Introduction
Random matrix theory studies the spectral properties of matrices with random entries and was initiated by the pioneering works of Wishart [43 ] and Wigner [41 , 42 ] . One of its main objectives is to understand the behavior of empirical spectral measures. Two foundational results in this direction are Wigner’s semicircle law for symmetric or Hermitian matrices and the Marchenko–Pastur law for sample covariance matrices, established by Wigner [42 ] and Marchenko–Pastur [26 ] , respectively. Random matrix methods are now widely used in statistics, finance, communications, and machine learning, see, e.g., [3 , 20 , 22 , 31 , 40 , 7 ] .
A fundamental dynamical model in random matrix theory is Dyson Brownian motion [10 ] , which describes the eigenvalue processes of symmetric or Hermitian matrix-valued Brownian motions. The eigenvalues evolve as interacting particles with pairwise repulsion generated by a logarithmic potential, thereby connecting matrix diffusions, Coulomb gases, and universality theory [11 ] . Another important class is given by Wishart processes [4 , 5 ] , which model evolving covariance matrices and give rise to nonnegative eigenvalue dynamics.
Graczyk and Małecki [14 , 15 ] introduced the so-called generalized Wishart processes as solutions to a class of matrix-valued SDEs that encompass, in particular, Dyson Brownian motion and the classical Wishart process. Subsequently, the high-dimensional spectral limits of these processes were investigated in [35 ] , and the fluctuations around the limiting spectral distributions were studied in [36 ] . We also refer to the survey [37 ] for an overview of matrix-valued stochastic processes.
The above-mentioned works are formulated on a probability space equipped with a single probability measure. In many applications, however, the probabilistic model itself is uncertain. For example, the volatility or covariance structure of a financial system may not be known precisely and may vary over a prescribed family of plausible models. This motivates the use of sublinear expectations in order to capture model ambiguity [2 , 6 ] . Denis and Martini [9 ] established one of the earliest general duality results for model uncertainty in continuous-time finance. Their work laid important foundations for the subsequent development of robust finance and is closely related in spirit to the theory of G G G -expectation. Peng’s G G G -expectation framework [28 , 29 ] provides a canonical setting for modeling volatility uncertainty and gives rise to G G G -Brownian motion and a nonlinear stochastic calculus [8 , 33 ] . We refer to the monograph [30 ] for a systematic treatment of the theory of G G G -expectation and its applications.
Compared with the extensive literature on (backward) G G G -stochastic differential equations (see, e.g., [12 , 13 , 17 , 23 , 16 , 30 ] ), random matrix theory under nonlinear expectations remains at an early stage. Initial work by Stihi, Boutabia, and Meradji [39 ] and Meradji, Boutabia, and Stihi [27 ] introduced matrix-valued G G G -Brownian motion and G G G -Wishart processes, deriving eigenvalue dynamics and noncollision conditions. However, large-dimensional spectral limits and nonlinear analogues of classical laws remain largely open.
In this paper, we investigate the spectral dynamics of generalized Wishart processes within the framework of nonlinear expectation and derive the McKean–Vlasov equations governing the high-dimensional limits of their empirical spectral measures. We focus in particular on two important special cases, the symmetric G G G -Brownian matrix and the G G G -Wishart process, and establish the corresponding semicircle law and Marchenko–Pastur law under G G G -expectation.
Main results on dynamics of spectral processes
We first present our main results on the equation satisfied by the eigenvalue processes of generalized Wishart process in the framework of G G G -expectation.
Let S d \mathbb{S}_d S d denote the space of d × d d \times d d × d symmetric matrices. A d d d -dimensional G G G -Brownian motion B = ( B 1 , … , B d ) B=(B^1,\ldots,B^d) B = ( B 1 , … , B d ) is characterized by its generator G : S d → R G:\mathbb{S}_d\to\mathbb{R} G : S d → R , which is a monotone sublinear function (see § 2 for details). In this work, we impose the following condition on G G G :
Assumption 1. Assume
G ( A ) = g ( tr ( A ) ) = g ( ∑ i = 1 d A i i ) , G(A)=g(\operatorname{tr}(A))=g\left(\sum_{i=1}^{d}A_{ii}\right), G ( A ) = g ( tr ( A )) = g ( i = 1 ∑ d A ii ) , for A = ( A i j ) 1 ≤ i , j ≤ d ∈ S d A=(A_{ij})_{1\leq i,j\leq d}\in\mathbb{S}_d A = ( A ij ) 1 ≤ i , j ≤ d ∈ S d , where
g ( x ) = 1 2 ( σ ‾ 2 x + − σ ‾ 2 x − ) , x ∈ R , (1) g(x)=\frac{1}{2}\left(\overline{\sigma}^{2}x^{+}-\underline{\sigma}^{2}x^{-}\right),\qquad x\in\mathbb{R}, \tag*{(1)} g ( x ) = 2 1 ( σ 2 x + − σ 2 x − ) , x ∈ R , (1) with 0 < σ ‾ ≤ σ ‾ < ∞ 0<\underline{\sigma}\leq\overline{\sigma}<\infty 0 < σ ≤ σ < ∞ .
When σ ‾ = σ ‾ = 1 \underline{\sigma}=\overline{\sigma}=1 σ = σ = 1 in Assumption 1, the G G G -Brownian motion B B B reduces to a classical d d d -dimensional Brownian motion with independent components. Moreover, by [19 Lemma 3.2] , under Assumption 1, we have for 1 ≤ i ≠ j ≤ d 1\leq i\ne j\leq d 1 ≤ i = j ≤ d ,
⟨ B 1 ⟩ t = ⟨ B i ⟩ t and ⟨ B i , B j ⟩ t = 0 , quasi-surely for all t > 0. (2) \langle B^{1}\rangle_t=\langle B^{i}\rangle_t\quad\text{and}\quad\langle B^{i},B^{j}\rangle_t=0,\qquad\text{quasi-surely for all }t>0. \tag*{(2)} ⟨ B 1 ⟩ t = ⟨ B i ⟩ t and ⟨ B i , B j ⟩ t = 0 , quasi-surely for all t > 0. (2)
Conversely, if a d d d -dimensional G G G -Brownian motion B B B satisfies (2) , then G G G satisfies Assumption 1. In § 3 , we provide a preliminary study of G G G -Dyson Brownian motion, which suggests that Assumption 1 is a natural condition for studying the generalized G G G -Wishart process (1.3)`` below (see also Section 3.3`).
To describe the underlying matrix dynamics, we consider the following generalized G G G -Wishart process. Let X = ( X t ) t ≥ 0 X=(X_t)_{t\geq0} X = ( X t ) t ≥ 0 be an S N \mathbb{S}_N S N -valued stochastic process satisfying
d X t N = g N ( X t N ) d B t h N ( X t N ) + h N ( X t N ) d B t T g N ( X t N ) + b N ( X t N ) d ⟨ B 11 ⟩ t , t ≥ 0 , (3) dX_t^N=g_N(X_t^N)\,dB_t h_N(X_t^N)+h_N(X_t^N)\,dB_t^{\mathsf{T}}g_N(X_t^N)+b_N(X_t^N)\,d\langle B^{11}\rangle_t,\qquad t\geq0, \tag*{(3)} d X t N = g N ( X t N ) d B t h N ( X t N ) + h N ( X t N ) d B t T g N ( X t N ) + b N ( X t N ) d ⟨ B 11 ⟩ t , t ≥ 0 , (3)
where g N , h N , b N : R → R g_N,h_N,b_N:\mathbb{R}\to\mathbb{R} g N , h N , b N : R → R are continuous functions acting on the spectrum of X t N X_t^N X t N and B t = ( B t k h ) 1 ≤ k , h ≤ N B_t=(B_t^{kh})_{1\leq k,h\leq N} B t = ( B t k h ) 1 ≤ k , h ≤ N is a d d d -dimensional G G G -Brownian motion satisfying Assumption 1 with d = N 2 d=N^2 d = N 2 .
The generalized G G G -Wishart process (3) includes two important special cases: the symmetric G G G -Brownian matrix, whose eigenvalue process is the G G G -Dyson Brownian motion, and the G G G -Wishart process, which has potential applications in statistics and finance. Indeed, if we take g N ( x ) = ( 2 N ) − 1 / 2 g_N(x)=(2N)^{-1/2} g N ( x ) = ( 2 N ) − 1/2 , h N ( x ) = 1 h_N(x)=1 h N ( x ) = 1 and b N ( x ) = 0 b_N(x)=0 b N ( x ) = 0 in (3) , then we get
d X t N = 1 2 N ( d B t + d B t T ) . (4) dX_t^N=\frac{1}{\sqrt{2N}}\left(dB_t+dB_t^{\mathsf{T}}\right). \tag*{(4)} d X t N = 2 N 1 ( d B t + d B t T ) . (4)
Denote
d B ~ k h = { 1 2 ( d B k h + d B h k ) , if k < h , d B h h , if k = h . d\widetilde{B}_{kh}=
\begin{cases}
\frac{1}{\sqrt{2}}(dB_{kh}+dB_{hk}), & \text{if } k<h,\\
dB_{hh}, & \text{if } k=h.
\end{cases} d B k h = { 2 1 ( d B k h + d B hk ) , d B hh , if k < h , if k = h .
Under Assumption 1 with d = N × N d=N\times N d = N × N , we have that ⟨ B ~ k h ⟩ t = ⟨ 1 2 ( B k h + B h k ) ⟩ t = ⟨ B 11 ⟩ t \langle\widetilde{B}_{kh}\rangle_t=\left\langle\frac{1}{\sqrt{2}}(B_{kh}+B_{hk})\right\rangle_t=\langle B_{11}\rangle_t ⟨ B k h ⟩ t = ⟨ 2 1 ( B k h + B hk ) ⟩ t = ⟨ B 11 ⟩ t for k < h k<h k < h . Then ⟨ B ~ k h , B ~ k ′ h ′ ⟩ t = I { ( k , h ) = ( k ′ , h ′ ) } ⟨ B ~ 11 ⟩ t \langle\widetilde{B}_{kh},\widetilde{B}_{k'h'}\rangle_t=I_{\{(k,h)=(k',h')\}}\langle\widetilde{B}_{11}\rangle_t ⟨ B k h , B k ′ h ′ ⟩ t = I {( k , h ) = ( k ′ , h ′ )} ⟨ B 11 ⟩ t , and by Corollary 2.14 it is easy to verify that B ~ \widetilde{B} B is a N ( N + 1 ) 2 \frac{N(N+1)}{2} 2 N ( N + 1 ) -dimensional G ‾ \overline{G} G -Brownian motion with G ‾ \overline{G} G satisfying Assumption 1 with d = N ( N + 1 ) 2 d=\frac{N(N+1)}{2} d = 2 N ( N + 1 ) .
Hence, X N X^N X N given by (4) can be represented the following N × N N \times N N × N symmetric G G G -Brownian matrix whose upper-triangular entries satisfy
d X N = { 1 N d B ~ k h , if k < h , 2 N d B ~ h h , if k = h . (5) \mathrm{d}X^N =
\begin{cases}
\frac{1}{\sqrt{N}}\mathrm{d}\widetilde{B}^{kh}, & \text{if } k<h,\\
\frac{\sqrt{2}}{\sqrt{N}}\mathrm{d}\widetilde{B}^{hh}, & \text{if } k=h.
\end{cases} \tag*{(5)} d X N = { N 1 d B k h , N 2 d B hh , if k < h , if k = h . (5)
If we instead choose g N ( x ) = x g_N(x)=\sqrt{x} g N ( x ) = x , h N ( x ) = 1 h_N(x)=1 h N ( x ) = 1 , and b N ( x ) = p b_N(x)=p b N ( x ) = p , then X t N X_t^N X t N is the G G G -Wishart process B t T B t B_t^{\mathsf{T}}B_t B t T B t (see Lemma (4.2)), where B t B_t B t is an p × N p \times N p × N G G G -Brownian matrix.
The well-posedness of the generalized Wishart process (3) is assumed throughout this paper and is not investigated in full generality. Existence and uniqueness are immediate in the symmetric G G G -Brownian matrix case and are established for the G G G -Wishart process in Theorem 1.5 .
The following result characterizes the dynamics of the eigenvalue vector λ = ( λ 1 , … , λ N ) \lambda=(\lambda_1,\ldots,\lambda_N) λ = ( λ 1 , … , λ N ) of X N X^N X N under Assumption 1 and establishes the quasi-sure non-collision property of the eigenvalue processes. The proof is provided in § 4.1 .
Theorem 1.1. Let X N X^N X N be given by (3) , assuming that the initial eigenvalues are distinct and that X t N X_t^N X t N has a quasi-continuous version for each t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] . Assume that b N b_N b N , g N 2 g_N^2 g N 2 , h N 2 h_N^2 h N 2 are Lipschitz continuous functions and that g N 2 g_N^2 g N 2 is either convex or belongs to the class C 1 , 1 ( R ) C^{1,1}(\mathbb{R}) C 1 , 1 ( R ) , where C 1 , 1 ( R ) C^{1,1}(\mathbb{R}) C 1 , 1 ( R ) denotes the space of differentiable functions whose derivatives are Lipschitz continuous. We further assume that the underlying G G G -Brownian motion satisfies Assumption 1 with d = N × N d=N \times N d = N × N .
Let λ 1 N ( t ) , … , λ N N ( t ) \lambda_1^N(t),\ldots,\lambda_N^N(t) λ 1 N ( t ) , … , λ N N ( t ) denote the eigenvalues of X t N X_t^N X t N , and define the first collision time by
τ : = inf { t ≥ 0 : λ i N ( t ) = λ j N ( t ) for some i ≠ j } . (6) \tau:=\inf\{t\geq0:\lambda_i^N(t)=\lambda_j^N(t)\text{ for some }i\ne j\}. \tag*{(6)} τ := inf { t ≥ 0 : λ i N ( t ) = λ j N ( t ) for some i = j } . (6)
Then, τ = ∞ \tau=\infty τ = ∞ quasi-surely. Moreover, the eigenvalue processes, arranged in increasing order as λ 1 N ( t ) < λ 2 N ( t ) < ⋯ < λ N N ( t ) \lambda_1^N(t)<\lambda_2^N(t)<\cdots<\lambda_N^N(t) λ 1 N ( t ) < λ 2 N ( t ) < ⋯ < λ N N ( t ) , satisfy the following G G G -SDE: for i = 1 , 2 , … , N i=1,2,\ldots,N i = 1 , 2 , … , N ,
d λ i N = 2 g N ( λ i N ) h N ( λ i N ) d B ‾ t i + b N ( λ i N ) d ⟨ B ‾ i ⟩ t + ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N d ⟨ B ‾ i ⟩ t , (7) \mathrm{d}\lambda_i^N=2g_N(\lambda_i^N)h_N(\lambda_i^N)\mathrm{d}\overline{B}_t^i+b_N(\lambda_i^N)\mathrm{d}\langle\overline{B}^i\rangle_t+\sum_{j:j\ne i}\frac{H_N(\lambda_i^N,\lambda_j^N)}{\lambda_i^N-\lambda_j^N}\mathrm{d}\langle\overline{B}^i\rangle_t, \tag*{(7)} d λ i N = 2 g N ( λ i N ) h N ( λ i N ) d B t i + b N ( λ i N ) d ⟨ B i ⟩ t + j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) d ⟨ B i ⟩ t , (7)
where B ‾ = ( B ‾ 1 , … , B ‾ N ) \overline{B}=(\overline{B}^1,\ldots,\overline{B}^N) B = ( B 1 , … , B N ) is an N N N -dimensional G ‾ \overline{G} G -Brownian motion with G ‾ \overline{G} G satisfying Assumption 1 with d = N d=N d = N , and
H N ( x , y ) : = g N 2 ( x ) h N 2 ( y ) + g N 2 ( y ) h N 2 ( x ) . (8) H_N(x,y):=g_N^2(x)h_N^2(y)+g_N^2(y)h_N^2(x). \tag*{(8)} H N ( x , y ) := g N 2 ( x ) h N 2 ( y ) + g N 2 ( y ) h N 2 ( x ) . (8)
The well-posedness of the G G G -SDE (7) is established below, of which the proof is given in § 4.1 .
Theorem 1.2. Assume the same conditions as in Theorem 1.1 . Then, there exists a unique strong solution to the G G G -SDE (7) .
As a consequence of Theorem 1.1 , we have the following result for G G G -Dyson Brownian motion:
Theorem 1.3. Let X N X^N X N be given by (4) and τ \tau τ be given by (6) . Suppose that the initial eigenvalues are distinct. We further assume that the underlying G G G -Brownian motion satisfies Assumption 1 with d = N ( N + 1 ) 2 d=\frac{N(N+1)}{2} d = 2 N ( N + 1 ) . Then, τ = ∞ \tau=\infty τ = ∞ quasi-surely. Moreover, the eigenvalue processes, arranged in increasing order as λ 1 N ( t ) < λ 2 N ( t ) < ⋯ < λ N N ( t ) \lambda_1^N(t)<\lambda_2^N(t)<\cdots<\lambda_N^N(t) λ 1 N ( t ) < λ 2 N ( t ) < ⋯ < λ N N ( t ) , satisfy the following G G G -SDE: for i = 1 , … , N i=1,\ldots,N i = 1 , … , N ,
d λ i N = 2 N d B ‾ i + 1 N ∑ j : j ≠ i 1 λ i N − λ j N d ⟨ B ‾ i ⟩ t , (9) \mathrm{d}\lambda_i^N=\frac{\sqrt{2}}{\sqrt{N}}\mathrm{d}\overline{B}^i+\frac{1}{N}\sum_{j:j\ne i}\frac{1}{\lambda_i^N-\lambda_j^N}\mathrm{d}\langle\overline{B}^i\rangle_t, \tag*{(9)} d λ i N = N 2 d B i + N 1 j : j = i ∑ λ i N − λ j N 1 d ⟨ B i ⟩ t , (9)
where B ‾ = ( B ‾ 1 , … , B ‾ N ) \overline{B}=(\overline{B}^1,\ldots,\overline{B}^N) B = ( B 1 , … , B N ) is an N N N -dimensional G ‾ \overline{G} G -Brownian motion with G ‾ \overline{G} G satisfying Assumption 1 with d = N d=N d = N .
Theorem 1.2 establishes the existence and uniqueness of a strong solution to the G G G -SDE (9) when the initial eigenvalues are distinct. In fact, this result remains valid even when the initial eigenvalues are allowed to coincide, as shown in the following proposition, of which the proof is given in Section 4.2. Set Δ N : = { x ∈ R N : x 1 < ⋯ < x N } \Delta_N := \{x \in\mathbb{R}^N : x_1 < \cdots< x_N\} Δ N := { x ∈ R N : x 1 < ⋯ < x N } and denote by Δ ‾ N \overline{\Delta}_N Δ N its closure.
Proposition 1.4. Assume Assumption 1 for the underlying G G G -Brownian motion. Then, there exists a unique solution λ N ∈ C ( R + ; Δ ‾ N ) \lambda^N \in C(\mathbb{R}_+;\overline{\Delta}_N) λ N ∈ C ( R + ; Δ N ) to the G G G -SDE (9) with initial condition λ N ( 0 ) ∈ Δ ‾ N \lambda^N(0) \in\overline{\Delta}_N λ N ( 0 ) ∈ Δ N . Moreover, λ N ( t ) ∈ Δ N \lambda^N(t) \in\Delta_N λ N ( t ) ∈ Δ N for all t > 0 t > 0 t > 0 quasi-surely.
An N × N N \times N N × N Wishart process with p p p degrees of freedom is defined by
X t N : = Y t T Y t ∈ S N + , (10) X_t^N := Y_t^{\mathsf{T}}Y_t \in S_N^+, \tag*{(10)} X t N := Y t T Y t ∈ S N + , (10)
where Y t Y_t Y t is an p × N p \times N p × N G ~ \widetilde{G} G -Brownian matrix and S N + S_N^+ S N + denotes the space of N × N N \times N N × N nonnegative definite matrices. Assuming that G ~ \widetilde{G} G satisfies Assumption 1 with d = p × N d = p \times N d = p × N , by Lemma (48) , X t N X_t^N X t N satisfies the following G G G -SDE:
d X t N = X t N d B t + d B t T X t N + p I N d ⟨ B 11 ⟩ t , (11) \mathrm{d}X_t^N = \sqrt{X_t^N}\,\mathrm{d}B_t + \mathrm{d}B_t^{\mathsf{T}}\sqrt{X_t^N} + pI_N\mathrm{d}\langle B^{11}\rangle_t, \tag*{(11)} d X t N = X t N d B t + d B t T X t N + p I N d ⟨ B 11 ⟩ t , (11)
where B B B is an N × N N \times N N × N G G G -Brownian matrix with G G G satisfying Assumption 1 with d = N × N d = N \times N d = N × N . We consider the following general G G G -Wishart process with parameter α \alpha α :
{ d X t N = X t N d B t + d B t T X t N + α I N d ⟨ B 11 ⟩ t , X 0 N = x 0 . (12) \begin{cases}
\mathrm{d}X_t^N = \sqrt{X_t^N}\,\mathrm{d}B_t + \mathrm{d}B_t^{\mathsf{T}}\sqrt{X_t^N} + \alpha I_N\mathrm{d}\langle B^{11}\rangle_t,\\
X_0^N = x_0. \tag*{(12)}
\end{cases} { d X t N = X t N d B t + d B t T X t N + α I N d ⟨ B 11 ⟩ t , X 0 N = x 0 . (12)
The following result establishes the existence and uniqueness of a positive-definite solution to the G G G -SDE (12) , the proof of which is given in Section 4.3.
Theorem 1.5. Assuming α ≥ N + 1 \alpha\ge N + 1 α ≥ N + 1 , the G G G -SDE (12) with x 0 ∈ S N + + x_0 \in S_N^{++} x 0 ∈ S N ++ , where S N + + S_N^{++} S N ++ denotes the space of N × N N \times N N × N positive definite matrices, admits a unique solution X t N ∈ S N + + X_t^N \in S_N^{++} X t N ∈ S N ++ q.s. for t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] . Moreover, we have X t N ∈ ( L G 2 ( Ω t ) ) d X_t^N \in(L_G^2(\Omega_t))^d X t N ∈ ( L G 2 ( Ω t ) ) d with d = N × N d = N \times N d = N × N for t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] .
Theorem 1.1 implies directly the following result for G G G -Whishart process.
Theorem 1.6. Let X N X^N X N be given by (12) , where x 0 ∈ S N + + x_0 \in S_N^{++} x 0 ∈ S N ++ and τ \tau τ be given by (6) . We further assume that the underlying G G G -Brownian motion satisfies Assumption 1 with d = N × N d = N \times N d = N × N . Then, τ = ∞ \tau= \infty τ = ∞ quasi-surely. Moreover, the eigenvalue processes, arranged in increasing order as λ 1 N ( t ) < λ 2 N ( t ) < ⋯ < λ N N ( t ) \lambda_1^N(t) < \lambda_2^N(t) < \cdots< \lambda_N^N(t) λ 1 N ( t ) < λ 2 N ( t ) < ⋯ < λ N N ( t ) , satisfy the following G G G -SDE: for i = 1 , 2 , … , N i = 1,2,\ldots,N i = 1 , 2 , … , N ,
d λ i N = 2 λ i N d B ‾ i + α d ⟨ B ‾ i ⟩ t + ∑ j : j ≠ i λ i N + λ j N λ i N − λ j N d ⟨ B ‾ i ⟩ t , (13) \mathrm{d}\lambda_i^N = 2\sqrt{\lambda_i^N}\,\mathrm{d}\overline{B}^i + \alpha\mathrm{d}\langle\overline{B}^i\rangle_t + \sum_{j:j\ne i}\frac{\lambda_i^N+\lambda_j^N}{\lambda_i^N-\lambda_j^N}\mathrm{d}\langle\overline{B}^i\rangle_t, \tag*{(13)} d λ i N = 2 λ i N d B i + α d ⟨ B i ⟩ t + j : j = i ∑ λ i N − λ j N λ i N + λ j N d ⟨ B i ⟩ t , (13) where B ‾ = ( B ‾ 1 , … , B ‾ N ) \overline{B} = (\overline{B}^1,\ldots,\overline{B}^N) B = ( B 1 , … , B N ) is an N N N -dimensional G ‾ \overline{G} G -Brownian motion with G ‾ \overline{G} G satisfying Assumption 1 with d = N d = N d = N .
Main results on high-dimensional spectral limits
Now we present our results on high-dimensional limits of the empirical eigenvalue distributions associated with generalized G G G -Wishart processes. Let C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathcal{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) denote the space of continuous functions mapping from [ 0 , T ] [0,T] [ 0 , T ] into P ( R ) \mathcal{P}(\mathbb{R}) P ( R ) , where P ( R ) \mathcal{P}(\mathbb{R}) P ( R ) is the space of probability measures on R \mathbb{R} R , equipped with the topology of weak convergence. We study the high-dimensional limit of the empirical spectral measure processes:
{ L N ( t ) : = 1 N ∑ i = 1 N δ λ i N ( t ) , t ∈ [ 0 , T ] } N ∈ N , (14) \left\{L_N(t) := \frac{1}{N}\sum_{i=1}^{N}\delta_{\lambda_i^N(t)},\ t \in[0,T]\right\}_{N\in\mathbb{N}}, \tag*{(14)} { L N ( t ) := N 1 i = 1 ∑ N δ λ i N ( t ) , t ∈ [ 0 , T ] } N ∈ N , (14)
which are viewed as C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathbb{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) -valued random elements. We first establish its tightness under the following assumption.
Assumption 2. We assume that the following conditions hold.
(1) There exists a positive function φ ( x ) ∈ C 2 ( R ) \varphi(x) \in C^{2}(\mathbb{R}) φ ( x ) ∈ C 2 ( R ) such that lim ∣ x ∣ → ∞ φ ( x ) = ∞ \lim_{|x|\to\infty}\varphi(x)=\infty lim ∣ x ∣ → ∞ φ ( x ) = ∞ , φ ′ ( x ) b N ( x ) \varphi'(x)b_N(x) φ ′ ( x ) b N ( x ) is bounded uniformly with respect to x ∈ R x\in\mathbb{R} x ∈ R and N ≥ 1 N\geq1 N ≥ 1 , and
∑ N = 1 ∞ ( ∥ φ ′ g N h N ∥ ∞ 2 N ) l 1 < ∞ \sum_{N=1}^{\infty}\left(\frac{\|\varphi'g_Nh_N\|_{\infty}^{2}}{N}\right)^{l_1}<\infty N = 1 ∑ ∞ ( N ∥ φ ′ g N h N ∥ ∞ 2 ) l 1 < ∞ for some positive integer l 1 l_1 l 1 . One typical choice is φ ( x ) = ln ( 1 + x 2 ) \varphi(x)=\ln(1+x^2) φ ( x ) = ln ( 1 + x 2 ) .
(2) The function N H N ( x , y ) φ ′ ( x ) − φ ′ ( y ) x − y NH_N(x,y)\frac{\varphi'(x)-\varphi'(y)}{x-y} N H N ( x , y ) x − y φ ′ ( x ) − φ ′ ( y ) is bounded uniformly with respect to x , y ∈ R x,y\in\mathbb{R} x , y ∈ R and N ≥ 1 N\geq1 N ≥ 1 .
(3) The initial empirical measure satisfies
C 0 : = sup N ≥ 1 ⟨ φ , L N ( 0 ) ⟩ = sup N ≥ 1 1 N ∑ i = 1 N φ ( λ i N ( 0 ) ) < ∞ . C_0:=\sup_{N\geq1}\langle\varphi,L_N(0)\rangle=\sup_{N\geq1}\frac{1}{N}\sum_{i=1}^{N}\varphi(\lambda_i^N(0))<\infty. C 0 := N ≥ 1 sup ⟨ φ , L N ( 0 )⟩ = N ≥ 1 sup N 1 i = 1 ∑ N φ ( λ i N ( 0 )) < ∞. (4) There exists a sequence { f ~ k } k ∈ N \{\widetilde{f}_k\}_{k\in\mathbb{N}} { f k } k ∈ N of C 2 ( R ) C^2(\mathbb{R}) C 2 ( R ) functions which is dense in the space C 0 ( R ) C_0(\mathbb{R}) C 0 ( R ) of continuous functions vanishing at infinity such that
ψ ( k ) : = ∑ N = 1 ∞ ( ∥ f ~ k ′ g N h N ∥ ∞ 2 N ) l 2 < ∞ , k ∈ N \psi(k):=\sum_{N=1}^{\infty}\left(\frac{\|\widetilde{f}'_kg_Nh_N\|_{\infty}^{2}}{N}\right)^{l_2}<\infty,\qquad k\in\mathbb{N} ψ ( k ) := N = 1 ∑ ∞ ( N ∥ f k ′ g N h N ∥ ∞ 2 ) l 2 < ∞ , k ∈ N for some integer $l_2\geq 2.
The following result characterizes the relative compactness of the sequence { L N ( t ) , t ∈ [ 0 , T ] } N ∈ N \{L_N(t),t\in[0,T]\}_{N\in\mathbb{N}} { L N ( t ) , t ∈ [ 0 , T ] } N ∈ N , and the proof is given in § 5.1 .
Theorem 1.7. Suppose that (7) has a solution that is nonexploding and noncolliding for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Then under Assumption 2, the sequence { L N ( t ) , t ∈ [ 0 , T ] } N ∈ N \{L_N(t),t\in[0,T]\}_{N\in\mathbb{N}} { L N ( t ) , t ∈ [ 0 , T ] } N ∈ N given by (14) is relatively compact in C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathbb{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) quasi-surely.
The following result characterizes the spectral limit, of which the proof is given in § 5.1 .
Theorem 1.8. Assume the same conditions as in Theorem 1.7 . Furthermore, assume that there exist continuous functions b ( x ) b(x) b ( x ) and H ( x , y ) H(x,y) H ( x , y ) , such that b N ( x ) b_N(x) b N ( x ) converges to b ( x ) b(x) b ( x ) and N H N ( x , y ) NH_N(x,y) N H N ( x , y ) converges to H ( x , y ) H(x,y) H ( x , y ) uniformly as N → ∞ N\to\infty N → ∞ , and that
∥ b ( x ) 1 + x 2 ∥ ∞ < ∞ , ∥ H ( x , y ) ( 1 + ∣ x ∣ ) ( 1 + y 2 ) ∥ ∞ < ∞ . \left\|\frac{b(x)}{1+x^2}\right\|_{\infty}<\infty,\qquad\left\|\frac{H(x,y)}{(1+|x|)(1+y^2)}\right\|_{\infty}<\infty. 1 + x 2 b ( x ) ∞ < ∞ , ( 1 + ∣ x ∣ ) ( 1 + y 2 ) H ( x , y ) ∞ < ∞. Suppose that L N ( 0 ) L_N(0) L N ( 0 ) converges weakly to μ 0 \mu_0 μ 0 , quasi-surely, as N → ∞ N\to\infty N → ∞ , and that { L N } \{L_N\} { L N } has a limit measure μ ∈ C ( [ 0 , T ] , P ( R ) ) \mu\in C([0,T],\mathbb{P}(\mathbb{R})) μ ∈ C ([ 0 , T ] , P ( R )) . Then, for every z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R ,
∫ μ t ( d x ) z − x = ∫ μ 0 ( d x ) z − x + ∫ 0 t [ ∫ b ( x ) ( z − x ) 2 μ s ( d x ) ] d ⟨ B ‾ 1 ⟩ s + ∫ 0 t [ ∬ H ( x , y ) ( z − x ) ( z − y ) 2 μ s ( d x ) μ s ( d y ) ] d ⟨ B ‾ 1 ⟩ s . (15) \begin{aligned}
\int\frac{\mu_t(\mathrm{d}x)}{z-x}
={}&\int\frac{\mu_0(\mathrm{d}x)}{z-x}
+\int_0^t\left[\int\frac{b(x)}{(z-x)^2}\mu_s(\mathrm{d}x)\right]\mathrm{d}\langle\overline{B}^{1}\rangle_s\\
&+\int_0^t\left[\iint\frac{H(x,y)}{(z-x)(z-y)^2}\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)\right]\mathrm{d}\langle\overline{B}^{1}\rangle_s. \tag*{(15)}
\end{aligned} ∫ z − x μ t ( d x ) = ∫ z − x μ 0 ( d x ) + ∫ 0 t [ ∫ ( z − x ) 2 b ( x ) μ s ( d x ) ] d ⟨ B 1 ⟩ s + ∫ 0 t [ ∬ ( z − x ) ( z − y ) 2 H ( x , y ) μ s ( d x ) μ s ( d y ) ] d ⟨ B 1 ⟩ s . (15)
sup σ ≤ b ≤ σ φ ( b 2 t ) \sup_{\sigma\le b\le\sigma}\varphi(b^{2}t) sup σ ≤ b ≤ σ φ ( b 2 t ) and the function v ( t , x ) = E ^ [ φ ( x + ⟨ B ‾ 1 ⟩ t ) ] v(t,x)=\widehat{\mathbb{E}}[\varphi(x+\langle\overline{B}^{1}\rangle_t)] v ( t , x ) = E [ φ ( x + ⟨ B 1 ⟩ t )] satisfies the nonlinear equation ∂ t v = 2 g ( ∂ x v ) \partial_t v=2g(\partial_x v) ∂ t v = 2 g ( ∂ x v ) with the initial condition v ( 0 , x ) = φ ( x ) v(0,x)=\varphi(x) v ( 0 , x ) = φ ( x ) .
In general, uniqueness of solutions to (15) , (1.16) and (1.17) has not been established. Nevertheless, for the special cases of G G G -Dyson Brownian motion and the G G G -Wishart process, we establish uniqueness of solutions to the corresponding limiting equations of (15) in Proposition 1.11 and Proposition 1.14 , respectively.
As a corollary of Theorems 1.7 and 1.8 , we have the following result for G G G -Dyson Brownian motion. Moreover, the uniqueness of the solution to the equation of the limiting measure can be obtained, the proof of which is given in § 5.2 .
Proposition 1.11. Assume that (9) has a solution λ N ( t ) \lambda^{N}(t) λ N ( t ) which is nonexploding and noncolliding for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] with initial condition λ N ( 0 ) \lambda^{N}(0) λ N ( 0 ) satisfying the corresponding condition (3) in Assumption 2. Assume that L N ( 0 ) L_N(0) L N ( 0 ) converges weakly as N N N goes to infinity to ν ∈ P ( R ) \nu\in\mathcal{P}(\mathbb{R}) ν ∈ P ( R ) . Then, { L N ( t ) , t ∈ [ 0 , T ] } \{L_N(t),t\in[0,T]\} { L N ( t ) , t ∈ [ 0 , T ]} converges quasi-surely in C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathcal{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) as N → ∞ N\to\infty N → ∞ . Its limit is the unique measure-valued process { μ t , t ∈ [ 0 , T ] } \{\mu_t,t\in[0,T]\} { μ t , t ∈ [ 0 , T ]} so that μ 0 = ν \mu_0=\nu μ 0 = ν and the function
V t ( z ) = ∫ ( z − x ) − 1 μ t ( d x ) (18) V_t(z)=\int(z-x)^{-1}\mu_t(\mathrm{d}x) \tag*{(18)} V t ( z ) = ∫ ( z − x ) − 1 μ t ( d x ) (18) satisfies the equation
V t ( z ) = V 0 ( z ) − ∫ 0 t V s ( z ) ∂ z V s ( z ) d ⟨ B ‾ 1 ⟩ s , (19) V_t(z)=V_0(z)-\int_0^t V_s(z)\partial_zV_s(z)\,\mathrm{d}\langle\overline{B}^{1}\rangle_s, \tag*{(19)} V t ( z ) = V 0 ( z ) − ∫ 0 t V s ( z ) ∂ z V s ( z ) d ⟨ B 1 ⟩ s , (19) for z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R .
and the density formulation (1.17) is
d ρ t ( x ) = − ∂ x [ ρ t ( x ) p.v. ∫ R ρ t ( y ) x − y d y ] d ⟨ B ‾ 1 ⟩ t . \mathrm{d}\rho_t(x)=-\partial_x\left[\rho_t(x)\operatorname{p.v.}\int_{\mathbb{R}}\frac{\rho_t(y)}{x-y}\,\mathrm{d}y\right]\mathrm{d}\langle\overline{B}^{1}\rangle_t. d ρ t ( x ) = − ∂ x [ ρ t ( x ) p.v. ∫ R x − y ρ t ( y ) d y ] d ⟨ B 1 ⟩ t .
Remark 1.13. For μ t \mu_t μ t satisfying (1.19), we have μ t = μ 0 ⊞ S ⟨ B ‾ 1 ⟩ t \mu_t=\mu_0\boxplus S_{\langle\overline{B}^{1}\rangle_t} μ t = μ 0 ⊞ S ⟨ B 1 ⟩ t , where ⊞ \boxplus ⊞ denotes free additive convolution, and
S a = 1 2 π a 4 a − x 2 1 { ∣ x ∣ ≤ 2 a } . S_a=\frac{1}{2\pi a}\sqrt{4a-x^2}\mathbf{1}_{\{|x|\leq2\sqrt{a}\}}. S a = 2 π a 1 4 a − x 2 1 { ∣ x ∣ ≤ 2 a } .
is the semicircle law with variance a a a . If μ 0 \mu_0 μ 0 is the point mass at 0 0 0 , then μ t \mu_t μ t coincides with the semicircle law with variance ⟨ B ‾ 1 ⟩ t \langle\overline{B}^{1}\rangle_t ⟨ B 1 ⟩ t , i.e., μ t = S ⟨ B ‾ 1 ⟩ t ( x ) \mu_t=S_{\langle\overline{B}^{1}\rangle_t}(x) μ t = S ⟨ B 1 ⟩ t ( x ) .
As a corollary of Theorems 1.7 and 1.8 , we have the following result for G G G -Wishart process, where the uniqueness of the solution to the equation of the limiting measure is also obtained. The proof is provided in § 5.3 .
Let X t N : = 1 N B t T B t X_t^N:=\frac{1}{N}B_t^{\mathsf{T}}B_t X t N := N 1 B t T B t , where B t B_t B t is a p × N p\times N p × N G G G -Brownian matrix satisfying Assumption 1. Then by (13) , it is easy to know that the eigenvalue processes λ N \lambda^N λ N of X N X^N X N satisfies the following system corresponding with (7) for g N ( x ) = x g_N(x)=\sqrt{x} g N ( x ) = x , h N ( x ) = 1 / N h_N(x)=1/\sqrt{N} h N ( x ) = 1/ N , and b N ( x ) = p N b_N(x)=\frac{p}{N} b N ( x ) = N p :
d λ i N = 2 N λ i N d B ‾ i + p N d ⟨ B ‾ i ⟩ t + 1 N ∑ j : j ≠ i λ i N + λ j N λ i N − λ j N d ⟨ B ‾ i ⟩ t . (24) \mathrm{d}\lambda_i^N=\frac{2}{\sqrt{N}}\sqrt{\lambda_i^N}\,\mathrm{d}\overline{B}_i+\frac{p}{N}\,\mathrm{d}\langle\overline{B}_i\rangle_t+\frac{1}{N}\sum_{j:j\ne i}\frac{\lambda_i^N+\lambda_j^N}{\lambda_i^N-\lambda_j^N}\,\mathrm{d}\langle\overline{B}_i\rangle_t. \tag*{(24)} d λ i N = N 2 λ i N d B i + N p d ⟨ B i ⟩ t + N 1 j : j = i ∑ λ i N − λ j N λ i N + λ j N d ⟨ B i ⟩ t . (24)
We further assume p ≥ N + 1 p\geq N+1 p ≥ N + 1 , which guarantees the existence and uniqueness of the solution to the G G G -SDE (12) with α = p \alpha=p α = p by Theorem 1.5. and
lim N → ∞ p N = c ≥ 1. (25) \lim_{N\to\infty}\frac{p}{N}=c\geq1. \tag*{(25)} N → ∞ lim N p = c ≥ 1. (25)
Then b N ( x ) b_N(x) b N ( x ) converges to b ( x ) = c b(x)=c b ( x ) = c and N H N ( x , y ) NH_N(x,y) N H N ( x , y ) converges to H ( x , y ) = x + y H(x,y)=x+y H ( x , y ) = x + y .
Proposition 1.14. Assume that (24) has a solution λ N ( t ) \lambda^N(t) λ N ( t ) which is nonexploding and noncolliding for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] with initial condition λ N ( 0 ) \lambda^N(0) λ N ( 0 ) satisfying the corresponding condition (3) in Assumption 2. Assume that the empirical measure L N ( 0 ) L_N(0) L N ( 0 ) converges weakly as N → ∞ N\to\infty N → ∞ to ν ∈ P ( R ) \nu\in\mathcal{P}(\mathbb{R}) ν ∈ P ( R ) . Then, { L N ( t ) , t ∈ [ 0 , T ] } \{L_N(t),t\in[0,T]\} { L N ( t ) , t ∈ [ 0 , T ]} converges quasi-surely in C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathcal{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) . Its limit is the unique measure-valued process { μ t , t ∈ [ 0 , T ] } \{\mu_t,t\in[0,T]\} { μ t , t ∈ [ 0 , T ]} so that μ 0 = ν \mu_0=\nu μ 0 = ν and the function
V t ( z ) = ∫ ( z − x ) − 1 μ t ( d x ) (26) V_t(z)=\int(z-x)^{-1}\mu_t(\mathrm{d}x) \tag*{(26)} V t ( z ) = ∫ ( z − x ) − 1 μ t ( d x ) (26) satisfies the equation
V t ( z ) = V 0 ( z ) − ∫ 0 t [ V s ( z ) 2 + ( c − 1 ) ∂ z V s ( z ) + 2 z V s ( z ) ∂ z V s ( z ) ] d ⟨ B ‾ 1 ⟩ s , (27) V_t(z)=V_0(z)-\int_0^t\left[V_s(z)^2+(c-1)\partial_zV_s(z)+2zV_s(z)\partial_zV_s(z)\right]\mathrm{d}\langle\overline{B}^{1}\rangle_s, \tag*{(27)} V t ( z ) = V 0 ( z ) − ∫ 0 t [ V s ( z ) 2 + ( c − 1 ) ∂ z V s ( z ) + 2 z V s ( z ) ∂ z V s ( z ) ] d ⟨ B 1 ⟩ s , (27) for z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , where c ≥ 1 c\geq1 c ≥ 1 is given in (25) .
Remark 1.15. For the G G G -Wishart process, the corresponding McKean–Vlasov equation of (27) is
⟨ μ t , f ⟩ = ⟨ μ 0 , f ⟩ + c ∫ 0 t ∫ R f ′ ( x ) μ s ( d x ) d ⟨ B ‾ 1 ⟩ s + 1 2 ∫ 0 t ∫ R 2 ( x + y ) f ′ ( x ) − f ′ ( y ) x − y μ s ( d x ) μ s ( d y ) d ⟨ B ‾ 1 ⟩ s . (28) \begin{aligned}
\langle\mu_t,f\rangle=\langle\mu_0,f\rangle+c\int_0^t\int_{\mathbb{R}}f'(x)\,\mu_s(\mathrm{d}x)\,\mathrm{d}\langle\overline{B}^{1}\rangle_s \\
&+\frac{1}{2}\int_0^t\int_{\mathbb{R}^2}(x+y)\frac{f'(x)-f'(y)}{x-y}\,\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)\,\mathrm{d}\langle\overline{B}^{1}\rangle_s. \tag*{(28)}
\end{aligned} ⟨ μ t , f ⟩ = ⟨ μ 0 , f ⟩ + c ∫ 0 t ∫ R f ′ ( x ) μ s ( d x ) d ⟨ B 1 ⟩ s + 2 1 ∫ 0 t ∫ R 2 ( x + y ) x − y f ′ ( x ) − f ′ ( y ) μ s ( d x ) μ s ( d y ) d ⟨ B 1 ⟩ s . (28)
Since the measures are supported on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , so the spatial integrals may be restricted accordingly. Assuming that μ t \mu_t μ t has a density ρ t ( x ) \rho_t(x) ρ t ( x ) , then the density formulation is
d ρ t ( x ) = − ∂ x [ ρ t ( x ) ( c − 1 + 2 x p.v. ∫ R ρ t ( y ) x − y d y ) ] d ⟨ B ‾ 1 ⟩ t . \mathrm{d}\rho_t(x)=-\partial_x\left[\rho_t(x)\left(c-1+2x\operatorname{p.v.}\int_{\mathbb{R}}\frac{\rho_t(y)}{x-y}\,\mathrm{d}y\right)\right]\mathrm{d}\langle\overline{B}^{1}\rangle_t. d ρ t ( x ) = − ∂ x [ ρ t ( x ) ( c − 1 + 2 x p.v. ∫ R x − y ρ t ( y ) d y ) ] d ⟨ B 1 ⟩ t .
Remark 1.16. If μ 0 \mu_0 μ 0 is the point mass at 0 0 0 , then we have
V t ( z ) = z + ( 1 − c ) ⟨ B ‾ 1 ⟩ t − ( z − ( c + 1 ) ⟨ B ‾ 1 ⟩ t ) 2 − 4 c ⟨ B ‾ 1 ⟩ t 2 2 z ⟨ B ‾ 1 ⟩ t V_t(z)=\frac{z+(1-c)\langle\overline{B}^{1}\rangle_t-\sqrt{\left(z-(c+1)\langle\overline{B}^{1}\rangle_t\right)^2-4c\langle\overline{B}^{1}\rangle_t^2}}{2z\langle\overline{B}^{1}\rangle_t} V t ( z ) = 2 z ⟨ B 1 ⟩ t z + ( 1 − c ) ⟨ B 1 ⟩ t − ( z − ( c + 1 ) ⟨ B 1 ⟩ t ) 2 − 4 c ⟨ B 1 ⟩ t 2
and consequently, μ t \mu_t μ t has the density
ρ t ( x ) = ( ⟨ B ‾ 1 ⟩ t ( 1 + c ) 2 − x ) ( x − ⟨ B ‾ 1 ⟩ t ( c − 1 ) 2 ) 2 π x ⟨ B ‾ 1 ⟩ t 1 [ ⟨ B ‾ 1 ⟩ t ( c − 1 ) 2 , ⟨ B ‾ 1 ⟩ t ( 1 + c ) 2 ] ( x ) , \rho_t(x)=\frac{\sqrt{\left(\langle\overline{B}^{1}\rangle_t(1+\sqrt{c})^2-x\right)\left(x-\langle\overline{B}^{1}\rangle_t(\sqrt{c}-1)^2\right)}}{2\pi x\langle\overline{B}^{1}\rangle_t}\mathbf{1}_{[\langle\overline{B}^{1}\rangle_t(\sqrt{c}-1)^2,\langle\overline{B}^{1}\rangle_t(1+\sqrt{c})^2]}(x), ρ t ( x ) = 2 π x ⟨ B 1 ⟩ t ( ⟨ B 1 ⟩ t ( 1 + c ) 2 − x ) ( x − ⟨ B 1 ⟩ t ( c − 1 ) 2 ) 1 [⟨ B 1 ⟩ t ( c − 1 ) 2 , ⟨ B 1 ⟩ t ( 1 + c ) 2 ] ( x ) ,
which is the Marchenko–Pastur law with mean c ⟨ B ‾ 1 ⟩ t c\langle\overline{B}^{1}\rangle_t c ⟨ B 1 ⟩ t .
Summary of our contributions
In this work, we investigate generalized Wishart processes defined by (3) within the framework of G G G -expectation, focusing on eigenvalue non-collision and high-dimensional spectral limits. Volatility uncertainty affects both the diffusion and the drift of the eigenvalue dynamics. Consequently, classical non-collision results need not carry over to this setting, and eigenvalue collisions may occur under certain covariance structures. Under these covariance structures, Lévy’s characterization and the GOE property may also fail; see § 3 for details.
We identify a sufficient condition for non-collision, formulated in Assumption 1, and study the empirical spectral measures of the generalized G G G -Wishart processes. We characterize their high-dimensional limits through an equation for the Stieltjes transform, or equivalently, a McKean–Vlasov equation for the limiting measure. In particular, we establish the semicircle law for G G G -Dyson Brownian motion and the Marchenko–Pastur law for the G G G -Wishart process.
In general, the limiting spectral distribution μ t \mu_t μ t of a generalized G G G -Wishart process is random rather than deterministic. It inherits the volatility uncertainty of the underlying G G G -expectation, encoded by the maximally distributed random variable ⟨ B ‾ 1 ⟩ t \langle\overline{B}^{1}\rangle_t ⟨ B 1 ⟩ t . For the two special cases above, the classical semicircle and Marchenko–Pastur laws are recovered when the volatility-uncertainty set reduces to a singleton, whereas genuine volatility uncertainty yields families of laws corresponding to the admissible models. These results describe how volatility uncertainty affects eigenvalue dynamics and extend classical spectral limit laws to the framework of nonlinear expectation.
The rest of the paper is organized as follows. In § 2 , we review the basic theory of nonlinear expectation and derive several corollaries related to the G G G -Lévy characterization theorem. § 3 examines alternative assumptions under which the GOE property and the non-collision property of G G G -Dyson Brownian motion may fail. In § 4 , we derive the spectral dynamics of generalized Wishart processes and establish the non-collision result under Assumption 1, as stated in Theorem 1.1 . We also investigate G G G -Dyson Brownian motion and the G G G -Wishart process as two important special cases. In § 5 , we establish the relative compactness of the empirical spectral measure processes and characterize their high-dimensional limits, as stated in Theorem 1.7 and Theorem 1.8 , respectively. Furthermore, for G G G -Dyson Brownian motion and the G G G -Wishart process, we prove uniqueness of solutions to the corresponding limiting equations, thereby obtaining the semicircle and Marchenko–Pastur laws under G G G -expectation, respectively.
Preliminaries on nonlinear expectations
In this section, we provide some preliminaries on nonlinear expectations, with particular emphasis on the theory of G G G -expectation and G G G -Brownian motion. We refer to [8 , 18 , 28 , 29 , 30 , 16 , 25 ] for further details.
Sublinear expectation and G G G -normal distribution
Let Ω \Omega Ω be a given set and H \mathcal{H} H be a linear space of real-valued functions on Ω \Omega Ω such that c ∈ H c \in\mathcal{H} c ∈ H for all constants c c c and ∣ X ∣ ∈ H |X| \in\mathcal{H} ∣ X ∣ ∈ H if X ∈ H X \in\mathcal{H} X ∈ H . We further suppose that if X 1 , ⋯ , X n ∈ H X_{1}, \cdots, X_{n} \in\mathcal{H} X 1 , ⋯ , X n ∈ H , then φ ( X 1 , ⋯ , X n ) ∈ H \varphi(X_{1}, \cdots, X_{n}) \in\mathcal{H} φ ( X 1 , ⋯ , X n ) ∈ H for each φ ∈ C b , Lip ( R n ) \varphi\in C_{b,\operatorname{Lip}}(\mathbb{R}^{n}) φ ∈ C b , Lip ( R n ) , where C b , Lip ( R n ) C_{b,\operatorname{Lip}}(\mathbb{R}^{n}) C b , Lip ( R n ) denotes the space of all bounded Lipschitz-continuous functions on R n \mathbb{R}^{n} R n .
Definition 2.1. A functional E : H → R \mathbb{E}: \mathcal{H} \to\mathbb{R} E : H → R is a sublinear expectation if the following properties are satisfied: for each X , Y ∈ H X,Y \in\mathcal{H} X , Y ∈ H ,
X ≥ Y X \ge Y X ≥ Y implies E [ X ] ≥ E [ Y ] \mathbb{E}[X] \ge\mathbb{E}[Y] E [ X ] ≥ E [ Y ] (monotonicity );
E [ c ] = c \mathbb{E}[c] = c E [ c ] = c for c ∈ R c \in\mathbb{R} c ∈ R (constant preserving );
E [ X + Y ] ≤ E [ X ] + E [ Y ] \mathbb{E}[X+Y] \le\mathbb{E}[X] + \mathbb{E}[Y] E [ X + Y ] ≤ E [ X ] + E [ Y ] (sub-additivity );
E [ λ X ] = λ E [ X ] \mathbb{E}[\lambda X] = \lambda\mathbb{E}[X] E [ λ X ] = λ E [ X ] for λ > 0 \lambda> 0 λ > 0 (positive homogeneity ).
The triple ( Ω , H , E ) (\Omega, \mathcal{H}, \mathbb{E}) ( Ω , H , E ) is called a sublinear expectation space .
When only (1) and (2) are satisfied, E \mathbb{E} E is called a nonlinear expectation and ( Ω , H , E ) (\Omega, \mathcal{H}, \mathbb{E}) ( Ω , H , E ) is therefore called a nonlinear expectation space. If equality holds in (3), then E \mathbb{E} E reduces to a classical linear expectation which is denoted by E E E .
For p ≥ 1 p \ge1 p ≥ 1 , suppose ∣ X ∣ p ∈ H |X|^{p} \in\mathcal{H} ∣ X ∣ p ∈ H for each X ∈ H X \in\mathcal{H} X ∈ H , and let L p ( Ω ) L^{p}(\Omega) L p ( Ω ) be the completion of H \mathcal{H} H under the norm ∥ X ∥ L p : = ( E [ ∣ X ∣ p ] ) 1 / p \|X\|_{L^{p}} := (\mathbb{E}[|X|^{p}])^{1/p} ∥ X ∥ L p := ( E [ ∣ X ∣ p ] ) 1/ p . Then, we have L p ( Ω ) ⊂ L p ′ ( Ω ) L^{p}(\Omega) \subset L^{p'}(\Omega) L p ( Ω ) ⊂ L p ′ ( Ω ) for p ≥ p ′ p \ge p' p ≥ p ′ . Clearly ( Ω , L 1 ( Ω ) , E ) (\Omega, L^{1}(\Omega), \mathbb{E}) ( Ω , L 1 ( Ω ) , E ) forms a sublinear expectation space, and we call it a complete sublinear expectation space.
Definition 2.2. Let X 1 X_{1} X 1 and X 2 X_{2} X 2 be two n n n -dimensional random vectors on complete sublinear expectation spaces ( Ω 1 , L 1 ( Ω 1 ) , E 1 ) (\Omega_{1}, L^{1}(\Omega_{1}), \mathbb{E}_{1}) ( Ω 1 , L 1 ( Ω 1 ) , E 1 ) and ( Ω 2 , L 1 ( Ω 2 ) , E 2 ) (\Omega_{2}, L^{1}(\Omega_{2}), \mathbb{E}_{2}) ( Ω 2 , L 1 ( Ω 2 ) , E 2 ) , respectively. They are called identically distributed, denoted by X 1 = d X 2 X_{1} \overset{d}{=} X_{2} X 1 = d X 2 , if E 1 [ φ ( X 1 ) ] = E 2 [ φ ( X 2 ) ] \mathbb{E}_{1}[\varphi(X_{1})] = \mathbb{E}_{2}[\varphi(X_{2})] E 1 [ φ ( X 1 )] = E 2 [ φ ( X 2 )] for each φ ∈ C b , Lip ( R n ) \varphi\in C_{b,\operatorname{Lip}}(\mathbb{R}^{n}) φ ∈ C b , Lip ( R n ) .
Definition 2.3. Let ( Ω , L 1 ( Ω ) , E ) (\Omega, L^{1}(\Omega), \mathbb{E}) ( Ω , L 1 ( Ω ) , E ) be a complete sublinear expectation space. An m m m -dimensional random vector Y Y Y is said to be independent from an n n n -dimensional random vector X X X if E [ φ ( X , Y ) ] = E [ E [ φ ( x , Y ) ] x = X ] \mathbb{E}[\varphi(X,Y)] = \mathbb{E}[\mathbb{E}[\varphi(x,Y)]_{x=X}] E [ φ ( X , Y )] = E [ E [ φ ( x , Y ) ] x = X ] , for each φ ∈ C b , Lip ( R n + m ) \varphi\in C_{b,\operatorname{Lip}}(\mathbb{R}^{n+m}) φ ∈ C b , Lip ( R n + m ) .
Denote by S d \mathbb{S}_{d} S d the space of all d × d d \times d d × d symmetric matrices, and let G : S d → R G:\mathbb{S}_{d} \to\mathbb{R} G : S d → R be a monotone sublinear function which satisfies, for A , A ′ ∈ S d A,A' \in\mathbb{S}_{d} A , A ′ ∈ S d ,
{ G ( A ) ≥ G ( A ′ ) if A ≥ A ′ , G ( A + A ′ ) ≤ G ( A ) + G ( A ′ ) , G ( λ A ) = λ G ( A ) for λ ≥ 0. \begin{cases}
G(A) \ge G(A') & \text{if } A \ge A',\\
G(A+A') \le G(A)+G(A'),\\
G(\lambda A) = \lambda G(A) & \text{for } \lambda\ge0.
\end{cases} ⎩ ⎨ ⎧ G ( A ) ≥ G ( A ′ ) G ( A + A ′ ) ≤ G ( A ) + G ( A ′ ) , G ( λ A ) = λ G ( A ) if A ≥ A ′ , for λ ≥ 0.
By the Hahn–Banach theorem, there exists a bounded, convex, and closed subset Γ ⊂ S d + \Gamma\subset\mathbb{S}_{d}^{+} Γ ⊂ S d + such that
G ( A ) = 1 2 sup γ ∈ Γ tr [ γ A ] , for A ∈ S d , G(A) = \frac{1}{2}\sup_{\gamma\in\Gamma}\operatorname{tr}[\gamma A], \quad\text{for } A \in\mathbb{S}_{d}, G ( A ) = 2 1 γ ∈ Γ sup tr [ γ A ] , for A ∈ S d ,
where S d + \mathbb{S}_{d}^{+} S d + denotes the space of all nonnegative definite matrices in S d \mathbb{S}_{d} S d .
The following G G G -normal distribution and maximal distribution arise naturally from the central limit theorem and the law of large numbers in the framework of sublinear expectations, respectively.
Definition 2.4 (G G G -normal distribution). Let ( Ω , L 1 ( Ω ) , E ) (\Omega, L^{1}(\Omega), \mathbb{E}) ( Ω , L 1 ( Ω ) , E ) be a complete sublinear expectation space and let G : S d → R G:\mathbb{S}_{d} \to\mathbb{R} G : S d → R be a monotone sublinear function. A d d d -dimensional random vector X X X is called G G G -normally distributed, if for each φ ∈ C b , Lip ( R d ) \varphi\in C_{b,\operatorname{Lip}}(\mathbb{R}^{d}) φ ∈ C b , Lip ( R d ) , E [ φ ( X ) ] = u φ ( 1 , 0 ) \mathbb{E}[\varphi(X)] = u^{\varphi}(1,0) E [ φ ( X )] = u φ ( 1 , 0 ) , where u φ u^{\varphi} u φ is the viscosity solution of the following G G G -heat equation:
{ ∂ t u ( t , x ) − G ( D x 2 u ( t , x ) ) = 0 , t > 0 , x ∈ R d , u ( 0 , x ) = φ ( x ) , (29) \begin{cases}
\partial_{t}u(t,x)-G(D_{x}^{2}u(t,x))=0, & t>0,\ x \in\mathbb{R}^{d},\\
u(0,x)=\varphi(x), \tag*{(29)}
\end{cases} { ∂ t u ( t , x ) − G ( D x 2 u ( t , x )) = 0 , u ( 0 , x ) = φ ( x ) , t > 0 , x ∈ R d , (29)
where D x 2 : = ( ∂ x i x j 2 ) d × d D_{x}^{2}:=(\partial_{x_{i}x_{j}}^{2})_{d \times d} D x 2 := ( ∂ x i x j 2 ) d × d denotes the Hessian matrix.
