MSC (2020) Classification: 60L10, 60J76, 68T05
Introduction
Score-based generative models have become a leading paradigm for sampling from complex distributions Sohl-Dickstein et al. [65 ] ; Song and Ermon [66 ] ; Ho et al. (2020); Song et al. [67 ] ; Tang and Zhao [69 ] . Extending them to time series data is challenging, as sequential data are path-valued. Indeed, applying standard score-based generative modeling via stochastic differential equations to time series generation requires considering time-reversed (jump) diffusions on path space, which are inherently infinite-dimensional and potentially intractable. A natural way to overcome this complexity is to handle paths through their signature , a universal feature map given by iterated integrals Lyons [54 ] ; Chevyrev and Kormilitzin [16 ] ; Cass and Salvi [14 ] , which – in its logarithmic form – has recently been applied by Barancikova et al. [5 ] in the context of generative time series modeling. In this paper, we construct score-based generative models for time series as well as finite-dimensional data, building on an explicit representation of the expected signature of a broad class of jump-diffusions. Our main contributions are the following.
Entire sig-jump-diffusions. We introduce the class of entire sig-jump-diffusions , whose expected signature admits an explicit representation via linear ordinary differential equations (Definition 3.3 , Theorem 3.13 ).
Score-based generative modeling with (lifted) entire sig-jump-diffusions for time series. We exploit this representation to construct generative models for time series, in which a generalized score function is learned through a deterministic optimization problem (Sections 4.1–4.4, Theorem 4.9 ).
Finite-dimensional data generation and inverse signature matching. For data in R d \mathbb{R}^{d} R d , we adapt this framework to the subclass of polynomial processes introduced in Cuchiero et al. [22 ] ; Filipović and Larsson [33 ] (Remark 4.10 ). Additionally we propose an alternative to score matching, called inverse signature matching (§ 4.5 ).
Our starting point is the expected signature, the path-space analogue of the sequence of moments of a random variable, which under suitable conditions characterizes the law of a stochastic process, Chevyrev and Lyons [17 ] ; Chevyrev and Oberhauser [18 ] . It is therefore a natural tool whenever laws on path space must be estimated or compared, as in parameter estimation Papavasiliou and Ladroue [61 ] ; Semnani et al. [64 ] , hypothesis testing Gasteratos et al. [36 ] and generative modeling of time series Liao et al. [52 ] ; Lou et al. [53 ] . It often also reduces stochastic control problems Kalsi et al. [48 ] ; Bayer et al. [9 ] ; Cuchiero and Möller [23 ] ; Gennaro et al. [38 ] ; Abi Jaber et al. [1 ] and financial derivative pricing Lyons et al. [55 ] ; Cuchiero et al. [21 ] ; Abi Jaber et al. [2 ] to simple deterministic solutions, since expected costs and prices become linear functionals of the expected signature.
For these reasons, the computation of expected signatures has received considerable attention. Explicit or recursive representations are available for Brownian motion Fawcett [31 ] ; Lyons and Victoir [57 ] ; Lyons and Ni [56 ] ; Boedihardjo et al. [10 ] , Gaussian processes Boedihardjo et al. [11 ] ; Cass and Ferrucci [13 ] , Lévy processes Friz and Shekhar [35 ] ; Cuchiero et al. [25 ] , and more general diffusion and semimartingale models Ni [60 ] ; Friz et al. [34 ] ; Cuchiero et al. [27 ] .
Entire sig-jump-diffusions. We contribute to this line of work by introducing a general class of jump-diffusions whose expected signature is obtained by solving a linear ODE in the extended tensor algebra. To this end, we define entire sig-jump-diffusions as semimartingales whose characteristics can be written as a linear combination of potentially infinitely many components of their own signature (Definition 3.3 ). This framework is very general: by the universal approximation properties of the signature, continuous path functionals can be approximated arbitrarily well, on suitable compact sets and with respect to suitable variation metrics, by linear functionals of the signature [48 , 9 , 25 ] , therefore approximating (in suitable metrics) any semimartingale with sufficiently regular characteristics. The class extends signature SDEs introduced in Cuchiero et al. [27 ] , where only semimartingales with continuous trajectories are considered. In particular, it includes the signature models for asset pricing of [62 , 21 , 20 b] .
Based on duality arguments for stochastic processes [47 , 30 ] , as in [27 a] , we show that expectations of potentially infinitely many linear combinations of the signature components of such processes can be computed by solving linear ODEs in the extended tensor algebra (Theorem 3.13 ), extending to the non-continuous setting the results of Cuchiero et al. [27 ] . The main technical challenge is the convergence of the resulting infinite series, which is considerably harder in the presence of jumps, as it requires control of the series and all its derivatives (equation (5) ).
For certain important cases, we can prove that these convergence conditions are satisfied. This applies to holomorphic jump-diffusions as of Cuchiero et al. [26 ] (Subsection 3.3.1), to polynomial processes introduced in [22 , 33 ] (Subsection 3.3.3), and to some jump-diffusions with possibly path-dependent characteristics (Remark 3.27), all of which embed in the entire sig-jump-diffusion framework. In particular, the truncated signature of a polynomial jump-diffusion is again a finite-dimensional polynomial jump-diffusion (Proposition 3.24 ), so that its expected signature reduces to an easily computable matrix exponential (Example 3.28 ). Another tractable subclass, whose truncated signature lift is a finite-dimensional Markov process, is that of entire sig-jump-diffusions with truncated signature-dependent coefficients (Subsection 3.3.2).
Score-based generative modeling with (lifted) entire sig-jump-diffusions for time series. We apply the framework of entire sig-jump-diffusions, specifically the subclass of Subsection 3.3.2, to score-based generative modeling of time series. We encode the distribution of the observed time series through the distribution of their truncated signatures and perturb it with a lifted entire sig-jump-diffusion converging to an invariant measure independent of the initial data distribution. Starting from this invariant distribution, a suitable reverse-time process generates new samples approximating the original signature distribution.
The key ingredient in this construction is a generalized score function, which extends the classical score function to jump-diffusions. In the Lévy setting it corresponds to the score considered in Baule [8 ] for finite jump activity and in Yoon et al. [71 ] for isotropic α \alpha α -stable noise. The latter has been taken up by Popov et al. [63 ] , who use a parametrized family of reverse-time processes sharing the prescribed marginal distributions (and not the whole law on path space), a principle that we extend beyond the α \alpha α -stable case to general jump-diffusions. The crucial difference to these previous works is that our definition of the generalized score function is formulated directly in terms of the generator and the marginal probability measures, without requiring a particular structure of the jump mechanism. As shown in Sections 4.1 and § 4.2 , this makes it applicable to state-dependent jump-diffusions and, in particular, to the lifted truncated signature dynamics of Subsection 3.3.2.
Our main contribution is to exploit the analytic structure of entire sig-jump-diffusions and their explicit expected signature representation to reduce the learning of this score to linear ODEs and a deterministic optimization problem (Section 4.3, see in particular Theorem 4.9 ). As already mentioned, our approach is related to the recent work of Barancikova et al. [5 ] , who introduced score-based generative models operating on log-signature representations of time series. While their approach relies on diffusion processes in the vector space of log-signatures, we consider lifted entire sig-jump-diffusions evolving directly on the space of signatures, which provides a natural connection to the stochastic dynamics of the underlying time series and allows the coefficients of the forward process to be calibrated to the observed dynamics. We refer to § 4 for a more detailed explanation. Beyond this, our approach differs from Barancikova et al. [5 ] in two further respects. First, we incorporate jumps into the generative dynamics. Second, thanks to the analytic tractability of entire sig-jump-diffusions, the score is learned through the explicit expected signature instead of neural network training.
To show that suitable forward processes exist, in Section 4.4 we construct an entire sig-diffusion whose truncated signature lift converges to a unique invariant measure, independently of the initial data distribution, and thus provides a natural forward perturbation process.
Since the framework operates on signatures rather than on time series, recovering a time series from a generated signature requires a separate reconstruction step, for which several inversion procedures exist Lyons and Xu [58 , 59 ] ; Fermanian et al. [32 ] ; Barancikova et al. [5 ] . As truncated signatures do not determine paths uniquely, this step is subject to approximation issues, which we do not address here.
Finite-dimensional data generation and inverse signature matching. The entire sig-jump-diffusion framework can also be used to generate finite-dimensional data, i.e., new samples of random variables with values in R d \mathbb{R}^{d} R d . To this end, we use polynomial processes Cuchiero et al. [22 ] ; Filipović and Larsson [33 ] , analyzed in Subsection 3.3.3, for the forward dynamics. In this setting, learning the generalized score function reduces to computing moments and solving, again, a deterministic optimization problem (Remark 4.10 ).
We also propose an alternative to standard score matching, which we call inverse signature matching , tailored to polynomial diffusions. Since, under suitable conditions, the expected signature characterizes the law of a stochastic process Chevyrev and Oberhauser [18 ] , the key idea is to learn the generalized score by matching the expected signature of the reverse process to that of the inverse forward signature. Under a suitable parametrization of the score map, the polynomial structure ensures that both expectations can be computed analytically, which reduces the problem again to a deterministic optimization problem (§ 4.5 ).
Finally, we stress that the goal of the paper is to establish the framework and the analytic learning procedure, while a numerical investigation of the generative models, on time series as well as on finite-dimensional data, is left to future work.
Related work on generative modeling. As the literature on expected signatures was already discussed above, we here only focus on generative modeling. In score-based generative models, the reverse-time drift involves the score of the marginal densities Haussmann and Pardoux [43 ] , which is usually learned by (denoising) score matching Hyvärinen [45 ] ; Vincent [70 ] . Besides the reverse SDE, the probability-flow ODE Song et al. [67 ] gives a deterministic sampler with the same marginal laws. Jump noise has been considered in Yoon et al. [71 ] ; Baule [8 ] , with a family of marginal-preserving reverse dynamics in Popov et al. [63 ] ; we compare with these works in the description of our contributions above. Other works build on pretrained diffusion models, for instance for conditional generation and fine-tuning Han et al. [42 ] ; Guo et al. [40 ] or distillation Ge et al. (2026); these are complementary to the construction of the forward and reverse dynamics studied here.
For time series, several generative models are based on signatures Liao et al. [52 ] ; Lou et al. [53 ] , the closest to our work being Barancikova et al. [5 ] , which we already discussed in detail above. Other approaches that do not rely on signatures include the Schrödinger bridge model of Hamdouche et al. [41 ] , whose path-dependent drift is estimated nonparametrically, its extension to jump-diffusions De Marco et al. [28 ] , and the sequential forward-backward diffusion of Cao et al. [12 ] , which generates adapted series by conditioning on the generated history and comes with statistical guarantees.
Organization of the paper. In Section 2, we introduce the algebraic setting of signature theory and the notion of signature for càdlàg semimartingales. In Section 3, we introduce the class of entire sig-jump-diffusions, present the main representation results (Theorem 3.13 ), and show that holomorphic jump-diffusions, entire sig-jump-diffusions with truncated signature-dependent coefficients, and polynomial processes can be embedded within this framework. § 4 is dedicated to the application to score-based generative modeling.
Preliminaries
Notation
Fix d ∈ N d \in\mathbb{N} d ∈ N and let R d \mathbb{R}^{d} R d be the Euclidean space. The extended tensor algebra over R d \mathbb{R}^{d} R d is defined by
T ( R d ) : = ∏ n = 0 ∞ ( R d ) ⊗ n , T\left(\mathbb{R}^{d}\right) := \prod_{n=0}^{\infty}\left(\mathbb{R}^{d}\right)^{\otimes n}, T ( R d ) := n = 0 ∏ ∞ ( R d ) ⊗ n ,
where ( R d ) ⊗ n \left(\mathbb{R}^{d}\right)^{\otimes n} ( R d ) ⊗ n denotes the n n n -fold tensor product of R d \mathbb{R}^{d} R d with the convention ( R d ) ⊗ 0 : = R \left(\mathbb{R}^{d}\right)^{\otimes0} := \mathbb{R} ( R d ) ⊗ 0 := R . We equip T ( ( R d ) ) T\left(\left(\mathbb{R}^{d}\right)\right) T ( ( R d ) ) with the standard addition + + + , tensor multiplication ⊗ \otimes ⊗ and scalar multiplication. For N ∈ N N \in\mathbb{N} N ∈ N , the truncated tensor algebra is defined by
T N ( R d ) : = ⨁ n = 0 N ( R d ) ⊗ n . T^{N}\left(\mathbb{R}^{d}\right) := \bigoplus_{n=0}^{N}\left(\mathbb{R}^{d}\right)^{\otimes n}. T N ( R d ) := n = 0 ⨁ N ( R d ) ⊗ n .
Let I = ( i 1 , … , i n ) I=(i_{1},\ldots,i_{n}) I = ( i 1 , … , i n ) be a multi-index with entries in { 1 , … , d } \{1,\ldots,d\} { 1 , … , d } . Denoting by ϵ 1 , … , ϵ d \epsilon_{1},\ldots,\epsilon_{d} ϵ 1 , … , ϵ d the canonical basis of R d \mathbb{R}^{d} R d , we use the notation ∣ I ∣ : = n |I|:=n ∣ I ∣ := n and ϵ I : = ϵ i 1 ⊗ ϵ i 2 ⊗ ⋯ ⊗ ϵ i n \epsilon_{I}:=\epsilon_{i_{1}}\otimes\epsilon_{i_{2}}\otimes\cdots\otimes\epsilon_{i_{n}} ϵ I := ϵ i 1 ⊗ ϵ i 2 ⊗ ⋯ ⊗ ϵ i n . Observe that ( ϵ I ) I (\epsilon_{I})_{I} ( ϵ I ) I is the canonical orthonormal basis of ( R d ) ⊗ n (\mathbb{R}^{d})^{\otimes n} ( R d ) ⊗ n . Furthermore, we denote by ϵ ∅ \epsilon_{\varnothing} ϵ ∅ the basis element of ( R d ) ⊗ 0 (\mathbb{R}^{d})^{\otimes0} ( R d ) ⊗ 0 and set ∣ ∅ ∣ : = 0 |\varnothing|:=0 ∣ ∅ ∣ := 0 . We also set I ′ : = ( i 1 , … , i n − 1 ) I^{\prime}:=(i_{1},\ldots,i_{n-1}) I ′ := ( i 1 , … , i n − 1 ) for n > 1 n>1 n > 1 , I ′ = ∅ I^{\prime}=\varnothing I ′ = ∅ for n = 1 n=1 n = 1 , I ′ ′ : = ( I ′ ) ′ I^{\prime\prime}:=(I^{\prime})^{\prime} I ′′ := ( I ′ ) ′ for n > 1 n>1 n > 1 , and use the convention ϵ I ′ = 0 \epsilon_{I^{\prime}}=0 ϵ I ′ = 0 for I = ∅ I=\varnothing I = ∅ . For u ∈ T ( ( R d ) ) \mathbf{u}\in T((\mathbb{R}^{d})) u ∈ T (( R d )) , we denote its projection on T N ( R d ) T^{N}(\mathbb{R}^{d}) T N ( R d ) by u ≤ N \mathbf{u}^{\leq N} u ≤ N and let u I u_{I} u I denote its component corresponding to ϵ I \epsilon_{I} ϵ I .
For two multi-indices I ∈ { 1 , … , d } ∣ I ∣ I\in\{1,\ldots,d\}^{|I|} I ∈ { 1 , … , d } ∣ I ∣ , J ∈ { 1 , … , d } ∣ J ∣ J\in\{1,\ldots,d\}^{|J|} J ∈ { 1 , … , d } ∣ J ∣ , and two indices of length 1, a , b ∈ { 1 , … , d } a,b\in\{1,\ldots,d\} a , b ∈ { 1 , … , d } , the shuffle product ⊔ \sqcup ⊔ is defined recursively by
ϵ I ⊔ ϵ ∅ = ϵ ∅ ⊔ ϵ I = ϵ I , \epsilon_{I}\sqcup\epsilon_{\varnothing}=\epsilon_{\varnothing}\sqcup\epsilon_{I}=\epsilon_{I}, ϵ I ⊔ ϵ ∅ = ϵ ∅ ⊔ ϵ I = ϵ I ,
ϵ I a ⊔ ϵ J b = ( ϵ I ⊔ ϵ J b ) ⊗ ϵ a + ( ϵ I a ⊔ ϵ J ) ⊗ ϵ b , \epsilon_{Ia}\sqcup\epsilon_{Jb}=(\epsilon_{I}\sqcup\epsilon_{Jb})\otimes\epsilon_{a}+(\epsilon_{Ia}\sqcup\epsilon_{J})\otimes\epsilon_{b}, ϵ I a ⊔ ϵ J b = ( ϵ I ⊔ ϵ J b ) ⊗ ϵ a + ( ϵ I a ⊔ ϵ J ) ⊗ ϵ b ,
where I a Ia I a denotes the concatenation of multi-indices.
For u ∈ T ( R d ) \mathbf{u}\in T(\mathbb{R}^{d}) u ∈ T ( R d ) and x ∈ T ( ( R d ) ) \mathbf{x}\in T((\mathbb{R}^{d})) x ∈ T (( R d )) , we write
⟨ u , x ⟩ : = ∑ n = 0 ∞ ∑ ∣ I ∣ = n u I x I . \langle\mathbf{u},\mathbf{x}\rangle:=\sum_{n=0}^{\infty}\sum_{|I|=n}u_{I}x_{I}. ⟨ u , x ⟩ := n = 0 ∑ ∞ ∣ I ∣ = n ∑ u I x I .
Define G N ( R d ) : = { u ≤ N ∈ T N ( R d ) : u ∈ G ( ( R d ) ) } G^{N}(\mathbb{R}^{d}):=\{\mathbf{u}^{\leq N}\in T^{N}(\mathbb{R}^{d}):\mathbf{u}\in G((\mathbb{R}^{d}))\} G N ( R d ) := { u ≤ N ∈ T N ( R d ) : u ∈ G (( R d ))} for
G ( ( R d ) ) : = { x ∈ T ( ( R d ) ) : ⟨ ϵ ∅ , x ⟩ = 1 , ⟨ ϵ I ⊔ ϵ J , x ⟩ = ⟨ ϵ I , x ⟩ ⟨ ϵ J , x ⟩ , for all I , J } , G((\mathbb{R}^{d})):=\{\mathbf{x}\in T((\mathbb{R}^{d})):\langle\epsilon_{\varnothing},\mathbf{x}\rangle=1,\ \langle\epsilon_{I}\sqcup\epsilon_{J},\mathbf{x}\rangle=\langle\epsilon_{I},\mathbf{x}\rangle\langle\epsilon_{J},\mathbf{x}\rangle,\ \text{for all }I,J\}, G (( R d )) := { x ∈ T (( R d )) : ⟨ ϵ ∅ , x ⟩ = 1 , ⟨ ϵ I ⊔ ϵ J , x ⟩ = ⟨ ϵ I , x ⟩ ⟨ ϵ J , x ⟩ , for all I , J } ,
and for S ⊆ R d S\subseteq\mathbb{R}^{d} S ⊆ R d , set
S : = { x ∈ G ( ( R d ) ) : ∑ ∣ I ∣ = 1 x I ϵ I ∈ S } . (1) S:=\left\{\mathbf{x}\in G((\mathbb{R}^{d})):\sum_{|I|=1}x_{I}\epsilon_{I}\in S\right\}. \tag*{(1)} S := ⎩ ⎨ ⎧ x ∈ G (( R d )) : ∣ I ∣ = 1 ∑ x I ϵ I ∈ S ⎭ ⎬ ⎫ . (1)
The set S S S is endowed with the sigma-algebra induced by the product sigma-algebra on T ( ( R d ) ) T((\mathbb{R}^{d})) T (( R d )) .
