Abstract
In this article, we consider the problem of obtaining general Hermite type approximations for densities of uniformly elliptic multi-dimensional diffusions with smooth coefficients in small time. These approximations resemble the Itô chaos expansions for square integrable random variables on Wiener space. The proofs are constructive and the approximations are explicit enough so that they can be used for constructing algorithms of approximation in the total variation sense. As examples, we provide some density expansions on some multi-dimensional examples.
Similar content being viewed by others
Notes
It is well known that all possible changes of measure in this setting are given by Girsanov’s theorem which require the diffusion term not to be approximated.
We call it conditional Gaussian as when it is used in the Euler-Maruyama scheme is used conditional to its initial point.
Constants will depend on the fixed value \( t_0\) which we assume without loss of generality to satisfy that \(t_0\in (0,1) \).
We will also still use the terminology “monomial” for a polynomial with only one term.
For an exception using tensorial notation, see Sect. 5.7 in [4].
References
Ait-Sahahia, Y.: Closed-form likelihood expansions for multivariate diffusions. Ann. Stat. 36, 906–937 (2008)
Azencott, R.: Densité des diffusions en temps petit: développements asymptotiques. I, Seminar on probability, XVIII, Lecture Notes in Math.1059, Springer, Berlin, 402–498, (1984)
Bally, V., Rey, C.: Approximation of Markov semigroups in total variation distance. Electron. J. Probab. 21(12), 1–44 (2016)
Barndorff-Nielsen, O.E., Cox, D.R.: Asymptotic Techniques for Use in Statistics, Chapman and Hall, (1989)
Bodnarchuk, S., Ivanenko, D., Kohatsu-Higa, A., Kulik, A.: Improved local approximation for multidimensional diffusions: the G-rates. Teoriya Ĭmovirnosteĭ ta Matematychna Statystyka 101(2), 15–39 (2019)
Bras, P., Pagés, G., Panloup, F.: Total variation convergence of the Euler-Maruyama scheme in small time with unbounded drift. Preprint, Arxiv (2022)
Corielli, F., Foschi, P., Pascucci, A.: Parametrix approximation of diffusion transition densities. SIAM J. Financial Math. 1, 833–867 (2010)
Friedman, A.: Partial differential equations of parabolic type. Dover Publications (1964)
Ivanenko, D.O., Kohatsu-Higa, A., Kulik, A.M.: LA(M)N property for diffusion models with low regularity coefficients, Manuscript
Konakov, V., Mammen, E.: Local limit theorems for transition densities of Markov chains converging to diffusions. Probab Theory Relat. Fields 117, 551–587 (2000)
Konakov, V., Mammen, E.: Small time Edgeworth-type expansions for weakly convergent nonhomogeneous Markov chains. Probab. Theory Relat. Fields 143, 137 (2009)
Lorig, M., Pagliarani, S., Pascucci, A.: Analytical expansions for parabolic equations SIAM. J. Appl. Math. 75(2), 468–491 (2015)
Pagés, G.: Numerical Probability: an introduction with applications to finance, p. 587. Springer, New york (2018)
Watanabe, S.: Short time asymptotic problems in Wiener functional integration theory. Applications to heat kernels and index theorems. Stochastic analysis and related topics, II (Silivri, 1988), Lecture Notes in Math., 1444, Springer, Berlin, 1–62, (1990)
Watanabe, S.: Analysis of Wiener functionals (Malliavin calculus) and its applications to heat kernels. Ann. Probab. 15, 1–39 (1987)
Yoshida, N.: Asymptotic expansions of maximum likelihood estimators for small diffusions via the theory of Malliavin-Watanabe. Probab. Theor. Related Fields. 92, 275–311 (1992)
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors of the present article certify that they have no affiliations with or involvement in any organization or entity with any financial interest or non-financial interest (except for the research support as declared in the first page of this article) in the subject matter or materials discussed in this manuscript.
Additional information
Communicated by Charles-Edouard Bréhier.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
A. Kohatsu-Higa has been supported by KAKENHI Grant 20K03666.
A. Kulik has been supported through the DFG-NCN Beethoven Classic 3 programme, contract no. 2018/31/G/ST1/02252 (National Science Center, Poland) and SCHI-419/11-1 (DFG, Germany).
Appendices
Algorithm for calculating the terms of the chaos decomposition
1.1 Description of the algorithm
The algorithm consists of finding an expansion of (4.1) of a certain time order. As our proof is constructive this expansion can be obtained from the proof. Suppose that we want an expansion with t-order N. By induction, we suppose that we have the chaos monomial expansion of t-order \( N-1 \), denoted by \( p^{(N-1)} \) in Proposition 4.2. The algorithm has two parts:
Part I: Chaos monomial decomposition of t-order N/2 for \(\varXi _t(x,y)\) (The first term in (4.1))
Step I.1: Collect all terms in Taylor’s formula for \(\varPhi _t(\xi ;x,y)\) w.r.t. \(\xi \) around x as in Corollary 4.12. That is, the terms where \( l+m/2\le N/2 \).
Step I.2: The Taylor formula for increments of a, b. Although this is mentioned in Corollary 4.2, the polynomials P and Q are not given explicitly. One has to recall that each polynomial of order k in the Taylor expansion is equivalent to a t-order of k/2 and therefore one needs to perform this expansion as long as the t-order of that term does not exceed N/2.
Step I.3: Transformation of the products into chaos monomials. This implies that each difference \( (y-x) \) has to be corrected using the mean \( \mp a(x)t \). Then use formula B.6 in order to convert the product of the polynomial and the derivative of \( \varPhi _t(x,y) \) into sums of chaos monomials. This procedure is also used repeatedly in Part II.
Part II: Chaos decomposition of \(p^{(N-1)}\circledast \Upsilon \) (the second term in (4.1)).
Step II.1: elimination of negligible terms: Note that the kernel \( \Upsilon _t(z,y)=\Upsilon ^{\textrm{drift}}_t(z,y)+\Upsilon ^{\textrm{diffusion}}_t(z,y) \) where each kernel has a different t-time order according to Lemma 4.1. Therefore the order of the required expansions will be different for each term. Still the procedure is similar as in Step I and one uses the Taylor formula for \(\varPhi _t^{(i)}(\xi ;z,y)\),\(\varPhi _t^{(i,j)}(\xi ;z,y)\) w.r.t. \(\xi \) around x.
Step II.2: Apply the Taylor formula for increments of a, b as in Step I.2 above.
Step II.3: Transformation of the products into 3-point monomials: Now we will have polynomials in the variables \( (z-x) \) and \( (z-y) \) due to the Step II.2. As in Step I.3, we add and subtract the mean term a(x)t and use B.6.
Step II.4: Integration of the 3-point monomials into chaos monomials using (B.10) and computing the integrals with respect to time in \(p^{(N-1)}\circledast \Upsilon \) .
Hermite functions and polynomials
In this section, we gather a series of identities and properties of general Hermite functions and polynomials as introduced in Section 2. These functions/polynomials do not appear often in the literature, which typically focuses on standardized Gaussian measures \(\mathscr {N}(0, \textrm{Id}_{{\mathbb {R}}^d})\) or \(\mathscr {N}(0, t\textrm{Id}_{{\mathbb {R}}^d})\); thus their properties can be hardly found explicitly writtenFootnote 5 in the present generality.
To provide a comprehensive discussion, in addition to the Hermite functions introduced in Sect. 2, we will also consider their standardized analogues
and corresponding Hermite polynomials defined by the equation
Note that the Hermite functions/polynomials introduced in Section 2 reduce to the above definition in the particular case that \( \textbf{a}=0 \) and \( \textbf{b}=I_{d}\). Furthermore, let \(\textbf{a}\in {\mathbb {R}}^d, \textbf{b}\in {\mathbb {R}}^{d\times d}_\textrm{sym}\) be given. As \( \textbf{b}\) is assumed to be a strictly positive definite matrix then there exists a \(d\times d\) matrix \(\sigma \) such that \(\textbf{b}=\sigma \sigma ^*\). From now on, \(\sigma \) is considered to be fixed and it is invertible since \(\textbf{b}\in {\mathbb {R}}^{d\times d}_\textrm{sym}\). It follows simply by the definition of \( \varPhi \) that if we let \( w=\sigma ^{-1}(y-x-t\textbf{a}) \) then
Thus using the chain rule for derivatives we have the following relation:
Furthermore, the above formulas show that any property on the standardized Hermite polynomials \( \mathscr {H}\) can be transferred to the general Hermite polynomials H.
For example,
the second identity can be either derived directly or obtained from the first one using (B.1) and the fact that \(\textbf{b}^{-1}= (\sigma ^*)^{-1}\sigma ^{-1}\). Similarly,
and
Next, we have the identities
which can be derived either directly or using (B.1) and the well known identities for canonical Hermite polynomials:
Combining (B.2) and (B.3) and re-arranging the notation for the multi-indices we get
Multiplying (B.4) by \(\textbf{b}_{j i}\) and taking the sum over i, we obtain for \( n\ge 0 \)
In calculations, it is better to simplify the notation as follows (here we use the summation over double indexes notation)
In the above equation, we have also used the following index deleting operation:
Extensions of this operation are defined by iteration. For example,
Also, we will need to use the following index addition operation
We remark that the above operations with indexes have to be carried out in an order from right to left.