Definition 2.5 (maximal distribution). Let ( Ω , L 1 ( Ω ) , E ) (\Omega, L^{1}(\Omega), \mathbb{E}) ( Ω , L 1 ( Ω ) , E ) be a complete sublinear expectation space. A d d d -dimensional random vector η \eta η is called maximally distributed, if there exists a bounded, closed and convex subset Γ ⊂ R d \Gamma\subset\mathbb{R}^{d} Γ ⊂ R d such that
E [ φ ( η ) ] = max y ∈ Γ φ ( y ) , φ ∈ C b , L i p ( R d ) . \mathbb{E}[\varphi(\eta)] = \max_{y \in\Gamma} \varphi(y), \qquad\varphi\in C_{b,\mathrm{Lip}}(\mathbb{R}^{d}). E [ φ ( η )] = y ∈ Γ max φ ( y ) , φ ∈ C b , Lip ( R d ) .
G-Brownian motion and Lévy’s characterization
Definition 2.6 (G-Brownian motion). Let ( Ω , L 1 ( Ω ) , E ) (\Omega, L^{1}(\Omega), \mathbb{E}) ( Ω , L 1 ( Ω ) , E ) be a complete sublinear expectation space and let G : S d → R G : \mathbb{S}_{d} \to\mathbb{R} G : S d → R be a monotone sublinear function. A d d d -dimensional process ( B t ) t ≥ 0 (B_{t})_{t\geq0} ( B t ) t ≥ 0 with B t ∈ ( L 1 ( Ω ) ) d B_{t} \in(L^{1}(\Omega))^{d} B t ∈ ( L 1 ( Ω ) ) d for each t ≥ 0 t \geq0 t ≥ 0 is called a G-Brownian motion if the following properties are satisfied:
B 0 = 0 B_{0}=0 B 0 = 0 ;
for t , s ≥ 0 t,s \geq0 t , s ≥ 0 , B t + s − B s B_{t+s}-B_{s} B t + s − B s is independent from ( B t 1 , ⋯ , B t n ) (B_{t_{1}},\cdots,B_{t_{n}}) ( B t 1 , ⋯ , B t n ) for each n ∈ N n \in\mathbb{N} n ∈ N and 0 ≤ t 1 ≤ ⋯ ≤ t n ≤ t 0 \leq t_{1} \leq\cdots\leq t_{n} \leq t 0 ≤ t 1 ≤ ⋯ ≤ t n ≤ t ;
for t , s ≥ 0 t,s \geq0 t , s ≥ 0 , B t + s − B t = s ξ B_{t+s}-B_{t} = \sqrt{s}\xi B t + s − B t = s ξ , where ξ \xi ξ is G-normally distributed.
Moreover, if E [ B t ] = E [ − B t ] = 0 \mathbb{E}[B_{t}] = \mathbb{E}[-B_{t}] = 0 E [ B t ] = E [ − B t ] = 0 , then we call ( B t ) t ≥ 0 (B_{t})_{t\geq0} ( B t ) t ≥ 0 a symmetric G-Brownian motion.
Theorem 2.7. Let ( B t ) t ≥ 0 (B_{t})_{t\geq0} ( B t ) t ≥ 0 be d d d -dimensional symmetric G-Brownian motion. Then, for each fixed φ ∈ C b , L i p ( R d ) \varphi\in C_{b,\mathrm{Lip}}(\mathbb{R}^{d}) φ ∈ C b , Lip ( R d ) , the function
u ( t , x ) : = E [ φ ( x + B t ) ] , ( t , x ) ∈ [ 0 , ∞ ) × R d u(t,x) := \mathbb{E}[\varphi(x+B_{t})], \qquad(t,x) \in[0,\infty) \times\mathbb{R}^{d} u ( t , x ) := E [ φ ( x + B t )] , ( t , x ) ∈ [ 0 , ∞ ) × R d is the viscosity solution to the G-heat equation (29) , with G G G given by
G ( A ) = 1 2 E [ ⟨ A B 1 , B 1 ⟩ ] , A ∈ S ( d ) . G(A) = \frac{1}{2}\mathbb{E}[\langle AB_{1},B_{1}\rangle], \qquad A \in\mathbb{S}(d). G ( A ) = 2 1 E [⟨ A B 1 , B 1 ⟩] , A ∈ S ( d ) .
Proposition 2.8. Let B B B be a d d d -dimensional G-Brownian motion. For any deterministic n × d n \times d n × d -dimensional matrix Q Q Q , define X t = Q B t X_{t}=QB_{t} X t = Q B t . Then X t X_{t} X t is a G Q G_{Q} G Q -Brownian motion with the generator given by G Q ( A ) = G ( Q T A Q ) G_{Q}(A)=G(Q^{\mathsf{T}}AQ) G Q ( A ) = G ( Q T A Q ) for A ∈ S n A \in\mathbb{S}_{n} A ∈ S n .
Proof. Let
G Q ( A ) : = 1 2 E [ ⟨ A Q B 1 , Q B 1 ⟩ ] , A ∈ S n . G_{Q}(A) := \frac{1}{2}\mathbb{E}[\langle AQB_{1},QB_{1}\rangle], \qquad A \in\mathbb{S}_{n}. G Q ( A ) := 2 1 E [⟨ A Q B 1 , Q B 1 ⟩] , A ∈ S n .
It is easily seen that X t X_{t} X t is a G Q G_{Q} G Q -Brownian motion, and we can verify that
G Q ( A ) = 1 2 E [ ⟨ Q T A Q B 1 , B 1 ⟩ ] = G ( Q T A Q ) . G_{Q}(A) = \frac{1}{2}\mathbb{E}[\langle Q^{\mathsf{T}}AQB_{1},B_{1}\rangle] = G(Q^{\mathsf{T}}AQ). G Q ( A ) = 2 1 E [⟨ Q T A Q B 1 , B 1 ⟩] = G ( Q T A Q ) .
For each t ≥ 0 t \geq0 t ≥ 0 , set
Lip ( Ω t ) : = { φ ( B t 1 , B t 2 , ⋯ , B t n ) : n ∈ N , 0 ≤ t 1 < ⋯ < t n ≤ t , φ ∈ C b , L i p ( R n × d ) } ; \operatorname{Lip}(\Omega_{t}) := \left\{\varphi(B_{t_{1}},B_{t_{2}},\cdots,B_{t_{n}}) : n \in\mathbb{N}, 0 \leq t_{1} < \cdots< t_{n} \leq t, \varphi\in C_{b,\mathrm{Lip}}(\mathbb{R}^{n\times d})\right\}; Lip ( Ω t ) := { φ ( B t 1 , B t 2 , ⋯ , B t n ) : n ∈ N , 0 ≤ t 1 < ⋯ < t n ≤ t , φ ∈ C b , Lip ( R n × d ) } ;
and
Lip ( Ω ) : = ⋃ m = 1 ∞ Lip ( Ω m ) . \operatorname{Lip}(\Omega) := \bigcup_{m=1}^{\infty}\operatorname{Lip}(\Omega_{m}). Lip ( Ω ) := m = 1 ⋃ ∞ Lip ( Ω m ) .
For each p ≥ 1 p \geq1 p ≥ 1 , we denote by L G p ( Ω t ) L^{p}_{G}(\Omega_{t}) L G p ( Ω t ) (resp., L G p ( Ω ) L^{p}_{G}(\Omega) L G p ( Ω ) ) the completion of Lip ( Ω t ) \operatorname{Lip}(\Omega_{t}) Lip ( Ω t ) (resp., Lip ( Ω ) \operatorname{Lip}(\Omega) Lip ( Ω ) ) under the norm ∥ X ∥ L G p : = ( E [ ∣ X ∣ p ] ) 1 p \lVert X\rVert_{L^{p}_{G}} := \left(\mathbb{E}[|X|^{p}]\right)^{\frac{1}{p}} ∥ X ∥ L G p := ( E [ ∣ X ∣ p ] ) p 1 for each t ≥ 0 t \geq0 t ≥ 0 . The sublinear expectation E : L G 1 ( Ω ) → R \mathbb{E}:L^{1}_{G}(\Omega)\to\mathbb{R} E : L G 1 ( Ω ) → R is now called a G-expectation.
For φ ( B t 1 , B t 2 − B t 1 , ⋯ , B t n − B t n − 1 ) ∈ Lip ( Ω ) \varphi(B_{t_1}, B_{t_2}-B_{t_1},\cdots,B_{t_n}-B_{t_{n-1}})\in\operatorname{Lip}(\Omega) φ ( B t 1 , B t 2 − B t 1 , ⋯ , B t n − B t n − 1 ) ∈ Lip ( Ω ) , its conditional expectation at time t i t_i t i is defined by
E ^ t i [ φ ( B t 1 , B t 2 − B t 1 , ⋯ , B t n − B t n − 1 ) ] : = ψ ( B t 1 , ⋯ , B t i − B t i − 1 ) , \widehat{\mathbb{E}}_{t_i}\left[\varphi(B_{t_1},B_{t_2}-B_{t_1},\cdots,B_{t_n}-B_{t_{n-1}})\right]:=\psi(B_{t_1},\cdots,B_{t_i}-B_{t_{i-1}}), E t i [ φ ( B t 1 , B t 2 − B t 1 , ⋯ , B t n − B t n − 1 ) ] := ψ ( B t 1 , ⋯ , B t i − B t i − 1 ) ,
for i = 1 , … , n i=1,\ldots,n i = 1 , … , n , where
ψ ( x 1 , ⋯ , x i ) : = E ^ [ φ ( x 1 , ⋯ , x i , B t i + 1 − B t i , ⋯ , B t n − B t n − 1 ) ] . \psi(x_1,\cdots,x_i):=\widehat{\mathbb{E}}\left[\varphi(x_1,\cdots,x_i,B_{t_{i+1}}-B_{t_i},\cdots,B_{t_n}-B_{t_{n-1}})\right]. ψ ( x 1 , ⋯ , x i ) := E [ φ ( x 1 , ⋯ , x i , B t i + 1 − B t i , ⋯ , B t n − B t n − 1 ) ] .
Then, the mapping E ^ t i : Lip ( Ω ) → Lip ( Ω t i ) \widehat{\mathbb{E}}_{t_i}:\operatorname{Lip}(\Omega)\to\operatorname{Lip}(\Omega_{t_i}) E t i : Lip ( Ω ) → Lip ( Ω t i ) satisfies conditions (1)–(4) in Definition 2.1 and can be continuously extended to the mapping which is still denoted by E ^ t i : L G 1 ( Ω ) → L G 1 ( Ω t i ) \widehat{\mathbb{E}}_{t_i}:L_G^1(\Omega)\to L_G^1(\Omega_{t_i}) E t i : L G 1 ( Ω ) → L G 1 ( Ω t i ) . We call ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ≥ 0 , ( E ^ t ) t ≥ 0 ) (\Omega,L_G^1(\Omega),(L_G^1(\Omega_t))_{t\geq0},(\widehat{\mathbb{E}}_t)_{t\geq0}) ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ≥ 0 , ( E t ) t ≥ 0 ) a G G G -expectation space.
Definition 2.9. For p ≥ 1 p\geq1 p ≥ 1 , let M G p , 0 ( 0 , T ) M_G^{p,0}(0,T) M G p , 0 ( 0 , T ) denote the space of processes of the following form:
η t ( ω ) = ∑ i = 0 n − 1 ξ i ( ω ) I [ t i , t i + 1 ) ( t ) , \eta_t(\omega)=\sum_{i=0}^{n-1}\xi_i(\omega)I_{[t_i,t_{i+1})}(t), η t ( ω ) = i = 0 ∑ n − 1 ξ i ( ω ) I [ t i , t i + 1 ) ( t ) , where { t 0 , t 1 , … , t n } \{t_0,t_1,\ldots,t_n\} { t 0 , t 1 , … , t n } is a partition of [ 0 , T ] [0,T] [ 0 , T ] and ξ i ∈ L G p ( Ω t i ) \xi_i\in L_G^p(\Omega_{t_i}) ξ i ∈ L G p ( Ω t i ) for i ∈ { 0 , 1 , … , n − 1 } i\in\{0,1,\ldots,n-1\} i ∈ { 0 , 1 , … , n − 1 } . We denote by M G p ( 0 , T ) M_G^p(0,T) M G p ( 0 , T ) the completion of M G p , 0 ( 0 , T ) M_G^{p,0}(0,T) M G p , 0 ( 0 , T ) under the norm
∥ η ∥ M G p : = ( E ^ [ ∫ 0 T ∣ η t ∣ p d t ] ) 1 p . \|\eta\|_{M_G^p}:=\left(\widehat{\mathbb{E}}\left[\int_0^T|\eta_t|^p\,dt\right]\right)^{\frac{1}{p}}. ∥ η ∥ M G p := ( E [ ∫ 0 T ∣ η t ∣ p d t ] ) p 1 .
Definition 2.10. A d d d -dimensional process M t = ( M t 1 , ⋯ , M t d ) T M_t=(M_t^1,\cdots,M_t^d)^{\mathsf{T}} M t = ( M t 1 , ⋯ , M t d ) T is called a martingale if for any t ∈ [ 0 , ∞ ) t\in[0,\infty) t ∈ [ 0 , ∞ ) , M t i ∈ L G 1 ( Ω t ) M_t^i\in L_G^1(\Omega_t) M t i ∈ L G 1 ( Ω t ) and E ^ s [ M t i ] = M s i \widehat{\mathbb{E}}_s[M_t^i]=M_s^i E s [ M t i ] = M s i for all s ≤ t s\leq t s ≤ t and i ∈ { 1 , ⋯ , d } i\in\{1,\cdots,d\} i ∈ { 1 , ⋯ , d } . Furthermore, a d d d -dimensional martingale M M M is called symmetric if E ^ [ M t i ] = − E ^ [ − M t i ] \widehat{\mathbb{E}}[M_t^i]=-\widehat{\mathbb{E}}[-M_t^i] E [ M t i ] = − E [ − M t i ] for all t t t and i i i .
Clearly, the G G G -Brownian motion B B B is a symmetric martingale. For a symmetric martingale M M M satisfying
M 0 = 0 , M t ∈ L G 2 ( Ω t ) and E ^ t [ ∣ M t + s − M t ∣ 2 ] ≤ C M s , (30) M_0=0,\quad M_t\in L_G^2(\Omega_t)\quad\text{and}\quad\widehat{\mathbb{E}}_t[|M_{t+s}-M_t|^2]\leq C_Ms, \tag*{(30)} M 0 = 0 , M t ∈ L G 2 ( Ω t ) and E t [ ∣ M t + s − M t ∣ 2 ] ≤ C M s , (30)
for some constant C M C_M C M for all $s,t\geq 0, we can define stochastic integrals with respect to M M M as follows.
For η t ( ω ) = ∑ i = 0 n − 1 ξ i ( ω ) I [ t i , t i + 1 ) ( t ) ∈ M G 2 , 0 ( 0 , T ) \eta_t(\omega)=\sum_{i=0}^{n-1}\xi_i(\omega)I_{[t_i,t_{i+1})}(t)\in M_G^{2,0}(0,T) η t ( ω ) = ∑ i = 0 n − 1 ξ i ( ω ) I [ t i , t i + 1 ) ( t ) ∈ M G 2 , 0 ( 0 , T ) , define
I ( η ) = ∫ 0 T η s d M s : = ∑ i = 0 n − 1 ξ i ( M t i + 1 − M t i ) . (31) I(\eta)=\int_0^T\eta_s\,dM_s:=\sum_{i=0}^{n-1}\xi_i(M_{t_{i+1}}-M_{t_i}). \tag*{(31)} I ( η ) = ∫ 0 T η s d M s := i = 0 ∑ n − 1 ξ i ( M t i + 1 − M t i ) . (31)
Then the mapping I : M G 2 , 0 ( 0 , T ) → L G 2 ( Ω T ) I:M_G^{2,0}(0,T)\to L_G^2(\Omega_T) I : M G 2 , 0 ( 0 , T ) → L G 2 ( Ω T ) can be continuously extended to the mapping from M G 2 ( 0 , T ) M_G^2(0,T) M G 2 ( 0 , T ) to L G 2 ( Ω T ) L_G^2(\Omega_T) L G 2 ( Ω T ) . Moreover, for each η ∈ M G 2 ( 0 , T ) \eta\in M_G^2(0,T) η ∈ M G 2 ( 0 , T ) , the stochastic integral ∫ 0 t η s d M s \int_0^t\eta_s\,dM_s ∫ 0 t η s d M s is a symmetric martingale and
E ^ [ ∣ ∫ 0 T η s d M s ∣ 2 ] ≤ C M E ^ [ ∫ 0 T ∣ η s ∣ 2 d s ] . \widehat{\mathbb{E}}\left[\left|\int_0^T\eta_s\,dM_s\right|^2\right]\leq C_M\widehat{\mathbb{E}}\left[\int_0^T|\eta_s|^2\,ds\right]. E ∫ 0 T η s d M s 2 ≤ C M E [ ∫ 0 T ∣ η s ∣ 2 d s ] .
Similar to the classical case, we have the following G G G -Itô isometry and and G G G -Burkholder-Davis-Gundy inequality for integrals against G G G -Brownian motion.
Theorem 2.11 (G G G -Itô isometry). Let η ∈ M G 2 ( 0 , T ) \eta\in M_G^2(0,T) η ∈ M G 2 ( 0 , T ) . Then, we have
E ^ [ ( ∫ 0 T η t d B t ) 2 ] = E ^ [ ∫ 0 T ∣ η t ∣ 2 d ⟨ B ⟩ t ] . \widehat{\mathbb{E}}\left[\left(\int_0^T\eta_t\,dB_t\right)^2\right]=\widehat{\mathbb{E}}\left[\int_0^T|\eta_t|^2\,d\langle B\rangle_t\right]. E ( ∫ 0 T η t d B t ) 2 = E [ ∫ 0 T ∣ η t ∣ 2 d ⟨ B ⟩ t ] .
Theorem 2.12 (G-Burkholder-Davis-Gundy inequality). For any η ∈ M G 2 ( 0 , T ) \eta\in M_G^2(0,T) η ∈ M G 2 ( 0 , T ) and p > 0 p>0 p > 0 , there exist constants c p c_p c p and C p C_p C p with 0 < c p < C p < ∞ 0<c_p<C_p<\infty 0 < c p < C p < ∞ , depending only on p p p , such that
c p E [ ( ∫ 0 T ∣ η s ∣ 2 d ⟨ B ⟩ s ) p / 2 ] ≤ E [ sup 0 ≤ t ≤ T ∣ ∫ 0 t η s d B s ∣ p ] ≤ C p E [ ( ∫ 0 T ∣ η s ∣ 2 d ⟨ B ⟩ s ) p / 2 ] . c_p\mathbb{E}\left[\left(\int_0^T|\eta_s|^2\mathrm{d}\langle B\rangle_s\right)^{p/2}\right]\leq\mathbb{E}\left[\sup_{0\leq t\leq T}\left|\int_0^t\eta_s\mathrm{d}B_s\right|^p\right]\leq C_p\mathbb{E}\left[\left(\int_0^T|\eta_s|^2\mathrm{d}\langle B\rangle_s\right)^{p/2}\right]. c p E ( ∫ 0 T ∣ η s ∣ 2 d ⟨ B ⟩ s ) p /2 ≤ E [ 0 ≤ t ≤ T sup ∫ 0 t η s d B s p ] ≤ C p E ( ∫ 0 T ∣ η s ∣ 2 d ⟨ B ⟩ s ) p /2 .
For a d d d -dimensional martingale M t = ( M t 1 , ⋯ , M t d ) T M_t=(M_t^1,\cdots,M_t^d)^{\mathsf{T}} M t = ( M t 1 , ⋯ , M t d ) T , we denote by ⟨ M ⟩ t = ( ⟨ M i , M j ⟩ t ) d × d \langle M\rangle_t=(\langle M^i,M^j\rangle_t)_{d\times d} ⟨ M ⟩ t = (⟨ M i , M j ⟩ t ) d × d its quadratic covariation matrix, where
⟨ M i , M j ⟩ t : = lim ∣ π n ∣ → 0 ∑ k = 0 n − 1 ( M t k + 1 i − M t k i ) ( M t k + 1 j − M t k j ) , \langle M^i,M^j\rangle_t:=\lim_{|\pi_n|\to0}\sum_{k=0}^{n-1}(M_{t_{k+1}}^i-M_{t_k}^i)(M_{t_{k+1}}^j-M_{t_k}^j), ⟨ M i , M j ⟩ t := ∣ π n ∣ → 0 lim k = 0 ∑ n − 1 ( M t k + 1 i − M t k i ) ( M t k + 1 j − M t k j ) ,
with ∣ π n ∣ : = max { ∣ t k + 1 − t k ∣ : k = 0 , 1 , … , n − 1 } |\pi_n|:=\max\{|t_{k+1}-t_k|:k=0,1,\ldots,n-1\} ∣ π n ∣ := max { ∣ t k + 1 − t k ∣ : k = 0 , 1 , … , n − 1 } , is the quadratic covariation between M i M^i M i and M j M^j M j . When i = j i=j i = j , we write ⟨ M i ⟩ t : = ⟨ M i , M i ⟩ t \langle M^i\rangle_t:=\langle M^i,M^i\rangle_t ⟨ M i ⟩ t := ⟨ M i , M i ⟩ t for the corresponding quadratic variation. We remark that ∫ 0 t η s d ⟨ M ⟩ s \int_0^t\eta_s\mathrm{d}\langle M\rangle_s ∫ 0 t η s d ⟨ M ⟩ s is well-defined for η ∈ M G 1 ( 0 , T ) \eta\in M_G^1(0,T) η ∈ M G 1 ( 0 , T ) if the condition (30) is satisfied.
Below is the Lévy’s characterization of G G G -Brownian motion established in [18 ] .
Theorem 2.13. On a G G G -expectation space ( Ω , L G 1 ( Ω ) , ( L G ( Ω t ) ) t ≥ 0 , ( E t ) t ≥ 0 ) (\Omega,L_G^1(\Omega),(L_G(\Omega_t))_{t\geq0},(\mathbb{E}_t)_{t\geq0}) ( Ω , L G 1 ( Ω ) , ( L G ( Ω t ) ) t ≥ 0 , ( E t ) t ≥ 0 ) , let ( M t ) t ≥ 0 (M_t)_{t\geq0} ( M t ) t ≥ 0 be a d d d -dimensional symmetric martingale satisfying M 0 = 0 M_0=0 M 0 = 0 and M t ∈ ( L G 2 ( Ω t ) ) d M_t\in(L_G^2(\Omega_t))^d M t ∈ ( L G 2 ( Ω t ) ) d for each t ≥ 0 t\geq0 t ≥ 0 . Given a monotone sublinear function G ˉ : S d → R \bar{G}:S_d\to\mathbb{R} G ˉ : S d → R , the following conditions are equivalent:
( M t ) t ≥ 0 (M_t)_{t\geq0} ( M t ) t ≥ 0 is a d d d -dimensional G ˉ \bar{G} G ˉ -Brownian motion;
E t [ ∣ M t + s − M t ∣ 2 ] ≤ C s \mathbb{E}_t[|M_{t+s}-M_t|^2]\leq Cs E t [ ∣ M t + s − M t ∣ 2 ] ≤ C s for all s , t ≥ 0 s,t\geq0 s , t ≥ 0 for some C > 0 C>0 C > 0 , and the process 1 2 tr [ A ⟨ M ⟩ t ] − G ˉ ( A ) t \frac{1}{2}\operatorname{tr}[A\langle M\rangle_t]-\bar{G}(A)t 2 1 tr [ A ⟨ M ⟩ t ] − G ˉ ( A ) t is a martingale for each A ∈ S ( d ) A\in S(d) A ∈ S ( d ) .
The following result provides a criterion for a symmetric martingale to be a G G G -Brownian motion satisfying Assumption 1.
Corollary 2.14. Let ( M t ) t ≥ 0 (M_t)_{t\geq0} ( M t ) t ≥ 0 be a d d d -dimensional symmetric martingale as in Theorem 2.13 . Let ( B t ) t ≥ 0 (B_t)_{t\geq0} ( B t ) t ≥ 0 be a 1-dimensional g g g -Brownian motion, where g g g is given by (1) . If ⟨ M i , M j ⟩ t = 1 { i = j } ⟨ B ⟩ t \langle M^i,M^j\rangle_t=\mathbf{1}_{\{i=j\}}\langle B\rangle_t ⟨ M i , M j ⟩ t = 1 { i = j } ⟨ B ⟩ t for all i , j ∈ { 1 , … , d } i,j\in\{1,\ldots,d\} i , j ∈ { 1 , … , d } and t ≥ 0 t\geq0 t ≥ 0 , then ( M t ) t ≥ 0 (M_t)_{t\geq0} ( M t ) t ≥ 0 is a d d d -dimensional G ˉ \bar{G} G ˉ -Brownian motion, with G ˉ \bar{G} G ˉ satisfying Assumption 1.
Proof. Noting that for s , t > 0 s,t>0 s , t > 0 we have
E t [ ∣ M ( t + s ) − M ( t ) ∣ 2 ] = E t [ ∑ i = 1 d ∣ M i ( t + s ) − M i ( t ) ∣ 2 ] ≤ ∑ i = 1 d E t [ ⟨ B ⟩ t + s − ⟨ B ⟩ t ] ≤ d σ ˉ 2 s , \mathbb{E}_t[|M(t+s)-M(t)|^2]=\mathbb{E}_t\left[\sum_{i=1}^d|M_i(t+s)-M_i(t)|^2\right]\leq\sum_{i=1}^d\mathbb{E}_t[\langle B\rangle_{t+s}-\langle B\rangle_t]\leq d\bar{\sigma}^2s, E t [ ∣ M ( t + s ) − M ( t ) ∣ 2 ] = E t [ i = 1 ∑ d ∣ M i ( t + s ) − M i ( t ) ∣ 2 ] ≤ i = 1 ∑ d E t [⟨ B ⟩ t + s − ⟨ B ⟩ t ] ≤ d σ ˉ 2 s ,
by Theorem 2.13 , it suffices to show that 1 2 tr [ A ⟨ M ⟩ t ] − G ˉ ( A ) t \frac{1}{2}\operatorname{tr}[A\langle M\rangle_t]-\bar{G}(A)t 2 1 tr [ A ⟨ M ⟩ t ] − G ˉ ( A ) t is a martingale for each A ∈ S d A\in S_d A ∈ S d .
Indeed, for s < t s<t s < t , we have
E s [ 1 2 tr [ A ⟨ M ⟩ t ] − G ˉ ( A ) t ] = E s [ 1 2 ∑ i = 1 d a i i ⟨ M i ⟩ t − G ˉ ( A ) t ] = E s [ 1 2 ∑ i = 1 d a i i ( ⟨ B ⟩ t − ⟨ B ⟩ s ) ] + 1 2 ( ∑ i = 1 d a i i ⟨ M i ⟩ s ) − G ˉ ( A ) t = 1 2 ( σ ˉ 2 ( ∑ i = 1 d a i i ) + − σ ‾ 2 ( ∑ i = 1 d a i i ) − ) ( t − s ) + 1 2 ( ∑ i = 1 d a i i ⟨ M i ⟩ s ) − G ˉ ( A ) t = 1 2 ( ∑ i = 1 d a i i ⟨ M i ⟩ s ) − G ( A ) s = 1 2 tr [ A ⟨ M ⟩ s ] − G ˉ ( A ) s . \begin{aligned}
\mathbb{E}_s\left[\frac{1}{2}\operatorname{tr}[A\langle M\rangle_t]-\bar{G}(A)t\right]=\mathbb{E}_s\left[\frac{1}{2}\sum_{i=1}^d a_{ii}\langle M_i\rangle_t-\bar{G}(A)t\right] \\
&=\mathbb{E}_s\left[\frac{1}{2}\sum_{i=1}^d a_{ii}\big(\langle B\rangle_t-\langle B\rangle_s\big)\right]+\frac{1}{2}\left(\sum_{i=1}^d a_{ii}\langle M_i\rangle_s\right)-\bar{G}(A)t \\
&=\frac{1}{2}\left(\bar{\sigma}^2\left(\sum_{i=1}^d a_{ii}\right)^+-\underline{\sigma}^2\left(\sum_{i=1}^d a_{ii}\right)^-\right)(t-s)+\frac{1}{2}\left(\sum_{i=1}^d a_{ii}\langle M_i\rangle_s\right)-\bar{G}(A)t \\
&=\frac{1}{2}\left(\sum_{i=1}^d a_{ii}\langle M_i\rangle_s\right)-G(A)s=\frac{1}{2}\operatorname{tr}[A\langle M\rangle_s]-\bar{G}(A)s.
\end{aligned} E s [ 2 1 tr [ A ⟨ M ⟩ t ] − G ˉ ( A ) t ] = E s [ 2 1 i = 1 ∑ d a ii ⟨ M i ⟩ t − G ˉ ( A ) t ] = E s [ 2 1 i = 1 ∑ d a ii ( ⟨ B ⟩ t − ⟨ B ⟩ s ) ] + 2 1 ( i = 1 ∑ d a ii ⟨ M i ⟩ s ) − G ˉ ( A ) t = 2 1 σ ˉ 2 ( i = 1 ∑ d a ii ) + − σ 2 ( i = 1 ∑ d a ii ) − ( t − s ) + 2 1 ( i = 1 ∑ d a ii ⟨ M i ⟩ s ) − G ˉ ( A ) t = 2 1 ( i = 1 ∑ d a ii ⟨ M i ⟩ s ) − G ( A ) s = 2 1 tr [ A ⟨ M ⟩ s ] − G ˉ ( A ) s .
This completes the proof. □
Similar to Assumption 1, it is also natural to impose the following diagonal structure assumption on the generator G G G .
Assumption 1′. Assume G ( A ) = ∑ i = 1 d g ( A i i ) G(A)=\sum_{i=1}^{d}g(A_{ii}) G ( A ) = ∑ i = 1 d g ( A ii ) for A = ( A i j ) 1 ≤ i , j ≤ d ∈ S d A=(A_{ij})_{1\leq i,j\leq d}\in\mathbb{S}_{d} A = ( A ij ) 1 ≤ i , j ≤ d ∈ S d , where g g g is given by (1) .
Corollary 2.15. Let ( B t ) t ≥ 0 (B_t)_{t\geq0} ( B t ) t ≥ 0 be a d d d -dimensional G G G -Brownian motion satisfying Assumption 1′, and α = ( α 1 , α 2 , … , α d ) ∈ M G 2 ( 0 , T ) d \alpha=(\alpha^{1},\alpha^{2},\ldots,\alpha^{d})\in M_{G}^{2}(0,T)^{d} α = ( α 1 , α 2 , … , α d ) ∈ M G 2 ( 0 , T ) d be a d d d -dimensional process satisfying ∥ α t ∥ 2 = ( α t 1 ) 2 + ( α t 2 ) 2 + ⋯ + ( α t d ) 2 = 1 \|\alpha_t\|^{2}=(\alpha_t^{1})^{2}+(\alpha_t^{2})^{2}+\cdots+(\alpha_t^{d})^{2}=1 ∥ α t ∥ 2 = ( α t 1 ) 2 + ( α t 2 ) 2 + ⋯ + ( α t d ) 2 = 1 for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Then, the process M t : = ∫ 0 t α r d B r M_t:=\int_{0}^{t}\alpha_r\mathrm{d}B_r M t := ∫ 0 t α r d B r is a one-dimensional g g g -Brownian motion.
Proof. Obviously, M M M is a symmetric martingale, noting the boundedness of the integrand α \alpha α . To show that M M M is a g g g -Brownian motion, by Theorem 2.13, it suffices to verify that both 1 2 ⟨ M ⟩ t − 1 2 σ ‾ 2 t \frac{1}{2}\langle M\rangle_t-\frac{1}{2}\overline{\sigma}^{2}t 2 1 ⟨ M ⟩ t − 2 1 σ 2 t and 1 2 σ ‾ 2 t − 1 2 ⟨ M ⟩ t \frac{1}{2}\underline{\sigma}^{2}t-\frac{1}{2}\langle M\rangle_t 2 1 σ 2 t − 2 1 ⟨ M ⟩ t are g g g -martingales.
Indeed, for 0 ≤ s < t 0\leq s<t 0 ≤ s < t ,
E s [ ⟨ M ⟩ t − σ ‾ 2 t ] = E s [ ⟨ M ⟩ t − ⟨ M ⟩ s ] + ⟨ M ⟩ s − σ ‾ 2 t = E s [ ∑ i = 1 d ∫ s t ( α r i ) 2 d ⟨ B i ⟩ r ] + ⟨ M ⟩ s − σ ‾ 2 t = σ ‾ 2 E s [ ∑ i = 1 d ∫ s t ( α r i ) 2 d r ] + ⟨ M ⟩ s − σ ‾ 2 t = ⟨ M ⟩ s − σ ‾ 2 s , \begin{aligned}
\mathbb{E}_{s}\left[\langle M\rangle_t-\overline{\sigma}^{2}t\right]=\mathbb{E}_{s}\left[\langle M\rangle_t-\langle M\rangle_s\right]+\langle M\rangle_s-\overline{\sigma}^{2}t \\
&=\mathbb{E}_{s}\left[\sum_{i=1}^{d}\int_{s}^{t}(\alpha_r^{i})^{2}\mathrm{d}\langle B^{i}\rangle_r\right]+\langle M\rangle_s-\overline{\sigma}^{2}t \\
&=\overline{\sigma}^{2}\mathbb{E}_{s}\left[\sum_{i=1}^{d}\int_{s}^{t}(\alpha_r^{i})^{2}\mathrm{d}r\right]+\langle M\rangle_s-\overline{\sigma}^{2}t \\
&=\langle M\rangle_s-\overline{\sigma}^{2}s,
\end{aligned} E s [ ⟨ M ⟩ t − σ 2 t ] = E s [ ⟨ M ⟩ t − ⟨ M ⟩ s ] + ⟨ M ⟩ s − σ 2 t = E s [ i = 1 ∑ d ∫ s t ( α r i ) 2 d ⟨ B i ⟩ r ] + ⟨ M ⟩ s − σ 2 t = σ 2 E s [ i = 1 ∑ d ∫ s t ( α r i ) 2 d r ] + ⟨ M ⟩ s − σ 2 t = ⟨ M ⟩ s − σ 2 s ,
where in the last step we use the assumption ∥ α r ∥ 2 = 1 \|\alpha_r\|^{2}=1 ∥ α r ∥ 2 = 1 . Similarly, we can that 1 2 σ ‾ 2 t − 1 2 ⟨ M ⟩ t \frac{1}{2}\underline{\sigma}^{2}t-\frac{1}{2}\langle M\rangle_t 2 1 σ 2 t − 2 1 ⟨ M ⟩ t is also a g g g -martingale. The proof is completed. □
Example 1. Let B B B be a 3 3 3 -dimensional G G G -Brownian motion and α ( 2 ) ( t ) : = A \alpha^{(2)}(t):=A α ( 2 ) ( t ) := A be a constant process with
A : = 1 2 ( 1 − 1 0 1 1 0 ) . (32) A:=\frac{1}{\sqrt{2}}\begin{pmatrix}1&-1&0\\1&1&0\end{pmatrix}. \tag*{(32)} A := 2 1 ( 1 1 − 1 1 0 0 ) . (32) Then, W t = ∫ 0 t α ( 2 ) ( r ) d B r = Q ( B t 1 , B t 2 ) T W_t=\int_{0}^{t}\alpha^{(2)}(r)\mathrm{d}B_r=Q(B_t^{1},B_t^{2})^{\mathsf{T}} W t = ∫ 0 t α ( 2 ) ( r ) d B r = Q ( B t 1 , B t 2 ) T , where Q = 1 2 ( 1 − 1 1 1 ) Q=\frac{1}{\sqrt{2}}\begin{pmatrix}1&-1\\1&1\end{pmatrix} Q = 2 1 ( 1 1 − 1 1 ) is an orthogonal matrix.
Under Assumption 1′, clearly ( B t 1 , B t 2 ) T (B_t^{1},B_t^{2})^{\mathsf{T}} ( B t 1 , B t 2 ) T is a 2 2 2 -dimensional G G G -Brownian motion with the same form of G G G . However, Proposition 2.8 implies that Q ( B t 1 , B t 2 ) T Q(B_t^{1},B_t^{2})^{\mathsf{T}} Q ( B t 1 , B t 2 ) T is a 2 2 2 -dimensional G G G -Brownian motion with generator G Q G_Q G Q , where G Q ( M ) = G ( Q T M Q ) G_Q(M)=G(Q^{\mathsf{T}}MQ) G Q ( M ) = G ( Q T M Q ) for M ∈ S 2 M\in\mathbb{S}_{2} M ∈ S 2 . Note that G Q ≠ G G_Q\ne G G Q = G since
G Q ( D ) = 0 ≠ 1 2 ( σ ‾ 2 − σ ‾ 2 ) = G ( D ) G_Q(D)=0\ne\frac{1}{2}(\overline{\sigma}^{2}-\underline{\sigma}^{2})=G(D) G Q ( D ) = 0 = 2 1 ( σ 2 − σ 2 ) = G ( D ) for D = ( 1 0 0 − 1 ) D=\begin{pmatrix}1&0\\0&-1\end{pmatrix} D = ( 1 0 0 − 1 ) . Now, let α ( 2 ) ( t ) = A I [ 0 , T / 2 ] ( t ) + B I [ T / 2 , T ] ( t ) \alpha^{(2)}(t)=A I_{[0,T/2]}(t)+B I_{[T/2,T]}(t) α ( 2 ) ( t ) = A I [ 0 , T /2 ] ( t ) + B I [ T /2 , T ] ( t ) with A A A given by (32) and B : = ( 1 0 0 0 1 0 ) B:=\begin{pmatrix}1&0&0\\0&1&0\end{pmatrix} B := ( 1 0 0 1 0 0 ) , then W W W is a G Q G_Q G Q -Brownian motion on [ 0 , T / 2 ] [0,T/2] [ 0 , T /2 ] while a G G G -Brownian motion on [ T / 2 , T ] [T/2,T] [ T /2 , T ] .
By contrast, if the generator G G G satisfies Assumption 1, then Corollary 2.14, together with the fact that the quadratic covariation is given by
⟨ ∫ 0 t α i ( n ) ( r ) d B r , ∫ 0 t α j ( n ) ( r ) d B r ⟩ = ∫ 0 t α i ( n ) ( r ) ( α j ( n ) ( r ) ) ⊤ d ⟨ B i ⟩ r = 1 { i = j } ⟨ B 1 ⟩ t , \left\langle\int_0^t \alpha_i^{(n)}(r)\,\mathrm{d}B_r,\int_0^t \alpha_j^{(n)}(r)\,\mathrm{d}B_r\right\rangle=\int_0^t \alpha_i^{(n)}(r)(\alpha_j^{(n)}(r))^\top\mathrm{d}\langle B^i\rangle_r=\mathbf{1}_{\{i=j\}}\langle B^1\rangle_t, ⟨ ∫ 0 t α i ( n ) ( r ) d B r , ∫ 0 t α j ( n ) ( r ) d B r ⟩ = ∫ 0 t α i ( n ) ( r ) ( α j ( n ) ( r ) ) ⊤ d ⟨ B i ⟩ r = 1 { i = j } ⟨ B 1 ⟩ t ,
where the second equality is due to (2) , implies that ∫ 0 t α ( n ) ( r ) d B r \int_0^t\alpha^{(n)}(r)\mathrm{d}B_r ∫ 0 t α ( n ) ( r ) d B r is an n n n -dimensional G ‾ \overline{G} G -Brownian motion with G ‾ \overline{G} G still satisfying Assumption 1.
on the path space
Let Ω : = C 0 d ( R + ) \Omega:= C_0^d(\mathbb{R}^+) Ω := C 0 d ( R + ) be the space of R d \mathbb{R}^d R d -valued continuous paths ( ω t ) t ≥ 0 (\omega_t)_{t\geq0} ( ω t ) t ≥ 0 with ω 0 = 0 \omega_0=0 ω 0 = 0 , endowed with the distance
ρ ( ω 1 , ω 2 ) : = ∑ i = 1 ∞ 2 − i [ ( max t ∈ [ 0 , i ] ∣ ω t 1 − ω t 2 ∣ ) ∧ 1 ] . \rho(\omega^1,\omega^2):=\sum_{i=1}^{\infty}2^{-i}\left[\left(\max_{t\in[0,i]}|\omega_t^1-\omega_t^2|\right)\wedge1\right]. ρ ( ω 1 , ω 2 ) := i = 1 ∑ ∞ 2 − i [ ( t ∈ [ 0 , i ] max ∣ ω t 1 − ω t 2 ∣ ) ∧ 1 ] .
Then ( Ω , ρ ) (\Omega,\rho) ( Ω , ρ ) is a Polish space, and we use B ( Ω ) \mathcal{B}(\Omega) B ( Ω ) to denote its Borel σ \sigma σ -algebra. For each fixed T ∈ [ 0 , ∞ ) T\in[0,\infty) T ∈ [ 0 , ∞ ) , denote Ω T = { ω ⋅ ∧ T : ω ∈ Ω } \Omega_T=\{\omega_{\cdot\wedge T}:\omega\in\Omega\} Ω T = { ω ⋅∧ T : ω ∈ Ω } . The canonical process ( B t ( ω ) = ω t ) t ≥ 0 (B_t(\omega)=\omega_t)_{t\geq0} ( B t ( ω ) = ω t ) t ≥ 0 for ω ∈ Ω \omega\in\Omega ω ∈ Ω is then a G G G -Brownian motion on the G G G -expectation space ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ≥ 0 , ( E ^ t ) t ≥ 0 ) (\Omega,L_G^1(\Omega),(L_G^1(\Omega_t))_{t\geq0},(\mathbb{\hat{E}}_t)_{t\geq0}) ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ≥ 0 , ( E ^ t ) t ≥ 0 ) .
In this subsection, the generator G G G is given by
G ( A ) : = 1 2 sup γ ∈ Θ tr [ γ γ ⊤ A ] , A ∈ S d . (33) G(A):=\frac{1}{2}\sup_{\gamma\in\Theta}\operatorname{tr}[\gamma\gamma^\top A],\qquad A\in\mathbb{S}_d. \tag*{(33)} G ( A ) := 2 1 γ ∈ Θ sup tr [ γ γ ⊤ A ] , A ∈ S d . (33)
where Θ \Theta Θ is a bounded and closed subset in R d × d \mathbb{R}^{d\times d} R d × d . It is easy to verify that one can choose
Θ : = { σ I d : σ ‾ ≤ σ ≤ σ ‾ } (34) \Theta:=\{\sigma I_d:\underline{\sigma}\leq\sigma\leq\overline{\sigma}\} \tag*{(34)} Θ := { σ I d : σ ≤ σ ≤ σ } (34)
for G G G satisfying Assumption 1 and
Θ : = { diag ( σ 1 , … , σ d ) : σ ‾ ≤ σ i ≤ σ ‾ , i = 1 , … , d } (35) \Theta:=\{\operatorname{diag}(\sigma_1,\ldots,\sigma_d):\underline{\sigma}\leq\sigma_i\leq\overline{\sigma},\ i=1,\ldots,d\} \tag*{(35)} Θ := { diag ( σ 1 , … , σ d ) : σ ≤ σ i ≤ σ , i = 1 , … , d } (35)
for G G G satisfying Assumption 1 ′ 1' 1 ′ .
Let ( Ω , F , F : = ( F t ) , P ) (\Omega,\mathcal{F},\mathbb{F}:=(\mathcal{F}_t),P) ( Ω , F , F := ( F t ) , P ) be the classical Wiener space, on which ( W t ) t ≥ 0 = ( W t i ) i = 1 , t ≥ 0 d (W_t)_{t\geq0}=(W_t^i)_{i=1,t\geq0}^d ( W t ) t ≥ 0 = ( W t i ) i = 1 , t ≥ 0 d is a d d d -dimensional Wiener process. We denote by A t , T Θ \mathcal{A}_{t,T}^{\Theta} A t , T Θ , the space of all Θ \Theta Θ -valued F \mathbb{F} F -adapted processes on an interval [ t , T ] [t,T] [ t , T ] . For each fixed θ ∈ A 0 , ∞ Θ : = ⋃ n = 1 ∞ A 0 , n Θ \theta\in\mathcal{A}_{0,\infty}^{\Theta}:=\bigcup_{n=1}^{\infty}\mathcal{A}_{0,n}^{\Theta} θ ∈ A 0 , ∞ Θ := ⋃ n = 1 ∞ A 0 , n Θ , we denote by P θ P_\theta P θ the law of the martingale ∫ 0 ⋅ θ s d W s \int_0^\cdot\theta_s\mathrm{d}W_s ∫ 0 ⋅ θ s d W s . Denote P : = { P θ : θ ∈ A 0 , ∞ Θ } \mathcal{P}:=\{P_\theta:\theta\in\mathcal{A}_{0,\infty}^{\Theta}\} P := { P θ : θ ∈ A 0 , ∞ Θ } and let P ‾ = P \overline{\mathcal{P}}=\mathcal{P} P = P be the closure of P \mathcal{P} P under the topology of weak convergence. Then, P \mathcal{P} P is tight and hence P ‾ \overline{\mathcal{P}} P is weakly compact.
Let L 0 ( Ω ) L^0(\Omega) L 0 ( Ω ) denote the space of all B ( Ω ) \mathcal{B}(\Omega) B ( Ω ) -measurable random variables. For X ∈ L 0 ( Ω ) X\in L^0(\Omega) X ∈ L 0 ( Ω ) satisfying E P [ X ] E_P[X] E P [ X ] exists for each P ∈ P ‾ P\in\overline{\mathcal{P}} P ∈ P , define the upper expectations as follows
E ‾ [ X ] : = sup P ∈ P E P [ X ] and E ^ [ X ] : = sup P ∈ P ‾ E P [ X ] . (36) \overline{\mathbb{E}}[X]:=\sup_{P\in\mathcal{P}}E_P[X]\quad\text{and}\quad\widehat{\mathbb{E}}[X]:=\sup_{P\in\overline{\mathcal{P}}}E_P[X]. \tag*{(36)} E [ X ] := P ∈ P sup E P [ X ] and E [ X ] := P ∈ P sup E P [ X ] . (36)
Then, we have
E [ X ] = E ‾ [ X ] = E ^ [ X ] , for all X ∈ L G 1 ( Ω ) , (37) \mathbb{E}[X]=\overline{\mathbb{E}}[X]=\widehat{\mathbb{E}}[X],\qquad\text{for all }X\in L_G^1(\Omega), \tag*{(37)} E [ X ] = E [ X ] = E [ X ] , for all X ∈ L G 1 ( Ω ) , (37)
where we recall that the space L G 1 ( Ω ) L_G^1(\Omega) L G 1 ( Ω ) is defined in Section 2.2. Similarly, we define the upper capacity
c ( A ) : = sup P ∈ P ‾ P ( A ) = E ^ ( 1 A ) , A ∈ B ( Ω ) , c(A):=\sup_{P\in\overline{\mathcal{P}}}P(A)=\widehat{\mathbb{E}}(\mathbf{1}_A),\qquad A\in\mathcal{B}(\Omega), c ( A ) := P ∈ P sup P ( A ) = E ( 1 A ) , A ∈ B ( Ω ) ,
which is indeed a Choquet capacity since P ‾ \overline{\mathcal{P}} P is weakly compact.
Definition 2.17. For a set A ∈ B ( Ω ) A\in\mathcal{B}(\Omega) A ∈ B ( Ω ) , it is called polar if c ( A ) = 0 c(A)=0 c ( A ) = 0 . We say that a property holds quasi-surely (q.s.) if it holds outside a polar set.
The following Borel–Cantelli lemma is useful for identifying polar sets.
Lemma 2.18. Let ( A n ) n ∈ N (A_n)_{n\in\mathbb{N}} ( A n ) n ∈ N be a sequence of Borel sets satisfying ∑ n = 1 ∞ c ( A n ) < ∞ \sum_{n=1}^{\infty}c(A_n)<\infty ∑ n = 1 ∞ c ( A n ) < ∞ , then lim sup n → ∞ A n \limsup_{n\to\infty}A_n lim sup n → ∞ A n is a polar set.
Definition 2.19. A random variable X ∈ L G 0 ( Ω ) X\in L_G^0(\Omega) X ∈ L G 0 ( Ω ) is said to be quasi-continuous (q.c.), if for any ε > 0 \varepsilon>0 ε > 0 , there exists an open set O O O with c ( O ) < ε c(O)<\varepsilon c ( O ) < ε such that X 1 O c : Ω → R X1_{O^c}:\Omega\to\mathbb{R} X 1 O c : Ω → R is a continuous mapping.
We say that a random variable X X X has a q.c. version if there exists a q.c. random variable Y Y Y such that X = Y X=Y X = Y q.s.
We have the following characterizations for L G 1 ( Ω T ) L_G^1(\Omega_T) L G 1 ( Ω T ) and L G 1 ( Ω ) L_G^1(\Omega) L G 1 ( Ω ) , respectively:
L G 1 ( Ω T ) = { X ∈ L 0 ( Ω T ) : X has a q.c. version , lim n → ∞ E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 } ; L G 1 ( Ω ) = { X ∈ L 0 ( Ω ) : X has a q.c. version , lim n → ∞ E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 } . (38) \begin{aligned}
L_G^1(\Omega_T)&=\left\{X\in L^0(\Omega_T):X\text{ has a q.c. version},\ \lim_{n\to\infty}\hat{\mathbb{E}}\left[|X|I_{\{|X|>n\}}\right]=0\right\};\\
L_G^1(\Omega)&=\left\{X\in L^0(\Omega):X\text{ has a q.c. version},\ \lim_{n\to\infty}\hat{\mathbb{E}}\left[|X|I_{\{|X|>n\}}\right]=0\right\}. \tag*{(38)}
\end{aligned} L G 1 ( Ω T ) L G 1 ( Ω ) = { X ∈ L 0 ( Ω T ) : X has a q.c. version , n → ∞ lim E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 } ; = { X ∈ L 0 ( Ω ) : X has a q.c. version , n → ∞ lim E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 } . (38)
The above equations remain valid with E ^ \hat{\mathbb{E}} E ^ replaced by E ‾ \overline{\mathbb{E}} E .
Itô's formula
Itô’s formula was first proved for C 2 C^2 C 2 functions whose second-order derivatives have polynomial growth (see, e.g., [28 ] ). The growth condition on the second-order derivatives was imposed due to the lack of a theory of stopping times, which made localization techniques unavailable. We remark that the stochastic analysis recalled in § 2.2 and 2.3 mainly concerns the space L G 1 ( Ω ) L_G^1(\Omega) L G 1 ( Ω ) , whose random variables are q.c., whereas stopping times do not, in general, possess the q.c. property.
In [24 ] , the Itô’s formula was extended to C 2 C^2 C 2 functions without any restriction on the growth of their second derivatives by analyzing the upper expectation E ^ \hat{\mathbb{E}} E ^ defined in (36) . Note that E ^ \hat{\mathbb{E}} E ^ coincides with the G G G -expectation E \mathbb{E} E on the space L G 1 ( Ω ) L_G^1(\Omega) L G 1 ( Ω ) (see eq. (37) ).
For p ≥ 1 p\geq1 p ≥ 1 , let M b p , 0 ( 0 , T ) M_b^{p,0}(0,T) M b p , 0 ( 0 , T ) denote the space of processes of the form:
η t ( ω ) = ∑ j = 0 n − 1 ξ j ( ω ) I [ t j , t j + 1 ) ( t ) . \eta_t(\omega)=\sum_{j=0}^{n-1}\xi_j(\omega)I_{[t_j,t_{j+1})}(t). η t ( ω ) = j = 0 ∑ n − 1 ξ j ( ω ) I [ t j , t j + 1 ) ( t ) .
In the above, ξ j ∈ B b ( Ω t j ) \xi_j\in B_b(\Omega_{t_j}) ξ j ∈ B b ( Ω t j ) for j ∈ { 0 , … , n − 1 } j\in\{0,\ldots,n-1\} j ∈ { 0 , … , n − 1 } , where B b ( Ω t ) B_b(\Omega_t) B b ( Ω t ) denotes the space of all bounded B ( Ω t ) \mathcal{B}(\Omega_t) B ( Ω t ) -measurable random variables. We denote by M ∗ p ( 0 , T ) M_*^p(0,T) M ∗ p ( 0 , T ) the completion of M b p , 0 ( 0 , T ) M_b^{p,0}(0,T) M b p , 0 ( 0 , T ) under the norm
∥ η ∥ M ∗ p = ( E ^ [ ∫ 0 T ∣ η t ∣ p d t ] ) 1 / p , \|\eta\|_{M_*^p}=\left(\hat{\mathbb{E}}\left[\int_0^T|\eta_t|^p\,\mathrm{d}t\right]\right)^{1/p}, ∥ η ∥ M ∗ p = ( E ^ [ ∫ 0 T ∣ η t ∣ p d t ] ) 1/ p ,
where E ^ \hat{\mathbb{E}} E ^ is given in (36) .
We denote by L ∗ p ( Ω ) L_*^p(\Omega) L ∗ p ( Ω ) (resp. L ∗ p ( Ω t ) L_*^p(\Omega_t) L ∗ p ( Ω t ) ) the completion of B b ( Ω ) B_b(\Omega) B b ( Ω ) under the norm ∥ X ∥ p : = ( E ^ [ ∣ X ∣ p ] ) 1 / p \|X\|_p:=\left(\hat{\mathbb{E}}[|X|^p]\right)^{1/p} ∥ X ∥ p := ( E ^ [ ∣ X ∣ p ] ) 1/ p . Then, L ∗ 1 ( Ω ) L_*^1(\Omega) L ∗ 1 ( Ω ) has the following characterization:
L ∗ 1 ( Ω ) = { X ∈ L 0 ( Ω ) : lim n → ∞ E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 } . L_*^1(\Omega)=\left\{X\in L^0(\Omega):\lim_{n\to\infty}\hat{\mathbb{E}}\left[|X|I_{\{|X|>n\}}\right]=0\right\}. L ∗ 1 ( Ω ) = { X ∈ L 0 ( Ω ) : n → ∞ lim E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 } .
In particular, random variables in L ∗ 1 ( Ω ) L_*^1(\Omega) L ∗ 1 ( Ω ) are not necessarily q.c. (see (38) for a comparison with L G 1 ( Ω ) L_G^1(\Omega) L G 1 ( Ω ) ). It is also worth noting that, given a general weakly compact family P \mathcal{P} P of probability measures, a random variable X X X satisfying E ^ [ ∣ X ∣ ] = sup P ∈ P E P [ ∣ X ∣ ] < ∞ \hat{\mathbb{E}}[|X|]=\sup_{P\in\mathcal{P}}E_P[|X|]<\infty E ^ [ ∣ X ∣ ] = sup P ∈ P E P [ ∣ X ∣ ] < ∞ need not satisfy the condition lim n → ∞ E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 \lim_{n\to\infty}\hat{\mathbb{E}}[|X|I_{\{|X|>n\}}]=0 lim n → ∞ E ^ [ ∣ X ∣ I { ∣ X ∣ > n } ] = 0 (see [9 Examples 20 and 21] ).
We refer to ( Ω , L ∗ 1 ( Ω ) , ( L ∗ 1 ( Ω t ) ) t ≥ 0 , ( E ^ t ) t ≥ 0 ) (\Omega,L_*^1(\Omega),(L_*^1(\Omega_t))_{t\geq0},(\hat{\mathbb{E}}_t)_{t\geq0}) ( Ω , L ∗ 1 ( Ω ) , ( L ∗ 1 ( Ω t ) ) t ≥ 0 , ( E ^ t ) t ≥ 0 ) as the extended G G G -expectation space . Clearly, its restriction to L G 1 ( Ω ) L_G^1(\Omega) L G 1 ( Ω ) coincides with the G G G -expectation space ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ≥ 0 , ( E t ) t ≥ 0 ) (\Omega,L_G^1(\Omega),(L_G^1(\Omega_t))_{t\geq0},(\mathbb{E}_t)_{t\geq0}) ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ≥ 0 , ( E t ) t ≥ 0 ) .
The mapping I : M b , 0 ( 0 , T ) → L G 2 ( Ω T ) I:\mathbb{M}^{b,0}(0,T)\to L^{2}_{G}(\Omega_{T}) I : M b , 0 ( 0 , T ) → L G 2 ( Ω T ) defined as in (31) with M M M replaced by G G G -Brownian motion B B B can be continuously extended to I : M ∗ 2 ( 0 , T ) → L G 2 ( Ω T ) I:\mathbb{M}^{2}_{*}(0,T)\to L^{2}_{G}(\Omega_{T}) I : M ∗ 2 ( 0 , T ) → L G 2 ( Ω T ) , which also defines a stochastic integral of η ∈ M ∗ 2 ( 0 , T ) \eta\in\mathbb{M}^{2}_{*}(0,T) η ∈ M ∗ 2 ( 0 , T ) . We remark that, in this setting, conditional expectation has not yet been properly defined, and hence a martingale theory is not available. Noting that now η t \eta_{t} η t is not necessarily a q.c. random variable, this extends the definition given in Section Section 2.3. In particular, this extended definition of stochastic integrals allows us to define stochastic integrals stopped at a stopping time τ \tau τ :
∫ 0 t ∧ τ η s d B s = ∫ 0 t 1 [ 0 , τ ] ( s ) η s d B s q.s. , \int_{0}^{t\wedge\tau}\eta_{s}\,\mathrm{d}B_{s}=\int_{0}^{t}\mathbf{1}_{[0,\tau]}(s)\eta_{s}\,\mathrm{d}B_{s}\quad\text{q.s.}, ∫ 0 t ∧ τ η s d B s = ∫ 0 t 1 [ 0 , τ ] ( s ) η s d B s q.s. ,
where the notion of a stopping time is defined below.