Define
S ∗ : = { u ∈ T ( ( R d ) ) + i T ( ( R d ) ) : ∣ u ∣ x < ∞ for all x ∈ S } , S^{*}:=\{\mathbf{u}\in T((\mathbb{R}^{d}))+iT((\mathbb{R}^{d})):\lvert\mathbf{u}\rvert_{\mathbf{x}}<\infty\text{ for all }\mathbf{x}\in S\}, S ∗ := { u ∈ T (( R d )) + i T (( R d )) : ∣ u ∣ x < ∞ for all x ∈ S } ,
where ∣ u ∣ x : = ∑ n = 0 ∞ ∑ ∣ I ∣ = n ∣ u I x I ∣ \lvert\mathbf{u}\rvert_{\mathbf{x}}:=\sum_{n=0}^{\infty}\sum_{\lvert I\rvert=n}\lvert u_{I}x_{I}\rvert ∣ u ∣ x := ∑ n = 0 ∞ ∑ ∣ I ∣ = n ∣ u I x I ∣ . For u ∈ S ∗ \mathbf{u}\in S^{*} u ∈ S ∗ , we again write
⟨ u , x ⟩ : = ∑ n = 0 ∞ ∑ ∣ I ∣ = n u I x I . \langle\mathbf{u},\mathbf{x}\rangle:=\sum_{n=0}^{\infty}\sum_{\lvert I\rvert=n}u_{I}x_{I}. ⟨ u , x ⟩ := n = 0 ∑ ∞ ∣ I ∣ = n ∑ u I x I .
Moreover, we set u ( 0 ) : = u \mathbf{u}^{(0)}:=\mathbf{u} u ( 0 ) := u and for k ∈ N k\in\mathbb{N} k ∈ N , ∣ I ∣ = k \lvert I\rvert=k ∣ I ∣ = k , we define u I ( k ) ∈ T ( ( R d ) ) + i T ( ( R d ) ) \mathbf{u}_{I}^{(k)}\in T((\mathbb{R}^{d}))+iT((\mathbb{R}^{d})) u I ( k ) ∈ T (( R d )) + i T (( R d )) via
u I ( k ) : = ∑ ∣ J ∣ ≥ 0 u J I . \mathbf{u}_{I}^{(k)}:=\sum_{\lvert J\rvert\geq0}u_{JI}. u I ( k ) := ∣ J ∣ ≥ 0 ∑ u J I .
where J I JI J I denotes the concatenation of the multi-indices J J J and I I I . Define V : = T ( ( R d ) ) + i T ( ( R d ) ) \mathcal{V}:=T((\mathbb{R}^{d}))+iT((\mathbb{R}^{d})) V := T (( R d )) + i T (( R d )) and note that u ( k ) ∈ ( V d ) ⊗ k \mathbf{u}^{(k)}\in(\mathcal{V}^{d})^{\otimes k} u ( k ) ∈ ( V d ) ⊗ k . Hence, setting ( V d ) ⊗ 0 : = V (\mathcal{V}^{d})^{\otimes0}:=\mathcal{V} ( V d ) ⊗ 0 := V ,
( u ( 0 ) , u ( 1 ) , u ( 2 ) , … ) ∈ T ( ( V d ) ) . (\mathbf{u}^{(0)},\mathbf{u}^{(1)},\mathbf{u}^{(2)},\ldots)\in T((\mathcal{V}^{d})). ( u ( 0 ) , u ( 1 ) , u ( 2 ) , … ) ∈ T (( V d )) .
Signature of càdlàg semimartingales
Definition 2.1. For a semimartingale X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] , we call the G ( ( R d ) ) G((\mathbb{R}^{d})) G (( R d )) -valued process X = ( X t ) t ∈ [ 0 , T ] \mathbf{X}=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] the signature of X X X if ⟨ ϵ ∅ , X t ⟩ = 1 \langle\epsilon_{\varnothing},\mathbf{X}_t\rangle=1 ⟨ ϵ ∅ , X t ⟩ = 1 and
⟨ ϵ I , X t ⟩ = ∫ 0 t ⟨ ϵ I ′ , X s − ⟩ d X s i ∣ I ∣ 1 { ∣ I ∣ > 0 } + 1 2 ∫ 0 t ⟨ ϵ I ′ ′ , X s − ⟩ d [ X i ∣ I ∣ − 1 , X i ∣ I ∣ ] s 1 { ∣ I ∣ > 1 } + ∑ 0 ≤ s ≤ t ∑ I 1 ⊛ I 2 = I ϵ I 1 ≠ ϵ 1 ∣ I 2 ∣ ! 1 { ∣ I 2 ∣ > 1 } ⟨ ϵ I 1 , X s − ⟩ ⟨ ϵ I 2 , ( Δ X s ) ⊗ ∣ I 2 ∣ ⟩ , (2) \begin{aligned}
\langle\epsilon_I,\mathbf{X}_t\rangle=\int_0^t\langle\epsilon_{I'},\mathbf{X}_{s-}\rangle dX_s^{i_{\lvert I\rvert}}1_{\{\lvert I\rvert>0\}}+\frac{1}{2}\int_0^t\langle\epsilon_{I''},\mathbf{X}_{s-}\rangle d[X^{i_{\lvert I\rvert-1}},X^{i_{\lvert I\rvert}}]_s1_{\{\lvert I\rvert>1\}} \\
&+\sum_{0\leq s\leq t}\sum_{\substack{I_1\mathbin{\circledast}I_2=I\\ \epsilon_{I_1}\ne\epsilon}} \frac{1}{\lvert I_2\rvert!}1_{\{\lvert I_2\rvert>1\}}\langle\epsilon_{I_1},\mathbf{X}_{s-}\rangle\langle\epsilon_{I_2},(\Delta X_s)^{\otimes\lvert I_2\rvert}\rangle, \tag*{(2)}
\end{aligned} ⟨ ϵ I , X t ⟩ = ∫ 0 t ⟨ ϵ I ′ , X s − ⟩ d X s i ∣ I ∣ 1 {∣ I ∣ > 0 } + 2 1 ∫ 0 t ⟨ ϵ I ′′ , X s − ⟩ d [ X i ∣ I ∣ − 1 , X i ∣ I ∣ ] s 1 {∣ I ∣ > 1 } + 0 ≤ s ≤ t ∑ I 1 ⊛ I 2 = I ϵ I 1 = ϵ ∑ ∣ I 2 ∣! 1 1 {∣ I 2 ∣ > 1 } ⟨ ϵ I 1 , X s − ⟩ ⟨ ϵ I 2 , ( Δ X s ) ⊗ ∣ I 2 ∣ ⟩ , (2)
for each ∣ I ∣ > 0 \lvert I\rvert>0 ∣ I ∣ > 0 .
Since X \mathbf{X} X takes values in the infinite-dimensional space G ( ( R d ) ) G((\mathbb{R}^{d})) G (( R d )) , it is often convenient - both for practical and technical purposes - to consider a finite-dimensional truncation thereof. Recall that for u ∈ T ( ( R d ) ) \mathbf{u}\in T((\mathbb{R}^{d})) u ∈ T (( R d )) , we denote its projection to T N ( R d ) T^{N}(\mathbb{R}^{d}) T N ( R d ) , N ∈ N N\in\mathbb{N} N ∈ N , by u ≤ N \mathbf{u}^{\leq N} u ≤ N .
Definition 2.2. The truncated signature of order N N N of a semimartingale X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] is the T N ( R d ) T^{N}(\mathbb{R}^{d}) T N ( R d ) -valued process X ≤ N = ( X t ≤ N ) t ∈ [ 0 , T ] \mathbf{X}^{\leq N}=(\mathbf{X}^{\leq N}_t)_{t\in[0,T]} X ≤ N = ( X t ≤ N ) t ∈ [ 0 , T ] defined by ⟨ ϵ I , X t ≤ N ⟩ = ⟨ ϵ I , X t ⟩ \langle\epsilon_I,\mathbf{X}^{\leq N}_t\rangle=\langle\epsilon_I,\mathbf{X}_t\rangle ⟨ ϵ I , X t ≤ N ⟩ = ⟨ ϵ I , X t ⟩ for every multi-index I I I with ∣ I ∣ ≤ N \lvert I\rvert\leq N ∣ I ∣ ≤ N , and ⟨ ϵ I , X t ≤ N ⟩ : = 0 \langle\epsilon_I,\mathbf{X}^{\leq N}_t\rangle:=0 ⟨ ϵ I , X t ≤ N ⟩ := 0 otherwise.
Remark 2.3. Since the right-hand side of (2) for a given multi-index I I I only involves coefficients ⟨ ϵ I ′ , X s − ⟩ \langle\epsilon_{I'},\mathbf{X}_{s-}\rangle ⟨ ϵ I ′ , X s − ⟩ indexed by I ′ I' I ′ with ∣ I ′ ∣ < ∣ I ∣ \lvert I'\rvert<\lvert I\rvert ∣ I ′ ∣ < ∣ I ∣ , the process X ≤ N \mathbf{X}^{\leq N} X ≤ N is itself characterized by the recursion (2.2) restricted to indices ∣ I ∣ ≤ N \lvert I\rvert\leq N ∣ I ∣ ≤ N , and in particular does not depend on the tensor levels of order greater than N N N .
Remark 2.4. This concept of signature corresponds to the solution of the Marcus SDE X t = X t − ⋄ d X t \mathbf{X}_t=\mathbf{X}_{t-}\diamond dX_t X t = X t − ⋄ d X t , Kurtz et al. [49 ] ; Cuchiero et al. [25 ] . This in particular implies that if X − X 0 X-X_0 X − X 0 is an S S S -valued semimartingale, for some S ⊆ R d S\subseteq\mathbb{R}^{d} S ⊆ R d , then X \mathbf{X} X takes values in S S S as in (1) , and thus satisfies the shuffle property. Moreover, observe that Δ X t = δ ( X t − , Δ X t ) \Delta\mathbf{X}_t=\delta(\mathbf{X}_{t-},\Delta X_t) Δ X t = δ ( X t − , Δ X t ) for
δ ( x , ξ ) I : = ∑ ϵ I 1 ⊛ ϵ I 2 = ϵ I 1 ∣ I 2 ∣ ! ⟨ ϵ I 1 , x ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ 1 { ∣ I 2 ∣ > 0 } . \delta(\mathbf{x},\xi)_{I}:=\sum_{\epsilon_{I_1}\mathbin{\circledast}\epsilon_{I_2}=\epsilon_{I}}\frac{1}{\lvert I_2\rvert!}\langle\epsilon_{I_1},\mathbf{x}\rangle\langle\epsilon_{I_2},\xi^{\otimes\lvert I_2\rvert}\rangle1_{\{\lvert I_2\rvert>0\}}. δ ( x , ξ ) I := ϵ I 1 ⊛ ϵ I 2 = ϵ I ∑ ∣ I 2 ∣! 1 ⟨ ϵ I 1 , x ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ 1 {∣ I 2 ∣ > 0 } .
This dual perspective leads to different ways of simulating trajectories of X \mathbf{X} X .
Fix a partition π = { 0 = T 0 < ⋯ < T n = T } \pi=\{0=T_0<\cdots<T_n=T\} π = { 0 = T 0 < ⋯ < T n = T } . Consider a trajectory ( X t ) t ∈ [ 0 , T ] (X_t)_{t\in[0,T]} ( X t ) t ∈ [ 0 , T ] and fix ( τ i ) i = 1 ∞ (\tau_i)_{i=1}^{\infty} ( τ i ) i = 1 ∞ and ( Y i ) i = 1 ∞ (Y^i)_{i=1}^{\infty} ( Y i ) i = 1 ∞ such that ∑ s ≤ t Δ X s = ∑ i = 1 ∞ Y i 1 { τ i ≤ t } \sum_{s\leq t}\Delta X_s=\sum_{i=1}^{\infty}Y^i1_{\{\tau_i\leq t\}} ∑ s ≤ t Δ X s = ∑ i = 1 ∞ Y i 1 { τ i ≤ t } . Set π τ = { 0 = t 0 < ⋯ < t n τ = T } : = π ∪ { τ i : τ i ≤ T } \pi^{\tau}=\{0=t_0<\cdots<t_{n_{\tau}}=T\}:=\pi\cup\{\tau_i:\tau_i\leq T\} π τ = { 0 = t 0 < ⋯ < t n τ = T } := π ∪ { τ i : τ i ≤ T } .
From X t 0 π τ = ϵ ∅ X_{t_0}^{\pi^\tau}=\epsilon_{\varnothing} X t 0 π τ = ϵ ∅ , one may use the coordinatewise Euler scheme
⟨ ϵ I , X t k + 1 π τ ⟩ = ⟨ ϵ I , X t k π τ ⟩ + ⟨ ϵ I ′ , X t k π τ ⟩ ( X t k + 1 i ∣ I ∣ − X t k i ∣ I ∣ ) + 1 2 ⟨ ϵ I ′ ′ , X t k π τ ⟩ ( [ X i ∣ I ∣ − 1 , X i ∣ I ∣ ] t k + 1 c − [ X i ∣ I ∣ − 1 , X i ∣ I ∣ ] t k c ) + ∑ i = 1 ∞ ⟨ ϵ I , Y i ⟩ 1 t k + 1 = τ i , \langle\epsilon_I,X_{t_{k+1}}^{\pi^\tau}\rangle=\langle\epsilon_I,X_{t_k}^{\pi^\tau}\rangle+\langle\epsilon_{I'},X_{t_k}^{\pi^\tau}\rangle\left(X_{t_{k+1}}^{i|I|}-X_{t_k}^{i|I|}\right)
+\frac{1}{2}\langle\epsilon_{I''},X_{t_k}^{\pi^\tau}\rangle\left([X^{i|I|-1},X^{i|I|}]_{t_{k+1}}^c-[X^{i|I|-1},X^{i|I|}]_{t_k}^c\right)+\sum_{i=1}^{\infty}\langle\epsilon_I,Y^i\rangle\mathbf{1}_{t_{k+1}=\tau_i}, ⟨ ϵ I , X t k + 1 π τ ⟩ = ⟨ ϵ I , X t k π τ ⟩ + ⟨ ϵ I ′ , X t k π τ ⟩ ( X t k + 1 i ∣ I ∣ − X t k i ∣ I ∣ ) + 2 1 ⟨ ϵ I ′′ , X t k π τ ⟩ ( [ X i ∣ I ∣ − 1 , X i ∣ I ∣ ] t k + 1 c − [ X i ∣ I ∣ − 1 , X i ∣ I ∣ ] t k c ) + i = 1 ∑ ∞ ⟨ ϵ I , Y i ⟩ 1 t k + 1 = τ i ,
where the contribution from the jumps is given by
⟨ ϵ I , Y i ⟩ : = ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I 1 ∣ I 2 ∣ ! 1 { ∣ I 2 ∣ > 1 } ( ⟨ ϵ I 1 , X τ i − π τ ⟩ − ⟨ ϵ I 1 , Y i ⟩ ) ⟨ ϵ I 2 , ( Y i ) ⊗ ∣ I 2 ∣ ⟩ . \langle\epsilon_I,Y^i\rangle:=\sum_{\epsilon_{I_1}\otimes\epsilon_{I_2}=\epsilon_I}\frac{1}{|I_2|!}\mathbf{1}_{\{|I_2|>1\}}\left(\langle\epsilon_{I_1},X_{\tau_i-}^{\pi^\tau}\rangle-\langle\epsilon_{I_1},Y^i\rangle\right)\langle\epsilon_{I_2},(Y^i)^{\otimes|I_2|}\rangle. ⟨ ϵ I , Y i ⟩ := ϵ I 1 ⊗ ϵ I 2 = ϵ I ∑ ∣ I 2 ∣ ! 1 1 { ∣ I 2 ∣ > 1 } ( ⟨ ϵ I 1 , X τ i − π τ ⟩ − ⟨ ϵ I 1 , Y i ⟩ ) ⟨ ϵ I 2 , ( Y i ) ⊗ ∣ I 2 ∣ ⟩ .
Alternatively, by Examples 4.20 and 4.21 in Chevrey and Friz (2019) one can compute the signature of X X X by any signature package based on a piecewise linear approximation of the time series. Given a sequence of deterministic nested partitions, the cited examples establish convergence in probability of the approximation at every time point contained in the union of the partitions.
Finally, one can replace each jump of X X X by a linear segment joining X τ i − X_{\tau_i-} X τ i − to X τ i X_{\tau_i} X τ i , thereby obtaining a continuous path whose signature can be computed using a standard signature package. Since such packages do not typically require a uniform time grid, this can be achieved by considering the time series ( X t k ) k = 0 n ˉ (X_{t_k})_{k=0}^{\bar n} ( X t k ) k = 0 n ˉ , where
t ˉ k ∈ π τ ∪ { τ i : τ i ≤ T } \bar t_k\in\pi^\tau\cup\{\tau_i:\tau_i\leq T\} t ˉ k ∈ π τ ∪ { τ i : τ i ≤ T }
and X τ i − : = X τ i − Y i X_{\tau_i-}:=X_{\tau_i}-Y^i X τ i − := X τ i − Y i . If the semimartingale admits deterministic jump times such that the sequence of partitions can be chosen deterministically, then the Wong-Zakai theorem implies uniform convergence in probability of the approximated signature to the true one.
Entire sig-jump-diffusions
Characteristics of entire sig-jump-diffusions
We introduce now the set of suitable coefficients of the processes which will be studied in this paper. Fix S ⊆ R d S\subseteq\mathbb{R}^d S ⊆ R d and recall the set S S S introduced in equation (1) .
