Appendix
1.1
\(\epsilon \)-Nash Equilibrium
We now establish that the solutions of Problems 2.1 and 2.2 is an \(\epsilon \)-Nash Equilibrium. Suppose that there are N representative agents behaving in similar manner, so that the state of the dominating player and the i-th agent satisfies the following SDE respectively:
$$\begin{aligned}&\left\{ \begin{array}{rcl} dy_0&{}=&{}g_0\Big (y_0(t),\dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{y_1^j(t)},u_0(t)\Big )dt+\sigma _0\Big (y_0(t)\Big )dW_0(t),\\ y_0(0)&{}=&{}\xi _0.\\ \end{array} \right. \nonumber \\&\left\{ \begin{array}{rcl} dy_1^i&{}=&{}g_1\Big (y_1^i(t),y_0(t),\dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{y_1^j(t)},u_1^i(t)\Big )dt+\sigma _1\Big (y_1^i(t)\Big )dW_1^i(t),\\ y_1^i(0)&{}=&{}\xi _1^i.\\ \end{array}\right. \nonumber \\ \end{aligned}$$
(22)
where \(\delta _y\) is Dirac measure with a unit mass at y. We call Equation (22) the empirical system. The corresponding objective functional for the i-th agent is:
$$\begin{aligned} \mathcal {J}^{N,i}(\mathbf{u})= & {} \mathbb {E}\bigg [\displaystyle \int _0^Tf_1\Big (y_1^i(t),y_0(t),\dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{y_1^j(t)},u^i(t)\Big )dt\\&+h_1\Big (y_1^i(T),y_0(T),\dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{y_1^j(T)}\Big )\bigg ], \end{aligned}$$
where \(\mathbf{u}=(u_1^1,u_1^2,\ldots ,u_1^N)\). We expect that when \(N \rightarrow \infty \), the hypothetical approximation is described by (1), that is:
$$\begin{aligned}&\left\{ \begin{array}{rcl} dx_0&{}=&{}g_0\Big (x_0(t),m_{x_0}(t),u_0(t)\Big )dt+\sigma _0\Big (x_0(t)\Big )dW_0(t),\\ x_0(0)&{}=&{}\xi _0.\\ \end{array}\right. \nonumber \\&\left\{ \begin{array}{rcl} dx_1^i&{}=&{}g_1\Big (x_1^i(t),x_0(t),m_{x_0}(t),u_1^i(t)\Big )dt+\sigma _1\Big (x_1^i(t)\Big )dW_1^i(t),\\ x_1^i(0)&{}=&{}\xi _1^i.\\ \end{array}\right. \end{aligned}$$
(23)
We call Eq. (23) the mean field system. The corresponding limiting objective functional for the i-th player is
$$\begin{aligned} \mathcal {J}^{i}(u_1^i)=\mathbb {E}\bigg [\displaystyle \int _0^Tf_1\Big (x_1^i(t),x_0(t),m_{x_0}(t),u_1^i(t)\Big )dt+h_1\Big (x_1^i(T),x_0(T),m_{x_0}(T)\Big )\bigg ].\nonumber \\ \end{aligned}$$
(24)
Using Corollary 2.5, the necessary condition for optimality is described by the SHJB-FP coupled Eq. (8). To proceed, we assume that the optimal control \(\hat{\mathbf{u}}=(\hat{u}_1^1,\hat{u}_1^2,\ldots ,\hat{u}_1^N)\) exists. To avoid ambiguity, denote \(\hat{x}_1^i\) and \(\hat{y}_1^i\) the states dynamics of \(x_1^i\) and \(y_1^i\) corresponding to the optimal control \(\hat{u}_1^i\). The mean field term \(m_{x_0}(\cdot ,t)\), is the probability measure of the optimal trajectory \(\hat{x}_1^i\) at time t, conditioning on \(\mathcal {F}^0_t\). Under this construction, being conditional on \(\mathcal {F}^0_t\), \(\{\hat{x}_1^i\}_i\) are independent and identically distributed processes; while \(\{\hat{y}_1^i\}_i\) are dependent on each other through the empirical distribution. For simplicity, for two density functions m and \(m'\), we write \(W_2(m d \lambda ,m' d \lambda )=W_2(m,m')\).