Definition 2.20. A stopping time τ \tau τ with respect to the filtration ( B ( Ω t ) ) t ≥ 0 (\mathcal{B}(\Omega_{t}))_{t\geq0} ( B ( Ω t ) ) t ≥ 0 is a map on Ω \Omega Ω with values in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that { τ ≤ t } ∈ B ( Ω t ) \{\tau\leq t\}\in\mathcal{B}(\Omega_{t}) { τ ≤ t } ∈ B ( Ω t ) for all t ≥ 0 t\geq0 t ≥ 0 .
For p ≥ 1 p\geq1 p ≥ 1 , let M l o c p ( 0 , T ) \mathbb{M}^{p}_{\mathrm{loc}}(0,T) M loc p ( 0 , T ) denote the space of stochastic processes η \eta η satisfying that there exists a sequence ( τ m ) m ∈ N + (\tau_{m})_{m\in\mathbb{N}^{+}} ( τ m ) m ∈ N + of increasing stopping times with τ m ↑ T \tau_{m}\uparrow T τ m ↑ T q.s., such that η 1 [ 0 , τ m ] ∈ M ∗ p ( 0 , T ) \eta\mathbf{1}_{[0,\tau_{m}]}\in\mathbb{M}^{p}_{*}(0,T) η 1 [ 0 , τ m ] ∈ M ∗ p ( 0 , T ) and ∫ 0 T ∣ η s ∣ p d s < ∞ \int_{0}^{T}|\eta_{s}|^{p}\,\mathrm{d}s<\infty ∫ 0 T ∣ η s ∣ p d s < ∞ , q.s. Then, the definition of ∫ 0 T η s d B s \int_{0}^{T}\eta_{s}\,\mathrm{d}B_{s} ∫ 0 T η s d B s and ∫ 0 T ξ s d ⟨ B ⟩ s \int_{0}^{T}\xi_{s}\,\mathrm{d}\langle B\rangle_{s} ∫ 0 T ξ s d ⟨ B ⟩ s can be extended to η ∈ M l o c 2 ( 0 , T ) \eta\in\mathbb{M}^{2}_{\mathrm{loc}}(0,T) η ∈ M loc 2 ( 0 , T ) and ξ ∈ M l o c 1 ( 0 , T ) \xi\in\mathbb{M}^{1}_{\mathrm{loc}}(0,T) ξ ∈ M loc 1 ( 0 , T ) , respectively.
We can now apply the localization technique to remove the growth condition on the second-order derivatives and obtain the following Itô’s formula (see [24 Theorem 5.4] ).
Proposition 2.21 (Itô’s formula). Let Φ ∈ C 1 , 2 ( [ 0 , T ] × R n ) \Phi\in C^{1,2}([0,T]\times\mathbb{R}^{n}) Φ ∈ C 1 , 2 ([ 0 , T ] × R n ) and X = ( X 1 , … , X n ) X=(X^{1},\ldots,X^{n}) X = ( X 1 , … , X n ) be an n n n -dimensional process on [ 0 , T ] [0,T] [ 0 , T ] of the form:
X t m = X 0 m + ∫ 0 t α s m d s + ∫ 0 t η s m i j d ⟨ B i , B j ⟩ s + ∫ 0 t β s m i d B s i , m = 1 , … , n , X^{m}_{t}=X^{m}_{0}+\int_{0}^{t}\alpha^{m}_{s}\,\mathrm{d}s+\int_{0}^{t}\eta^{mij}_{s}\,\mathrm{d}\langle B^{i},B^{j}\rangle_{s}+\int_{0}^{t}\beta^{mi}_{s}\,\mathrm{d}B^{i}_{s},\quad m=1,\ldots,n, X t m = X 0 m + ∫ 0 t α s m d s + ∫ 0 t η s mij d ⟨ B i , B j ⟩ s + ∫ 0 t β s mi d B s i , m = 1 , … , n , where α m , η m i j ∈ M l o c 1 ( 0 , T ) \alpha^{m},\eta^{mij}\in\mathbb{M}^{1}_{\mathrm{loc}}(0,T) α m , η mij ∈ M loc 1 ( 0 , T ) and β m i ∈ M l o c 2 ( 0 , T ) \beta^{mi}\in\mathbb{M}^{2}_{\mathrm{loc}}(0,T) β mi ∈ M loc 2 ( 0 , T ) . Then, for each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , we have q.s.,
Φ ( t , X t ) − Φ ( 0 , X 0 ) = ∫ 0 t ∂ x m Φ ( u , X u ) β u m i d B u i + ∫ 0 t [ ∂ t Φ ( u , X u ) + ∂ x m Φ ( u , X u ) α u m ] d u + ∫ 0 t [ ∂ x m Φ ( u , X u ) η u m i j + 1 2 ∂ x m x k 2 Φ ( u , X u ) β u m i β u k j ] d ⟨ B i , B j ⟩ u . \Phi(t,X_{t})-\Phi(0,X_{0})=\int_{0}^{t}\partial_{x^{m}}\Phi(u,X_{u})\beta^{mi}_{u}\,\mathrm{d}B^{i}_{u}+\int_{0}^{t}\left[\partial_{t}\Phi(u,X_{u})+\partial_{x^{m}}\Phi(u,X_{u})\alpha^{m}_{u}\right]\,\mathrm{d}u
+\int_{0}^{t}\left[\partial_{x^{m}}\Phi(u,X_{u})\eta^{mij}_{u}+\frac{1}{2}\partial^{2}_{x^{m}x^{k}}\Phi(u,X_{u})\beta^{mi}_{u}\beta^{kj}_{u}\right]\,\mathrm{d}\langle B^{i},B^{j}\rangle_{u}. Φ ( t , X t ) − Φ ( 0 , X 0 ) = ∫ 0 t ∂ x m Φ ( u , X u ) β u mi d B u i + ∫ 0 t [ ∂ t Φ ( u , X u ) + ∂ x m Φ ( u , X u ) α u m ] d u + ∫ 0 t [ ∂ x m Φ ( u , X u ) η u mij + 2 1 ∂ x m x k 2 Φ ( u , X u ) β u mi β u k j ] d ⟨ B i , B j ⟩ u . In the above equations we use the Einstein summation convention.
Weak and strong solutions Consider the following G G G -SDE:
X t = X 0 + ∫ 0 t b ( s , X s ) d s + ∫ 0 t h i j ( s , X s ) d ⟨ B i , B j ⟩ s + ∫ 0 t σ j ( s , X s ) d B s j , t ∈ [ 0 , T ] , (39) X_{t}=X_{0}+\int_{0}^{t}b(s,X_{s})\,\mathrm{d}s+\int_{0}^{t}h_{ij}(s,X_{s})\,\mathrm{d}\langle B^{i},B^{j}\rangle_{s}+\int_{0}^{t}\sigma_{j}(s,X_{s})\,\mathrm{d}B^{j}_{s},\quad t\in[0,T], \tag*{(39)} X t = X 0 + ∫ 0 t b ( s , X s ) d s + ∫ 0 t h ij ( s , X s ) d ⟨ B i , B j ⟩ s + ∫ 0 t σ j ( s , X s ) d B s j , t ∈ [ 0 , T ] , (39) where the initial value X 0 X_{0} X 0 is a constant.
Definition 2.22 (Weak solution). A triple ( G , B , X ) (\mathcal{G},B,X) ( G , B , X ) is called a weak solution to the G G G -SDE (39) if G : = ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ∈ [ 0 , T ] , ( E t ) t ∈ [ 0 , T ] ) \mathcal{G}:=(\Omega,L^{1}_{G}(\Omega),(L^{1}_{G}(\Omega_{t}))_{t\in[0,T]},(\mathbb{E}_{t})_{t\in[0,T]}) G := ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ∈ [ 0 , T ] , ( E t ) t ∈ [ 0 , T ] ) is a G G G -expectation space, B B B is a G G G -Brownian motion on G \mathcal{G} G , X t X_{t} X t is B ( Ω t ) \mathcal{B}(\Omega_{t}) B ( Ω t ) -measurable for each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , and (39) holds q.s.
Consider (39) on a prescribed G G G -expectation space G : = ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ∈ [ 0 , T ] , ( E t ) t ∈ [ 0 , T ] ) \mathcal{G}:=(\Omega,L^{1}_{G}(\Omega),(L^{1}_{G}(\Omega_{t}))_{t\in[0,T]},(\mathbb{E}_{t})_{t\in[0,T]}) G := ( Ω , L G 1 ( Ω ) , ( L G 1 ( Ω t ) ) t ∈ [ 0 , T ] , ( E t ) t ∈ [ 0 , T ] ) , with a given G G G -Brownian motion B B B on G \mathcal{G} G . We define a strong solution as follows.
Definition 2.23 (Strong solution). A process X X X is called a strong solution to the G G G -SDE (39) relative to B B B if ( G , B , X ) (\mathcal{G},B,X) ( G , B , X ) is a weak solution to (39) and X t X_{t} X t is σ { B s : s ∈ [ 0 , t ] } \sigma\{B_{s}:s\in[0,t]\} σ { B s : s ∈ [ 0 , t ]} -measurable for each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
Definition 2.24 (Pathwise uniqueness). We say that pathwise uniqueness holds for the G G G -SDE (39) if, whenever X 1 X^1 X 1 and X 2 X^2 X 2 are two (weak) solutions of (39) on the same G G G -expectation space, driven by the same G G G -Brownian motion B B B and starting from the same initial condition, we have X t 1 = X t 2 X_t^1=X_t^2 X t 1 = X t 2 for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , q.s.
A preliminary study on G G G -Dyson Brownian motion
In this section, we provide a preliminary analysis on G G G -Dyson Brownian motion, i.e., the eigenvalue processes λ N = ( λ 1 N , … , λ N N ) \lambda^N=(\lambda_1^N,\ldots,\lambda_N^N) λ N = ( λ 1 N , … , λ N N ) of the symmetric G G G -Brownian matrix X N X^N X N given by (5) .
For each fixed t t t and i i i , λ i N ( t ) \lambda_i^N(t) λ i N ( t ) is a function of d d d variables ( X k h N ( t ) , 1 ≤ k ≤ h ≤ N ) (X_{kh}^N(t),1\leq k\leq h\leq N) ( X k h N ( t ) , 1 ≤ k ≤ h ≤ N ) , which is denoted by Φ i ( X N ( t ) ) \Phi_i(X^N(t)) Φ i ( X N ( t )) . Then, we have (see, e.g., [34 Appendix] )
∂ Φ i ( X N ) ∂ X k h N = 2 U k i U h i I [ k ≠ h ] + U k i 2 I [ k = h ] ; ∂ 2 Φ i ( X N ) ∂ ( X k h N ) 2 = 2 ∑ j : j ≠ i ∣ U h i U h j + U h i U k j ∣ 2 λ i N − λ j N I [ k ≠ h ] + 2 ∑ j : j ≠ i ∣ U k i U k j ∣ 2 λ i N − λ j N I [ k = h ] ; ∂ 2 Φ i ( X N ) ∂ X k h N ∂ X k ′ h ′ N = 2 ∑ j : j ≠ i ( U k i U h j + U h i U k j ) ( I [ k ≠ h ] + I [ k = h ] / 2 ) ( U k ′ i U h ′ j + U h ′ i U k ′ j ) ( I [ k ′ ≠ h ′ ] + I [ k ′ = h ′ ] / 2 ) λ i N − λ j N . (40) \begin{aligned}
\frac{\partial\Phi_i(X^N)}{\partial X_{kh}^N}&=2U_{ki}U_{hi}\mathbb{I}_{[k\ne h]}+U_{ki}^2\mathbb{I}_{[k=h]};\\
\frac{\partial^2\Phi_i(X^N)}{\partial(X_{kh}^N)^2}&=2\sum_{j:j\ne i}\frac{\lvert U_{hi}U_{hj}+U_{hi}U_{kj}\rvert^2}{\lambda_i^N-\lambda_j^N}\mathbb{I}_{[k\ne h]}+2\sum_{j:j\ne i}\frac{\lvert U_{ki}U_{kj}\rvert^2}{\lambda_i^N-\lambda_j^N}\mathbb{I}_{[k=h]};\\
\frac{\partial^2\Phi_i(X^N)}{\partial X_{kh}^N\partial X_{k'h'}^N}&=2\sum_{j:j\ne i}\frac{(U_{ki}U_{hj}+U_{hi}U_{kj})(\mathbb{I}_{[k\ne h]}+\mathbb{I}_{[k=h]}/2)(U_{k'i}U_{h'j}+U_{h'i}U_{k'j})(\mathbb{I}_{[k'\ne h']}+\mathbb{I}_{[k'=h']}/2)}{\lambda_i^N-\lambda_j^N}. \tag*{(40)}
\end{aligned} ∂ X k h N ∂ Φ i ( X N ) ∂ ( X k h N ) 2 ∂ 2 Φ i ( X N ) ∂ X k h N ∂ X k ′ h ′ N ∂ 2 Φ i ( X N ) = 2 U k i U hi I [ k = h ] + U k i 2 I [ k = h ] ; = 2 j : j = i ∑ λ i N − λ j N ∣ U hi U hj + U hi U k j ∣ 2 I [ k = h ] + 2 j : j = i ∑ λ i N − λ j N ∣ U k i U k j ∣ 2 I [ k = h ] ; = 2 j : j = i ∑ λ i N − λ j N ( U k i U hj + U hi U k j ) ( I [ k = h ] + I [ k = h ] /2 ) ( U k ′ i U h ′ j + U h ′ i U k ′ j ) ( I [ k ′ = h ′ ] + I [ k ′ = h ′ ] /2 ) . (40)
Remark 3.1. For a stochastic process X N X^N X N taking values in S N \mathbb{S}_N S N , denote by Λ t N = diag [ λ i N ( t ) ] \Lambda_t^N=\operatorname{diag}[\lambda_i^N(t)] Λ t N = diag [ λ i N ( t )] the diagonal matrix of eigenvalues of X t N X_t^N X t N , arranged in non-decreasing order. Let U t U_t U t be the matrix whose columns are the corresponding orthonormal eigenvectors of X t N X_t^N X t N . Then, X N X^N X N has the orthogonal decomposition X N = U Λ N U ⊤ X^N=U\Lambda^N U^\top X N = U Λ N U ⊤ .
Assume further that X 0 N X_0^N X 0 N has N N N distinct eigenvalues. By [21 Chapter II, Theorem 1.8] , in a neighbourhood of any real symmetric matrix with simple spectrum, the ordered eigenvalues and the corresponding spectral projections depend analytically, and hence smoothly, on the matrix entries. Moreover, the orthogonal eigenvector matrix constructed by spectral projections can be chosen locally smoothly in the matrix argument. Therefore, matrices Λ N \Lambda^N Λ N and U U U can be chosen as smooth functions of X N X^N X N until the first collision time
τ = inf { t : λ i N ( t ) = λ j N ( t ) for some i ≠ j } . (41) \tau=\inf\{t:\lambda_i^N(t)=\lambda_j^N(t)\text{ for some }i\ne j\}. \tag*{(41)} τ = inf { t : λ i N ( t ) = λ j N ( t ) for some i = j } . (41)
Now we show that the random time τ \tau τ is a stopping time in the sense of Definition 2.20 . Denote F = { x ∈ R N : ∃ i ≠ j , x i = x j } F=\{x\in\mathbb{R}^N:\exists i\ne j,x_i=x_j\} F = { x ∈ R N : ∃ i = j , x i = x j } and d ( x , F ) : = inf y ∈ F ∣ x − y ∣ d(x,F):=\inf_{y\in F}|x-y| d ( x , F ) := inf y ∈ F ∣ x − y ∣ . Then, τ = inf { t > 0 : λ N ( t ) ∈ F } \tau=\inf\{t>0:\lambda^N(t)\in F\} τ = inf { t > 0 : λ N ( t ) ∈ F } . Since X N ( t ) X^N(t) X N ( t ) is B ( Ω t ) \mathcal{B}(\Omega_t) B ( Ω t ) -measurable, so are λ N ( t ) \lambda^N(t) λ N ( t ) and d ( λ N ( t ) , F ) d(\lambda^N(t),F) d ( λ N ( t ) , F ) . By the Hoffman-Wielandt inequality (see, e.g., [1 ] ), λ N \lambda^N λ N is continuous with respect to X N X^N X N , and hence λ N ( t ) \lambda^N(t) λ N ( t ) is a continuous function of t t t . Thus, for each t > 0 t>0 t > 0 ,
{ τ ≤ t } = ⋂ n = 1 ∞ ⋃ s ∈ Q 0 ≤ s ≤ t { d ( λ N ( s ) , F ) < 1 n } ∈ B ( Ω t ) , \{\tau\leq t\}=\bigcap_{n=1}^{\infty}\bigcup_{\substack{s\in\mathbb{Q}\\0\leq s\leq t}}\left\{d(\lambda^N(s),F)<\frac{1}{n}\right\}\in\mathcal{B}(\Omega_t), { τ ≤ t } = n = 1 ⋂ ∞ s ∈ Q 0 ≤ s ≤ t ⋃ { d ( λ N ( s ) , F ) < n 1 } ∈ B ( Ω t ) ,
and hence τ \tau τ is a stopping time.
The following result provides the SDEs satisfied by the G G G -Dyson Brownian motion:
Theorem 3.2. Let X N X^N X N be given by (5) . Assume that the initial eigenvalues of X 0 N X_0^N X 0 N are distinct and strictly ordered as λ 1 N ( 0 ) < λ 2 N ( 0 ) < ⋯ < λ N N ( 0 ) \lambda_1^N(0)<\lambda_2^N(0)<\cdots<\lambda_N^N(0) λ 1 N ( 0 ) < λ 2 N ( 0 ) < ⋯ < λ N N ( 0 ) . For R > 0 R>0 R > 0 , define
τ R : = inf { t > 0 : min i ≠ j ∣ λ i N ( t ) − λ j N ( t ) ∣ < R − 1 } . (42) \tau_R:=\inf\left\{t>0:\min_{i\ne j}\left|\lambda_i^N(t)-\lambda_j^N(t)\right|<R^{-1}\right\}. \tag*{(42)} τ R := inf { t > 0 : i = j min λ i N ( t ) − λ j N ( t ) < R − 1 } . (42)
Then, for t < τ R t<\tau_R t < τ R ,
d λ i N ( t ) = 1 N ∑ j : j ≠ i 1 λ i N − λ j N ( ∑ 1 ≤ k < h ≤ N ∣ U k i U h j + U h i U k j ∣ 2 d ⟨ B k h ⟩ t + ∑ k = 1 N 2 ∣ U k i U k j ∣ 2 d ⟨ B k k ⟩ t ) + 2 N ( ∑ 1 ≤ k < h ≤ N 2 U k i U h i d B k h + ∑ k = 1 N U k i 2 d B k k ) , i = 1 , … , N . (43) \mathrm{d}\lambda_i^N(t)=\frac{1}{N}\sum_{j:j\ne i}\frac{1}{\lambda_i^N-\lambda_j^N}\left(\sum_{1\le k<h\le N}\lvert U_{ki}U_{hj}+U_{hi}U_{kj}\rvert^2\mathrm{d}\langle B_{kh}\rangle_t+\sum_{k=1}^N2\lvert U_{ki}U_{kj}\rvert^2\mathrm{d}\langle B_{kk}\rangle_t\right)
+\frac{\sqrt{2}}{\sqrt{N}}\left(\sum_{1\le k<h\le N}\sqrt{2}U_{ki}U_{hi}\mathrm{d}B_{kh}+\sum_{k=1}^NU_{ki}^2\mathrm{d}B_{kk}\right),\quad i=1,\ldots,N. \tag*{(43)} d λ i N ( t ) = N 1 j : j = i ∑ λ i N − λ j N 1 ( 1 ≤ k < h ≤ N ∑ ∣ U k i U hj + U hi U k j ∣ 2 d ⟨ B k h ⟩ t + k = 1 ∑ N 2 ∣ U k i U k j ∣ 2 d ⟨ B k k ⟩ t ) + N 2 ( 1 ≤ k < h ≤ N ∑ 2 U k i U hi d B k h + k = 1 ∑ N U k i 2 d B k k ) , i = 1 , … , N . (43)
Proof. For t < τ R t<\tau_R t < τ R , Φ i ∈ C 2 ( R n ) \Phi_i\in C^2(\mathbb{R}^n) Φ i ∈ C 2 ( R n ) for i = 1 , … , N i=1,\ldots,N i = 1 , … , N , where n = N ( N + 1 ) 2 n=\frac{N(N+1)}{2} n = 2 N ( N + 1 ) . Then, by Proposition 2.21 and (40) , we get
d λ i N ( t ) = d Φ i ( X N ( t ) ) = ∑ 1 ≤ k < h ≤ N ∂ Φ i ( X N ) ∂ x k h 1 N d B k h + ∑ k = 1 N ∂ Φ i ( X N ) ∂ x k k 2 N d B k k + 1 2 ( ∑ 1 ≤ k < h ≤ N ∂ 2 Φ i ( X N ) ∂ x k h 2 1 N d ⟨ B k h ⟩ t + ∑ k = 1 N ∂ 2 Φ i ( X N ) ∂ x k k 2 2 N d ⟨ B k k ⟩ t ) = ∑ 1 ≤ k < h ≤ N 2 U k i U h i 1 N d B k h + ∑ k = 1 N U k i 2 2 N d B k k + 1 2 ( ∑ 1 ≤ k < h ≤ N 2 ∑ j : j ≠ i ∣ U k i U h j + U h i U k j ∣ 2 λ i N − λ j N 1 N d ⟨ B k h ⟩ t + ∑ k = 1 N 2 ∑ j : j ≠ i ∣ U k i U k j ∣ 2 λ i N − λ j N 2 N d ⟨ B k k ⟩ t ) = ∑ 1 ≤ k < h ≤ N 2 U k i U h i 1 N d B k h + ∑ k = 1 N U k i 2 2 N d B k k + 1 N ∑ j : j ≠ i 1 λ i N − λ j N ( ∑ 1 ≤ k < h ≤ N ∣ U k i U h j + U h i U k j ∣ 2 d ⟨ B k h ⟩ t + ∑ k = 1 N 2 ∣ U k i U k j ∣ 2 d ⟨ B k k ⟩ t ) . \begin{aligned}
\mathrm{d}\lambda_i^N(t)=\mathrm{d}\Phi_i(X^N(t)) \\
&=\sum_{1\le k<h\le N}\frac{\partial\Phi_i(X^N)}{\partial x_{kh}}\frac{1}{\sqrt{N}}\mathrm{d}B_{kh}+\sum_{k=1}^N\frac{\partial\Phi_i(X^N)}{\partial x_{kk}}\frac{\sqrt{2}}{\sqrt{N}}\mathrm{d}B_{kk} \\
&+\frac{1}{2}\left(\sum_{1\le k<h\le N}\frac{\partial^2\Phi_i(X^N)}{\partial x_{kh}^2}\frac{1}{N}\mathrm{d}\langle B_{kh}\rangle_t+\sum_{k=1}^N\frac{\partial^2\Phi_i(X^N)}{\partial x_{kk}^2}\frac{2}{N}\mathrm{d}\langle B_{kk}\rangle_t\right) \\
&=\sum_{1\le k<h\le N}2U_{ki}U_{hi}\frac{1}{\sqrt{N}}\mathrm{d}B_{kh}+\sum_{k=1}^NU_{ki}^2\frac{\sqrt{2}}{\sqrt{N}}\mathrm{d}B_{kk} \\
&+\frac{1}{2}\left(\sum_{1\le k<h\le N}2\sum_{j:j\ne i}\frac{\lvert U_{ki}U_{hj}+U_{hi}U_{kj}\rvert^2}{\lambda_i^N-\lambda_j^N}\frac{1}{N}\mathrm{d}\langle B_{kh}\rangle_t+\sum_{k=1}^N2\sum_{j:j\ne i}\frac{\lvert U_{ki}U_{kj}\rvert^2}{\lambda_i^N-\lambda_j^N}\frac{2}{N}\mathrm{d}\langle B_{kk}\rangle_t\right) \\
&=\sum_{1\le k<h\le N}2U_{ki}U_{hi}\frac{1}{\sqrt{N}}\mathrm{d}B_{kh}+\sum_{k=1}^NU_{ki}^2\frac{\sqrt{2}}{\sqrt{N}}\mathrm{d}B_{kk} \\
&+\frac{1}{N}\sum_{j:j\ne i}\frac{1}{\lambda_i^N-\lambda_j^N}\left(\sum_{1\le k<h\le N}\lvert U_{ki}U_{hj}+U_{hi}U_{kj}\rvert^2\mathrm{d}\langle B_{kh}\rangle_t+\sum_{k=1}^N2\lvert U_{ki}U_{kj}\rvert^2\mathrm{d}\langle B_{kk}\rangle_t\right).
\end{aligned} d λ i N ( t ) = d Φ i ( X N ( t )) = 1 ≤ k < h ≤ N ∑ ∂ x k h ∂ Φ i ( X N ) N 1 d B k h + k = 1 ∑ N ∂ x k k ∂ Φ i ( X N ) N 2 d B k k + 2 1 ( 1 ≤ k < h ≤ N ∑ ∂ x k h 2 ∂ 2 Φ i ( X N ) N 1 d ⟨ B k h ⟩ t + k = 1 ∑ N ∂ x k k 2 ∂ 2 Φ i ( X N ) N 2 d ⟨ B k k ⟩ t ) = 1 ≤ k < h ≤ N ∑ 2 U k i U hi N 1 d B k h + k = 1 ∑ N U k i 2 N 2 d B k k + 2 1 1 ≤ k < h ≤ N ∑ 2 j : j = i ∑ λ i N − λ j N ∣ U k i U hj + U hi U k j ∣ 2 N 1 d ⟨ B k h ⟩ t + k = 1 ∑ N 2 j : j = i ∑ λ i N − λ j N ∣ U k i U k j ∣ 2 N 2 d ⟨ B k k ⟩ t = 1 ≤ k < h ≤ N ∑ 2 U k i U hi N 1 d B k h + k = 1 ∑ N U k i 2 N 2 d B k k + N 1 j : j = i ∑ λ i N − λ j N 1 ( 1 ≤ k < h ≤ N ∑ ∣ U k i U hj + U hi U k j ∣ 2 d ⟨ B k h ⟩ t + k = 1 ∑ N 2 ∣ U k i U k j ∣ 2 d ⟨ B k k ⟩ t ) .
□
Denote the martingale part on the right-hand side of (43) by,
d M i ( t ) : = ∑ 1 ≤ k < h ≤ N 2 U k i U h i d B k h + ∑ k = 1 N U k i 2 d B k k , i = 1 , … , N . \mathrm{d}M_i(t):=\sum_{1\le k<h\le N}\sqrt{2}U_{ki}U_{hi}\mathrm{d}B_{kh}+\sum_{k=1}^NU_{ki}^2\mathrm{d}B_{kk},\quad i=1,\ldots,N. d M i ( t ) := 1 ≤ k < h ≤ N ∑ 2 U k i U hi d B k h + k = 1 ∑ N U k i 2 d B k k , i = 1 , … , N .
We have
∑ 1 ≤ k < h ≤ N ( 2 U k i U h i ) 2 + ∑ k = 1 N U k i 4 = ( ∑ k = 1 N U k i 2 ) 2 = 1 , \sum_{1\le k<h\le N}(\sqrt{2}U_{ki}U_{hi})^2+\sum_{k=1}^NU_{ki}^4=\left(\sum_{k=1}^NU_{ki}^2\right)^2=1, 1 ≤ k < h ≤ N ∑ ( 2 U k i U hi ) 2 + k = 1 ∑ N U k i 4 = ( k = 1 ∑ N U k i 2 ) 2 = 1 ,
and thus Corollary 2.15 yields that, under the diagonal structure assumption 1 ′ 1' 1 ′ , each component M i ( t ) M_i(t) M i ( t ) is a G ~ \widetilde{G} G -Brownian motion, for i = 1 , 2 , … , N i=1,2,\ldots,N i = 1 , 2 , … , N . However, M = ( M 1 , … , M N ) ⊤ M=(M_1,\ldots,M_N)^\top M = ( M 1 , … , M N ) ⊤ may fail to be a G G G -Brownian motion for any G G G . By contrast, under Assumption 1 for G G G , by Corollary 2.14 M M M is a G ‾ \overline{G} G -Brownian motion with G ‾ \overline{G} G still satisfying Assumption 1. See Remark 2.16 for detailed discussions. Moreover, under Assumption 1, the G G G -Dyson Brownian motion given by (43) admits a more concise representation (9) , as shown in Theorem 1.3.
Orthogonal invariance and the GOE property
In the classical setting, a Gaussian Orthogonal Ensemble (GOE) matrix H H H satisfies O T H O = d H O^{\mathsf{T}}HO \mathrel{\stackrel{d}{=}} H O T H O = d H for any orthogonal matrix O O O . Let { ξ i j , 1 ≤ i ≤ j ≤ N } \{\xi_{ij}, 1 \le i \le j \le N\} { ξ ij , 1 ≤ i ≤ j ≤ N } be a family of i.i.d. standard Gaussian random variables. Let P N P_N P N be the law of the symmetric random matrix
H = [ 2 ξ 11 ξ 12 ⋯ ξ 1 N ξ 12 2 ξ 22 ⋯ ξ 2 N ⋮ ⋮ ⋱ ⋮ ξ 1 N ξ 2 N ⋯ 2 ξ N N ] . H=\begin{bmatrix}
\sqrt{2}\xi_{11} & \xi_{12} & \cdots& \xi_{1N}\\
\xi_{12} & \sqrt{2}\xi_{22} & \cdots& \xi_{2N}\\
\vdots& \vdots& \ddots& \vdots\\
\xi_{1N} & \xi_{2N} & \cdots& \sqrt{2}\xi_{NN}
\end{bmatrix}. H = 2 ξ 11 ξ 12 ⋮ ξ 1 N ξ 12 2 ξ 22 ⋮ ξ 2 N ⋯ ⋯ ⋱ ⋯ ξ 1 N ξ 2 N ⋮ 2 ξ N N .
Then, O T H O O^{\mathsf{T}}HO O T H O still has the law P N P_N P N for any N × N N \times N N × N orthogonal matrix O O O and thus H H H is a GOE matrix (see, e.g., [1 ] ).
In this subsection, we investigate the GOE property under G G G -expectation. We first identify a symmetric matrix H = ( H i j ) 1 ≤ i , j ≤ N H=(H_{ij})_{1\le i,j\le N} H = ( H ij ) 1 ≤ i , j ≤ N with the vector
H ~ = ( H 11 , H 12 , … , H 1 N , H 22 , … , H 2 N , … , H N N ) T . (44) \widetilde{H}=(H_{11},H_{12},\ldots,H_{1N},H_{22},\ldots,H_{2N},\ldots,H_{NN})^{\mathsf{T}}. \tag*{(44)} H = ( H 11 , H 12 , … , H 1 N , H 22 , … , H 2 N , … , H N N ) T . (44)
Let D D D be an d × d d \times d d × d diagonal matrix, recalling that d : = N ( N + 1 ) 2 d:=\frac{N(N+1)}{2} d := 2 N ( N + 1 ) , whose diagonal entries corresponding to the diagonal entries of H H H are 1 / 2 1/\sqrt{2} 1/ 2 , and whose remaining diagonal entries are equal to 1. Thus, we have
D H ~ = ( ξ 11 , ξ 12 , … , ξ 1 N , ξ 22 , … , ξ 2 N , … , ξ N N ) T = : ξ . D\widetilde{H}=(\xi_{11},\xi_{12},\ldots,\xi_{1N},\xi_{22},\ldots,\xi_{2N},\ldots,\xi_{NN})^{\mathsf{T}}=:\xi. D H = ( ξ 11 , ξ 12 , … , ξ 1 N , ξ 22 , … , ξ 2 N , … , ξ N N ) T =: ξ .
For any N × N N \times N N × N orthogonal matrix O O O , there exists a linear map Q : R d ⟶ R d Q:\mathbb{R}^{d}\longrightarrow\mathbb{R}^{d} Q : R d ⟶ R d such that D O T H O ~ = Q ξ D\widetilde{O^{\mathsf{T}}HO}=Q\xi D O T H O = Q ξ , where O T H O ~ \widetilde{O^{\mathsf{T}}HO} O T H O denotes the vector consisting of the the upper-triangular entries of O T H O O^{\mathsf{T}}HO O T H O , as in (44) . We show that the matrix Q Q Q is orthogonal.
Given any two symmetric matrices H 1 H^{1} H 1 and H 2 H^{2} H 2 , we have
⟨ D H ~ 1 , D H ~ 2 ⟩ = 1 2 ∑ i = 1 N H i i 1 H i i 2 + ∑ 1 ≤ i < j ≤ N H i j 1 H i j 2 = 1 2 ∑ i = 1 N ∑ j = 1 N H i j 1 H i j 2 = 1 2 tr ( H 1 H 2 ) . \left\langle D\widetilde{H}^{1},D\widetilde{H}^{2}\right\rangle=\frac{1}{2}\sum_{i=1}^{N}H_{ii}^{1}H_{ii}^{2}+\sum_{1\le i<j\le N}H_{ij}^{1}H_{ij}^{2}=\frac{1}{2}\sum_{i=1}^{N}\sum_{j=1}^{N}H_{ij}^{1}H_{ij}^{2}=\frac{1}{2}\operatorname{tr}(H^{1}H^{2}). ⟨ D H 1 , D H 2 ⟩ = 2 1 i = 1 ∑ N H ii 1 H ii 2 + 1 ≤ i < j ≤ N ∑ H ij 1 H ij 2 = 2 1 i = 1 ∑ N j = 1 ∑ N H ij 1 H ij 2 = 2 1 tr ( H 1 H 2 ) .
Using the above identity, we obtain
∥ Q ξ ∥ 2 = ⟨ D O T H O ~ , D O T H O ~ ⟩ = 1 2 tr ( ( O T H O ) 2 ) = 1 2 tr ( H 2 ) = ⟨ D H ~ , D H ~ ⟩ = ∥ ξ ∥ 2 . \|Q\xi\|^{2}=\left\langle D\widetilde{O^{\mathsf{T}}HO},D\widetilde{O^{\mathsf{T}}HO}\right\rangle=\frac{1}{2}\operatorname{tr}\left((O^{\mathsf{T}}HO)^{2}\right)=\frac{1}{2}\operatorname{tr}(H^{2})=\left\langle D\widetilde{H},D\widetilde{H}\right\rangle=\|\xi\|^{2}. ∥ Q ξ ∥ 2 = ⟨ D O T H O , D O T H O ⟩ = 2 1 tr ( ( O T H O ) 2 ) = 2 1 tr ( H 2 ) = ⟨ D H , D H ⟩ = ∥ ξ ∥ 2 .
This implies that Q Q Q is an orthogonal matrix.
Now let ξ i j = B i j ( 1 ) \xi_{ij}=B_{ij}(1) ξ ij = B ij ( 1 ) , where B = ( B 11 , B 12 , … , B 1 N , B 22 , … , B 2 N , … , B N N ) T B=(B_{11},B_{12},\ldots,B_{1N},B_{22},\ldots,B_{2N},\ldots,B_{NN})^{\mathsf{T}} B = ( B 11 , B 12 , … , B 1 N , B 22 , … , B 2 N , … , B N N ) T is an d d d -dimensional G G G -Brownian motion. By the construction of D D D , we have D H ~ = B ( 1 ) D\widetilde{H}=B(1) D H = B ( 1 ) . Our goal is to determine whether O T H O O^{\mathsf{T}}HO O T H O and H H H are identically distributed under the G G G -expectation for any orthogonal matrix O O O . That is, the question is whether O T H O = d H ~ O^{\mathsf{T}}HO\mathrel{\stackrel{d}{=}}\widetilde{H} O T H O = d H , which is equivalent to D O T H O ~ = d D H ~ D\widetilde{O^{\mathsf{T}}HO}\mathrel{\stackrel{d}{=}}D\widetilde{H} D O T H O = d D H , or, equivalently, to Q B ( 1 ) = d B ( 1 ) QB(1)\mathrel{\stackrel{d}{=}}B(1) QB ( 1 ) = d B ( 1 ) . Thus, the orthogonal invariance of the matrix H H H reduces to the invariance in distribution of the underlying G G G -Brownian motion under orthogonal transformations Q Q Q .
By Proposition 2.8 we know that Q B QB QB is a G Q G_Q G Q -Brownian motion, where G Q ( A ) = G ( Q T A Q ) G_Q(A)=G(Q^{\mathsf{T}}AQ) G Q ( A ) = G ( Q T A Q ) . Assuming that G G G satisfies Assumption 1, it is easy to verify that G Q = G G_Q=G G Q = G . Thus, Q B = d B QB\mathrel{\stackrel{d}{=}}B QB = d B and hence the corresponding matrix H H H is a GOE matrix under G G G -expectation.
However, the orthogonal invariance may fail under Assumption 1’ if σ ‾ < σ ‾ \underline{\sigma}<\overline{\sigma} σ < σ , as indicated by the following example.
Example 2. When N = 2 N=2 N = 2 , for an orthogonal 2 × 2 2 \times2 2 × 2 matrix O = 1 2 ( 1 − 1 1 1 ) O=\frac{1}{\sqrt{2}}\begin{pmatrix} 1 & -1 \\ 1 & 1 \end{pmatrix} O = 2 1 ( 1 1 − 1 1 ) , the corresponding 3 × 3 3 \times3 3 × 3 orthogonal matrix Q Q Q is given by
Q = ( 1 2 1 2 1 2 − 1 2 0 1 2 1 2 − 1 2 1 2 ) . Q=\begin{pmatrix}
\frac{1}{2} & \frac{1}{\sqrt{2}} & \frac{1}{2}\\
-\frac{1}{\sqrt{2}} & 0 & \frac{1}{\sqrt{2}}\\
\frac{1}{2} & -\frac{1}{\sqrt{2}} & \frac{1}{2}
\end{pmatrix}. Q = 2 1 − 2 1 2 1 2 1 0 − 2 1 2 1 2 1 2 1 . Under Assumption 1 ′ 1' 1 ′ with σ < σ ˉ \sigma< \bar{\sigma} σ < σ ˉ , it can be verified that
G Q ( A ) = G ( Q T A Q ) = 0 ≠ 1 2 ( σ ˉ 2 − σ 2 ) = G ( A ) , G_Q(A)=G(Q^{\mathsf{T}}AQ)=0\ne\frac{1}{2}\left(\bar{\sigma}^{2}-\sigma^{2}\right)=G(A), G Q ( A ) = G ( Q T A Q ) = 0 = 2 1 ( σ ˉ 2 − σ 2 ) = G ( A ) , where
A = ( 1 0 0 0 0 0 0 0 − 1 ) , and Q T A Q = ( 0 1 2 0 1 2 0 1 2 0 1 2 0 ) . A=\begin{pmatrix}
1 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & -1
\end{pmatrix},\quad\text{and}\quad Q^{\mathsf{T}}AQ=\begin{pmatrix}
0 & \frac{1}{\sqrt{2}} & 0\\
\frac{1}{\sqrt{2}} & 0 & \frac{1}{\sqrt{2}}\\
0 & \frac{1}{\sqrt{2}} & 0
\end{pmatrix}. A = 1 0 0 0 0 0 0 0 − 1 , and Q T A Q = 0 2 1 0 2 1 0 2 1 0 2 1 0 . This indicates that Q B QB QB and B B B have different distributions under Assumption 1 ′ 1' 1 ′ .
A colliding G G G -Dyson Brownian motion: an example
In this subsection, we provide an example demonstrating the non-collision property may fail for G G G -Dyson Brownian motion.
When Ω = C 0 d ( R + ) \Omega=C^d_0(\mathbb{R}^{+}) Ω = C 0 d ( R + ) , the G G G -expectation E \mathbb{E} E coincides with the upper expectation E ^ [ ⋅ ] = sup P ∈ P E P [ ⋅ ] \widehat{\mathbb{E}}[\cdot]=\sup_{P\in\mathcal{P} }E_P[\cdot] E [ ⋅ ] = sup P ∈ P E P [ ⋅ ] on L G 1 ( Ω ) L_G^1(\Omega) L G 1 ( Ω ) , where P = P ‾ \mathcal{P}=\overline{\mathcal{P}} P = P with P = { P θ : θ ∈ A 0 , ∞ Θ } \mathcal{P}=\{P_\theta:\theta\in\mathcal{A}^{\Theta}_{0,\infty}\} P = { P θ : θ ∈ A 0 , ∞ Θ } . Here, P θ P_\theta P θ denotes the law of ∫ 0 ⋅ θ d W s \int_0^\cdot\theta dW_s ∫ 0 ⋅ θ d W s , where W W W is a d d d -dimensional Wiener process on Ω \Omega Ω , and A 0 , ∞ Θ \mathcal{A}^{\Theta}_{0,\infty} A 0 , ∞ Θ denotes the space of Θ \Theta Θ -valued adapted processes, with Θ \Theta Θ being a given bounded closed subset of R d × d \mathbb{R}^{d\times d} R d × d . We refer to Section 2.3 for more details.
Consider the case N = 2 N=2 N = 2 (i.e., d = 3 d=3 d = 3 ). Assume that the generator G G G of the underlying G G G -Brownian motion is given by (33) , with
Θ = { θ = ( θ i j ) 3 × 3 ∈ R 3 × 3 : 0 < σ ‾ 2 ≤ θ i 1 2 + θ i 2 2 + θ i 3 2 ≤ σ ˉ 2 , i = 1 , 2 , 3 } . (45) \Theta=\left\{\theta=(\theta_{ij})_{3\times3}\in\mathbb{R}^{3\times3}:0<\underline{\sigma}^{2}\le\theta_{i1}^{2}+\theta_{i2}^{2}+\theta_{i3}^{2}\le\bar{\sigma}^{2},\ i=1,2,3\right\}. \tag*{(45)} Θ = { θ = ( θ ij ) 3 × 3 ∈ R 3 × 3 : 0 < σ 2 ≤ θ i 1 2 + θ i 2 2 + θ i 3 2 ≤ σ ˉ 2 , i = 1 , 2 , 3 } . (45)
Assuming λ 1 2 ( 0 ) < λ 2 2 ( 0 ) \lambda_1^2(0)<\lambda_2^2(0) λ 1 2 ( 0 ) < λ 2 2 ( 0 ) , we show below that there exists θ ∗ ∈ A 0 , ∞ Θ \theta^*\in\mathcal{A}^{\Theta}_{0,\infty} θ ∗ ∈ A 0 , ∞ Θ , such that λ 2 2 ( t ) − λ 1 2 ( t ) \lambda_2^2(t)-\lambda_1^2(t) λ 2 2 ( t ) − λ 1 2 ( t ) hits zero in finite time P θ ∗ P_{\theta^*} P θ ∗ -almost surely. This indicates that the 2-dimensional G G G -Dyson Brownian motion collides q.s.
Let
X t = ( X t 11 X t 12 X t 12 X t 22 ) = ( 2 B t 11 + X 0 11 B t 12 + X 0 12 B t 12 + X 0 12 2 B t 22 + X 0 22 ) X_t=\begin{pmatrix}X_t^{11} & X_t^{12}\\ X_t^{12} & X_t^{22}\end{pmatrix}
=\begin{pmatrix}\sqrt{2}B_t^{11}+X_0^{11} & B_t^{12}+X_0^{12}\\ B_t^{12}+X_0^{12} & \sqrt{2}B_t^{22}+X_0^{22}\end{pmatrix} X t = ( X t 11 X t 12 X t 12 X t 22 ) = ( 2 B t 11 + X 0 11 B t 12 + X 0 12 B t 12 + X 0 12 2 B t 22 + X 0 22 )
be a symmetric matrix-valued process with distinct initial eigenvalues, where B = ( B 11 , B 12 , B 22 ) B=(B^{11},B^{12},B^{22}) B = ( B 11 , B 12 , B 22 ) is a 3-dimensional G G G -Brownian motion. Define
Z t : = X t 11 − X t 22 , Y t : = 2 X t 12 , R t : = Z t 2 + Y t 2 , τ ∗ : = inf { t ≥ 0 : R t = 0 } . Z_t:=X_t^{11}-X_t^{22},\quad Y_t:=2X_t^{12},\quad R_t:=\sqrt{Z_t^2+Y_t^2},\quad\tau^*:=\inf\{t\ge0:R_t=0\}. Z t := X t 11 − X t 22 , Y t := 2 X t 12 , R t := Z t 2 + Y t 2 , τ ∗ := inf { t ≥ 0 : R t = 0 } .
Then, two eigenvalues of X t X_t X t collide if and only if R t = 0 R_t=0 R t = 0 , and τ ∗ \tau^* τ ∗ is the first spectral collision time.
Given the bounds 0 < σ < σ ˉ < ∞ 0<\sigma<\bar{\sigma}<\infty 0 < σ < σ ˉ < ∞ , we introduce
α = 4 σ ˉ 2 , β = 4 σ 2 . \alpha=4\bar{\sigma}^{2},\qquad\beta=4\sigma^{2}. α = 4 σ ˉ 2 , β = 4 σ 2 .
::: We now construct the process θ ∗ ( t ) = ( θ 1 ∗ ( t ) , θ 2 ∗ ( t ) , θ 3 ∗ ( t ) ) T \theta^*(t)=(\theta_1^*(t),\theta_2^*(t),\theta_3^*(t))^\mathsf{T} θ ∗ ( t ) = ( θ 1 ∗ ( t ) , θ 2 ∗ ( t ) , θ 3 ∗ ( t ) ) T :
θ 1 ∗ ( t ) : = ( α 2 2 Z t R t , − β 2 2 Y t R t , β 2 2 ) T , θ 2 ∗ ( t ) : = ( α 2 Y t R t , β 2 Z t R t , 0 ) T , θ 3 ∗ ( t ) : = ( − α 2 2 Z t R t , β 2 2 Y t R t , β 2 2 ) T . (46) \begin{aligned}
\theta_1^*(t)&:=\left(\frac{\sqrt{\alpha}}{2\sqrt{2}}\frac{Z_t}{R_t},-\frac{\sqrt{\beta}}{2\sqrt{2}}\frac{Y_t}{R_t},\frac{\sqrt{\beta}}{2\sqrt{2}}\right)^\mathsf{T},\\
\theta_2^*(t)&:=\left(\frac{\sqrt{\alpha}}{2}\frac{Y_t}{R_t},\frac{\sqrt{\beta}}{2}\frac{Z_t}{R_t},0\right)^\mathsf{T},\\
\theta_3^*(t)&:=\left(-\frac{\sqrt{\alpha}}{2\sqrt{2}}\frac{Z_t}{R_t},\frac{\sqrt{\beta}}{2\sqrt{2}}\frac{Y_t}{R_t},\frac{\sqrt{\beta}}{2\sqrt{2}}\right)^\mathsf{T}. \tag*{(46)}
\end{aligned} θ 1 ∗ ( t ) θ 2 ∗ ( t ) θ 3 ∗ ( t ) := ( 2 2 α R t Z t , − 2 2 β R t Y t , 2 2 β ) T , := ( 2 α R t Y t , 2 β R t Z t , 0 ) T , := ( − 2 2 α R t Z t , 2 2 β R t Y t , 2 2 β ) T . (46)
We have σ ‾ 2 = β 4 ≤ ∣ θ 2 ∗ ( t ) ∣ 2 = α 4 Y t 2 R t 2 + β 4 Z t 2 R t 2 ≤ α 4 = σ ‾ 2 \underline{\sigma}^2=\frac{\beta}{4}\leq\lvert\theta_2^*(t)\rvert^2=\frac{\alpha}{4}\frac{Y_t^2}{R_t^2}+\frac{\beta}{4}\frac{Z_t^2}{R_t^2}\leq\frac{\alpha}{4}=\overline{\sigma}^2 σ 2 = 4 β ≤ ∣ θ 2 ∗ ( t ) ∣ 2 = 4 α R t 2 Y t 2 + 4 β R t 2 Z t 2 ≤ 4 α = σ 2 and σ ‾ 2 = β + γ 8 ≤ ∣ θ 1 ∗ ( t ) ∣ 2 = ∣ θ 3 ∗ ( t ) ∣ 2 = α 8 Z t 2 R t 2 + β 8 Y t 2 R t 2 + β 8 ≤ α + β 8 < σ ‾ 2 \underline{\sigma}^2=\frac{\beta+\gamma}{8}\leq\lvert\theta_1^*(t)\rvert^2=\lvert\theta_3^*(t)\rvert^2=\frac{\alpha}{8}\frac{Z_t^2}{R_t^2}+\frac{\beta}{8}\frac{Y_t^2}{R_t^2}+\frac{\beta}{8}\leq\frac{\alpha+\beta}{8}<\overline{\sigma}^2 σ 2 = 8 β + γ ≤ ∣ θ 1 ∗ ( t ) ∣ 2 = ∣ θ 3 ∗ ( t ) ∣ 2 = 8 α R t 2 Z t 2 + 8 β R t 2 Y t 2 + 8 β ≤ 8 α + β < σ 2 . Thus, we get that ∣ θ i ∗ ( t ) ∣ 2 ∈ [ σ ‾ 2 , σ ‾ 2 ] \lvert\theta_i^*(t)\rvert^2\in[\underline{\sigma}^2,\overline{\sigma}^2] ∣ θ i ∗ ( t ) ∣ 2 ∈ [ σ 2 , σ 2 ] , i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 for all t ∈ [ 0 , τ ∗ ) t\in[0,\tau^*) t ∈ [ 0 , τ ∗ ) .
Then, under P θ ∗ P_{\theta^*} P θ ∗ we have,
d Z t = 2 ( θ 1 ∗ ( t ) − θ 3 ∗ ( t ) ) T d W t = α Z t R t d W t 11 − β Y t R t d W t 12 , dZ_t=\sqrt{2}(\theta_1^*(t)-\theta_3^*(t))^\mathsf{T}dW_t=\sqrt{\alpha}\frac{Z_t}{R_t}dW_t^{11}-\sqrt{\beta}\frac{Y_t}{R_t}dW_t^{12}, d Z t = 2 ( θ 1 ∗ ( t ) − θ 3 ∗ ( t ) ) T d W t = α R t Z t d W t 11 − β R t Y t d W t 12 ,
d Y t = 2 θ 2 ∗ ( t ) T d W t = α Y t R t d W t 11 + β Z t R t d W t 12 , dY_t=2\theta_2^*(t)^\mathsf{T}dW_t=\sqrt{\alpha}\frac{Y_t}{R_t}dW_t^{11}+\sqrt{\beta}\frac{Z_t}{R_t}dW_t^{12}, d Y t = 2 θ 2 ∗ ( t ) T d W t = α R t Y t d W t 11 + β R t Z t d W t 12 ,
where W = ( W 11 , W 12 , W 22 ) W=(W^{11},W^{12},W^{22}) W = ( W 11 , W 12 , W 22 ) is a 3-dimensional Wiener process. Thus, we get
d ⟨ Z ⟩ t = ( α Z t 2 R t 2 + β Y t 2 R t 2 ) d t , d ⟨ Y ⟩ t = ( α Y t 2 R t 2 + β Z t 2 R t 2 ) d t , d ⟨ Z , Y ⟩ t = ( α − β ) Z t Y t R t 2 d t . d\langle Z\rangle_t=\left(\alpha\frac{Z_t^2}{R_t^2}+\beta\frac{Y_t^2}{R_t^2}\right)dt,\qquad d\langle Y\rangle_t=\left(\alpha\frac{Y_t^2}{R_t^2}+\beta\frac{Z_t^2}{R_t^2}\right)dt,\qquad d\langle Z,Y\rangle_t=(\alpha-\beta)\frac{Z_tY_t}{R_t^2}dt. d ⟨ Z ⟩ t = ( α R t 2 Z t 2 + β R t 2 Y t 2 ) d t , d ⟨ Y ⟩ t = ( α R t 2 Y t 2 + β R t 2 Z t 2 ) d t , d ⟨ Z , Y ⟩ t = ( α − β ) R t 2 Z t Y t d t .
Applying Itô’s formula to R t = Z t 2 + Y t 2 R_t=\sqrt{Z_t^2+Y_t^2} R t = Z t 2 + Y t 2 , we obtain for t < τ ∗ t<\tau^* t < τ ∗ ,
d R t = Z t R t d Z t + Y t R t d Y t + 1 2 ( ∂ 2 R t ∂ Z t 2 d ⟨ Z ⟩ t + 2 ∂ 2 R t ∂ Z t ∂ Y t d ⟨ Z , Y ⟩ t + ∂ 2 R t ∂ Y t 2 d ⟨ Y ⟩ t ) = Z t R t d Z t + Y t R t d Y t + 1 2 ( Y t 2 R t 3 d ⟨ Z ⟩ t − 2 Z t Y t R t 3 d ⟨ Z , Y ⟩ t + Z t 2 R t 3 d ⟨ Y ⟩ t ) = Z t R t ( α Z t R t d W t 11 − β Y t R t d W t 12 ) + Y t R t ( α Y t R t d W t 11 + β Z t R t d W t 12 ) + 1 2 R t [ Y t 2 R t 2 ( α Z t 2 R t 2 + β Y t 2 R t 2 ) + Z t 2 R t 2 ( α Y t 2 R t 2 + β Z t 2 R t 2 ) − 2 Y t 2 R t 2 Z t 2 R t 2 ( α − β ) ] d t = α d W t 11 + β 2 R t d t . \begin{aligned}
dR_t&=\frac{Z_t}{R_t}dZ_t+\frac{Y_t}{R_t}dY_t+\frac{1}{2}\left(\frac{\partial^2R_t}{\partial Z_t^2}d\langle Z\rangle_t+2\frac{\partial^2R_t}{\partial Z_t\partial Y_t}d\langle Z,Y\rangle_t+\frac{\partial^2R_t}{\partial Y_t^2}d\langle Y\rangle_t\right)\\
&=\frac{Z_t}{R_t}dZ_t+\frac{Y_t}{R_t}dY_t+\frac{1}{2}\left(\frac{Y_t^2}{R_t^3}d\langle Z\rangle_t-2\frac{Z_tY_t}{R_t^3}d\langle Z,Y\rangle_t+\frac{Z_t^2}{R_t^3}d\langle Y\rangle_t\right)\\
&=\frac{Z_t}{R_t}\left(\sqrt{\alpha}\frac{Z_t}{R_t}dW_t^{11}-\sqrt{\beta}\frac{Y_t}{R_t}dW_t^{12}\right)+\frac{Y_t}{R_t}\left(\sqrt{\alpha}\frac{Y_t}{R_t}dW_t^{11}+\sqrt{\beta}\frac{Z_t}{R_t}dW_t^{12}\right)\\
&\quad+\frac{1}{2R_t}\left[\frac{Y_t^2}{R_t^2}\left(\alpha\frac{Z_t^2}{R_t^2}+\beta\frac{Y_t^2}{R_t^2}\right)+\frac{Z_t^2}{R_t^2}\left(\alpha\frac{Y_t^2}{R_t^2}+\beta\frac{Z_t^2}{R_t^2}\right)-2\frac{Y_t^2}{R_t^2}\frac{Z_t^2}{R_t^2}(\alpha-\beta)\right]dt\\
&=\sqrt{\alpha}dW_t^{11}+\frac{\beta}{2R_t}dt.
\end{aligned} d R t = R t Z t d Z t + R t Y t d Y t + 2 1 ( ∂ Z t 2 ∂ 2 R t d ⟨ Z ⟩ t + 2 ∂ Z t ∂ Y t ∂ 2 R t d ⟨ Z , Y ⟩ t + ∂ Y t 2 ∂ 2 R t d ⟨ Y ⟩ t ) = R t Z t d Z t + R t Y t d Y t + 2 1 ( R t 3 Y t 2 d ⟨ Z ⟩ t − 2 R t 3 Z t Y t d ⟨ Z , Y ⟩ t + R t 3 Z t 2 d ⟨ Y ⟩ t ) = R t Z t ( α R t Z t d W t 11 − β R t Y t d W t 12 ) + R t Y t ( α R t Y t d W t 11 + β R t Z t d W t 12 ) + 2 R t 1 [ R t 2 Y t 2 ( α R t 2 Z t 2 + β R t 2 Y t 2 ) + R t 2 Z t 2 ( α R t 2 Y t 2 + β R t 2 Z t 2 ) − 2 R t 2 Y t 2 R t 2 Z t 2 ( α − β ) ] d t = α d W t 11 + 2 R t β d t .
Define the scaled process ρ s : = R s / α \rho_s:=R_{s/\alpha} ρ s := R s / α . By the Brownian scaling property, W ~ s : = α W s / α 11 \widetilde{W}_s:=\sqrt{\alpha}W_{s/\alpha}^{11} W s := α W s / α 11 is a standard 1-dimensional Brownian motion. Then we have
d ρ s = d R s / α = α d W s / α 11 + β 2 R s / α d ( s α ) = d W ~ s + β / α 2 ρ s d s . d\rho_s=dR_{s/\alpha}=\sqrt{\alpha}dW_{s/\alpha}^{11}+\frac{\beta}{2R_{s/\alpha}}d\left(\frac{s}{\alpha}\right)=d\widetilde{W}_s+\frac{\beta/\alpha}{2\rho_s}ds. d ρ s = d R s / α = α d W s / α 11 + 2 R s / α β d ( α s ) = d W s + 2 ρ s β / α d s .
Thus, ρ s \rho_s ρ s is a ( 1 + β α ) (1+\frac{\beta}{\alpha}) ( 1 + α β ) -dimensional Bessel process. Since β < α \beta<\alpha β < α , the dimension 1 + β α < 2 1+\frac{\beta}{\alpha}<2 1 + α β < 2 and hence P θ ∗ ( τ ∗ < ∞ ) = 1 P_{\theta^*}(\tau^*<\infty)=1 P θ ∗ ( τ ∗ < ∞ ) = 1 . Consequently, we have c ( τ ∗ < ∞ ) = 1 c(\tau^*<\infty)=1 c ( τ ∗ < ∞ ) = 1 , i.e., a spectral collision for X X X occurs q.s.
A concluding remark
3.3. A concluding remark. The analysis in this section shows that both Lévy’s characterization and the GOE property may fail under Assumption 1 ′ 1' 1 ′ . Moreover, if the generator G G G loses the diagonal structure, as in the example constructed in § 3.2 , then the corresponding G G G -Dyson Brownian motion may exhibit collisions q.s. These observations motivate our choice of the diagonal structure in Assumption 1 for the underlying G G G -Brownian motion when investigating the spectral behavior of the generalized Wishart process.
Spectral non-collision property for generalized G G G -Wishart processes
In this section, we prove that under Assumption 1, the eigenvalue processes of the generalized G G G -Wishart process (3) do not collide q.s. and the G G G -SDE (7) admits a unique strong solution. For G G G -Dyson Brownian motion, we obtain the comparison principle and further establish the existence and uniqueness of a strong solution to the G G G -SDE (9) satisfied by the eigenvalue processes even when the initial eigenvalues are allowed to coincide. For G G G -Wishart process, we derive the G G G -SDE (11) satisfied by an N × N N \times N N × N G G G -Wishart process and establish the existence and uniqueness of a positive-definite solution to the G G G -SDE (12) satisfied by G G G -Wishart process with parameter α \alpha α when α ≥ N + 1 \alpha\ge N+1 α ≥ N + 1 .