Definition 3.1. We call sig-coefficients any triplet ( b , a , K ) (b,a,K) ( b , a , K ) such that
b : S → R d , a : S → S + d , b:S\to\mathbb{R}^d,\qquad a:S\to S_+^d, b : S → R d , a : S → S + d , and K K K is a kernel from S S S to R d \mathbb{R}^d R d such that the following conditions are verified. For all x ∈ S \mathbf{x}\in S x ∈ S ,
b ( x ) = ⟨ b , x ⟩ b(\mathbf{x})=\langle\mathbf{b},\mathbf{x}\rangle b ( x ) = ⟨ b , x ⟩ , for some b ∈ S ∗ d \mathbf{b}\in S^{*d} b ∈ S ∗ d ;
a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) = ⟨ a + k 2 , x ⟩ a(\mathbf{x})+\int_{\mathbb{R}^d}\xi^{\otimes2}K(\mathbf{x},d\xi)=\langle\mathbf{a}+\mathbf{k}^2,\mathbf{x}\rangle a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) = ⟨ a + k 2 , x ⟩ , for some a + k 2 ∈ ( S ∗ d ) ⊗ 2 \mathbf{a}+\mathbf{k}^2\in(S^{*d})^{\otimes2} a + k 2 ∈ ( S ∗ d ) ⊗ 2 ;
K ( x , { 0 } ) = 0 K(\mathbf{x},\{0\})=0 K ( x , { 0 }) = 0 and
∫ R d ξ ⊗ n K ( x , d ξ ) = ⟨ k n , x ⟩ ∈ ( R d ) ⊗ n , \int_{\mathbb{R}^d}\xi^{\otimes n}K(\mathbf{x},d\xi)=\langle\mathbf{k}^n,\mathbf{x}\rangle\in(\mathbb{R}^d)^{\otimes n}, ∫ R d ξ ⊗ n K ( x , d ξ ) = ⟨ k n , x ⟩ ∈ ( R d ) ⊗ n , for some k n ∈ ( S ∗ d ) ⊗ n \mathbf{k}^n\in(S^{*d})^{\otimes n} k n ∈ ( S ∗ d ) ⊗ n , n ≥ 3 n\geq3 n ≥ 3 .
Definition 3.3. Fix some sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) with K K K such that ∫ R d ( ∥ ξ ∥ ∧ ∥ ξ ∥ 2 ) K ( x , d ξ ) < ∞ \int_{\mathbb{R}^{d}}(\lVert\xi\rVert\wedge\lVert\xi\rVert^{2})K(x,d\xi)<\infty ∫ R d (∥ ξ ∥ ∧ ∥ ξ ∥ 2 ) K ( x , d ξ ) < ∞ , for all x ∈ S x\in S x ∈ S . Let X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be a ( S + x 0 ) (S+x_0) ( S + x 0 ) -valued semimartingale with X 0 = x 0 ∈ R d X_0=x_0\in\mathbb{R}^{d} X 0 = x 0 ∈ R d on some filtered probability space ( Ω , F , ( F t ) , P ) (\Omega,\mathcal{F},(\mathcal{F}_t),\mathbb{P}) ( Ω , F , ( F t ) , P ) and let X : = ( X t ) t ∈ [ 0 , T ] \mathbf{X}:=(\mathbf{X}_t)_{t\in[0,T]} X := ( X t ) t ∈ [ 0 , T ] denote its signature. Assume that the maps s ↦ b ( X s − ) , a ( X s − ) , ∫ R d g ( ξ ) K ( X s − , d ξ ) s\mapsto b(X_{s-}),a(X_{s-}),\int_{\mathbb{R}^{d}}g(\xi)K(X_{s-},d\xi) s ↦ b ( X s − ) , a ( X s − ) , ∫ R d g ( ξ ) K ( X s − , d ξ ) are a.s. left-continuous for g = 1 A g=\mathbf{1}_{A} g = 1 A and A A A being a Borel set in R d \mathbb{R}^{d} R d , and g ( ξ ) = ξ ⊗ n g(\xi)=\xi^{\otimes n} g ( ξ ) = ξ ⊗ n for n ≥ 2 n\ge2 n ≥ 2 .
We call X X X an entire sig-jump-diffusion with sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) if for all bounded f ∈ C 2 ( R d ) f\in C^{2}(\mathbb{R}^{d}) f ∈ C 2 ( R d ) , the process N f : = ( N t f ) t ∈ [ 0 , T ] N^{f}:=(N^{f}_{t})_{t\in[0,T]} N f := ( N t f ) t ∈ [ 0 , T ] given by
N t f : = f ( X t ) − f ( x 0 ) − ∫ 0 t ( ⟨ ∇ f ( X s − ) , b ( X s − ) ⟩ + 1 2 Tr ( a ( X s − ) ∇ 2 f ( X s − ) ) N^{f}_{t}:=f(X_{t})-f(x_{0})-\int_{0}^{t}\left(\langle\nabla f(X_{s-}),b(X_{s-})\rangle+\frac{1}{2}\operatorname{Tr}(a(X_{s-})\nabla^{2}f(X_{s-}))\right. N t f := f ( X t ) − f ( x 0 ) − ∫ 0 t ( ⟨ ∇ f ( X s − ) , b ( X s − )⟩ + 2 1 Tr ( a ( X s − ) ∇ 2 f ( X s − )) + ∫ R d ( f ( X s − + ξ ) − f ( X s − ) − ∇ f ( X s − ) T ξ ) K ( X s − , d ξ ) ) d s \left.+\int_{\mathbb{R}^{d}}\left(f(X_{s-}+\xi)-f(X_{s-})-\nabla f(X_{s-})^{\mathsf{T}}\xi\right)K(X_{s-},d\xi)\right)ds + ∫ R d ( f ( X s − + ξ ) − f ( X s − ) − ∇ f ( X s − ) T ξ ) K ( X s − , d ξ ) ) d s defines a local martingale. We also denote by X g X^{g} X g the entire sig-jump-diffusion with sig-coefficients
b g ( x ) : = b ( g ⊗ x ) , a g ( x ) : = a ( g ⊗ x ) , K g ( x , d ξ ) : = K ( g ⊗ x , d ξ ) . b^{g}(x):=b(g\otimes x),\qquad a^{g}(x):=a(g\otimes x),\qquad K^{g}(x,d\xi):=K(g\otimes x,d\xi). b g ( x ) := b ( g ⊗ x ) , a g ( x ) := a ( g ⊗ x ) , K g ( x , d ξ ) := K ( g ⊗ x , d ξ ) .
Remark 3.4. (i) Notice that by Proposition II.2.29 and Theorem II.2.42 in Jacod and Shiryaev [46 ] , X X X is a special semimartingale with characteristics with respect to the truncation function χ ( ξ ) = ξ \chi(\xi)=\xi χ ( ξ ) = ξ ,
( ∫ 0 ⋅ b ( X s − ) d s , ∫ 0 ⋅ a ( X s − ) d s , K ( X t − , d ξ ) d t ) . \left(\int_{0}^{\cdot}b(X_{s-})ds,\int_{0}^{\cdot}a(X_{s-})ds,K(X_{t-},d\xi)dt\right). ( ∫ 0 ⋅ b ( X s − ) d s , ∫ 0 ⋅ a ( X s − ) d s , K ( X t − , d ξ ) d t ) .
It is important to note that X X X is not necessarily a jump-diffusion in the sense of Definition II.2.18 in Jacod and Shiryaev [46 ] . This property can nevertheless be shown to hold for the signature process X \mathbf{X} X , when viewed as an infinite-dimensional process (Lemma 3.5).
(ii) The class of entire sig-jump-diffusions constitutes an extension of the class of signature SDEs introduced in Cuchiero et al. [27 ] , where only processes with continuous trajectories are considered. In particular, it includes all the continuous signature models for asset pricing that have been proposed so far Perez Arribas et al. [62 ] ; ); Cuchiero et al. [21 , 20 ] where only diffusive drivers and finite sums of signature components are considered, as well as the Lévy-type signature models introduced in Cuchiero et al. [25 ] .
Expected signature of entire sig-jump-diffusions
In this section, we exploit duality methods for stochastic processes to show that the expected (or rather expected quantities of the) signature of entire sig-jump-diffusions can be computed analytically by solving certain linear ODEs in the extended tensor algebra.
As a first step, we study the dynamics of the signature of such processes.
Lemma 3.5. Let X = ( X t ) t ∈ [ 0 , T ] X=(X_{t})_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be an entire sig-jump-diffusion with sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) and fix a multiindex I I I . Then, for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
⟨ ε I , X t ⟩ = ∫ 0 t ( ⟨ ε I , X s − ⟩ ⟨ b ( i I ) , X s − ⟩ 1 { ∣ I ∣ > 0 } + 1 2 ⟨ ε I ′ ′ , X s − ⟩ ⟨ ( a + k 2 ) ( i I − 1 , i I ) , X s − ⟩ 1 { ∣ I ∣ > 1 } + ∑ I 1 ⊗ I 2 = I ε I 1 ⊗ ε I 2 = ε I 1 ∣ I 2 ∣ ! 1 { ∣ I 2 ∣ > 2 } ⟨ ε I 1 , X s − ⟩ ∫ R d ⟨ ε I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ K ( X s − , d ξ ) ) d s + local martingale. \begin{aligned}
\langle\varepsilon_{I},\mathbf{X}_{t}\rangle=\int_{0}^{t}(\langle\varepsilon_{I},\mathbf{X}_{s-}\rangle\langle b^{(i_{I})},\mathbf{X}_{s-}\rangle\mathbf{1}_{\{|I|>0\}}+\frac{1}{2}\langle\varepsilon_{I''},\mathbf{X}_{s-}\rangle\langle(a+\mathbf{k}^{2})^{(i_{I-1},i_{I})},\mathbf{X}_{s-}\rangle\mathbf{1}_{\{|I|>1\}} \\
&+\sum_{\substack{I_{1}\otimes I_{2}=I\\ \varepsilon_{I_{1}}\otimes\varepsilon_{I_{2}}=\varepsilon_{I}}}\frac{1}{|I_{2}|!}\mathbf{1}_{\{|I_{2}|>2\}}\langle\varepsilon_{I_{1}},\mathbf{X}_{s-}\rangle\int_{\mathbb{R}^{d}}\langle\varepsilon_{I_{2}},\xi^{\otimes|I_{2}|}\rangle K(X_{s-},d\xi))ds \\
&+\text{local martingale.}
\end{aligned} ⟨ ε I , X t ⟩ = ∫ 0 t (⟨ ε I , X s − ⟩ ⟨ b ( i I ) , X s − ⟩ 1 { ∣ I ∣ > 0 } + 2 1 ⟨ ε I ′′ , X s − ⟩ ⟨( a + k 2 ) ( i I − 1 , i I ) , X s − ⟩ 1 { ∣ I ∣ > 1 } + I 1 ⊗ I 2 = I ε I 1 ⊗ ε I 2 = ε I ∑ ∣ I 2 ∣ ! 1 1 { ∣ I 2 ∣ > 2 } ⟨ ε I 1 , X s − ⟩ ∫ R d ⟨ ε I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ K ( X s − , d ξ )) d s + local martingale.
Proof. First recall that the characteristics of X X X are given by
∫ 0 ⋅ b ( X s − ) d s , ∫ 0 ⋅ a ( X s − ) d s , K ( X s − , d ξ ) d s , \int_{0}^{\cdot}b(X_{s-})ds,\qquad\int_{0}^{\cdot}a(X_{s-})ds,\qquad K(X_{s-},d\xi)ds, ∫ 0 ⋅ b ( X s − ) d s , ∫ 0 ⋅ a ( X s − ) d s , K ( X s − , d ξ ) d s ,
Remark 3.4. Since by definition of entire sig-jump-diffusion the map s ↦ ∫ R d ξ ⊗ n K ( X s − , d ξ ) s \mapsto\int_{\mathbb{R}^{d}} \xi^{\otimes n} K(\mathsf{X}_{s-},d\xi) s ↦ ∫ R d ξ ⊗ n K ( X s − , d ξ ) is a.s. left-continuous for n ≥ 2 n \ge2 n ≥ 2 , we have that
∫ 0 T ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I ∣ I 2 ∣ ≥ 2 1 ∣ I 2 ∣ ! ⟨ ϵ I 1 , X t − ⟩ ∫ R d ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ K ( X t − , d ξ ) d t < ∞ a.s. \int_{0}^{T} \sum_{\substack{\epsilon_{I_{1}}\otimes\epsilon_{I_{2}}=\epsilon_{I}\\ |I_{2}|\ge2}} \frac{1}{|I_{2}|!}\left\langle\epsilon_{I_{1}},\mathsf{X}_{t-}\right\rangle\int_{\mathbb{R}^{d}}\left\langle\epsilon_{I_{2}},\xi^{\otimes|I_{2}|}\right\rangle K(\mathsf{X}_{t-},d\xi)\,dt < \infty\quad\text{a.s.} ∫ 0 T ϵ I 1 ⊗ ϵ I 2 = ϵ I ∣ I 2 ∣ ≥ 2 ∑ ∣ I 2 ∣ ! 1 ⟨ ϵ I 1 , X t − ⟩ ∫ R d ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ K ( X t − , d ξ ) d t < ∞ a.s.
By Theorem II.1.8 in Jacob and Shiryaev [46 ] , the compensator of the second line of equation (2) is given by
∫ 0 T ∫ R d ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I ∣ I 2 ∣ > 1 1 ∣ I 2 ∣ ! ⟨ ϵ I 1 , X s − ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ K ( X s − , d ξ ) d s . \int_{0}^{T}\int_{\mathbb{R}^{d}}\sum_{\substack{\epsilon_{I_{1}}\otimes\epsilon_{I_{2}}=\epsilon_{I}\\ |I_{2}|>1}} \frac{1}{|I_{2}|!}\left\langle\epsilon_{I_{1}},\mathsf{X}_{s-}\right\rangle\left\langle\epsilon_{I_{2}},\xi^{\otimes|I_{2}|}\right\rangle K(\mathsf{X}_{s-},d\xi)\,ds. ∫ 0 T ∫ R d ϵ I 1 ⊗ ϵ I 2 = ϵ I ∣ I 2 ∣ > 1 ∑ ∣ I 2 ∣ ! 1 ⟨ ϵ I 1 , X s − ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ K ( X s − , d ξ ) d s .
The claim follows by definition of characteristics.
Lemma 3.5 can be reformulated in a more compact way with the following notation.
Definition 3.6. Define
Q : = ( 0 , b , 1 2 ( a + k 2 ) , 1 3 ! k 3 , … , 1 n ! k n , … ) . \mathbf{Q}:=\left(0,b,\frac{1}{2}(a+k^{2}),\frac{1}{3!}k^{3},\ldots,\frac{1}{n!}k^{n},\ldots\right). Q := ( 0 , b , 2 1 ( a + k 2 ) , 3 ! 1 k 3 , … , n ! 1 k n , … ) .
such that for each I I I and all x ∈ S \mathbf{x}\in\mathcal{S} x ∈ S , it holds ⟨ Q I , x ⟩ = Q ( x ) I \left\langle\mathbf{Q}^{I},\mathbf{x}\right\rangle=Q(\mathbf{x})_{I} ⟨ Q I , x ⟩ = Q ( x ) I , for Q : S → T ( R d ) Q:\mathcal{S}\to T(\mathbb{R}^{d}) Q : S → T ( R d ) defined via
Q ( x ) : = ( 0 , b ( x ) , 1 2 ( a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) ) , … , 1 k ! ∫ R d ξ ⊗ k K ( x , d ξ ) , … ) . (3) Q(\mathbf{x}):=\left(0,b(\mathbf{x}),\frac{1}{2}\left(a(\mathbf{x})+\int_{\mathbb{R}^{d}}\xi^{\otimes2}K(\mathbf{x},d\xi)\right),\ldots,\frac{1}{k!}\int_{\mathbb{R}^{d}}\xi^{\otimes k}K(\mathbf{x},d\xi),\ldots\right). \tag*{(3)} Q ( x ) := ( 0 , b ( x ) , 2 1 ( a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) ) , … , k ! 1 ∫ R d ξ ⊗ k K ( x , d ξ ) , … ) . (3)
For each k ∈ N 0 k\in\mathbb{N}_{0} k ∈ N 0 , u ∈ S ∗ \mathbf{u}\in\mathcal{S}^{*} u ∈ S ∗ , set
u ( k ) ⊡ Q : = ∑ ∣ I ∣ = k u I ( k ) ⊡ Q I . \mathbf{u}^{(k)}\boxdot\mathbf{Q}:=\sum_{|I|=k}u_{I}^{(k)}\boxdot\mathbf{Q}^{I}. u ( k ) ⊡ Q := ∣ I ∣ = k ∑ u I ( k ) ⊡ Q I .
Remark 3.7. (i) Notice that setting ( S d ) ⊗ 0 : = S (\mathcal{S}^{d})^{\otimes0}:=\mathcal{S} ( S d ) ⊗ 0 := S , Q ∈ T ( ( S ∗ ) d ) \mathbf{Q}\in T((\mathcal{S}^{*})^{d}) Q ∈ T (( S ∗ ) d ) .
(ii) By Lemma 4.3(ii) in Cuchiero et al. [27 ] , if for k ∈ N 0 k\in\mathbb{N}_{0} k ∈ N 0 and all I I I , u I ( k ) ∈ S ∗ u_{I}^{(k)}\in\mathcal{S}^{*} u I ( k ) ∈ S ∗ , then u ( k ) ⊡ Q ∈ S ∗ \mathbf{u}^{(k)}\boxdot\mathbf{Q}\in\mathcal{S}^{*} u ( k ) ⊡ Q ∈ S ∗ .
Corollary 3.8. Let X = ( X t ) t ∈ [ 0 , T ] X=(X_{t})_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be an entire sig-jump-diffusion with sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) . Then the process M = ( M t ) t ∈ [ 0 , T ] M=(M_{t})_{t\in[0,T]} M = ( M t ) t ∈ [ 0 , T ] defined via
M t : = ⟨ u , X t ⟩ − ⟨ u , ϵ ∅ ⟩ − ∫ 0 t ⟨ L ( u ) , X s ⟩ d s , t ∈ [ 0 , T ] , (4) M_{t}:=\left\langle\mathbf{u},X_{t}\right\rangle-\left\langle\mathbf{u},\epsilon_{\varnothing}\right\rangle-\int_{0}^{t}\left\langle L(\mathbf{u}),X_{s}\right\rangle ds,\quad t\in[0,T], \tag*{(4)} M t := ⟨ u , X t ⟩ − ⟨ u , ϵ ∅ ⟩ − ∫ 0 t ⟨ L ( u ) , X s ⟩ d s , t ∈ [ 0 , T ] , (4)
is a local martingale with M 0 = 0 M_{0}=0 M 0 = 0 , for each u ∈ T ( R d ) \mathbf{u}\in T(\mathbb{R}^{d}) u ∈ T ( R d ) , where
L ( u ) : = ∑ k = 0 ∞ u ( k ) ⊡ Q . (5) L(\mathbf{u}):=\sum_{k=0}^{\infty}\mathbf{u}^{(k)}\boxdot\mathbf{Q}. \tag*{(5)} L ( u ) := k = 0 ∑ ∞ u ( k ) ⊡ Q . (5)
Proof. Observe that for each I I I by Lemma 3.5 it holds L ( ϵ I ) = ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I ϵ I 1 ⊡ Q I 2 L(\epsilon_{I})=\sum_{\epsilon_{I_{1}}\otimes\epsilon_{I_{2}}=\epsilon_{I}}\epsilon_{I_{1}}\boxdot\mathbf{Q}^{I_{2}} L ( ϵ I ) = ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I ϵ I 1 ⊡ Q I 2 . The result follows by linearity.