Lemma 5.1
Suppose that the assumptions (A.1–A.3) hold. If \(m_{x_0}(\cdot ,t)\) is chosen to be the density function of \(\hat{x}_1^i\) conditional on \(\mathcal {F}_t^0\), then
$$\begin{aligned} \mathbb {E}\Big [\displaystyle \sup _{u \le T}|y_0(u)-x_0(u)|^2\Big ]+\mathbb {E}\Big [\displaystyle \sup _{u \le T}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\Big ]=o(1). \end{aligned}$$
Proof
Observe that for any \(t\in [0,T]\)
$$\begin{aligned} \mathbb {E}\displaystyle \sup _{u\le t}|y_0(u)-x_0(u)|^2\le & {} C \bigg \{t\mathbb {E}\displaystyle \int _0^t \bigg |g_0\Bigg (y_0(s),\dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{y}_1^j(s)},u_0(s)\Bigg )\\&-g_0\Big (x_0(s),m_{x_0}(s),u_0(s)\Big )\bigg |^2ds\\&+\mathbb {E}\displaystyle \int _0^t\bigg |\sigma _0\Big (y_0(s)\Big )-\sigma _0\Big (x_0(s)\Big )\bigg |^2ds\bigg \}, \end{aligned}$$
and
$$\begin{aligned} \mathbb {E}\displaystyle \sup _{u\le t}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\le & {} C \bigg \{t\mathbb {E}\displaystyle \int _0^t \bigg |g_1\Big (\hat{y}_1^i(s),y_0(s),\dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{y}_1^j(s)},\hat{u}_1^i(s)\Big )\\&-g_1\Big (\hat{x}_1^i(s),x_0(s),m_{x_0}(s),\hat{u}_1^i(s)\Big )\bigg |^2ds\\&+\mathbb {E}\displaystyle \int _0^t\bigg |\sigma _1\Big (\hat{y}_1^i(s)\Big )-\sigma _1\Big (\hat{x}_1^i(s)\Big )\bigg |^2ds\bigg \}, \end{aligned}$$
By the Lipschitz assumptions, we have
$$\begin{aligned}&\mathbb {E}\displaystyle \sup _{u\le t}|y_0(u)-x_0(u)|^2+\mathbb {E}\displaystyle \sup _{u\le t}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\nonumber \\&\quad \le C\bigg \{t\mathbb {E}\displaystyle \int _0^t \Big |y_0(s)-x_0(s)\Big |^2+W_2^2\left( \dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{y}_1^j(s)},m_{x_0}(s)\right) ds\nonumber \\&\qquad +\mathbb {E}\displaystyle \int _0^t\Big |y_0(s)-x_0(s)\Big |^2 ds\bigg \}\nonumber \\&\qquad +C\bigg \{t\mathbb {E}\displaystyle \int _0^t \Big |\hat{y}_1^i(s)-\hat{x}_1^i(s)\Big |^2+\Big |y_0(s)-x_0(s)\Big |^2\nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{y}_1^j(s)},m_{x_0}(t)\right) ds+\mathbb {E}\displaystyle \int _0^t\Big |y_1^i(s)-x_1^i(s)\Big |^2 ds\bigg \}\nonumber \\&\quad \le C\bigg \{\mathbb {E}\displaystyle \int _0^t \displaystyle \sup _{u\le s}\Big |y_0(u)-x_0(u)\Big |^2+\displaystyle \sup _{u\le s}\Big |\hat{y}_1^i(u)-\hat{x}_1^i(u)\Big |^2\nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{x}_1^j(s)}\right) +W_2^2\left( \dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{x}_1^j(s)}\right) \nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) ds\bigg \}, \end{aligned}$$
(25)
where \(C>0\) is a constant, changing line by line, that depends only on T and K. By definition, for any Dirac measures \(\delta _y\) on \(\mathbb {R}^{n_1}\) and density function m, we have
$$\begin{aligned} W_2^2(\delta _y,m)=\int _{\mathbb {R}^{n_1}}|y-x|^2 dm(x). \end{aligned}$$
Also observe that the joint measure \(\frac{1}{N}\sum _{j=1}^N\delta _{(\hat{y}_1^j(s), \hat{x}_1^j(s))}\) on \({\mathbb {R}^{n_1}\times {\mathbb {R}^{n_1}}}\) has respective marginals \(\frac{1}{N}\sum _{j=1}^N\delta _{\hat{y}_1^j(s)}\) and \(\frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)}\) on \({\mathbb {R}^{n_1}}\). Using the definition of Wasserstein metric, we evaluate
$$\begin{aligned}&\mathbb {E}\left[ W_2^2\left( \dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{x}_1^j(s)}\right) \right] \nonumber \\&\quad \le \mathbb {E}\left[ \displaystyle \int _{\mathbb {R}^{n_1}\times {\mathbb {R}^{n_1}}}|y-x|^2 d\left( \dfrac{1}{N}\sum _{j=1}^N\delta _{(\hat{y}_1^j(s), \hat{x}_1^j(s))}(y,x)\right) \right] \nonumber \\&\quad \le \dfrac{1}{N}\displaystyle \sum _{j=1}^N\mathbb {E}\Big |\hat{y}_1^j(s)-\hat{x}_1^j(s)\Big |^2\nonumber \\&\quad =\mathbb {E}\Big |\hat{y}_1^i(s)-\hat{x}_1^i(s)\Big |^2, \end{aligned}$$