We will work on the extended G G G -expectation space ( Ω , L ∗ 1 ( Ω ) , ( L ∗ 1 ( Ω t ) ) t ≥ 0 , ( E ^ t ) t ≥ 0 ) (\Omega, L^1_*(\Omega), (L^1_*(\Omega_t))_{t\ge0},(\widehat{\mathbb{E}}_t)_{t\ge0}) ( Ω , L ∗ 1 ( Ω ) , ( L ∗ 1 ( Ω t ) ) t ≥ 0 , ( E t ) t ≥ 0 ) , throughout the rest of the paper. This allows us to employ the Itô’s formula for C 2 C^2 C 2 functions. See § 2.4 for details.
Non-collision for generalized Wishart process
In this subsection, we study the eigenvalue process (7) of the generalized Wishart process X N X^N X N given by (3) .
Recall that the functions g N , h N , b N : R → R g_N,h_N,b_N:\mathbb{R}\to\mathbb{R} g N , h N , b N : R → R act on the spectrum of X N X^N X N . That is, f ( X N ) = U f ( Λ N ) U ⊤ f(X^N)=Uf(\Lambda^N)U^\top f ( X N ) = U f ( Λ N ) U ⊤ for f = g N , h N f=g_N,h_N f = g N , h N and b N b_N b N , where X N = U Λ N U ⊤ X^N=U\Lambda^NU^\top X N = U Λ N U ⊤ , with Λ N \Lambda^N Λ N and U U U defined as in Remark 3.1. It is clear that we have
∑ 1 ≤ k , h ≤ N U k i f ( X N ) k h U h j = f ( λ i N ) I { i = j } and ( f ( X N ) ) n = f n ( X N ) for any n ∈ N . (47) \sum_{1\le k,h\le N} U_{ki}f(X^N)_{kh}U_{hj}=f(\lambda_i^N)\mathrm{I}_{\{i=j\}}\quad\text{and}\quad(f(X^N))^n=f^n(X^N)\quad\text{for any }n\in\mathbb{N}. \tag*{(47)} 1 ≤ k , h ≤ N ∑ U k i f ( X N ) k h U hj = f ( λ i N ) I { i = j } and ( f ( X N ) ) n = f n ( X N ) for any n ∈ N . (47)
Proof of Theorem 1.1. Applying Itô’s formula (Proposition 2.21) to Φ i ( X N ( t ) ) \Phi_i(X^N(t)) Φ i ( X N ( t )) together with (40) , we get for t < τ t<\tau t < τ ,
d λ i N ( t ) = d Φ i ( X N ( t ) ) = ∑ 1 ≤ k ≤ h ≤ N ∂ Φ i ( X N ) ∂ x k h d X k h N + 1 2 ∑ 1 ≤ k ≤ h ≤ N , 1 ≤ k ′ ≤ h ′ ≤ N ∂ 2 Φ i ( X N ) ∂ x k h ∂ x k ′ h ′ d ⟨ X k h N , X k ′ h ′ N ⟩ t = ∑ 1 ≤ k ≤ h ≤ N ( 2 U k i U h i I [ k ≠ h ] + U k i 2 I [ k = h ] ) ( ∑ p , q g N k p d B t p q h N q h + ∑ p , q h N k p d B t q p g N q h + b N k h d ⟨ B h h ⟩ t ) + 1 2 ∑ 1 ≤ k ≤ h ≤ N , 1 ≤ k ′ < h ′ ≤ N ∂ 2 Φ i ( X N ) ∂ x k h ∂ x k ′ h ′ d ⟨ X k h N , X k ′ h ′ N ⟩ t = d M i + ∑ k , h U k i b N k h U h i d ⟨ B h h ⟩ t + ∑ j : j ≠ i 1 λ i N − λ j N ∑ k , k ′ , h , h ′ U k i U h j U k ′ i U h ′ j d ⟨ X k h N , X k ′ h ′ N ⟩ t , \begin{aligned}
d\lambda_i^N(t)=d\Phi_i(X^N(t)) \\
&=\sum_{1\le k\le h\le N}\frac{\partial\Phi_i(X^N)}{\partial x_{kh}}dX^N_{kh}+\frac{1}{2}\sum_{\substack{1\le k\le h\le N,\\1\le k'\le h'\le N}}\frac{\partial^2\Phi_i(X^N)}{\partial x_{kh}\partial x_{k'h'}}d\langle X^N_{kh},X^N_{k'h'}\rangle_t \\
&=\sum_{1\le k\le h\le N}(2U_{ki}U_{hi}\mathrm{I}_{[k\ne h]}+U_{ki}^2\mathrm{I}_{[k=h]})\left(\sum_{p,q}g_N^{kp}dB_t^{pq}h_N^{qh}+\sum_{p,q}h_N^{kp}dB_t^{qp}g_N^{qh}+b_N^{kh}d\langle B^{hh}\rangle_t\right) \\
&+\frac{1}{2}\sum_{\substack{1\le k\le h\le N,\\1\le k'<h'\le N}}\frac{\partial^2\Phi_i(X^N)}{\partial x_{kh}\partial x_{k'h'}}d\langle X^N_{kh},X^N_{k'h'}\rangle_t \\
&=dM_i+\sum_{k,h}U_{ki}b_N^{kh}U_{hi}d\langle B^{hh}\rangle_t+\sum_{j:j\ne i}\frac{1}{\lambda_i^N-\lambda_j^N}\sum_{k,k',h,h'}U_{ki}U_{hj}U_{k'i}U_{h'j}d\langle X^N_{kh},X^N_{k'h'}\rangle_t,
\end{aligned} d λ i N ( t ) = d Φ i ( X N ( t )) = 1 ≤ k ≤ h ≤ N ∑ ∂ x k h ∂ Φ i ( X N ) d X k h N + 2 1 1 ≤ k ≤ h ≤ N , 1 ≤ k ′ ≤ h ′ ≤ N ∑ ∂ x k h ∂ x k ′ h ′ ∂ 2 Φ i ( X N ) d ⟨ X k h N , X k ′ h ′ N ⟩ t = 1 ≤ k ≤ h ≤ N ∑ ( 2 U k i U hi I [ k = h ] + U k i 2 I [ k = h ] ) ( p , q ∑ g N k p d B t pq h N q h + p , q ∑ h N k p d B t q p g N q h + b N k h d ⟨ B hh ⟩ t ) + 2 1 1 ≤ k ≤ h ≤ N , 1 ≤ k ′ < h ′ ≤ N ∑ ∂ x k h ∂ x k ′ h ′ ∂ 2 Φ i ( X N ) d ⟨ X k h N , X k ′ h ′ N ⟩ t = d M i + k , h ∑ U k i b N k h U hi d ⟨ B hh ⟩ t + j : j = i ∑ λ i N − λ j N 1 k , k ′ , h , h ′ ∑ U k i U hj U k ′ i U h ′ j d ⟨ X k h N , X k ′ h ′ N ⟩ t ,
where
d M i = ∑ k , h U k i ( ∑ p , q g N k p d B t p q h N q h + ∑ p , q h N k p d B t q p g N q h ) U h i . (48) dM_i=\sum_{k,h}U_{ki}\left(\sum_{p,q}g_N^{kp}dB_t^{pq}h_N^{qh}+\sum_{p,q}h_N^{kp}dB_t^{qp}g_N^{qh}\right)U_{hi}. \tag*{(48)} d M i = k , h ∑ U k i ( p , q ∑ g N k p d B t pq h N q h + p , q ∑ h N k p d B t q p g N q h ) U hi . (48)
Now we calculate the quadratic co-variation of X N X^N X N . Noting that
d X k h N = ∑ p , q g N k p d B t p q g N q h + ∑ p , q h N k p d B t q p g N q h + b N k h d ⟨ B h h ⟩ t , dX^N_{kh}=\sum_{p,q}g_N^{kp}dB_t^{pq}g_N^{qh}+\sum_{p,q}h_N^{kp}dB_t^{qp}g_N^{qh}+b_N^{kh}d\langle B_{hh}\rangle_t, d X k h N = p , q ∑ g N k p d B t pq g N q h + p , q ∑ h N k p d B t q p g N q h + b N k h d ⟨ B hh ⟩ t ,
and similarly for d X k ′ h ′ N \mathrm{d}X^{N}_{k'h'} d X k ′ h ′ N , we get
d ⟨ X k h N , X k ′ h ′ N ⟩ t = ∑ p , q ( g N k p g N q h g N k ′ p g N q ′ h ′ + g N k p g N q h g N q ′ p h N k ′ h ′ ) d ⟨ B p q ⟩ t + ∑ p , q ( h N k p g N q h g N k ′ q h N p ′ h ′ + h N k p g N q h h N k ′ p g N q ′ h ′ ) d ⟨ B p q ⟩ t = [ ( g N 2 ) k k ′ ( h N 2 ) h h ′ + ( g N 2 ) k h ′ ( h N 2 ) h k ′ + ( g N 2 ) h k ′ ( h N 2 ) k h ′ + ( g N 2 ) h h ′ ( h N 2 ) k k ′ ] d ⟨ B 11 ⟩ t . \mathrm{d}\langle X^{N}_{kh},X^{N}_{k'h'}\rangle_t
= \sum_{p,q}\left(g^{kp}_{N}g^{qh}_{N}g^{k'p}_{N}g^{q'h'}_{N}+g^{kp}_{N}g^{qh}_{N}g^{q'p}_{N}h^{k'h'}_{N}\right)\mathrm{d}\langle B^{pq}\rangle_t+\sum_{p,q}\left(h^{kp}_{N}g^{qh}_{N}g^{k'q}_{N}h^{p'h'}_{N}+h^{kp}_{N}g^{qh}_{N}h^{k'p}_{N}g^{q'h'}_{N}\right)\mathrm{d}\langle B^{pq}\rangle_t
= \left[(g^{2}_{N})_{kk'}(h^{2}_{N})_{hh'}+(g^{2}_{N})_{kh'}(h^{2}_{N})_{hk'}+(g^{2}_{N})_{hk'}(h^{2}_{N})_{kh'}+(g^{2}_{N})_{hh'}(h^{2}_{N})_{kk'}\right]\mathrm{d}\langle B^{11}\rangle_t. d ⟨ X k h N , X k ′ h ′ N ⟩ t = p , q ∑ ( g N k p g N q h g N k ′ p g N q ′ h ′ + g N k p g N q h g N q ′ p h N k ′ h ′ ) d ⟨ B pq ⟩ t + p , q ∑ ( h N k p g N q h g N k ′ q h N p ′ h ′ + h N k p g N q h h N k ′ p g N q ′ h ′ ) d ⟨ B pq ⟩ t = [ ( g N 2 ) k k ′ ( h N 2 ) h h ′ + ( g N 2 ) k h ′ ( h N 2 ) h k ′ + ( g N 2 ) h k ′ ( h N 2 ) k h ′ + ( g N 2 ) h h ′ ( h N 2 ) k k ′ ] d ⟨ B 11 ⟩ t .
Using the above identity together with (47) , we get for t < τ t<\tau t < τ ,
d λ i N ( t ) = d M i + b N ( λ i N ) d ⟨ B h h ⟩ t + ∑ j : j ≠ i 1 λ i N − λ j N [ g N 2 ( λ i N ) h N 2 ( λ j N ) + g N 2 ( λ j N ) h N 2 ( λ i N ) ] d ⟨ B 11 ⟩ t = d M i + b N ( λ i N ) d ⟨ B 11 ⟩ t + ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N d ⟨ B 11 ⟩ t . \mathrm{d}\lambda^{N}_{i}(t)=\mathrm{d}M_i+b_N(\lambda^{N}_{i})\mathrm{d}\langle B_{hh}\rangle_t
+\sum_{j:j\ne i}\frac{1}{\lambda^{N}_{i}-\lambda^{N}_{j}}\left[g^{2}_{N}(\lambda^{N}_{i})h^{2}_{N}(\lambda^{N}_{j})+g^{2}_{N}(\lambda^{N}_{j})h^{2}_{N}(\lambda^{N}_{i})\right]\mathrm{d}\langle B^{11}\rangle_t
=\mathrm{d}M_i+b_N(\lambda^{N}_{i})\mathrm{d}\langle B^{11}\rangle_t+\sum_{j:j\ne i}\frac{H_N(\lambda^{N}_{i},\lambda^{N}_{j})}{\lambda^{N}_{i}-\lambda^{N}_{j}}\mathrm{d}\langle B^{11}\rangle_t. d λ i N ( t ) = d M i + b N ( λ i N ) d ⟨ B hh ⟩ t + j : j = i ∑ λ i N − λ j N 1 [ g N 2 ( λ i N ) h N 2 ( λ j N ) + g N 2 ( λ j N ) h N 2 ( λ i N ) ] d ⟨ B 11 ⟩ t = d M i + b N ( λ i N ) d ⟨ B 11 ⟩ t + j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) d ⟨ B 11 ⟩ t .
Furthermore, d M i \mathrm{d}M_i d M i given by (48) can be rewritten as
d M i = ( U T g N ( X N ) d B t h N ( X N ) U ) i i + ( U T h N ( X N ) d B t T g N ( X N ) U ) i i = ( U T g N ( X N ) U U T d B t U U T h N ( X N ) U ) i i + ( U T h N ( X N ) U U T d B t T U U T g N ( X N ) U ) i i = g N ( λ i N ) h N ( λ i N ) ( U T d B t U ) i i + g N ( λ i N ) h N ( λ i N ) ( U T d B t T U ) i i = g N ( λ i N ) h N ( λ i N ) ∑ k , h = 1 N U i k T ( d B t k h + d B t h k ) U h i . \mathrm{d}M_i=(U^{\mathsf T}g_N(X^N)\mathrm{d}B_t h_N(X^N)U)_{ii}+(U^{\mathsf T}h_N(X^N)\mathrm{d}B_t^{\mathsf T}g_N(X^N)U)_{ii}
=(U^{\mathsf T}g_N(X^N)UU^{\mathsf T}\mathrm{d}B_tUU^{\mathsf T}h_N(X^N)U)_{ii}+(U^{\mathsf T}h_N(X^N)UU^{\mathsf T}\mathrm{d}B_t^{\mathsf T}UU^{\mathsf T}g_N(X^N)U)_{ii}
=g_N(\lambda^{N}_{i})h_N(\lambda^{N}_{i})(U^{\mathsf T}\mathrm{d}B_tU)_{ii}+g_N(\lambda^{N}_{i})h_N(\lambda^{N}_{i})(U^{\mathsf T}\mathrm{d}B_t^{\mathsf T}U)_{ii}
=g_N(\lambda^{N}_{i})h_N(\lambda^{N}_{i})\sum_{k,h=1}^{N}U^{\mathsf T}_{ik}(\mathrm{d}B^{kh}_t+\mathrm{d}B^{hk}_t)U_{hi}. d M i = ( U T g N ( X N ) d B t h N ( X N ) U ) ii + ( U T h N ( X N ) d B t T g N ( X N ) U ) ii = ( U T g N ( X N ) U U T d B t U U T h N ( X N ) U ) ii + ( U T h N ( X N ) U U T d B t T U U T g N ( X N ) U ) ii = g N ( λ i N ) h N ( λ i N ) ( U T d B t U ) ii + g N ( λ i N ) h N ( λ i N ) ( U T d B t T U ) ii = g N ( λ i N ) h N ( λ i N ) k , h = 1 ∑ N U ik T ( d B t k h + d B t hk ) U hi .
Denote for i = 1 , … , N i=1,\ldots,N i = 1 , … , N ,
d B ‾ i ( t ) : = 1 2 ∑ k , h = 1 N U i k T ( d B t k h + d B t h k ) U h i . (49) \mathrm{d}\overline{B}_i(t):=\frac{1}{2}\sum_{k,h=1}^{N}U^{\mathsf T}_{ik}(\mathrm{d}B^{kh}_t+\mathrm{d}B^{hk}_t)U_{hi}. \tag*{(49)} d B i ( t ) := 2 1 k , h = 1 ∑ N U ik T ( d B t k h + d B t hk ) U hi . (49)
Then, we have
d ⟨ B ‾ i , B ‾ j ⟩ t = 1 4 ∑ k , h = 1 N U i k T U j k T U h i U h j d ⟨ B k h ⟩ t + 1 4 ∑ k , h = 1 N U i k T U j h T U h i U k j d ⟨ B h k ⟩ t + 1 4 ∑ k , h = 1 N U i h T U j k T U h i U k j d ⟨ B k h ⟩ t + 1 4 ∑ k , h = 1 N U i h T U j h T U h i U h j d ⟨ B h k ⟩ t = 1 4 ( U T U ) i j 2 d ⟨ B 11 ⟩ t + 1 4 ( U T U ) i j ( U T U ) j i d ⟨ B 11 ⟩ t + 1 4 ( U T U ) i j ( U T U ) j i d ⟨ B 11 ⟩ t + 1 4 ( U T U ) i j 2 d ⟨ B 11 ⟩ t = 1 { i = j } d ⟨ B 11 ⟩ t . (50) \begin{aligned}
\mathrm{d}\langle\overline{B}_i,\overline{B}_j\rangle_t
&=\frac{1}{4}\sum_{k,h=1}^{N}U^{\mathsf T}_{ik}U^{\mathsf T}_{jk}U_{hi}U_{hj}\mathrm{d}\langle B^{kh}\rangle_t+\frac{1}{4}\sum_{k,h=1}^{N}U^{\mathsf T}_{ik}U^{\mathsf T}_{jh}U_{hi}U_{kj}\mathrm{d}\langle B^{hk}\rangle_t\\
&\quad+\frac{1}{4}\sum_{k,h=1}^{N}U^{\mathsf T}_{ih}U^{\mathsf T}_{jk}U_{hi}U_{kj}\mathrm{d}\langle B^{kh}\rangle_t+\frac{1}{4}\sum_{k,h=1}^{N}U^{\mathsf T}_{ih}U^{\mathsf T}_{jh}U_{hi}U_{hj}\mathrm{d}\langle B^{hk}\rangle_t\\
&=\frac{1}{4}(U^{\mathsf T}U)_{ij}^{2}\mathrm{d}\langle B^{11}\rangle_t+\frac{1}{4}(U^{\mathsf T}U)_{ij}(U^{\mathsf T}U)_{ji}\mathrm{d}\langle B^{11}\rangle_t\\
&\quad+\frac{1}{4}(U^{\mathsf T}U)_{ij}(U^{\mathsf T}U)_{ji}\mathrm{d}\langle B^{11}\rangle_t+\frac{1}{4}(U^{\mathsf T}U)_{ij}^{2}\mathrm{d}\langle B^{11}\rangle_t\\
&=1_{\{i=j\}}\mathrm{d}\langle B^{11}\rangle_t. \tag*{(50)}
\end{aligned} d ⟨ B i , B j ⟩ t = 4 1 k , h = 1 ∑ N U ik T U j k T U hi U hj d ⟨ B k h ⟩ t + 4 1 k , h = 1 ∑ N U ik T U j h T U hi U k j d ⟨ B hk ⟩ t + 4 1 k , h = 1 ∑ N U ih T U j k T U hi U k j d ⟨ B k h ⟩ t + 4 1 k , h = 1 ∑ N U ih T U j h T U hi U hj d ⟨ B hk ⟩ t = 4 1 ( U T U ) ij 2 d ⟨ B 11 ⟩ t + 4 1 ( U T U ) ij ( U T U ) j i d ⟨ B 11 ⟩ t + 4 1 ( U T U ) ij ( U T U ) j i d ⟨ B 11 ⟩ t + 4 1 ( U T U ) ij 2 d ⟨ B 11 ⟩ t = 1 { i = j } d ⟨ B 11 ⟩ t . (50)
Consequently, we get that for t < τ t<\tau t < τ , and i = 1 , 2 , … , N i=1,2,\ldots,N i = 1 , 2 , … , N ,
d λ i N = 2 g N ( λ i N ) h N ( λ i N ) d B ‾ i + b N ( λ i N ) d ⟨ B ‾ i ⟩ t + ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N d ⟨ B ‾ i ⟩ t . \mathrm{d}\lambda^{N}_{i}=2g_N(\lambda^{N}_{i})h_N(\lambda^{N}_{i})\mathrm{d}\overline{B}_i+b_N(\lambda^{N}_{i})\mathrm{d}\langle\overline{B}_i\rangle_t+\sum_{j:j\ne i}\frac{H_N(\lambda^{N}_{i},\lambda^{N}_{j})}{\lambda^{N}_{i}-\lambda^{N}_{j}}\mathrm{d}\langle\overline{B}_i\rangle_t. d λ i N = 2 g N ( λ i N ) h N ( λ i N ) d B i + b N ( λ i N ) d ⟨ B i ⟩ t + j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) d ⟨ B i ⟩ t .
Now we prove that the first collision time τ \tau τ is infinite q.s. We first introduce some auxiliary functions.
We consider the following Lyapunov function, which is inspired by [1 ] ,
f ( x ) = f ( x 1 , … , x N ) = 1 N ∑ i = 1 N x i 2 − 1 N 2 ∑ i ≠ j log ∣ x i − x j ∣ , f(x)=f(x_1,\ldots,x_N)=\frac{1}{N}\sum_{i=1}^{N}x_i^2-\frac{1}{N^2}\sum_{i\ne j}\log|x_i-x_j|, f ( x ) = f ( x 1 , … , x N ) = N 1 i = 1 ∑ N x i 2 − N 2 1 i = j ∑ log ∣ x i − x j ∣ ,
defined for x = ( x 1 , … , x N ) ∈ Δ N x=(x_1,\ldots,x_N)\in\Delta_N x = ( x 1 , … , x N ) ∈ Δ N . It is not hard to verify the following property of f f f , for i ≠ j i\ne j i = j ,
f ( x 1 , … , x N ) ≥ − 4 , − 1 N 2 log ∣ x i − x j ∣ ≤ f ( x ) + 4 , 1 N ∑ k = 1 N x k 2 ≤ 2 f ( x ) + 4. (51) f(x_1,\ldots,x_N)\ge-4,\qquad-\frac{1}{N^2}\log|x_i-x_j|\le f(x)+4,\qquad\frac{1}{N}\sum_{k=1}^{N}x_k^2\le2f(x)+4. \tag*{(51)} f ( x 1 , … , x N ) ≥ − 4 , − N 2 1 log ∣ x i − x j ∣ ≤ f ( x ) + 4 , N 1 k = 1 ∑ N x k 2 ≤ 2 f ( x ) + 4. (51)
For any M > 0 M>0 M > 0 , set R : = R ( N , M ) = e N 2 ( 4 + M ) R:=R(N,M)=e^{N^2(4+M)} R := R ( N , M ) = e N 2 ( 4 + M ) and define the stopping times:
τ R : = inf { t > 0 : min i ≠ j ∣ λ i N ( t ) − λ j N ( t ) ∣ < R − 1 } and T M : = inf { t ≥ 0 : f ( λ N ( t ) ) ≥ M } . (52) \tau_R:=\inf\left\{t>0:\min_{i\ne j}|\lambda_i^N(t)-\lambda_j^N(t)|<R^{-1}\right\}\quad\text{and}\quad T_M:=\inf\left\{t\ge0:f(\lambda^N(t))\ge M\right\}. \tag*{(52)} τ R := inf { t > 0 : i = j min ∣ λ i N ( t ) − λ j N ( t ) ∣ < R − 1 } and T M := inf { t ≥ 0 : f ( λ N ( t )) ≥ M } . (52)
Then, by (51) we have for any i ≠ j i\ne j i = j ,
log ∣ λ i N ( t ) − λ j N ( t ) ∣ ≥ − N 2 ( f ( λ N ( t ) ) + 4 ) ≥ − N 2 ( M + 4 ) , for t ≤ T M . \log|\lambda_i^N(t)-\lambda_j^N(t)|\ge-N^2(f(\lambda^N(t))+4)\ge-N^2(M+4),\quad\text{for }t\le T_M. log ∣ λ i N ( t ) − λ j N ( t ) ∣ ≥ − N 2 ( f ( λ N ( t )) + 4 ) ≥ − N 2 ( M + 4 ) , for t ≤ T M .
This implies that T M ≤ τ R T_M\le\tau_R T M ≤ τ R . Note that when t ≤ T M t\le T_M t ≤ T M , λ i N ( t ) \lambda_i^N(t) λ i N ( t ) is bounded by (51) , and hence g N ( λ i N ) g_N(\lambda_i^N) g N ( λ i N ) , h N ( λ i N ) h_N(\lambda_i^N) h N ( λ i N ) , b N ( λ i N ) b_N(\lambda_i^N) b N ( λ i N ) and H N ( λ i N , λ j N ) H_N(\lambda_i^N,\lambda_j^N) H N ( λ i N , λ j N ) are also bounded.
For i = 1 , … , N i=1,\ldots,N i = 1 , … , N , define the functions
u i , 1 ( x ) = ∑ k : k ≠ i 1 x i − x k , u i , 2 ( x ) = ∑ k : k ≠ i 1 ( x i − x k ) 2 , x ∈ Δ N . u_{i,1}(x)=\sum_{k:k\ne i}\frac{1}{x_i-x_k},\qquad u_{i,2}(x)=\sum_{k:k\ne i}\frac{1}{(x_i-x_k)^2},\qquad x\in\Delta_N. u i , 1 ( x ) = k : k = i ∑ x i − x k 1 , u i , 2 ( x ) = k : k = i ∑ ( x i − x k ) 2 1 , x ∈ Δ N .
Note that for x ∈ Δ N x\in\Delta_N x ∈ Δ N ,
A : = ∑ i ≠ k , i ≠ l , k ≠ l 1 ( x i − x l ) ( x i − x k ) = ∑ i ≠ k , i ≠ l , k ≠ l 1 x l − x k ( 1 x i − x l − 1 x i − x k ) = − 2 A , A:=\sum_{\substack{i\ne k,i\ne l,\\ k\ne l}}\frac{1}{(x_i-x_l)(x_i-x_k)}
=\sum_{\substack{i\ne k,i\ne l,\\ k\ne l}}\frac{1}{x_l-x_k}\left(\frac{1}{x_i-x_l}-\frac{1}{x_i-x_k}\right)=-2A, A := i = k , i = l , k = l ∑ ( x i − x l ) ( x i − x k ) 1 = i = k , i = l , k = l ∑ x l − x k 1 ( x i − x l 1 − x i − x k 1 ) = − 2 A ,
and hence A = 0 A=0 A = 0 . Thus, we get
∑ i = 1 N ( u i , 1 ( x ) 2 − u i , 2 ( x ) ) = 0. \sum_{i=1}^{N}\left(u_{i,1}(x)^2-u_{i,2}(x)\right)=0. i = 1 ∑ N ( u i , 1 ( x ) 2 − u i , 2 ( x ) ) = 0.
Similarly, noting that
B : = ∑ i = 1 N ∑ k ≠ i x i x i − x k = ∑ i = 1 N ∑ k ≠ i ( 1 + x k x i − x k ) = N ( N − 1 ) − B , B:=\sum_{i=1}^{N}\sum_{k\ne i}\frac{x_i}{x_i-x_k}
=\sum_{i=1}^{N}\sum_{k\ne i}\left(1+\frac{x_k}{x_i-x_k}\right)=N(N-1)-B, B := i = 1 ∑ N k = i ∑ x i − x k x i = i = 1 ∑ N k = i ∑ ( 1 + x i − x k x k ) = N ( N − 1 ) − B ,
we get
∑ i = 1 N u i , 1 ( x ) x i = N ( N − 1 ) 2 . \sum_{i=1}^{N}u_{i,1}(x)x_i=\frac{N(N-1)}{2}. i = 1 ∑ N u i , 1 ( x ) x i = 2 N ( N − 1 ) .
Noting that
∂ ∂ x i f ( x ) = 2 N x i − 1 N 2 u i , 1 ( x ) , ∂ 2 ∂ x i 2 f ( x ) = 2 N + 1 N 2 u i , 2 ( x ) , \frac{\partial}{\partial x_i}f(x)=\frac{2}{N}x_i-\frac{1}{N^2}u_{i,1}(x),\qquad\frac{\partial^2}{\partial x_i^2}f(x)=\frac{2}{N}+\frac{1}{N^2}u_{i,2}(x), ∂ x i ∂ f ( x ) = N 2 x i − N 2 1 u i , 1 ( x ) , ∂ x i 2 ∂ 2 f ( x ) = N 2 + N 2 1 u i , 2 ( x ) ,
we get for t ≤ T M t\le T_M t ≤ T M ,
d f ( λ N ( t ) ) = ∑ i = 1 N 2 N λ i N ( t ) ( ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N d ⟨ B ˉ i ⟩ t + b N ( λ i N ) d ⟨ B ˉ i ⟩ t ) \mathrm{d}f(\lambda^N(t))=\sum_{i=1}^{N}\frac{2}{N}\lambda_i^N(t)\left(\sum_{j:j\ne i}\frac{H_N(\lambda_i^N,\lambda_j^N)}{\lambda_i^N-\lambda_j^N}\mathrm{d}\langle\bar B_i\rangle_t+b_N(\lambda_i^N)\mathrm{d}\langle\bar B_i\rangle_t\right) d f ( λ N ( t )) = i = 1 ∑ N N 2 λ i N ( t ) j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) d ⟨ B ˉ i ⟩ t + b N ( λ i N ) d ⟨ B ˉ i ⟩ t
− ∑ i = 1 N 1 N 2 u i , 1 ( λ N ( t ) ) ( ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N d ⟨ B ˉ i ⟩ t + b N ( λ i N ) d ⟨ B ˉ i ⟩ t ) + 1 2 ∑ i = 1 N ( 2 N + 1 N 2 u i , 2 ( λ N ( t ) ) ) g 2 N 2 ( λ i N ) h N 2 ( λ i N ) d ⟨ B ˉ i ⟩ t + d M ~ ( t ) , \begin{aligned}
-\sum_{i=1}^{N}\frac{1}{N^{2}}u_{i,1}\left(\lambda^{N}(t)\right)\left(\sum_{j:j\ne i}\frac{H_{N}\left(\lambda_{i}^{N},\lambda_{j}^{N}\right)}{\lambda_{i}^{N}-\lambda_{j}^{N}}\mathrm{d}\langle\bar{B}_{i}\rangle_{t}+b_{N}\left(\lambda_{i}^{N}\right)\mathrm{d}\langle\bar{B}_{i}\rangle_{t}\right) \\
&+\frac{1}{2}\sum_{i=1}^{N}\left(\frac{2}{N}+\frac{1}{N^{2}}u_{i,2}\left(\lambda^{N}(t)\right)\right)g_{2N}^{2}\left(\lambda_{i}^{N}\right)h_{N}^{2}\left(\lambda_{i}^{N}\right)\mathrm{d}\langle\bar{B}_{i}\rangle_{t}+\mathrm{d}\widetilde{M}(t),
\end{aligned} − i = 1 ∑ N N 2 1 u i , 1 ( λ N ( t ) ) j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) d ⟨ B ˉ i ⟩ t + b N ( λ i N ) d ⟨ B ˉ i ⟩ t + 2 1 i = 1 ∑ N ( N 2 + N 2 1 u i , 2 ( λ N ( t ) ) ) g 2 N 2 ( λ i N ) h N 2 ( λ i N ) d ⟨ B ˉ i ⟩ t + d M ( t ) ,
where
d M ~ ( t ) = ∑ i = 1 N ( 2 N λ i N ( t ) − 1 N 2 u i , 1 ( λ N ( t ) ) ) 2 g N ( λ i N ) h N ( λ i N ) d B ˉ i , \mathrm{d}\widetilde{M}(t)=\sum_{i=1}^{N}\left(\frac{2}{N}\lambda_{i}^{N}(t)-\frac{1}{N^{2}}u_{i,1}\left(\lambda^{N}(t)\right)\right)2g_{N}\left(\lambda_{i}^{N}\right)h_{N}\left(\lambda_{i}^{N}\right)\mathrm{d}\bar{B}_{i}, d M ( t ) = i = 1 ∑ N ( N 2 λ i N ( t ) − N 2 1 u i , 1 ( λ N ( t ) ) ) 2 g N ( λ i N ) h N ( λ i N ) d B ˉ i ,
for which we have
E ^ [ ∫ 0 t ∧ T M d M ~ ( s ) ] = E ^ [ ∫ 0 t ∧ T M ∑ i = 1 N ( 2 N λ i N ( s ) − 1 N 2 u i , 1 ( λ N ( s ) ) ) 2 g N ( λ i N ) h N ( λ i N ) d B ˉ i ] = 0. (53) \widehat{\mathbb{E}}\left[\int_{0}^{t\wedge T_{M}}\mathrm{d}\widetilde{M}(s)\right]=\widehat{\mathbb{E}}\left[\int_{0}^{t\wedge T_{M}}\sum_{i=1}^{N}\left(\frac{2}{N}\lambda_{i}^{N}(s)-\frac{1}{N^{2}}u_{i,1}\left(\lambda^{N}(s)\right)\right)2g_{N}\left(\lambda_{i}^{N}\right)h_{N}\left(\lambda_{i}^{N}\right)\mathrm{d}\bar{B}_{i}\right]=0. \tag*{(53)} E [ ∫ 0 t ∧ T M d M ( s ) ] = E [ ∫ 0 t ∧ T M i = 1 ∑ N ( N 2 λ i N ( s ) − N 2 1 u i , 1 ( λ N ( s ) ) ) 2 g N ( λ i N ) h N ( λ i N ) d B ˉ i ] = 0. (53)
We write d f ( λ N ( t ) ) \mathrm{d}f\left(\lambda^{N}(t)\right) d f ( λ N ( t ) ) as follows
d f ( λ N ( t ) ) = [ B ( λ N ( t ) ) + ∑ i = 1 3 S i ( λ N ( t ) ) ] d ⟨ B 11 ⟩ t + d M ~ t (54) \mathrm{d}f\left(\lambda^{N}(t)\right)=\left[B\left(\lambda^{N}(t)\right)+\sum_{i=1}^{3}S_{i}\left(\lambda^{N}(t)\right)\right]\mathrm{d}\langle B^{11}\rangle_{t}+\mathrm{d}\widetilde{M}_{t} \tag*{(54)} d f ( λ N ( t ) ) = [ B ( λ N ( t ) ) + i = 1 ∑ 3 S i ( λ N ( t ) ) ] d ⟨ B 11 ⟩ t + d M t (54)
where
B ( λ N ( t ) ) = ∑ i = 1 N ( 2 N λ i N ( t ) − 1 N 2 u i , 1 ( λ N ( t ) ) ) b N ( λ i N ) ; B\left(\lambda^{N}(t)\right)=\sum_{i=1}^{N}\left(\frac{2}{N}\lambda_{i}^{N}(t)-\frac{1}{N^{2}}u_{i,1}\left(\lambda^{N}(t)\right)\right)b_{N}\left(\lambda_{i}^{N}\right); B ( λ N ( t ) ) = i = 1 ∑ N ( N 2 λ i N ( t ) − N 2 1 u i , 1 ( λ N ( t ) ) ) b N ( λ i N ) ;
S 1 ( λ N ( t ) ) = ∑ i = 1 N 2 N λ i N ( t ) ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N ; S_{1}\left(\lambda^{N}(t)\right)=\sum_{i=1}^{N}\frac{2}{N}\lambda_{i}^{N}(t)\sum_{j:j\ne i}\frac{H_{N}\left(\lambda_{i}^{N},\lambda_{j}^{N}\right)}{\lambda_{i}^{N}-\lambda_{j}^{N}}; S 1 ( λ N ( t ) ) = i = 1 ∑ N N 2 λ i N ( t ) j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) ;
S 2 ( λ N ( t ) ) = − ∑ i = 1 N 1 N 2 u i , 1 ( λ N ( t ) ) ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N + ∑ i = 1 N 2 N 2 u i , 2 ( λ N ( t ) ) g N 2 ( λ i N ) h N 2 ( λ i N ) ; S_{2}\left(\lambda^{N}(t)\right)=-\sum_{i=1}^{N}\frac{1}{N^{2}}u_{i,1}\left(\lambda^{N}(t)\right)\sum_{j:j\ne i}\frac{H_{N}\left(\lambda_{i}^{N},\lambda_{j}^{N}\right)}{\lambda_{i}^{N}-\lambda_{j}^{N}}+\sum_{i=1}^{N}\frac{2}{N^{2}}u_{i,2}\left(\lambda^{N}(t)\right)g_{N}^{2}\left(\lambda_{i}^{N}\right)h_{N}^{2}\left(\lambda_{i}^{N}\right); S 2 ( λ N ( t ) ) = − i = 1 ∑ N N 2 1 u i , 1 ( λ N ( t ) ) j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) + i = 1 ∑ N N 2 2 u i , 2 ( λ N ( t ) ) g N 2 ( λ i N ) h N 2 ( λ i N ) ;
S 3 ( λ N ( t ) ) = ∑ i = 1 N 4 N g N 2 ( λ i N ) h N 2 ( λ i N ) . S_{3}\left(\lambda^{N}(t)\right)=\sum_{i=1}^{N}\frac{4}{N}g_{N}^{2}\left(\lambda_{i}^{N}\right)h_{N}^{2}\left(\lambda_{i}^{N}\right). S 3 ( λ N ( t ) ) = i = 1 ∑ N N 4 g N 2 ( λ i N ) h N 2 ( λ i N ) .
We first estimate B ( λ N ( t ) ) B\left(\lambda^{N}(t)\right) B ( λ N ( t ) ) . Noting that
∑ i = 1 N u i , 1 ( λ N ( t ) ) b N ( λ i N ) = ∑ i ≠ k b N ( λ i N ) λ i N − λ k N = ∑ i < k b N ( λ i N ) − b N ( λ k N ) λ i N − λ k N , \sum_{i=1}^{N}u_{i,1}\left(\lambda^{N}(t)\right)b_{N}\left(\lambda_{i}^{N}\right)=\sum_{i\ne k}\frac{b_{N}\left(\lambda_{i}^{N}\right)}{\lambda_{i}^{N}-\lambda_{k}^{N}}=\sum_{i<k}\frac{b_{N}\left(\lambda_{i}^{N}\right)-b_{N}\left(\lambda_{k}^{N}\right)}{\lambda_{i}^{N}-\lambda_{k}^{N}}, i = 1 ∑ N u i , 1 ( λ N ( t ) ) b N ( λ i N ) = i = k ∑ λ i N − λ k N b N ( λ i N ) = i < k ∑ λ i N − λ k N b N ( λ i N ) − b N ( λ k N ) ,
we get
∣ ∑ i = 1 N u i , 1 ( λ N ( t ) ) b N ( λ i N ) ∣ ≤ N ( N − 1 ) 2 L b N , \left|\sum_{i=1}^{N}u_{i,1}\left(\lambda^{N}(t)\right)b_{N}\left(\lambda_{i}^{N}\right)\right|\le\frac{N(N-1)}{2}L_{b_{N}}, i = 1 ∑ N u i , 1 ( λ N ( t ) ) b N ( λ i N ) ≤ 2 N ( N − 1 ) L b N ,
where we use L f L_{f} L f to denote the Lipschitz constant for a Lipschitz-continuous function f f f . We also have
∣ ∑ i = 1 N λ i N ( t ) b N ( λ i N ) ∣ ≤ ∑ i = 1 N ∣ λ i N ( t ) ∣ max { L b N , b N ( 0 ) } ( λ i N ( t ) + 1 ) , \left|\sum_{i=1}^{N}\lambda_{i}^{N}(t)b_{N}\left(\lambda_{i}^{N}\right)\right|\le\sum_{i=1}^{N}\left|\lambda_{i}^{N}(t)\right|\max\{L_{b_{N}},b_{N}(0)\}\left(\lambda_{i}^{N}(t)+1\right), i = 1 ∑ N λ i N ( t ) b N ( λ i N ) ≤ i = 1 ∑ N λ i N ( t ) max { L b N , b N ( 0 )} ( λ i N ( t ) + 1 ) ,
and thus
∣ B ( λ N ( t ) ) ∣ ≤ C ∑ i = 1 N ( 1 + ( λ i N ( t ) ) 2 ) for some C > 0. (55) \lvert B(\lambda^N(t))\rvert\le C\sum_{i=1}^{N}\left(1+(\lambda_i^N(t))^2\right)\text{ for some }C>0. \tag*{(55)} ∣ B ( λ N ( t ))∣ ≤ C i = 1 ∑ N ( 1 + ( λ i N ( t ) ) 2 ) for some C > 0. (55)
For the term S 1 ( λ N ( t ) ) S_1(\lambda^N(t)) S 1 ( λ N ( t )) , noting that
∑ i = 1 N λ i N ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N = ∑ i < j H N ( λ i N , λ j N ) ( λ i N λ i N − λ j N + λ j N λ j N − λ i N ) = ∑ i < j H N ( λ i N , λ j N ) , \sum_{i=1}^{N}\lambda_i^N\sum_{j:j\ne i}\frac{H_N(\lambda_i^N,\lambda_j^N)}{\lambda_i^N-\lambda_j^N}
=\sum_{i<j}H_N(\lambda_i^N,\lambda_j^N)\left(\frac{\lambda_i^N}{\lambda_i^N-\lambda_j^N}+\frac{\lambda_j^N}{\lambda_j^N-\lambda_i^N}\right)
=\sum_{i<j}H_N(\lambda_i^N,\lambda_j^N), i = 1 ∑ N λ i N j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) = i < j ∑ H N ( λ i N , λ j N ) ( λ i N − λ j N λ i N + λ j N − λ i N λ j N ) = i < j ∑ H N ( λ i N , λ j N ) ,
and that H N ( λ i N , λ j N ) ≤ C ( 1 + ( λ i N ( t ) ) 2 + ( λ j N ( t ) ) 2 ) H_N(\lambda_i^N,\lambda_j^N)\le C(1+(\lambda_i^N(t))^2+(\lambda_j^N(t))^2) H N ( λ i N , λ j N ) ≤ C ( 1 + ( λ i N ( t ) ) 2 + ( λ j N ( t ) ) 2 ) for C > 0 C>0 C > 0 , we get
S 1 ( λ N ( t ) ) = 2 N ∑ i < j H N ( λ i N , λ j N ) ≤ C ∑ i = 1 N ( 1 + ( λ i N ( t ) ) 2 ) (56) S_1(\lambda^N(t))=\frac{2}{N}\sum_{i<j}H_N(\lambda_i^N,\lambda_j^N)\le C\sum_{i=1}^{N}\left(1+(\lambda_i^N(t))^2\right) \tag*{(56)} S 1 ( λ N ( t )) = N 2 i < j ∑ H N ( λ i N , λ j N ) ≤ C i = 1 ∑ N ( 1 + ( λ i N ( t ) ) 2 ) (56)
for some C > 0 C>0 C > 0 . For S 3 ( λ N ( t ) ) S_3(\lambda^N(t)) S 3 ( λ N ( t )) , we have the same estimation:
S 3 ( λ N ( t ) ) ≤ C ∑ i = 1 N ( 1 + ( λ i N ( t ) ) 2 ) . (57) S_3(\lambda^N(t))\le C\sum_{i=1}^{N}\left(1+(\lambda_i^N(t))^2\right). \tag*{(57)} S 3 ( λ N ( t )) ≤ C i = 1 ∑ N ( 1 + ( λ i N ( t ) ) 2 ) . (57)
Now, we estimate S 2 ( λ N ( t ) ) S_2(\lambda^N(t)) S 2 ( λ N ( t )) .
S 2 ( λ N ( t ) ) = − 1 N 2 ∑ i = 1 N ∑ k : k ≠ i 1 λ i N − λ k N ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N + 2 N 2 ∑ i = 1 N ∑ k : k ≠ i g N 2 ( λ i N ) h N 2 ( λ i N ) ( λ i N − λ k N ) 2 = − 1 N 2 ∑ i = 1 N ∑ k , j : k ≠ i ≠ j H N ( λ i N , λ j N ) ( λ i N − λ k N ) ( λ i N − λ j N ) − 1 N 2 ∑ i = 1 N ∑ k : k ≠ i H N ( λ i N , λ k N ) ( λ i N − λ k N ) 2 + 2 N 2 ∑ i = 1 N ∑ k : k ≠ i g N 2 ( λ i N ) h N 2 ( λ i N ) ( λ i N − λ k N ) 2 = − 1 N 2 ∑ i = 1 N ∑ k , j : k ≠ i ≠ j H N ( λ i N , λ j N ) ( λ i N − λ k N ) ( λ i N − λ j N ) − 1 N 2 ∑ i = 1 N ∑ k : k ≠ i H N ( λ i N , λ k N ) − 2 g 2 ( λ i N ) h N 2 ( λ i N ) ( λ i N − λ k N ) 2 , \begin{aligned}
S_2(\lambda^N(t))
={}&-\frac{1}{N^2}\sum_{i=1}^{N}\sum_{k:k\ne i}\frac{1}{\lambda_i^N-\lambda_k^N}\sum_{j:j\ne i}\frac{H_N(\lambda_i^N,\lambda_j^N)}{\lambda_i^N-\lambda_j^N}
+\frac{2}{N^2}\sum_{i=1}^{N}\sum_{k:k\ne i}\frac{g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)}{(\lambda_i^N-\lambda_k^N)^2}\\
={}&-\frac{1}{N^2}\sum_{i=1}^{N}\sum_{k,j:k\ne i\ne j}\frac{H_N(\lambda_i^N,\lambda_j^N)}{(\lambda_i^N-\lambda_k^N)(\lambda_i^N-\lambda_j^N)}
-\frac{1}{N^2}\sum_{i=1}^{N}\sum_{k:k\ne i}\frac{H_N(\lambda_i^N,\lambda_k^N)}{(\lambda_i^N-\lambda_k^N)^2}
+\frac{2}{N^2}\sum_{i=1}^{N}\sum_{k:k\ne i}\frac{g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)}{(\lambda_i^N-\lambda_k^N)^2}\\
={}&-\frac{1}{N^2}\sum_{i=1}^{N}\sum_{k,j:k\ne i\ne j}\frac{H_N(\lambda_i^N,\lambda_j^N)}{(\lambda_i^N-\lambda_k^N)(\lambda_i^N-\lambda_j^N)}
-\frac{1}{N^2}\sum_{i=1}^{N}\sum_{k:k\ne i}\frac{H_N(\lambda_i^N,\lambda_k^N)-2g^2(\lambda_i^N)h_N^2(\lambda_i^N)}{(\lambda_i^N-\lambda_k^N)^2},
\end{aligned} S 2 ( λ N ( t )) = = = − N 2 1 i = 1 ∑ N k : k = i ∑ λ i N − λ k N 1 j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) + N 2 2 i = 1 ∑ N k : k = i ∑ ( λ i N − λ k N ) 2 g N 2 ( λ i N ) h N 2 ( λ i N ) − N 2 1 i = 1 ∑ N k , j : k = i = j ∑ ( λ i N − λ k N ) ( λ i N − λ j N ) H N ( λ i N , λ j N ) − N 2 1 i = 1 ∑ N k : k = i ∑ ( λ i N − λ k N ) 2 H N ( λ i N , λ k N ) + N 2 2 i = 1 ∑ N k : k = i ∑ ( λ i N − λ k N ) 2 g N 2 ( λ i N ) h N 2 ( λ i N ) − N 2 1 i = 1 ∑ N k , j : k = i = j ∑ ( λ i N − λ k N ) ( λ i N − λ j N ) H N ( λ i N , λ j N ) − N 2 1 i = 1 ∑ N k : k = i ∑ ( λ i N − λ k N ) 2 H N ( λ i N , λ k N ) − 2 g 2 ( λ i N ) h N 2 ( λ i N ) ,
where in the above k ≠ i ≠ j k\ne i\ne j k = i = j actually means k ≠ i k\ne i k = i , i ≠ j i\ne j i = j , k ≠ j k\ne j k = j .
The second term on the right-hand side equals
− 1 N 2 ∑ i < k H N ( λ i N , λ k N ) − 2 g 2 ( λ i N ) h N 2 ( λ i N ) + H N ( λ i N , λ k N ) − 2 g 2 ( λ k N ) h N 2 ( λ k N ) ( λ i N − λ k N ) 2 = 2 N 2 ∑ i < k ( g N 2 ( λ i N ) − g N 2 ( λ k N ) ) ( h N 2 ( λ i N ) − h N 2 ( λ k N ) ) ( λ i N − λ k N ) 2 , \begin{aligned}
-\frac{1}{N^2}\sum_{i<k}\frac{H_N(\lambda_i^N,\lambda_k^N)-2g^2(\lambda_i^N)h_N^2(\lambda_i^N)+H_N(\lambda_i^N,\lambda_k^N)-2g^2(\lambda_k^N)h_N^2(\lambda_k^N)}{(\lambda_i^N-\lambda_k^N)^2} \\
&=\frac{2}{N^2}\sum_{i<k}\frac{\left(g_N^2(\lambda_i^N)-g_N^2(\lambda_k^N)\right)\left(h_N^2(\lambda_i^N)-h_N^2(\lambda_k^N)\right)}{(\lambda_i^N-\lambda_k^N)^2},
\end{aligned} − N 2 1 i < k ∑ ( λ i N − λ k N ) 2 H N ( λ i N , λ k N ) − 2 g 2 ( λ i N ) h N 2 ( λ i N ) + H N ( λ i N , λ k N ) − 2 g 2 ( λ k N ) h N 2 ( λ k N ) = N 2 2 i < k ∑ ( λ i N − λ k N ) 2 ( g N 2 ( λ i N ) − g N 2 ( λ k N ) ) ( h N 2 ( λ i N ) − h N 2 ( λ k N ) ) ,
whose absolute value is bounded by 2 N 2 N ( N − 1 ) 2 L g N 2 L h N 2 ≤ L g N 2 L h N 2 \frac{2}{N^2}\frac{N(N-1)}{2}L_{g_N}^2L_{h_N}^2\le L_{g_N}^2L_{h_N}^2 N 2 2 2 N ( N − 1 ) L g N 2 L h N 2 ≤ L g N 2 L h N 2 .
For the first term of S 2 ( λ N ( t ) ) S_2(\lambda^N(t)) S 2 ( λ N ( t )) , by symmetry we have
− 1 N 2 ∑ i = 1 N ∑ k , j : k ≠ i ≠ j H N ( λ i N , λ j N ) ( λ i N − λ k N ) ( λ i N − λ j N ) = − 1 N 2 ∑ 1 ≤ i < j < k ≤ N [ H N ( λ i N , λ j N ) + H N ( λ i N , λ k N ) ( λ i N − λ k N ) ( λ i N − λ j N ) + H N ( λ j N , λ i N ) + H N ( λ j N , λ k N ) ( λ j N − λ k N ) ( λ j N − λ i N ) \begin{aligned}
-\frac{1}{N^2}\sum_{i=1}^{N}\sum_{k,j:k\ne i\ne j}\frac{H_N(\lambda_i^N,\lambda_j^N)}{(\lambda_i^N-\lambda_k^N)(\lambda_i^N-\lambda_j^N)} \\
&=-\frac{1}{N^2}\sum_{1\le i<j<k\le N}[
\frac{H_N(\lambda_i^N,\lambda_j^N)+H_N(\lambda_i^N,\lambda_k^N)}{(\lambda_i^N-\lambda_k^N)(\lambda_i^N-\lambda_j^N)}
+\frac{H_N(\lambda_j^N,\lambda_i^N)+H_N(\lambda_j^N,\lambda_k^N)}{(\lambda_j^N-\lambda_k^N)(\lambda_j^N-\lambda_i^N)}
\end{aligned} − N 2 1 i = 1 ∑ N k , j : k = i = j ∑ ( λ i N − λ k N ) ( λ i N − λ j N ) H N ( λ i N , λ j N ) = − N 2 1 1 ≤ i < j < k ≤ N ∑ [ ( λ i N − λ k N ) ( λ i N − λ j N ) H N ( λ i N , λ j N ) + H N ( λ i N , λ k N ) + ( λ j N − λ k N ) ( λ j N − λ i N ) H N ( λ j N , λ i N ) + H N ( λ j N , λ k N )
H N ( λ i N , λ k N ) + H N ( λ i N , λ k N ) ( λ k N − λ i N ) ( λ k N − λ j N ) = − 1 N 2 ∑ 1 ≤ i < j < k ≤ N H N ( λ i N , λ k N ) ( λ k N − λ i N ) − H N ( λ i N , λ j N ) ( λ j N − λ i N ) − H N ( λ j N , λ k N ) ( λ k N − λ j N ) ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) \begin{aligned}
H_N(\lambda_i^N,\lambda_k^N)+H_N(\lambda_i^N,\lambda_k^N)
\over
(\lambda_k^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N) \\
&=-\frac{1}{N^2}\sum_{1\le i<j<k\le N}
\frac{H_N(\lambda_i^N,\lambda_k^N)(\lambda_k^N-\lambda_i^N)-H_N(\lambda_i^N,\lambda_j^N)(\lambda_j^N-\lambda_i^N)-H_N(\lambda_j^N,\lambda_k^N)(\lambda_k^N-\lambda_j^N)}
{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N)}
\end{aligned} ( λ k N − λ i N ) ( λ k N − λ j N ) H N ( λ i N , λ k N ) + H N ( λ i N , λ k N ) = − N 2 1 1 ≤ i < j < k ≤ N ∑ ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) H N ( λ i N , λ k N ) ( λ k N − λ i N ) − H N ( λ i N , λ j N ) ( λ j N − λ i N ) − H N ( λ j N , λ k N ) ( λ k N − λ j N )
For the numerator, we have
H N ( λ i N , λ k N ) ( λ k N − λ i N ) − H N ( λ i N , λ j N ) ( λ j N − λ i N ) − H N ( λ j N , λ k N ) ( λ k N − λ j N ) = [ g N 2 ( λ i N ) h N 2 ( λ i N ) + g N 2 ( λ k N ) h N 2 ( λ k N ) − ( g N 2 ( λ i N ) − g N 2 ( λ k N ) ) ( h N 2 ( λ i N ) − h N 2 ( λ k N ) ) ] ( λ k N − λ i N ) − [ g N 2 ( λ i N ) h N 2 ( λ i N ) + g N 2 ( λ j N ) h N 2 ( λ j N ) − ( g N 2 ( λ i N ) − g N 2 ( λ j N ) ) ( h N 2 ( λ i N ) − h N 2 ( λ j N ) ) ] ( λ j N − λ i N ) − [ g N 2 ( λ k N ) h N 2 ( λ k N ) + g N 2 ( λ j N ) h N 2 ( λ j N ) − ( g N 2 ( λ k N ) − g N 2 ( λ j N ) ) ( h N 2 ( λ k N ) − h N 2 ( λ j N ) ) ] ( λ k N − λ j N ) = g N 2 ( λ i N ) h N 2 ( λ i N ) ( λ k N − λ j N ) − g N 2 ( λ j N ) h N 2 ( λ j N ) ( λ k N − λ i N ) + g N 2 ( λ k N ) h N 2 ( λ k N ) ( λ j N − λ i N ) − ( g N 2 ( λ i N ) − g N 2 ( λ k N ) ) ( h N 2 ( λ i N ) − h N 2 ( λ k N ) ) ( λ k N − λ i N ) + ( g N 2 ( λ i N ) − g N 2 ( λ j N ) ) ( h N 2 ( λ i N ) − h N 2 ( λ j N ) ) ( λ j N − λ i N ) + ( g N 2 ( λ k N ) − g N 2 ( λ j N ) ) ( h N 2 ( λ k N ) − h N 2 ( λ j N ) ) ( λ k N − λ j N ) : = I + I I + I I I + I V . \begin{aligned}
&H_N(\lambda_i^N,\lambda_k^N)(\lambda_k^N-\lambda_i^N)-H_N(\lambda_i^N,\lambda_j^N)(\lambda_j^N-\lambda_i^N)-H_N(\lambda_j^N,\lambda_k^N)(\lambda_k^N-\lambda_j^N)\\
={}&[g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)+g_N^2(\lambda_k^N)h_N^2(\lambda_k^N)-(g_N^2(\lambda_i^N)-g_N^2(\lambda_k^N))(h_N^2(\lambda_i^N)-h_N^2(\lambda_k^N))](\lambda_k^N-\lambda_i^N)\\
&-[g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)+g_N^2(\lambda_j^N)h_N^2(\lambda_j^N)-(g_N^2(\lambda_i^N)-g_N^2(\lambda_j^N))(h_N^2(\lambda_i^N)-h_N^2(\lambda_j^N))](\lambda_j^N-\lambda_i^N)\\
&-[g_N^2(\lambda_k^N)h_N^2(\lambda_k^N)+g_N^2(\lambda_j^N)h_N^2(\lambda_j^N)-(g_N^2(\lambda_k^N)-g_N^2(\lambda_j^N))(h_N^2(\lambda_k^N)-h_N^2(\lambda_j^N))](\lambda_k^N-\lambda_j^N)\\
={}&g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)(\lambda_k^N-\lambda_j^N)-g_N^2(\lambda_j^N)h_N^2(\lambda_j^N)(\lambda_k^N-\lambda_i^N)+g_N^2(\lambda_k^N)h_N^2(\lambda_k^N)(\lambda_j^N-\lambda_i^N)\\
&-(g_N^2(\lambda_i^N)-g_N^2(\lambda_k^N))(h_N^2(\lambda_i^N)-h_N^2(\lambda_k^N))(\lambda_k^N-\lambda_i^N)\\
&+(g_N^2(\lambda_i^N)-g_N^2(\lambda_j^N))(h_N^2(\lambda_i^N)-h_N^2(\lambda_j^N))(\lambda_j^N-\lambda_i^N)\\
&+(g_N^2(\lambda_k^N)-g_N^2(\lambda_j^N))(h_N^2(\lambda_k^N)-h_N^2(\lambda_j^N))(\lambda_k^N-\lambda_j^N)\\
:={}&I+II+III+IV.
