Viscosity Solutions of Fully second-order HJB Equations in the Wasserstein Space
Erhan Bayraktar, Hang Cheung, Ibrahim Ekren, Jinniao Qiu, Ho Man Tai, Xin Zhang · 2025
Read Viscosity Solutions of Fully second-order HJB Equations in the Wasserstein Space by Erhan Bayraktar, Hang Cheung, Ibrahim Ekren, Jinniao Qiu, Ho Man Tai, Xin Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-01612.txt.
Introduction
Erhan Bayraktar and five colleagues prove a claim about mean field control problems with shared noise. The value functions of these problems are the solutions of a second-order equation. That equation lives on the space of probability distributions. Mean field control steers huge populations where everyone feels the same random shocks. The governing equation is of Hamilton-Jacobi-Bellman type, and making sense of its solutions in this infinite-dimensional setting needs the viscosity framework. Writing in twenty twenty-five, the team allows the second derivative in the distribution variable to depend on the state itself, the case they call fully second order.
Their proof borrows smooth approximations built from finite particle systems by earlier authors and a compactness trick based on penalizing distribution moments. Unbounded dynamics and state-dependent shared noise are covered. The authors state that, to their knowledge, no result of this kind existed before.
Catalog record
| CREATOR | Erhan Bayraktar, Hang Cheung, Ibrahim Ekren, Jinniao Qiu, Ho Man Tai, Xin Zhang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 151 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-01612
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Viscosity Solutions of Fully second-order HJB Equations in the Wasserstein Space Thanks: I. Ekren is partially supported by the NSF grant DMS-2406240. H. Cheung and J. Qiu are partially supported by Discovery Grant from the Natural Sciences and Engineering Research Council of Canada (NSERC). Erhan Bayraktar, Hang Cheung, Ibrahim Ekren, Jinniao Qiu, Ho Man Tai, Xin Zhang Abstract In this paper, we show that the value functions of mean field control problems with common noise are the unique viscosity solutions to fully second-order Hamilton-Jacobi-Bellman equations, in a Crandall-Lions-like framework. We allow the second-order derivative in measure to be state-dependent and thus infinite-dimensional, rather than derived from a finite-dimensional operator, hence the term “fully”. Our argument leverages the construction of smooth approximations from particle systems developed by Cosso, Gozzi, Kharroubi, Pham, and Rosestolato [Trans. Amer. Math. Soc., 2023], and the compactness argument via penalization of measure moments in Soner and Yan [Appl. Math. Optim., 2024]. Our work addresses unbounded dynamics and state-dependent common noise volatility, and to our knowledge, this is the first result of its kind in the literature. Keywords: mean field type control, Wasserstein space, second-order HJB equation, viscosity solutions, Bellman equation, comparison theorem. Mathematics Subject Classification (2020): 49L25, 35Q93, 35B51, 58E30. 1 Introduction This paper establishes the existence and uniqueness of viscosity solution to the following Hamilton-Jacobi-Bellman (HJB) equation in the Wasserstein space arising from mean field control problems with common noise \displaystyle\left\{\begin{aligned} &\partial_{t}u(t,\mu)+\int_{\mathbb{R}^{d}}\sup_{a\in A}\Bigg\{f(t,x,\mu,a)+b(t,x,\mu,a)\cdot\partial_{\mu}u(t,\mu)(x)\\ &\hskip 90.0pt+\dfrac{1}{2}\text{tr}\Big((\sigma(t,x,a)\big[\sigma(t,x,a)\big]^{\top}+\sigma^{0}(t,x)[\sigma^{0}(t,x)]^{\top})\nabla_{x}\partial_{\mu}u(t,\mu)(x)\Big)\Bigg\}\mu(dx)\\ &+\dfrac{1}{2}\int_{\mathbb{R}^{d}\times\mathbb{R}^{d}}\text{tr}\Big[\sigma^{0}(t,x)[\sigma^{0}(t,y)]^{\top}\partial_{\mu}^{2}u(t,\mu)(x,y)\Big]\,\mu^{\otimes 2}(dx,dy)=0,\hskip 5.0pt\forall\,(t,\mu)\in[0,T)\times\mathcal{P}_{2}(\mathbb{R}^{d}),\\ &u(T,\mu)=\int_{\mathbb{R}^{d}}g(x,\mu)\mu(dx)\hskip 10.0pt\text{for $\mu\in\mathcal{P}_{2}(\mathbb{R}^{d})$},\end{aligned}\right. (1.1) where A is the control space, and b,\sigma,\sigma^{0},f,g are coefficients of the associated optimization problem. Here \partial_{\mu}u and \partial^{2}_{\mu}u denote the respective first and second-order L -derivatives of u [ 13 , 11 ] , and ( 1.1 ) is a degenerate fully second-order equation in the Wasserstein space. By “fully” second-order, we mean that the second-order term is infinite-dimensional, rather than a finite-dimensional operator as in [ 4 , 24 , 17 ] . Our manuscript is the first work in the literature to address the viscosity solution directly in the Wasserstein space and to allow for unbounded dynamics and state-dependent common noise volatility. Over the past two decades, the mean field games and mean field control problems, as models for strategic decision making among a large population of symmetric agents, have attracted considerable attention. This method was independently introduced by Huang, Malhamé, and Caines [ 32 ] , and Lasry and Lions [ 34 ] . In this framework, it is assumed that