Corollary 3.9. Fix g ∈ S g\in\mathcal{S} g ∈ S , some sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) and let X g = ( X t g ) t ∈ [ 0 , T ] X^{g}=(X_{t}^{g})_{t\in[0,T]} X g = ( X t g ) t ∈ [ 0 , T ] be as in Definition 3.3. Then the process M g : = ( M t g ) t ∈ [ 0 , T ] M^{g}:=(M_{t}^{g})_{t\in[0,T]} M g := ( M t g ) t ∈ [ 0 , T ] defined via
M t g : = ⟨ u , g ⊗ X t g ⟩ − ⟨ u , g ⟩ − ∫ 0 t ⟨ L ( u ) , g ⊗ X s g ⟩ d s , t ∈ [ 0 , T ] , M_{t}^{g}:=\left\langle\mathbf{u},g\otimes X_{t}^{g}\right\rangle-\left\langle\mathbf{u},g\right\rangle-\int_{0}^{t}\left\langle L(\mathbf{u}),g\otimes X_{s}^{g}\right\rangle ds,\quad t\in[0,T], M t g := ⟨ u , g ⊗ X t g ⟩ − ⟨ u , g ⟩ − ∫ 0 t ⟨ L ( u ) , g ⊗ X s g ⟩ d s , t ∈ [ 0 , T ] ,
is a local martingale with M 0 g = 0 M_{0}^{g}=0 M 0 g = 0 , for each u ∈ T ( R d ) \mathbf{u}\in T(\mathbb{R}^{d}) u ∈ T ( R d ) , where
L ( u ) : = ∑ k = 0 ∞ u ( k ) ⊡ Q . L(\mathbf{u}):=\sum_{k=0}^{\infty}\mathbf{u}^{(k)}\boxdot\mathbf{Q}. L ( u ) := k = 0 ∑ ∞ u ( k ) ⊡ Q .
Proof. For each u ∈ T ( R d ) \mathbf{u}\in T(\mathbb{R}^{d}) u ∈ T ( R d ) define u g ∈ T ( R d ) \mathbf{u}^{\mathbf{g}}\in T(\mathbb{R}^{d}) u g ∈ T ( R d ) as u I g : = ⟨ u , g ⊗ e I ⟩ \mathbf{u}^{\mathbf{g}}_{I}:=\langle\mathbf{u},\mathbf{g}\otimes\mathbf{e}_{I}\rangle u I g := ⟨ u , g ⊗ e I ⟩ , and note that ⟨ u , g ⊗ x ⟩ = ⟨ u g , x ⟩ \langle\mathbf{u},\mathbf{g}\otimes\mathbf{x}\rangle=\langle\mathbf{u}^{\mathbf{g}},\mathbf{x}\rangle ⟨ u , g ⊗ x ⟩ = ⟨ u g , x ⟩ . Noting that the operator Q g \mathbf{Q}^{\mathbf{g}} Q g corresponding to the sig-coefficients ( b g , a g , K g ) (b^{\mathbf{g}},a^{\mathbf{g}},K^{\mathbf{g}}) ( b g , a g , K g ) is given by
Q g : = ( 0 , b g , 1 2 ( a g + ( k 2 ) g ) , 1 3 ! ( k 3 ) g , … , 1 n ! ( k n ) g , … ) , \mathbf{Q}^{\mathbf{g}}:=\left(0,b^{\mathbf{g}},\frac{1}{2}(a^{\mathbf{g}}+(\mathbf{k}^{2})^{\mathbf{g}}),\frac{1}{3!}(\mathbf{k}^{3})^{\mathbf{g}},\ldots,\frac{1}{n!}(\mathbf{k}^{n})^{\mathbf{g}},\ldots\right), Q g := ( 0 , b g , 2 1 ( a g + ( k 2 ) g ) , 3 ! 1 ( k 3 ) g , … , n ! 1 ( k n ) g , … ) ,
by Corollary 3.8 we obtain that the process ( N t g ) t ∈ [ 0 , T ] (N^{\mathbf{g}}_{t})_{t\in[0,T]} ( N t g ) t ∈ [ 0 , T ] is a local martingale with N 0 g = 0 N^{\mathbf{g}}_{0}=0 N 0 g = 0 , where
N t g : = ⟨ u g , X t ⟩ − ⟨ u g , ϵ ⟩ − ∫ 0 t ⟨ L g ( u g ) , X s ⟩ d s , N^{\mathbf{g}}_{t}:=\langle\mathbf{u}^{\mathbf{g}},X_{t}\rangle-\langle\mathbf{u}^{\mathbf{g}},\epsilon\rangle-\int_{0}^{t}\langle L^{\mathbf{g}}(\mathbf{u}^{\mathbf{g}}),X_{s}\rangle\,ds, N t g := ⟨ u g , X t ⟩ − ⟨ u g , ϵ ⟩ − ∫ 0 t ⟨ L g ( u g ) , X s ⟩ d s ,
for L g ( u ) : = ∑ k = 0 ∞ u ( k ) ⊳ Q g L^{\mathbf{g}}(\mathbf{u}):=\sum_{k=0}^{\infty}\mathbf{u}^{(k)}\mathbin{\rhd}\mathbf{Q}^{\mathbf{g}} L g ( u ) := ∑ k = 0 ∞ u ( k ) ⊳ Q g . To conclude the proof we now just need to show that N g = M g N^{\mathbf{g}}=M^{\mathbf{g}} N g = M g , which, by definition of u g \mathbf{u}^{\mathbf{g}} u g , reduces to showing that L g ( u g ) = ( L ( u ) ) g L^{\mathbf{g}}(\mathbf{u}^{\mathbf{g}})=(L(\mathbf{u}))^{\mathbf{g}} L g ( u g ) = ( L ( u ) ) g . For ∣ I ∣ = k |I|=k ∣ I ∣ = k , we have
⟨ ( u g ) I ( k ) , x ⟩ = ⟨ u g , x ⊗ e I ⟩ = ⟨ u , g ⊗ x ⊗ e I ⟩ = ⟨ u I ( k ) , g ⊗ x ⟩ = ⟨ ( u I ( k ) ) g , x ⟩ \langle(\mathbf{u}^{\mathbf{g}})^{(k)}_{I},\mathbf{x}\rangle=\langle\mathbf{u}^{\mathbf{g}},\mathbf{x}\otimes\mathbf{e}_{I}\rangle=\langle\mathbf{u},\mathbf{g}\otimes\mathbf{x}\otimes\mathbf{e}_{I}\rangle=\langle\mathbf{u}^{(k)}_{I},\mathbf{g}\otimes\mathbf{x}\rangle=\langle(\mathbf{u}^{(k)}_{I})^{\mathbf{g}},\mathbf{x}\rangle ⟨( u g ) I ( k ) , x ⟩ = ⟨ u g , x ⊗ e I ⟩ = ⟨ u , g ⊗ x ⊗ e I ⟩ = ⟨ u I ( k ) , g ⊗ x ⟩ = ⟨( u I ( k ) ) g , x ⟩
and
⟨ u g ⊔ v g , x ⟩ = ⟨ u g , x ⟩ ⟨ v g , x ⟩ = ⟨ u , g ⊗ x ⟩ ⟨ v , g ⊗ x ⟩ = ⟨ u ⊔ v , g ⊗ x ⟩ = ⟨ ( u ⊔ v ) g , x ⟩ , \langle\mathbf{u}^{\mathbf{g}}\mathbin{\sqcup}\mathbf{v}^{\mathbf{g}},\mathbf{x}\rangle=\langle\mathbf{u}^{\mathbf{g}},\mathbf{x}\rangle\langle\mathbf{v}^{\mathbf{g}},\mathbf{x}\rangle=\langle\mathbf{u},\mathbf{g}\otimes\mathbf{x}\rangle\langle\mathbf{v},\mathbf{g}\otimes\mathbf{x}\rangle=\langle\mathbf{u}\mathbin{\sqcup}\mathbf{v},\mathbf{g}\otimes\mathbf{x}\rangle=\langle(\mathbf{u}\mathbin{\sqcup}\mathbf{v})^{\mathbf{g}},\mathbf{x}\rangle, ⟨ u g ⊔ v g , x ⟩ = ⟨ u g , x ⟩ ⟨ v g , x ⟩ = ⟨ u , g ⊗ x ⟩ ⟨ v , g ⊗ x ⟩ = ⟨ u ⊔ v , g ⊗ x ⟩ = ⟨( u ⊔ v ) g , x ⟩ ,
the claim follows.
Remark 3.10. Notice that in (5) the sum is finite and thus well defined. In particular for each I I I it holds ⟨ L ( ϵ I ) , x ⟩ = ⟨ ϵ I , x ∘ Q ( x ) ⟩ \langle L(\epsilon_{I}),\mathbf{x}\rangle=\langle\epsilon_{I},\mathbf{x}\mathbin{\circ}Q(\mathbf{x})\rangle ⟨ L ( ϵ I ) , x ⟩ = ⟨ ϵ I , x ∘ Q ( x )⟩ , where for each x ∈ S \mathbf{x}\in S x ∈ S , Q ( x ) Q(\mathbf{x}) Q ( x ) is given in equation (3) .
Now, we aim to extend the decomposition (4) to some process ⟨ u , X ⟩ \langle\mathbf{u},\mathbb{X}\rangle ⟨ u , X ⟩ , for u ∈ S ∗ \mathbf{u}\in S^{*} u ∈ S ∗ with u ∈ T ( ( R d ) ) ∖ T ( R d ) + i T ( ( R d ) ) ∖ T ( R d ) \mathbf{u}\in T((\mathbb{R}^{d}))\setminus T(\mathbb{R}^{d})+iT((\mathbb{R}^{d}))\setminus T(\mathbb{R}^{d}) u ∈ T (( R d )) ∖ T ( R d ) + i T (( R d )) ∖ T ( R d ) . As a first step, we extend the operator L L L in (5) to a subset of its pointwise closure.
Definition 3.11. For each u ∈ S ∗ \mathbf{u}\in S^{*} u ∈ S ∗ , we write u ∈ D ( L ‾ ) \mathbf{u}\in\mathcal{D}(\overline{L}) u ∈ D ( L ) and
⟨ L ‾ ( u ) , x ⟩ : = lim N → ∞ ⟨ L ( u ≤ N ) , x ⟩ \langle\overline{L}(\mathbf{u}),\mathbf{x}\rangle:=\lim_{N\to\infty}\langle L(\mathbf{u}^{\leq N}),\mathbf{x}\rangle ⟨ L ( u ) , x ⟩ := N → ∞ lim ⟨ L ( u ≤ N ) , x ⟩
whenever the limit is well defined for each x ∈ S \mathbf{x}\in S x ∈ S .
Theorem 3.12. Let X = ( X t ) t ∈ [ 0 , T ] X=(X_{t})_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be an entire sig-jump-diffusion with sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) and denote by X \mathbb{X} X its signature. Fix U ⊂ D ( L ‾ ) \mathcal{U}\subset\mathcal{D}(\overline{L}) U ⊂ D ( L ) and for each u ∈ U \mathbf{u}\in\mathcal{U} u ∈ U , define V 1 : = ( V t 1 ) t ∈ [ 0 , T ] V^{1}:=(V^{1}_{t})_{t\in[0,T]} V 1 := ( V t 1 ) t ∈ [ 0 , T ] and V 2 : = ( V t 2 ) t ∈ [ 0 , T ] V^{2}:=(V^{2}_{t})_{t\in[0,T]} V 2 := ( V t 2 ) t ∈ [ 0 , T ] via
V t 1 : = sup N ∈ N ∣ ⟨ u ≤ N , X t ⟩ ∣ , V t 2 : = sup N ∈ N ∣ ⟨ L ( u ≤ N ) , X t ⟩ ∣ , V^{1}_{t}:=\sup_{N\in\mathbb{N}}|\langle\mathbf{u}^{\leq N},X_{t}\rangle|,\qquad V^{2}_{t}:=\sup_{N\in\mathbb{N}}|\langle L(\mathbf{u}^{\leq N}),X_{t}\rangle|, V t 1 := N ∈ N sup ∣ ⟨ u ≤ N , X t ⟩ ∣ , V t 2 := N ∈ N sup ∣ ⟨ L ( u ≤ N ) , X t ⟩ ∣ , for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Suppose that ( sup s ≤ t V s 1 ) t ∈ [ 0 , T ] (\sup_{s\leq t}V^{1}_{s})_{t\in[0,T]} ( sup s ≤ t V s 1 ) t ∈ [ 0 , T ] is locally integrable, and V 2 V^{2} V 2 and ⟨ u , X ⟩ \langle\mathbf{u},\mathbb{X}\rangle ⟨ u , X ⟩ are càdlàg processes, for each u ∈ U \mathbf{u}\in\mathcal{U} u ∈ U . Then the process M : = ( M t ) t ∈ [ 0 , T ] M:=(M_{t})_{t\in[0,T]} M := ( M t ) t ∈ [ 0 , T ] defined via
M t : = ⟨ u , X t ⟩ − ⟨ u , ϵ ⟩ − ∫ 0 t ⟨ L ‾ ( u ) , X s ⟩ d s , t ∈ [ 0 , T ] , (6) M_{t}:=\langle\mathbf{u},X_{t}\rangle-\langle\mathbf{u},\epsilon\rangle-\int_{0}^{t}\langle\overline{L}(\mathbf{u}),X_{s}\rangle\,ds,\qquad t\in[0,T], \tag*{(6)} M t := ⟨ u , X t ⟩ − ⟨ u , ϵ ⟩ − ∫ 0 t ⟨ L ( u ) , X s ⟩ d s , t ∈ [ 0 , T ] , (6) is a local martingale with M 0 = 0 M_{0}=0 M 0 = 0 , for each u ∈ U \mathbf{u}\in\mathcal{U} u ∈ U .
Proof. Fix u ∈ U \mathbf{u}\in\mathcal{U} u ∈ U . By Corollary 3.8, for all N ∈ N N\in\mathbb{N} N ∈ N the process M N : = ( M t N ) t ∈ [ 0 , T ] M^{N}:=(M^{N}_{t})_{t\in[0,T]} M N := ( M t N ) t ∈ [ 0 , T ] given by
M t N : = ⟨ u ≤ N , X t ⟩ − ⟨ u ≤ N , ϵ ⟩ − ∫ 0 t ⟨ L ( u ≤ N ) , X u ⟩ d u , M^{N}_{t}:=\langle\mathbf{u}^{\leq N},X_{t}\rangle-\langle\mathbf{u}^{\leq N},\epsilon\rangle-\int_{0}^{t}\langle L(\mathbf{u}^{\leq N}),X_{u}\rangle\,du, M t N := ⟨ u ≤ N , X t ⟩ − ⟨ u ≤ N , ϵ ⟩ − ∫ 0 t ⟨ L ( u ≤ N ) , X u ⟩ d u , is a local martingale. We show that sup N ∈ N sup s ≤ T ∣ M s ∧ τ n N ∣ \sup_{N\in\mathbb{N}}\sup_{s\leq T}|M^{N}_{s\wedge\tau_{n}}| sup N ∈ N sup s ≤ T ∣ M s ∧ τ n N ∣ is integrable for τ n : = τ n 1 ∧ τ n 2 \tau_{n}:=\tau^{1}_{n}\wedge\tau^{2}_{n} τ n := τ n 1 ∧ τ n 2 , where ( τ n i ) n ∈ N (\tau^{i}_{n})_{n\in\mathbb{N}} ( τ n i ) n ∈ N is a localizing sequence for the locally integrable process V 1 V^{1} V 1 and
τ n 2 : = inf { t > 0 : V t 2 > n } . \tau^{2}_{n}:=\inf\{t>0:V^{2}_{t}>n\}. τ n 2 := inf { t > 0 : V t 2 > n } .
Indeed, this implies that M ∧ τ n N M^{N}_{\wedge\tau_n} M ∧ τ n N is a true martingale for each N ∈ N N \in\mathbb{N} N ∈ N and by the dominated convergence theorem the same holds for the limit process lim N → ∞ M ∧ τ n N \lim_{N\to\infty}M^{N}_{\wedge\tau_n} lim N → ∞ M ∧ τ n N , which is equal to M = ( M t ) t ∈ [ 0 , T ] M = (M_t)_{t\in[0,T]} M = ( M t ) t ∈ [ 0 , T ] given in (6) , as lim N → ∞ ⟨ u ≤ N , X t ⟩ = ⟨ u , X t ⟩ \lim_{N\to\infty}\langle\mathbf{u}^{\leq N},\mathbf{X}_t\rangle= \langle\mathbf{u},\mathbf{X}_t\rangle lim N → ∞ ⟨ u ≤ N , X t ⟩ = ⟨ u , X t ⟩ and
lim N → ∞ ∫ 0 t ∧ τ n ⟨ L ‾ ( u ≤ N ) , X s ⟩ d s = ∫ 0 t ∧ τ n ⟨ L ‾ ( u ) , X s ⟩ d s , \lim_{N\to\infty}\int_0^{t\wedge\tau_n}\langle\overline{L}(\mathbf{u}^{\leq N}),\mathbf{X}_s\rangle\,ds=\int_0^{t\wedge\tau_n}\langle\overline{L}(\mathbf{u}),\mathbf{X}_s\rangle\,ds, N → ∞ lim ∫ 0 t ∧ τ n ⟨ L ( u ≤ N ) , X s ⟩ d s = ∫ 0 t ∧ τ n ⟨ L ( u ) , X s ⟩ d s ,
a.s. To see that sup N ∈ N sup s ≤ T ∣ M s ∧ τ n N ∣ \sup_{N\in\mathbb{N}}\sup_{s\leq T}|M^{N}_{s\wedge\tau_n}| sup N ∈ N sup s ≤ T ∣ M s ∧ τ n N ∣ is integrable it suffices to note that
sup s ≤ T ∣ M s ∧ τ n N ∣ ≤ C + sup s ≤ T ∧ τ n V s 1 + ∫ 0 T ∧ τ n V u 2 d u . \sup_{s\leq T}|M^{N}_{s\wedge\tau_n}|\leq C+\sup_{s\leq T\wedge\tau_n}V_s^1+\int_0^{T\wedge\tau_n}V_u^2\,du. s ≤ T sup ∣ M s ∧ τ n N ∣ ≤ C + s ≤ T ∧ τ n sup V s 1 + ∫ 0 T ∧ τ n V u 2 d u .