\end{aligned} = = := H N ( λ i N , λ k N ) ( λ k N − λ i N ) − H N ( λ i N , λ j N ) ( λ j N − λ i N ) − H N ( λ j N , λ k N ) ( λ k N − λ j N ) [ g N 2 ( λ i N ) h N 2 ( λ i N ) + g N 2 ( λ k N ) h N 2 ( λ k N ) − ( g N 2 ( λ i N ) − g N 2 ( λ k N )) ( h N 2 ( λ i N ) − h N 2 ( λ k N ))] ( λ k N − λ i N ) − [ g N 2 ( λ i N ) h N 2 ( λ i N ) + g N 2 ( λ j N ) h N 2 ( λ j N ) − ( g N 2 ( λ i N ) − g N 2 ( λ j N )) ( h N 2 ( λ i N ) − h N 2 ( λ j N ))] ( λ j N − λ i N ) − [ g N 2 ( λ k N ) h N 2 ( λ k N ) + g N 2 ( λ j N ) h N 2 ( λ j N ) − ( g N 2 ( λ k N ) − g N 2 ( λ j N )) ( h N 2 ( λ k N ) − h N 2 ( λ j N ))] ( λ k N − λ j N ) g N 2 ( λ i N ) h N 2 ( λ i N ) ( λ k N − λ j N ) − g N 2 ( λ j N ) h N 2 ( λ j N ) ( λ k N − λ i N ) + g N 2 ( λ k N ) h N 2 ( λ k N ) ( λ j N − λ i N ) − ( g N 2 ( λ i N ) − g N 2 ( λ k N )) ( h N 2 ( λ i N ) − h N 2 ( λ k N )) ( λ k N − λ i N ) + ( g N 2 ( λ i N ) − g N 2 ( λ j N )) ( h N 2 ( λ i N ) − h N 2 ( λ j N )) ( λ j N − λ i N ) + ( g N 2 ( λ k N ) − g N 2 ( λ j N )) ( h N 2 ( λ k N ) − h N 2 ( λ j N )) ( λ k N − λ j N ) I + I I + I I I + I V .
For the term I I I we have the following estimate: for i < j < k i<j<k i < j < k ,
I ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) = [ g N 2 ( λ i N ) h N 2 ( λ i N ) − g N 2 ( λ j N ) h N 2 ( λ j N ) ] ( λ k N − λ j N ) + [ g N 2 ( λ k N ) h N 2 ( λ k N ) − g N 2 ( λ j N ) h N 2 ( λ j N ) ] ( λ j N − λ i N ) ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) = 1 λ k N − λ i N [ g N 2 ( λ i N ) h N 2 ( λ i N ) − g N 2 ( λ j N ) h N 2 ( λ j N ) λ k N − λ j N − g N 2 ( λ j N ) h N 2 ( λ j N ) − g N 2 ( λ i N ) h N 2 ( λ i N ) λ j N − λ i N ] ≥ − C , \begin{aligned}
\frac{I}{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N)} \\
&=\frac{[g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)-g_N^2(\lambda_j^N)h_N^2(\lambda_j^N)](\lambda_k^N-\lambda_j^N)+[g_N^2(\lambda_k^N)h_N^2(\lambda_k^N)-g_N^2(\lambda_j^N)h_N^2(\lambda_j^N)](\lambda_j^N-\lambda_i^N)}
{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N)} \\
&=\frac{1}{\lambda_k^N-\lambda_i^N}\left[
\frac{g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)-g_N^2(\lambda_j^N)h_N^2(\lambda_j^N)}{\lambda_k^N-\lambda_j^N}
-\frac{g_N^2(\lambda_j^N)h_N^2(\lambda_j^N)-g_N^2(\lambda_i^N)h_N^2(\lambda_i^N)}{\lambda_j^N-\lambda_i^N}
\right]\ge-C,
\end{aligned} ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) I = ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) [ g N 2 ( λ i N ) h N 2 ( λ i N ) − g N 2 ( λ j N ) h N 2 ( λ j N )] ( λ k N − λ j N ) + [ g N 2 ( λ k N ) h N 2 ( λ k N ) − g N 2 ( λ j N ) h N 2 ( λ j N )] ( λ j N − λ i N ) = λ k N − λ i N 1 [ λ k N − λ j N g N 2 ( λ i N ) h N 2 ( λ i N ) − g N 2 ( λ j N ) h N 2 ( λ j N ) − λ j N − λ i N g N 2 ( λ j N ) h N 2 ( λ j N ) − g N 2 ( λ i N ) h N 2 ( λ i N ) ] ≥ − C ,
for some constant C > 0 C>0 C > 0 , noting that g N 2 h N 2 g_N^2h_N^2 g N 2 h N 2 is either convex or of class C 1 , 1 C^{1,1} C 1 , 1 and λ i N < λ j N < λ k N \lambda_i^N<\lambda_j^N<\lambda_k^N λ i N < λ j N < λ k N .
For the remaining terms, we have for i < j < k i<j<k i < j < k ,
∣ I I + I I I + I V ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) ∣ = ∣ − [ ( g N 2 ( λ k N ) − g N 2 ( λ j N ) ) + ( g N 2 ( λ j N ) − g N 2 ( λ i N ) ) ] [ ( h N 2 ( λ k N ) − h N 2 ( λ j N ) ) + ( h N 2 ( λ j N ) − h N 2 ( λ i N ) ) ] ( λ j N − λ i N ) ( λ k N − λ j N ) + ( g N 2 ( λ j N ) − g N 2 ( λ i N ) ) ( h N 2 ( λ j N ) − h N 2 ( λ i N ) ) ( λ j N − λ i N ) ( λ k N − λ i N ) + ( g N 2 ( λ k N ) − g N 2 ( λ j N ) ) ( h N 2 ( λ k N ) − h N 2 ( λ j N ) ) ( λ k N − λ j N ) ( λ k N − λ i N ) ∣ = ∣ − ( g N 2 ( λ j N ) − g N 2 ( λ i N ) ) ( h N 2 ( λ j N ) − h N 2 ( λ i N ) ) ( λ j N − λ i N ) ( λ k N − λ i N ) − ( g N 2 ( λ k N ) − g N 2 ( λ j N ) ) ( h N 2 ( λ k N ) − h N 2 ( λ j N ) ) ( λ j N − λ i N ) ( λ k N − λ j N ) − ( g N 2 ( λ j N ) − g N 2 ( λ i N ) ) ( h N 2 ( λ k N ) − h N 2 ( λ j N ) ) ( λ j N − λ i N ) ( λ k N − λ i N ) − ( g N 2 ( λ k N ) − g N 2 ( λ j N ) ) ( h N 2 ( λ j N ) − h N 2 ( λ i N ) ) ( λ k N − λ i N ) ( λ k N − λ j N ) ∣ ≤ L g N 2 L h N 2 λ j N − λ i N λ k N − λ i N + 2 L g N 2 L h N 2 + L g N 2 L h N 2 λ k N − λ j N λ k N − λ i N = 3 L g N 2 L h N 2 . \begin{aligned}
\left|\frac{II+III+IV}{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N)}\right|
={}&\left|-\frac{[(g_N^2(\lambda_k^N)-g_N^2(\lambda_j^N))+(g_N^2(\lambda_j^N)-g_N^2(\lambda_i^N))][(h_N^2(\lambda_k^N)-h_N^2(\lambda_j^N))+(h_N^2(\lambda_j^N)-h_N^2(\lambda_i^N))]}
{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N)}\right.\\
&\left.+\frac{(g_N^2(\lambda_j^N)-g_N^2(\lambda_i^N))(h_N^2(\lambda_j^N)-h_N^2(\lambda_i^N))}
{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_i^N)}
+\frac{(g_N^2(\lambda_k^N)-g_N^2(\lambda_j^N))(h_N^2(\lambda_k^N)-h_N^2(\lambda_j^N))}
{(\lambda_k^N-\lambda_j^N)(\lambda_k^N-\lambda_i^N)}\right|\\
={}&\left|-\frac{(g_N^2(\lambda_j^N)-g_N^2(\lambda_i^N))(h_N^2(\lambda_j^N)-h_N^2(\lambda_i^N))}
{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_i^N)}
-\frac{(g_N^2(\lambda_k^N)-g_N^2(\lambda_j^N))(h_N^2(\lambda_k^N)-h_N^2(\lambda_j^N))}
{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N)}\right.\\
&\left.-\frac{(g_N^2(\lambda_j^N)-g_N^2(\lambda_i^N))(h_N^2(\lambda_k^N)-h_N^2(\lambda_j^N))}
{(\lambda_j^N-\lambda_i^N)(\lambda_k^N-\lambda_i^N)}
-\frac{(g_N^2(\lambda_k^N)-g_N^2(\lambda_j^N))(h_N^2(\lambda_j^N)-h_N^2(\lambda_i^N))}
{(\lambda_k^N-\lambda_i^N)(\lambda_k^N-\lambda_j^N)}\right|\\
\le{}&L_{g_N^2}L_{h_N^2}\frac{\lambda_j^N-\lambda_i^N}{\lambda_k^N-\lambda_i^N}+2L_{g_N^2}L_{h_N^2}+L_{g_N^2}L_{h_N^2}\frac{\lambda_k^N-\lambda_j^N}{\lambda_k^N-\lambda_i^N}
=3L_{g_N^2}L_{h_N^2}.
\end{aligned} ( λ j N − λ i N ) ( λ k N − λ i N ) ( λ k N − λ j N ) I I + I I I + I V = = ≤ − ( λ j N − λ i N ) ( λ k N − λ j N ) [( g N 2 ( λ k N ) − g N 2 ( λ j N )) + ( g N 2 ( λ j N ) − g N 2 ( λ i N ))] [( h N 2 ( λ k N ) − h N 2 ( λ j N )) + ( h N 2 ( λ j N ) − h N 2 ( λ i N ))] + ( λ j N − λ i N ) ( λ k N − λ i N ) ( g N 2 ( λ j N ) − g N 2 ( λ i N )) ( h N 2 ( λ j N ) − h N 2 ( λ i N )) + ( λ k N − λ j N ) ( λ k N − λ i N ) ( g N 2 ( λ k N ) − g N 2 ( λ j N )) ( h N 2 ( λ k N ) − h N 2 ( λ j N )) − ( λ j N − λ i N ) ( λ k N − λ i N ) ( g N 2 ( λ j N ) − g N 2 ( λ i N )) ( h N 2 ( λ j N ) − h N 2 ( λ i N )) − ( λ j N − λ i N ) ( λ k N − λ j N ) ( g N 2 ( λ k N ) − g N 2 ( λ j N )) ( h N 2 ( λ k N ) − h N 2 ( λ j N )) − ( λ j N − λ i N ) ( λ k N − λ i N ) ( g N 2 ( λ j N ) − g N 2 ( λ i N )) ( h N 2 ( λ k N ) − h N 2 ( λ j N )) − ( λ k N − λ i N ) ( λ k N − λ j N ) ( g N 2 ( λ k N ) − g N 2 ( λ j N )) ( h N 2 ( λ j N ) − h N 2 ( λ i N )) L g N 2 L h N 2 λ k N − λ i N λ j N − λ i N + 2 L g N 2 L h N 2 + L g N 2 L h N 2 λ k N − λ i N λ k N − λ j N = 3 L g N 2 L h N 2 .
By the above analysis for I I I and I I + I I I + I V II+III+IV I I + I I I + I V , we get the following estimate for the first term of S 2 ( λ N ( t ) ) S_2(\lambda^N(t)) S 2 ( λ N ( t )) :
− 1 N 2 ∑ 1 ≤ i ≠ k ≠ j ≤ N H N ( λ i N , λ j N ) ( λ i N − λ k N ) ( λ i N − λ j N ) ≤ 1 N 2 ( N 3 ) C , -\frac{1}{N^2}\sum_{1\leq i\ne k\ne j\leq N}\frac{H_N(\lambda_i^N,\lambda_j^N)}{(\lambda_i^N-\lambda_k^N)(\lambda_i^N-\lambda_j^N)}\leq\frac{1}{N^2}\binom{N}{3}C, − N 2 1 1 ≤ i = k = j ≤ N ∑ ( λ i N − λ k N ) ( λ i N − λ j N ) H N ( λ i N , λ j N ) ≤ N 2 1 ( 3 N ) C ,
for some C > 0 C>0 C > 0 . Then, combining the estimates for the two terms of S 2 ( λ N ( t ) ) S_2(\lambda^N(t)) S 2 ( λ N ( t )) , we get
S 2 ( λ N ( t ) ) ≤ C , (58) S_2(\lambda^N(t))\leq C, \tag*{(58)} S 2 ( λ N ( t )) ≤ C , (58)
for some constant C C C that may depending on N N N .
Therefore, by the equation (54) of d f ( λ N ( t ) ) \mathrm{d}f(\lambda^N(t)) d f ( λ N ( t )) and the estimations (55) , (56) , (57) and (58) , we obtain
d f ( λ N ( t ) ) ≤ C ∑ i = 1 N ( 1 + ( λ i N ( t ) ) 2 ) d t + d M ~ ( t ) = [ C N + C ∑ i = 1 N ( λ i N ( t ) ) 2 ] d t + d M ~ ( t ) ≤ C ( f ( λ N ( t ) ) + 1 ) d t + d M ~ ( t ) , \begin{aligned}
\mathrm{d}f(\lambda^N(t))\leq C\sum_{i=1}^{N}(1+(\lambda_i^N(t))^2)\mathrm{d}t+\mathrm{d}\widetilde{M}(t)=[CN+C\sum_{i=1}^{N}(\lambda_i^N(t))^2]\mathrm{d}t+\mathrm{d}\widetilde{M}(t) \\
&\leq C(f(\lambda^N(t))+1)\mathrm{d}t+\mathrm{d}\widetilde{M}(t),
\end{aligned} d f ( λ N ( t )) ≤ C i = 1 ∑ N ( 1 + ( λ i N ( t ) ) 2 ) d t + d M ( t ) = [ C N + C i = 1 ∑ N ( λ i N ( t ) ) 2 ] d t + d M ( t ) ≤ C ( f ( λ N ( t )) + 1 ) d t + d M ( t ) ,
where the last step follows from (51) . Taking the integral and then the expectation, by (53) we obtain
E ^ [ f ( λ N ( t ∧ T M ) ) ] ≤ C E ^ [ ∫ 0 t ∧ T M f ( λ N ( s ) ) d s ] + C t + f ( λ N ( 0 ) ) ≤ C ∫ 0 t E ^ [ f ( λ N ( s ∧ T M ) ) ] d s + C t + f ( λ N ( 0 ) ) . \begin{aligned}
\widehat{\mathbb{E}}[f(\lambda^N(t\wedge T_M))]\leq C\widehat{\mathbb{E}}\left[\int_0^{t\wedge T_M}f(\lambda^N(s))\,\mathrm{d}s\right]+C_t+f(\lambda^N(0)) \\
&\leq C\int_0^t\widehat{\mathbb{E}}[f(\lambda^N(s\wedge T_M))]\,\mathrm{d}s+C_t+f(\lambda^N(0)).
\end{aligned} E [ f ( λ N ( t ∧ T M ))] ≤ C E [ ∫ 0 t ∧ T M f ( λ N ( s )) d s ] + C t + f ( λ N ( 0 )) ≤ C ∫ 0 t E [ f ( λ N ( s ∧ T M ))] d s + C t + f ( λ N ( 0 )) .
Then, Gronwall’s inequality yields E ^ [ f ( λ N ( t ∧ T M ) ) ] ≤ C t \widehat{\mathbb{E}}[f(\lambda^N(t\wedge T_M))]\leq C_t E [ f ( λ N ( t ∧ T M ))] ≤ C t , where C t C_t C t is a finite number only depending on t t t , N N N and λ N ( 0 ) \lambda^N(0) λ N ( 0 ) . Noting that 0 ≤ f ( λ N ( t ∧ T M ) ) + 4 ≤ M + 4 0\leq f(\lambda^N(t\wedge T_M))+4\leq M+4 0 ≤ f ( λ N ( t ∧ T M )) + 4 ≤ M + 4 q.s. and that f ( λ N ( t ∧ T M ) ) = M f(\lambda^N(t\wedge T_M))=M f ( λ N ( t ∧ T M )) = M on the event [ T M ≤ t ] [T_M\leq t] [ T M ≤ t ] , we get
c ( T M ≤ t ) ( M + 4 ) ≤ E ^ [ f ( λ N ( t ∧ T M ) ) + 4 ] ≤ C t + 4 , c(T_M\leq t)(M+4)\leq\widehat{\mathbb{E}}[f(\lambda^N(t\wedge T_M))+4]\leq C_t+4, c ( T M ≤ t ) ( M + 4 ) ≤ E [ f ( λ N ( t ∧ T M )) + 4 ] ≤ C t + 4 ,
and hence
c ( T M ≤ t ) ≤ C t + 4 M + 4 . c(T_M\leq t)\leq\frac{C_t+4}{M+4}. c ( T M ≤ t ) ≤ M + 4 C t + 4 .
Denote A n t = { T n 2 ≤ t } A_n^t=\{T_{n^2}\leq t\} A n t = { T n 2 ≤ t } for n ∈ N n\in\mathbb{N} n ∈ N . Then, we have
∑ n = 1 ∞ c ( A n t ) = ∑ n = 1 ∞ c ( T n 2 ≤ t ) ≤ ∑ n = 1 ∞ C t + 4 n 2 + 4 < ∞ , \sum_{n=1}^{\infty}c(A_n^t)=\sum_{n=1}^{\infty}c(T_{n^2}\leq t)\leq\sum_{n=1}^{\infty}\frac{C_t+4}{n^2+4}<\infty, n = 1 ∑ ∞ c ( A n t ) = n = 1 ∑ ∞ c ( T n 2 ≤ t ) ≤ n = 1 ∑ ∞ n 2 + 4 C t + 4 < ∞ ,
and Lemma 2.18 yields that lim sup n → ∞ A n t = ⋂ k = 1 ∞ ⋃ n = 1 ∞ A n t \limsup_{n\to\infty}A_n^t=\bigcap_{k=1}^{\infty}\bigcup_{n=1}^{\infty}A_n^t lim sup n → ∞ A n t = ⋂ k = 1 ∞ ⋃ n = 1 ∞ A n t is a polar set, i.e., c ( ⋂ n = 1 ∞ A n t ) = 0 c(\bigcap_{n=1}^{\infty}A_n^t)=0 c ( ⋂ n = 1 ∞ A n t ) = 0 . Thus, we have c ( ⋃ k = 1 ∞ ⋂ n = 1 ∞ A k n ) = lim k → ∞ c ( ⋂ n = 1 ∞ A k n ) = 0 c(\bigcup_{k=1}^{\infty}\bigcap_{n=1}^{\infty}A_k^n)=\lim_{k\to\infty}c(\bigcap_{n=1}^{\infty}A_k^n)=0 c ( ⋃ k = 1 ∞ ⋂ n = 1 ∞ A k n ) = lim k → ∞ c ( ⋂ n = 1 ∞ A k n ) = 0 . Consequently, we get
c ( { lim n → ∞ T n 2 = ∞ } ) = c ( ⋂ k = 1 ∞ ⋃ n = 1 ∞ ( A n k ) c ) = 1 , c\left(\left\{\lim_{n\to\infty}T_{n^2}=\infty\right\}\right)=c\left(\bigcap_{k=1}^{\infty}\bigcup_{n=1}^{\infty}(A_n^k)^c\right)=1, c ( { n → ∞ lim T n 2 = ∞ } ) = c ( k = 1 ⋂ ∞ n = 1 ⋃ ∞ ( A n k ) c ) = 1 ,
and thus
c ( { lim R → ∞ τ R = ∞ } ) ≥ c ( { lim n → ∞ T n 2 = ∞ } ) = 1 , c\left(\left\{\lim_{R\to\infty}\tau_R=\infty\right\}\right)\geq c\left(\left\{\lim_{n\to\infty}T_{n^2}=\infty\right\}\right)=1, c ( { R → ∞ lim τ R = ∞ } ) ≥ c ( { n → ∞ lim T n 2 = ∞ } ) = 1 ,
since T n 2 ≤ τ R T_{n^2}\leq\tau_R T n 2 ≤ τ R with R = e N 2 ( 4 + n 2 ) R=e^{N^2(4+n^2)} R = e N 2 ( 4 + n 2 ) . This implies τ = ∞ \tau=\infty τ = ∞ q.s., since τ R ≤ τ \tau_R\leq\tau τ R ≤ τ for all R ∈ R R\in\mathbb{R} R ∈ R .
Finally, noting that τ = ∞ \tau=\infty τ = ∞ q.s., U U U can be chosen as a smooth function of X t N X_t^N X t N . By assumption, X t N X_t^N X t N has a q.c. version which is still denoted by X t N X_t^N X t N , then U ( X t N ) U(X_t^N) U ( X t N ) is also q.c. Thus, it is easy to check that B ‾ ( t ) = ( B ‾ i ( t ) , i = 1 , … , N ) \overline{B}(t)=(\overline{B}_i(t),i=1,\ldots,N) B ( t ) = ( B i ( t ) , i = 1 , … , N ) given by (49) is a symmetric martingale. Then, by Corollary 2.14 and (50) , we get that B ˉ ( t ) \bar{B}(t) B ˉ ( t ) is an N N N -dimensional G ˉ \bar{G} G ˉ -Brownian motion satisfying Assumption 1 with d = N d=N d = N .
Now we are ready to show that the G G G -SDE (7) admits a unique strong solution.
Proof of Theorem 1.2 . Let W ˉ \bar{W} W ˉ be an N N N -dimensional Wiener process, and we denote d W ˉ t θ = θ t d W ˉ t d\bar{W}^{\theta}_{t}=\theta_{t}d\bar{W}_{t} d W ˉ t θ = θ t d W ˉ t , where θ \theta θ is a Θ \Theta Θ -valued adapted process with Θ \Theta Θ given by (34) . Assuming θ t = σ t I N \theta_{t}=\sigma_{t}I_{N} θ t = σ t I N , where ( σ t ) t ≥ 0 (\sigma_{t})_{t\geq0} ( σ t ) t ≥ 0 is a process which is adapted to the filtration generated by W ˉ \bar{W} W ˉ and takes values in [ σ ‾ , σ ˉ ] [\underline{\sigma},\bar{\sigma}] [ σ , σ ˉ ] , we have d W ˉ t θ , i = σ t d W ˉ t i d\bar{W}^{\theta,i}_{t}=\sigma_{t}d\bar{W}^{i}_{t} d W ˉ t θ , i = σ t d W ˉ t i , for i = 1 , … , N i=1,\ldots,N i = 1 , … , N . Then, the G G G -SDE (7) under P θ P_{\theta} P θ becomes
d λ i N = 2 g N ( λ i N ) h N ( λ i N ) σ t d W ˉ t i + b N ( λ i N ) σ t 2 d t + ∑ j : j ≠ i H N ( λ i N , λ j N ) λ i N − λ j N σ t 2 d t . (59) d\lambda_{i}^{N}=2g_{N}(\lambda_{i}^{N})h_{N}(\lambda_{i}^{N})\sigma_{t}d\bar{W}^{i}_{t}+b_{N}(\lambda_{i}^{N})\sigma_{t}^{2}dt+\sum_{j:j\neq i}\frac{H_{N}(\lambda_{i}^{N},\lambda_{j}^{N})}{\lambda_{i}^{N}-\lambda_{j}^{N}}\sigma_{t}^{2}dt. \tag*{(59)} d λ i N = 2 g N ( λ i N ) h N ( λ i N ) σ t d W ˉ t i + b N ( λ i N ) σ t 2 d t + j : j = i ∑ λ i N − λ j N H N ( λ i N , λ j N ) σ t 2 d t . (59)
Clearly, Theorem 1.1 guarantees the existence of a weak solution to the G G G -SDE (7) . To show that this weak solution is in fact the unique strong solution, it suffices to prove that it is the unique strong solution to 4.13 under P θ P_{\theta} P θ for every θ ∈ Θ \theta\in\Theta θ ∈ Θ . By the Yamada–Watanabe theorem, this reduces to establishing pathwise uniqueness under each P θ P_{\theta} P θ .
Now, suppose λ N = ( λ 1 N , … , λ N N ) \lambda^{N}=(\lambda_{1}^{N},\ldots,\lambda_{N}^{N}) λ N = ( λ 1 N , … , λ N N ) and λ ~ N = ( λ ~ 1 N , … , λ ~ N N ) \widetilde{\lambda}^{N}=(\widetilde{\lambda}_{1}^{N},\ldots,\widetilde{\lambda}_{N}^{N}) λ N = ( λ 1 N , … , λ N N ) are two solutions of (59) driven by the same Wiener process W ˉ \bar{W} W ˉ . It suffices to show that λ N \lambda^{N} λ N and λ ~ N \widetilde{\lambda}^{N} λ N coincide P θ P_{\theta} P θ -a.s.
Denote, for i = 1 , … , N i=1,\ldots,N i = 1 , … , N ,
λ ˉ i N ( t ) : = λ i N ( t ) − λ ~ i N ( t ) , f ( x ) : = g N ( x ) h N ( x ) , F i ( x ) : = ∑ j : j ≠ i H N ( x i , x j ) x i − x j , \bar{\lambda}_{i}^{N}(t):=\lambda_{i}^{N}(t)-\widetilde{\lambda}_{i}^{N}(t),\qquad f(x):=g_{N}(x)h_{N}(x),\qquad F_{i}(x):=\sum_{j:j\neq i}\frac{H_{N}(x_{i},x_{j})}{x_{i}-x_{j}}, λ ˉ i N ( t ) := λ i N ( t ) − λ i N ( t ) , f ( x ) := g N ( x ) h N ( x ) , F i ( x ) := j : j = i ∑ x i − x j H N ( x i , x j ) ,
and
f ˉ i ( t ) : = f ( λ i N ( t ) ) − f ( λ ~ i N ( t ) ) , b ˉ i ( t ) : = b N ( λ i N ( t ) ) − b N ( λ ~ i N ( t ) ) , F ˉ i ( t ) : = F i ( λ N ( t ) ) − F i ( λ ~ N ( t ) ) . \bar{f}_{i}(t):=f(\lambda_{i}^{N}(t))-f(\widetilde{\lambda}_{i}^{N}(t)),\qquad\bar{b}_{i}(t):=b_{N}(\lambda_{i}^{N}(t))-b_{N}(\widetilde{\lambda}_{i}^{N}(t)),\qquad\bar{F}_{i}(t):=F_{i}(\lambda^{N}(t))-F_{i}(\widetilde{\lambda}^{N}(t)). f ˉ i ( t ) := f ( λ i N ( t )) − f ( λ i N ( t )) , b ˉ i ( t ) := b N ( λ i N ( t )) − b N ( λ i N ( t )) , F ˉ i ( t ) := F i ( λ N ( t )) − F i ( λ N ( t )) .
Then,
λ ˉ i N ( t ) = 2 ∫ 0 t f ˉ i ( s ) σ s d W ˉ s i + ∫ 0 t ( b ˉ i ( s ) + F ˉ i ( s ) ) σ s 2 d s . (60) \bar{\lambda}_{i}^{N}(t)=2\int_{0}^{t}\bar{f}_{i}(s)\sigma_{s}d\bar{W}_{s}^{i}+\int_{0}^{t}\left(\bar{b}_{i}(s)+\bar{F}_{i}(s)\right)\sigma_{s}^{2}ds. \tag*{(60)} λ ˉ i N ( t ) = 2 ∫ 0 t f ˉ i ( s ) σ s d W ˉ s i + ∫ 0 t ( b ˉ i ( s ) + F ˉ i ( s ) ) σ s 2 d s . (60)
For some fixed T > 0 T>0 T > 0 , we define
η m : = inf { t ≥ 0 : min i ≠ j ( ∣ λ i N ( t ) − λ j N ( t ) ∣ ∧ ∣ λ ~ i N ( t ) − λ ~ j N ( t ) ∣ ) ≤ 1 m or max 1 ≤ i ≤ N ( ∣ λ i N ( t ) ∣ ∨ ∣ λ ~ i N ( t ) ∣ ) ≥ m } ∧ T , m ∈ N . \eta_{m}:=\inf\left\{t\geq0:\min_{i\neq j}\left(\left|\lambda_{i}^{N}(t)-\lambda_{j}^{N}(t)\right|\wedge\left|\widetilde{\lambda}_{i}^{N}(t)-\widetilde{\lambda}_{j}^{N}(t)\right|\right)\leq\frac{1}{m}\ \text{or}\ \max_{1\leq i\leq N}\left(\left|\lambda_{i}^{N}(t)\right|\vee\left|\widetilde{\lambda}_{i}^{N}(t)\right|\right)\geq m\right\}\wedge T,\quad m\in\mathbb{N}. η m := inf { t ≥ 0 : i = j min ( λ i N ( t ) − λ j N ( t ) ∧ λ i N ( t ) − λ j N ( t ) ) ≤ m 1 or 1 ≤ i ≤ N max ( λ i N ( t ) ∨ λ i N ( t ) ) ≥ m } ∧ T , m ∈ N .
On [ 0 , η m ] [0,\eta_{m}] [ 0 , η m ] , λ N \lambda^{N} λ N and λ ~ N \widetilde{\lambda}^{N} λ N take values in a compact subset of the non-collision region. Since b N b_{N} b N and F i F_{i} F i are locally Lipschitz, there exists a constant C m > 0 C_{m}>0 C m > 0 such that
( b ˉ i ( s ) + F ˉ i ( s ) ) σ s 2 ≤ C m ∣ λ ˉ N ( s ) ∣ , for s ∈ [ 0 , η m ] . (61) \left(\bar{b}_{i}(s)+\bar{F}_{i}(s)\right)\sigma_{s}^{2}\leq C_{m}\left|\bar{\lambda}^{N}(s)\right|,\quad\text{for }s\in[0,\eta_{m}]. \tag*{(61)} ( b ˉ i ( s ) + F ˉ i ( s ) ) σ s 2 ≤ C m λ ˉ N ( s ) , for s ∈ [ 0 , η m ] . (61)
Moreover, we claim that the local time of λ ˉ i N \bar{\lambda}_{i}^{N} λ ˉ i N at zero on [ 0 , η m ] [0,\eta_{m}] [ 0 , η m ] vanishes. Indeed, since f ( x ) 2 = g N 2 ( x ) h N 2 ( x ) f(x)^{2}=g_{N}^{2}(x)h_{N}^{2}(x) f ( x ) 2 = g N 2 ( x ) h N 2 ( x ) is locally Lipschitz and hence ∣ f ( x ) ∣ |f(x)| ∣ f ( x ) ∣ is locally 1 2 \frac{1}{2} 2 1 -continuous and thus there exists ρ : ( 0 , ∞ ) → ( 0 , ∞ ) \rho:(0,\infty)\to(0,\infty) ρ : ( 0 , ∞ ) → ( 0 , ∞ ) satisfying ∫ 0 + ρ − 1 ( x ) d x = ∞ \int_{0+}\rho^{-1}(x)dx=\infty ∫ 0 + ρ − 1 ( x ) d x = ∞ such that ∣ ∣ f ( x ) ∣ − ∣ f ( y ) ∣ ∣ 2 ≤ ρ ( ∣ x − y ∣ ) \bigl||f(x)|-|f(y)|\bigr|^{2}\leq\rho(|x-y|) ∣ f ( x ) ∣ − ∣ f ( y ) ∣ 2 ≤ ρ ( ∣ x − y ∣ ) for x , y ∈ R x,y\in\mathbb{R} x , y ∈ R . Then, we have
∫ 0 t ∧ η m 1 { λ ˉ i N ( s ) ≠ 0 } ρ ( ∣ λ ˉ i N ( s ) ∣ ) d ⟨ λ ˉ i N ⟩ s = 4 ∫ 0 t ∧ η m 1 { λ ˉ i N ( s ) ≠ 0 } ∣ f ˉ i ( s ) ∣ 2 σ s 2 ρ ( ∣ λ ˉ i N ( s ) ∣ ) d s ≤ 4 σ ˉ 2 t < ∞ . \int_{0}^{t\wedge\eta_{m}}\frac{\mathbf{1}_{\{\bar{\lambda}_{i}^{N}(s)\neq0\}}}{\rho(|\bar{\lambda}_{i}^{N}(s)|)}d\langle\bar{\lambda}_{i}^{N}\rangle_{s}=4\int_{0}^{t\wedge\eta_{m}}\frac{\mathbf{1}_{\{\bar{\lambda}_{i}^{N}(s)\neq0\}}|\bar{f}_{i}(s)|^{2}\sigma_{s}^{2}}{\rho(|\bar{\lambda}_{i}^{N}(s)|)}ds\leq4\bar{\sigma}^{2}t<\infty. ∫ 0 t ∧ η m ρ ( ∣ λ ˉ i N ( s ) ∣ ) 1 { λ ˉ i N ( s ) = 0 } d ⟨ λ ˉ i N ⟩ s = 4 ∫ 0 t ∧ η m ρ ( ∣ λ ˉ i N ( s ) ∣ ) 1 { λ ˉ i N ( s ) = 0 } ∣ f ˉ i ( s ) ∣ 2 σ s 2 d s ≤ 4 σ ˉ 2 t < ∞.
By [32 Lemma 3.3, Chapter IX] , we get that L t ∧ η m 0 ( λ ˉ i N ) = 0 L_{t\wedge\eta_{m}}^{0}(\bar{\lambda}_{i}^{N})=0 L t ∧ η m 0 ( λ ˉ i N ) = 0 , t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , P θ P_{\theta} P θ -a.s.
Consequently, Tanaka’s formula and (60) imply
∣ λ ˉ i N ( t ∧ η m ) ∣ = 2 ∫ 0 t ∧ η m sgn ( λ ˉ i N ( s ) ) f ˉ i ( s ) σ s d W ˉ s i + ∫ 0 t ∧ η m sgn ( λ ˉ i N ( s ) ) ( b ˉ i ( s ) + F ˉ i ( s ) ) σ s 2 d s . (62) \left|\bar{\lambda}^{N}_{i}(t\wedge\eta_m)\right|=2\int_{0}^{t\wedge\eta_m}\operatorname{sgn}\left(\bar{\lambda}^{N}_{i}(s)\right)\bar{f}_{i}(s)\sigma_s\,d\bar{W}^{i}_s+\int_{0}^{t\wedge\eta_m}\operatorname{sgn}\left(\bar{\lambda}^{N}_{i}(s)\right)\left(\bar{b}_{i}(s)+\bar{F}_{i}(s)\right)\sigma_s^2ds. \tag*{(62)} λ ˉ i N ( t ∧ η m ) = 2 ∫ 0 t ∧ η m sgn ( λ ˉ i N ( s ) ) f ˉ i ( s ) σ s d W ˉ s i + ∫ 0 t ∧ η m sgn ( λ ˉ i N ( s ) ) ( b ˉ i ( s ) + F ˉ i ( s ) ) σ s 2 d s . (62)
Since ∣ λ ˉ i N ( s ) ∣ ∨ ∣ λ ˉ ~ i N ( s ) ∣ ≤ m \left|\bar{\lambda}^{N}_{i}(s)\right|\vee\left|\widetilde{\bar{\lambda}}^{N}_{i}(s)\right|\le m λ ˉ i N ( s ) ∨ λ ˉ i N ( s ) ≤ m on [ 0 , η m ] [0,\eta_m] [ 0 , η m ] , then ∣ f ˉ i ( s ) ∣ 2 ≤ ρ ( ∣ λ ˉ i N ( s ) ∣ ) ≤ ρ ( 2 m ) \left|\bar{f}_{i}(s)\right|^2\le\rho\left(\left|\bar{\lambda}^{N}_{i}(s)\right|\right)\le\rho(2m) f ˉ i ( s ) 2 ≤ ρ ( λ ˉ i N ( s ) ) ≤ ρ ( 2 m ) , s ≤ η m s\le\eta_m s ≤ η m , and we get
E P θ [ ∫ 0 T ∧ η m 4 f ˉ i 2 ( s ) σ s 2 d s ] ≤ 4 σ 2 ρ ( 2 m ) T < ∞ . E_{P_\theta}\left[\int_{0}^{T\wedge\eta_m}4\bar{f}_i^2(s)\sigma_s^2ds\right]\le4\sigma^2\rho(2m)T<\infty. E P θ [ ∫ 0 T ∧ η m 4 f ˉ i 2 ( s ) σ s 2 d s ] ≤ 4 σ 2 ρ ( 2 m ) T < ∞.
Thus the stochastic integral in (62) is a martingale. Taking expectation of (62) , we obtain by (61) ,
E P θ [ ∣ λ ˉ i N ( t ∧ η m ) ∣ ] ≤ C m ∫ 0 t E P θ [ ∣ λ ˉ i N ( s ∧ η m ) ∣ ] d s < ∞ . E_{P_\theta}\left[\left|\bar{\lambda}^{N}_{i}(t\wedge\eta_m)\right|\right]\le C_m\int_{0}^{t}E_{P_\theta}\left[\left|\bar{\lambda}^{N}_{i}(s\wedge\eta_m)\right|\right]ds<\infty. E P θ [ λ ˉ i N ( t ∧ η m ) ] ≤ C m ∫ 0 t E P θ [ λ ˉ i N ( s ∧ η m ) ] d s < ∞.
Summing over i = 1 , … , N i=1,\ldots,N i = 1 , … , N , we get
E P θ [ ∣ λ ˉ N ( t ∧ η m ) ∣ ] ≤ N C m ∫ 0 t E P θ [ ∣ λ ˉ N ( s ∧ η m ) ∣ ] d s < ∞ , E_{P_\theta}\left[\left|\bar{\lambda}^{N}(t\wedge\eta_m)\right|\right]\le NC_m\int_{0}^{t}E_{P_\theta}\left[\left|\bar{\lambda}^{N}(s\wedge\eta_m)\right|\right]ds<\infty, E P θ [ λ ˉ N ( t ∧ η m ) ] ≤ N C m ∫ 0 t E P θ [ λ ˉ N ( s ∧ η m ) ] d s < ∞ ,
and thus E P θ [ ∣ λ ˉ N ( t ∧ η m ) ∣ ] = 0 E_{P_\theta}\left[\left|\bar{\lambda}^{N}(t\wedge\eta_m)\right|\right]=0 E P θ [ λ ˉ N ( t ∧ η m ) ] = 0 for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] by Gronwall’s inequality. Therefore,
λ i N ( t ∧ η m ) = λ ~ i N ( t ∧ η m ) , i = 1 , … , N , 0 ≤ t ≤ T , P θ -a.s. \lambda_i^N(t\wedge\eta_m)=\widetilde{\lambda}_i^N(t\wedge\eta_m),\quad i=1,\ldots,N,\quad0\le t\le T,\quad P_\theta\text{-a.s.} λ i N ( t ∧ η m ) = λ i N ( t ∧ η m ) , i = 1 , … , N , 0 ≤ t ≤ T , P θ -a.s.
Letting m → ∞ m\to\infty m → ∞ , we get that that λ N \lambda^N λ N and λ ~ N \widetilde{\lambda}^N λ N coincide P θ P_\theta P θ -a.s. This completes the proof.
Discussion of -Dyson Brownian motion
By Theorem 1.2 , the G G G -Dyson Brownian motion (9) with λ N ( 0 ) ∈ Δ N \lambda^N(0)\in\Delta_N λ N ( 0 ) ∈ Δ N has a unique strong solution. Below we will prove that, similar to classical Dyson Brownian motion (see [1 ] ), it also admits a unique strong solution when λ N ( 0 ) ∈ Δ ‾ N \lambda^N(0)\in\overline{\Delta}_N λ N ( 0 ) ∈ Δ N . To this end, we first establish the following comparison result.
Lemma 4.1. Assume Assumption 1 for the underlying G G G -Brownian motion. Let λ N ( t ) \lambda^N(t) λ N ( t ) and η N ( t ) \eta^N(t) η N ( t ) be two solutions to the G G G -SDE (9) starting from λ N ( 0 ) ∈ Δ N \lambda^N(0)\in\Delta_N λ N ( 0 ) ∈ Δ N and η N ( 0 ) ∈ Δ N \eta^N(0)\in\Delta_N η N ( 0 ) ∈ Δ N , respectively. Assume λ i N ( 0 ) < η i N ( 0 ) \lambda_i^N(0)<\eta_i^N(0) λ i N ( 0 ) < η i N ( 0 ) for each i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } . Then,
c ( λ i N ( t ) < η i N ( t ) for all t ≥ 0 and i = 1 , … , N ) = 1. c\left(\lambda_i^N(t)<\eta_i^N(t)\text{ for all }t\ge0\text{ and }i=1,\ldots,N\right)=1. c ( λ i N ( t ) < η i N ( t ) for all t ≥ 0 and i = 1 , … , N ) = 1.
Proof. The argument essentially follows from the proof of [1 Lemma 4.3.6] , which is provided here for the reader’s convenience.
First, we have d ( ∑ i λ i N ( t ) − ∑ i η i N ( t ) ) = 0 d\left(\sum_i\lambda_i^N(t)-\sum_i\eta_i^N(t)\right)=0 d ( ∑ i λ i N ( t ) − ∑ i η i N ( t ) ) = 0 , since the interaction terms cancel pairwise in the summation, noting that ⟨ B ˉ i ⟩ t = ⟨ B ˉ 1 ⟩ t \langle\bar{B}_i\rangle_t=\langle\bar{B}_1\rangle_t ⟨ B ˉ i ⟩ t = ⟨ B ˉ 1 ⟩ t for all i = 1 , … , N i=1,\ldots,N i = 1 , … , N . Consequently, we get
∑ i ( λ i N ( t ) − η i N ( t ) ) = ∑ i ( λ i N ( 0 ) − η i N ( 0 ) ) < 0. (63) \sum_i\left(\lambda_i^N(t)-\eta_i^N(t)\right)=\sum_i\left(\lambda_i^N(0)-\eta_i^N(0)\right)<0. \tag*{(63)} i ∑ ( λ i N ( t ) − η i N ( t ) ) = i ∑ ( λ i N ( 0 ) − η i N ( 0 ) ) < 0. (63)
On the other hand, for each i = 1 , … , N i=1,\ldots,N i = 1 , … , N , we have
d ( λ i N ( t ) − η i N ( t ) ) = 1 N ∑ j : j ≠ i η i N ( t ) − λ i N ( t ) − η j N ( t ) + λ j N ( t ) ( η i N ( t ) − η j N ( t ) ) ( λ i N ( t ) − λ j N ( t ) ) d ⟨ B ˉ 1 ⟩ t . (64) d\left(\lambda_i^N(t)-\eta_i^N(t)\right)=\frac{1}{N}\sum_{j:j\ne i}\frac{\eta_i^N(t)-\lambda_i^N(t)-\eta_j^N(t)+\lambda_j^N(t)}{\left(\eta_i^N(t)-\eta_j^N(t)\right)\left(\lambda_i^N(t)-\lambda_j^N(t)\right)}\,d\langle\bar{B}_1\rangle_t. \tag*{(64)} d ( λ i N ( t ) − η i N ( t ) ) = N 1 j : j = i ∑ ( η i N ( t ) − η j N ( t ) ) ( λ i N ( t ) − λ j N ( t ) ) η i N ( t ) − λ i N ( t ) − η j N ( t ) + λ j N ( t ) d ⟨ B ˉ 1 ⟩ t . (64)
Define the first contact time
T : = inf { t > 0 : λ i N ( t ) − η i N ( t ) = 0 for some i ∈ { 1 , … , N } } . T:=\inf\left\{t>0:\lambda_i^N(t)-\eta_i^N(t)=0\text{ for some }i\in\{1,\ldots,N\}\right\}. T := inf { t > 0 : λ i N ( t ) − η i N ( t ) = 0 for some i ∈ { 1 , … , N } } .
If T < ∞ T<\infty T < ∞ , then by the continuity of the paths, we get λ i N ( t ) − η i N ( t ) < 0 \lambda_i^N(t)-\eta_i^N(t)<0 λ i N ( t ) − η i N ( t ) < 0 , t < T t<T t < T and for all i = 1 , … , N i=1,\ldots,N i = 1 , … , N . Let i 0 i_0 i 0 be such that λ i 0 N ( T ) − η i 0 N ( T ) = 0 \lambda_{i_0}^N(T)-\eta_{i_0}^N(T)=0 λ i 0 N ( T ) − η i 0 N ( T ) = 0 . Then λ j N ( T ) − η j N ( T ) ≤ 0 \lambda_j^N(T)-\eta_j^N(T)\le0 λ j N ( T ) − η j N ( T ) ≤ 0 , j = 1 , … , N j=1,\ldots,N j = 1 , … , N . Moreover, by (63) , these differences cannot be all zero. Hence, there exists j 0 ≠ i 0 j_0\ne i_0 j 0 = i 0 such that λ j 0 N ( T ) − η j 0 N ( T ) < 0 \lambda_{j_0}^{N}(T)-\eta_{j_0}^{N}(T)<0 λ j 0 N ( T ) − η j 0 N ( T ) < 0 . Since both λ N \lambda^{N} λ N and η N \eta^{N} η N remain in the strictly ordered region Δ \Delta Δ , the denominator in (64) is strictly positive. At t = T t=T t = T , the numerator corresponding to j 0 j_0 j 0 is given by
λ i 0 N ( T ) − λ i 0 N ( T ) − η j 0 N ( T ) + λ j 0 N ( T ) = λ j 0 N ( T ) − η j 0 N ( T ) < 0. \lambda_{i_0}^{N}(T)-\lambda_{i_0}^{N}(T)-\eta_{j_0}^{N}(T)+\lambda_{j_0}^{N}(T)=\lambda_{j_0}^{N}(T)-\eta_{j_0}^{N}(T)<0. λ i 0 N ( T ) − λ i 0 N ( T ) − η j 0 N ( T ) + λ j 0 N ( T ) = λ j 0 N ( T ) − η j 0 N ( T ) < 0.
Therefore, by continuity we have, for t < T t<T t < T sufficiently close to T T T , d ( λ i 0 N ( t ) − η j 0 N ( t ) ) < 0 \mathrm{d}(\lambda_{i_0}^{N}(t)-\eta_{j_0}^{N}(t))<0 d ( λ i 0 N ( t ) − η j 0 N ( t )) < 0 and thus λ i 0 N ( t ) − η j 0 N ( t ) > 0 \lambda_{i_0}^{N}(t)-\eta_{j_0}^{N}(t)>0 λ i 0 N ( t ) − η j 0 N ( t ) > 0 . This contradicts with the definition of T T T . Thus, T = ∞ T=\infty T = ∞ q.s., and the desired result follows. □
We now establish the existence and uniqueness of a solution to (9) for any initial condition λ N ( 0 ) ∈ Δ ‾ N \lambda^{N}(0)\in\overline{\Delta}_{N} λ N ( 0 ) ∈ Δ N , thereby relaxing the assumption λ N ( 0 ) ∈ Δ N \lambda^{N}(0)\in\Delta_{N} λ N ( 0 ) ∈ Δ N in Theorem 1.3 . It turns out that, as in the classical Dyson Brownian motion, the particles of the G G G -Dyson Brownian motion separate instantaneously after the initial time.
Proof of Proposition 1.4. The proof essentially follows the proof of [1 Proposition 4.3.5] . For n ∈ N n\in\mathbb{N} n ∈ N , denote
λ i N , n ( 0 ) : = λ i N ( 0 ) + i n , for i = 1 , … , N . \lambda_{i}^{N,n}(0):=\lambda_{i}^{N}(0)+\frac{i}{n},\qquad\text{for }i=1,\ldots,N. λ i N , n ( 0 ) := λ i N ( 0 ) + n i , for i = 1 , … , N .
Then λ N , n ( 0 ) ∈ Δ N \lambda^{N,n}(0)\in\Delta_{N} λ N , n ( 0 ) ∈ Δ N . Moreover, for every n ≥ 1 n\geq1 n ≥ 1 ,
λ i N , − n ( 0 ) < λ i N , − n − 1 ( 0 ) < λ i N , n + 1 ( 0 ) < λ i N , n ( 0 ) , i = 1 , … , N . \lambda_{i}^{N,-n}(0)<\lambda_{i}^{N,-n-1}(0)<\lambda_{i}^{N,n+1}(0)<\lambda_{i}^{N,n}(0),\qquad i=1,\ldots,N. λ i N , − n ( 0 ) < λ i N , − n − 1 ( 0 ) < λ i N , n + 1 ( 0 ) < λ i N , n ( 0 ) , i = 1 , … , N .
Let λ N , n \lambda^{N,n} λ N , n denote the solution to (9) starting from λ N , n ( 0 ) \lambda^{N,n}(0) λ N , n ( 0 ) . By Lemma 4.1 , we have
λ i N , − n ( t ) < λ i N , − n − 1 ( t ) < λ i N , n + 1 ( t ) < λ i N , n ( t ) , i = 1 , … , N , q.s. , \lambda_{i}^{N,-n}(t)<\lambda_{i}^{N,-n-1}(t)<\lambda_{i}^{N,n+1}(t)<\lambda_{i}^{N,n}(t),\qquad i=1,\ldots,N,\quad\text{q.s.}, λ i N , − n ( t ) < λ i N , − n − 1 ( t ) < λ i N , n + 1 ( t ) < λ i N , n ( t ) , i = 1 , … , N , q.s. ,
for each t > 0 t>0 t > 0 . Furthermore, since the interaction terms cancel after summation, we have
∑ i = 1 N ( λ i N , n ( t ) − λ i N , − n ( t ) ) = ∑ i = 1 N ( λ i N , n ( 0 ) − λ i N , − n ( 0 ) ) , t ≥ 0. (65) \sum_{i=1}^{N}\left(\lambda_{i}^{N,n}(t)-\lambda_{i}^{N,-n}(t)\right)=\sum_{i=1}^{N}\left(\lambda_{i}^{N,n}(0)-\lambda_{i}^{N,-n}(0)\right),\qquad t\geq0. \tag*{(65)} i = 1 ∑ N ( λ i N , n ( t ) − λ i N , − n ( t ) ) = i = 1 ∑ N ( λ i N , n ( 0 ) − λ i N , − n ( 0 ) ) , t ≥ 0. (65)
the right-hand side of which converges to zero as n → ∞ n\to\infty n → ∞ . Since
0 ≤ λ i N , n ( t ) − λ i N , − n ( t ) ≤ ∑ j = 1 N ( λ j N , n ( t ) − λ j N , − n ( t ) ) , 0\leq\lambda_{i}^{N,n}(t)-\lambda_{i}^{N,-n}(t)\leq\sum_{j=1}^{N}\left(\lambda_{j}^{N,n}(t)-\lambda_{j}^{N,-n}(t)\right), 0 ≤ λ i N , n ( t ) − λ i N , − n ( t ) ≤ j = 1 ∑ N ( λ j N , n ( t ) − λ j N , − n ( t ) ) ,
it follows that both { λ N , n } n ≥ 1 \{\lambda^{N,n}\}_{n\geq1} { λ N , n } n ≥ 1 and { λ N , − n } n ≥ 1 \{\lambda^{N,-n}\}_{n\geq1} { λ N , − n } n ≥ 1 are Cauchy sequences on any compact time interval, converging uniformly to the same limit denoted by λ N \lambda^{N} λ N . In particular, λ N ∈ C ( R + ; Δ ‾ N ) \lambda^{N}\in C(\mathbb{R}_{+};\overline{\Delta}_{N}) λ N ∈ C ( R + ; Δ N ) .
Now, we show that
λ N ( t ) ∈ Δ N for all t > 0 q.s. (66) \lambda^{N}(t)\in\Delta_{N}\quad\text{for all }t>0\text{ q.s.} \tag*{(66)} λ N ( t ) ∈ Δ N for all t > 0 q.s. (66)
Noting that by Theorem 1.1 , once ( λ i N , i = 1 , … , N ) (\lambda_{i}^{N},i=1,\ldots,N) ( λ i N , i = 1 , … , N ) separate, they do not contact again, suppose on the contrary that there exists t ′ > 0 t'>0 t ′ > 0 such that λ N ( s ) ∈ Δ ‾ N ∖ Δ N \lambda^{N}(s)\in\overline{\Delta}_{N}\setminus\Delta_{N} λ N ( s ) ∈ Δ N ∖ Δ N , 0 < s ≤ t ′ 0<s\leq t' 0 < s ≤ t ′ . Then, we can find i 0 ∈ { 2 , … , N } i_{0}\in\{2,\ldots,N\} i 0 ∈ { 2 , … , N } and t ∈ ( 0 , t ′ ] t\in(0,t'] t ∈ ( 0 , t ′ ] such that λ k N ( s ) < λ k + 1 N ( s ) \lambda_{k}^{N}(s)<\lambda_{k+1}^{N}(s) λ k N ( s ) < λ k + 1 N ( s ) for k ≥ i 0 k\geq i_{0} k ≥ i 0 and λ i 0 − 1 N ( s ) = λ i 0 N ( s ) \lambda_{i_{0}-1}^{N}(s)=\lambda_{i_{0}}^{N}(s) λ i 0 − 1 N ( s ) = λ i 0 N ( s ) , 0 < s ≤ t 0<s\leq t 0 < s ≤ t . More precisely, there exist a constant C > 0 C>0 C > 0 and a sequence ε n ↓ 0 \varepsilon_{n}\downarrow0 ε n ↓ 0 such that, for n n n sufficiently large,
∣ λ k N , n ( s ) − λ k + 1 N , n ( s ) ∣ ≥ C , k ≥ i 0 , 0 < s ≤ t , \left|\lambda_{k}^{N,n}(s)-\lambda_{k+1}^{N,n}(s)\right|\geq C,\qquad k\geq i_{0},\quad0<s\leq t, λ k N , n ( s ) − λ k + 1 N , n ( s ) ≥ C , k ≥ i 0 , 0 < s ≤ t ,
and
∣ λ i 0 N , n ( s ) − λ i 0 − 1 N , n ( s ) ∣ ≤ ε n , 0 < s ≤ t . \left|\lambda_{i_{0}}^{N,n}(s)-\lambda_{i_{0}-1}^{N,n}(s)\right|\leq\varepsilon_{n},\qquad0<s\leq t. λ i 0 N , n ( s ) − λ i 0 − 1 N , n ( s ) ≤ ε n , 0 < s ≤ t .
By (9) we get
λ i 0 − 1 N , n ( s ) ≥ λ i 0 − 1 N , n ( 0 ) + 2 N B i 0 − 1 ( s ) + 1 N ( 1 ε n − N − i 0 C ) ⟨ B i 0 − 1 ⟩ s . \lambda_{i_{0}-1}^{N,n}(s)\geq\lambda_{i_{0}-1}^{N,n}(0)+\sqrt{\frac{2}{N}}\mathcal{B}_{i_{0}-1}(s)+\frac{1}{N}\left(\frac{1}{\varepsilon_{n}}-\frac{N-i_{0}}{C}\right)\langle\mathcal{B}_{i_{0}-1}\rangle_{s}. λ i 0 − 1 N , n ( s ) ≥ λ i 0 − 1 N , n ( 0 ) + N 2 B i 0 − 1 ( s ) + N 1 ( ε n 1 − C N − i 0 ) ⟨ B i 0 − 1 ⟩ s .
Letting n → ∞ n \to\infty n → ∞ , we obtain λ i 0 − 1 N , n ( s ) → ∞ \lambda_{i_0-1}^{N,n}(s) \to\infty λ i 0 − 1 N , n ( s ) → ∞ , q.s., which contradicts the uniform bound (67) established below.
Denote C N ( n , t ) : = 1 N ∑ i = 1 N ( λ i N , n ( t ) ) 2 C_N(n,t) := \frac{1}{N}\sum_{i=1}^{N}\left(\lambda_i^{N,n}(t)\right)^2 C N ( n , t ) := N 1 ∑ i = 1 N ( λ i N , n ( t ) ) 2 . We claim
sup n ≥ n ′ sup s ∈ [ 0 , t ] C N ( n , s ) < ∞ , q.s. (67) \sup_{n\ge n'}\sup_{s\in[0,t]} \sqrt{C_N(n,s)} < \infty,\qquad\text{q.s.} \tag*{(67)} n ≥ n ′ sup s ∈ [ 0 , t ] sup C N ( n , s ) < ∞ , q.s. (67)
Let n ≥ n ′ n \ge n' n ≥ n ′ . By Lemma 4.1, λ i N , n ( s ) ≥ λ i N , n ′ ( s ) \lambda_i^{N,n}(s) \ge\lambda_i^{N,n'}(s) λ i N , n ( s ) ≥ λ i N , n ′ ( s ) , i = 1 , … , N i=1,\ldots,N i = 1 , … , N . Moreover, using the triangle inequality in R N \mathbb{R}^N R N , together with (65) , we obtain
∣ C N ( n , s ) − C N ( n ′ , s ) ∣ ≤ ( 1 N ∑ i = 1 N ( λ i N , n ( s ) − λ i N , n ′ ( s ) ) 2 ) 1 / 2 ≤ 1 N ∑ i = 1 N ( λ i N , n ( s ) − λ i N , n ′ ( s ) ) = 1 N ∑ i = 1 N ( λ i N , n ( 0 ) − λ i N , n ′ ( 0 ) ) . \begin{aligned}
\left|\sqrt{C_N(n,s)}-\sqrt{C_N(n',s)}\right|
\le\left(\frac{1}{N}\sum_{i=1}^{N}\left(\lambda_i^{N,n}(s)-\lambda_i^{N,n'}(s)\right)^2\right)^{1/2} \\
&\le\frac{1}{\sqrt{N}}\sum_{i=1}^{N}\left(\lambda_i^{N,n}(s)-\lambda_i^{N,n'}(s)\right) \\
&= \frac{1}{\sqrt{N}}\sum_{i=1}^{N}\left(\lambda_i^{N,n}(0)-\lambda_i^{N,n'}(0)\right).
\end{aligned} C N ( n , s ) − C N ( n ′ , s ) ≤ ( N 1 i = 1 ∑ N ( λ i N , n ( s ) − λ i N , n ′ ( s ) ) 2 ) 1/2 ≤ N 1 i = 1 ∑ N ( λ i N , n ( s ) − λ i N , n ′ ( s ) ) = N 1 i = 1 ∑ N ( λ i N , n ( 0 ) − λ i N , n ′ ( 0 ) ) .
Consequently,
sup n ≥ n ′ sup s ∈ [ 0 , t ] C N ( n , s ) ≤ sup s ∈ [ 0 , t ] C N ( n ′ , s ) + 1 N sup n ≥ n ′ ∑ i = 1 N ( λ i N , n ( 0 ) − λ i N , n ′ ( 0 ) ) . \sup_{n\ge n'}\sup_{s\in[0,t]}\sqrt{C_N(n,s)}
\le\sup_{s\in[0,t]}\sqrt{C_N(n',s)}
+\frac{1}{\sqrt{N}}\sup_{n\ge n'}\sum_{i=1}^{N}\left(\lambda_i^{N,n}(0)-\lambda_i^{N,n'}(0)\right). n ≥ n ′ sup s ∈ [ 0 , t ] sup C N ( n , s ) ≤ s ∈ [ 0 , t ] sup C N ( n ′ , s ) + N 1 n ≥ n ′ sup i = 1 ∑ N ( λ i N , n ( 0 ) − λ i N , n ′ ( 0 ) ) .
Thus, it remains to bound sup s ∈ [ 0 , t ] C N ( n ′ , s ) \sup_{s\in[0,t]}C_N(n',s) sup s ∈ [ 0 , t ] C N ( n ′ , s ) for a fixed n ′ n' n ′ .