symmetric agents interact with each other through a medium known as the mean field term, which represents the collective influence of all agents’ decisions. For a comprehensive introduction to mean field theory, we refer readers to Bensoussan, Frehse, and Yam [ 6 ] , Carmona and Delarue [ 13 ] , and Gomes, Pimentel, and Voskanyan [ 31 ] . Due to the presence of the mean field term, the HJB equations of mean field control problems are defined in the Wasserstein space and involve derivatives of measures, making them infinite-dimensional problems. For the well-posedness of classical solutions to HJB equations in the Wasserstein space, we refer the reader to [ 11 ] by Cardaliaguet, Delarue, Lasry, and Lions, [ 16 ] by Chassagneux, Crisan, and Delarue, [ 29 ] by Gangbo, Mészáros, Mou, and Zhang, and the references therein. However, the existence of classical solutions requires stringent regularity conditions on the coefficient functions and a monotonicity structure of the model. To analyze the HJB equations to mean-field control problems with general cost functions and state dynamics, the theory of viscosity solutions is indispensable, as it typically requires only continuity conditions and natural growth rates of the coefficient functions. To overcome the non-smooth structure of the Wasserstein space, one can lift equations in the Wasserstein space to Hilbert spaces and make use of the existing theory [ 26 , 1 , 36 ] . However, the relation between the solutions to the lifted PDEs and the original ones is unclear; see [ 19 , Remark 3.6] for more details. The lifting of a smooth function in the Wasserstein space may not be second-order Fréchet differentiable; see e.g. [ 9 , Example 2.3] . Moreover, there is no suitable lift version of \nabla_{x}\partial_{\mu} in the L^{2} space. In this paper, we address the HJB equations directly in the Wasserstein space, establishing the existence and uniqueness of viscosity solution to equation ( 1.1 ) under a Crandall-Lions-like Definition 2.5 . A viscosity solution to equation ( 1.1 ) is provided by the value function v of the corresponding mean-field control problem; see Theorem 4.2 . To prove uniqueness, we adopt the idea of smooth approximation v_{n}\to v , as n\to\infty from [ 19 , 17 ] . For any subsolution u_{1} , we aim to show that u_{1}\leq v . Suppose the contrary, u_{1}(t_{0},\mu_{0})>v(t_{0},\mu_{0}) at some (t_{0},\mu_{0}) in the interior. Therefore, for sufficiently large n , \sup_{(t,\mu)}u_{1}(t,\mu)-v_{n}(t,\mu)>0 . To show a contradiction, we want to find a maximum point of (t,\mu)\mapsto u_{1}(t,\mu)-v_{n}(t,\mu) and apply the definition of viscosity solution. However, the Wasserstein space is not locally compact and the maximizer may not be obtained. Therefore, [ 19 , 17 ] applied the smooth variational principle so that perturbed functions achieve their global maximum. In [ 19 , 17 , 3 ] , the perturbation terms are not second-order L -differentiable, and hence they were not able to prove uniqueness for fully second-order HJB equations in the Wasserstein space. In this paper, to overcome the compactness issue of the Wasserstein space, instead of applying the smooth variational principle, we adopt the idea of moment penalization from [ 40 ] . For any \mu\in\mathcal{P}_{2}(\mathbb{R}^{d}) , we define M_{2}(\mu):=\int x^{2}\,\mu(dx) . We can always find a small \delta>0 such that u_{1}(t_{0},\mu_{0})-v_{n}(t_{0},\mu_{0})-\delta M_{2}(\mu_{0})>0 . This implies that the maximum point of u_{1}-v_{n}-\delta M_{2} , if exists, must be attained in the set of measures with bounded second moments. As shown in Lemma 2.1 , such a set is compact with respect to the finer topology induced by 1 -Wasserstein metric. Moreover, the coefficient functions are assumed to be Lipschitz continuous under the 1-Wasserstein metric in our framework, which implies the same property for the value function. Therefore, the maximum point of u_{1}-v_{n}-\delta M_{2} is attained, and we derive the desired contradiction using the definition of viscosity solution. Another advantage of this approach is that M_{2}(\mu) possesses L -derivatives up to second-order in a very clean form, \partial_{\mu}M_{2}(\mu)(x)=2x and \partial_{\mu}^{2}M_{2}(\mu)(x,y)=0 , which allows unbounded state dynamics and state dependent common noise in our estimate. We note that the definition of viscosity subsolution in this paper follows the standard Crandall-Lions framework, but we need to modify the definition of viscosity supersolution in a stronger sense when dealing with equations with mean field terms. This modification aims to address the technical difficulty posed by the supremum in equation ( 1.1 ), which leads to scarce choice of test functions in the proof of the comparison for the supersolution. Similarly to [ 19 ] , we construct smooth test functions for supersolutions using the cost functional with fixed control, but with the domain of [0,T]\times\mathcal{P}_{2}(\mathbb{R}^{d}\times A) rather than [0,T]\times\mathcal{P}_{2}(\mathbb{R}^{d}) . The strong dependence on the initial random variable and