Finally, we present the main result of the section. Building on an approach used in Cuchiero et al. [27 ] and Cuchiero et al. [26 ] , and grounded in duality methods for stochastic processes Janssen and Kurt (2014); Etheer and Kurtz (1986), we show that certain expected signature quantities of entire sig-jump-diffusions can be represented as solutions to linear ODEs in the extended tensor algebra.
Theorem 3.13. Let X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be an entire sig-jump-diffusion with sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) , and U ⊆ D ( L ‾ ) \mathcal{U}\subseteq\mathcal{D}(\overline{L}) U ⊆ D ( L ) so that all the conditions of Theorem 3.12 are verified. Fix u ∈ U \mathbf{u}\in\mathcal{U} u ∈ U and assume that there exists a U \mathcal{U} U -valued solution ( c ( t ) ) t ∈ [ 0 , T ] (\mathbf{c}(t))_{t\in[0,T]} ( c ( t ) ) t ∈ [ 0 , T ] of the linear ODE
⟨ c ( t ) , x ⟩ = ⟨ u , x ⟩ + ∫ 0 t ⟨ L ‾ ( c ( s ) ) , x ⟩ d s , t ∈ [ 0 , T ] , (7) \langle\mathbf{c}(t),\mathbf{x}\rangle=\langle\mathbf{u},\mathbf{x}\rangle+\int_0^t\langle\overline{L}(\mathbf{c}(s)),\mathbf{x}\rangle\,ds,\qquad t\in[0,T], \tag*{(7)} ⟨ c ( t ) , x ⟩ = ⟨ u , x ⟩ + ∫ 0 t ⟨ L ( c ( s )) , x ⟩ d s , t ∈ [ 0 , T ] , (7) for all x ∈ S \mathbf{x}\in\mathcal{S} x ∈ S . If furthermore
E [ sup s , t ≤ T ∣ ⟨ c ( s ) , X t ⟩ ∣ ] < ∞ and E [ sup s , t ≤ T ∣ ⟨ L ‾ ( c ( s ) ) , X t ⟩ ∣ ] < ∞ , (8) \mathbb{E}\left[\sup_{s,t\leq T}|\langle\mathbf{c}(s),\mathbf{X}_t\rangle|\right]<\infty
\qquad\text{and}\qquad
\mathbb{E}\left[\sup_{s,t\leq T}|\langle\overline{L}(\mathbf{c}(s)),\mathbf{X}_t\rangle|\right]<\infty, \tag*{(8)} E [ s , t ≤ T sup ∣ ⟨ c ( s ) , X t ⟩ ∣ ] < ∞ and E [ s , t ≤ T sup ∣ ⟨ L ( c ( s )) , X t ⟩ ∣ ] < ∞ , (8) then it holds that
E [ ⟨ u , X T ⟩ ] = ⟨ c ( T ) , ϵ ⟩ . \mathbb{E}[\langle\mathbf{u},\mathbf{X}_T\rangle]=\langle\mathbf{c}(T),\boldsymbol{\epsilon}\rangle. E [⟨ u , X T ⟩] = ⟨ c ( T ) , ϵ ⟩ . Proof. The proof follows by Lemma A.1 in Cuchiero et al. [27 ] .
Remark 3.14. Notice that the conditions in (8) ensure that the process ( M t ) t ∈ [ 0 , T ] (M_t)_{t\in[0,T]} ( M t ) t ∈ [ 0 , T ] introduced in equation (6) is a true martingale for each coefficient c ( s ) \mathbf{c}(s) c ( s ) , s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] .
Combining Corollary 3.9 with the procedure followed in the proof of Theorem 3.12 and Theorem 3.13 we obtain the following result.
Corollary 3.15. Fix g ∈ S g\in\mathcal{S} g ∈ S , some sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) and let X g = ( X t g ) t ∈ [ 0 , T ] X^g=(X_t^g)_{t\in[0,T]} X g = ( X t g ) t ∈ [ 0 , T ] be an entire sig-jump-diffusion with sig-coefficients
b g ( x ) : = b ( g ⊗ x ) , a g ( x ) : = a ( g ⊗ x ) , K g ( x , d ξ ) : = K ( g ⊗ x , d ξ ) . b^g(\mathbf{x}):=b(g\otimes\mathbf{x}),\qquad
a^g(\mathbf{x}):=a(g\otimes\mathbf{x}),\qquad
K^g(\mathbf{x},d\xi):=K(g\otimes\mathbf{x},d\xi). b g ( x ) := b ( g ⊗ x ) , a g ( x ) := a ( g ⊗ x ) , K g ( x , d ξ ) := K ( g ⊗ x , d ξ ) . Fix U ⊆ D ( L ‾ ) \mathcal{U}\subseteq\mathcal{D}(\overline{L}) U ⊆ D ( L ) so that all the conditions of Theorem 3.12 are verified. Fix u ∈ U \mathbf{u}\in\mathcal{U} u ∈ U and assume that there exists a U \mathcal{U} U -valued solution ( c ( t ) ) t ∈ [ 0 , T ] (\mathbf{c}(t))_{t\in[0,T]} ( c ( t ) ) t ∈ [ 0 , T ] of the linear ODE
⟨ c ( t ) , x ⟩ = ⟨ u , x ⟩ + ∫ 0 t ⟨ L ‾ ( c ( s ) ) , x ⟩ d s , t ∈ [ 0 , T ] , \langle\mathbf{c}(t),\mathbf{x}\rangle=\langle\mathbf{u},\mathbf{x}\rangle+\int_0^t\langle\overline{L}(\mathbf{c}(s)),\mathbf{x}\rangle\,ds,\qquad t\in[0,T], ⟨ c ( t ) , x ⟩ = ⟨ u , x ⟩ + ∫ 0 t ⟨ L ( c ( s )) , x ⟩ d s , t ∈ [ 0 , T ] , for all x ∈ S \mathbf{x}\in\mathcal{S} x ∈ S . If furthermore
E [ sup s , t ≤ T ∣ ⟨ c ( s ) , g ⊗ X t g ⟩ ∣ ] < ∞ and E [ sup s , t ≤ T ∣ ⟨ L ‾ ( c ( s ) ) , g ⊗ X t g ⟩ ∣ ] < ∞ , \mathbb{E}\left[\sup_{s,t\leq T}|\langle\mathbf{c}(s),g\otimes\mathbf{X}_t^g\rangle|\right]<\infty
\qquad\text{and}\qquad
\mathbb{E}\left[\sup_{s,t\leq T}|\langle\overline{L}(\mathbf{c}(s)),g\otimes\mathbf{X}_t^g\rangle|\right]<\infty, E [ s , t ≤ T sup ∣ ⟨ c ( s ) , g ⊗ X t g ⟩ ∣ ] < ∞ and E [ s , t ≤ T sup ∣ ⟨ L ( c ( s )) , g ⊗ X t g ⟩ ∣ ] < ∞ , then it holds that E [ ⟨ u , g ⊗ X T g ⟩ ] = ⟨ c ( T ) , g ⟩ \mathbb{E}[\langle\mathbf{u},g\otimes\mathbf{X}_T^g\rangle]=\langle\mathbf{c}(T),g\rangle E [⟨ u , g ⊗ X T g ⟩] = ⟨ c ( T ) , g ⟩ .
Examples
Holomorphic jump-diffusions
In this section we demonstrate that a broad class of holomorphic jump-diffusions, introduced in Cuchiero et al. [26 ] , can be embedded within the framework of entire sig-jump-diffusions. The defining property of this class of processes concerns the extended generator, which is required to map a subset of holomorphic functions to themselves. If this subset includes in particular the set of polynomials on R d \mathbb{R}^d R d , by Lemma 3.3 in Cuchiero et al. [26 ] , one can show that the characteristics ( b ˉ , a ˉ , K ˉ ) (\bar{b},\bar{a},\bar{K}) ( b ˉ , a ˉ , K ˉ ) of such processes with respect to the truncation function χ ( ξ ) = ξ \chi(\xi)=\xi χ ( ξ ) = ξ satisfy
b ˉ i ∈ O ( R d ) , a ˉ i 1 i 2 + m ˉ i 1 i 2 ∈ O ( R d ) , m ˉ i 1 ⋯ i n ∈ O ( R d ) for n ≥ 3 , (9) \bar{b}_i \in\mathcal{O}(\mathbb{R}^d), \qquad\bar{a}_{i_1i_2}+\bar{m}_{i_1i_2} \in\mathcal{O}(\mathbb{R}^d), \qquad\bar{m}_{i_1\cdots i_n} \in\mathcal{O}(\mathbb{R}^d)\text{ for }n\geq3, \tag*{(9)} b ˉ i ∈ O ( R d ) , a ˉ i 1 i 2 + m ˉ i 1 i 2 ∈ O ( R d ) , m ˉ i 1 ⋯ i n ∈ O ( R d ) for n ≥ 3 , (9)
where, for each n ≥ 2 n\geq2 n ≥ 2 ,
m ˉ i 1 ⋯ i n : = ∫ R d ξ i 1 ⋯ ξ i n K ˉ ( ⋅ , d ξ ) , \bar{m}_{i_1\cdots i_n}:=\int_{\mathbb{R}^d}\xi_{i_1}\cdots\xi_{i_n}\,\bar{K}(\cdot,\mathrm{d}\xi), m ˉ i 1 ⋯ i n := ∫ R d ξ i 1 ⋯ ξ i n K ˉ ( ⋅ , d ξ ) ,
and O ( R d ) : = { f : R d → C : f = h ∣ R d for some entire map h on C d } \mathcal{O}(\mathbb{R}^d):=\{f:\mathbb{R}^d\to\mathbb{C}:f=h|_{\mathbb{R}^d}\text{ for some entire map }h\text{ on }\mathbb{C}^d\} O ( R d ) := { f : R d → C : f = h ∣ R d for some entire map h on C d } . Note that the argument of each of these maps is the process itself and not its signature.
To simplify the notation, for a sufficiently differentiable map f : R d → R f:\mathbb{R}^d\to\mathbb{R} f : R d → R we set
f ( ∅ ) ( x ) : = f ( x ) and f ( J ) ( x ) : = ∂ x j 1 ⋯ ∂ x j n f ( x ) , (10) f^{(\varnothing)}(x):=f(x)\qquad\text{and}\qquad f^{(J)}(x):=\partial_{x_{j_1}}\cdots\partial_{x_{j_n}}f(x), \tag*{(10)} f ( ∅ ) ( x ) := f ( x ) and f ( J ) ( x ) := ∂ x j 1 ⋯ ∂ x j n f ( x ) , (10)
for J = ( j 1 , … , j n ) J=(j_1,\ldots,j_n) J = ( j 1 , … , j n ) .
Lemma 3.16. Fix S = R d S=\mathbb{R}^d S = R d and let X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be a holomorphic jump-diffusion such that X − X 0 X-X_0 X − X 0 takes values in S S S and its characteristics ( b ˉ , a ˉ , K ˉ ) (\bar{b},\bar{a},\bar{K}) ( b ˉ , a ˉ , K ˉ ) satisfy the conditions in equation (9) . Then X X X is an entire sig-jump-diffusion whose sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) satisfy the following conditions.
For all x ∈ S x\in S x ∈ S ,
b ( x ) = ⟨ b , x ⟩ b(x)=\langle b,x\rangle b ( x ) = ⟨ b , x ⟩ , for some b ∈ S ∗ d b\in S^{*d} b ∈ S ∗ d given by
b J I = b ˉ i 1 ( J ) ( X 0 ) for each I = ( i 1 ) and multiindex J ; b^I_J=\bar{b}^{(J)}_{i_1}(X_0)\qquad\text{for each }I=(i_1)\text{ and multiindex }J; b J I = b ˉ i 1 ( J ) ( X 0 ) for each I = ( i 1 ) and multiindex J ; a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) = ⟨ a + k 2 , x ⟩ a(x)+\int_{\mathbb{R}^d}\xi^{\otimes2}K(x,\mathrm{d}\xi)=\langle a+k^2,x\rangle a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) = ⟨ a + k 2 , x ⟩ , for some a + k 2 ∈ ( ( S ∗ ) d ) ⊗ 2 a+k^2\in((S^*)^d)^{\otimes2} a + k 2 ∈ (( S ∗ ) d ) ⊗ 2 given by
( a + k 2 ) J I = ( a ˉ i 1 i 2 + m ˉ i 1 i 2 ) ( J ) ( X 0 ) for each I = ( i 1 , i 2 ) and multiindex J ; (a+k^2)^I_J=(\bar{a}_{i_1i_2}+\bar{m}_{i_1i_2})^{(J)}(X_0)\qquad\text{for each }I=(i_1,i_2)\text{ and multiindex }J; ( a + k 2 ) J I = ( a ˉ i 1 i 2 + m ˉ i 1 i 2 ) ( J ) ( X 0 ) for each I = ( i 1 , i 2 ) and multiindex J ; for all n ≥ 3 n\geq3 n ≥ 3 , ∫ R d ξ ⊗ n K ( x , d ξ ) = ⟨ k n , x ⟩ \int_{\mathbb{R}^d}\xi^{\otimes n}K(x,\mathrm{d}\xi)=\langle k^n,x\rangle ∫ R d ξ ⊗ n K ( x , d ξ ) = ⟨ k n , x ⟩ , for some k n ∈ ( ( S ∗ ) d ) ⊗ n k^n\in((S^*)^d)^{\otimes n} k n ∈ (( S ∗ ) d ) ⊗ n given by
( k n ) J I = m ˉ i 1 ⋯ i n ( J ) ( X 0 ) for each I = ( i 1 , … , i n ) and multiindex J . (k^n)^I_J=\bar{m}^{(J)}_{i_1\cdots i_n}(X_0)\qquad\text{for each }I=(i_1,\ldots,i_n)\text{ and multiindex }J. ( k n ) J I = m ˉ i 1 ⋯ i n ( J ) ( X 0 ) for each I = ( i 1 , … , i n ) and multiindex J .
Proof. First of all notice that the maps in O ( R d ) \mathcal{O}(\mathbb{R}^d) O ( R d ) admit a power series expansion centered at any point in R d \mathbb{R}^d R d ([26 Section D] ). Then, the proof follows by performing the Taylor expansion of the maps in (9) centered at X 0 X_0 X 0 , and noticing that by the properties of the shuffle product ([27 Remark 2.8 (iii] ), for each x ∈ S ⊆ G ( ( R d ) ) x\in S\subseteq G((\mathbb{R}^d)) x ∈ S ⊆ G (( R d )) and each multiindex ∣ I ∣ ≥ 1 |I|\geq1 ∣ I ∣ ≥ 1 it holds
∑ J : J ord = I ord ⟨ ϵ J , x ⟩ = 1 I ! ⟨ ϵ i 1 , x ⟩ ⋯ ⟨ ϵ i ∣ I ∣ , x ⟩ , \sum_{J:J^{\operatorname{ord}}=I^{\operatorname{ord}}}\langle\epsilon_J,x\rangle=\frac{1}{I!}\langle\epsilon_{i_1},x\rangle\cdots\langle\epsilon_{i_{|I|}},x\rangle, J : J ord = I ord ∑ ⟨ ϵ J , x ⟩ = I ! 1 ⟨ ϵ i 1 , x ⟩ ⋯ ⟨ ϵ i ∣ I ∣ , x ⟩ ,
where I ord I^{\operatorname{ord}} I ord is the multi-index obtained by sorting the entries of I I I and I ! : = ∏ k = 1 d I ( k ) ! I!:=\prod_{k=1}^d I(k)! I ! := ∏ k = 1 d I ( k )! for I ( k ) : = ∑ j = 1 ∣ I ∣ 1 { i j = k } I(k):=\sum_{j=1}^{|I|}1_{\{i_j=k\}} I ( k ) := ∑ j = 1 ∣ I ∣ 1 { i j = k } .
Remark 3.17. For holomorphic jump-diffusions in the class of entire sig-jump-diffusions, Theorem 3.13 corresponds to Theorem 3.20 in Cuchiero et al. [26 ] whenever for each u ∈ U \mathbf{u} \in\mathcal{U} u ∈ U and each multi-index I I I , it holds u I = u I o r d \mathbf{u}_I = \mathbf{u}_{I^{ord}} u I = u I or d , for I o r d I^{ord} I or d introduced in the proof of Lemma 3.16. In such a case the local martingality of the process M M M in equation (6) follows directly from the definition of holomorphic jump-diffusion (Definition 3.1 in Cuchiero et al. [26 ] ). Moreover, the existence of a solution of the sequence-valued ODE (3.5) can be derived exploiting some tools from complex analysis (Cuchiero et al., [26 ] , Section 3.5).
Entire sig-jump-diffusions with truncated signature-dependent coefficients
Fix two entire maps b ˉ : T N ( R d ) → R d \bar{b}: T^{N}(\mathbb{R}^{d}) \to\mathbb{R}^{d} b ˉ : T N ( R d ) → R d and σ ˉ : T N ( R d ) → R d × d \bar{\sigma}: T^{N}(\mathbb{R}^{d}) \to\mathbb{R}^{d \times d} σ ˉ : T N ( R d ) → R d × d bounded on their real domain. Let ( E , E , F ) (E,\mathcal{E},F) ( E , E , F ) be a finite measure space and let
δ ˉ : T N ( R d ) × E → R d \bar{\delta}: T^{N}(\mathbb{R}^{d}) \times E \to\mathbb{R}^{d} δ ˉ : T N ( R d ) × E → R d
be a bounded measurable map such that δ ˉ ( ⋅ , z ) \bar{\delta}(\cdot,z) δ ˉ ( ⋅ , z ) is entire for every z ∈ E z \in E z ∈ E , with local Lipschitz constants that can be chosen uniformly in z z z . The assumption F ( E ) < ∞ F(E) < \infty F ( E ) < ∞ can be relaxed, but it simplifies the assumptions needed for the proof of Lemma 3.18 later on. Fix g ∈ G N ( R d ) g \in G^{N}(\mathbb{R}^{d}) g ∈ G N ( R d ) and set
b ˉ g ( x ) : = b ˉ ( g ⊗ x ) , σ ˉ g ( x ) : = σ ˉ ( g ⊗ x ) , δ ˉ g ( x , z ) : = δ ˉ ( g ⊗ x , z ) . \bar{b}^{g}(\mathbf{x}) := \bar{b}(g \otimes\mathbf{x}), \qquad\bar{\sigma}^{g}(\mathbf{x}) := \bar{\sigma}(g \otimes\mathbf{x}), \qquad\bar{\delta}^{g}(\mathbf{x},z) := \bar{\delta}(g \otimes\mathbf{x},z). b ˉ g ( x ) := b ˉ ( g ⊗ x ) , σ ˉ g ( x ) := σ ˉ ( g ⊗ x ) , δ ˉ g ( x , z ) := δ ˉ ( g ⊗ x , z ) .