G G G -Itô’s formula yields
C N ( n , t ) = D N ( n , t ) + 2 2 N N ∑ i = 1 N ∫ 0 t λ i N , n ( s ) d B ˉ i ( s ) . C_N(n,t)=D_N(n,t)+\frac{2\sqrt{2}}{N\sqrt{N}}\sum_{i=1}^{N}\int_0^t\lambda_i^{N,n}(s)\,\mathrm{d}\bar{B}_i(s). C N ( n , t ) = D N ( n , t ) + N N 2 2 i = 1 ∑ N ∫ 0 t λ i N , n ( s ) d B ˉ i ( s ) .
where
D N ( n , t ) : = C N ( n , 0 ) + 2 N ∑ i = 1 N ∫ 0 t λ i N , n ( s ) 1 N ∑ j : j ≠ i 1 λ i N , n ( s ) − λ j N , n ( s ) d ⟨ B ˉ i ⟩ s + 1 N ∑ i = 1 N ∫ 0 t 2 N d ⟨ B ˉ i ⟩ s . D_N(n,t):=C_N(n,0)+\frac{2}{N}\sum_{i=1}^{N}\int_0^t\lambda_i^{N,n}(s)\frac{1}{N}\sum_{j:j\ne i}\frac{1}{\lambda_i^{N,n}(s)-\lambda_j^{N,n}(s)}\,\mathrm{d}\langle\bar{B}_i\rangle_s+\frac{1}{N}\sum_{i=1}^{N}\int_0^t\frac{2}{N}\,\mathrm{d}\langle\bar{B}_i\rangle_s. D N ( n , t ) := C N ( n , 0 ) + N 2 i = 1 ∑ N ∫ 0 t λ i N , n ( s ) N 1 j : j = i ∑ λ i N , n ( s ) − λ j N , n ( s ) 1 d ⟨ B ˉ i ⟩ s + N 1 i = 1 ∑ N ∫ 0 t N 2 d ⟨ B ˉ i ⟩ s .
By the symmetry of the interaction terms, we get
D N ( n , t ) = C N ( n , 0 ) + N + 1 N ⟨ B ˉ 1 ⟩ t ≤ C N ( n , 0 ) + N + 1 N σ ˉ 2 t . D_N(n,t)=C_N(n,0)+\frac{N+1}{N}\langle\bar{B}_1\rangle_t
\le C_N(n,0)+\frac{N+1}{N}\bar{\sigma}_2t. D N ( n , t ) = C N ( n , 0 ) + N N + 1 ⟨ B ˉ 1 ⟩ t ≤ C N ( n , 0 ) + N N + 1 σ ˉ 2 t .
For R > 0 R>0 R > 0 , define S R : = inf { s ≥ 0 : C N ( n , s ) ≥ R } S_R:=\inf\{s\ge0:C_N(n,s)\ge R\} S R := inf { s ≥ 0 : C N ( n , s ) ≥ R } . By the G G G -Burkholder–Davis–Gundy inequality, there exists a constant C > 0 C>0 C > 0 , independent of R R R , such that
E ^ [ sup s ∈ [ 0 , t ] C N ( n , s ∧ S R ) 2 ] ≤ 2 E ^ [ D N ( n , t ) 2 ] + 2 N − 2 C ∫ 0 t E ^ [ sup r ∈ [ 0 , u ] C N ( n , r ∧ S R ) 2 ] d u \widehat{\mathbb{E}}\left[\sup_{s\in[0,t]}C_N(n,s\wedge S_R)^2\right]
\le2\widehat{\mathbb{E}}\left[D_N(n,t)^2\right]
+2N^{-2}C\int_0^t\widehat{\mathbb{E}}\left[\sup_{r\in[0,u]}C_N(n,r\wedge S_R)^2\right]\mathrm{d}u E [ s ∈ [ 0 , t ] sup C N ( n , s ∧ S R ) 2 ] ≤ 2 E [ D N ( n , t ) 2 ] + 2 N − 2 C ∫ 0 t E [ r ∈ [ 0 , u ] sup C N ( n , r ∧ S R ) 2 ] d u
≤ 2 E ^ [ D N ( n , t ) 2 ] + N − 2 C t + N − 2 C ∫ 0 t E ^ [ sup r ∈ [ 0 , u ] C N ( n , r ∧ S R ) 2 ] d u . \le2\widehat{\mathbb{E}}\left[D_N(n,t)^2\right]+N^{-2}Ct+N^{-2}C\int_0^t\widehat{\mathbb{E}}\left[\sup_{r\in[0,u]}C_N(n,r\wedge S_R)^2\right]\mathrm{d}u. ≤ 2 E [ D N ( n , t ) 2 ] + N − 2 C t + N − 2 C ∫ 0 t E [ r ∈ [ 0 , u ] sup C N ( n , r ∧ S R ) 2 ] d u .
Gronwall’s inequality therefore yields that E ^ [ sup s ∈ [ 0 , t ] C N ( n , s ∧ S R ) 2 ] \widehat{\mathbb{E}}\left[\sup_{s\in[0,t]}C_N(n,s\wedge S_R)^2\right] E [ sup s ∈ [ 0 , t ] C N ( n , s ∧ S R ) 2 ] has a finite upper bound which is independent of R R R . Letting R → ∞ R\to\infty R → ∞ , we conclude that E ^ [ sup s ∈ [ 0 , t ] C N ( n , s ) 2 ] < ∞ \widehat{\mathbb{E}}\left[\sup_{s\in[0,t]}C_N(n,s)^2\right]<\infty E [ sup s ∈ [ 0 , t ] C N ( n , s ) 2 ] < ∞ . This proves (67) , and hence (66) follows.
Finally, we show that λ N \lambda^N λ N is a unique solution to (9) . For any fixed s > 0 s>0 s > 0 , λ N ( s ) ∈ Δ N \lambda^N(s)\in\Delta_N λ N ( s ) ∈ Δ N by (66) , and thus Theorem 1.3 implies that there is a unique solution to (9) starting from λ N ( s ) \lambda^N(s) λ N ( s ) . By the uniform convergence of λ N , n ( t ) → λ N ( t ) \lambda^{N,n}(t)\to\lambda^N(t) λ N , n ( t ) → λ N ( t ) on any compact interval, this solution coincides with ( λ N ( t ) ) t ≥ s (\lambda^N(t))_{t\geq s} ( λ N ( t ) ) t ≥ s . Since s > 0 s>0 s > 0 is arbitrary, λ N \lambda^N λ N is a solution on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , with initial condition λ N ( 0 ) \lambda^N(0) λ N ( 0 ) . The uniqueness follows directly from Theorem 1.3 and (66) . □
Discussion of G G G -Wishart process
Lemma 4.2. Let X t N X^N_t X t N be given by (10) . Assume that the underlying G ~ \widetilde G G -Brownian motion Y Y Y satisfies Assumption 1 with d = p × N \mathrm{d}=p\times N d = p × N . Then X t N X^N_t X t N satisfies the G G G -SDE (11) .
Proof. Noting X i j N ( t ) = ∑ k = 1 p Y k i ( t ) Y k j ( t ) X^N_{ij}(t)=\sum_{k=1}^{p}Y_{ki}(t)Y_{kj}(t) X ij N ( t ) = ∑ k = 1 p Y k i ( t ) Y k j ( t ) , G G G -Itô’s formula yields
d X i j N ( t ) = ∑ k = 1 p ( Y k i ( t ) d Y k j ( t ) + Y k j ( t ) d Y k i ( t ) + d ⟨ Y k i , Y k j ⟩ t ) = ∑ k = 1 p ( Y k i ( t ) d Y k j ( t ) + Y k j ( t ) d Y k i ( t ) ) + p 1 { i = j } d ⟨ Y k i ⟩ t , \begin{aligned}
\mathrm{d}X^N_{ij}(t)
&=\sum_{k=1}^{p}\left(Y_{ki}(t)\mathrm{d}Y_{kj}(t)+Y_{kj}(t)\mathrm{d}Y_{ki}(t)+\mathrm{d}\langle Y_{ki},Y_{kj}\rangle_t\right)\\
&=\sum_{k=1}^{p}\left(Y_{ki}(t)\mathrm{d}Y_{kj}(t)+Y_{kj}(t)\mathrm{d}Y_{ki}(t)\right)+p\mathbf{1}_{\{i=j\}}\mathrm{d}\langle Y_{ki}\rangle_t,
\end{aligned} d X ij N ( t ) = k = 1 ∑ p ( Y k i ( t ) d Y k j ( t ) + Y k j ( t ) d Y k i ( t ) + d ⟨ Y k i , Y k j ⟩ t ) = k = 1 ∑ p ( Y k i ( t ) d Y k j ( t ) + Y k j ( t ) d Y k i ( t ) ) + p 1 { i = j } d ⟨ Y k i ⟩ t ,
which in the matrix form is
d X N ( t ) = Y t T d Y t + d Y t T Y t + p I N d ⟨ Y 11 ⟩ t . (68) \mathrm{d}X^N(t)=Y_t^{\mathsf T}\mathrm{d}Y_t+\mathrm{d}Y_t^{\mathsf T}Y_t+pI_N\mathrm{d}\langle Y_{11}\rangle_t. \tag*{(68)} d X N ( t ) = Y t T d Y t + d Y t T Y t + p I N d ⟨ Y 11 ⟩ t . (68)
Define an N × N N\times N N × N matrix-valued process B ^ t \widehat B_t B t by
d B ^ t : = ( X t N ) † Y t T d Y t , (69) \mathrm{d}\widehat B_t:=\left(\sqrt{X^N_t}\right)^\dagger Y_t^{\mathsf T}\mathrm{d}Y_t, \tag*{(69)} d B t := ( X t N ) † Y t T d Y t , (69)
where ( X t N ) † \left(\sqrt{X^N_t}\right)^\dagger ( X t N ) † is the pseudoinverse of X t N \sqrt{X^N_t} X t N defined as follows: given X t N = Q t T diag ( λ 1 , … , λ N ) Q t X^N_t=Q_t^{\mathsf T}\operatorname{diag}(\lambda_1,\ldots,\lambda_N)Q_t X t N = Q t T diag ( λ 1 , … , λ N ) Q t , the pseudoinverse is X t N † = Q t T diag ( λ 1 − 1 2 1 { λ 1 ≠ 0 } , … , λ N − 1 2 1 { λ N ≠ 0 } ) Q t \sqrt{X^N_t}^{\dagger}=Q_t^{\mathsf T}\operatorname{diag}(\lambda_1^{-\frac12}\mathbf{1}_{\{\lambda_1\ne0\}},\ldots,\lambda_N^{-\frac12}\mathbf{1}_{\{\lambda_N\ne0\}})Q_t X t N † = Q t T diag ( λ 1 − 2 1 1 { λ 1 = 0 } , … , λ N − 2 1 1 { λ N = 0 } ) Q t . Denote
P t : = ( X t N ) † X t N ( X t N ) † = X t N ( X t N ) † . P_t:=\left(\sqrt{X^N_t}\right)^\dagger X^N_t\left(\sqrt{X^N_t}\right)^\dagger=\sqrt{X^N_t}\left(\sqrt{X^N_t}\right)^\dagger. P t := ( X t N ) † X t N ( X t N ) † = X t N ( X t N ) † .
Then, P t = P t T P_t=P_t^{\mathsf T} P t = P t T and P t 2 = P t P_t^2=P_t P t 2 = P t . That is, P t P_t P t is an orthogonal projection. Indeed, P t P_t P t is the orthogonal projection onto ran ( X t ) = ran ( Y t ) = ( ker X t ) ⊥ = ( ker Y t ) ⊥ \operatorname{ran}(X_t)=\operatorname{ran}(Y_t)=(\ker X_t)^\perp=(\ker Y_t)^\perp ran ( X t ) = ran ( Y t ) = ( ker X t ) ⊥ = ( ker Y t ) ⊥ , where ran ( X t ) \operatorname{ran}(X_t) ran ( X t ) denotes the range of X t X_t X t and ker ( X t ) \ker(X_t) ker ( X t ) means the kernel of X t X_t X t . Thus, we have P t X t = X t P_tX_t=X_t P t X t = X t and P t Y t = Y t P_tY_t=Y_t P t Y t = Y t . Therefore, we get
X t N d B ^ t = X t N ( X t N ) † Y t T d Y t = P t Y t T d Y t = Y t T d Y t , d B ^ t T X t N = d Y t T Y t . (70) \sqrt{X^N_t}\mathrm{d}\widehat B_t=\sqrt{X^N_t}\left(\sqrt{X^N_t}\right)^\dagger Y_t^{\mathsf T}\mathrm{d}Y_t=P_tY_t^{\mathsf T}\mathrm{d}Y_t=Y_t^{\mathsf T}\mathrm{d}Y_t,\qquad\mathrm{d}\widehat B_t^{\mathsf T}\sqrt{X^N_t}=\mathrm{d}Y_t^{\mathsf T}Y_t. \tag*{(70)} X t N d B t = X t N ( X t N ) † Y t T d Y t = P t Y t T d Y t = Y t T d Y t , d B t T X t N = d Y t T Y t . (70)
Noting that (69) yields
d B ^ k i = ∑ s = 1 N ∑ n = 1 p ( X t N ) k s † Y n s d Y n i \mathrm{d}\widehat B_{ki}=\sum_{s=1}^{N}\sum_{n=1}^{p}\left(\sqrt{X^N_t}\right)^\dagger_{ks}Y_{ns}\mathrm{d}Y_{ni} d B k i = s = 1 ∑ N n = 1 ∑ p ( X t N ) k s † Y n s d Y ni
and that the underlying G G G -Brownian motion satisfies Assumption 1 with d = p × N d=p\times N d = p × N , we obtain
d ⟨ B ^ i j , B ^ h l ⟩ t = ∑ s = 1 N ∑ n = 1 p ( X t N ) i s † Y n s ∑ r = 1 N ∑ q = 1 p ( X t N ) h r † Y q r d ⟨ Y n j , Y q l ⟩ t = ∑ s = 1 N ∑ r = 1 N ( ∑ n = 1 p Y n s Y n r ) ( X t N ) h r † ( X t N ) i s † d ⟨ Y n j , Y n l ⟩ t = ∑ s = 1 N ∑ r = 1 N ( X t N ) i s † X s r N ( X t N ) h r † d ⟨ Y n j , Y n l ⟩ t = 1 { j = l } ( P t ) i h d ⟨ Y 11 ⟩ t . (71) \begin{aligned}
\mathrm{d}\langle\widehat{B}_{ij},\widehat{B}_{hl}\rangle_t
&=\sum_{s=1}^{N}\sum_{n=1}^{p}\left(\sqrt{X_t^N}\right)_{is}^{\dagger}Y_{ns}\sum_{r=1}^{N}\sum_{q=1}^{p}\left(\sqrt{X_t^N}\right)_{hr}^{\dagger}Y_{qr}\mathrm{d}\langle Y_{nj},Y_{ql}\rangle_t\\
&=\sum_{s=1}^{N}\sum_{r=1}^{N}\left(\sum_{n=1}^{p}Y_{ns}Y_{nr}\right)\left(\sqrt{X_t^N}\right)_{hr}^{\dagger}\left(\sqrt{X_t^N}\right)_{is}^{\dagger}\mathrm{d}\langle Y_{nj},Y_{nl}\rangle_t\\
&=\sum_{s=1}^{N}\sum_{r=1}^{N}\left(\sqrt{X_t^N}\right)_{is}^{\dagger}X_{sr}^N\left(\sqrt{X_t^N}\right)_{hr}^{\dagger}\mathrm{d}\langle Y_{nj},Y_{nl}\rangle_t\\
&=\mathbf{1}_{\{j=l\}}(P_t)_{ih}\mathrm{d}\langle Y_{11}\rangle_t. \tag*{(71)}
\end{aligned} d ⟨ B ij , B h l ⟩ t = s = 1 ∑ N n = 1 ∑ p ( X t N ) i s † Y n s r = 1 ∑ N q = 1 ∑ p ( X t N ) h r † Y q r d ⟨ Y nj , Y q l ⟩ t = s = 1 ∑ N r = 1 ∑ N ( n = 1 ∑ p Y n s Y n r ) ( X t N ) h r † ( X t N ) i s † d ⟨ Y nj , Y n l ⟩ t = s = 1 ∑ N r = 1 ∑ N ( X t N ) i s † X sr N ( X t N ) h r † d ⟨ Y nj , Y n l ⟩ t = 1 { j = l } ( P t ) ih d ⟨ Y 11 ⟩ t . (71)
Now we define
d B t : = d B ^ t + ( I N − P t ) d B ~ t . (72) \mathrm{d}B_t:=\mathrm{d}\widehat{B}_t+(I_N-P_t)\mathrm{d}\widetilde{B}_t. \tag*{(72)} d B t := d B t + ( I N − P t ) d B t . (72)
where B ~ t \widetilde{B}_t B t is an N × N N\times N N × N G ~ \widetilde{G} G -Brownian matrix with G ~ \widetilde{G} G satisfying Assumption 1 with d = N × N d=N\times N d = N × N and satisfies ⟨ B ^ i j , B ~ h l ⟩ t = 0 \langle\widehat{B}_{ij},\widetilde{B}_{hl}\rangle_t=0 ⟨ B ij , B h l ⟩ t = 0 and ⟨ B ~ i j , B ~ h l ⟩ t = 1 { ( i , j ) = ( h , l ) } ⟨ Y 11 ⟩ t \langle\widetilde{B}_{ij},\widetilde{B}_{hl}\rangle_t=\mathbf{1}_{\{(i,j)=(h,l)\}}\langle Y_{11}\rangle_t ⟨ B ij , B h l ⟩ t = 1 {( i , j ) = ( h , l )} ⟨ Y 11 ⟩ t for 1 ≤ i , j , h , l ≤ N 1\leq i,j,h,l\leq N 1 ≤ i , j , h , l ≤ N . It is easy to know that ( I N − P t ) 2 = I N − P t (I_N-P_t)^2=I_N-P_t ( I N − P t ) 2 = I N − P t and X t N ( I N − P t ) = 0 \sqrt{X_t^N}(I_N-P_t)=0 X t N ( I N − P t ) = 0 . By (70) we get
X t N d B t = X t N d B ^ t = Y t T d Y t , d B t T X t N = d B ^ t T X t N = d Y t T Y t . \sqrt{X_t^N}\mathrm{d}B_t=\sqrt{X_t^N}\mathrm{d}\widehat{B}_t=Y_t^{\mathsf{T}}\mathrm{d}Y_t,\qquad
\mathrm{d}B_t^{\mathsf{T}}\sqrt{X_t^N}=\mathrm{d}\widehat{B}_t^{\mathsf{T}}\sqrt{X_t^N}=\mathrm{d}Y_t^{\mathsf{T}}Y_t. X t N d B t = X t N d B t = Y t T d Y t , d B t T X t N = d B t T X t N = d Y t T Y t .
This together with (68) implies the desired (11) , and it remains to show that B B B given by (72) is a G G G -Brownian motion satisfying Assumption 1.
Noting that (72) yields d B i j = d B ^ i j + ∑ k = 1 N ( I N − P t ) i k d B ~ k j \mathrm{d}B_{ij}=\mathrm{d}\widehat{B}_{ij}+\sum_{k=1}^{N}(I_N-P_t)_{ik}\mathrm{d}\widetilde{B}_{kj} d B ij = d B ij + ∑ k = 1 N ( I N − P t ) ik d B k j , by (71) we get
d ⟨ B i j , B h l ⟩ t = d ⟨ B ^ i j , B ^ h l ⟩ t + ∑ k = 1 N ∑ s = 1 N ( I N − P t ) i k ( I N − P t ) h s d ⟨ B ~ k j , B ~ s l ⟩ t = 1 { j = l } ( P t ) i h d ⟨ Y 11 ⟩ t + 1 { j = l } ∑ k = 1 N ( I N − P t ) i k ( I N − P t ) h k d ⟨ Y 11 ⟩ t = 1 { j = l } ( P t ) i h d ⟨ Y 11 ⟩ t + 1 { j = l } ( I N − P t ) i h d ⟨ Y 11 ⟩ t = 1 { ( i , j ) = ( h , l ) } ⟨ Y 11 ⟩ t . \begin{aligned}
\mathrm{d}\langle B_{ij},B_{hl}\rangle_t
&=\mathrm{d}\langle\widehat{B}_{ij},\widehat{B}_{hl}\rangle_t+\sum_{k=1}^{N}\sum_{s=1}^{N}(I_N-P_t)_{ik}(I_N-P_t)_{hs}\mathrm{d}\langle\widetilde{B}_{kj},\widetilde{B}_{sl}\rangle_t\\
&=\mathbf{1}_{\{j=l\}}(P_t)_{ih}\mathrm{d}\langle Y_{11}\rangle_t+\mathbf{1}_{\{j=l\}}\sum_{k=1}^{N}(I_N-P_t)_{ik}(I_N-P_t)_{hk}\mathrm{d}\langle Y_{11}\rangle_t\\
&=\mathbf{1}_{\{j=l\}}(P_t)_{ih}\mathrm{d}\langle Y_{11}\rangle_t+\mathbf{1}_{\{j=l\}}(I_N-P_t)_{ih}\mathrm{d}\langle Y_{11}\rangle_t\\
&=\mathbf{1}_{\{(i,j)=(h,l)\}}\langle Y_{11}\rangle_t.
\end{aligned} d ⟨ B ij , B h l ⟩ t = d ⟨ B ij , B h l ⟩ t + k = 1 ∑ N s = 1 ∑ N ( I N − P t ) ik ( I N − P t ) h s d ⟨ B k j , B s l ⟩ t = 1 { j = l } ( P t ) ih d ⟨ Y 11 ⟩ t + 1 { j = l } k = 1 ∑ N ( I N − P t ) ik ( I N − P t ) hk d ⟨ Y 11 ⟩ t = 1 { j = l } ( P t ) ih d ⟨ Y 11 ⟩ t + 1 { j = l } ( I N − P t ) ih d ⟨ Y 11 ⟩ t = 1 {( i , j ) = ( h , l )} ⟨ Y 11 ⟩ t .
By Corollary 2.14, B B B is an N × N N\times N N × N G G G -Brownian matrix with G G G satisfying Assumption 1 with d = N × N d=N\times N d = N × N . This completes the proof. □
The following result shows that when α ≥ N + 1 \alpha\geq N+1 α ≥ N + 1 , the solution to the G G G -SDE (12) with positive initial eigenvalues is positive definite.
Proposition 4.3. Let X N X^N X N be a solution to the G G G -SDE (12) with x 0 ∈ S N + + x_0\in S_N^{++} x 0 ∈ S N ++ . When α ≥ N + 1 \alpha\geq N+1 α ≥ N + 1 , we have E ( ∃ t ∈ [ 0 , T ] : det X t N ≤ 0 ) = 0 \mathop{\mathbb{E}}(\exists t\in[0,T]:\det X_t^N\leq0)=0 E ( ∃ t ∈ [ 0 , T ] : det X t N ≤ 0 ) = 0 for any T > 0 T>0 T > 0 .
Proof. Let κ m : = inf { t ≥ 0 : det X t N ≤ 1 m or Tr ( X t N ) ≥ m } ∧ T \kappa_m:=\inf\{t\geq0:\det X_t^N\leq\frac{1}{m}\text{ or }\operatorname{Tr}(X_t^N)\geq m\}\wedge T κ m := inf { t ≥ 0 : det X t N ≤ m 1 or Tr ( X t N ) ≥ m } ∧ T . Then, to prove the desired result, it suffices to prove lim m → ∞ E ( κ m < T ) = 0 \lim_{m\to\infty}\mathop{\mathbb{E}(\kappa_m<T)}=0 lim m → ∞ E ( κ m < T ) = 0 .
Noting that
d X i j N = ∑ k = 1 N X i k N d B k j + ∑ k = 1 N d B k i X k j N + α 1 { i = j } d ⟨ B 11 ⟩ t , \mathrm{d}X_{ij}^N=\sum_{k=1}^{N}\sqrt{X_{ik}^N}\,\mathrm{d}B_{kj}+\sum_{k=1}^{N}\mathrm{d}B_{ki}\sqrt{X_{kj}^N}+\alpha\mathbf{1}_{\{i=j\}}\mathrm{d}\langle B_{11}\rangle_t, d X ij N = k = 1 ∑ N X ik N d B k j + k = 1 ∑ N d B k i X k j N + α 1 { i = j } d ⟨ B 11 ⟩ t ,
we have
d ⟨ X i j N , X k l N ⟩ t = ∑ m , n = 1 N ⟨ X i m N d B m j + X j m N d B m i , X k n N d B n l + X l n N d B n k ⟩ t = ∑ m , n = 1 N [ X i m N X k n N d ⟨ B m j , B n l ⟩ t + X i m N X l n N d ⟨ B m j , B n k ⟩ t + X j m N X k n N d ⟨ B m i , B n l ⟩ t + X j m N X l n N d ⟨ B m i , B n k ⟩ t ] = ∑ m = 1 N [ X i m N X k m N 1 { j = l } + X i m N X l m N 1 { j = k } + X j m N X k m N 1 { i = l } + X j m N X l m N 1 { i = k } ] d ⟨ B 11 ⟩ t = ( X i k N 1 { j = l } + X i l N 1 { j = k } + X j k N 1 { i = l } + X j l N 1 { i = k } ) d ⟨ B 11 ⟩ t . \begin{aligned}
\mathrm{d}\langle X^N_{ij},X^N_{kl}\rangle_t
&=\sum_{m,n=1}^{N}\left\langle\sqrt{X^N_{im}}\mathrm{d}B_{mj}+\sqrt{X^N_{jm}}\mathrm{d}B_{mi},\sqrt{X^N_{kn}}\mathrm{d}B_{nl}+\sqrt{X^N_{ln}}\mathrm{d}B_{nk}\right\rangle_t\\
&=\sum_{m,n=1}^{N}\left[\sqrt{X^N_{im}}\sqrt{X^N_{kn}}\mathrm{d}\langle B_{mj},B_{nl}\rangle_t+\sqrt{X^N_{im}}\sqrt{X^N_{ln}}\mathrm{d}\langle B_{mj},B_{nk}\rangle_t\right.\\
&\qquad\left.+\sqrt{X^N_{jm}}\sqrt{X^N_{kn}}\mathrm{d}\langle B_{mi},B_{nl}\rangle_t+\sqrt{X^N_{jm}}\sqrt{X^N_{ln}}\mathrm{d}\langle B_{mi},B_{nk}\rangle_t\right]\\
&=\sum_{m=1}^{N}\left[\sqrt{X^N_{im}}\sqrt{X^N_{km}}\mathbf{1}_{\{j=l\}}+\sqrt{X^N_{im}}\sqrt{X^N_{lm}}\mathbf{1}_{\{j=k\}}\right.\\
&\qquad\left.+\sqrt{X^N_{jm}}\sqrt{X^N_{km}}\mathbf{1}_{\{i=l\}}+\sqrt{X^N_{jm}}\sqrt{X^N_{lm}}\mathbf{1}_{\{i=k\}}\right]\mathrm{d}\langle B_{11}\rangle_t\\
&=\left(X^N_{ik}\mathbf{1}_{\{j=l\}}+X^N_{il}\mathbf{1}_{\{j=k\}}+X^N_{jk}\mathbf{1}_{\{i=l\}}+X^N_{jl}\mathbf{1}_{\{i=k\}}\right)\mathrm{d}\langle B_{11}\rangle_t.
\end{aligned} d ⟨ X ij N , X k l N ⟩ t = m , n = 1 ∑ N ⟨ X im N d B mj + X j m N d B mi , X k n N d B n l + X l n N d B nk ⟩ t = m , n = 1 ∑ N [ X im N X k n N d ⟨ B mj , B n l ⟩ t + X im N X l n N d ⟨ B mj , B nk ⟩ t + X j m N X k n N d ⟨ B mi , B n l ⟩ t + X j m N X l n N d ⟨ B mi , B nk ⟩ t ] = m = 1 ∑ N [ X im N X k m N 1 { j = l } + X im N X l m N 1 { j = k } + X j m N X k m N 1 { i = l } + X j m N X l m N 1 { i = k } ] d ⟨ B 11 ⟩ t = ( X ik N 1 { j = l } + X i l N 1 { j = k } + X j k N 1 { i = l } + X j l N 1 { i = k } ) d ⟨ B 11 ⟩ t .
Denote
V ( X N ) : = Tr ( X N ) − ln ( det X N ) = ∑ i = 1 N ( λ i N − ln λ i N ) . V(X^N):=\operatorname{Tr}(X^N)-\ln(\det X^N)=\sum_{i=1}^{N}\left(\lambda_i^N-\ln\lambda_i^N\right). V ( X N ) := Tr ( X N ) − ln ( det X N ) = i = 1 ∑ N ( λ i N − ln λ i N ) .
We have
∂ V ∂ ( X N ) i j = 1 { i = j } − ( X N ) j i − 1 , ∂ 2 V ∂ X i j N ∂ X k l N = ( X N ) i l − 1 ( X N ) k j − 1 . \frac{\partial V}{\partial(X^N)_{ij}}=\mathbf{1}_{\{i=j\}}-(X^N)^{-1}_{ji},\qquad
\frac{\partial^2V}{\partial X^N_{ij}\,\partial X^N_{kl}}=(X^N)^{-1}_{il}(X^N)^{-1}_{kj}. ∂ ( X N ) ij ∂ V = 1 { i = j } − ( X N ) j i − 1 , ∂ X ij N ∂ X k l N ∂ 2 V = ( X N ) i l − 1 ( X N ) k j − 1 .
Applying G G G -Itô’s formula to V ( X t N ) V(X^N_t) V ( X t N ) , we get
d V ( X t N ) = ∑ i , j = 1 N ∂ V ∂ X i j N ( ∑ k = 1 N X i k N d B k j + ∑ k = 1 N d B k i X k j N + α 1 { i = j } d ⟨ B i i ⟩ t ) + 1 2 ∑ i , j , k , l = 1 N ∂ 2 V ∂ X i j N ∂ X k l N ( X i k N 1 { j = l } + X i l N 1 { j = k } + X j k N 1 { i = l } + X j l N 1 { i = k } ) d ⟨ B 11 ⟩ t = ∑ i , j = 1 N ∂ V ∂ X i j N ( ∑ k = 1 N X i k N d B k j + ∑ k = 1 N d B k i X k j N ) + α ∑ i = 1 N ( 1 − ( X N ) i i − 1 ) d ⟨ B 11 ⟩ t + 1 2 ∑ i , j , k , l = 1 N ( X N ) i l − 1 ( X N ) k j − 1 ( X i k N 1 { j = l } + X i l N 1 { j = k } + X j k N 1 { i = l } + X j l N 1 { i = k } ) d ⟨ B 11 ⟩ t . \begin{aligned}
\mathrm{d}V(X^N_t)
&=\sum_{i,j=1}^{N}\frac{\partial V}{\partial X^N_{ij}}\left(\sum_{k=1}^{N}\sqrt{X^N_{ik}}\mathrm{d}B_{kj}+\sum_{k=1}^{N}\mathrm{d}B_{ki}\sqrt{X^N_{kj}}+\alpha\mathbf{1}_{\{i=j\}}\mathrm{d}\langle B_{ii}\rangle_t\right)\\
&\quad+\frac{1}{2}\sum_{i,j,k,l=1}^{N}\frac{\partial^2V}{\partial X^N_{ij}\partial X^N_{kl}}\left(X^N_{ik}\mathbf{1}_{\{j=l\}}+X^N_{il}\mathbf{1}_{\{j=k\}}+X^N_{jk}\mathbf{1}_{\{i=l\}}+X^N_{jl}\mathbf{1}_{\{i=k\}}\right)\mathrm{d}\langle B_{11}\rangle_t\\
&=\sum_{i,j=1}^{N}\frac{\partial V}{\partial X^N_{ij}}\left(\sum_{k=1}^{N}\sqrt{X^N_{ik}}\mathrm{d}B_{kj}+\sum_{k=1}^{N}\mathrm{d}B_{ki}\sqrt{X^N_{kj}}\right)+\alpha\sum_{i=1}^{N}\left(1-(X^N)^{-1}_{ii}\right)\mathrm{d}\langle B_{11}\rangle_t\\
&\quad+\frac{1}{2}\sum_{i,j,k,l=1}^{N}(X^N)^{-1}_{il}(X^N)^{-1}_{kj}\left(X^N_{ik}\mathbf{1}_{\{j=l\}}+X^N_{il}\mathbf{1}_{\{j=k\}}+X^N_{jk}\mathbf{1}_{\{i=l\}}+X^N_{jl}\mathbf{1}_{\{i=k\}}\right)\mathrm{d}\langle B_{11}\rangle_t.
\end{aligned} d V ( X t N ) = i , j = 1 ∑ N ∂ X ij N ∂ V ( k = 1 ∑ N X ik N d B k j + k = 1 ∑ N d B k i X k j N + α 1 { i = j } d ⟨ B ii ⟩ t ) + 2 1 i , j , k , l = 1 ∑ N ∂ X ij N ∂ X k l N ∂ 2 V ( X ik N 1 { j = l } + X i l N 1 { j = k } + X j k N 1 { i = l } + X j l N 1 { i = k } ) d ⟨ B 11 ⟩ t = i , j = 1 ∑ N ∂ X ij N ∂ V ( k = 1 ∑ N X ik N d B k j + k = 1 ∑ N d B k i X k j N ) + α i = 1 ∑ N ( 1 − ( X N ) ii − 1 ) d ⟨ B 11 ⟩ t + 2 1 i , j , k , l = 1 ∑ N ( X N ) i l − 1 ( X N ) k j − 1 ( X ik N 1 { j = l } + X i l N 1 { j = k } + X j k N 1 { i = l } + X j l N 1 { i = k } ) d ⟨ B 11 ⟩ t .
The second term can simplified as
1 2 ∑ i , j , k = 1 N ( X N ) i j − 1 ( X N ) k j − 1 X i k N + 1 2 ∑ i , j , l = 1 N ( X N ) i l − 1 ( X N ) j j − 1 X i l N + 1 2 ∑ i , j , k = 1 N ( X N ) i i − 1 ( X N ) k j − 1 X j k N + 1 2 ∑ i , j , l = 1 N ( X N ) i l − 1 ( X N ) i j − 1 X j l N = 1 2 ∑ j , k = 1 N ( I N ) j k ( X N ) k j − 1 + 1 2 Tr ( I N ) Tr ( ( X N ) − 1 ) + 1 2 Tr ( ( X N ) − 1 ) Tr ( I N ) + 1 2 ∑ i , l = 1 N ( I N ) i l ( X N ) l i − 1 = ( N + 1 ) Tr ( ( X N ) − 1 ) . \begin{aligned}
&\frac{1}{2}\sum_{i,j,k=1}^{N}(X^N)^{-1}_{ij}(X^N)^{-1}_{kj}X^N_{ik}
+\frac{1}{2}\sum_{i,j,l=1}^{N}(X^N)^{-1}_{il}(X^N)^{-1}_{jj}X^N_{il}\\
&\quad+\frac{1}{2}\sum_{i,j,k=1}^{N}(X^N)^{-1}_{ii}(X^N)^{-1}_{kj}X^N_{jk}
+\frac{1}{2}\sum_{i,j,l=1}^{N}(X^N)^{-1}_{il}(X^N)^{-1}_{ij}X^N_{jl}\\
&=\frac{1}{2}\sum_{j,k=1}^{N}(I_N)_{jk}(X^N)^{-1}_{kj}
+\frac{1}{2}\operatorname{Tr}(I_N)\operatorname{Tr}((X^N)^{-1})
+\frac{1}{2}\operatorname{Tr}((X^N)^{-1})\operatorname{Tr}(I_N)
+\frac{1}{2}\sum_{i,l=1}^{N}(I_N)_{il}(X^N)^{-1}_{li} \\
&=(N+1)\operatorname{Tr}\left((X^{N})^{-1}\right).
\end{aligned} 2 1 i , j , k = 1 ∑ N ( X N ) ij − 1 ( X N ) k j − 1 X ik N + 2 1 i , j , l = 1 ∑ N ( X N ) i l − 1 ( X N ) j j − 1 X i l N + 2 1 i , j , k = 1 ∑ N ( X N ) ii − 1 ( X N ) k j − 1 X j k N + 2 1 i , j , l = 1 ∑ N ( X N ) i l − 1 ( X N ) ij − 1 X j l N = 2 1 j , k = 1 ∑ N ( I N ) j k ( X N ) k j − 1 + 2 1 Tr ( I N ) Tr (( X N ) − 1 ) + 2 1 Tr (( X N ) − 1 ) Tr ( I N ) + 2 1 i , l = 1 ∑ N ( I N ) i l ( X N ) l i − 1 = ( N + 1 ) Tr ( ( X N ) − 1 ) .
Thus, we get
d V ( X t N ) = ∑ i , j = 1 N ∂ V ∂ X i j N ( ∑ k = 1 N X i k N d B k j + ∑ k = 1 N d B k i X k j N ) + [ α N − ( α − ( N + 1 ) ) Tr ( ( X t N ) − 1 ) ] d ⟨ B 11 ⟩ t . \mathrm{d}V(X_t^N)=\sum_{i,j=1}^{N}\frac{\partial V}{\partial X_{ij}^N}\left(\sum_{k=1}^{N}\sqrt{X_{ik}^N}\,\mathrm{d}B_{kj}+\sum_{k=1}^{N}\mathrm{d}B_{ki}\sqrt{X_{kj}^N}\right)+\left[\alpha N-(\alpha-(N+1))\operatorname{Tr}\left((X_t^N)^{-1}\right)\right]\mathrm{d}\langle B_{11}\rangle_t. d V ( X t N ) = i , j = 1 ∑ N ∂ X ij N ∂ V ( k = 1 ∑ N X ik N d B k j + k = 1 ∑ N d B k i X k j N ) + [ α N − ( α − ( N + 1 )) Tr ( ( X t N ) − 1 ) ] d ⟨ B 11 ⟩ t .
Note that X t N X_t^N X t N is positive definite for t < κ m t<\kappa_m t < κ m and so is its inverse ( X t N ) − 1 (X_t^N)^{-1} ( X t N ) − 1 . Thus, when α ≥ N + 1 \alpha\geq N+1 α ≥ N + 1 , we have for t < κ m t<\kappa_m t < κ m ,
d V ( X t N ) ≤ d M t V + α N d ⟨ B 11 ⟩ t , \mathrm{d}V(X_t^N)\leq\mathrm{d}M_t^V+\alpha N\mathrm{d}\langle B_{11}\rangle_t, d V ( X t N ) ≤ d M t V + α N d ⟨ B 11 ⟩ t ,
where
d M t V = ∑ i , j = 1 N ∂ V ∂ X i j N ( ∑ k = 1 N X i k N d B k j + ∑ k = 1 N d B k i X k j N ) . \mathrm{d}M_t^V=\sum_{i,j=1}^{N}\frac{\partial V}{\partial X_{ij}^N}\left(\sum_{k=1}^{N}\sqrt{X_{ik}^N}\,\mathrm{d}B_{kj}+\sum_{k=1}^{N}\mathrm{d}B_{ki}\sqrt{X_{kj}^N}\right). d M t V = i , j = 1 ∑ N ∂ X ij N ∂ V ( k = 1 ∑ N X ik N d B k j + k = 1 ∑ N d B k i X k j N ) .
Thus, we have
E [ V ( X κ m N ) ] ≤ E [ V ( X 0 N ) ] + α N σ ˉ 2 T . (73) \mathbb{E}\left[V(X_{\kappa_m}^N)\right]\leq\mathbb{E}\left[V(X_0^N)\right]+\alpha N\bar{\sigma}^{2}T. \tag*{(73)} E [ V ( X κ m N ) ] ≤ E [ V ( X 0 N ) ] + α N σ ˉ 2 T . (73)
Now we study the behavior of V ( X t N ) V(X_t^N) V ( X t N ) on the event { κ m < T } \{\kappa_m<T\} { κ m < T } , where either det X t N = 1 m \det X_t^N=\frac{1}{m} det X t N = m 1 or Tr ( X t N ) = m \operatorname{Tr}(X_t^N)=m Tr ( X t N ) = m for some t = κ m < T t=\kappa_m<T t = κ m < T . If det X t N = 1 m \det X_t^N=\frac{1}{m} det X t N = m 1 , then we have V ( X t N ) = Tr ( X t N ) − ln det X t N > ln m V(X_t^N)=\operatorname{Tr}(X_t^N)-\ln\det X_t^N>\ln m V ( X t N ) = Tr ( X t N ) − ln det X t N > ln m , since X t N ∈ S + + N X_t^N\in S_{++}^N X t N ∈ S ++ N . If Tr ( X t N ) = m \operatorname{Tr}(X_t^N)=m Tr ( X t N ) = m , noting that 1 N ln det X t N = 1 N ∑ i = 1 N ln λ i N ≤ ln ( 1 N ∑ i = 1 N λ i N ) = ln ( Tr ( X t N ) N ) \frac{1}{N}\ln\det X_t^N=\frac{1}{N}\sum_{i=1}^{N}\ln\lambda_i^N\leq\ln\left(\frac{1}{N}\sum_{i=1}^{N}\lambda_i^N\right)=\ln\left(\frac{\operatorname{Tr}(X_t^N)}{N}\right) N 1 ln det X t N = N 1 ∑ i = 1 N ln λ i N ≤ ln ( N 1 ∑ i = 1 N λ i N ) = ln ( N Tr ( X t N ) ) , we have V ( X t N ) ≥ m − N ln m N V(X_t^N)\geq m-N\ln\frac{m}{N} V ( X t N ) ≥ m − N ln N m .
Thus, we have V ( X κ m N ) ≥ g ( m ) V(X_{\kappa_m}^N)\geq g(m) V ( X κ m N ) ≥ g ( m ) on the event { κ m < T } \{\kappa_m<T\} { κ m < T } , where g ( m ) : = min { ln m , m − N ln ( m / N ) } g(m):=\min\{\ln m,m-N\ln(m/N)\} g ( m ) := min { ln m , m − N ln ( m / N )} . Consequently, by (73) we get
P ( κ m < T ) ≤ P ( V ( X κ m N ) ≥ g ( m ) ) ≤ E [ V ( X 0 N ) ] + α N σ ˉ 2 T g ( m ) . \mathbb{P}(\kappa_m<T)\leq\mathbb{P}\left(V(X_{\kappa_m}^N)\geq g(m)\right)\leq\frac{\mathbb{E}[V(X_0^N)]+\alpha N\bar{\sigma}^{2}T}{g(m)}. P ( κ m < T ) ≤ P ( V ( X κ m N ) ≥ g ( m ) ) ≤ g ( m ) E [ V ( X 0 N )] + α N σ ˉ 2 T .
Since lim m → ∞ g ( m ) = ∞ \lim_{m\to\infty}g(m)=\infty lim m → ∞ g ( m ) = ∞ , taking the limit yields lim m → ∞ P ( κ m < T ) = 0 \lim_{m\to\infty}\mathbb{P}(\kappa_m<T)=0 lim m → ∞ P ( κ m < T ) = 0 . This completes the proof. □ \square □
Now, we are ready to prove Theorem 1.5 .
Proof of Theorem 1.5. For m ≥ 1 m\geq1 m ≥ 1 , define
D m : = { Y N ∈ S + + N : det Y N > 1 m and Tr ( Y N ) < m } . D_m:=\left\{Y^N\in S_{++}^N:\det Y^N>\frac{1}{m}\ \text{and}\ \operatorname{Tr}(Y^N)<m\right\}. D m := { Y N ∈ S ++ N : det Y N > m 1 and Tr ( Y N ) < m } .
Let ψ m : S N → [ 0 , 1 ] \psi_m:S_N\to[0,1] ψ m : S N → [ 0 , 1 ] be a Lipschitz continuous cutoff function satisfying:
ψ m ( Y N ) = { 1 , Y N ∈ D m , 0 , Y N ∉ D m + 1 . \psi_m(Y^N)=
\begin{cases}
1, & Y^N\in D_m,\\
0, & Y^N\notin D_{m+1}.
\end{cases} ψ m ( Y N ) = { 1 , 0 , Y N ∈ D m , Y N ∈ / D m + 1 .
We define Φ m ( Y N ) : = ψ m ( Y N ) Y N \Phi_m(Y^N):=\psi_m(Y^N)\sqrt{Y^N} Φ m ( Y N ) := ψ m ( Y N ) Y N . Clearly, Φ m \Phi_m Φ m is globally Lipschitz continuous on S N S_N S N . Then the following G G G -SDE
d X ( m ) N = Φ m ( X ( m ) N ) d B t + d B t T Φ m ( X ( m ) N ) + α I N d ⟨ B 11 ⟩ t , X ( m ) N ( 0 ) = x 0 , (74) \mathrm{d}X_{(m)}^N=\Phi_m(X_{(m)}^N)\mathrm{d}B_t+\mathrm{d}B_t^{\mathsf{T}}\Phi_m(X_{(m)}^N)+\alpha I_N\mathrm{d}\langle B_{11}\rangle_t,\qquad X_{(m)}^N(0)=x_0, \tag*{(74)} d X ( m ) N = Φ m ( X ( m ) N ) d B t + d B t T Φ m ( X ( m ) N ) + α I N d ⟨ B 11 ⟩ t , X ( m ) N ( 0 ) = x 0 , (74)
admits a unique solution ( X ( m ) N ( t ) ) t ∈ [ 0 , T ] ∈ ( M G 2 ( 0 , T ) ) d (X_{(m)}^N(t))_{t\in[0,T]}\in(M_G^2(0,T))^d ( X ( m ) N ( t ) ) t ∈ [ 0 , T ] ∈ ( M G 2 ( 0 , T ) ) d with d = N × N d=N\times N d = N × N by Theorem 5.1.3 in [30 ] .
Denote κ m ( m ) : = inf { t ≥ 0 : X ( m ) N ( t ) ∉ D m } ∧ T \kappa_m^{(m)}:=\inf\{t\geq0:X_{(m)}^N(t)\notin D_m\}\wedge T κ m ( m ) := inf { t ≥ 0 : X ( m ) N ( t ) ∈ / D m } ∧ T . Then we have X ( m ) N ( t ) = X ( m + k ) N ( t ) X_{(m)}^N(t)=X_{(m+k)}^N(t) X ( m ) N ( t ) = X ( m + k ) N ( t ) for t ∈ [ 0 , κ m ( m ) ] t\in[0,\kappa_m^{(m)}] t ∈ [ 0 , κ m ( m ) ] for all k ∈ N k\in\mathbb{N} k ∈ N by the uniqueness of the solution, and thus κ m ( m ) \kappa_m^{(m)} κ m ( m ) is increasing in m m m .
Then, we set X t N : = lim m → ∞ X ( m ) N ( t ) X_t^N:=\lim_{m\to\infty}X_{(m)}^N(t) X t N := lim m → ∞ X ( m ) N ( t ) . By a similar argument as in the proof of Proposition 4.3 , we can prove that lim m → ∞ κ m ( m ) = T \lim_{m\to\infty}\kappa_m^{(m)}=T lim m → ∞ κ m ( m ) = T q.s., and hence X t N X_t^N X t N is defined on the total interval [ 0 , T ] [0,T] [ 0 , T ] . Set X ~ ( m ) N ( t ) : = X ( m ) N ( t ) 1 [ 0 , κ m ( m ) ] ( t ) \widetilde{X}_{(m)}^{N}(t):=X_{(m)}^{N}(t)\mathbf{1}_{[0,\kappa_{m}^{(m)}]}(t) X ( m ) N ( t ) := X ( m ) N ( t ) 1 [ 0 , κ m ( m ) ] ( t ) . Then, X ~ ( m ) N ( t ) ∈ ( L G 2 ( Ω t ) ) d \widetilde{X}_{(m)}^{N}(t)\in(L_{G}^{2}(\Omega_t))^{d} X ( m ) N ( t ) ∈ ( L G 2 ( Ω t ) ) d with d = N × N d=N\times N d = N × N , noting that κ m ( m ) \kappa_{m}^{(m)} κ m ( m ) is q.c. by [25 Example 5.1] . Thus, in order to prove X t N ∈ ( L G 2 ( Ω t ) ) d X_t^{N}\in(L_{G}^{2}(\Omega_t))^{d} X t N ∈ ( L G 2 ( Ω t ) ) d with d = N × N d=N\times N d = N × N , it suffices to show
lim m → ∞ E [ ∣ X t N − X ~ ( m ) N ( t ) ∣ 2 ] = 0. (75) \lim_{m\to\infty}\mathbb{E}\left[\left|X_t^{N}-\widetilde{X}_{(m)}^{N}(t)\right|^{2}\right]=0. \tag*{(75)} m → ∞ lim E [ X t N − X ( m ) N ( t ) 2 ] = 0. (75)
For a matrix A = ( A i j ) ∈ R N × N A=(A_{ij})\in\mathbb{R}^{N\times N} A = ( A ij ) ∈ R N × N , denote its Frobenius norm by ∣ A ∣ 2 : = Tr ( A T A ) = ∑ i , j = 1 N ∣ A i j ∣ 2 |A|^{2}:=\operatorname{Tr}(A^{\mathsf{T}}A)=\sum_{i,j=1}^{N}|A_{ij}|^{2} ∣ A ∣ 2 := Tr ( A T A ) = ∑ i , j = 1 N ∣ A ij ∣ 2 . Then, we have
E [ ∣ X t N − X ~ ( m ) N ( t ) ∣ 2 ] = E [ ∣ ( X t N − X ~ ( m ) N ( t ) ) 1 [ κ m ( m ) = T ] + ( X t N − X ~ ( m ) N ( t ) ) 1 [ κ m ( m ) < T ] ∣ 2 ] = E [ ∣ ( X t N − X ~ ( m ) N ( t ) ) 1 [ κ m ( m ) < T ] ∣ 2 ] ≤ E [ ∣ X t N ∣ 2 1 [ κ m ( m ) < T ] ] ≤ ( E [ ∣ X N ( t ) ∣ 4 ] ) 1 2 ( c ( κ m ( m ) < T ) ) 1 2 ≤ ( lim inf m → ∞ E [ ∣ X ( m ) N ( t ) ∣ 4 ] ) 1 2 ( c ( κ m ( m ) < T ) ) 1 2 . (76) \begin{aligned}
\mathbb{E}\left[\left|X_t^{N}-\widetilde{X}_{(m)}^{N}(t)\right|^{2}\right]
&=\mathbb{E}\left[\left|\left(X_t^{N}-\widetilde{X}_{(m)}^{N}(t)\right)\mathbf{1}_{[\kappa_{m}^{(m)}=T]}+\left(X_t^{N}-\widetilde{X}_{(m)}^{N}(t)\right)\mathbf{1}_{[\kappa_{m}^{(m)}<T]}\right|^{2}\right]\\
&=\mathbb{E}\left[\left|\left(X_t^{N}-\widetilde{X}_{(m)}^{N}(t)\right)\mathbf{1}_{[\kappa_{m}^{(m)}<T]}\right|^{2}\right]\\
&\leq\mathbb{E}\left[|X_t^{N}|^{2}\mathbf{1}_{[\kappa_{m}^{(m)}<T]}\right]\\
&\leq\left(\mathbb{E}\left[|X^{N}(t)|^{4}\right]\right)^{\frac{1}{2}}\left(c(\kappa_{m}^{(m)}<T)\right)^{\frac{1}{2}}\\
&\leq\left(\liminf_{m\to\infty}\mathbb{E}\left[|X_{(m)}^{N}(t)|^{4}\right]\right)^{\frac{1}{2}}\left(c(\kappa_{m}^{(m)}<T)\right)^{\frac{1}{2}}. \tag*{(76)}
\end{aligned} E [ X t N − X ( m ) N ( t ) 2 ] = E [ ( X t N − X ( m ) N ( t ) ) 1 [ κ m ( m ) = T ] + ( X t N − X ( m ) N ( t ) ) 1 [ κ m ( m ) < T ] 2 ] = E [ ( X t N − X ( m ) N ( t ) ) 1 [ κ m ( m ) < T ] 2 ] ≤ E [ ∣ X t N ∣ 2 1 [ κ m ( m ) < T ] ] ≤ ( E [ ∣ X N ( t ) ∣ 4 ] ) 2 1 ( c ( κ m ( m ) < T ) ) 2 1 ≤ ( m → ∞ lim inf E [ ∣ X ( m ) N ( t ) ∣ 4 ] ) 2 1 ( c ( κ m ( m ) < T ) ) 2 1 . (76)
Similar as in in the proof of Proposition 4.3 , we can get that c ( κ m ( m ) < T ) → 0 c(\kappa_{m}^{(m)}<T)\to0 c ( κ m ( m ) < T ) → 0 as m → ∞ m\to\infty m → ∞ . Thus, it suffices to show that E [ ∣ X ( m ) N ( t ) ∣ 4 ] \mathbb{E}[|X_{(m)}^{N}(t)|^{4}] E [ ∣ X ( m ) N ( t ) ∣ 4 ] is uniformly bounded in m m m .
By G-Itô’s formula, we have
d ( ( X ( m ) N ) i j ) 4 ( t ) = 4 ( ( X ( m ) N ) i j ) 3 d ( X ( m ) N ) i j + 6 ( ( X ( m ) N ) i j ) 2 d ⟨ ( X ( m ) N ) i j ⟩ t , \mathrm{d}\left((X_{(m)}^{N})^{ij}\right)^{4}(t)=4\left((X_{(m)}^{N})^{ij}\right)^{3}\mathrm{d}(X_{(m)}^{N})^{ij}+6\left((X_{(m)}^{N})^{ij}\right)^{2}\mathrm{d}\left\langle(X_{(m)}^{N})^{ij}\right\rangle_t, d ( ( X ( m ) N ) ij ) 4 ( t ) = 4 ( ( X ( m ) N ) ij ) 3 d ( X ( m ) N ) ij + 6 ( ( X ( m ) N ) ij ) 2 d ⟨ ( X ( m ) N ) ij ⟩ t ,
where
d ( X ( m ) N ) i j ( t ) = ∑ k = 1 N Φ m i k ( X ( m ) N ) d B k j ( t ) + ∑ k = 1 N Φ m k j ( X ( m ) N ) d B k i ( t ) + α 1 { i = j } d ⟨ B 11 ⟩ t , \mathrm{d}(X_{(m)}^{N})^{ij}(t)=\sum_{k=1}^{N}\Phi_{m}^{ik}(X_{(m)}^{N})\mathrm{d}B_{kj}(t)+\sum_{k=1}^{N}\Phi_{m}^{kj}(X_{(m)}^{N})\mathrm{d}B_{ki}(t)+\alpha\mathbf{1}_{\{i=j\}}\mathrm{d}\langle B_{11}\rangle_t, d ( X ( m ) N ) ij ( t ) = k = 1 ∑ N Φ m ik ( X ( m ) N ) d B k j ( t ) + k = 1 ∑ N Φ m k j ( X ( m ) N ) d B k i ( t ) + α 1 { i = j } d ⟨ B 11 ⟩ t ,
and
d ⟨ ( X ( m ) N ) i j ⟩ t = ∑ k = 1 N ( Φ m i k ( X ( m ) N ) ) 2 d ⟨ B k j ⟩ t + ∑ k = 1 N ( Φ m k j ( X ( m ) N ) ) 2 d ⟨ B k i ⟩ t + 2 ∑ k = 1 N Φ m i k ( X ( m ) N ) Φ m k j ( X ( m ) N ) d ⟨ B k j , B k i ⟩ t = ∑ k = 1 N Φ m i k ( X ( m ) N ) Φ m k i ( X ( m ) N ) d ⟨ B k j ⟩ t + ∑ k = 1 N Φ m j k ( X ( m ) N ) Φ m k j ( X ( m ) N ) d ⟨ B k i ⟩ t + 21 { i = j } ∑ k = 1 N Φ m i k ( X ( m ) N ) Φ m k i ( X ( m ) N ) d ⟨ B k i ⟩ t = [ ( Φ m 2 ( X ( m ) N ) ) i i + ( Φ m 2 ( X ( m ) N ) ) j j + 21 { i = j } ( Φ m 2 ( X ( m ) N ) ) i i ] d ⟨ B 11 ⟩ t . \begin{aligned}
\mathrm{d}\left\langle(X_{(m)}^{N})^{ij}\right\rangle_t
={}&\sum_{k=1}^{N}\left(\Phi_{m}^{ik}(X_{(m)}^{N})\right)^{2}\mathrm{d}\langle B_{kj}\rangle_t+\sum_{k=1}^{N}\left(\Phi_{m}^{kj}(X_{(m)}^{N})\right)^{2}\mathrm{d}\langle B_{ki}\rangle_t\\
&+2\sum_{k=1}^{N}\Phi_{m}^{ik}(X_{(m)}^{N})\Phi_{m}^{kj}(X_{(m)}^{N})\mathrm{d}\langle B_{kj},B_{ki}\rangle_t\\
={}&\sum_{k=1}^{N}\Phi_{m}^{ik}(X_{(m)}^{N})\Phi_{m}^{ki}(X_{(m)}^{N})\mathrm{d}\langle B_{kj}\rangle_t+\sum_{k=1}^{N}\Phi_{m}^{jk}(X_{(m)}^{N})\Phi_{m}^{kj}(X_{(m)}^{N})\mathrm{d}\langle B_{ki}\rangle_t\\
&+2\mathbf{1}_{\{i=j\}}\sum_{k=1}^{N}\Phi_{m}^{ik}(X_{(m)}^{N})\Phi_{m}^{ki}(X_{(m)}^{N})\mathrm{d}\langle B_{ki}\rangle_t\\
={}&\left[(\Phi_{m}^{2}(X_{(m)}^{N}))_{ii}+(\Phi_{m}^{2}(X_{(m)}^{N}))_{jj}+2\mathbf{1}_{\{i=j\}}(\Phi_{m}^{2}(X_{(m)}^{N}))_{ii}\right]\mathrm{d}\langle B_{11}\rangle_t.
\end{aligned} d ⟨ ( X ( m ) N ) ij ⟩ t = = = k = 1 ∑ N ( Φ m ik ( X ( m ) N ) ) 2 d ⟨ B k j ⟩ t + k = 1 ∑ N ( Φ m k j ( X ( m ) N ) ) 2 d ⟨ B k i ⟩ t + 2 k = 1 ∑ N Φ m ik ( X ( m ) N ) Φ m k j ( X ( m ) N ) d ⟨ B k j , B k i ⟩ t k = 1 ∑ N Φ m ik ( X ( m ) N ) Φ m k i ( X ( m ) N ) d ⟨ B k j ⟩ t + k = 1 ∑ N Φ m j k ( X ( m ) N ) Φ m k j ( X ( m ) N ) d ⟨ B k i ⟩ t + 2 1 { i = j } k = 1 ∑ N Φ m ik ( X ( m ) N ) Φ m k i ( X ( m ) N ) d ⟨ B k i ⟩ t [ ( Φ m 2 ( X ( m ) N ) ) ii + ( Φ m 2 ( X ( m ) N ) ) j j + 2 1 { i = j } ( Φ m 2 ( X ( m ) N ) ) ii ] d ⟨ B 11 ⟩ t .