(26)
where the last equality results from the fact that \(\{\hat{y}^j - \hat{x}^j\}_{j=1}^N\) are symmetric. Similarly, we also have
$$\begin{aligned} \mathbb {E}\left[ W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{x}_1^j(s)}\right) \right] \le \mathbb {E}\Big |\hat{y}_1^i(s)-\hat{x}_1^i(s)\Big |^2. \end{aligned}$$
Combining with (25) and applying Gronwall’s inequality, we have
$$\begin{aligned}&\mathbb {E}\displaystyle \sup _{u\le t}|y_0(u)-x_0(u)|^2+\mathbb {E}\displaystyle \sup _{u\le t}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\nonumber \\&\quad \le Ce^{Ct}\mathbb {E}\left[ \displaystyle \int _0^t W_2^2\left( \dfrac{1}{N}\displaystyle \sum _{j=1}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1,j\ne i}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) ds\right] .\nonumber \\ \end{aligned}$$
(27)
First observe that
$$\begin{aligned} \mathbb {E}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] =\mathbb {E}\left[ \mathbb {E}^{\mathcal {F}_s^0}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] \right] . \end{aligned}$$
Given \(\mathcal {F}^0_s\), \(\{\hat{x}_1^j\}_j\) are i.i.d. with the common conditional density function \(m_{x_0}(s)\), which is \(\mathcal {F}^0_s\) measurable. We first have the empirical measure \(\dfrac{1}{N}\displaystyle \sum _{j=1}^N\delta _{\hat{x}_1^j(s)}\) converges weakly to \( m_{x_0}(s)\); while by the strong law of larger number, the empirical second moment \(\dfrac{1}{N}\displaystyle \sum _{j=1}^N |\hat{x}_1^j(s)|^2\) converges almost surely to the theoretical second moment \(\mathbb {E}^{\mathcal {F}_s^0}\left| \hat{x}_1^1(s)\right| ^2\). By the equivalence between Wasserstein metric and both (1) the weak* convergence in measure and (2) the second moment convergence, we have
$$\begin{aligned} W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \rightarrow 0 \text { a.s. on } \mathcal {F}^0_s, s\in [0,T]. \end{aligned}$$
(28)
By the definition of Wasserstein metric, we obtain
$$\begin{aligned} W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \le \left[ \frac{2}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2+2\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2\right] . \end{aligned}$$
To check the uniform integrability with respect to the conditional expectation on \(\mathcal {F}_s^0\),
$$\begin{aligned}&\mathbb {E}^{\mathcal {F}_s^0}\left[ \frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2\mathbb {I}_{\{\frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2+\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2>C\}}\right] \nonumber \\&\quad \le \mathbb {E}^{\mathcal {F}_s^0}\left[ \frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2\mathbb {I}_{\{\frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2>\frac{C}{2}\}}\right] +\mathbb {E}^{\mathcal {F}_s^0}\left[ \frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2\mathbb {I}_{\{\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2>\frac{C}{2}\}}\right] \nonumber \\&\quad =\frac{1}{N}\sum _{j=1}^N \mathbb {E}^{\mathcal {F}_s^0}\Big [|\hat{x}_1^j(s)|^2\mathbb {I}_{\{\frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2>\frac{C}{2}\}}\Big ]+\frac{1}{N}\sum _{j=1}^N\mathbb {E}^{\mathcal {F}_s^0}\Big [|\hat{x}_1^j(s)|^2\Big ]\mathbb {I}_{\{\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2>\frac{C}{2}\}}\nonumber \\&\quad =\mathbb {E}^{\mathcal {F}_s^0}\Big [|\hat{x}_1^1(s)|^2\mathbb {I}_{\{\frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2>\frac{C}{2}\}}\Big ]+\mathbb {E}^{\mathcal {F}_s^0}\Big [|\hat{x}_1^j(s)|^2\Big ]\mathbb {I}_{\{\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2>\frac{C}{2}\}}, \end{aligned}$$
