Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)
Papers
Topics
Authors
Recent
Search
2000 character limit reached

Risk-Sensitive Stochastic Optimal Control

Updated 6 July 2026
  • Risk-sensitive stochastic optimal control is a framework that integrates risk measures (e.g., CVaR, exponential-of-cost criteria) into stochastic systems to address variability and tail risks.
  • It employs techniques such as dynamic programming, nonlinear HJB equations, and Pontryagin’s maximum principle to derive optimal policies under uncertainty.
  • The approaches are applied in wireless scheduling, robotic control, financial optimization, and mean-field games to ensure robust performance in uncertain environments.

Searching arXiv for recent and foundational papers on risk-sensitive stochastic optimal control. Risk-sensitive stochastic optimal control studies control laws for stochastic systems when the objective is not limited to an expected cumulative cost or reward, but instead incorporates tail behavior, variability, or dynamically updated notions of risk. In the literature represented here, the subject appears in several mathematically distinct forms: exponential-of-cost or entropic criteria for finite- and infinite-horizon problems, Conditional Value-at-Risk (CVaR), dynamic convex or coherent risk measures, and explicit risk constraints such as bounded probabilities of failure. Across discrete-time MDPs, controlled diffusions, mean-field systems, and forward-backward stochastic systems, a recurrent theme is that the risk-neutral formulation is recovered in small-parameter limits such as θ→0\theta\to0, μ→0\mu\to0, or γ→0\gamma\to0, where variance enters at first order in the expansion of the criterion (Bensoussan et al., 2017, Broek et al., 2012, Djehiche et al., 2014).

1. Risk criteria and problem classes

A central formulation uses an exponential transform of cumulative cost. In finite-horizon discrete time, one example is

JTπ=Eπexp⁡{γ∑t=1Tc(t)},J_T^\pi=\mathbb E^\pi\exp\Bigl\{\gamma\sum_{t=1}^T c(t)\Bigr\},

with infinite-horizon average analogue

Jπ=lim sup⁡T→∞1γTlog⁡Eπexp⁡{γ∑t=1Tc(t)}.J^\pi=\limsup_{T\to\infty}\frac1{\gamma T}\log\mathbb E^\pi\exp\Bigl\{\gamma\sum_{t=1}^T c(t)\Bigr\}.

This multiplicative criterion is used, for example, in queueing control for wireless transmission scheduling (Singh et al., 2017). In continuous time, a standard cost-minimization form is

J(t,x;u)=1μln⁡E[exp⁡{μ ⁣∫tTf(s,Xt,x;u(s),u(s)) ds+μ h(Xt,x;u(T))}],J(t,x;u)=\frac1\mu\ln \mathbb E\Bigl[\exp\Bigl\{\mu\!\int_t^T f\bigl(s,X^{t,x;u}(s),u(s)\bigr)\,ds+\mu\,h\bigl(X^{t,x;u}(T)\bigr)\Bigr\}\Bigr],

which is equivalent to a recursive formulation through a backward component of an FBSDE (Dong et al., 5 Feb 2026). Mean-field variants replace the state law by its marginal distribution or a functional of that law, yielding risk-sensitive McKean–Vlasov control problems (Bensoussan et al., 2017).

Long-run formulations on finite state-action MDPs are often expressed through entropic utility. For a bounded random variable ZZ and risk parameter α≠0\alpha\neq0,

Ent(Z,α)=1αln⁡E[eαZ],\mathrm{Ent}(Z,\alpha)=\frac1\alpha\ln\mathbb E[e^{\alpha Z}],

and the long-run average reward is

Jα(x,π)=lim inf⁡N→∞1N Entxπ(∑t=0N−1c(Xt,at),α).J_\alpha(x,\pi)=\liminf_{N\to\infty}\frac1N\,\mathrm{Ent}_x^\pi\Bigl(\sum_{t=0}^{N-1}c(X_t,a_t),\alpha\Bigr).

