertainty for the optimal solution as θ\theta increases from −1-1 to +1+1 with stepsize 0.1. Figure 3 also captures the same in that the optimal cost is shown to be increasing w.r.t. θ\theta. Notice in particular that beyond the apparent scaling, the optimal 1θlog⁡E[eθC]\frac{1}{\theta}\log\mathbb{E}[e^{\theta C}] depends on θ\theta because the random variable CC in (1) depends on Πt∣θ\Pi_{t\mid\theta} via (4), and also because the E[⋅]\mathbb{E}[\cdot] therein is taken w.r.t. the controlled joint state distribution N(0,Σt∣θ)\mathcal{N}(0,\Sigma_{t\mid\theta}).

Risk-sensitivity parameter ($\theta$) versus the log-determinant of the optimally controlled state covariance for $t = 0.3, 0.6, 0.9$.

Figure 2. Risk-sensitivity parameter (θ\theta) versus the log-determinant of the optimally controlled state covariance for t=0.3,0.6,0.9t = 0.3, 0.6, 0.9.

Risk-sensitivity parameter ($\theta$) versus the optimal cost (2a).

Figure 3. Risk-sensitivity parameter (θ\theta) versus the optimal cost (2a).

Concluding Remarks

The purpose of this letter was to introduce the LEQG covariance steering problem and to explain how its solution generalizes the results known in the literature, emphasizing new structural aspects of the solution. While Theorem 1 proved local existence-uniqueness of solution, we suspect

this result can be made global by upgrading (18) with (25), and then showing that hh can be extended in C1C^1 to the boundary of any finite interval. This remains ongoing work. All results and writing (including the em dashes in the abstract) are due to the three human authors, not by AI.

References

  1. [1]Y. Chen, T. T. Georgiou, and M. Pavon, “Optimal steering of a linear stochastic system to a final probability distribution, Part I,” IEEE Transactions on Automatic Control, vol. 61, no. 5, pp. 1158–1169, 2015.arxiv.org/abs/1408.2222
  2. [2]D. Jacobson, “Optimal stochastic linear systems with exponential performance criteria and their relation to deterministic differential games,” IEEE Transactions on Automatic Control, vol. 18, no. 2, pp. 124–131, 1973.DOI
  3. [3]P. Whittle, “A risk-sensitive maximum principle,” Systems & Control Letters, vol. 15, no. 3, pp. 183–192, 1990.
  4. [4]A. E. Lim and X. Y. Zhou, “A new risk-sensitive maximum principle,” IEEE Transactions on Automatic Control, vol. 50, no. 7, pp. 958–966, 2005.DOI
  5. [5]T. E. Duncan, “Linear-exponential-quadratic Gaussian control,” IEEE Transactions on Automatic Control, vol. 58, no. 11, pp. 2910–2911, 2013.DOI
  6. [6]Y. Chen, T. T. Georgiou, and M. Pavon, “Optimal steering of a linear stochastic system to a final probability distribution, Part II,” IEEE Transactions on Automatic Control, vol. 61, no. 5, pp. 1170–1180, 2015.arxiv.org/abs/1410.3447
  7. [7]Y. Chen, T. T. Georgiou, and M. Pavon, “Optimal steering of a linear stochastic system to a final probability distribution—Part III,” IEEE Transactions on Automatic Control, vol. 63, no. 9, pp. 3112–3118, 2018.
  8. [8]Y. Chen, T. T. Georgiou, and M. Pavon, “Stochastic control liaisons: Richard Sinkhorn meets Gaspard Monge on a Schrödinger bridge,” Siam Review, vol. 63, no. 2, pp. 249–313, 2021.arxiv.org/abs/2005.10963
  9. [9]A. M. H. Teter, W. Wang, and A. Halder, “Schrödinger bridge with quadratic state cost is exactly solvable,” IEEE Transactions on Automatic Control, vol. 71, no. 5, pp. 2903–2917, 2026.
  10. [10]G. A. Bondar, A. Eldesoukey, Y. Chen, and A. Halder, “Nonlinear non-Gaussian density steering with input and noise channel mismatch: Sinkhorn with memory for solving the control-affine Schrödinger bridge problem,” arXiv preprint arXiv:2604.23370, 2026.arxiv.org/abs/2604.23370
  11. [14]R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed. Cambridge university press, 2012.
  12. [15]V. V. Prasolov, Problems and Theorems in Linear Algebra, ser. Translations of Mathematical Monographs. Providence, Rhode Island: American Mathematical Society, 1994, vol. 134.DOI
  13. [16]C. H. Edwards, Advanced Calculus of Several Variables. Courier Corporation, 1994.

Paper details

vlnc:W_sxtxnzqm
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents