Abstract
This article aims to give a formula for differentiating, with respect to , an expression of the form , where and is a diffusion process starting from , taking values in a manifold, and the expectation is taken with respect to the law of this process. is a trace class operator defined by , where , are locally Lipschitz, positive matrices.
1. Introduction
Suppose we have a differentiable manifold of dimension . By the Whitney's embedding theorem, there exists an embedding such that is a closed subset of . It turns out that will do. We will identify with the image and assume that itself is a closed submanifold of . We will also assume that does not have a boundary. Let be a one-point compactification of .
Definition 1.1. An -valued path with explosion time is a continuous map such that for and for all if . The space of -valued paths with explosion time is called the path space of and is denoted by .
Let be a filtered probability space and let be a smooth second-order elliptic operator on . Using the coordinates of the ambient space , and extending smoothly to in the ambient space, we may writewith , a positive matrix. Since is smooth, its square root is locally Lipschitz. Construct a time homogeneous Itô diffusion process which solves the following stochastic differential equation:in the ambient space , where is -dimensional Euclidean-Brownian motion and such thatfor some constant dependent on an open ball centered at with radius . Till the explosion time , for . On , . Furthermore, is an -diffusion measure on . As a result, we use to be the probability measure on . Refer to [1] for a more detailed description.
Fix some positive integer . Let and be the spaces of real-valued matrices and symmetric real-valued matrices, respectively. Also let be the space of nonnegative matrices. Suppose that is locally Lipschitz with if , , and is locally Lipschitz. Also assume that is bounded, where is the operator norm. Define as the multiplication operator with and as the integral operator with kernel , that is, for any ,where is the minimum of and . Note that under the assumptions on , is a positive operator and is trace class (see Proposition 3.1.).
Consider the following integral operator: ,It is a fact that for any trace class operator , if we let denote the trace of , thenfor some constant . Here, means taking the trace of a matrix. Thus is trace class. Therefore,Hence the Fredholm determinant is bounded for each .
Let be continuous bounded functions taking . Fix some number . Define a function bywhere the expectation is taken with respect to and the paths start from . Note that is finite for any from the above discussion. Let such thatThe main result is as follows.
Theorem 1.2. Let . Then where Here, is the component of the matrix .
Clearly, is not in the Feynman-Kac formula form using the process . The idea is to construct a diffusion process such thatand is given byIf we can achieve this, then our result follows from a simple application of the Feynman-Kac formula. Proving (1.12) requires the following 2 steps.
First, we have to prove an essential formula for the derivative of with respect to , given bywhere is a complex number and is some adapted process. For a precise definition of , see (4.1), with replaced by . When , . The goal is to show that the formula holds for by analytic continuation.
Fix a time . By making small such that , we can use the perturbation formula and apply it to the determinant; see (2.3). Differentiating this equation with respect to will give us (1.14). By analytic continuation, we can extend the formula in some domain containing the origin, provided we avoid the poles of the resolvent of for all . If , then (1.14) holds with . By integrating both sides and raising to the exponent, we will get (1.12). Note that if for some time , (1.14) and hence (1.12) hold. The details are given in Sections 2 and 3.
Now assume that (1.14) holds with . The second step consists of constructing a diffusion process from by using a stochastic differential equation. To do this, we differentiate with respect to and show that it satisfies the differential equationand hence satisfies the following stochastic differential equation:with explosion time . From this stochastic differential equation, it is clear that is a diffusion process and by replacing the Fredholm determinant by the formula in (1.12), can be written as a Feynman-Kac form using this process . However, if , then (1.12) may fail to hold.
The positivity of and are used to show that exists for all time and hence (1.14) holds at . This will also imply that . In particular, only the positivity of is required to show that is a trace class operator. To avoid , we can restrict ourselves to small time such that (2.24) holds true.
We can weaken our assumptions on and by not insisting that they are symmetric matrices. If we only assume that is trace class, then we can replace with . Under these weaker assumptions, we have the following result.
Theorem 1.3. Suppose that, for a given locally Lipschitz and , is trace class. Assume that there exists some constant such that Let . Then for all , where
From Lemma 2.4, using the assumptions on and , . If , then the norm is less than 1. Hence (1.14) holds and thus (1.12) holds true.
2. Functional Analytic Tools
Notation 2.1. Suppose that is an integral operator, acting on . We will write to mean where . To distinguish the operator from its kernel, we will write to refer to its kernel. This may be confusing, but it is used to avoid too many symbols. In this article, our integral operator is always trace class and the kernel is a continuous matrix-valued function. By abuse of notation, refers to the kernel of the integral operator .