At α=0\alpha=0 this reduces to the classical expected average reward (Bäuerle et al., 2024). The same entropic structure is used in another finite-MDP treatment with risk parameter γ\gamma, where the discounted approximation

Jγ(s,π;β)=1γln⁡Esπ[exp⁡{γ∑i=0∞βic(Xi,ai)}]J_\gamma(s,\pi;\beta)=\frac1\gamma\ln\mathbb E_s^\pi\Bigl[\exp\Bigl\{\gamma\sum_{i=0}^\infty \beta^i c(X_i,a_i)\Bigr\}\Bigr]

serves as the basis for vanishing-discount analysis (Pitera et al., 19 Jan 2026).

Risk sensitivity is not restricted to exponential criteria. A second major line of work optimizes CVaR: CVaRα(Z)=min⁡η∈R{η+11−α E[(Z−η)+]},\mathrm{CVaR}_\alpha(Z)=\min_{\eta\in\mathbb R}\Bigl\{\eta+\frac1{1-\alpha}\,\mathbb E[(Z-\eta)_+]\Bigr\}, with uncertainty entering through initial conditions, process noise, and uncertain parameters (Wang et al., 2020). A third line uses dynamic convex or coherent one-step risk maps ρt\rho_t and composes them recursively,

ρt,T(Zt,…,ZT)=Zt+ρt(Zt+1+ρt+1(Zt+2+⋯+ρT−1(ZT))),\rho_{t,T}(Z_t,\dots,Z_T)=Z_t+\rho_t\Bigl(Z_{t+1}+\rho_{t+1}(Z_{t+2}+\cdots+\rho_{T-1}(Z_T))\Bigr),

to obtain time-consistent multi-period risk evaluation (Coache et al., 2021, Chow et al., 2015). A fourth line imposes explicit chance constraints, such as

Px0=z{xτ∈Γ}≤η,P_{x_0=z}\{x_\tau\in\Gamma\}\le \eta,

and then reformulates them through an auxiliary martingale state variable (Huynh et al., 2013).

A common simplification is to equate risk sensitivity with a static mean-variance penalty. The literature here is broader. It includes entropic criteria, CVaR optimization, recursive dynamic risk measures, and bounded failure probabilities, and several works treat time consistency as a separate structural requirement rather than an automatic consequence of aversion to variance (Coache et al., 2021, Chow et al., 2015, Huynh et al., 2013).

2. Dynamic programming, Bellman equations, and HJB structure

The dynamic-programming equations for risk-sensitive control differ from the additive Bellman recursion of the risk-neutral case. In discrete-time finite-state models with exponential criteria, one obtains multiplicative Bellman equations. For the wireless queueing model with buffer size BB, there exist α>0\alpha>0 and positive V(i)V(i) satisfying

α V(i)=min⁡u∈{0,1}∑j=0Beγ c(i,u,j) p(j∣i,u) V(j),i=0,1,…,B,\alpha\,V(i)=\min_{u\in\{0,1\}}\sum_{j=0}^B e^{\gamma\,c(i,u,j)}\,p(j\mid i,u)\,V(j), \quad i=0,1,\dots,B,

and the minimizing action is optimal. Relative value iteration normalizes the iterates by Vk+1(0)V_{k+1}(0) and provides a constructive route to the optimal policy (Singh et al., 2017).

For long-run average entropic control on finite MDPs, Bäuerle, Pitera, and Stettner analyze an averaged risk-sensitive Bellman equation

w(x,α)=max⁡a∈U{c(x,a)−λ(α)+1αln⁡∑y∈Eeαw(y,α)Pa(x,y)},w(x,\alpha)=\max_{a\in U}\Bigl\{c(x,a)-\lambda(\alpha)+\frac1\alpha\ln\sum_{y\in E}e^{\alpha w(y,\alpha)}P^a(x,y)\Bigr\},

with ww defined up to an additive constant and λ(α)\lambda(\alpha) equal to the optimal average reward. The associated Bellman operator is a contraction in the span-seminorm under the stated mixing assumptions, and stationary maximizers are optimal for the long-run criterion (Bäuerle et al., 2024).