(29)
where the last equality results from the symmetry on \(\{|\hat{x}_1^j(s)|^2\}_j\). By a simple application of Chebyshev’s Inequality gives
$$\begin{aligned} \mathbb {P}^{\mathcal {F}_s^0}\left[ \frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2>\frac{C}{2}\right] \le \frac{2\mathbb {E}^{\mathcal {F}_s^0}\Big [\frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2\Big ]}{C}\le \frac{2\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2}{C}, \end{aligned}$$
which is independent of N, and goes to 0 as \(C\rightarrow \infty \) on \(\mathcal {F}^0_s\), and hence the first term converges to 0. On the other hand, the second term in (29) vanishes if \(\frac{C}{2}\ge \mathbb {E}^{\mathcal {F}_s^0} |\hat{x}_1^1(s)|^2\triangleq \int _{\mathbb {R}^{n_1}}|x|^2 dm_{x_0}(x,s)\) which is finite almost surely on \(\mathcal {F}^0_s\). We conclude that
$$\begin{aligned} \mathbb {E}^{\mathcal {F}_s^0}\left[ \frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2\mathbb {I}_{\{\frac{1}{N}\sum _{j=1}^N |\hat{x}_1^j(s)|^2+\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2>C\}}\right] \rightarrow 0, \end{aligned}$$
and hence the family \(\{W_2^2(\frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s))\}_N\) is uniformly integrable with respect to the conditional expectation \(\mathbb {E}^{\mathcal {F}_s^0}\). Together with (28), we obtain
$$\begin{aligned} \mathbb {E}^{\mathcal {F}_s^0}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] \rightarrow 0 \text { a.s.} \end{aligned}$$
Next we have
$$\begin{aligned} \mathbb {E}^{\mathcal {F}_s^0}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] \le 4\mathbb {E}^{\mathcal {F}_s^0}|\hat{x}_1^1(s)|^2\le 4\mathbb {E}^{\mathcal {F}_s^0}\left[ \sup _{u\le T}|\hat{x}_1^1(u)|^2\right] . \end{aligned}$$
(30)
On the left hand side, we obtain the standard estimate
$$\begin{aligned} \mathbb {E}\left[ \mathbb {E}^{\mathcal {F}_s^0}\left[ \sup _{u\le T}|\hat{x}_1^1(u)|^2\right] \right] =\mathbb {E}\left[ \sup _{u\le T}|\hat{x}_1^1(u)|^2\right] <\infty \end{aligned}$$
(31)
by the linear growth assumptions on functional coefficients in the evolution of \(\hat{x}_1^i\) and \(x_0\). We conclude that the left hand side of (30) is integrable, and hence by the Dominated Convergence Theorem we have
$$\begin{aligned}&\mathbb {E}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] \\&\quad =\mathbb {E}\left[ \mathbb {E}^{\mathcal {F}_s^0}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] \right] \rightarrow 0, \text { for all } s\in [0,t]. \end{aligned}$$
Finally, observe that the estimate (31) is uniformly in s, another application of the Dominated Convergence Theorem gives
$$\begin{aligned}&\lim _{N\rightarrow \infty }\int _0^t \mathbb {E}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] ds\\&\quad =\int _0^t \lim _{N\rightarrow \infty } \mathbb {E}\left[ W_2^2\left( \frac{1}{N}\sum _{j=1}^N\delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] ds=0. \end{aligned}$$
Similar estimate applies on the second term in Equation (27). Put \(t=T\), we finally have
$$\begin{aligned} \mathbb {E}\displaystyle \sup _{u\le T}|y_0(u)-x_0(u)|^2+\mathbb {E}\displaystyle \sup _{u\le T}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\rightarrow 0, \text { as }N\rightarrow \infty . \end{aligned}$$
\(\square \)
We also have approximation for the cost functionals.