Notation 2.2. We will use to denote the trace of a matrix and to denote taking the trace of a trace class operator. will denote the operator norm.
It is well known that for a trace class operator and , is a meromorphic function and has singularities at points such that . Define the determinant , given byHowever, this determinant, also known as the Fredholm determinant of , is analytic in because the singularities such that are removable; see [2, Lemma 16].
We want to differentiate the function with respect to , where we write to denote the dependence on the domain . If the kernel of is given by , then for small such that , using the perturbation formula,If we let , thenThus the series in the exponent converges absolutely.
We will define the resolvent by . Since we can write , the kernel of can be written asWhen we write , we meanWe will also write the resolvent of the operator , as . One more point to note is that we assume that .
Now the operator is an operator defined on different Hilbert spaces . Therefore, we will now think of our operator as acting on , defined aswhere is the characteristic function. Hence now our operator has a kernel dependent on the parameter . Note that our is continuous in the variable but is discontinuous at . Thus when we write , we mean
Definition 2.3. Let be a continuous matrix-valued function and let be the matrix norm of . Define to be the maximum value of on .
The next lemma allows us to control the operator norm of the operator by controlling the sup norm of the kernel.
Lemma 2.4. For ,
Proof. For and any ,Hence,for all .
Lemma 2.5. Fix a . For any such that and if is continuous, then
Proof. Since is fixed, we will replace by and hence assume that . Lemma 2.4 tells us that and thus (2.3) holds true. Taking on both sides of (2.3), we haveNowwhere . Differentiate with respect to and using the fundamental theorem of calculus, we getLet . Thus,Thus
Lemma 2.6. Let be continuous. For all such that for all , then is continuous.
Proof. Fix a and write as . By assumption, is invertible for all . By Lemma 2.4, as . Note thatand if we let , thenBy the open mapping theorem, because is a surjective continuous map, it is an open map. Therefore, its inverse is a bounded operator. Since , thus and hence converges to as . This shows that is continuous in . Note thatSince , and are continuous in , hence is continuous in .
Lemma 2.7. Let be continuous. If there exists an open-connected set containing 0 such that for , is analytic for all , then for all .
Proof. For all , is invertible for all and hence is analytic in . Therefore, it follows that is analytic in , becauseBy Lemma 2.4, for all . Thus if we choose , then is an open set containing 0 and for , for all . From Lemma 2.5, for ,Since is also analytic in and agrees with in , it follows that both functions are equal for all .
The proof in the previous theorem gives us the existence of a small neighbourhood containing 0 such that (2.21) holds. Hence we have the following corollary.
Corollary 2.8. Fix and . There exists an open set such that (2.21) holds for all .
Corollary 2.9. Let be an open-connected set as in Lemma 2.7 such that . Then for ,
Proof. The corollary follows from differentiating (2.21). By Lemma 2.6, is continuous and hence the fundamental theorem of calculus applies.
3. Fredholm Determinant
The kernel we are interested in is , for some process . More generally, the kernel we are interested in is of the form for some continuous matrix-valued . The Hilbert space is for some positive number . Without any ambiguity, we will in future write this space as . We will also use to denote the norm.
Proposition 3.1. If is continuous, for any and continuous, then as defined in Section 1 is a trace class operator.
To prove this result, we need the following theorem, which is [3, Theorem 2.12].
Theorem 3.2. Let be a Baire measure on a locally compact space . Let be a function on which is continuous and Hermitian positive, that is, for any , and for any . Then for all . Suppose that, in addition, Then there exists a unique trace class integral operator such that
Proof of Proposition 3.1. Let and be Lebesgue measure. Using Theorem 3.2, it suffices to show that is Hermitian positive. Let be any complex column vectors. Note that there are entries in each column and the entries are complex valued. Let . The proof is obtained using induction. Clearly, when , it is trivial. Suppose it is true for all values from . By relabelling, we can assume that , . Hence for any . Let be the usual dot product. ThenTherefore,Since and hence by assumption on . Thus by induction hypothesis, (replace by ),and hence
Notation 3.3. By abuse of notation, will denote integration over , where should be interpreted as matrix multiplication or inner product, depending on the context. To ease the notation, we will write in future .