In continuous time, dynamic programming produces nonlinear HJB equations. For the finite-horizon diffusion problem with exponential criterion, the value function V(t,x)V(t,x) solves, in the viscosity sense,

Vt(t,x)+inf⁡u∈U{f(t,x,u)+⟨Vx,b(t,x,u)⟩+μ2 ∣σ(t,x,u)⊤Vx∣2+12 tr(σσ⊤Vxx)}=0,V_t(t,x)+\inf_{u\in U}\Bigl\{ f(t,x,u)+\langle V_x,b(t,x,u)\rangle +\frac\mu2\,|\sigma(t,x,u)^\top V_x|^2 +\frac12\,\mathrm{tr}\bigl(\sigma\sigma^\top V_{xx}\bigr) \Bigr\}=0,

with terminal condition V(T,x)=h(x)V(T,x)=h(x) (Dong et al., 5 Feb 2026). The quadratic term in σ⊤Vx\sigma^\top V_x is the characteristic risk-sensitive modification.

Mean-field control lifts this structure to the space of probability measures. The infinite-dimensional HJB equation for the value functional Vθ(t,m)V^\theta(t,m) involves Lions derivatives δV/δm\delta V/\delta m and integrates drift, diffusion, and running cost against the current law mm. In the linear-exponential-quadratic case, the infinite-dimensional problem reduces to coupled Riccati ODEs for Pθ(t)P^\theta(t) and Πθ(t)\Pi^\theta(t) (Bensoussan et al., 2017).

Under a specific matching condition, path-integral methods linearize the risk-sensitive HJB. For a diffusion

dXs=b(s,Xs) ds+B(s,Xs)(us ds+σ dWs),dX_s=b(s,X_s)\,ds+B(s,X_s)\bigl(u_s\,ds+\sigma\,dW_s\bigr),

quadratic control cost, and

σσ⊤=λ0(R⊤R)−1,\sigma\sigma^\top=\lambda_0(R^\top R)^{-1},

the logarithmic transform Jθ=−λlog⁡ψJ_\theta=-\lambda\log\psi with λ=1−θλ0\lambda=1-\theta\lambda_0 converts the nonlinear PDE into the linear backward equation

−∂tψ=L0ψ−1λV(x)ψ,-\partial_t\psi=\mathcal L_0\psi-\frac1\lambda V(x)\psi,

which admits a Feynman–Kac representation (Broek et al., 2012). This suggests that, for that class of models, risk sensitivity can be handled without abandoning sampling-based forward simulation.

3. Maximum principle, adjoint systems, and the DPP–MP connection

Pontryagin-type analysis provides an alternative to dynamic programming and is particularly important when the value function is nonsmooth, the control domain is nonconvex, or the problem is not Markovian. For the finite-horizon diffusion problem with cost

J(t,x;u)=1μln⁡E[exp⁡{μ ⁣∫tTf(s,Xt,x;u(s),u(s)) ds+μ h(Xt,x;u(T))}],J(t,x;u)=\frac1\mu\ln \mathbb E\Bigl[\exp\Bigl\{\mu\!\int_t^Tf\bigl(s,X^{t,x;u}(s),u(s)\bigr)\,ds+\mu\,h\bigl(X^{t,x;u}(T)\bigr)\Bigr\}\Bigr],

the problem is equivalent to a decoupled FBSDE in which the backward equation has quadratic generator

dY(s)=−[f(s,X(s),u(s))+μ2∣Z(s)∣2] ds+Z(s) dW(s),Y(T)=h(X(T)).dY(s)=-\Bigl[f\bigl(s,X(s),u(s)\bigr)+\frac\mu2|Z(s)|^2\Bigr]\,ds+Z(s)\,dW(s), \quad Y(T)=h(X(T)).

First- and second-order adjoint BSDEs then produce a generalized Hamiltonian H\mathcal H, and the optimal control satisfies a pointwise maximum condition involving both first- and second-order adjoints (Dong et al., 5 Feb 2026).