Lemma 5.2
$$\begin{aligned} \mathcal {J}^{N,i}(\hat{\mathbf{u}})-\mathcal {J}^{i}(\hat{u}_1^i)=o(1). \end{aligned}$$
Proof
With the quadratic assumptions (2), we have
$$\begin{aligned}&|\mathcal {J}^{N,i}(\hat{\mathbf{u}})-\mathcal {J}^{i}(\hat{u}^i)|\le \mathbb {E}\left[ \displaystyle \int _0^Tf_1\left( \hat{y}_1^i(t),y_0(t),\dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{y}_1^j(t)},\hat{u}^i(t)\right) \right. \\&\qquad -f_1\left( \hat{x}_1^i(t),x_0(t),m_{x_0}(t),\hat{u}_1^i(t)\right) dt\\&\qquad \left. +h_1\left( \hat{y}_1^i(T),y_0(T),\dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{y}_1^j(T)}\right) -h_1\Big (\hat{x}_1^i(T),\hat{x}_0(T),m_{x_0}(T)\Big )\right] \\&\quad \le C\mathbb {E}\left[ \displaystyle \int _0^T \Big [1+|\hat{y}_1^i(t)|+|\hat{x}_1^i(t)|+|y_0(t)|+|x_0(t)|+\left( \dfrac{\sum _{j=1, j \ne i}^N |\hat{y}_1^j(t)|^2}{N-1}\right) ^{\frac{1}{2}}\right. \\&\qquad \qquad \left. +\Big (\mathbb {E}^{\mathcal {F}_t^0}|\hat{x}_1^i(t)|^2\Big )^{\frac{1}{2}}+2|\hat{u}^i(t)|\right] \\&\qquad \qquad \cdot \left[ |\hat{y}_1^i(t)-\hat{x}_1^i(t)|+|y_0(t)-x_0(t)|+W_2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{y}_1^j(t)},m_{x_0}(t)\right) \right] dt\\&\qquad +\left[ 1+|\hat{y}_1^i(T)|+|\hat{x}_1^i(T)|+|y_0(T)|+|x_0(T)|+\left( \dfrac{\sum _{j=1, j \ne i}^N |\hat{y}_1^j(T)|^2}{N-1}\right) ^{\frac{1}{2}}\right. \\&\qquad \qquad \left. +\Big (\mathbb {E}^{\mathcal {F}_t^0}|\hat{x}_1^i(T)|^2\Big )^{\frac{1}{2}}\right] \\&\qquad \qquad \cdot \left[ |\hat{y}_1^i(T)-\hat{x}_1^i(T)|+|y_0(T)-x_0(T)|\right. \\&\qquad \qquad \left. \left. +W_2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{y}_1^j(T)},m_{x_0}(T)\right) \right] \right] . \end{aligned}$$
An application of Hölder’s inequality, and the symmetry on \(\{\hat{x}_1^j\}_j\) gives
$$\begin{aligned}&|\mathcal {J}^{N,i}(\hat{\mathbf{u}})-\mathcal {J}^{i}(\hat{u}^i)| \le C\left\{ \left[ \mathbb {E}\displaystyle \int _0^T \left[ 1+|\hat{y}_1^i(t)|^2+|\hat{x}_1^i(t)|^2+|y_0(t)|^2+|x_0(t)|^2\right. \right. \right. \\&\quad \left. \left. +\,\dfrac{\sum _{j=1, j \ne i}^N |\hat{y}_1^j(t)|^2}{N-1}+\mathbb {E}^{\mathcal {F}_t^0}|\hat{x}_1^i(t)|^2+|\hat{u}^i(t)|^2\right] dt\right] ^{\frac{1}{2}}\\&\qquad \qquad \cdot \left[ \mathbb {E}\displaystyle \int _0^T\left[ |\hat{y}_1^i(t)-\hat{x}_1^i(t)|^2+|y_0(t)-x_0(t)|^2\right. \right. \\&\qquad \qquad \left. \left. +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{y}_1^j(t)},m_{x_0}(t)\right) \right] dt\right] ^{\frac{1}{2}}\\&\qquad +\left[ \mathbb {E}\left[ 1+|\hat{y}_1^i(T)|^2+|\hat{x}_1^i(T)|^2+|y_0(T)|^2+|x_0(T)|^2\right. \right. \\&\qquad \left. \left. +\dfrac{\sum _{j=1, j \ne i}^N |\hat{y}_1^j(T)|^2}{N-1}+\mathbb {E}^{\mathcal {F}_t^0}|\hat{x}_1^i(T)|^2\right] \right] ^{\frac{1}{2}}\\&\qquad \qquad \qquad \cdot \left[ \mathbb {E}\left[ |\hat{y}_1^i(T)-\hat{x}_1^i(T)|^2+|y_0(T)-x_0(T)|^2\right. \right. \\&\qquad \qquad \left. \left. \left. +W^2_2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{y}_1^j(T)},m_{x_0}(T)\right) \right] \right] ^{\frac{1}{2}}\right\} \\= & {} C\left\{ \bigg [\mathbb {E}\displaystyle \int _0^T \Big [1+|\hat{y}_1^i(t)|^2+|\hat{x}_1^i(t)|^2+|y_0(t)|^2+|x_0(t)|^2+|\hat{u}^i(t)|^2\Big ]dt\bigg ]^{\frac{1}{2}}\right. \\&\qquad \qquad \qquad \cdot \left[ \mathbb {E}\displaystyle \int _0^T\left[ |\hat{y}_1^i(t)-\hat{x}_1^i(t)|^2+|y_0(t)-x_0(t)|^2\right. \right. \\&\qquad \qquad \left. \left. +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{x}_1^j(t)},m_{x_0}(t)\right) \right] dt\right] ^{\frac{1}{2}}\\&\qquad +\bigg [\mathbb {E}\Big [1+|\hat{y}_1^i(T)|^2+|\hat{x}_1^i(T)|^2+|y_0(T)|^2+|x_0(T)|^2\Big ]\bigg ]^{\frac{1}{2}}\\&\qquad \qquad \qquad \cdot \left[ \mathbb {E}\Big [|\hat{y}_1^i(T)-\hat{x}_1^i(T)|^2+|y_0(T)-x_0(T)|^2\right. \\&\qquad \qquad \left. \left. \left. +W^2_2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{x}_1^j(T)},m_{x_0}(T)\right) \right] \right] ^{\frac{1}{2}}\right\} \end{aligned}$$
By the linear growth assumptions on \(g_0\), \(\sigma _0\), \(g_1\) and \(\sigma _1\), it is easy to show that
$$\begin{aligned} \mathbb {E}\displaystyle \int _0^T \Big [1+|\hat{y}_1^i(t)|^2+|\hat{x}_1^i(t)|^2+|y_0(t)|^2+|x_0(t)|^2+|\hat{u}^i(t)|^2\Big ]dt \end{aligned}$$
and
$$\begin{aligned} \mathbb {E}\Big [1+|\hat{y}_1^i(T)|^2+|\hat{x}_1^i(T)|^2+|y_0(T)|^2+|x_0(T)|^2\Big ] \end{aligned}$$
are bounded (independent of N). We finally arrive at the estimates
$$\begin{aligned}&|\mathcal {J}^{N,i}(\hat{\mathbf{u}})-\mathcal {J}^{i}(\hat{u}^i)|\\&\quad \le C\left\{ \left[ \mathbb {E}\displaystyle \int _0^T\left[ |\hat{y}_1^i(t)-\hat{x}_1^i(t)|^2+|y_0(t)-x_0(t)|^2\right. \right. \right. \\&\qquad \left. \left. +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{x}_1^j(t)},m_{x_0}(t)\right) \right] dt\right] ^{\frac{1}{2}}\\&\qquad +\left[ \mathbb {E}\left[ |\hat{y}_1^i(T)-\hat{x}_1^i(T)|^2+|y_0(T)-x_0(T)|^2\right. \right. \\&\qquad \left. \left. \left. +W^2_2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=1, j \ne i}^N\delta _{\hat{x}_1^j(T)},m_{x_0}(T)\right) \right] \right] ^{\frac{1}{2}}\right\} , \end{aligned}$$
which goes to 0 as \(N\rightarrow \infty \), as shown in Lemma 5.1. Hence
$$\begin{aligned} |\mathcal {J}^{N,i}(\hat{\mathbf{u}})-\mathcal {J}^{i}(\hat{u}_1^i)|=o(1). \end{aligned}$$
\(\square \)
In the previous lemmas, we assumed that all players adopt their corresponding mean field optimal controls. By symmetry, the convergences of state dynamics and the cost functionals are then established. To show that the mean field optimal controls \(\mathbf{u}\) indeed constitute a \(\epsilon \)-Nash equilibrium on the empirical system, without loss of generality, we assume that the first player did not obey the mean field optimal control. In particular, let \(u_1^1\) be an arbitrary control in \(\mathcal {A}_1\), define \(\mathbf{u}{:}=(u_1^1,\hat{u}_1^2,\dots ,\hat{u}_1^N)\). We then have the following empirical and mean field SDEs for the dominating player, the 1-st player and the i-th player (\(i>1\)) respectively:
$$\begin{aligned}&\left\{ \begin{array}{rcl} dy_0&{}=&{}g_0\left( y_0(t),\dfrac{1}{N}\left( \delta _{ y_1^1(t)}+\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(t)}\right) ,u_0(t)\right) dt+\sigma _0\left( y_0(t)\right) dW_0(t),\\ y_0(0)&{}=&{}\xi _0.\\ dx_0&{}=&{}g_0\left( x_0(t),m_{x_0}(t),u_0(t)\right) dt+\sigma _0\left( x_0(t)\right) dW_0(t),\\ x_0(0)&{}=&{}\xi _0.\\ \end{array}\right. \\&\left\{ \begin{array}{rcl} dy_1^1&{}=&{}g_1\left( y_1^1(t),y_0(t),\dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(t)},u_1^1(t)\right) dt+\sigma _1\left( y_1^1(t)\right) dW_1^1(t),\\ y_1^1(0)&{}=&{}\xi _1^1.\\ dx_1^1&{}=&{}g_1\Big (x_1^1(t),x_0(t),m_{x_0}(t),u_1^1(t)\Big )dt+\sigma _1\Big (x_1^1(t)\Big )dW_1^1(t),\\ x_1^1(0)&{}=&{}\xi _1^1.\\ \end{array}\right. \\&\left\{ \begin{array}{rcl} d\hat{y}_1^i&{}=&{}g_1\left( \hat{y}_1^i(t),y_0(t),\dfrac{1}{N-1}\left( \delta _{ y_1^1(t)}+\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(t)}\right) ,\hat{u}_1^i(t)\right) dt\\ &{}&{}+\sigma _1\left( \hat{y}_1^i(t)\right) dW_1^i(t),\\ \hat{y}_1^i(0)&{}=&{}\xi _1^i.\\ d\hat{x}_1^i&{}=&{}g_1\left( \hat{x}_1^i(t),x_0(t),m_{x_0}(t),\hat{u}_1^i(t)\right) dt+\sigma _1\left( \hat{x}_1^i(t)\right) dW_1^i(t),\\ \hat{x}_1^i(0)&{}=&{}\xi _1^i.\\ \end{array}\right. \end{aligned}$$
We claim that if \(m_{x_0}\) is the density function of \(\hat{x}_1^i\) conditioning on \(\mathcal {F}^0\), then we have the convergence \(y_0\rightarrow x_0\), \(y_1^1\rightarrow x_1^1\) and \(\hat{y}_1^i \rightarrow \hat{x}_1^i\) in the sense of the following lemma
Lemma 5.3
$$\begin{aligned}&\mathbb {E}\Big [\displaystyle \sup _{u \le T}|y_0(u)-x_0(u)|^2\Big ]+\mathbb {E}\Big [\displaystyle \sup _{u \le T}|y_1^1(u)-x_1^1(u)|^2\Big ]\\&\quad +\mathbb {E}\Big [\displaystyle \sup _{u \le T}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\Big ]\rightarrow 0, \quad \text { as } N \rightarrow \infty . \end{aligned}$$
Proof
We first show the convergence of the dominating player and the i-th player. Similar to the proof of Lemma 5.1, we first have
$$\begin{aligned}&\mathbb {E}\displaystyle \sup _{u\le t}|y_0(u)-x_0(u)|^2+\mathbb {E}\displaystyle \sup _{u\le t}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\nonumber \\&\quad \le C\left\{ t\mathbb {E}\displaystyle \int _0^t \Big |y_0(s)-x_0(s)\Big |^2+W_2^2\left( \dfrac{1}{N}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) ,m_{x_0}(s)\right) ds\right. \nonumber \\&\qquad +\mathbb {E}\displaystyle \int _0^t\Big |y_0(s)-x_0(s)\Big |^2\bigg \}ds\nonumber \\&\qquad +C\bigg \{t\mathbb {E}\displaystyle \int _0^t \Big |\hat{y}_1^i(s)-\hat{x}_1^i(s)\Big |^2+\Big |y_0(s)-x_0(s)\Big |^2\nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(s)}\right) ,m_{x_0}(s)\right) ds+\mathbb {E}\displaystyle \int _0^t\Big |\hat{y}_1^i(s)-\hat{x}_1^i(s)\Big |^2\bigg \}ds\nonumber \\&\quad \le C\mathbb {E}\displaystyle \int _0^t \left[ \displaystyle \sup _{u\le s}\Big |y_0(u)-x_0(u)\Big |^2+\displaystyle \sup _{u\le s}\Big |\hat{y}_1^i(u)-\hat{x}_1^i(u)\Big |^2\nonumber \right. \\&\qquad +W_2^2\left( \dfrac{1}{N}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) ,\dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) \nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{x}_1^j(s)}\right) +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(s)}\right) ,\dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(s)}\right) \nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{x}_1^j(s)}\right) \nonumber \\&\qquad \left. +W_2^2\left( \dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] ds \end{aligned}$$
(32)
By the same argument used in Equation (26) in Lemma 5.1, we have
$$\begin{aligned}&\mathbb {E}\left[ W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{x}_1^j(s)}\right) \right] \\&\qquad +\mathbb {E}\left[ W_2^2\left( \dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(s)},\dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{x}_1^j(s)}\right) \right] \le 2\mathbb {E}|\hat{y}_1^i(s)-\hat{x}_1^i|^2. \end{aligned}$$