The Hermite functions \(\varPhi _t^{(i_1, \dots , i_n)}(\textbf{a}, \textbf{b};x,y)\) also have nice properties as functions of the parameters \(\textbf{a}\in {\mathbb {R}}^d, \textbf{b}\in {\mathbb {R}}^{d\times d}_\textrm{sym}\):
These properties are difficult to prove using the canonical Hermite functions, thus we provide the direct proof based on the Fourier transform. Namely, let
then
and
Then, we conclude that
which proves (B.8). The proof of (B.7) is similar and omitted.
Finally, we mention the following convolution property of the Hermite functions: for any \(t,s>0\) and \(m,n\ge 0\), \(i_1, \dots , i_n, j_1, \dots , j_m\)
Identity (B.10) easily follows from (B.9) and the fact that the Fourier transform of a convolution is a product.
Carrying out the multi-dimensional algorithm
Here we will give the steps carried out in order to obtain Example 3.2. In fact, recall that the existence of the residue \( \varDelta _t^2(x,y) \) is of order \( 1+\rho /2 \) is already assured by Theorem 3.1. Therefore, we will only find expansions up to order 1 for all terms. For the one dimensional case, we refer the reader to [5].
In order to simplify the expressions, we use the summation over double indexes convention.
Step 1: Consider Proposition 4.2 with \( N=2 \). Therefore we now need to find \( \varGamma ^n_t(x,y) \) for \( n=0,1,2 \). Instead of this approach, one uses an iterative approach where one computes \( p^{(j)}_t(x,y) \) for \( j=0,1 \).
Step 1.0: From the statement of Theorem 3.1 in the case \( n=0 \), we have \(\varGamma ^0_t(x,y)=\varPhi _t(x,y) \) and \( p^{(0)}_t(x,y)=\varPhi _t(x,y) \). Therefore we need to find an expansion for \( \varXi _t(x,y)= \varPhi _t(y;x,y)\) around \( \varPhi _t(x;x,y)\equiv \varPhi _t(x,y) \).
In order to understand how this step should be carried out it is important to note the following two rules:
-
Each derivative of \( \varPhi _t(\textbf{a}, \textbf{b}; x,y) \) with respect to an element of Åleads to a term which increases its t-order in 1/2 while a derivative with respect to an element of \( \textbf{b}\) increases the t-order in 1.
-
Then the expansion is obtained from the equality \(\varPhi _t(\xi ;x,y)\!=\!\varPhi _t(a(\xi ),b(\xi ) ; x,y)\).
Carrying out these steps one obtains
This finishes Step I.1. Now we proceed with Step I.2, which is the Taylor expansions of the coefficients. That is,
From now on, in order to shorten expressions we will use the usual convention for products of functions: \( (fg)(x)\equiv f(x)g(x) \)
For Step I.3 we now add and subtract \( -ta(x) \). Putting these results together and using (B.6), we obtain:
Deleting the terms of high time order, we obtain
Further algebraic simplifications and arranging the terms give:
Now, we need to find the expansion of t-order 1/2 for \( \left( \varXi \circledast \Upsilon \right) _t(x,y)\). For this, given Lemma 4.1 and Proposition 4.1, we just need to compute the terms of this order in
We start with the expansion of \( \Upsilon ^{\textrm{diffusion}} \). Applying Taylor’s expansion of b at y, B.6 and (C.2), we have
After integration in \( \circledast \) and we obtain that
Putting this together with the result of Step I.3, we have
Step 1.1: We need to find the expansions described in Part II. We start with Step II.1:
Repeating similar steps as in the one dimensional case, one obtains at the end of Step II.2:
Step II.3. gives
Now we integrate in the time and space variables (z) to obtain that
Further algebraic simplifications give:
As in the one dimensional case, adding the above result to the expansion of \( \varXi _t(x,y)= \varPhi _t(y;x,y)\) around \( \varPhi _t(x;x,y)\equiv \varPhi _t(x,y) \) obtained in \( \mathbf{Step\ I.3} \), we have
Remark C.1
Note that a natural approach with lower orders of expansion may be obtained as follows. In order to write an expansion of order \(1+\rho /2 \), we need to use Lemma 4.1 and Proposition 4.1 to see that
-
1.
The t-order of \(\Upsilon \) is \( -1/2 \) and \( \Upsilon ^{\circledast {2}}\) is 0.
-
2.
Therefore in order to obtain \( \left( \varXi \circledast \Upsilon ^{\circledast {k}}\right) _t(x,y)\) of t-order \( 1+\rho /2 \), we need to expand \( \varXi \) up to t-order 1/2 in the case \( k=1 \) and t-order 0 in the case \( k=2 \).
This proposal can be cumbersome when general orders are required. For this reason, we prefer the iterative scheme.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Ivanenko, D., Kohatsu-Higa, A. & Kulik, A. Small time chaos approximations for heat kernels of multidimensional diffusions. Bit Numer Math 63, 16 (2023). https://doi.org/10.1007/s10543-023-00949-z
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s10543-023-00949-z