Let μ ( d t , d z ) \mu(dt,dz) μ ( d t , d z ) be a Poisson random measure on R + × E \mathbb{R}_{+} \times E R + × E with compensator d t F ( d z ) dtF(dz) d tF ( d z ) , independent of W W W , and consider the SDE
d X t g = b ˉ g ( X t − g ) d t + σ ˉ g ( X t − g ) d W t + ∫ E δ ˉ g ( X t − g , z ) ( μ ( d t , d z ) − F ( d z ) d t ) , X 0 g = x 0 . (11) dX_{t}^{g}=\bar{b}^{g}(X_{t-}^{g})dt+\bar{\sigma}^{g}(X_{t-}^{g})dW_{t}+\int_{E}\bar{\delta}^{g}(X_{t-}^{g},z)\bigl(\mu(dt,dz)-F(dz)dt\bigr), \qquad X_{0}^{g}=x_{0}. \tag*{(11)} d X t g = b ˉ g ( X t − g ) d t + σ ˉ g ( X t − g ) d W t + ∫ E δ ˉ g ( X t − g , z ) ( μ ( d t , d z ) − F ( d z ) d t ) , X 0 g = x 0 . (11)
where W W W is a d d d -dimensional Brownian motion and X t g \mathbf{X}_{t}^{g} X t g is the truncated signature of X g X^{g} X g . Note that we omit ≤ N \leq N ≤ N for notational convenience.
By Remark 2.4 we know that Δ X t g = δ ( X t − g , Δ X t g ) \Delta\mathbf{X}_{t}^{g}=\delta(\mathbf{X}_{t-}^{g},\Delta X_{t}^{g}) Δ X t g = δ ( X t − g , Δ X t g ) for
δ ( x , ξ ) I : = ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I 1 ∣ I 2 ∣ ! ⟨ ϵ I 1 , x ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ 1 { ∣ I 2 ∣ > 0 } . \delta(\mathbf{x},\xi)_{I}:=\sum_{\substack{\epsilon_{I_{1}}\otimes\epsilon_{I_{2}}=\epsilon_{I}}}\frac{1}{|I_{2}|!}\langle\epsilon_{I_{1}},\mathbf{x}\rangle\langle\epsilon_{I_{2}},\xi^{\otimes|I_{2}|}\rangle\mathbf{1}_{\{|I_{2}|>0\}}. δ ( x , ξ ) I := ϵ I 1 ⊗ ϵ I 2 = ϵ I ∑ ∣ I 2 ∣ ! 1 ⟨ ϵ I 1 , x ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ 1 { ∣ I 2 ∣ > 0 } .
An application of Itô’s formula for semimartingales with jumps shows that, setting
b I g ( x ) : = ⟨ ϵ I ′ , x ⟩ b I ∣ I ∣ g ( x ) + 1 2 ⟨ ϵ I ′ ′ , x ⟩ ( σ ˉ g ( x ) σ ˉ g ( x ) T ) i ∣ I ∣ − 1 i ∣ I ∣ + ∫ E ( δ ( x , δ ˉ g ( x , z ) ) I − ⟨ ϵ I ′ , x ⟩ δ ˉ g ( x , z ) i ∣ I ∣ ) F ( d z ) , b_{I}^{g}(\mathbf{x}):=\langle\epsilon_{I'},\mathbf{x}\rangle b_{I_{|I|}^{g}}(\mathbf{x})+\frac{1}{2}\langle\epsilon_{I''},\mathbf{x}\rangle\bigl(\bar{\sigma}^{g}(\mathbf{x})\bar{\sigma}^{g}(\mathbf{x})^{\mathsf{T}}\bigr)_{i_{|I|-1}i_{|I|}}
+\int_{E}\bigl(\delta(\mathbf{x},\bar{\delta}^{g}(\mathbf{x},z))_{I}-\langle\epsilon_{I'},\mathbf{x}\rangle\bar{\delta}^{g}(\mathbf{x},z)_{i_{|I|}}\bigr)F(dz), b I g ( x ) := ⟨ ϵ I ′ , x ⟩ b I ∣ I ∣ g ( x ) + 2 1 ⟨ ϵ I ′′ , x ⟩ ( σ ˉ g ( x ) σ ˉ g ( x ) T ) i ∣ I ∣ − 1 i ∣ I ∣ + ∫ E ( δ ( x , δ ˉ g ( x , z ) ) I − ⟨ ϵ I ′ , x ⟩ δ ˉ g ( x , z ) i ∣ I ∣ ) F ( d z ) ,
σ I g ( x ) : = ⟨ ϵ I ′ , x ⟩ σ ˉ g ( x ) i ∣ I ∣ , γ g ( x , z ) : = δ ( x , δ ˉ g ( x , z ) ) \sigma_{I}^{g}(\mathbf{x}):=\langle\epsilon_{I'},\mathbf{x}\rangle\bar{\sigma}^{g}(\mathbf{x})_{i_{|I|}}, \qquad\gamma^{g}(\mathbf{x},z):=\delta(\mathbf{x},\bar{\delta}^{g}(\mathbf{x},z)) σ I g ( x ) := ⟨ ϵ I ′ , x ⟩ σ ˉ g ( x ) i ∣ I ∣ , γ g ( x , z ) := δ ( x , δ ˉ g ( x , z ))
the dynamics of the truncated signature satisfy
d X t g = b g ( X t − g ) d t + σ g ( X t − g ) d W t + ∫ E γ g ( X t − g , z ) ( μ ( d t , d z ) − F ( d z ) d t ) , (12) d\mathbf{X}_{t}^{g}=b^{g}(\mathbf{X}_{t-}^{g})dt+\sigma^{g}(\mathbf{X}_{t-}^{g})dW_{t}+\int_{E}\gamma^{g}(\mathbf{X}_{t-}^{g},z)\bigl(\mu(dt,dz)-F(dz)dt\bigr), \tag*{(12)} d X t g = b g ( X t − g ) d t + σ g ( X t − g ) d W t + ∫ E γ g ( X t − g , z ) ( μ ( d t , d z ) − F ( d z ) d t ) , (12)
for X 0 g = ϵ ∅ \mathbf{X}_{0}^{g}=\epsilon_{\varnothing} X 0 g = ϵ ∅ . Setting b : = b ϵ ∅ b:=b^{\epsilon_{\varnothing}} b := b ϵ ∅ , σ : = σ ϵ ∅ \sigma:=\sigma^{\epsilon_{\varnothing}} σ := σ ϵ ∅ and γ : = γ ϵ ∅ \gamma:=\gamma^{\epsilon_{\varnothing}} γ := γ ϵ ∅ and noting that
g ⊗ b g ( x ) = b ( g ⊗ x ) , g ⊗ σ g ( x ) = σ ( g ⊗ x ) , g ⊗ γ g ( x , z ) = γ ( g ⊗ x , z ) , g\otimes b^{g}(\mathbf{x})=b(g\otimes\mathbf{x}), \qquad g\otimes\sigma^{g}(\mathbf{x})=\sigma(g\otimes\mathbf{x}), \qquad g\otimes\gamma^{g}(\mathbf{x},z)=\gamma(g\otimes\mathbf{x},z), g ⊗ b g ( x ) = b ( g ⊗ x ) , g ⊗ σ g ( x ) = σ ( g ⊗ x ) , g ⊗ γ g ( x , z ) = γ ( g ⊗ x , z ) ,
we can conclude that the process ( g ⊗ X t g ) t ∈ [ 0 , T ] (g\otimes\mathbf{X}_{t}^{g})_{t\in[0,T]} ( g ⊗ X t g ) t ∈ [ 0 , T ] is a solution of the SDE
d Z t ≤ N = b ( Z t − ≤ N ) d t + σ ( Z t − ≤ N ) d W t + ∫ E γ ( Z t − ≤ N , z ) ( μ ( d t , d z ) − F ( d z ) d t ) , Z 0 ≤ N = g . (13) d\mathbf{Z}_{t}^{\leq N}=b(\mathbf{Z}_{t-}^{\leq N})dt+\sigma(\mathbf{Z}_{t-}^{\leq N})dW_{t}+\int_{E}\gamma(\mathbf{Z}_{t-}^{\leq N},z)\bigl(\mu(dt,dz)-F(dz)dt\bigr), \qquad\mathbf{Z}_{0}^{\leq N}=g. \tag*{(13)} d Z t ≤ N = b ( Z t − ≤ N ) d t + σ ( Z t − ≤ N ) d W t + ∫ E γ ( Z t − ≤ N , z ) ( μ ( d t , d z ) − F ( d z ) d t ) , Z 0 ≤ N = g . (13)
The finite-dimensional nature of this process gives access to classical existence and uniqueness results. Here the pair ( X g , Z ≤ N ) (X^{g},\mathbf{Z}^{\leq N}) ( X g , Z ≤ N ) is regarded as the unknown of the coupled system (11) –(13) .
Lemma 3.18. The system of SDEs (11) –(13) admits a unique non-explosive strong càdlàg solution on [ 0 , T ] [0,T] [ 0 , T ] .
Proof. We prove the result for g = ϵ ∅ \mathbf{g}=\epsilon_{\varnothing} g = ϵ ∅ ; the general case is analogous. Since b ˉ \bar{b} b ˉ and σ ˉ \bar{\sigma} σ ˉ are entire, and δ ˉ ( ⋅ , z ) \bar{\delta}(\cdot,z) δ ˉ ( ⋅ , z ) is entire with local Lipschitz constants that are uniform in z z z , all these maps are locally Lipschitz in their state variable. Since the coordinate maps x ↦ ⟨ ϵ I , x ⟩ \mathbf{x}\mapsto\langle\epsilon_I,\mathbf{x}\rangle x ↦ ⟨ ϵ I , x ⟩ are linear, it follows that b b b and σ \sigma σ are locally Lipschitz and that, for every R > 0 R>0 R > 0 , there exists a constant L R > 0 L_R>0 L R > 0 such that
∫ E ∥ γ ( x , z ) − γ ( y , z ) ∥ 2 F ( d z ) ≤ L R ∥ x − y ∥ 2 , ∥ x ∥ , ∥ y ∥ ≤ R . \int_E\|\gamma(\mathbf{x},z)-\gamma(\mathbf{y},z)\|^2 F(dz)\leq L_R\|\mathbf{x}-\mathbf{y}\|^2,\qquad\|\mathbf{x}\|,\|\mathbf{y}\|\leq R. ∫ E ∥ γ ( x , z ) − γ ( y , z ) ∥ 2 F ( d z ) ≤ L R ∥ x − y ∥ 2 , ∥ x ∥ , ∥ y ∥ ≤ R .
Moreover, the boundedness of b ˉ \bar{b} b ˉ , σ ˉ \bar{\sigma} σ ˉ , and δ ˉ \bar{\delta} δ ˉ , together with the finiteness of F F F , implies that
∥ b ( x ) ∥ + ∥ σ ( x ) ∥ + ( ∫ E ∥ γ ( x , z ) ∥ 2 F ( d z ) ) 1 / 2 ≤ C ( 1 + ∥ x ∥ ) , x ∈ T N ( R d ) . \|b(\mathbf{x})\|+\|\sigma(\mathbf{x})\|+\left(\int_E\|\gamma(\mathbf{x},z)\|^2F(dz)\right)^{1/2}\leq C(1+\|\mathbf{x}\|),\qquad\mathbf{x}\in T^N(\mathbb{R}^d). ∥ b ( x ) ∥ + ∥ σ ( x ) ∥ + ( ∫ E ∥ γ ( x , z ) ∥ 2 F ( d z ) ) 1/2 ≤ C ( 1 + ∥ x ∥ ) , x ∈ T N ( R d ) .
for some constant C > 0 C>0 C > 0 . Thus, b b b , σ \sigma σ , and γ \gamma γ satisfy the local Lipschitz and linear-growth conditions for stochastic differential equations with jumps. By Jacod and Shiryaev, [46 ] , the system (11) –(13) admits a unique non-explosive strong càdlàg solution on [ 0 , T ] [0,T] [ 0 , T ] .
Remark 3.19.
(i) Note that the first level Z ( 1 ) \mathbb{Z}^{(1)} Z ( 1 ) of (13) is equal to X g \mathbf{X}^{\mathsf{g}} X g if and only if g ( 1 ) = x 0 \mathbf{g}^{(1)}=\mathbf{x}_0 g ( 1 ) = x 0 . This is just a consequence of Z ≤ N = g ⊗ X g \mathbb{Z}^{\leq N}=\mathbf{g}\otimes\mathbf{X}^{\mathsf{g}} Z ≤ N = g ⊗ X g , which at the first level reads as Z ( 1 ) = g ( 1 ) + X g − x 0 \mathbb{Z}^{(1)}=\mathbf{g}^{(1)}+\mathbf{X}^{\mathsf{g}}-\mathbf{x}_0 Z ( 1 ) = g ( 1 ) + X g − x 0 .
(ii) Observe furthermore that the translation invariance of the signature implies that
X g , y 0 = X g , x 0 − x 0 + y 0 , \mathbf{X}^{\mathsf{g},y_0}=\mathbf{X}^{\mathsf{g},x_0}-x_0+y_0, X g , y 0 = X g , x 0 − x 0 + y 0 ,
where X g , x 0 \mathbf{X}^{\mathsf{g},x_0} X g , x 0 denotes the solution of (11) when started at x 0 x_0 x 0 . This in turn means that X g \mathbf{X}^{\mathsf{g}} X g and Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N are independent of the initial value x 0 x_0 x 0 in (11) .
Remark 3.20. As an example, note that for fixed s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] we have
d X s + t = b ˉ ( X ( s + t ) − ≤ N ) d t + σ ˉ ( X ( s + t ) − ≤ N ) d W s + t + ∫ E ( δ ˉ ( X ( s + t ) − ≤ N , z ) ( μ s ( d t , d z ) − F ( d z ) d t ) ) = b ˉ ( X s ≤ N ⊗ X s , s + t − ≤ N ) d t + σ ˉ ( X s ≤ N ⊗ X s , s + t − ≤ N ) d ( W s + t − W s ) \begin{aligned}
d\mathbf{X}_{s+t}=\bar{b}(\mathbf{X}^{\leq N}_{(s+t)-})dt+\bar{\sigma}(\mathbf{X}^{\leq N}_{(s+t)-})dW_{s+t}+\int_E\left(\bar{\delta}(\mathbf{X}^{\leq N}_{(s+t)-},z)(\mu_s(dt,dz)-F(dz)dt)\right) \\
&=\bar{b}(\mathbf{X}^{\leq N}_s\otimes\mathbf{X}^{\leq N}_{s,s+t-})dt+\bar{\sigma}(\mathbf{X}^{\leq N}_s\otimes\mathbf{X}^{\leq N}_{s,s+t-})d(W_{s+t}-W_s)
\end{aligned} d X s + t = b ˉ ( X ( s + t ) − ≤ N ) d t + σ ˉ ( X ( s + t ) − ≤ N ) d W s + t + ∫ E ( δ ˉ ( X ( s + t ) − ≤ N , z ) ( μ s ( d t , d z ) − F ( d z ) d t ) ) = b ˉ ( X s ≤ N ⊗ X s , s + t − ≤ N ) d t + σ ˉ ( X s ≤ N ⊗ X s , s + t − ≤ N ) d ( W s + t − W s )
+ ∫ E δ ˉ ( X s ≤ N ⊗ X s , s + t − ≤ N , z ) ( μ s ( d t , d z ) − F ( d z ) d t ) , \qquad+\int_E\bar{\delta}(\mathbf{X}^{\leq N}_s\otimes\mathbf{X}^{\leq N}_{s,s+t-},z)(\mu_s(dt,dz)-F(dz)dt), + ∫ E δ ˉ ( X s ≤ N ⊗ X s , s + t − ≤ N , z ) ( μ s ( d t , d z ) − F ( d z ) d t ) ,
showing that the solution of (13) for g = X s ≤ N \mathbf{g}=\mathbf{X}^{\leq N}_s g = X s ≤ N is given by X s ≤ N ⊗ X s , s + t ≤ N = X s + t ≤ N \mathbf{X}^{\leq N}_s\otimes\mathbf{X}^{\leq N}_{s,s+t}=\mathbf{X}^{\leq N}_{s+t} X s ≤ N ⊗ X s , s + t ≤ N = X s + t ≤ N .
Remark 3.21. Observe that each component of b ˉ \bar{b} b ˉ , a ˉ g : = σ ˉ g ( σ ˉ g ) T \bar{a}^{\mathsf{g}}:=\bar{\sigma}^{\mathsf{g}}(\bar{\sigma}^{\mathsf{g}})^\mathsf{T} a ˉ g := σ ˉ g ( σ ˉ g ) T , and δ ˉ g ( ⋅ , z ) \bar{\delta}^{\mathsf{g}}(\cdot,z) δ ˉ g ( ⋅ , z ) is an entire map in x ∈ T N ( R d ) \mathbf{x}\in T^N(\mathbb{R}^d) x ∈ T N ( R d ) , for each z ∈ E z\in E z ∈ E . Since these maps depend on finitely many signature coordinates, their power series and the shuffle identity give the required sig-coefficient representations for the drift and covariance. Write then the jump kernel as
K g ( x , B ) : = ∫ E 1 B ∖ { 0 } ( δ ˉ g ( x , z ) ) F ( d z ) . K^{\mathsf{g}}(\mathbf{x},B):=\int_E\mathbf{1}_{B\setminus\{0\}}(\bar{\delta}^{\mathsf{g}}(\mathbf{x},z))F(dz). K g ( x , B ) := ∫ E 1 B ∖ { 0 } ( δ ˉ g ( x , z )) F ( d z ) .
To ensure that ( b ˉ g , a ˉ g , K g ) (\bar{b}^{\mathsf{g}},\bar{a}^{\mathsf{g}},K^{\mathsf{g}}) ( b ˉ g , a ˉ g , K g ) are sig-coefficients and thus (11) is an entire sig-jump-diffusion it remains to obtain sig-coefficient representations for the jump moments
m n g ( x ) : = ∫ E ( δ ˉ g ( x , z ) ) ⊗ n F ( d z ) , n ≥ 2. m_n^{\mathsf{g}}(\mathbf{x}):=\int_E(\bar{\delta}^{\mathsf{g}}(\mathbf{x},z))^{\otimes n}F(dz),\qquad n\geq2. m n g ( x ) := ∫ E ( δ ˉ g ( x , z ) ) ⊗ n F ( d z ) , n ≥ 2.