Hence, by applying Gronwall’s inequality on Equation (32), we have
$$\begin{aligned}&\mathbb {E}\displaystyle \sup _{u\le t}|y_0(u)-x_0(u)|^2+\mathbb {E}\displaystyle \sup _{u\le t}|\hat{y}_1^i(u)-\hat{x}_1^i(u)|^2\nonumber \\&\quad \le Ce^{Ct}\mathbb {E}\displaystyle \int _0^t \left[ W_2^2\left( \dfrac{1}{N}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) ,\dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) \right. \nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \nonumber \\&\qquad +W_2^2\left( \dfrac{1}{N-1}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(s)}\right) ,\dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{y}_1^j(s)}\right) \nonumber \\&\qquad \left. +W_2^2\left( \dfrac{1}{N-2}\displaystyle \sum _{j=2,j\ne i}^N \delta _{\hat{x}_1^j(s)},m_{x_0}(s)\right) \right] ds \end{aligned}$$
(33)
For the first term in (33), consider the following joint measure on \(\mathbb {R}^{n_1}\times \mathbb {R}^{n_1}\)
$$\begin{aligned} \mu (x,y)=\dfrac{1}{N}\displaystyle \sum _{j=2}^N\delta _{(\hat{y}_1^j(s),\hat{y}_1^j(s))}(x,y)+\dfrac{1}{N(N-1)}\displaystyle \sum _{j=2}^N\delta _{(y_1^1(s),\hat{y}_1^j(s))}(x,y), \end{aligned}$$
which has respective marginals
$$\begin{aligned} \dfrac{1}{N}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) \quad \text {and}\quad \dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}. \end{aligned}$$
By the definition of Wasserstein metric,
$$\begin{aligned}&\mathbb {E}\left[ W_2^2\left( \dfrac{1}{N}\left( \delta _{ y_1^1(s)}+\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) ,\dfrac{1}{N-1}\displaystyle \sum _{j=2}^N \delta _{\hat{y}_1^j(s)}\right) \right. \\&\quad \le \mathbb {E}\left[ \displaystyle \int _{\mathbb {R}^{n_1}\times \mathbb {R}^{n_1}}|x-y|^2d\mu (x,y)\right] \\&\quad =\mathbb {E}\left[ \dfrac{1}{N(N-1)}\displaystyle \sum _{j=2}^N|y_1^1(s)-\hat{y}_1^j(s)|^2\right] \\&\quad =\dfrac{1}{N}\mathbb {E}|y_1^1(s)-\hat{y}_1^2(s)|^2, \end{aligned}$$
where the last equality results from symmetry on \(\{y_1^j\}_j\), clearly goes to 0 as \(N\rightarrow \infty \). Similar argument applies for the third term in (33). For the convergence of the second and the fourth term, we refer to the argument in the last part of Lemma 5.1 and the results follow.
For the convergence of the 1-st player, the procedure is similar and we do not provide here. \(\square \)
We conclude from the similar procedures to show the convergence of the cost functional. In particular, we have
$$\begin{aligned} |\mathcal {J}^{N,1}(\mathbf{u})-\mathcal {J}^{1}(u_1^1)|=o(1). \end{aligned}$$
Theorem 5.4
\(\hat{\mathbf{u}}\) is an \(\epsilon \)-Nash equilibrium.
Proof
Summarizing all the obtained results in this section, we can conclude
$$\begin{aligned}&|\mathcal {J}^{N,1}(\hat{\mathbf{u}})-\mathcal {J}^{1}(\hat{u}_1^1)|=o(1);\\&|\mathcal {J}^{N,1}(\mathbf{u})-\mathcal {J}^{1}( u_1^1)|=o(1). \end{aligned}$$
Since \(\hat{u}_1^1\) is optimal control, we have \(\mathcal {J}^{1}(\hat{u}_1^1) \le \mathcal {J}^{1}( u_1^1)\). We deduce
$$\begin{aligned} \mathcal {J}^{N,i}(\hat{\mathbf{u}}) \le \mathcal {J}^{N,1}(\mathbf{u})+o(1). \end{aligned}$$
Hence, \(\hat{\mathbf{u}}\) is an \(\epsilon \)-Nash equilibrium. \(\square \)