Noting that ( Φ m 2 ( X ( m ) N ) ) i i ≤ ( X ( m ) N ) i i (\Phi_{m}^{2}(X_{(m)}^{N}))_{ii}\leq(X_{(m)}^{N})_{ii} ( Φ m 2 ( X ( m ) N ) ) ii ≤ ( X ( m ) N ) ii , we get
E [ ( ( X ( m ) N ) i j ) 4 ( t ) ] − E [ ( ( X ( m ) N ) i j ) 4 ( 0 ) ] ≤ E ∫ 0 t [ 4 ( ( X ( m ) N ) i j ) 3 α 1 { i = j } + 6 ( ( X ( m ) N ) i j ) 2 [ ( Φ m 2 ( X ( m ) N ) ) i i + ( Φ m 2 ( X ( m ) N ) ) j j + 21 { i = j } ( Φ m 2 ( X ( m ) N ) ) i i ] ] d ⟨ B 11 ⟩ s ≤ C σ ˉ − 2 ∫ 0 t E [ ∥ X ( m ) N ∥ 3 ] d t ≤ C σ ˉ − 2 t + C σ ˉ − 2 ∫ 0 t E [ ∥ X ( m ) N ∥ 4 ] d t . \begin{aligned}
\mathbb{E}\left[\left((X_{(m)}^{N})^{ij}\right)^{4}(t)\right]-\mathbb{E}\left[\left((X_{(m)}^{N})^{ij}\right)^{4}(0)\right]
\leq\mathbb{E}\int_{0}^{t}\bigg[4\left((X_{(m)}^{N})^{ij}\right)^{3}\alpha\mathbf{1}_{\{i=j\}}+6\left((X_{(m)}^{N})^{ij}\right)^{2}\big[ (\Phi_{m}^{2}(X_{(m)}^{N}))_{ii}\\
+(\Phi_{m}^{2}(X_{(m)}^{N}))_{jj}+2\mathbf{1}_{\{i=j\}}(\Phi_{m}^{2}(X_{(m)}^{N}))_{ii}\big]\bigg]\mathrm{d}\langle B_{11}\rangle_s \\
&\le C\bar{\sigma}^{-2}\int_0^t \mathbb{E}\left[\left\|X_{(m)}^N\right\|^3\right]\mathrm{d}t \\
&\le C\bar{\sigma}^{-2}t+C\bar{\sigma}^{-2}\int_0^t \mathbb{E}\left[\left\|X_{(m)}^N\right\|^4\right]\mathrm{d}t.
\end{aligned} E [ ( ( X ( m ) N ) ij ) 4 ( t ) ] − E [ ( ( X ( m ) N ) ij ) 4 ( 0 ) ] ≤ E ∫ 0 t [ 4 ( ( X ( m ) N ) ij ) 3 α 1 { i = j } + 6 ( ( X ( m ) N ) ij ) 2 [ ( Φ m 2 ( X ( m ) N ) ) ii + ( Φ m 2 ( X ( m ) N ) ) j j + 2 1 { i = j } ( Φ m 2 ( X ( m ) N ) ) ii ] ] d ⟨ B 11 ⟩ s ≤ C σ ˉ − 2 ∫ 0 t E [ X ( m ) N 3 ] d t ≤ C σ ˉ − 2 t + C σ ˉ − 2 ∫ 0 t E [ X ( m ) N 4 ] d t .
Then, we have
E [ ∥ X ( m ) N ∥ 4 ] ≤ C + C σ ˉ − 2 ∫ 0 t E [ ∥ X ( m ) N ∥ 4 ] d t , t ∈ [ 0 , T ] \mathbb{E}\left[\left\|X_{(m)}^N\right\|^4\right]\le C+C\bar{\sigma}^{-2}\int_0^t\mathbb{E}\left[\left\|X_{(m)}^N\right\|^4\right]\mathrm{d}t,\qquad t\in[0,T] E [ X ( m ) N 4 ] ≤ C + C σ ˉ − 2 ∫ 0 t E [ X ( m ) N 4 ] d t , t ∈ [ 0 , T ]
for some constants C C C independent of m m m , and Gronwall’s inequality yields that E [ ∥ X ( m ) N ( t ) ∥ 4 ] \mathbb{E}\left[\left\|X_{(m)}^N(t)\right\|^4\right] E [ X ( m ) N ( t ) 4 ] is uniformly bounded in m m m . This with (76) proves (75) , and hence X t N ∈ ( L G 2 ( Ω t ) ) d X_t^N\in\left(L_G^2(\Omega_t)\right)^d X t N ∈ ( L G 2 ( Ω t ) ) d with d = N × N d=N\times N d = N × N .
Finally, the uniqueness of the solution follows directly from the uniqueness of the solution on each [ 0 , κ m ( m ) ] [0,\kappa_m^{(m)}] [ 0 , κ m ( m ) ] and the fact that κ m ( m ) → T \kappa_m^{(m)}\to T κ m ( m ) → T q.s. as m → ∞ m\to\infty m → ∞ . □
High-dimensional spectral limits
In this section, we investigate the high-dimensional limits of the empirical spectral measures of generalized Wishart processes, providing proofs for Theorem 1.7 , Theorem 1.8 , Proposition 1.11, and Proposition 1.14 .
Generalized Wishart Processes
First, we prove the tightness of the empirical spectral measures.
Proof of Theorem 1.7 . Let f ∈ C 2 ( R ) f\in C^2(\mathbb{R}) f ∈ C 2 ( R ) be fixed. For each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , we have
⟨ f , L N ( t ) ⟩ = ∫ R f ( x ) L N ( t ) ( d x ) = 1 N ∑ i = 1 N ∫ R f ( x ) δ λ i N ( t ) ( d x ) = 1 N ∑ i = 1 N f ( λ i N ( t ) ) . \langle f,L_N(t)\rangle=\int_{\mathbb{R}}f(x)L_N(t)(\mathrm{d}x)=\frac{1}{N}\sum_{i=1}^N\int_{\mathbb{R}}f(x)\delta_{\lambda_i^N(t)}(\mathrm{d}x)=\frac{1}{N}\sum_{i=1}^N f\left(\lambda_i^N(t)\right). ⟨ f , L N ( t )⟩ = ∫ R f ( x ) L N ( t ) ( d x ) = N 1 i = 1 ∑ N ∫ R f ( x ) δ λ i N ( t ) ( d x ) = N 1 i = 1 ∑ N f ( λ i N ( t ) ) .
By G G G -Itô’s formula, we get for i = 1 , … , N i=1,\ldots,N i = 1 , … , N ,
f ( λ i N ( t ) ) − f ( λ i N ( 0 ) ) = ∫ 0 t f ′ ( λ i N ( s ) ) 2 g N ( λ i N ) h N ( λ i N ) d B ‾ i + ∫ 0 t f ′ ( λ i N ( s ) ) b N ( λ i N ) d ⟨ B ‾ i ⟩ s + ∫ 0 t f ′ ( λ i N ( s ) ) ∑ k : k ≠ i H N ( λ i N , λ k N ) λ i N − λ k N d ⟨ B ‾ i ⟩ s + 2 ∫ 0 t f ′ ′ ( λ i N ( s ) ) g N ( λ i N ) 2 h N ( λ i N ) 2 d ⟨ B ‾ i ⟩ s . \begin{aligned}
f\left(\lambda_i^N(t)\right)-f\left(\lambda_i^N(0)\right)
={}&\int_0^t f'\left(\lambda_i^N(s)\right)2g_N\left(\lambda_i^N\right)h_N\left(\lambda_i^N\right)\,\mathrm{d}\overline{B}_i\\
&+\int_0^t f'\left(\lambda_i^N(s)\right)b_N\left(\lambda_i^N\right)\,\mathrm{d}\langle\overline{B}_i\rangle_s\\
&+\int_0^t f'\left(\lambda_i^N(s)\right)\sum_{k:k\ne i}\frac{H_N\left(\lambda_i^N,\lambda_k^N\right)}{\lambda_i^N-\lambda_k^N}\,\mathrm{d}\langle\overline{B}_i\rangle_s\\
&+2\int_0^t f''\left(\lambda_i^N(s)\right)g_N\left(\lambda_i^N\right)^2h_N\left(\lambda_i^N\right)^2\,\mathrm{d}\langle\overline{B}_i\rangle_s.
\end{aligned} f ( λ i N ( t ) ) − f ( λ i N ( 0 ) ) = ∫ 0 t f ′ ( λ i N ( s ) ) 2 g N ( λ i N ) h N ( λ i N ) d B i + ∫ 0 t f ′ ( λ i N ( s ) ) b N ( λ i N ) d ⟨ B i ⟩ s + ∫ 0 t f ′ ( λ i N ( s ) ) k : k = i ∑ λ i N − λ k N H N ( λ i N , λ k N ) d ⟨ B i ⟩ s + 2 ∫ 0 t f ′′ ( λ i N ( s ) ) g N ( λ i N ) 2 h N ( λ i N ) 2 d ⟨ B i ⟩ s .
Summing over i i i and dividing by N N N , we obtain
⟨ f , L N ( t ) ⟩ − ⟨ f , L N ( 0 ) ⟩ = 1 N ∑ i = 1 N ∫ 0 t f ′ ( λ i N ( s ) ) 2 g N ( λ i N ) h N ( λ i N ) d B ‾ i + 1 N ∑ i = 1 N ∫ 0 t f ′ ( λ i N ( s ) ) b N ( λ i N ) d ⟨ B ‾ i ⟩ s + 1 N ∑ k ≠ i ∫ 0 t f ′ ( λ i N ( s ) ) H N ( λ i N , λ k N ) λ i N − λ k N d ⟨ B ‾ i ⟩ s + 2 N ∑ i = 1 N ∫ 0 t f ′ ′ ( λ i N ( s ) ) g N ( λ i N ) 2 h N ( λ i N ) 2 d ⟨ B ‾ i ⟩ s = M f N ( t ) + ∫ 0 t ⟨ f ′ b N , L N ( s ) ⟩ d ⟨ B ‾ i ⟩ s + 2 ∫ 0 t ⟨ f ′ ′ g N 2 h N 2 , L N ( s ) ⟩ d ⟨ B ‾ i ⟩ s + 1 N ∑ k ≠ i ∫ 0 t f ′ ( λ i N ( s ) ) H N ( λ i N , λ k N ) λ i N − λ k N d ⟨ B ‾ i ⟩ s . (77) \begin{aligned}
\langle f,L_N(t)\rangle-\langle f,L_N(0)\rangle
={}&\frac{1}{N}\sum_{i=1}^N\int_0^t f'\left(\lambda_i^N(s)\right)2g_N\left(\lambda_i^N\right)h_N\left(\lambda_i^N\right)\,\mathrm{d}\overline{B}_i\\
&+\frac{1}{N}\sum_{i=1}^N\int_0^t f'\left(\lambda_i^N(s)\right)b_N\left(\lambda_i^N\right)\,\mathrm{d}\langle\overline{B}_i\rangle_s\\
&+\frac{1}{N}\sum_{k\ne i}\int_0^t f'\left(\lambda_i^N(s)\right)\frac{H_N\left(\lambda_i^N,\lambda_k^N\right)}{\lambda_i^N-\lambda_k^N}\,\mathrm{d}\langle\overline{B}_i\rangle_s\\
&+\frac{2}{N}\sum_{i=1}^N\int_0^t f''\left(\lambda_i^N(s)\right)g_N\left(\lambda_i^N\right)^2h_N\left(\lambda_i^N\right)^2\,\mathrm{d}\langle\overline{B}_i\rangle_s\\
={}&M_f^N(t)+\int_0^t\langle f'b_N,L_N(s)\rangle\,\mathrm{d}\langle\overline{B}_i\rangle_s+2\int_0^t\langle f''g_N^2h_N^2,L_N(s)\rangle\,\mathrm{d}\langle\overline{B}_i\rangle_s\\
&+\frac{1}{N}\sum_{k\ne i}\int_0^t f'\left(\lambda_i^N(s)\right)\frac{H_N\left(\lambda_i^N,\lambda_k^N\right)}{\lambda_i^N-\lambda_k^N}\,\mathrm{d}\langle\overline{B}_i\rangle_s. \tag*{(77)}
\end{aligned} ⟨ f , L N ( t )⟩ − ⟨ f , L N ( 0 )⟩ = = N 1 i = 1 ∑ N ∫ 0 t f ′ ( λ i N ( s ) ) 2 g N ( λ i N ) h N ( λ i N ) d B i + N 1 i = 1 ∑ N ∫ 0 t f ′ ( λ i N ( s ) ) b N ( λ i N ) d ⟨ B i ⟩ s + N 1 k = i ∑ ∫ 0 t f ′ ( λ i N ( s ) ) λ i N − λ k N H N ( λ i N , λ k N ) d ⟨ B i ⟩ s + N 2 i = 1 ∑ N ∫ 0 t f ′′ ( λ i N ( s ) ) g N ( λ i N ) 2 h N ( λ i N ) 2 d ⟨ B i ⟩ s M f N ( t ) + ∫ 0 t ⟨ f ′ b N , L N ( s )⟩ d ⟨ B i ⟩ s + 2 ∫ 0 t ⟨ f ′′ g N 2 h N 2 , L N ( s )⟩ d ⟨ B i ⟩ s + N 1 k = i ∑ ∫ 0 t f ′ ( λ i N ( s ) ) λ i N − λ k N H N ( λ i N , λ k N ) d ⟨ B i ⟩ s . (77)
where we denote
M f N ( t ) : = 1 N ∑ i = 1 N ∫ 0 t f ′ ( λ i N ( s ) ) 2 g N ( λ i N ( s ) ) h N ( λ i N ( s ) ) d B ˉ i . M_f^N(t) := \frac{1}{N}\sum_{i=1}^{N}\int_0^t f'(\lambda_i^N(s))\,2g_N(\lambda_i^N(s))h_N(\lambda_i^N(s))\,\mathrm{d}\bar{B}_i. M f N ( t ) := N 1 i = 1 ∑ N ∫ 0 t f ′ ( λ i N ( s )) 2 g N ( λ i N ( s )) h N ( λ i N ( s )) d B ˉ i .
We adopt the convention f ′ ( x ) − f ′ ( y ) x − y = f ′ ′ ( x ) \frac{f'(x)-f'(y)}{x-y}=f''(x) x − y f ′ ( x ) − f ′ ( y ) = f ′′ ( x ) , whenever x = y x=y x = y . Using the symmetry of H N H_N H N , the interaction term in (77) can be symmetrized as follows:
1 N ∑ k ≠ i ∫ 0 t f ′ ( λ i N ( s ) ) H N ( λ i N , λ k N ) λ i N − λ k N d ⟨ B ˉ i ⟩ s = 1 2 N ∑ k ≠ i ∫ 0 t f ′ ( λ i N ( s ) ) − f ′ ( λ k N ( s ) ) λ i N ( s ) − λ k N ( s ) H N ( λ i N , λ k N ) d ⟨ B ˉ i ⟩ s = 1 2 N ∑ k ≠ i ∫ 0 t ∬ f ′ ( x ) − f ′ ( y ) x − y H N ( x , y ) δ λ i N ( s ) ( d x ) δ λ k N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s = N 2 ∫ 0 t ∬ f ′ ( x ) − f ′ ( y ) x − y H N ( x , y ) L N ( s ) ( d x ) L N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s \begin{aligned}
\frac{1}{N}\sum_{k\ne i}\int_0^t f'(\lambda_i^N(s))\frac{H_N(\lambda_i^N,\lambda_k^N)}{\lambda_i^N-\lambda_k^N}\,\mathrm{d}\langle\bar{B}_i\rangle_s \\
&=\frac{1}{2N}\sum_{k\ne i}\int_0^t\frac{f'(\lambda_i^N(s))-f'(\lambda_k^N(s))}{\lambda_i^N(s)-\lambda_k^N(s)}H_N(\lambda_i^N,\lambda_k^N)\,\mathrm{d}\langle\bar{B}_i\rangle_s \\
&=\frac{1}{2N}\sum_{k\ne i}\int_0^t\iint\frac{f'(x)-f'(y)}{x-y}H_N(x,y)\,\delta_{\lambda_i^N(s)}(\mathrm{d}x)\delta_{\lambda_k^N(s)}(\mathrm{d}y)\,\mathrm{d}\langle\bar{B}_i\rangle_s \\
&=\frac{N}{2}\int_0^t\iint\frac{f'(x)-f'(y)}{x-y}H_N(x,y)L_N(s)(\mathrm{d}x)L_N(s)(\mathrm{d}y)\,\mathrm{d}\langle\bar{B}_i\rangle_s
\end{aligned} N 1 k = i ∑ ∫ 0 t f ′ ( λ i N ( s )) λ i N − λ k N H N ( λ i N , λ k N ) d ⟨ B ˉ i ⟩ s = 2 N 1 k = i ∑ ∫ 0 t λ i N ( s ) − λ k N ( s ) f ′ ( λ i N ( s )) − f ′ ( λ k N ( s )) H N ( λ i N , λ k N ) d ⟨ B ˉ i ⟩ s = 2 N 1 k = i ∑ ∫ 0 t ∬ x − y f ′ ( x ) − f ′ ( y ) H N ( x , y ) δ λ i N ( s ) ( d x ) δ λ k N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s = 2 N ∫ 0 t ∬ x − y f ′ ( x ) − f ′ ( y ) H N ( x , y ) L N ( s ) ( d x ) L N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s
− 1 2 N ∑ i = 1 N ∫ 0 t ∬ f ′ ( x ) − f ′ ( y ) x − y H N ( x , y ) δ λ i N ( s ) ( d x ) δ λ i N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s = N 2 ∫ 0 t ∬ f ′ ( x ) − f ′ ( y ) x − y H N ( x , y ) L N ( s ) ( d x ) L N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s − ∫ 0 t ⟨ f ′ ′ g N 2 h N 2 , L N ( s ) ⟩ d ⟨ B ˉ i ⟩ s . \begin{aligned}
-\frac{1}{2N}\sum_{i=1}^{N}\int_0^t\iint\frac{f'(x)-f'(y)}{x-y}H_N(x,y)\,\delta_{\lambda_i^N(s)}(\mathrm{d}x)\delta_{\lambda_i^N(s)}(\mathrm{d}y)\,\mathrm{d}\langle\bar{B}_i\rangle_s \\
&=\frac{N}{2}\int_0^t\iint\frac{f'(x)-f'(y)}{x-y}H_N(x,y)L_N(s)(\mathrm{d}x)L_N(s)(\mathrm{d}y)\,\mathrm{d}\langle\bar{B}_i\rangle_s
-\int_0^t\langle f''g_N^2h_N^2,L_N(s)\rangle\,\mathrm{d}\langle\bar{B}_i\rangle_s.
\end{aligned} − 2 N 1 i = 1 ∑ N ∫ 0 t ∬ x − y f ′ ( x ) − f ′ ( y ) H N ( x , y ) δ λ i N ( s ) ( d x ) δ λ i N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s = 2 N ∫ 0 t ∬ x − y f ′ ( x ) − f ′ ( y ) H N ( x , y ) L N ( s ) ( d x ) L N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s − ∫ 0 t ⟨ f ′′ g N 2 h N 2 , L N ( s )⟩ d ⟨ B ˉ i ⟩ s .
Substituting this identity into (77) yields
⟨ f , L N ( t ) ⟩ = ⟨ f , L N ( 0 ) ⟩ + N 2 ∫ 0 t ∬ f ′ ( x ) − f ′ ( y ) x − y H N ( x , y ) L N ( s ) ( d x ) L N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s + ∫ 0 t ⟨ f ′ b N , L N ( s ) ⟩ d ⟨ B ˉ i ⟩ s + ∫ 0 t ⟨ f ′ ′ g N 2 h N 2 , L N ( s ) ⟩ d ⟨ B ˉ i ⟩ s + M f N ( t ) . (78) \begin{aligned}
\langle f,L_N(t)\rangle=\langle f,L_N(0)\rangle+\frac{N}{2}\int_0^t\iint\frac{f'(x)-f'(y)}{x-y}H_N(x,y)L_N(s)(\mathrm{d}x)L_N(s)(\mathrm{d}y)\,\mathrm{d}\langle\bar{B}_i\rangle_s \\
&+\int_0^t\langle f'b_N,L_N(s)\rangle\,\mathrm{d}\langle\bar{B}_i\rangle_s+\int_0^t\langle f''g_N^2h_N^2,L_N(s)\rangle\,\mathrm{d}\langle\bar{B}_i\rangle_s+M_f^N(t). \tag*{(78)}
\end{aligned} ⟨ f , L N ( t )⟩ = ⟨ f , L N ( 0 )⟩ + 2 N ∫ 0 t ∬ x − y f ′ ( x ) − f ′ ( y ) H N ( x , y ) L N ( s ) ( d x ) L N ( s ) ( d y ) d ⟨ B ˉ i ⟩ s + ∫ 0 t ⟨ f ′ b N , L N ( s )⟩ d ⟨ B ˉ i ⟩ s + ∫ 0 t ⟨ f ′′ g N 2 h N 2 , L N ( s )⟩ d ⟨ B ˉ i ⟩ s + M f N ( t ) . (78)
The quadratic variation of M f N ( t ) M_f^N(t) M f N ( t ) has the following estimation:
⟨ M f N ⟩ t = 4 N ∫ 0 t ⟨ ( f ′ ) 2 g N 2 h N 2 , L N ( s ) ⟩ d ⟨ B ˉ i ⟩ s ≤ 4 σ ˉ 2 t ∥ f ′ g N h N ∥ ∞ 2 N . (79) \langle M_f^N\rangle_t=\frac{4}{N}\int_0^t\langle(f')^2g_N^2h_N^2,L_N(s)\rangle\,\mathrm{d}\langle\bar{B}_i\rangle_s\le\frac{4\bar{\sigma}^2t\lVert f'g_Nh_N\rVert_\infty^2}{N}. \tag*{(79)} ⟨ M f N ⟩ t = N 4 ∫ 0 t ⟨( f ′ ) 2 g N 2 h N 2 , L N ( s )⟩ d ⟨ B ˉ i ⟩ s ≤ N 4 σ ˉ 2 t ∥ f ′ g N h N ∥ ∞ 2 . (79)
By Assumption 2, the drift and interaction terms in (78) , with f = φ f=\varphi f = φ , are bounded uniformly with respect to N N N . Let C C C be a constant such that
∥ φ ′ b N ∥ ∞ + ∥ φ ′ ′ g N 2 h N 2 ∥ ∞ + ∥ N H N ( x , y ) φ ′ ( x ) − φ ′ ( y ) x − y ∥ ∞ ≤ C , \lVert\varphi'b_N\rVert_\infty+\lVert\varphi''g_N^2h_N^2\rVert_\infty+\left\lVert NH_N(x,y)\frac{\varphi'(x)-\varphi'(y)}{x-y}\right\rVert_\infty\le C, ∥ φ ′ b N ∥ ∞ + ∥ φ ′′ g N 2 h N 2 ∥ ∞ + N H N ( x , y ) x − y φ ′ ( x ) − φ ′ ( y ) ∞ ≤ C ,
uniformly in N ≥ 1 N\ge1 N ≥ 1 . Hence, by (78) there exists a constant C 1 = C 1 ( T , σ ˉ 2 , C ) C_1=C_1(T,\bar{\sigma}^2,C) C 1 = C 1 ( T , σ ˉ 2 , C ) such that
sup 0 ≤ t ≤ T ⟨ φ , L N ( t ) ⟩ ≤ C 0 + C 1 T + sup 0 ≤ t ≤ T ∣ M φ N ( t ) ∣ . (80) \sup_{0\le t\le T}\langle\varphi,L_N(t)\rangle\le C_0+C_1T+\sup_{0\le t\le T}|M_\varphi^N(t)|. \tag*{(80)} 0 ≤ t ≤ T sup ⟨ φ , L N ( t )⟩ ≤ C 0 + C 1 T + 0 ≤ t ≤ T sup ∣ M φ N ( t ) ∣. (80)
By (79) we get
c ( sup 0 ≤ t ≤ T ∣ M φ N ( t ) ∣ ≥ ε ) ≤ C l 1 ε 2 l 1 E ^ [ ⟨ M φ N ⟩ T l 1 ] ≤ C l 1 , T , σ ˉ ε 2 l 1 ( ∥ φ ′ g N h N ∥ ∞ 2 N ) l 1 . c\left(\sup_{0\le t\le T}|M_\varphi^N(t)|\ge\varepsilon\right)\le\frac{C_{l_1}}{\varepsilon^{2l_1}}\widehat{\mathbb{E}}\left[\langle M_\varphi^N\rangle_T^{l_1}\right]\le\frac{C_{l_1,T,\bar{\sigma}}}{\varepsilon^{2l_1}}\left(\frac{\lVert\varphi'g_Nh_N\rVert_\infty^2}{N}\right)^{l_1}. c ( 0 ≤ t ≤ T sup ∣ M φ N ( t ) ∣ ≥ ε ) ≤ ε 2 l 1 C l 1 E [ ⟨ M φ N ⟩ T l 1 ] ≤ ε 2 l 1 C l 1 , T , σ ˉ ( N ∥ φ ′ g N h N ∥ ∞ 2 ) l 1 .
where the right-hand side is summable in N N N by Assumption 2. Then, Lemma 2.18 together with the estimate (80) implies that there exists a finite constant M M M such that sup 0 ≤ t ≤ T ⟨ φ , L N ( t ) ⟩ ≤ M \sup_{0\le t\le T}\langle\varphi,L_N(t)\rangle\le M sup 0 ≤ t ≤ T ⟨ φ , L N ( t )⟩ ≤ M , for all sufficiently large N N N . Thus,
P ( ⋃ N 0 ≥ 1 ⋂ N ≥ N 0 { L N ( t ) ∈ K M for every t ∈ [ 0 , T ] } ) = 1 , (81) \mathbb{P}\left(\bigcup_{N_0\ge1}\bigcap_{N\ge N_0}\{L_N(t)\in K_M\text{ for every }t\in[0,T]\}\right)=1, \tag*{(81)} P ( N 0 ≥ 1 ⋃ N ≥ N 0 ⋂ { L N ( t ) ∈ K M for every t ∈ [ 0 , T ]} ) = 1 , (81)
where
K M : = { μ ∈ P ( R ) : ⟨ φ , μ ⟩ ≤ M } K_M:=\{\mu\in\mathcal{P}(\mathbb{R}):\langle\varphi,\mu\rangle\le M\} K M := { μ ∈ P ( R ) : ⟨ φ , μ ⟩ ≤ M }
is a compact subset in P ( R ) \mathcal{P}(\mathbb{R}) P ( R ) by Lemma B.2 in [38 ] .
For f ∈ C 2 ( R ) f\in C^2(\mathbb{R}) f ∈ C 2 ( R ) , by (78) we have
∣ ⟨ f , L N ( t ) ⟩ − ⟨ f , L N ( s ) ⟩ ∣ ≤ C 2 σ ˉ 2 ∣ t − s ∣ + ∣ M f N ( t ) − M f N ( s ) ∣ , |\langle f,L_N(t)\rangle-\langle f,L_N(s)\rangle|\le C_2\bar{\sigma}^2|t-s|+|M_f^N(t)-M_f^N(s)|, ∣ ⟨ f , L N ( t )⟩ − ⟨ f , L N ( s )⟩ ∣ ≤ C 2 σ ˉ 2 ∣ t − s ∣ + ∣ M f N ( t ) − M f N ( s ) ∣ ,
where
C 2 : = sup N > 0 { ∥ f ′ b N ∥ ∞ + ∥ f ′ ′ g N 2 h N 2 ∥ ∞ + 1 2 ∥ N H N ( x , y ) f ′ ( x ) − f ′ ( y ) x − y ∥ ∞ } . C_2:=\sup_{N>0}\left\{\|f'b_N\|_\infty+\|f''g_N^2h_N^2\|_\infty+\frac{1}{2}\left\|NH_N(x,y)\frac{f'(x)-f'(y)}{x-y}\right\|_\infty\right\}. C 2 := N > 0 sup { ∥ f ′ b N ∥ ∞ + ∥ f ′′ g N 2 h N 2 ∥ ∞ + 2 1 N H N ( x , y ) x − y f ′ ( x ) − f ′ ( y ) ∞ } .
Thus, when C 2 σ ˉ 2 δ 7 / 8 ≤ 1 C_2\bar{\sigma}^2\delta^{7/8}\le1 C 2 σ ˉ 2 δ 7/8 ≤ 1 for some fixed δ > 0 \delta>0 δ > 0 , we have
P ( sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ ∣ ⟨ f , L N ( t ) ⟩ − ⟨ f , L N ( s ) ⟩ ∣ ≥ M δ 1 / 8 ) ≤ P ( sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ ∣ M f N ( t ) − M f N ( s ) ∣ ≥ ( M − 1 ) δ 1 / 8 ) . (82) \mathbb{P}\left(\sup_{\substack{s,t\in[0,T]\\|t-s|\le\delta}}|\langle f,L_N(t)\rangle-\langle f,L_N(s)\rangle|\ge M\delta^{1/8}\right)\le\mathbb{P}\left(\sup_{\substack{s,t\in[0,T]\\|t-s|\le\delta}}|M_f^N(t)-M_f^N(s)|\ge(M-1)\delta^{1/8}\right). \tag*{(82)} P s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ sup ∣ ⟨ f , L N ( t )⟩ − ⟨ f , L N ( s )⟩ ∣ ≥ M δ 1/8 ≤ P s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ sup ∣ M f N ( t ) − M f N ( s ) ∣ ≥ ( M − 1 ) δ 1/8 . (82)
Let J : = ⌈ T δ − 1 ⌉ J:=\lceil T\delta^{-1}\rceil J := ⌈ T δ − 1 ⌉ for some fixed δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) . Set I k : = [ k δ , ( k + 1 ) δ ] ∩ [ 0 , T ] I_k:=[k\delta,(k+1)\delta]\cap[0,T] I k := [ k δ , ( k + 1 ) δ ] ∩ [ 0 , T ] , k = 0 , … , J − 1 k=0,\ldots,J-1 k = 0 , … , J − 1 . For 0 ≤ t < s ≤ T 0\le t<s\le T 0 ≤ t < s ≤ T with ∣ t − s ∣ ≤ δ |t-s|\le\delta ∣ t − s ∣ ≤ δ , assuming t ∈ I k t\in I_k t ∈ I k for some k k k , we have either, if s s s also lies in I k I_k I k ,
∣ M f N ( t ) − M f N ( s ) ∣ ≤ ∣ M f N ( t ) − M f N ( k δ ) ∣ + ∣ M f N ( k δ ) − M f N ( s ) ∣ ≤ 2 sup t ∈ I k ∣ M f N ( t ) − M f N ( k δ ) ∣ , |M_f^N(t)-M_f^N(s)|\le|M_f^N(t)-M_f^N(k\delta)|+|M_f^N(k\delta)-M_f^N(s)|
\le2\sup_{t\in I_k}|M_f^N(t)-M_f^N(k\delta)|, ∣ M f N ( t ) − M f N ( s ) ∣ ≤ ∣ M f N ( t ) − M f N ( k δ ) ∣ + ∣ M f N ( k δ ) − M f N ( s ) ∣ ≤ 2 t ∈ I k sup ∣ M f N ( t ) − M f N ( k δ ) ∣ ,
or, if s ∈ I k + 1 s\in I_{k+1} s ∈ I k + 1 ,
∣ M f N ( t ) − M f N ( s ) ∣ ≤ ∣ M f N ( t ) − M f N ( k δ ) ∣ + ∣ M f N ( k δ ) − M f N ( ( k + 1 ) δ ) ∣ + ∣ M f N ( s ) − M f N ( ( k + 1 ) δ ) ∣ ≤ 2 sup t ∈ I k ∣ M f N ( t ) − M f N ( k δ ) ∣ + sup t ∈ I k + 1 ∣ M f N ( t ) − M f N ( k δ ) ∣ . |M_f^N(t)-M_f^N(s)|\le|M_f^N(t)-M_f^N(k\delta)|+|M_f^N(k\delta)-M_f^N((k+1)\delta)|+|M_f^N(s)-M_f^N((k+1)\delta)|
\le2\sup_{t\in I_k}|M_f^N(t)-M_f^N(k\delta)|+\sup_{t\in I_{k+1}}|M_f^N(t)-M_f^N(k\delta)|. ∣ M f N ( t ) − M f N ( s ) ∣ ≤ ∣ M f N ( t ) − M f N ( k δ ) ∣ + ∣ M f N ( k δ ) − M f N (( k + 1 ) δ ) ∣ + ∣ M f N ( s ) − M f N (( k + 1 ) δ ) ∣ ≤ 2 t ∈ I k sup ∣ M f N ( t ) − M f N ( k δ ) ∣ + t ∈ I k + 1 sup ∣ M f N ( t ) − M f N ( k δ ) ∣.
Thus, we get
P ( sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ ∣ M f N ( t ) − M f N ( s ) ∣ ≥ ( M − 1 ) δ 1 / 8 ) ≤ ∑ k = 0 J − 1 P ( sup t ∈ I k ∣ M f N ( t ) − M f N ( k δ ) ∣ ≥ M − 1 3 δ 1 / 8 ) . \mathbb{P}\left(\sup_{\substack{s,t\in[0,T]\\|t-s|\le\delta}}|M_f^N(t)-M_f^N(s)|\ge(M-1)\delta^{1/8}\right)
\le\sum_{k=0}^{J-1}\mathbb{P}\left(\sup_{t\in I_k}|M_f^N(t)-M_f^N(k\delta)|\ge\frac{M-1}{3}\delta^{1/8}\right). P s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ sup ∣ M f N ( t ) − M f N ( s ) ∣ ≥ ( M − 1 ) δ 1/8 ≤ k = 0 ∑ J − 1 P ( t ∈ I k sup ∣ M f N ( t ) − M f N ( k δ ) ∣ ≥ 3 M − 1 δ 1/8 ) .
Hence, by (82) and the G G G -Burkholder–Davis–Gundy inequality, we get
P ( sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ ∣ ⟨ f , L N ( t ) ⟩ − ⟨ f , L N ( s ) ⟩ ∣ ≥ M δ 1 / 8 ) ≤ C T , σ ˉ , l 2 δ 3 l 2 / 4 − 1 ( M − 1 ) 2 l 2 N l 2 ∥ f ′ g N h N ∥ ∞ 2 l 2 . (83) \mathbb{P}\left(\sup_{\substack{s,t\in[0,T]\\|t-s|\le\delta}}|\langle f,L_N(t)\rangle-\langle f,L_N(s)\rangle|\ge M\delta^{1/8}\right)
\le\frac{C_{T,\bar{\sigma},l^2}\delta^{3l^2/4-1}}{(M-1)^{2l^2}N^{l^2}}\|f'g_Nh_N\|_\infty^{2l^2}. \tag*{(83)} P s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ δ sup ∣ ⟨ f , L N ( t )⟩ − ⟨ f , L N ( s )⟩ ∣ ≥ M δ 1/8 ≤ ( M − 1 ) 2 l 2 N l 2 C T , σ ˉ , l 2 δ 3 l 2 /4 − 1 ∥ f ′ g N h N ∥ ∞ 2 l 2 . (83)
For each f ∈ C 2 ( R ) f\in C^2(\mathbb{R}) f ∈ C 2 ( R ) and each ε > 0 \varepsilon>0 ε > 0 , define
C T ( f , ε ) : = ⋂ n = 1 ∞ { μ ∈ C ( [ 0 , T ] , P ( R ) ) : sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ n − 4 ∣ μ t ( f ) − μ s ( f ) ∣ ≤ 1 ε n } . C_T(f,\varepsilon):=\bigcap_{n=1}^{\infty}\left\{\mu\in C([0,T],\mathcal{P}(\mathbb{R})):\sup_{\substack{s,t\in[0,T]\\|t-s|\le n^{-4}}}|\mu_t(f)-\mu_s(f)|\le\frac{1}{\varepsilon\sqrt{n}}\right\}. C T ( f , ε ) := n = 1 ⋂ ∞ ⎩ ⎨ ⎧ μ ∈ C ([ 0 , T ] , P ( R )) : s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ n − 4 sup ∣ μ t ( f ) − μ s ( f ) ∣ ≤ ε n 1 ⎭ ⎬ ⎫ .
By (83) we get
c ( L N ∉ C T ( f , ε ) ) ≤ ∑ n = 1 ∞ c ( sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ n − 4 ∣ ⟨ f , L N ( t ) ⟩ − ⟨ f , L N ( s ) ⟩ ∣ ≥ 1 ε n ) ≤ ∑ n = 1 ∞ C T , σ ˉ , l 2 n − 3 l 2 + 4 ( ε − 1 − 1 ) 2 l 2 N l 2 ∥ f ′ g N h N ∥ ∞ 2 l 2 = C T , σ ˉ , l 2 ∥ f ′ g N h N ∥ ∞ 2 l 2 ( ε − 1 − 1 ) 2 l 2 N l 2 ∑ n = 1 ∞ n − 3 l 2 + 4 . (84) \begin{aligned}
c(L_N \notin C_T(f,\varepsilon)) &\le\sum_{n=1}^{\infty} c\left(\sup_{\substack{s,t\in[0,T]\\ |t-s|\le n^{-4}}}|\langle f,L_N(t)\rangle-\langle f,L_N(s)\rangle|\ge\frac{1}{\varepsilon\sqrt{n}}\right)\\
&\le\sum_{n=1}^{\infty}\frac{C_{T,\bar{\sigma},l_2}n^{-3l_2+4}}{(\varepsilon^{-1}-1)^{2l_2}N^{l_2}}\|f'g_Nh_N\|_{\infty}^{2l_2}\\
&=\frac{C_{T,\bar{\sigma},l_2}\|f'g_Nh_N\|_{\infty}^{2l_2}}{(\varepsilon^{-1}-1)^{2l_2}N^{l_2}}\sum_{n=1}^{\infty}n^{-3l_2+4}. \tag*{(84)}
\end{aligned} c ( L N ∈ / C T ( f , ε )) ≤ n = 1 ∑ ∞ c s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ n − 4 sup ∣ ⟨ f , L N ( t )⟩ − ⟨ f , L N ( s )⟩ ∣ ≥ ε n 1 ≤ n = 1 ∑ ∞ ( ε − 1 − 1 ) 2 l 2 N l 2 C T , σ ˉ , l 2 n − 3 l 2 + 4 ∥ f ′ g N h N ∥ ∞ 2 l 2 = ( ε − 1 − 1 ) 2 l 2 N l 2 C T , σ ˉ , l 2 ∥ f ′ g N h N ∥ ∞ 2 l 2 n = 1 ∑ ∞ n − 3 l 2 + 4 . (84)
Note that l 2 ≥ 2 l_2\ge2 l 2 ≥ 2 , and hence ∑ n = 1 ∞ n − 3 l 2 + 4 < ∞ \sum_{n=1}^{\infty}n^{-3l_2+4}<\infty ∑ n = 1 ∞ n − 3 l 2 + 4 < ∞ . Let ( ε k ) k ∈ N (\varepsilon_k)_{k\in\mathbb{N}} ( ε k ) k ∈ N be a sequence of positive numbers converging to 0 0 0 satisfying ε k − 1 > 1 + k ψ ( k ) 1 2 l 2 \varepsilon_k^{-1}>1+k\psi(k)^{\frac{1}{2l_2}} ε k − 1 > 1 + k ψ ( k ) 2 l 2 1 . Then by Assumption 2 and (84) , we get
∑ k ≥ 1 ∑ N = 1 ∞ c ( L N ∉ C T ( f ~ k , ε k ) ) ≤ ∑ k ≥ 1 ∑ N = 1 ∞ C ∥ f ~ k ′ g N h N ∥ ∞ 2 l 2 ( ε k − 1 − 1 ) 2 l 2 N l 2 ≤ ∑ k ≥ 1 C k 2 l 2 < ∞ . (85) \sum_{k\ge1}\sum_{N=1}^{\infty}c(L_N\notin C_T(\widetilde f_k,\varepsilon_k))
\le\sum_{k\ge1}\sum_{N=1}^{\infty}\frac{C\|\widetilde f'_kg_Nh_N\|_{\infty}^{2l_2}}{(\varepsilon_k^{-1}-1)^{2l_2}Nl_2}
\le\sum_{k\ge1}\frac{C}{k^{2l_2}}<\infty. \tag*{(85)} k ≥ 1 ∑ N = 1 ∑ ∞ c ( L N ∈ / C T ( f k , ε k )) ≤ k ≥ 1 ∑ N = 1 ∑ ∞ ( ε k − 1 − 1 ) 2 l 2 N l 2 C ∥ f k ′ g N h N ∥ ∞ 2 l 2 ≤ k ≥ 1 ∑ k 2 l 2 C < ∞. (85)
Denote K : = K M ∩ ⋂ k ≥ 1 C T ( f ~ k , ε k ) \mathcal{K}:=K_M\cap\bigcap_{k\ge1}C_T(\widetilde f_k,\varepsilon_k) K := K M ∩ ⋂ k ≥ 1 C T ( f k , ε k ) , which is a compact subset of C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathbb{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) by [1 Lemma 4.3.13] . Combining (81) , (85) , and Lemma 2.18, we conclude that
c ( ⋃ N 0 ≥ 1 ⋂ N ≥ N 0 { L N ∈ K } ) = 1. c\left(\bigcup_{N_0\ge1}\bigcap_{N\ge N_0}\{L_N\in\mathcal{K}\}\right)=1. c ( N 0 ≥ 1 ⋃ N ≥ N 0 ⋂ { L N ∈ K } ) = 1.
Therefore, the sequence { L N } N ≥ 1 \{L_N\}_{N\ge1} { L N } N ≥ 1 is relatively compact in C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathbb{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) , quasi-surely. □
Now we derive the equation satisfied by the spectral limit.
Proof of Theorem 1.8 . For any limit point μ \mu μ , we can find a subsequence { N i } \{N_i\} { N i } , such that L N i L_{N_i} L N i converges to μ \mu μ in C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathbb{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) as N i N_i N i tends to infinity.
Fix z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R and define f z ( x ) : = 1 z − x f_z(x):=\frac{1}{z-x} f z ( x ) := z − x 1 , x ∈ R x\in\mathbb{R} x ∈ R . Then f z ′ ( x ) = 1 ( z − x ) 2 f'_z(x)=\frac{1}{(z-x)^2} f z ′ ( x ) = ( z − x ) 2 1 , f z ′ ′ ( x ) = 2 ( z − x ) 3 f''_z(x)=\frac{2}{(z-x)^3} f z ′′ ( x ) = ( z − x ) 3 2 . By (78) with f = f z f=f_z f = f z , N = N i N=N_i N = N i , we have
⟨ f z , L N i ( t ) ⟩ = ⟨ f z , L N i ( 0 ) ⟩ + M f z N i ( t ) + ∫ 0 t ∫ R b N i ( x ) ( z − x ) 2 L N i ( s ) ( d x ) d ⟨ B i ‾ ⟩ s + N i 2 ∫ 0 t ∫ R 2 ( z − x ) − 2 − ( z − y ) − 2 x − y H N i ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) d ⟨ B i ‾ ⟩ s + ∫ 0 t ∫ R 2 g N i 2 ( x ) h N i 2 ( x ) ( z − x ) 3 L N i ( s ) ( d x ) d ⟨ B i ‾ ⟩ s . \begin{aligned}
\langle f_z,L_{N_i}(t)\rangle
={}&\langle f_z,L_{N_i}(0)\rangle+M_{f_z}^{N_i}(t)
+\int_0^t\int_{\mathbb{R}}\frac{b_{N_i}(x)}{(z-x)^2}L_{N_i}(s)(\mathrm{d}x)\,\mathrm{d}\langle\overline{B_i}\rangle_s\\
&+\frac{N_i}{2}\int_0^t\int_{\mathbb{R}^2}\frac{(z-x)^{-2}-(z-y)^{-2}}{x-y}H_{N_i}(x,y)L_{N_i}(s)(\mathrm{d}x)L_{N_i}(s)(\mathrm{d}y)\,\mathrm{d}\langle\overline{B_i}\rangle_s\\
&+\int_0^t\int_{\mathbb{R}}\frac{2g_{N_i}^2(x)h_{N_i}^2(x)}{(z-x)^3}L_{N_i}(s)(\mathrm{d}x)\,\mathrm{d}\langle\overline{B_i}\rangle_s.
\end{aligned} ⟨ f z , L N i ( t )⟩ = ⟨ f z , L N i ( 0 )⟩ + M f z N i ( t ) + ∫ 0 t ∫ R ( z − x ) 2 b N i ( x ) L N i ( s ) ( d x ) d ⟨ B i ⟩ s + 2 N i ∫ 0 t ∫ R 2 x − y ( z − x ) − 2 − ( z − y ) − 2 H N i ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) d ⟨ B i ⟩ s + ∫ 0 t ∫ R ( z − x ) 3 2 g N i 2 ( x ) h N i 2 ( x ) L N i ( s ) ( d x ) d ⟨ B i ⟩ s .
Here
M f z N i ( t ) : = 2 N i ∑ j = 1 N i ∫ 0 t g N i ( λ j N i ( s ) ) h N i ( λ j N i ( s ) ) ( z − λ j N i ( s ) ) 2 d B ‾ j ( s ) . M_{f_z}^{N_i}(t):=\frac{2}{N_i}\sum_{j=1}^{N_i}\int_0^t\frac{g_{N_i}(\lambda_j^{N_i}(s))h_{N_i}(\lambda_j^{N_i}(s))}{(z-\lambda_j^{N_i}(s))^2}\,\mathrm{d}\overline{B}_j(s). M f z N i ( t ) := N i 2 j = 1 ∑ N i ∫ 0 t ( z − λ j N i ( s ) ) 2 g N i ( λ j N i ( s )) h N i ( λ j N i ( s )) d B j ( s ) .
By the G G G -Burkholder-Davis-Gundy inequality and (79) , we get
c ( sup 0 ≤ t ≤ T ∣ M f z N ( t ) ∣ ≥ ε ) ≤ C ε 2 E ^ [ ⟨ M f z N ⟩ T ] ≤ 4 C σ ˉ 2 T ∥ f z ′ g N h N ∥ ∞ 2 ε 2 N . c\left(\sup_{0\le t\le T}|M_{f_z}^{N}(t)|\ge\varepsilon\right)\le\frac{C}{\varepsilon^2}\widehat{\mathbb{E}}\left[\langle M_{f_z}^{N}\rangle_T\right]\le\frac{4C\bar{\sigma}^2T\|f'_zg_Nh_N\|_{\infty}^2}{\varepsilon^2N}. c ( 0 ≤ t ≤ T sup ∣ M f z N ( t ) ∣ ≥ ε ) ≤ ε 2 C E [ ⟨ M f z N ⟩ T ] ≤ ε 2 N 4 C σ ˉ 2 T ∥ f z ′ g N h N ∥ ∞ 2 .
Therefore,
∑ N = 1 ∞ c ( sup 0 ≤ t ≤ T ∣ M f N ( t ) ∣ ≥ ε ) ≤ ∑ N = 1 ∞ 4 C σ ˉ 2 T ε 2 N ∥ g N h N ( z − x ) 2 ∥ ∞ 2 < ∞ , \sum_{N=1}^{\infty} c\left(\sup_{0\leq t\leq T}\left|M_f^N(t)\right|\geq\varepsilon\right)\leq\sum_{N=1}^{\infty}\frac{4C\bar{\sigma}^{2}T}{\varepsilon^{2}N}\left\|\frac{g_Nh_N}{(z-x)^{2}}\right\|_{\infty}^{2}<\infty, N = 1 ∑ ∞ c ( 0 ≤ t ≤ T sup M f N ( t ) ≥ ε ) ≤ N = 1 ∑ ∞ ε 2 N 4 C σ ˉ 2 T ( z − x ) 2 g N h N ∞ 2 < ∞ ,
where the last step is due to the fact ∥ g N h N ( z − x ) 2 ∥ ∞ 2 ≤ C / N \left\|\frac{g_Nh_N}{(z-x)^{2}}\right\|_{\infty}^{2}\leq C/N ( z − x ) 2 g N h N ∞ 2 ≤ C / N for some constant C C C by [35 Remark 2.3] . Thus, by Lemma 2.18, we can verify that, for any ε > 0 \varepsilon>0 ε > 0 ,
c ( lim inf n → ∞ { sup 0 ≤ t ≤ T ∣ M f N ( t ) ∣ < ε } ) = 1 , c\left(\liminf_{n\to\infty}\left\{\sup_{0\leq t\leq T}\left|M_f^N(t)\right|<\varepsilon\right\}\right)=1, c ( n → ∞ lim inf { 0 ≤ t ≤ T sup M f N ( t ) < ε } ) = 1 ,
which yields the q.s. convergence of M f N i M_f^{N_i} M f N i towards zero, uniformly on compact time intervals.
Next, noting that
∣ ∫ b N i ( x ) ( z − x ) 2 L N i ( s ) ( d x ) − ∫ b ( x ) ( z − x ) 2 μ s ( d x ) ∣ ≤ ∣ ∫ b N i ( x ) − b ( x ) ( z − x ) 2 L N i ( s ) ( d x ) ∣ + ∣ ∫ b ( x ) ( z − x ) 2 ( L N i ( s ) ( d x ) − μ s ( d x ) ) ∣ ≤ sup x ∣ b N i ( x ) − b ( x ) ∣ ( Im ( z ) ) 2 + ∣ ∫ b ( x ) ( z − x ) 2 ( L N i ( s ) ( d x ) − μ s ( d x ) ) ∣ → 0 , N i → ∞ , \left|\int\frac{b_{N_i}(x)}{(z-x)^{2}}L_{N_i}(s)(\mathrm{d}x)-\int\frac{b(x)}{(z-x)^{2}}\mu_s(\mathrm{d}x)\right|
\leq\left|\int\frac{b_{N_i}(x)-b(x)}{(z-x)^{2}}L_{N_i}(s)(\mathrm{d}x)\right|+\left|\int\frac{b(x)}{(z-x)^{2}}\left(L_{N_i}(s)(\mathrm{d}x)-\mu_s(\mathrm{d}x)\right)\right|
\leq\frac{\sup_x|b_{N_i}(x)-b(x)|}{(\operatorname{Im}(z))^{2}}+\left|\int\frac{b(x)}{(z-x)^{2}}\left(L_{N_i}(s)(\mathrm{d}x)-\mu_s(\mathrm{d}x)\right)\right|\to0,\ N_i\to\infty, ∫ ( z − x ) 2 b N i ( x ) L N i ( s ) ( d x ) − ∫ ( z − x ) 2 b ( x ) μ s ( d x ) ≤ ∫ ( z − x ) 2 b N i ( x ) − b ( x ) L N i ( s ) ( d x ) + ∫ ( z − x ) 2 b ( x ) ( L N i ( s ) ( d x ) − μ s ( d x ) ) ≤ ( Im ( z ) ) 2 sup x ∣ b N i ( x ) − b ( x ) ∣ + ∫ ( z − x ) 2 b ( x ) ( L N i ( s ) ( d x ) − μ s ( d x ) ) → 0 , N i → ∞ ,
we have
lim N i → ∞ ∫ 0 t ∫ b N i ( x ) ( z − x ) 2 L N i ( s ) ( d x ) d ⟨ B ˉ i ⟩ s = ∫ 0 t ∫ b ( x ) ( z − x ) 2 μ s ( d x ) d ⟨ B ˉ 1 ⟩ s . \lim_{N_i\to\infty}\int_0^t\int\frac{b_{N_i}(x)}{(z-x)^{2}}L_{N_i}(s)(\mathrm{d}x)\mathrm{d}\langle\bar{B}_i\rangle_s=\int_0^t\int\frac{b(x)}{(z-x)^{2}}\mu_s(\mathrm{d}x)\mathrm{d}\langle\bar{B}^{1}\rangle_s. N i → ∞ lim ∫ 0 t ∫ ( z − x ) 2 b N i ( x ) L N i ( s ) ( d x ) d ⟨ B ˉ i ⟩ s = ∫ 0 t ∫ ( z − x ) 2 b ( x ) μ s ( d x ) d ⟨ B ˉ 1 ⟩ s .
For the last term on the right-hand side of (86) , we have
∣ ∫ 2 g N i 2 ( x ) h N i 2 ( x ) ( z − x ) 3 L N i ( s ) ( d x ) ∣ = 1 N i ∣ ∫ N i H N i ( x , x ) ( z − x ) 3 L N i ( s ) ( d x ) ∣ ≤ 1 N i ∣ ∫ N i H N i ( x , x ) − H ( x , x ) ( z − x ) 3 L N i ( s ) ( d x ) ∣ + 1 N i ∣ ∫ H ( x , x ) ( z − x ) 3 L N i ( s ) ( d x ) ∣ ≤ sup x , y ∣ N i H N i ( x , y ) − H ( x , y ) ∣ N i ( Im ( z ) ) 3 + C z N i ∥ H ( x , x ) 1 + ∣ x ∣ 3 ∥ L ∞ ( d x ) → 0 , N i → ∞ , \left|\int\frac{2g_{N_i}^{2}(x)h_{N_i}^{2}(x)}{(z-x)^{3}}L_{N_i}(s)(\mathrm{d}x)\right|
=\frac{1}{N_i}\left|\int\frac{N_iH_{N_i}(x,x)}{(z-x)^{3}}L_{N_i}(s)(\mathrm{d}x)\right|
\leq\frac{1}{N_i}\left|\int\frac{N_iH_{N_i}(x,x)-H(x,x)}{(z-x)^{3}}L_{N_i}(s)(\mathrm{d}x)\right|+\frac{1}{N_i}\left|\int\frac{H(x,x)}{(z-x)^{3}}L_{N_i}(s)(\mathrm{d}x)\right|
\leq\frac{\sup_{x,y}|N_iH_{N_i}(x,y)-H(x,y)|}{N_i(\operatorname{Im}(z))^{3}}+\frac{C_z}{N_i}\left\|\frac{H(x,x)}{1+|x|^{3}}\right\|_{L^\infty(\mathrm{d}x)}\to0,\ N_i\to\infty, ∫ ( z − x ) 3 2 g N i 2 ( x ) h N i 2 ( x ) L N i ( s ) ( d x ) = N i 1 ∫ ( z − x ) 3 N i H N i ( x , x ) L N i ( s ) ( d x ) ≤ N i 1 ∫ ( z − x ) 3 N i H N i ( x , x ) − H ( x , x ) L N i ( s ) ( d x ) + N i 1 ∫ ( z − x ) 3 H ( x , x ) L N i ( s ) ( d x ) ≤ N i ( Im ( z ) ) 3 sup x , y ∣ N i H N i ( x , y ) − H ( x , y ) ∣ + N i C z 1 + ∣ x ∣ 3 H ( x , x ) L ∞ ( d x ) → 0 , N i → ∞ ,
where C z C_z C z is a constant depending only on z z z .
Now, it remains to identify the limit of the interaction term. A direct calculation gives
( z − x ) − 2 − ( z − y ) − 2 x − y = ( z − y ) 2 − ( z − x ) 2 ( z − x ) 2 ( z − y ) 2 ( x − y ) = 2 z − x − y ( z − x ) 2 ( z − y ) 2 = 1 ( z − x ) ( z − y ) 2 + 1 ( z − x ) 2 ( z − y ) . (86) \begin{aligned}
\frac{(z-x)^{-2}-(z-y)^{-2}}{x-y}=\frac{(z-y)^{2}-(z-x)^{2}}{(z-x)^{2}(z-y)^{2}(x-y)}=\frac{2z-x-y}{(z-x)^{2}(z-y)^{2}} \\
&=\frac{1}{(z-x)(z-y)^{2}}+\frac{1}{(z-x)^{2}(z-y)}. \tag*{(86)}
\end{aligned} x − y ( z − x ) − 2 − ( z − y ) − 2 = ( z − x ) 2 ( z − y ) 2 ( x − y ) ( z − y ) 2 − ( z − x ) 2 = ( z − x ) 2 ( z − y ) 2 2 z − x − y = ( z − x ) ( z − y ) 2 1 + ( z − x ) 2 ( z − y ) 1 . (86)
Therefore, using the symmetry of H N H_N H N and exchanging x x x and y y y in the second term on the right-hand side of (86) , we obtain
1 2 ∬ R 2 ( z − x ) − 2 − ( z − y ) − 2 x − y N i H N i ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) = ∬ R 2 N i H N i ( x , y ) ( z − x ) ( z − y ) 2 L N i ( s ) ( d x ) L N i ( s ) ( d y ) . \frac{1}{2}\iint_{\mathbb{R}^{2}}\frac{(z-x)^{-2}-(z-y)^{-2}}{x-y}N_iH_{N_i}(x,y)L_{N_i}(s)(\mathrm{d}x)L_{N_i}(s)(\mathrm{d}y)
=\iint_{\mathbb{R}^{2}}\frac{N_iH_{N_i}(x,y)}{(z-x)(z-y)^{2}}L_{N_i}(s)(\mathrm{d}x)L_{N_i}(s)(\mathrm{d}y). 2 1 ∬ R 2 x − y ( z − x ) − 2 − ( z − y ) − 2 N i H N i ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) = ∬ R 2 ( z − x ) ( z − y ) 2 N i H N i ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) .
As H ( x , y ) ( 1 + ∣ x ∣ ) ( 1 + y 2 ) \frac{H(x,y)}{(1+|x|)(1+y^{2})} ( 1 + ∣ x ∣ ) ( 1 + y 2 ) H ( x , y ) is bounded and by the uniform convergence of N H N ( x , y ) NH_N(x,y) N H N ( x , y ) to H ( x , y ) H(x,y) H ( x , y ) , we can obtain that N H N ( x , y ) ( z − x ) ( z − y ) 2 \frac{NH_N(x,y)}{(z-x)(z-y)^2} ( z − x ) ( z − y ) 2 N H N ( x , y ) is bounded and as N i → ∞ N_i \to\infty N i → ∞ ,
∣ ∬ N i H N i ( x , y ) ( z − x ) ( z − y ) 2 L N i ( s ) ( d x ) L N i ( s ) ( d y ) − ∬ H ( x , y ) ( z − x ) ( z − y ) 2 μ s ( d x ) μ s ( d y ) ∣ ≤ ∣ ∬ N i H N i ( x , y ) − H ( x , y ) ( z − x ) ( z − y ) 2 L N i ( s ) ( d x ) L N i ( s ) ( d y ) ∣ + ∣ ∬ H ( x , y ) ( z − x ) ( z − y ) 2 ( L N i ( s ) ( d x ) L N i ( s ) ( d y ) − μ s ( d x ) μ s ( d y ) ) ∣ ≤ sup x , y ∣ N i H N i ( x , y ) − H ( x , y ) ∣ ∣ Im ( z ) ∣ 3 + ∣ ∬ H ( x , y ) ( z − x ) ( z − y ) 2 ( L N i ( s ) ( d x ) L N i ( s ) ( d y ) − μ s ( d x ) μ s ( d y ) ) ∣ → 0. \begin{aligned}
\left|\iint\frac{N_iH_{N_i}(x,y)}{(z-x)(z-y)^2}L_{N_i}(s)(dx)L_{N_i}(s)(dy)-\iint\frac{H(x,y)}{(z-x)(z-y)^2}\mu_s(dx)\mu_s(dy)\right| \\
&\leq\left|\iint\frac{N_iH_{N_i}(x,y)-H(x,y)}{(z-x)(z-y)^2}L_{N_i}(s)(dx)L_{N_i}(s)(dy)\right| \\
&+\left|\iint\frac{H(x,y)}{(z-x)(z-y)^2}\left(L_{N_i}(s)(dx)L_{N_i}(s)(dy)-\mu_s(dx)\mu_s(dy)\right)\right| \\
&\leq\frac{\sup_{x,y}|N_iH_{N_i}(x,y)-H(x,y)|}{|\operatorname{Im}(z)|^3}+\left|\iint\frac{H(x,y)}{(z-x)(z-y)^2}\left(L_{N_i}(s)(dx)L_{N_i}(s)(dy)-\mu_s(dx)\mu_s(dy)\right)\right|\to0.