The same paper establishes the link between the maximum principle and dynamic programming in a viscosity framework. Along an optimal trajectory, the adjoint processes enter the super- and sub-jets of the value function: {p(s)}×[P(s),∞)⊆Dx2,+V(s,Xˉ(s)),Dx2,−V(s,Xˉ(s))⊆{p(s)}×(−∞,P(s)].\{p(s)\}\times[P(s),\infty)\subseteq D_x^{2,+}V\bigl(s,\bar X(s)\bigr),\qquad D_x^{2,-}V\bigl(s,\bar X(s)\bigr)\subseteq \{p(s)\}\times(-\infty,P(s)]. This jet-based formulation extends the DPP–MP relation beyond the C1,2C^{1,2} setting (Dong et al., 5 Feb 2026). Under smoothness of the value function, the connection becomes pointwise: ps=Vx(s,Xˉs),qs=Vxx(s,Xˉs) σ(s,Xˉs,uˉs),p_s=V_x(s,\bar X_s),\qquad q_s=V_{xx}(s,\bar X_s)\,\sigma(s,\bar X_s,\bar u_s), and the HJB minimizer coincides with the Hamiltonian minimizer (Dong et al., 9 Jul 2025).

Mean-field systems introduce derivatives with respect to the law. In risk-sensitive mean-field-type control, the adjoint process satisfies a backward SDE containing both ∂xHθ\partial_x H^\theta and the Lions derivative ∂mHθ\partial_m H^\theta, while the dynamic-programming side uses an infinite-dimensional HJB master equation. Bensoussan, Djehiche, Tembine, and Yam show the relation

Y(t)=DxδVθδm(t,mt∗)(x∗(t)),Y(t)=D_x\frac{\delta V^\theta}{\delta m}\bigl(t,m_t^*\bigr)\bigl(x^*(t)\bigr),

connecting the adjoint to the gradient of the value functional along the optimal measure flow (Bensoussan et al., 2017).

When mean-field dependence induces time inconsistency and Bellman optimality is unavailable, the stochastic maximum principle remains applicable. Djehiche, Tembine, and Tempone derive a risk-sensitive SMP for mean-field-type dynamics without requiring smoothness of a value function. The resulting Hamiltonian contains an exponential-tilt term ℓ\ell, and the adjoint BSDEs are driven under a tilted measure (Djehiche et al., 2014). For partially observed coupled FBSDEs with Brownian and Poisson noise, the recursive risk-sensitive problem can be transformed into a complete-information problem with quadratic-exponential generator, leading to a global stochastic maximum principle and a modified Zakai equation in the associated filtering problem (Lin et al., 7 Apr 2025).

4. Structural policies, monotonicity, and long-run stability

Risk-sensitive criteria often preserve or induce strong structural properties. In the single-client wireless queueing model, the optimal scheduling policy is of threshold type: u⋆(n)={0,n<τ⋆,1,n≥τ⋆.u^\star(n)= \begin{cases} 0,& n<\tau^\star,\\ 1,& n\ge \tau^\star. \end{cases} The proof proceeds through the monotonicity of the decision differential

ΔJk+1(n)=Jk+1(n,0)−Jk+1(n,1),\Delta J_{k+1}(n)=J_{k+1}(n,0)-J_{k+1}(n,1),

which is shown by induction to be nondecreasing in the queue length. The optimal threshold increases when the transmission cost CC increases, and a fixed threshold τ\tau remains optimal throughout an interval [Cl,Cu][C_l,C_u] characterized by the inequalities ΔJ(τ−1)≤0\Delta J(\tau-1)\le0 and ΔJ(τ)≥0\Delta J(\tau)\ge0 (Singh et al., 2017).

Long-run average risk-sensitive control on finite MDPs exhibits a different structural issue: robustness of optimal stationary policies with respect to the risk-aversion parameter. For each stationary policy uu, the region of optimality

Γ(u)={α∈R: u∈arg⁡max⁡v∈Π′λv(α)}\Gamma(u)=\{\alpha\in\mathbb R:\ u\in\arg\max_{v\in\Pi'}\lambda^v(\alpha)\}

is closed and, in the formulation of Bäuerle, Pitera, and Stettner, a finite union of closed intervals. Moreover, λu(α)\lambda^u(\alpha) is continuous, bounded, non-decreasing in α\alpha, and real-analytic on R\mathbb R. Thus, if a unique stationary optimizer exists at α0\alpha_0, it remains optimal on a neighborhood of α0\alpha_0 (Bäuerle et al., 2024).