Remark 3.4. If we assume that is strictly positive if , which is the case we are interested in this article, then the proof of Proposition 3.1 shows that the operator with kernel is strictly positive, that is, if . This follows using a Riemann lower sum approximation on a double integral and that for any complex vectors ,under the strict positivity assumptions.
The next proposition is a crucial statement. For the time being, we will assume that exists for any time without any further justification. Later on, we will prove that for our operator , this is true; (see Proposition 5.2.). Writing in our new notation, we obtain the next proposition from Corollary 2.9.
Proposition 3.5. Let be an integral operator with kernel for some continuous matrix-valued :
Proof. We will use (2.5). SoTaking trace completes the proof.
Definition 3.6. Let
To ease the notation, we will now write and to be the kernel of . Note that in future we will drop the subscript from the operator and it should be understood that is dependent on . Operators with a prime will denote its derivative with respect to . Our task now is to differentiate .
Define a distributional kernelwhere is the Dirac delta function and is the resolvent.
For any operator depending smoothly on some parameter , we have the differentiation formulaFor the integral operator , its kernel is given, by the fundamental theorem of calculus, byand hence combining with (3.14), we have
Notation 3.7. Let be an integral operator with kernel . We define the adjoint by Here, is an matrix-valued function. We will also write and .
The following lemma defines the relationship between and .
Lemma 3.8. It holds that
Proof. We write the identity operator as , where is Dirac delta function. Then
Theorem 3.9. satisfies the following differential equation:
Proof. Now by definition of ,Using (3.16) and from Lemma 3.8, Hence differentiating with respect to , using the fundamental theorem and (3.22), givesButHenceThereforeThis completes the proof.
4. Integral Operator Driven by a Diffusion Process
Now back to the integral operator defined in Section 1. Define a process ,where (see Notation 3.3 for the definition of the angle brackets). From the definition, it is clear that is adapted. In fact, is a symmetric matrix under the usual assumptions on and .
Proposition 4.1. If and are symmetric matrices, then is symmetric as a matrix.
Proof. Since is fixed, we will drop the subscript . Also fix an , so we will also drop the subscript . Let be the adjoint of with kernel . By assumption of symmetry and by definition,
Theorem 4.2. Let be an -diffusion process satisfying (1.2) and and let be continuous. Further assume that for . Let be an integral operator defined by (1.5) and Let be the explosion time of . Then for , satisfies the following stochastic differential equation: where
Proof. In the ambient space , is a diffusion satisfying the stochastic differential equation of the formNow by Theorem 3.9, satisfies the following differential equation:where is defined by (4.5) and by Proposition 4.1, . Thus we can writewhich is (4.4). The existence of for small time is guaranteed by Lemma 2.4.
Lemma 4.3. If is locally Lipschitz, then is the unique solution (path-wise) to (4.4).
Proof. Now by the definition of , and are locally Lipschitz. However, since is locally Lipschitz on the manifold with bounded operator norm, it follows that is locally Lipschitz. Therefore (4.4) has a unique solution and is given by .
5. Long-Time Existence of
We had addressed the existence and uniqueness of the solution to (4.5), given by , with . We will now give sufficient conditions for .
Proposition 5.1. Suppose that the integral operator with kernel is a strictly positive operator and is a symmetric nonnegative matrix. Then for such that , exists for all .
Proof. When is trivial. So assume . Fix an and any . Since is a compact operator, it suffices to show that the kernel of is 0. Write and . Let such that and . Then is nonzero. Hence we can assume that . Note that and is a symmetric operator (see Section 1 for definitions of and .) Therefore,Since is a nonnegative matrix, , and because is a strictly positive operator, . If ,Otherwise, and hence we haveEither way, if ,is nonzero, and therefore is nonzero. Thus for any nonzero function , is never zero and since is arbitrary, hence is invertible for any .
Proposition 5.2. Suppose that the usual assumptions on and hold. Then exists for all . Furthermore, (3.10) holds for all .
Proof. Under the assumptions on and , Proposition 3.1 and Remark 3.4 will imply Proposition 5.1. Fix an , a and let be as defined in Lemma 2.4. Then on , is invertible for all . Hence is an open-connected set containing 0, and exists for all . In particular, at . Then the assumptions in Corollary 2.9 are met, and hence (3.10) holds.
6. Proof of Main Result
The proof of Theorem 1.2 now follows from Theorem 4.2 and Proposition 5.2. Integrating (3.10), we have for ,For , defineand observe thatBy the Feynman-Kac formula, satisfies the following partial differential equation:Thusand therefore from (6.3),