A sufficient condition is that, for every n ≥ 2 n\geq2 n ≥ 2 and every compact C ⊂ T N ( C d ) C\subset T^N(\mathbb{C}^d) C ⊂ T N ( C d ) ,
∫ E sup x ∈ C ∥ δ ˉ ( x , z ) ∥ n F ( d z ) < ∞ . \int_E\sup_{\mathbf{x}\in C}\|\bar{\delta}(\mathbf{x},z)\|^nF(dz)<\infty. ∫ E x ∈ C sup ∥ δ ˉ ( x , z ) ∥ n F ( d z ) < ∞.
It holds, in particular, if δ ˉ \bar{\delta} δ ˉ is locally bounded on the complex domain uniformly in z ∈ E z \in E z ∈ E , since F ( E ) < ∞ F(E) < \infty F ( E ) < ∞ . Under this condition each m n g m_n^{\mathsf{g}} m n g is entire in Cuchiero et al. [26 ] and the claim follows by the shuffle identity.
Such condition holds, for example, when δ ˉ ( x , z ) = v ( z ) \bar{\delta}(\mathbf{x},z)=v(z) δ ˉ ( x , z ) = v ( z ) for some bounded map v v v . It also holds when δ ˉ ( x , z ) = z + h ( x ) \bar{\delta}(\mathbf{x},z)=z+h(\mathbf{x}) δ ˉ ( x , z ) = z + h ( x ) , F F F has finite moments, and h h h is entire and bounded on its real domain. Under the same condition on δ ˉ \bar{\delta} δ ˉ , the finite dimensional process Z ≤ N = g ⊗ g \mathbb{Z}^{\leq N}=\mathbf{g}\otimes\mathbf{g} Z ≤ N = g ⊗ g is also a holomorphic jump-diffusion in the sense of Definition 3.1 in Cuchiero et al. [26 ] , with the space of test functions chosen as the polynomials on T N ( R d ) T^N(\mathbb{R}^d) T N ( R d ) . Indeed, its drift and covariance have entire components, and the same holds for all its jump moments: the lifted jump size γ ( x , z ) \gamma(\mathbf{x},z) γ ( x , z ) is a polynomial in δ ˉ ( x , z ) \bar{\delta}(\mathbf{x},z) δ ˉ ( x , z ) of degree at most N N N , with coefficients depending linearly on x \mathbf{x} x , so the assumed local integrability condition also applies to its powers. Larger spaces of holomorphic test functions can be considered whenever the conditions of Theorem 3.12 in Cuchiero et al. [26 ] are satisfied.
In particular, provided the corresponding sequence-valued linear ODE admits a solution satisfying the compatibility, martingale, and integrability assumptions of Theorem 3.20 in Cuchiero et al. [26 ] , the holomorphic formula established therein represents expectations of holomorphic functions of Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N in terms of that solution. This formula then coincides with the one provided in Corollary 3.15, by choosing the coefficients u \mathbf{u} u in line with the corresponding holomorphic function.
Polynomial jump-diffusions and their expected signature
We now focus on the class of polynomial jump-diffusions introduced in Cuchiero et al. [22 ] and further analyzed in Filipović and Larsson [33 ] . These processes form a subclass of holomorphic jump-diffusions and naturally fit within the entire sig-jump-diffusion framework. For this class, we explicitly compute the sig-coefficients and show that the infinite-dimensional ODE (7) reduces to a finite-dimensional system of linear ODEs. This reduction allows for efficient and explicit computation of the expected signature via matrix exponentiation.
First, recall that by Lemma 1 in Filipović and Larsson [33 ] a polynomial jump-diffusion is a special jump-diffusion such that the coefficients ( a ˉ , b ˉ , K ˉ ) (\bar{a},\bar{b},\bar{K}) ( a ˉ , b ˉ , K ˉ ) determining its characteristics with respect to the truncation function χ ( ξ ) = ξ \chi(\xi)=\xi χ ( ξ ) = ξ satisfy
b ˉ i ∈ Pol 1 ( R d ) , a ˉ i 1 i 2 + m ˉ i 1 i 2 ∈ Pol 2 ( R d ) , m ˉ i 1 ⋯ i n ∈ Pol n ( R d ) for n ≥ 3 , (14) \bar{b}_i \in\operatorname{Pol}_1(\mathbb{R}^d), \qquad\bar{a}_{i_1i_2}+\bar{m}_{i_1i_2}\in\operatorname{Pol}_2(\mathbb{R}^d), \qquad\bar{m}_{i_1\cdots i_n}\in\operatorname{Pol}_n(\mathbb{R}^d)\ \text{ for } n\geq3, \tag*{(14)} b ˉ i ∈ Pol 1 ( R d ) , a ˉ i 1 i 2 + m ˉ i 1 i 2 ∈ Pol 2 ( R d ) , m ˉ i 1 ⋯ i n ∈ Pol n ( R d ) for n ≥ 3 , (14)
where, for each n ≥ 2 n\geq2 n ≥ 2 ,
m ˉ i 1 ⋯ i n : = ∫ R d ξ i 1 ⋯ ξ i n K ˉ ( ⋅ , d ξ ) , \bar{m}_{i_1\cdots i_n}:=\int_{\mathbb{R}^d}\xi_{i_1}\cdots\xi_{i_n}\bar{K}(\cdot,\mathrm{d}\xi), m ˉ i 1 ⋯ i n := ∫ R d ξ i 1 ⋯ ξ i n K ˉ ( ⋅ , d ξ ) ,
and Pol i ( R d ) \operatorname{Pol}_i(\mathbb{R}^d) Pol i ( R d ) denotes the set of polynomials on R d \mathbb{R}^d R d up to degree i ∈ N i\in\mathbb{N} i ∈ N .
In the following lemma, we adopt the same notation for partial derivatives (equation (10) ) as in Lemma 3.16, and the proof proceeds along similar lines.
Lemma 3.22. A process X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] is a polynomial jump-diffusion such that X − X 0 X-X_0 X − X 0 takes values in S S S if and only if X X X is an entire sig-jump-diffusion whose sig-coefficients ( b , a , K ) (b,a,K) ( b , a , K ) satisfy the following conditions.
For all x ∈ S \mathbf{x}\in S x ∈ S ,
b ( x ) = ⟨ b , x ⟩ b(\mathbf{x})=\langle\mathbf{b},\mathbf{x}\rangle b ( x ) = ⟨ b , x ⟩ , for some b ∈ ( T 1 ( R d ) ) d \mathbf{b}\in(T^1(\mathbb{R}^d))^d b ∈ ( T 1 ( R d ) ) d given by b J I = b i 1 ( J ) ( X 0 ) b^I_J=b^{(J)}_{i_1}(X_0) b J I = b i 1 ( J ) ( X 0 ) for each I = ( i 1 ) I=(i_1) I = ( i 1 ) and ∣ J ∣ ≤ 1 |J|\leq1 ∣ J ∣ ≤ 1 ;
a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) = ⟨ a + k 2 , x ⟩ a(\mathbf{x})+\int_{\mathbb{R}^d}\xi^{\otimes2}K(\mathbf{x},\mathrm{d}\xi)=\langle\mathbf{a}+\mathbf{k}^2,\mathbf{x}\rangle a ( x ) + ∫ R d ξ ⊗ 2 K ( x , d ξ ) = ⟨ a + k 2 , x ⟩ , for some a + k 2 ∈ ( ( T 2 ( R d ) ) d ) ⊗ 2 \mathbf{a}+\mathbf{k}^2\in((T^2(\mathbb{R}^d))^d)^{\otimes2} a + k 2 ∈ (( T 2 ( R d ) ) d ) ⊗ 2 given by ( a + k 2 ) J I = ( a ˉ i 1 i 2 + m ˉ i 1 i 2 ) ( J ) ( X 0 ) (\mathbf{a}+\mathbf{k}^2)^I_J=(\bar{a}_{i_1i_2}+\bar{m}_{i_1i_2})^{(J)}(X_0) ( a + k 2 ) J I = ( a ˉ i 1 i 2 + m ˉ i 1 i 2 ) ( J ) ( X 0 ) for each I = ( i 1 , i 2 ) I=(i_1,i_2) I = ( i 1 , i 2 ) and ∣ J ∣ ≤ 2 |J|\leq2 ∣ J ∣ ≤ 2 ;
for all n ≥ 3 n\geq3 n ≥ 3 , ∫ R d ξ ⊗ n K ( x , d ξ ) = ⟨ k n , x ⟩ \int_{\mathbb{R}^d}\xi^{\otimes n}K(\mathbf{x},\mathrm{d}\xi)=\langle\mathbf{k}^n,\mathbf{x}\rangle ∫ R d ξ ⊗ n K ( x , d ξ ) = ⟨ k n , x ⟩ , for some k n ∈ ( ( T n ( R d ) ) d ) ⊗ n \mathbf{k}^n\in((T^n(\mathbb{R}^d))^d)^{\otimes n} k n ∈ (( T n ( R d ) ) d ) ⊗ n given by ( k n ) J I = m ˉ i 1 ⋯ i n ( J ) ( X 0 ) (\mathbf{k}^n)^I_J=\bar{m}^{(J)}_{i_1\cdots i_n}(X_0) ( k n ) J I = m ˉ i 1 ⋯ i n ( J ) ( X 0 ) for each I = ( i 1 , … , i n ) I=(i_1,\ldots,i_n) I = ( i 1 , … , i n ) and ∣ J ∣ ≤ n |J|\leq n ∣ J ∣ ≤ n .
:::`
Remark 3.23. A polynomial jump-diffusion is fully characterized by its sig-coefficients. This is not true for the broader class of holomorphic jump-diffusions (Lemma 3.16), where expressing the characteristics in terms of sig-coefficients is not sufficient to guarantee the process is a holomorphic jump-diffusion (Cuchiero et al., [26 ] , Lemma 3.3).
Next, we show that for a polynomial jump-diffusion and any initial condition u ∈ T ( R d ) \mathbf{u} \in T(\mathbb{R}^{d}) u ∈ T ( R d ) , the infinite-dimensional ODE ((7) ) reduces to a finite-dimensional system of linear ODEs allowing the explicit computation of expected signature terms via matrix exponentiation.
Proposition 3.24. Let X = ( X t ) t ∈ [ 0 , T ] X = (X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be a polynomial jump-diffusion with coefficients ( b ˉ , a ˉ , K ˉ ) (\bar{b},\bar{a},\bar{K}) ( b ˉ , a ˉ , K ˉ ) such that X − X 0 X-X_0 X − X 0 takes values in S S S . Then, for each N ∈ N N \in\mathbb{N} N ∈ N , the truncated signature X ≤ N X^{\leq N} X ≤ N of X X X is a T N ( R d ) T^N(\mathbb{R}^{d}) T N ( R d ) -valued polynomial jump-diffusion. This in particular implies that all the moments of X ≤ N X^{\leq N} X ≤ N exist and are finite.
Let G G G denote the matrix representation, with respect to a basis h n h_n h n , of the restriction of the infinitesimal generator of X ≤ N X^{\leq N} X ≤ N to the invariant space of polynomials of degree at most n n n . The polynomial property then yields the following conditional moment formula:
E [ p ⃗ T h n ( X T ≤ N ) ∣ F t ] = p ⃗ T exp ( ( T − t ) G T ) h n ( X t ≤ N ) , (15) \mathbb{E}[\vec{p}^{\mathsf{T}}h_n(X_T^{\leq N})\mid\mathcal{F}_t]=\vec{p}^{\mathsf{T}}\exp((T-t)G^{\mathsf{T}})h_n(X_t^{\leq N}), \tag*{(15)} E [ p T h n ( X T ≤ N ) ∣ F t ] = p T exp (( T − t ) G T ) h n ( X t ≤ N ) , (15)
where exp ( ( T − t ) G ) \exp((T-t)G) exp (( T − t ) G ) denotes the matrix exponential.
Proof. The proof of the polynomial property of X ≤ N X^{\leq N} X ≤ N is similar to the proof of Proposition 4.5 in Cuchiero et al. [25 ] . In particular, we get that X ≤ N X^{\leq N} X ≤ N is a jump-diffusion whose coefficients
( β , ζ , K ) : = ( ( β I ) ∣ I ∣ ≤ N , ( ζ I , J ) ∣ I ∣ , ∣ J ∣ ≤ N , K ) (\beta,\zeta,\mathcal{K}) := ((\beta_I)_{|I|\leq N},(\zeta_{I,J})_{|I|,|J|\leq N},\mathcal{K}) ( β , ζ , K ) := (( β I ) ∣ I ∣ ≤ N , ( ζ I , J ) ∣ I ∣ , ∣ J ∣ ≤ N , K )
satisfy
β I ( ⋅ ) = ⟨ ϵ I ′ , ⋅ ⟩ b ˉ i ∣ I ∣ ( π 1 X 0 ( ⋅ ) ) + 1 2 ⟨ ϵ I ′ ′ , ⋅ ⟩ ( a ˉ ( i ∣ I ∣ − 1 , i ∣ I ∣ ) + m ˉ ( i ∣ I ∣ − 1 , i ∣ I ∣ ) ) ( π 1 X 0 ( ⋅ ) ) + ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I 1 { ∣ I 2 ∣ > 2 } 1 ∣ I 2 ∣ ! ⟨ ϵ I 1 , ⋅ ⟩ m ˉ I 2 ( π 1 X 0 ( ⋅ ) ) , \beta_I(\cdot)=\langle\epsilon_{I'},\cdot\rangle\bar{b}_{i_{|I|}}\left(\pi_1^{X_0}(\cdot)\right)+\frac{1}{2}\langle\epsilon_{I''},\cdot\rangle\left(\bar{a}_{(i_{|I|-1},i_{|I|})}+\bar{m}_{(i_{|I|-1},i_{|I|})}\right)\left(\pi_1^{X_0}(\cdot)\right)
+\sum_{\epsilon_{I_1}\otimes\epsilon_{I_2}=\epsilon_I}\mathbf{1}_{\{|I_2|>2\}}\frac{1}{|I_2|!}\langle\epsilon_{I_1},\cdot\rangle\bar{m}_{I_2}\left(\pi_1^{X_0}(\cdot)\right), β I ( ⋅ ) = ⟨ ϵ I ′ , ⋅ ⟩ b ˉ i ∣ I ∣ ( π 1 X 0 ( ⋅ ) ) + 2 1 ⟨ ϵ I ′′ , ⋅ ⟩ ( a ˉ ( i ∣ I ∣ − 1 , i ∣ I ∣ ) + m ˉ ( i ∣ I ∣ − 1 , i ∣ I ∣ ) ) ( π 1 X 0 ( ⋅ ) ) + ϵ I 1 ⊗ ϵ I 2 = ϵ I ∑ 1 { ∣ I 2 ∣ > 2 } ∣ I 2 ∣ ! 1 ⟨ ϵ I 1 , ⋅ ⟩ m ˉ I 2 ( π 1 X 0 ( ⋅ ) ) ,
K ( ⋅ , A ) = ∫ R d 1 A ∖ { 0 } ( δ ( ⋅ , ξ ) ) K ˉ ( π 1 X 0 ( ⋅ ) , d ξ ) , \mathcal{K}(\cdot,A)=\int_{\mathbb{R}^{d}}\mathbf{1}_{A\setminus\{0\}}(\delta(\cdot,\xi))\bar{K}\left(\pi_1^{X_0}(\cdot),d\xi\right), K ( ⋅ , A ) = ∫ R d 1 A ∖ { 0 } ( δ ( ⋅ , ξ )) K ˉ ( π 1 X 0 ( ⋅ ) , d ξ ) ,
ζ I , J ( ⋅ ) = ⟨ ϵ I ′ , ⋅ ⟩ ⟨ ϵ J ′ , ⋅ ⟩ a ˉ ( i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( ⋅ ) ) , \zeta_{I,J}(\cdot)=\langle\epsilon_{I'},\cdot\rangle\langle\epsilon_{J'},\cdot\rangle\bar{a}_{(i_{|I|},j_{|J|})}\left(\pi_1^{X_0}(\cdot)\right), ζ I , J ( ⋅ ) = ⟨ ϵ I ′ , ⋅ ⟩ ⟨ ϵ J ′ , ⋅ ⟩ a ˉ ( i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( ⋅ ) ) ,
where π 1 X 0 \pi_1^{X_0} π 1 X 0 satisfies π 1 X 0 ( X t ) = X t \pi_1^{X_0}(X_t)=X_t π 1 X 0 ( X t ) = X t , and δ : T N ( R d ) × R d → T N ( R d ) \delta:T^N(\mathbb{R}^{d})\times\mathbb{R}^{d}\to T^N(\mathbb{R}^{d}) δ : T N ( R d ) × R d → T N ( R d ) is given by
δ ( x , ξ ) : = ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I 1 ∣ I 2 ∣ ! ⟨ ϵ I 1 , x ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ 1 { ∣ I 2 ∣ > 0 } . \delta(\mathbf{x},\xi):=\sum_{\epsilon_{I_1}\otimes\epsilon_{I_2}=\epsilon_I}\frac{1}{|I_2|!}\langle\epsilon_{I_1},\mathbf{x}\rangle\langle\epsilon_{I_2},\xi^{\otimes|I_2|}\rangle\mathbf{1}_{\{|I_2|>0\}}. δ ( x , ξ ) := ϵ I 1 ⊗ ϵ I 2 = ϵ I ∑ ∣ I 2 ∣ ! 1 ⟨ ϵ I 1 , x ⟩ ⟨ ϵ I 2 , ξ ⊗ ∣ I 2 ∣ ⟩ 1 { ∣ I 2 ∣ > 0 } .