The vanishing-discount limit yields a risk-sensitive Blackwell property. For each fixed coordinate nn of the discounted optimal policy πβ=(u0β,u1β,… )\pi^\beta=(u_0^\beta,u_1^\beta,\dots), there exists βn<1\beta_n<1 such that, for all β∈(βn,1)\beta\in(\beta_n,1), the stationary rule unβu_n^\beta is optimal for the average problem at the fixed risk parameter (Bäuerle et al., 2024, Pitera et al., 19 Jan 2026). This does not imply that discounted optimal policies themselves are stationary. On the contrary, the entropic discounted problem may require non-stationary optimal policies, and one of the cited examples shows that no discounted-optimal stationary policy exists for β<1\beta<1 even though the Blackwell limit identifies a stationary average-optimal policy (Bäuerle et al., 2024).

A plausible implication is that policy stability in risk-sensitive control must be interpreted along at least two different axes: perturbation of the risk parameter and vanishing discount. The distinction from ultimate stationarity, where β\beta is fixed and the risk parameter tends to zero, is explicit in the finite-MDP analysis and is not merely terminological (Pitera et al., 19 Jan 2026).

5. Time consistency, risk constraints, and partial information

Time consistency is a defining concern in multi-stage risk-sensitive control. In the dynamic-risk-measure framework, time consistency is enforced by recursive composition of one-step conditional risk measures. Under a policy π\pi, the value function becomes a conditional risk-to-go,

Vtπ(s)=ρt ⁣(ct(s,at,s′)+Vt+1π(s′)∣st=s),V_t^\pi(s)=\rho_t\!\Bigl(c_t(s,a_t,s')+V_{t+1}^\pi(s')\mid s_t=s\Bigr),

and policy optimization proceeds through a dynamic programming principle that respects the recursive risk semantics (Coache et al., 2021). Chow and Pavone use the same principle for constrained control with a second cost stream dtd_t and a risk budget rkr_k, augmenting the state by the residual budget and solving

Vk(x,r)=inf⁡(u,r′)∈Fk(x,r){ck(x,u)+∑x′Q(x′∣x,u) Vk+1(x′,r′(x′))},V_k(x,r)=\inf_{(u,r')\in F_k(x,r)} \Bigl\{ c_k(x,u)+\sum_{x'}Q(x'|x,u)\,V_{k+1}(x',r'(x')) \Bigr\},

subject to the single-period update

dk(x,u)+ρk(r′(⋅))≤r.d_k(x,u)+\rho_k(r'(\cdot))\le r.

The feasible state becomes (xk,rk)(x_k,r_k) rather than xkx_k alone (Chow et al., 2015).

Chance constraints reveal why this augmentation is necessary. In continuous-time diffusion control with failure boundary Γ\Gamma, the constraint

Px0=z{xτ∈Γ}≤ηP_{x_0=z}\{x_\tau\in\Gamma\}\le \eta

can be embedded into an auxiliary martingale qtq_t satisfying q0=ηq_0=\eta, dqt=ct⊤dWtdq_t=c_t^\top dW_t, qt∈[0,1]q_t\in[0,1], and 1xτ∈Γ≤qτ1_{x_\tau\in\Gamma}\le q_\tau almost surely. The enlarged state (xt,qt)(x_t,q_t) converts the problem into a stochastic target problem. Because

qt=E[1xτ∈Γ∣At],q_t=\mathbb E[1_{x_\tau\in\Gamma}\mid \mathcal A_t],

residual risk budgets observed at intermediate times remain exactly aligned with the original global chance constraint, thereby restoring time consistency (Huynh et al., 2013).

Partial observation introduces another layer. In adaptive CVaR-MPC, the unknown parameters are appended to the state, a particle filter propagates particles y=[x;ϕ]y=[x;\phi], and the CVaR objective is evaluated over the joint uncertainty in (xt,ϕ)(x_t,\phi) and process noise (Wang et al., 2020). In the partially observed FBSDE with Poisson jumps, a Girsanov transformation and an exponential transformation reduce the original risk-sensitive problem to a full-information recursive control problem with a quadratic-exponential backward generator; in the associated filtering problem, the unnormalized filter satisfies a modified Zakai equation (Lin et al., 7 Apr 2025).

A common misconception is that time consistency is automatically inherited from any one-period risk metric. The cited literature indicates the opposite: time consistency generally requires either recursive composition of one-step risk maps or an explicit state augmentation that tracks residual risk tolerance (Chow et al., 2015, Huynh et al., 2013).