\end{aligned} ∬ ( z − x ) ( z − y ) 2 N i H N i ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) − ∬ ( z − x ) ( z − y ) 2 H ( x , y ) μ s ( d x ) μ s ( d y ) ≤ ∬ ( z − x ) ( z − y ) 2 N i H N i ( x , y ) − H ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) + ∬ ( z − x ) ( z − y ) 2 H ( x , y ) ( L N i ( s ) ( d x ) L N i ( s ) ( d y ) − μ s ( d x ) μ s ( d y ) ) ≤ ∣ Im ( z ) ∣ 3 sup x , y ∣ N i H N i ( x , y ) − H ( x , y ) ∣ + ∬ ( z − x ) ( z − y ) 2 H ( x , y ) ( L N i ( s ) ( d x ) L N i ( s ) ( d y ) − μ s ( d x ) μ s ( d y ) ) → 0.
Then we have
lim N i → ∞ ∫ 0 t ∬ N i H N i ( x , y ) ( z − x ) ( z − y ) 2 L N i ( s ) ( d x ) L N i ( s ) ( d y ) d ⟨ B ˉ 1 ⟩ s = ∫ 0 t ∬ H ( x , y ) ( z − x ) ( z − y ) 2 μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s . \begin{aligned}
\lim_{N_i\to\infty}\int_0^t\iint\frac{N_iH_{N_i}(x,y)}{(z-x)(z-y)^2}L_{N_i}(s)(dx)L_{N_i}(s)(dy)d\langle\bar{B}^{1}\rangle_s \\
&=\int_0^t\iint\frac{H(x,y)}{(z-x)(z-y)^2}\mu_s(dx)\mu_s(dy)d\langle\bar{B}^{1}\rangle_s.
\end{aligned} N i → ∞ lim ∫ 0 t ∬ ( z − x ) ( z − y ) 2 N i H N i ( x , y ) L N i ( s ) ( d x ) L N i ( s ) ( d y ) d ⟨ B ˉ 1 ⟩ s = ∫ 0 t ∬ ( z − x ) ( z − y ) 2 H ( x , y ) μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s .
This finishes the proof. □
Dyson Brownian motion
Proof of Proposition 1.11. By Theorems 1.7 and 1.8 , the sequence ( L N ( t ) ) t ∈ [ 0 , T ] (L_N(t))_{t\in[0,T]} ( L N ( t ) ) t ∈ [ 0 , T ] is q.s. pre-compact in C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathcal{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) and that any limit point ( μ t , t ∈ [ 0 , T ] ) (\mu_t,t\in[0,T]) ( μ t , t ∈ [ 0 , T ]) of L N L_N L N satisfies the equation
∫ f ( x ) μ t ( d x ) = ∫ f ( x ) μ 0 ( d x ) + 1 2 ∫ 0 t ∬ f ′ ( x ) − f ′ ( y ) x − y μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s . \int f(x)\mu_t(dx)=\int f(x)\mu_0(dx)+\frac{1}{2}\int_0^t\iint\frac{f'(x)-f'(y)}{x-y}\mu_s(dx)\mu_s(dy)d\langle\bar{B}^{1}\rangle_s. ∫ f ( x ) μ t ( d x ) = ∫ f ( x ) μ 0 ( d x ) + 2 1 ∫ 0 t ∬ x − y f ′ ( x ) − f ′ ( y ) μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s .
Set f ( x ) = ( z − x ) − 1 f(x)=(z-x)^{-1} f ( x ) = ( z − x ) − 1 for some z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , and let
V t ( z ) = ∫ 1 z − x μ t ( d x ) , for t ≥ 0 , V_t(z)=\int\frac{1}{z-x}\mu_t(dx),\qquad\text{for }t\geq0, V t ( z ) = ∫ z − x 1 μ t ( d x ) , for t ≥ 0 ,
as given in (1.18). Then,
f ′ ( x ) − f ′ ( y ) x − y = ( z − x ) − 2 − ( z − y ) − 2 x − y = 1 ( z − x ) ( z − y ) 2 + 1 ( z − y ) ( z − x ) 2 , \frac{f'(x)-f'(y)}{x-y}=\frac{(z-x)^{-2}-(z-y)^{-2}}{x-y}=\frac{1}{(z-x)(z-y)^2}+\frac{1}{(z-y)(z-x)^2}, x − y f ′ ( x ) − f ′ ( y ) = x − y ( z − x ) − 2 − ( z − y ) − 2 = ( z − x ) ( z − y ) 2 1 + ( z − y ) ( z − x ) 2 1 ,
and by the symmetry we get
V t ( z ) = V 0 ( z ) + 1 2 ∫ 0 t ∬ [ 1 ( z − x ) ( z − y ) 2 + 1 ( z − y ) ( z − x ) 2 ] μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s = V 0 ( z ) + ∫ 0 t ∬ 1 ( z − x ) ( z − y ) 2 μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s = V 0 ( z ) − ∫ 0 t V s ( z ) ∂ z V s ( z ) d ⟨ B ˉ 1 ⟩ s . \begin{aligned}
V_t(z)=V_0(z)+\frac{1}{2}\int_0^t\iint\left[\frac{1}{(z-x)(z-y)^2}+\frac{1}{(z-y)(z-x)^2}\right]\mu_s(dx)\mu_s(dy)d\langle\bar{B}^{1}\rangle_s \\
&=V_0(z)+\int_0^t\iint\frac{1}{(z-x)(z-y)^2}\mu_s(dx)\mu_s(dy)d\langle\bar{B}^{1}\rangle_s \\
&=V_0(z)-\int_0^tV_s(z)\partial_zV_s(z)d\langle\bar{B}^{1}\rangle_s.
\end{aligned} V t ( z ) = V 0 ( z ) + 2 1 ∫ 0 t ∬ [ ( z − x ) ( z − y ) 2 1 + ( z − y ) ( z − x ) 2 1 ] μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s = V 0 ( z ) + ∫ 0 t ∬ ( z − x ) ( z − y ) 2 1 μ s ( d x ) μ s ( d y ) d ⟨ B ˉ 1 ⟩ s = V 0 ( z ) − ∫ 0 t V s ( z ) ∂ z V s ( z ) d ⟨ B ˉ 1 ⟩ s .
This proves (1.19).
Clearly, for each fixed t t t , V t ( z ) V_t(z) V t ( z ) is an analytic function on C + : = { z ∈ C : Im ( z ) > 0 } \mathbb{C}_+:=\{z\in\mathbb{C}:\operatorname{Im}(z)>0\} C + := { z ∈ C : Im ( z ) > 0 } . Now we prove the uniqueness of an analytic solution to (1.19), which yields the uniqueness of μ \mu μ .
Fix a solution V V V of (1.18) and (1.19). Consider the characteristic equation
d z t = V t ( z t ) d ⟨ B ˉ 1 ⟩ t , z 0 ∈ C + . (87) dz_t=V_t(z_t)d\langle\bar{B}^{1}\rangle_t,\qquad z_0\in\mathbb{C}_+. \tag*{(87)} d z t = V t ( z t ) d ⟨ B ˉ 1 ⟩ t , z 0 ∈ C + . (87)
To justify the existence of a solution to (87) , we first introduce a regularization. For ε > 0 \varepsilon> 0 ε > 0 , define
V t ε ( z ) : = ∫ R z ‾ − x ∣ z − x ∣ 2 + ε μ t ( d x ) , z ∈ C + , V_t^\varepsilon(z) := \int_{\mathbb{R}} \frac{\overline{z}-x}{\lvert z-x\rvert^2+\varepsilon}\,\mu_t(dx), \quad z \in\mathbb{C}_+, V t ε ( z ) := ∫ R ∣ z − x ∣ 2 + ε z − x μ t ( d x ) , z ∈ C + ,
and consider the regularized characteristic equation
d z t ε = V t ε ( z t ε ) d ⟨ B ‾ 1 ⟩ t , z 0 ε = z 0 . (88) dz_t^\varepsilon= V_t^\varepsilon(z_t^\varepsilon)\,d\langle\overline{B}^1\rangle_t,\quad z_0^\varepsilon=z_0. \tag*{(88)} d z t ε = V t ε ( z t ε ) d ⟨ B 1 ⟩ t , z 0 ε = z 0 . (88)
Since V t ε V_t^\varepsilon V t ε is smooth and uniformly Lipschitz on C + \mathbb{C}_+ C + , equation (88) admits a unique solution. Denote z t ε = u t ε + i v t ε z_t^\varepsilon=u_t^\varepsilon+iv_t^\varepsilon z t ε = u t ε + i v t ε . The imaginary part of (88) yields
d v t ε d ⟨ B ‾ 1 ⟩ t = − v t ε ∫ R 1 ∣ z t ε − x ∣ 2 + ε μ t ( d x ) , \frac{dv_t^\varepsilon}{d\langle\overline{B}^1\rangle_t}=-v_t^\varepsilon\int_{\mathbb{R}}\frac{1}{\lvert z_t^\varepsilon-x\rvert^2+\varepsilon}\,\mu_t(dx), d ⟨ B 1 ⟩ t d v t ε = − v t ε ∫ R ∣ z t ε − x ∣ 2 + ε 1 μ t ( d x ) ,
and consequently,
d ( v t ε ) 2 d ⟨ B ‾ 1 ⟩ t = − 2 ( v t ε ) 2 ∫ R 1 ∣ z t ε − x ∣ 2 + ε μ t ( d x ) ≥ − 2. \frac{d(v_t^\varepsilon)^2}{d\langle\overline{B}^1\rangle_t}=-2(v_t^\varepsilon)^2\int_{\mathbb{R}}\frac{1}{\lvert z_t^\varepsilon-x\rvert^2+\varepsilon}\,\mu_t(dx)\geq-2. d ⟨ B 1 ⟩ t d ( v t ε ) 2 = − 2 ( v t ε ) 2 ∫ R ∣ z t ε − x ∣ 2 + ε 1 μ t ( d x ) ≥ − 2.
Noting ⟨ B ‾ 1 ⟩ t ≤ σ ‾ 2 t \langle\overline{B}^1\rangle_t\leq\overline{\sigma}^2t ⟨ B 1 ⟩ t ≤ σ 2 t q.s., we deduce that
∣ Im z t ε ∣ 2 = ( v t ε ) 2 ≥ ( Im z 0 ) 2 − 2 σ ‾ 2 t . (89) \lvert\operatorname{Im}z_t^\varepsilon\rvert^2=(v_t^\varepsilon)^2\geq(\operatorname{Im}z_0)^2-2\overline{\sigma}^2t. \tag*{(89)} ∣ Im z t ε ∣ 2 = ( v t ε ) 2 ≥ ( Im z 0 ) 2 − 2 σ 2 t . (89)
In particular, for 0 ≤ t ≤ ( Im z 0 ) 2 4 σ ‾ 2 0\leq t\leq\frac{(\operatorname{Im}z_0)^2}{4\overline{\sigma}^2} 0 ≤ t ≤ 4 σ 2 ( Im z 0 ) 2 , we have Im z t ε ≥ Im z 0 2 > 0 \operatorname{Im}z_t^\varepsilon\geq\frac{\operatorname{Im}z_0}{\sqrt{2}}>0 Im z t ε ≥ 2 Im z 0 > 0 , and
∣ V t ε ( z t ε ) ∣ ≤ ∫ R 1 Im z t ε μ t ( d x ) ≤ 2 Im z 0 , \lvert V_t^\varepsilon(z_t^\varepsilon)\rvert\leq\int_{\mathbb{R}}\frac{1}{\operatorname{Im}z_t^\varepsilon}\,\mu_t(dx)\leq\frac{\sqrt{2}}{\operatorname{Im}z_0}, ∣ V t ε ( z t ε )∣ ≤ ∫ R Im z t ε 1 μ t ( d x ) ≤ Im z 0 2 ,
which implies that d v t ε d ⟨ B ‾ ⟩ t ≤ 2 Im z 0 \frac{dv_t^\varepsilon}{d\langle\overline{B}\rangle_t}\leq\frac{\sqrt{2}}{\operatorname{Im}z_0} d ⟨ B ⟩ t d v t ε ≤ Im z 0 2 . For the real part, it follows from the definition of V t ε V_t^\varepsilon V t ε that
∣ d u t ε d ⟨ B ‾ 1 ⟩ t ∣ ≤ ∫ R ∣ u t ε − x ∣ ∣ u t ε − x ∣ 2 + ( v t ε ) 2 + ε μ t ( d x ) ≤ 1 2 ( v t ε ) 2 + ε . \left\lvert\frac{du_t^\varepsilon}{d\langle\overline{B}^1\rangle_t}\right\rvert\leq\int_{\mathbb{R}}\frac{\lvert u_t^\varepsilon-x\rvert}{\lvert u_t^\varepsilon-x\rvert^2+(v_t^\varepsilon)^2+\varepsilon}\,\mu_t(dx)\leq\frac{1}{2\sqrt{(v_t^\varepsilon)^2+\varepsilon}}. d ⟨ B 1 ⟩ t d u t ε ≤ ∫ R ∣ u t ε − x ∣ 2 + ( v t ε ) 2 + ε ∣ u t ε − x ∣ μ t ( d x ) ≤ 2 ( v t ε ) 2 + ε 1 .
The above estimates imply that the family { z ε } ε > 0 \{z^\varepsilon\}_{\varepsilon>0} { z ε } ε > 0 is equicontinuous on [ 0 , ( Im z 0 ) 2 4 σ ‾ 2 ] [0,\frac{(\operatorname{Im}z_0)^2}{4\overline{\sigma}^2}] [ 0 , 4 σ 2 ( Im z 0 ) 2 ] . Hence, by the Arzelà–Ascoli theorem, there exists a subsequence which converges uniformly to a continuous process z t z_t z t . Passing to the limit in (88) , we find that z t z_t z t solves (87) . In addition, Im z t ≥ Im z 0 2 > 0 \operatorname{Im}z_t\geq\frac{\operatorname{Im}z_0}{\sqrt{2}}>0 Im z t ≥ 2 Im z 0 > 0 for t ∈ [ 0 , ( Im z 0 ) 2 4 σ ‾ 2 ] t\in[0,\frac{(\operatorname{Im}z_0)^2}{4\overline{\sigma}^2}] t ∈ [ 0 , 4 σ 2 ( Im z 0 ) 2 ] . On the domain where Im z t \operatorname{Im}z_t Im z t is bounded away from zero, the function V t V_t V t is uniformly Lipschitz. Therefore, the solution of (87) is unique on the above time interval. Applying the chain rule to V t ( z t ) V_t(z_t) V t ( z t ) and using (1.19) and (87) , we obtain
d ( V t ( z t ) ) = − V t ( z t ) ∂ z V t ( z t ) d ⟨ B ‾ 1 ⟩ t + ∂ z V t ( z t ) d z t = − V t ( z t ) ∂ z V t ( z t ) d ⟨ B ‾ 1 ⟩ t + ∂ z V t ( z t ) V t ( z t ) d ⟨ B ‾ 1 ⟩ t = 0. \begin{aligned}
d(V_t(z_t))&=-V_t(z_t)\partial_zV_t(z_t)\,d\langle\overline{B}^1\rangle_t+\partial_zV_t(z_t)\,dz_t\\
&=-V_t(z_t)\partial_zV_t(z_t)\,d\langle\overline{B}^1\rangle_t+\partial_zV_t(z_t)V_t(z_t)\,d\langle\overline{B}^1\rangle_t\\
&=0.
\end{aligned} d ( V t ( z t )) = − V t ( z t ) ∂ z V t ( z t ) d ⟨ B 1 ⟩ t + ∂ z V t ( z t ) d z t = − V t ( z t ) ∂ z V t ( z t ) d ⟨ B 1 ⟩ t + ∂ z V t ( z t ) V t ( z t ) d ⟨ B 1 ⟩ t = 0.
Thus, V t ( z t ) = V 0 ( z 0 ) V_t(z_t)=V_0(z_0) V t ( z t ) = V 0 ( z 0 ) . Substituting this identity into (87) gives d z t = V 0 ( z 0 ) d ⟨ B ‾ 1 ⟩ t dz_t=V_0(z_0)d\langle\overline{B}^1\rangle_t d z t = V 0 ( z 0 ) d ⟨ B 1 ⟩ t . Hence, we get
z t = z 0 + V 0 ( z 0 ) ⟨ B ‾ 1 ⟩ t = : Φ t ( z 0 ) , (90) z_t=z_0+V_0(z_0)\langle\overline{B}^1\rangle_t=:\Phi_t(z_0), \tag*{(90)} z t = z 0 + V 0 ( z 0 ) ⟨ B 1 ⟩ t =: Φ t ( z 0 ) , (90)
and V t ( z t ) = V 0 ( z 0 ) V_t(z_t)=V_0(z_0) V t ( z t ) = V 0 ( z 0 ) .
Note that ∣ V 0 ( z ) ∣ ≤ 1 Im z \lvert V_0(z)\rvert\leq\frac{1}{\operatorname{Im}z} ∣ V 0 ( z )∣ ≤ Im z 1 and ∣ V 0 ′ ( z ) ∣ ≤ 1 ( Im z ) 2 \lvert V_0'(z)\rvert\leq\frac{1}{(\operatorname{Im}z)^2} ∣ V 0 ′ ( z )∣ ≤ ( Im z ) 2 1 . Then, for z 0 ∈ D t : = { z ∈ C + : Im z > 2 σ ‾ 2 t } z_0\in D_t:=\{z\in\mathbb{C}_+:\operatorname{Im}z>2\sqrt{\overline{\sigma}^2t}\} z 0 ∈ D t := { z ∈ C + : Im z > 2 σ 2 t } , eq. (90) yields Im z t ≥ Im z 0 − σ ‾ 2 t Im z 0 > 0 \operatorname{Im}z_t\geq\operatorname{Im}z_0-\frac{\overline{\sigma}^2t}{\operatorname{Im}z_0}>0 Im z t ≥ Im z 0 − Im z 0 σ 2 t > 0 and ∣ ∂ z Φ t ( z 0 ) ∣ = ∣ 1 + V 0 ′ ( z 0 ) ⟨ B ‾ 1 ⟩ t ∣ ≥ 1 − σ ‾ 2 t ( Im z 0 ) 2 > 0 \lvert\partial_z\Phi_t(z_0)\rvert=\lvert1+V_0'(z_0)\langle\overline{B}^1\rangle_t\rvert\geq1-\frac{\overline{\sigma}^2t}{(\operatorname{Im}z_0)^2}>0 ∣ ∂ z Φ t ( z 0 )∣ = ∣ 1 + V 0 ′ ( z 0 ) ⟨ B 1 ⟩ t ∣ ≥ 1 − ( Im z 0 ) 2 σ 2 t > 0 . Thus, U t : = Φ t ( D t ) U_t:=\Phi_t(D_t) U t := Φ t ( D t ) is a nonempty open subset of C + \mathbb{C}_+ C + . Now suppose V V V and V ~ \widetilde{V} V are two solutions with the same initial transform V 0 V_0 V 0 . Then V t ( Φ t ( z ) ) = V 0 ( z ) = V ~ t ( Φ t ( z ) ) V_t(\Phi_t(z))=V_0(z)=\widetilde{V}_t(\Phi_t(z)) V t ( Φ t ( z )) = V 0 ( z ) = V t ( Φ t ( z )) for all z ∈ D t z\in D_t z ∈ D t . Therefore, V t ( w ) = V ~ t ( w ) V_t(w)=\widetilde{V}_t(w) V t ( w ) = V t ( w ) for all w ∈ U t w\in U_t w ∈ U t . Since both functions are analytic on the connected domain C + \mathbb{C}_+ C + , the identity theorem yields V t ( w ) = V ~ t ( w ) V_t(w)=\widetilde{V}_t(w) V t ( w ) = V t ( w ) , w ∈ C + w\in\mathbb{C}_+ w ∈ C + . This yields the uniqueness of the solution to (1.18) and (1.19).
Wishart process
Proof of Proposition 1.14. By Theorem 1.7 and Theorem 1.8, we can obtain that the sequence ( L N ( t ) ) t ∈ [ 0 , T ] (L_N(t))_{t\in[0,T]} ( L N ( t ) ) t ∈ [ 0 , T ] is q.s. pre-compact in C ( [ 0 , T ] , P ( R ) ) C([0,T],\mathcal{P}(\mathbb{R})) C ([ 0 , T ] , P ( R )) and that any limit point ( μ t , t ∈ [ 0 , T ] ) (\mu_t,t\in[0,T]) ( μ t , t ∈ [ 0 , T ]) of L N L_N L N satisfies
∫ μ t ( d x ) z − x = ∫ μ 0 ( d x ) z − x + ∫ 0 t [ ∫ c ( z − x ) 2 μ s ( d x ) ] d ⟨ B ˉ 1 ⟩ s + ∫ 0 t [ ∬ x + y ( z − x ) ( z − y ) μ s ( d x ) μ s ( d y ) ] d ⟨ B ˉ 1 ⟩ s \int\frac{\mu_t(\mathrm{d}x)}{z-x}=\int\frac{\mu_0(\mathrm{d}x)}{z-x}+\int_0^t\left[\int\frac{c}{(z-x)^2}\mu_s(\mathrm{d}x)\right]\mathrm{d}\langle\bar{B}^1\rangle_s+\int_0^t\left[\iint\frac{x+y}{(z-x)(z-y)}\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)\right]\mathrm{d}\langle\bar{B}^1\rangle_s ∫ z − x μ t ( d x ) = ∫ z − x μ 0 ( d x ) + ∫ 0 t [ ∫ ( z − x ) 2 c μ s ( d x ) ] d ⟨ B ˉ 1 ⟩ s + ∫ 0 t [ ∬ ( z − x ) ( z − y ) x + y μ s ( d x ) μ s ( d y ) ] d ⟨ B ˉ 1 ⟩ s
for z ∈ C \ R z\in\mathbb{C}\backslash\mathbb{R} z ∈ C \ R . Denote
V t ( z ) = ∫ R 1 z − x μ t ( d x ) , z ∈ C + , V_t(z)=\int_{\mathbb{R}}\frac{1}{z-x}\mu_t(\mathrm{d}x),\qquad z\in\mathbb{C}_+, V t ( z ) = ∫ R z − x 1 μ t ( d x ) , z ∈ C + ,
and we have ∂ z V t ( z ) = − ∫ R 1 ( z − x ) 2 μ t ( d x ) \partial_zV_t(z)=-\int_{\mathbb{R}}\frac{1}{(z-x)^2}\mu_t(\mathrm{d}x) ∂ z V t ( z ) = − ∫ R ( z − x ) 2 1 μ t ( d x ) . Moreover,
∬ x + y ( z − x ) ( z − y ) μ s ( d x ) μ s ( d y ) = ∬ ( x − z ) + ( y − z ) + 2 z ( z − x ) ( z − y ) μ s ( d x ) μ s ( d y ) = ∬ 2 z ( z − x ) ( z − y ) μ s ( d x ) μ s ( d y ) − ∬ z − x ( z − x ) ( z − y ) μ s ( d x ) μ s ( d y ) − ∬ z − y ( z − x ) ( z − y ) μ s ( d x ) μ s ( d y ) = − 2 z V s ( z ) ∂ z V s ( z ) + ∂ z V s ( z ) − V s ( z ) 2 . \begin{aligned}
\iint\frac{x+y}{(z-x)(z-y)}\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)
&=\iint\frac{(x-z)+(y-z)+2z}{(z-x)(z-y)}\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)\\
&=\iint\frac{2z}{(z-x)(z-y)}\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)-\iint\frac{z-x}{(z-x)(z-y)}\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)\\
&\quad-\iint\frac{z-y}{(z-x)(z-y)}\mu_s(\mathrm{d}x)\mu_s(\mathrm{d}y)\\
&=-2zV_s(z)\partial_zV_s(z)+\partial_zV_s(z)-V_s(z)^2.
\end{aligned} ∬ ( z − x ) ( z − y ) x + y μ s ( d x ) μ s ( d y ) = ∬ ( z − x ) ( z − y ) ( x − z ) + ( y − z ) + 2 z μ s ( d x ) μ s ( d y ) = ∬ ( z − x ) ( z − y ) 2 z μ s ( d x ) μ s ( d y ) − ∬ ( z − x ) ( z − y ) z − x μ s ( d x ) μ s ( d y ) − ∬ ( z − x ) ( z − y ) z − y μ s ( d x ) μ s ( d y ) = − 2 z V s ( z ) ∂ z V s ( z ) + ∂ z V s ( z ) − V s ( z ) 2 .
Consequently, V t V_t V t satisfies
d V t ( z ) = − c ∂ z V t ( z ) d ⟨ B ˉ 1 ⟩ t − ( V t ( z ) 2 − ∂ z V t ( z ) + 2 z V t ( z ) ∂ z V t ( z ) ) d ⟨ B ˉ 1 ⟩ t = ∂ z V t ( z ) ( 1 − c − 2 z V t ( z ) ) d ⟨ B ˉ 1 ⟩ t − V t ( z ) 2 d ⟨ B ˉ 1 ⟩ t . (91) \begin{aligned}
\mathrm{d}V_t(z)&=-c\partial_zV_t(z)\mathrm{d}\langle\bar{B}^1\rangle_t-\left(V_t(z)^2-\partial_zV_t(z)+2zV_t(z)\partial_zV_t(z)\right)\mathrm{d}\langle\bar{B}^1\rangle_t\\
&=\partial_zV_t(z)\left(1-c-2zV_t(z)\right)\mathrm{d}\langle\bar{B}^1\rangle_t-V_t(z)^2\mathrm{d}\langle\bar{B}^1\rangle_t. \tag*{(91)}
\end{aligned} d V t ( z ) = − c ∂ z V t ( z ) d ⟨ B ˉ 1 ⟩ t − ( V t ( z ) 2 − ∂ z V t ( z ) + 2 z V t ( z ) ∂ z V t ( z ) ) d ⟨ B ˉ 1 ⟩ t = ∂ z V t ( z ) ( 1 − c − 2 z V t ( z ) ) d ⟨ B ˉ 1 ⟩ t − V t ( z ) 2 d ⟨ B ˉ 1 ⟩ t . (91)
We now prove the uniqueness of an analytic solution to (91) . The uniqueness of the Stieltjes transform V t ( z ) V_t(z) V t ( z ) then implies the uniqueness of the corresponding probability measure μ t \mu_t μ t .
Consider the characteristic equation
d z t = ( c − 1 + 2 z t V t ( z t ) ) d ⟨ B ˉ 1 ⟩ t , z 0 ∈ C + . (92) \mathrm{d}z_t=\left(c-1+2z_tV_t(z_t)\right)\mathrm{d}\langle\bar{B}^1\rangle_t,\qquad z_0\in\mathbb{C}_+. \tag*{(92)} d z t = ( c − 1 + 2 z t V t ( z t ) ) d ⟨ B ˉ 1 ⟩ t , z 0 ∈ C + . (92)
Denote z = u + i v ∈ C + z=u+iv\in\mathbb{C}_+ z = u + i v ∈ C + . It is easy to know that ∣ V t ( z ) ∣ ≤ 1 v \lvert V_t(z)\rvert\le\frac{1}{v} ∣ V t ( z )∣ ≤ v 1 and ∣ ∂ z V t ( z ) ∣ ≤ 1 v 2 \lvert\partial_zV_t(z)\rvert\le\frac{1}{v^2} ∣ ∂ z V t ( z )∣ ≤ v 2 1 . Furthermore, Im V t ( z ) = − ∫ R v ( u − x ) 2 + v 2 μ t ( d x ) ≤ 0 \operatorname{Im}V_t(z)=-\int_{\mathbb{R}}\frac{v}{(u-x)^2+v^2}\mu_t(\mathrm{d}x)\le0 Im V t ( z ) = − ∫ R ( u − x ) 2 + v 2 v μ t ( d x ) ≤ 0 , and ∣ Im V t ( z ) ∣ ≤ 1 v \lvert\operatorname{Im}V_t(z)\rvert\le\frac{1}{v} ∣ Im V t ( z )∣ ≤ v 1 . We also have ∣ Re V t ( z ) ∣ ≤ 1 2 v \lvert\operatorname{Re}V_t(z)\rvert\le\frac{1}{2v} ∣ Re V t ( z )∣ ≤ 2 v 1 . Let y ∗ y^* y ∗ and ρ ∗ \rho^* ρ ∗ be two positive numbers which are to be determined later and define
D : = { z ∈ C + : Im z ≥ y ∗ , ∣ Re z ∣ Im z < ρ ∗ } . (93) D:=\left\{z\in\mathbb{C}_+:\operatorname{Im}z\ge y^*,\ \frac{\lvert\operatorname{Re}z\rvert}{\operatorname{Im}z}<\rho^*\right\}. \tag*{(93)} D := { z ∈ C + : Im z ≥ y ∗ , Im z ∣ Re z ∣ < ρ ∗ } . (93)
Then we have for all t ≥ 0 t\ge0 t ≥ 0 ,
∣ V t ( z ) ∣ ≤ 1 y ∗ , ∣ ∂ z V t ( z ) ∣ ≤ 1 ∣ y ∗ ∣ 2 on D . (94) \lvert V_t(z)\rvert\le\frac{1}{y^*},\qquad\lvert\partial_zV_t(z)\rvert\le\frac{1}{\lvert y^*\rvert^2}\text{ on }D. \tag*{(94)} ∣ V t ( z )∣ ≤ y ∗ 1 , ∣ ∂ z V t ( z )∣ ≤ ∣ y ∗ ∣ 2 1 on D . (94)
We now construct a solution to (92) by Picard iteration. Set z t 0 ≡ z 0 = x 0 + i y 0 ∈ C + z_t^0\equiv z_0=x_0+iy_0\in\mathbb{C}_+ z t 0 ≡ z 0 = x 0 + i y 0 ∈ C + , and for n ≥ 0 n\ge0 n ≥ 0 , define z n + 1 z^{n+1} z n + 1 by
d z t n + 1 = ( c − 1 + 2 z t n V t ( z t n ) ) d ⟨ B ˉ 1 ⟩ t , z 0 n + 1 = z 0 . (95) \mathrm{d}z_t^{n+1}=\left(c-1+2z_t^nV_t(z_t^n)\right)\mathrm{d}\langle\bar{B}^1\rangle_t,\qquad z_0^{n+1}=z_0. \tag*{(95)} d z t n + 1 = ( c − 1 + 2 z t n V t ( z t n ) ) d ⟨ B ˉ 1 ⟩ t , z 0 n + 1 = z 0 . (95)
Denote z t n = x t n + i y t n z_t^n=x_t^n+iy_t^n z t n = x t n + i y t n , ρ t n : = x t n y t n \rho_t^n:=\frac{x_t^n}{y_t^n} ρ t n := y t n x t n , V t ( z t n ) = a t n + i b t n V_t(z_t^n)=a_t^n+ib_t^n V t ( z t n ) = a t n + i b t n . If y t n > 0 y_t^n>0 y t n > 0 , we have b t n = Im V t ( z t n ) ≤ 0 b_t^n=\operatorname{Im}V_t(z_t^n)\le0 b t n = Im V t ( z t n ) ≤ 0 , and, by the estimates established above, ∣ a t n ∣ ≤ 1 2 y t n |a_t^n|\le\frac{1}{2y_t^n} ∣ a t n ∣ ≤ 2 y t n 1 , ∣ b t n ∣ ≤ 1 y t n |b_t^n|\le\frac{1}{y_t^n} ∣ b t n ∣ ≤ y t n 1 . Furthermore,
z t n V t ( z t n ) = ( x t n + i y t n ) ( a t n + i b t n ) = x t n a t n − y t n b t n + i ( x t n b t n + y t n a t n ) . z_t^nV_t(z_t^n)=(x_t^n+iy_t^n)(a_t^n+ib_t^n)=x_t^na_t^n-y_t^nb_t^n+i(x_t^nb_t^n+y_t^na_t^n). z t n V t ( z t n ) = ( x t n + i y t n ) ( a t n + i b t n ) = x t n a t n − y t n b t n + i ( x t n b t n + y t n a t n ) .
The real and imaginary parts in (95) yields respectively,
d x t n + 1 = ( 2 x t n a t n − 2 y t n b t n − 1 + c ) d ⟨ B ˉ 1 ⟩ t (96) dx_t^{n+1}=(2x_t^na_t^n-2y_t^nb_t^n-1+c)\,d\langle\bar{B}^{1}\rangle_t \tag*{(96)} d x t n + 1 = ( 2 x t n a t n − 2 y t n b t n − 1 + c ) d ⟨ B ˉ 1 ⟩ t (96)
d y t n + 1 = 2 ( x t n b t n + y t n a t n ) d ⟨ B ˉ 1 ⟩ t . (97) dy_t^{n+1}=2(x_t^nb_t^n+y_t^na_t^n)\,d\langle\bar{B}^{1}\rangle_t. \tag*{(97)} d y t n + 1 = 2 ( x t n b t n + y t n a t n ) d ⟨ B ˉ 1 ⟩ t . (97)
Assume that y t n > 0 y_t^n>0 y t n > 0 and ∣ ρ t n ∣ ≤ ρ ∗ |\rho_t^n|\le\rho^* ∣ ρ t n ∣ ≤ ρ ∗ . Then ∣ x t n a t n ∣ ≤ ∣ x t n ∣ 2 y t n = 1 2 ∣ ρ t n ∣ ≤ ρ ∗ 2 |x_t^na_t^n|\le\frac{|x_t^n|}{2y_t^n}=\frac{1}{2}|\rho_t^n|\le\frac{\rho^*}{2} ∣ x t n a t n ∣ ≤ 2 y t n ∣ x t n ∣ = 2 1 ∣ ρ t n ∣ ≤ 2 ρ ∗ , and
0 ≤ − y t n b t n ≤ 1 , ∣ x t n b t n ∣ ≤ ∣ x t n ∣ y t n = ∣ ρ t n ∣ , ∣ y t n a t n ∣ ≤ 1 2 . 0\le-y_t^nb_t^n\le1,\qquad|x_t^nb_t^n|\le\frac{|x_t^n|}{y_t^n}=|\rho_t^n|,\qquad|y_t^na_t^n|\le\frac{1}{2}. 0 ≤ − y t n b t n ≤ 1 , ∣ x t n b t n ∣ ≤ y t n ∣ x t n ∣ = ∣ ρ t n ∣ , ∣ y t n a t n ∣ ≤ 2 1 .
It follows from (96) that
( c − ρ ∗ − 1 ) d ⟨ B ˉ 1 ⟩ t ≤ d x t n + 1 ≤ ( c + ρ ∗ + 1 ) d ⟨ B ˉ 1 ⟩ t (c-\rho^*-1)\,d\langle\bar{B}^{1}\rangle_t\le dx_t^{n+1}\le(c+\rho^*+1)\,d\langle\bar{B}^{1}\rangle_t ( c − ρ ∗ − 1 ) d ⟨ B ˉ 1 ⟩ t ≤ d x t n + 1 ≤ ( c + ρ ∗ + 1 ) d ⟨ B ˉ 1 ⟩ t
and from (97) that
− ( 2 ρ ∗ + 1 ) d ⟨ B ˉ 1 ⟩ t ≤ d y t n + 1 ≤ ( 2 ρ ∗ + 1 ) d ⟨ B ˉ 1 ⟩ t . -(2\rho^*+1)\,d\langle\bar{B}^{1}\rangle_t\le dy_t^{n+1}\le(2\rho^*+1)\,d\langle\bar{B}^{1}\rangle_t. − ( 2 ρ ∗ + 1 ) d ⟨ B ˉ 1 ⟩ t ≤ d y t n + 1 ≤ ( 2 ρ ∗ + 1 ) d ⟨ B ˉ 1 ⟩ t .
Then we obtain that
∣ x t n + 1 ∣ ≤ ∣ x 0 ∣ + ( c + ρ ∗ + 1 ) σ ˉ 2 T (98) |x_t^{n+1}|\le|x_0|+(c+\rho^*+1)\bar{\sigma}^2T \tag*{(98)} ∣ x t n + 1 ∣ ≤ ∣ x 0 ∣ + ( c + ρ ∗ + 1 ) σ ˉ 2 T (98)
y t n + 1 ≥ y 0 − ( 2 ρ ∗ + 1 ) σ ˉ 2 T . (99) y_t^{n+1}\ge y_0-(2\rho^*+1)\bar{\sigma}^2T. \tag*{(99)} y t n + 1 ≥ y 0 − ( 2 ρ ∗ + 1 ) σ ˉ 2 T . (99)
Given ρ ∗ > 0 \rho^*>0 ρ ∗ > 0 , we can choose y 0 > 0 y_0>0 y 0 > 0 sufficiently large such that y ∗ : = y 0 − ( 2 ρ ∗ + 1 ) σ ˉ 2 T > 0 y^*:=y_0-(2\rho^*+1)\bar{\sigma}^2T>0 y ∗ := y 0 − ( 2 ρ ∗ + 1 ) σ ˉ 2 T > 0 and C 0 y ∗ ≤ ρ ∗ \frac{C_0}{y^*}\le\rho^* y ∗ C 0 ≤ ρ ∗ , where C 0 : = ∣ x 0 ∣ + c σ ˉ 2 T + ( ρ ∗ + 1 ) σ ˉ 2 T C_0:=|x_0|+c\bar{\sigma}^2T+(\rho^*+1)\bar{\sigma}^2T C 0 := ∣ x 0 ∣ + c σ ˉ 2 T + ( ρ ∗ + 1 ) σ ˉ 2 T . Then (99) implies that y t n + 1 ≥ y ∗ y_t^{n+1}\ge y^* y t n + 1 ≥ y ∗ for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , and consequently, ∣ ρ t n + 1 ∣ = ∣ x t n + 1 ∣ y t n + 1 ≤ C 0 y ∗ ≤ ρ ∗ |\rho_t^{n+1}|=\frac{|x_t^{n+1}|}{y_t^{n+1}}\le\frac{C_0}{y^*}\le\rho^* ∣ ρ t n + 1 ∣ = y t n + 1 ∣ x t n + 1 ∣ ≤ y ∗ C 0 ≤ ρ ∗ . Thus, we get y t n ≥ y ∗ y_t^n\ge y^* y t n ≥ y ∗ , ∣ ρ t n ∣ ≤ ρ ∗ |\rho_t^n|\le\rho^* ∣ ρ t n ∣ ≤ ρ ∗ for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] and n ∈ N n\in\mathbb{N} n ∈ N . Moreover, there exist constants C x 0 , C y 0 > 0 C_{x_0},C_{y_0}>0 C x 0 , C y 0 > 0 , independent of n n n , such that ∣ x t n ∣ ≤ C x 0 |x_t^n|\le C_{x_0} ∣ x t n ∣ ≤ C x 0 , y ∗ ≤ y t n ≤ C y 0 y^*\le y_t^n\le C_{y_0} y ∗ ≤ y t n ≤ C y 0 , t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Therefore, we have z t n ∈ D z_t^n\in D z t n ∈ D and ∣ z t n ∣ ≤ C T : = C x 0 2 + C y 0 2 |z_t^n|\le C_T:=\sqrt{C_{x_0}^2+C_{y_0}^2} ∣ z t n ∣ ≤ C T := C x 0 2 + C y 0 2 for all n ≥ 0 n\ge0 n ≥ 0 and t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
We now prove the convergence of the Picard iteration. Define F t ( w ) : = 2 w V t ( w ) − 1 + c F_t(w):=2wV_t(w)-1+c F t ( w ) := 2 w V t ( w ) − 1 + c . For w 1 , w 2 ∈ D w_1,w_2\in D w 1 , w 2 ∈ D with ∣ w 2 ∣ ≤ C T |w_2|\le C_T ∣ w 2 ∣ ≤ C T , we have by (94) ,
∣ F t ( w 1 ) − F t ( w 2 ) ∣ = 2 ∣ w 1 V t ( w 1 ) − w 2 V t ( w 2 ) ∣ ≤ 2 ∣ w 1 − w 2 ∣ ∣ V t ( w 1 ) ∣ + 2 ∣ w 2 ∣ ∣ V t ( w 1 ) − V t ( w 2 ) ∣ ≤ 2 ( 1 y ∗ + C T ( y ∗ ) 2 ) ∣ w 1 − w 2 ∣ . (100) \begin{aligned}
|F_t(w_1)-F_t(w_2)|&=2|w_1V_t(w_1)-w_2V_t(w_2)|\\
&\le2|w_1-w_2||V_t(w_1)|+2|w_2||V_t(w_1)-V_t(w_2)|\\
&\le2\left(\frac{1}{y^*}+\frac{C_T}{(y^*)^2}\right)|w_1-w_2|. \tag*{(100)}
\end{aligned} ∣ F t ( w 1 ) − F t ( w 2 ) ∣ = 2∣ w 1 V t ( w 1 ) − w 2 V t ( w 2 ) ∣ ≤ 2∣ w 1 − w 2 ∣∣ V t ( w 1 ) ∣ + 2∣ w 2 ∣∣ V t ( w 1 ) − V t ( w 2 ) ∣ ≤ 2 ( y ∗ 1 + ( y ∗ ) 2 C T ) ∣ w 1 − w 2 ∣. (100)
By (95) , we get
∣ z t n + 1 − z t n ∣ ≤ ∫ 0 t ∣ F s ( z s n ) − F s ( z s n − 1 ) ∣ d ⟨ B ˉ 1 ⟩ s ≤ L ∫ 0 t ∣ z s n − z s n − 1 ∣ d ⟨ B ˉ 1 ⟩ s , |z_t^{n+1}-z_t^n|\le\int_0^t|F_s(z_s^n)-F_s(z_s^{n-1})|\,d\langle\bar{B}^{1}\rangle_s\le L\int_0^t|z_s^n-z_s^{n-1}|\,d\langle\bar{B}^{1}\rangle_s, ∣ z t n + 1 − z t n ∣ ≤ ∫ 0 t ∣ F s ( z s n ) − F s ( z s n − 1 ) ∣ d ⟨ B ˉ 1 ⟩ s ≤ L ∫ 0 t ∣ z s n − z s n − 1 ∣ d ⟨ B ˉ 1 ⟩ s ,
where L : = 2 ( 1 y ∗ + C T ( y ∗ ) 2 ) L:=2\left(\frac{1}{y^*}+\frac{C_T}{(y^*)^2}\right) L := 2 ( y ∗ 1 + ( y ∗ ) 2 C T ) . Thus, we have, denoting A n ( t ) : = sup 0 ≤ s ≤ t ∣ z s n + 1 − z s n ∣ A_n(t):=\sup_{0\le s\le t}|z_s^{n+1}-z_s^n| A n ( t ) := sup 0 ≤ s ≤ t ∣ z s n + 1 − z s n ∣ ,
A n ( t ) ≤ L ∫ 0 t A n − 1 ( s ) d ⟨ B ˉ 1 ⟩ s ≤ σ ˉ 2 L ∫ 0 t A n − 1 ( s ) d s , A_n(t)\le L\int_0^t A_{n-1}(s)\,d\langle\bar{B}^{1}\rangle_s\le\bar{\sigma}^2L\int_0^t A_{n-1}(s)\,ds, A n ( t ) ≤ L ∫ 0 t A n − 1 ( s ) d ⟨ B ˉ 1 ⟩ s ≤ σ ˉ 2 L ∫ 0 t A n − 1 ( s ) d s ,
which yields
A n ( t ) ≤ ( L σ ˉ 2 t ) n n ! A 0 ( T ) , t ∈ [ 0 , T ] . A_n(t)\le\frac{(L\bar{\sigma}^2t)^n}{n!}A_0(T),\qquad t\in[0,T]. A n ( t ) ≤ n ! ( L σ ˉ 2 t ) n A 0 ( T ) , t ∈ [ 0 , T ] .
Therefore, the sequence { z n } n ≥ 0 \{z^n\}_{n\ge0} { z n } n ≥ 0 is uniformly Cauchy on [ 0 , T ] [0,T] [ 0 , T ] . It converges uniformly to a continuous function z t z_t z t satisfying
z t = z 0 + ∫ 0 t F s ( z s ) d ⟨ B ˉ 1 ⟩ s . z_t=z_0+\int_0^t F_s(z_s)\,\mathrm{d}\langle\bar{B}^1\rangle_s. z t = z 0 + ∫ 0 t F s ( z s ) d ⟨ B ˉ 1 ⟩ s .
Thus, the existence of the solution to (92) is established.
Let z t z_t z t be a solution to (92) and V t ( z ) V_t(z) V t ( z ) a solution to (27) . By the chain rule, we get
d ( V t ( z t ) ) = ( − c ∂ z V t ( z t ) − V t ( z t ) 2 + ∂ z V t ( z t ) − 2 z t ∂ z V t ( z t ) ) d ⟨ B ˉ 1 ⟩ t + ∂ z V t ( z t ) d z t = ( − c ∂ z V t ( z t ) − V t ( z t ) 2 + ∂ z V t ( z t ) − 2 z t ∂ z V t ( z t ) ) d ⟨ B ˉ 1 ⟩ t + ∂ z V t ( z t ) ( 2 z t ∂ z V t ( z t ) − 1 + c ) d ⟨ B ˉ 1 ⟩ t = − V t ( z t ) 2 d ⟨ B ˉ 1 ⟩ t . \begin{aligned}
\mathrm{d}(V_t(z_t))={}&\left(-c\partial_zV_t(z_t)-V_t(z_t)^2+\partial_zV_t(z_t)-2z_t\partial_zV_t(z_t)\right)\mathrm{d}\langle\bar{B}^1\rangle_t+\partial_zV_t(z_t)\,\mathrm{d}z_t\\
={}&\left(-c\partial_zV_t(z_t)-V_t(z_t)^2+\partial_zV_t(z_t)-2z_t\partial_zV_t(z_t)\right)\mathrm{d}\langle\bar{B}^1\rangle_t+\partial_zV_t(z_t)\left(2z_t\partial_zV_t(z_t)-1+c\right)\mathrm{d}\langle\bar{B}^1\rangle_t\\
={}&-V_t(z_t)^2\mathrm{d}\langle\bar{B}^1\rangle_t.
\end{aligned} d ( V t ( z t )) = = = ( − c ∂ z V t ( z t ) − V t ( z t ) 2 + ∂ z V t ( z t ) − 2 z t ∂ z V t ( z t ) ) d ⟨ B ˉ 1 ⟩ t + ∂ z V t ( z t ) d z t ( − c ∂ z V t ( z t ) − V t ( z t ) 2 + ∂ z V t ( z t ) − 2 z t ∂ z V t ( z t ) ) d ⟨ B ˉ 1 ⟩ t + ∂ z V t ( z t ) ( 2 z t ∂ z V t ( z t ) − 1 + c ) d ⟨ B ˉ 1 ⟩ t − V t ( z t ) 2 d ⟨ B ˉ 1 ⟩ t .
Then, d V t ( z t ) V t ( z t ) 2 = − d ⟨ B ˉ 1 ⟩ t \frac{\mathrm{d}V_t(z_t)}{V_t(z_t)^2}=-\mathrm{d}\langle\bar{B}^1\rangle_t V t ( z t ) 2 d V t ( z t ) = − d ⟨ B ˉ 1 ⟩ t , which yields − 1 V t ( z t ) + 1 V 0 ( z 0 ) = − ⟨ B ˉ 1 ⟩ t -\frac{1}{V_t(z_t)}+\frac{1}{V_0(z_0)}=-\langle\bar{B}^1\rangle_t − V t ( z t ) 1 + V 0 ( z 0 ) 1 = − ⟨ B ˉ 1 ⟩ t . Therefore, we get
V t ( z t ) = V 0 ( z 0 ) 1 + V 0 ( z 0 ) ⟨ B ˉ 1 ⟩ t . (101) V_t(z_t)=\frac{V_0(z_0)}{1+V_0(z_0)\langle\bar{B}^1\rangle_t}. \tag*{(101)} V t ( z t ) = 1 + V 0 ( z 0 ) ⟨ B ˉ 1 ⟩ t V 0 ( z 0 ) . (101)
Now, one can verify directly that
z t = z 0 ( 1 + V 0 ( z 0 ) ⟨ B ˉ 1 ⟩ t ) 2 + ( c − 1 ) ⟨ B ˉ 1 ⟩ t ( 1 + V 0 ( z 0 ) ⟨ B ˉ 1 ⟩ t ) z_t=z_0\left(1+V_0(z_0)\langle\bar{B}^1\rangle_t\right)^2+(c-1)\langle\bar{B}^1\rangle_t\left(1+V_0(z_0)\langle\bar{B}^1\rangle_t\right) z t = z 0 ( 1 + V 0 ( z 0 ) ⟨ B ˉ 1 ⟩ t ) 2 + ( c − 1 ) ⟨ B ˉ 1 ⟩ t ( 1 + V 0 ( z 0 ) ⟨ B ˉ 1 ⟩ t )
is the unique solution to (92) .
Denote
Ψ t ( z ) : = z ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) 2 + ( c − 1 ) ⟨ B ˉ 1 ⟩ t ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) . \Psi_t(z):=z\left(1+V_0(z)\langle\bar{B}^1\rangle_t\right)^2+(c-1)\langle\bar{B}^1\rangle_t\left(1+V_0(z)\langle\bar{B}^1\rangle_t\right). Ψ t ( z ) := z ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) 2 + ( c − 1 ) ⟨ B ˉ 1 ⟩ t ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) .
Then, we get
∂ z Ψ t ( z ) = ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) 2 + [ 2 z ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) ⟨ B ˉ 1 ⟩ t + ( c − 1 ) ⟨ B ˉ 1 ⟩ t 2 ] V 0 ′ ( z ) , \partial_z\Psi_t(z)=\left(1+V_0(z)\langle\bar{B}^1\rangle_t\right)^2+\left[2z\left(1+V_0(z)\langle\bar{B}^1\rangle_t\right)\langle\bar{B}^1\rangle_t+(c-1)\langle\bar{B}^1\rangle_t^2\right]V_0'(z), ∂ z Ψ t ( z ) = ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) 2 + [ 2 z ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) ⟨ B ˉ 1 ⟩ t + ( c − 1 ) ⟨ B ˉ 1 ⟩ t 2 ] V 0 ′ ( z ) ,
and for z ∈ D z\in D z ∈ D , by (94) ,
∣ ∂ z Ψ t ( z ) − 1 ∣ = ∣ ( V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) 2 + 2 V 0 ( z ) ⟨ B ˉ 1 ⟩ t + [ 2 z ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) ⟨ B ˉ 1 ⟩ t + ( c − 1 ) ⟨ B ˉ 1 ⟩ t 2 ] V 0 ′ ( z ) ∣ ≤ ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 + 2 ⟨ B ˉ 1 ⟩ t Im z + 2 ∣ z ∣ ⟨ B ˉ 1 ⟩ t ( Im z ) 2 + 2 ∣ z ∣ ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 3 + ( c − 1 ) ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 = c ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 + 2 ( ∣ z ∣ + Im z ) ⟨ B ˉ 1 ⟩ t ( Im z ) 2 + 2 ∣ z ∣ ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 3 ≤ c ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 + 2 ( 1 + ( ρ ∗ ) 2 Im z + Im z ) ⟨ B ˉ 1 ⟩ t ( Im z ) 2 + 2 1 + ( ρ ∗ ) 2 Im z ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 3 = c ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 + 2 ( 1 + ( ρ ∗ ) 2 + 1 ) ⟨ B ˉ 1 ⟩ t Im z + 2 1 + ( ρ ∗ ) 2 ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 = : E ( z ) . \begin{aligned}
|\partial_z\Psi_t(z)-1|
={}&\left|\left(V_0(z)\langle\bar{B}^1\rangle_t\right)^2+2V_0(z)\langle\bar{B}^1\rangle_t+\left[2z\left(1+V_0(z)\langle\bar{B}^1\rangle_t\right)\langle\bar{B}^1\rangle_t+(c-1)\langle\bar{B}^1\rangle_t^2\right]V_0'(z)\right|\\
\le{}&\frac{\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^2}+\frac{2\langle\bar{B}^1\rangle_t}{\operatorname{Im}z}+\frac{2|z|\langle\bar{B}^1\rangle_t}{(\operatorname{Im}z)^2}+\frac{2|z|\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^3}+\frac{(c-1)\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^2}\\
={}&\frac{c\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^2}+\frac{2(|z|+\operatorname{Im}z)\langle\bar{B}^1\rangle_t}{(\operatorname{Im}z)^2}+\frac{2|z|\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^3}\\
\le{}&\frac{c\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^2}+\frac{2\left(\sqrt{1+(\rho^*)^2}\operatorname{Im}z+\operatorname{Im}z\right)\langle\bar{B}^1\rangle_t}{(\operatorname{Im}z)^2}+\frac{2\sqrt{1+(\rho^*)^2}\operatorname{Im}z\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^3}\\
={}&\frac{c\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^2}+\frac{2\left(\sqrt{1+(\rho^*)^2}+1\right)\langle\bar{B}^1\rangle_t}{\operatorname{Im}z}+\frac{2\sqrt{1+(\rho^*)^2}\langle\bar{B}^1\rangle_t^2}{(\operatorname{Im}z)^2}\\
={}&:E(z).
\end{aligned} ∣ ∂ z Ψ t ( z ) − 1∣ = ≤ = ≤ = = ( V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) 2 + 2 V 0 ( z ) ⟨ B ˉ 1 ⟩ t + [ 2 z ( 1 + V 0 ( z ) ⟨ B ˉ 1 ⟩ t ) ⟨ B ˉ 1 ⟩ t + ( c − 1 ) ⟨ B ˉ 1 ⟩ t 2 ] V 0 ′ ( z ) ( Im z ) 2 ⟨ B ˉ 1 ⟩ t 2 + Im z 2 ⟨ B ˉ 1 ⟩ t + ( Im z ) 2 2∣ z ∣ ⟨ B ˉ 1 ⟩ t + ( Im z ) 3 2∣ z ∣ ⟨ B ˉ 1 ⟩ t 2 + ( Im z ) 2 ( c − 1 ) ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 c ⟨ B ˉ 1 ⟩ t 2 + ( Im z ) 2 2 ( ∣ z ∣ + Im z ) ⟨ B ˉ 1 ⟩ t + ( Im z ) 3 2∣ z ∣ ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 c ⟨ B ˉ 1 ⟩ t 2 + ( Im z ) 2 2 ( 1 + ( ρ ∗ ) 2 Im z + Im z ) ⟨ B ˉ 1 ⟩ t + ( Im z ) 3 2 1 + ( ρ ∗ ) 2 Im z ⟨ B ˉ 1 ⟩ t 2 ( Im z ) 2 c ⟨ B ˉ 1 ⟩ t 2 + Im z 2 ( 1 + ( ρ ∗ ) 2 + 1 ) ⟨ B ˉ 1 ⟩ t + ( Im z ) 2 2 1 + ( ρ ∗ ) 2 ⟨ B ˉ 1 ⟩ t 2 : E ( z ) .
We can choose an interior point z ∗ ∈ D z^*\in D z ∗ ∈ D with Im z ∗ \operatorname{Im}z^* Im z ∗ sufficiently large such that E ( z ∗ ) < 1 E(z^*)<1 E ( z ∗ ) < 1 . Then, ∂ z Ψ t ( z ∗ ) ≠ 0 \partial_z\Psi_t(z^*)\ne0 ∂ z Ψ t ( z ∗ ) = 0 and hence there is an open neighborhood U U U of z ∗ z^* z ∗ satisfying U ⊂ D U\subset D U ⊂ D such that Ψ t ( U ) ⊂ C + \Psi_t(U)\subset\mathbb{C}_+ Ψ t ( U ) ⊂ C + is a nonempty open set. Let V V V and V ~ \widetilde{V} V be two solutions with the same initial condition V 0 V_0 V 0 . Then, by (101) we must have V t ( w ) = V ~ t ( w ) V_t(w)=\widetilde{V}_t(w) V t ( w ) = V t ( w ) for all w ∈ Ψ t ( U ) w\in\Psi_t(U) w ∈ Ψ t ( U ) . Noting that V t V_t V t and V ~ t \widetilde{V}_t V t are analytic on C + \mathbb{C}_+ C + , the identity theorem implies that V t ( w ) = V ~ t ( w ) V_t(w)=\widetilde{V}_t(w) V t ( w ) = V t ( w ) , w ∈ C + w\in\mathbb{C}_+ w ∈ C + . This yields the uniqueness of the solution of (27) . □
Acknowledgment
J. Song wishes to thank Mingshang Hu and Yongsheng Song for helpful discussions. J. Song is partially supported by National Key R&D Program of China (Grant No. 2024YFA1013501) and the Fundamental Research Funds for the Central Universities. S. Yang’s work supported by the National Natural Science Foundation of China (Grant No.12471450) and Taishan Scholar Talent Project Youth Project.
References
Research Center for Mathematics and Interdisciplinary Sciences; Frontiers Science Center for Nonlinear Expectations, Ministry of Education, Shandong University, Qingdao 266237, China
Email address: lixinpeng@sdu.edu.cn
Research Center for Mathematics and Interdisciplinary Sciences; Frontiers Science Center for Nonlinear Expectations, Ministry of Education, Shandong University, Qingdao 266237, China
Email address: 202317071@mail.sdu.edu.cn
Research Center for Mathematics and Interdisciplinary Sciences; Frontiers Science Center for Nonlinear Expectations, Ministry of Education, Shandong University, Qingdao 266237, China
Email address: txjsong@sdu.edu.cn
Zhongtai Securities Institute for Financial Studies, Shandong University, Jinan, Shandong, 250100, China.
Email address: yangsz@sdu.edu.cn