6. Numerical methods and application domains

Computational methods for risk-sensitive control range from exact transforms in special models to approximate dynamic programming, stochastic search, and inference-based policy optimization. Path-integral control solves the transformed linear PDE by forward simulation of uncontrolled dynamics with Feynman–Kac weights, and the optimal feedback becomes

u∗(x,t)=λ (R⊤R)−1B⊤∇xlog⁡ψ(x,t).u^*(x,t)=\lambda\,(R^\top R)^{-1}B^\top\nabla_x\log\psi(x,t).

In multimodal problems, the resulting control law superposes modes through probability-weighted contributions, and the cited analysis shows how risk sensitivity changes the weighting without changing the symmetry-breaking time in the symmetric two-slit example (Broek et al., 2012).

For nonlinear stochastic systems with Gaussian noise, the ILEG method iteratively linearizes the dynamics, quadratizes the cost, and solves a linear exponential-quadratic subproblem. The backward pass is governed by a risk-sensitive Riccati equation containing the term

θ St Ct Σ Ct⊤St.\theta\,S_t\,C_t\,\Sigma\,C_t^\top S_t.

Positive risk sensitivity leads to larger feedback gains, whereas negative values produce lower gains and a more feedforward strategy (Farshidian et al., 2015). In discrete-time convex control, RS-MPC modifies prescient planning by optimizing over disturbance realizations. For γ<0\gamma<0 the resulting planning problem is convex; for γ>0\gamma>0 it becomes a difference-of-convex program handled approximately by the convex-concave procedure, and the same calculation yields a lower bound on the optimal cost (Moehle, 2021).

Sampling-based methods now cover both model-predictive and policy-gradient settings. The RS3 framework for CVaR-MPC parameterizes control trajectories by an exponential-family sampling distribution and performs stochastic-gradient ascent on a shaped objective involving empirical CVaR. Under the stated stochastic-approximation conditions, the parameter sequence converges almost surely to a stationary point, and the empirical CVaR estimator obeys an exponential concentration bound. The implementation details reported include typical values N=200N=200, M=50M=50, horizon H=20H=20, confidence level α=0.9\alpha=0.9, Gaussian policies with fixed covariance, Polyak averaging, and GPU-parallel Monte Carlo rollouts (Wang et al., 2020). In a different direction, Rao-Blackwellized Markovian score climbing formulates entropic control as maximum likelihood in a state-space model, uses conditional SMC as an MCMC kernel over trajectories, and obtains unbiased policy-gradient estimates with variance decaying as O(1/N)O(1/N) (Abdulsamad et al., 2023).

Applications are correspondingly varied. Queueing control yields threshold transmission policies in wireless scheduling (Singh et al., 2017). CVaR-MPC and dynamic-risk RL are demonstrated on pendulum, cartpole, quadcopter, statistical arbitrage, option hedging with friction, and cliff-walking robot control (Wang et al., 2020, Coache et al., 2021). Risk-sensitive impedance optimization for uncertain contact interactions in legged locomotion uses an EKF-augmented risk-sensitive DDP and reports, among other outcomes, a 25% lower peak impact force in the single-leg swing example and improved trotting success rates from 47% to 65% and from 72% to 91.6% under different terrain uncertainties (Hammoud et al., 2020). Financial applications include HARA-utility portfolio choice under correlated noises, reformulated as an exponential-of-integral control problem and solved by a Riccati equation in the regime 0<γ<10<\gamma<1 (Yang et al., 2019), as well as linear-quadratic portfolio optimization in the SMP–DPP comparison framework (Dong et al., 9 Jul 2025).

Taken together, these methods show that risk-sensitive stochastic optimal control is not a single algorithmic doctrine but a family of mathematically distinct control paradigms. What unifies them is the replacement of linear expectation by nonlinear risk evaluation, the resulting modification of Bellman or Pontryagin structures, and the systematic appearance of additional state variables, adjoint terms, or transformed dynamics needed to preserve optimality, tractability, or time consistency.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (18)

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to Risk-Sensitive Stochastic Optimal Control.