By Lemma 1 in Filipović and Larsson [33 ] it suffices to check that
β I ( ⋅ ) ∈ Pol 1 ( T N ( R d ) ) , \beta_I(\cdot)\in\operatorname{Pol}_1(T^N(\mathbb{R}^{d})), β I ( ⋅ ) ∈ Pol 1 ( T N ( R d )) ,
ζ I , J ( ⋅ ) + ∫ T N ( R d ) y I y J K ( ⋅ , d y ) ∈ Pol 2 ( T N ( R d ) ) , \zeta_{I,J}(\cdot)+\int_{T^N(\mathbb{R}^{d})}y_Iy_J\mathcal{K}(\cdot,dy)\in\operatorname{Pol}_2(T^N(\mathbb{R}^{d})), ζ I , J ( ⋅ ) + ∫ T N ( R d ) y I y J K ( ⋅ , d y ) ∈ Pol 2 ( T N ( R d )) ,
∫ T N ( R d ) y I 1 ⋯ y I n K ( ⋅ , d y ) ∈ Pol n ( T N ( R d ) ) , for n ≥ 3 , \int_{T^N(\mathbb{R}^{d})}y_{I_1}\cdots y_{I_n}\mathcal{K}(\cdot,dy)\in\operatorname{Pol}_n(T^N(\mathbb{R}^{d})),\qquad\text{for }n\geq3, ∫ T N ( R d ) y I 1 ⋯ y I n K ( ⋅ , d y ) ∈ Pol n ( T N ( R d )) , for n ≥ 3 ,
where Pol i ( R d ) \operatorname{Pol}_i(\mathbb{R}^{d}) Pol i ( R d ) denotes the set of polynomials on R d \mathbb{R}^{d} R d up to degree i ∈ N i\in\mathbb{N} i ∈ N . We now show how to conclude the proof of the second condition; the other conditions can be shown analogously. Observe that
L I , J ( ⋅ ) + ∫ T N ( R d ) y I y J K ( ⋅ , d y ) = ⟨ ϵ I ′ , ⋄ ϵ J ′ , ⟩ a ‾ ( i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( x ) ) + ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I ϵ J 1 ⊗ ϵ J 2 = ϵ J 1 ∣ I 2 ∣ ! ∣ J 2 ∣ ! ⟨ ϵ I 1 , x N ⟩ ⟨ ϵ J 1 , x N ⟩ 1 { ∣ I 2 ∣ > 0 } 1 { ∣ J 2 ∣ > 0 } m ‾ I 2 J 2 ( π 1 X 0 ( ⋅ ) ) = ⟨ ϵ I ′ , ⋄ ϵ J ′ , ⟩ ( a ‾ ( i ∣ I ∣ , j ∣ J ∣ ) + m ‾ i ∣ I ∣ j ∣ J ∣ ) ( π 1 X 0 ( x ) ) + ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I ϵ J 1 ⊗ ϵ J 2 = ϵ J 1 ∣ I 2 ∣ ! ∣ J 2 ∣ ! ⟨ ϵ I 1 , x N ⟩ ⟨ ϵ J 1 , x N ⟩ 1 { ∣ I 1 ∣ + ∣ J 2 ∣ > 2 } 1 { ∣ I 2 ∣ > 0 } 1 { ∣ J 2 ∣ > 0 } m ‾ I 2 J 2 ( π 1 X 0 ( ⋅ ) ) . \begin{aligned}
\mathcal{L}_{I,J}(\cdot)+\int_{T^N(\mathbb{R}^d)}y_Iy_JK(\cdot,\mathrm{d}y)=\langle\epsilon_{I'},\diamond\epsilon_{J'},\rangle\overline{a}_{(i_{|I|},j_{|J|})}(\pi_1^{X_0}(\mathbf{x})) \\
&+\sum_{\substack{\epsilon_{I_1}\otimes\epsilon_{I_2}=\epsilon_I\\ \epsilon_{J_1}\otimes\epsilon_{J_2}=\epsilon_J}}\frac{1}{|I_2|!\,|J_2|!}\langle\epsilon_{I_1},\mathbf{x}^N\rangle\langle\epsilon_{J_1},\mathbf{x}^N\rangle\mathbf{1}_{\{|I_2|>0\}}\mathbf{1}_{\{|J_2|>0\}}\overline{m}_{I_2J_2}(\pi_1^{X_0}(\cdot)) \\
&=\langle\epsilon_{I'},\diamond\epsilon_{J'},\rangle\left(\overline{a}_{(i_{|I|},j_{|J|})}+\overline{m}_{i_{|I|}j_{|J|}}\right)(\pi_1^{X_0}(\mathbf{x})) \\
&+\sum_{\substack{\epsilon_{I_1}\otimes\epsilon_{I_2}=\epsilon_I\\ \epsilon_{J_1}\otimes\epsilon_{J_2}=\epsilon_J}}\frac{1}{|I_2|!\,|J_2|!}\langle\epsilon_{I_1},\mathbf{x}^N\rangle\langle\epsilon_{J_1},\mathbf{x}^N\rangle\mathbf{1}_{\{|I_1|+|J_2|>2\}}\mathbf{1}_{\{|I_2|>0\}}\mathbf{1}_{\{|J_2|>0\}}\overline{m}_{I_2J_2}(\pi_1^{X_0}(\cdot)).
\end{aligned} L I , J ( ⋅ ) + ∫ T N ( R d ) y I y J K ( ⋅ , d y ) = ⟨ ϵ I ′ , ⋄ ϵ J ′ , ⟩ a ( i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( x )) + ϵ I 1 ⊗ ϵ I 2 = ϵ I ϵ J 1 ⊗ ϵ J 2 = ϵ J ∑ ∣ I 2 ∣ ! ∣ J 2 ∣ ! 1 ⟨ ϵ I 1 , x N ⟩ ⟨ ϵ J 1 , x N ⟩ 1 { ∣ I 2 ∣ > 0 } 1 { ∣ J 2 ∣ > 0 } m I 2 J 2 ( π 1 X 0 ( ⋅ )) = ⟨ ϵ I ′ , ⋄ ϵ J ′ , ⟩ ( a ( i ∣ I ∣ , j ∣ J ∣ ) + m i ∣ I ∣ j ∣ J ∣ ) ( π 1 X 0 ( x )) + ϵ I 1 ⊗ ϵ I 2 = ϵ I ϵ J 1 ⊗ ϵ J 2 = ϵ J ∑ ∣ I 2 ∣ ! ∣ J 2 ∣ ! 1 ⟨ ϵ I 1 , x N ⟩ ⟨ ϵ J 1 , x N ⟩ 1 { ∣ I 1 ∣ + ∣ J 2 ∣ > 2 } 1 { ∣ I 2 ∣ > 0 } 1 { ∣ J 2 ∣ > 0 } m I 2 J 2 ( π 1 X 0 ( ⋅ )) .
This implies that for ∣ I ∣ , ∣ J ∣ ≤ N |I|,|J|\leq N ∣ I ∣ , ∣ J ∣ ≤ N , setting
( a ‾ i ∣ I ∣ , j ∣ J ∣ + m ‾ i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( ⋅ ) ) = ∑ ∣ H 1 ∣ , ∣ H 2 ∣ ≤ 1 α H 1 H 2 ⟨ ϵ H 1 , ⋄ ϵ H 2 , ⟩ , \left(\overline{a}_{i_{|I|},j_{|J|}}+\overline{m}_{i_{|I|},j_{|J|}}\right)(\pi_1^{X_0}(\cdot))=\sum_{|H_1|,|H_2|\leq1}\alpha_{H_1H_2}\langle\epsilon_{H_1},\diamond\epsilon_{H_2},\rangle, ( a i ∣ I ∣ , j ∣ J ∣ + m i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( ⋅ )) = ∣ H 1 ∣ , ∣ H 2 ∣ ≤ 1 ∑ α H 1 H 2 ⟨ ϵ H 1 , ⋄ ϵ H 2 , ⟩ ,
we get
⟨ ϵ I ′ , ⋄ ⟨ ϵ J ′ , ⋄ ( a ‾ i ∣ I ∣ , j ∣ J ∣ + m ‾ i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( ⋅ ) ) ⟩ = ∑ ∣ H 1 ∣ , ∣ H 2 ∣ ≤ 1 α H 1 H 2 ⟨ ϵ I 1 □ ϵ H 1 , ⋄ ϵ J ′ □ ϵ H 2 , ⋄ ⟩ ∈ Pol 2 ( T N ( R d ) ) . \begin{aligned}
\langle\epsilon_{I'},\diamond\langle\epsilon_{J'},\diamond\left(\overline{a}_{i_{|I|},j_{|J|}}+\overline{m}_{i_{|I|},j_{|J|}}\right)(\pi_1^{X_0}(\cdot))\rangle\\
&=\sum_{|H_1|,|H_2|\leq1}\alpha_{H_1H_2}\langle\epsilon_{I_1}\mathbin{\Box}\epsilon_{H_1},\diamond\epsilon_{J'}\mathbin{\Box}\epsilon_{H_2},\diamond\rangle\in\operatorname{Pol}_2(T^N(\mathbb{R}^d)).
\end{aligned} ⟨ ϵ I ′ , ⋄ ⟨ ϵ J ′ , ⋄ ( a i ∣ I ∣ , j ∣ J ∣ + m i ∣ I ∣ , j ∣ J ∣ ) ( π 1 X 0 ( ⋅ ))⟩ = ∣ H 1 ∣ , ∣ H 2 ∣ ≤ 1 ∑ α H 1 H 2 ⟨ ϵ I 1 □ ϵ H 1 , ⋄ ϵ J ′ □ ϵ H 2 , ⋄ ⟩ ∈ Pol 2 ( T N ( R d )) .
Similarly, exploiting the polynomial property of m ‾ i 1 ⋯ i n \overline{m}_{i_1\cdots i_n} m i 1 ⋯ i n , for n ≥ 3 n\geq3 n ≥ 3 the claim follows. Finally, the representation in equation (15) follows then by an application of Theorem 2.5 in Filipović and Larsson [33 ] .
As a corollary of Proposition 3.24 , the expected signature of a polynomial process, which corresponds to a polynomial of degree 1 on the truncated signature, can be computed analytically.
Corollary 3.25. Let X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] be a polynomial jump-diffusion with coefficients ( b , a ‾ , K ‾ ) (b,\overline{a},\overline{K}) ( b , a , K ) such that X − X 0 X-X_0 X − X 0 takes values in S S S . In the special case of the invariant space given by linear polynomials one can choose the basis { ⟨ ϵ I , ⋅ ⟩ : ∣ I ∣ ≤ N } \{\langle\epsilon_I,\cdot\rangle:|I|\leq N\} {⟨ ϵ I , ⋅ ⟩ : ∣ I ∣ ≤ N } and we get G = L G=L G = L for L L L as in Corollary 3.8 leading to
E [ X T ≤ N ∣ F t ] = exp ( ( T − t ) L T ) X t ≤ N . \mathbb{E}[\mathbf{X}_T^{\leq N}\mid\mathcal{F}_t]=\exp((T-t)L^\mathsf{T})\mathbf{X}_t^{\leq N}. E [ X T ≤ N ∣ F t ] = exp (( T − t ) L T ) X t ≤ N .
Note in particular that the column of L L L corresponding to I I I has the representation
L ϵ I = ϵ I ′ □ b ( i ∣ I ∣ ) + 1 2 ϵ I ′ ′ □ ( a + k 2 ) ( i ∣ I ∣ − 1 , i ∣ I ∣ ) + ∑ ϵ I 1 ⊗ ϵ I 2 = ϵ I 1 { ∣ I 2 ∣ > 2 } 1 ∣ I 2 ∣ ! ϵ I 1 □ ( k ∣ I 2 ∣ ) I 2 , (16) L_{\epsilon_I}=\epsilon_{I'}\mathbin{\Box}\mathbf{b}^{(i_{|I|})}+\frac{1}{2}\epsilon_{I''}\mathbin{\Box}\left(\mathbf{a}+\mathbf{k}^2\right)^{(i_{|I|-1},i_{|I|})}+\sum_{\substack{\epsilon_{I_1}\otimes\epsilon_{I_2}=\epsilon_I}}\mathbf{1}_{\{|I_2|>2\}}\frac{1}{|I_2|!}\epsilon_{I_1}\mathbin{\Box}\left(\mathbf{k}^{|I_2|}\right)_{I_2}, \tag*{(16)} L ϵ I = ϵ I ′ □ b ( i ∣ I ∣ ) + 2 1 ϵ I ′′ □ ( a + k 2 ) ( i ∣ I ∣ − 1 , i ∣ I ∣ ) + ϵ I 1 ⊗ ϵ I 2 = ϵ I ∑ 1 { ∣ I 2 ∣ > 2 } ∣ I 2 ∣ ! 1 ϵ I 1 □ ( k ∣ I 2 ∣ ) I 2 , (16)
for b , a , k \mathbf{b},\mathbf{a},\mathbf{k} b , a , k computed in Lemma 3.22 .
Remark 3.26. Let us remark that Proposition 3.24 can be extended beyond the polynomial jump-diffusion setting. To make this precise, let us introduce the following set of graded polynomials. For N ∈ N N\in\mathbb{N} N ∈ N we define the (ordered) tuples of multi-indices with individual length between 1 and N N N and combined length at most N N N as
I_N=\left\{(I_1,\ldots,I_n)\mathrel{\left|}n\in\mathbb{N},1\leq|I_1|,\ldots,|I_n|\leq N,\sum_{j=1}^{n}|I_j|\leq N\right.\right\}.
Graded polynomials are then given as follows:
P N ( R d ) = { T N ( R d ) ∋ x ↦ c + ∑ J ∈ I N c J ∏ I ∈ J x I ∣ c , c J ∈ R , J ∈ I N } . \mathcal{P}_N(\mathbb{R}^d)=\left\{T^N(\mathbb{R}^d)\ni\mathbf{x}\mapsto c+\sum_{\mathcal{J}\in I_N}c_{\mathcal{J}}\prod_{I\in\mathcal{J}}x_I\mid c,c_{\mathcal{J}}\in\mathbb{R},\mathcal{J}\in I_N\right\}. P N ( R d ) = { T N ( R d ) ∋ x ↦ c + J ∈ I N ∑ c J I ∈ J ∏ x I ∣ c , c J ∈ R , J ∈ I N } .
For N = 1 N=1 N = 1 , P 1 ( R d ) \mathcal{P}_1(\mathbb{R}^d) P 1 ( R d ) coincides with Pol 1 ( R d ) \operatorname{Pol}_1(\mathbb{R}^d) Pol 1 ( R d ) , but for N ≥ 2 N\geq2 N ≥ 2 the notions are different. For simplicity, let d = 2 d=2 d = 2 . Then, for N = 2 N=2 N = 2 , P 2 \mathcal{P}_2 P 2 is spanned by { 1 , x 1 , x 2 , x 11 , x 12 , x 21 , x 22 , x 1 2 , x 1 x 2 , x 2 2 } \{1,x_1,x_2,x_{11},x_{12},x_{21},x_{22},x_1^2,x_1x_2,x_2^2\} { 1 , x 1 , x 2 , x 11 , x 12 , x 21 , x 22 , x 1 2 , x 1 x 2 , x 2 2 } which clearly differs from Pol ( R d ) \operatorname{Pol}(\mathbb{R}^{d}) Pol ( R d ) and also from the set of polynomials with degree at most two in the variables x 1 , x 2 , x 11 , x 12 , x 21 x_{1},x_{2},x_{11},x_{12},x_{21} x 1 , x 2 , x 11 , x 12 , x 21 and x 22 x_{22} x 22 , as for example the term x 11 2 x_{11}^{2} x 11 2 is missing.
If the characteristics of a special semimartingale X X X with respect to the truncation function χ ( ξ ) = ξ \chi(\xi)=\xi χ ( ξ ) = ξ now satisfy
b ˉ i ∈ P 1 ( R d ) = Pol 1 ( R d ) , a ˉ i 1 i 2 + m ˉ i 1 i 2 ∈ P 2 ( R d ) , and m ˉ i 1 ⋯ i n ∈ P n ( R d ) for n ≥ 3 , (17) \bar{b}_{i}\in\mathcal{P}_{1}(\mathbb{R}^{d})=\operatorname{Pol}_{1}(\mathbb{R}^{d}),\quad\bar{a}_{i_{1}i_{2}}+\bar{m}_{i_{1}i_{2}}\in\mathcal{P}_{2}(\mathbb{R}^{d}),\quad\text{and}\quad\bar{m}_{i_{1}\cdots i_{n}}\in\mathcal{P}_{n}(\mathbb{R}^{d})\text{ for }n\geq3, \tag*{(17)} b ˉ i ∈ P 1 ( R d ) = Pol 1 ( R d ) , a ˉ i 1 i 2 + m ˉ i 1 i 2 ∈ P 2 ( R d ) , and m ˉ i 1 ⋯ i n ∈ P n ( R d ) for n ≥ 3 , (17)
then we clearly stay in the entire sig-jump-diffusion setup, but go beyond the polynomial jump-diffusion case. Nevertheless, a conclusion similar to that of Proposition 3.24 remains true. That is, for each N ∈ N N\in\mathbb{N} N ∈ N , the truncated signature X ≤ N \mathbf{X}^{\leq N} X ≤ N of X X X is a T N ( R d ) T^{N}(\mathbb{R}^{d}) T N ( R d ) -valued graded polynomial jump-diffusion. This means that there is a conditional moment formula for every graded polynomial in P N ( R d ) \mathcal{P}_{N}(\mathbb{R}^{d}) P N ( R d ) similar to (15) , with G G G being the matrix representation, with respect to a basis h n h_{n} h n , of the restriction of the infinitesimal generator of X ≤ N \mathbf{X}^{\leq N} X ≤ N to the invariant space P N ( R d ) \mathcal{P}_{N}(\mathbb{R}^{d}) P N ( R d ) . The proof proceeds as in Proposition 3.24 , but for the conclusion of Lemma 2.4 in Filipović and Larsson [33 ] , the true martingale property of N f N^{f} N f (needed to apply Theorem 2.5 in Filipović and Larsson [33 ] ), certain estimates based on a graded norm have to be used. We refer to Lemma 3.19, Lemma 3.20 and Theorem 3.21 in Cuchiero and Pabschschwöll (2026), where the details in the pure diffusion case are carried out.
Remark 3.27. For a concrete specification of characteristics of the form (17) in the pure diffusion case with the second characteristic in P 2 ( R d ) ∖ Pol 2 ( R d ) \mathcal{P}_{2}(\mathbb{R}^{d})\setminus\operatorname{Pol}_{2}(\mathbb{R}^{d}) P 2 ( R d ) ∖ Pol 2 ( R d ) we refer to Example 3.10 in Cuchiero and Pabschschwöll (2026). This pure diffusion example can be extended by adding jumps of a certain form. Indeed, consider the case of a jump kernel explicitly given via
K ( ⋅ , A ) = λ ( ⋅ ) ∫ E 1 A ∖ { 0 } ( j ( ⋅ , y ) ) F ( d y ) , K(\cdot,A)=\lambda(\cdot)\int_{E}\mathbf{1}_{A\setminus\{0\}}(j(\cdot,y))F(\mathrm{d}y), K ( ⋅ , A ) = λ ( ⋅ ) ∫ E 1 A ∖ { 0 } ( j ( ⋅ , y )) F ( d y ) ,
where F F F is a non-negative measure on a measurable space E E E , the jump size map j j j is given by j ( ⋅ , y ) : = ( ⟨ j 1 ( y ) , ⋅ ⟩ , … , ⟨ j d ( y ) , ⋅ ⟩ ) j(\cdot,y):=(\langle j_{1}(y),\cdot\rangle,\ldots,\langle j_{d}(y),\cdot\rangle) j ( ⋅ , y ) := (⟨ j 1 ( y ) , ⋅ ⟩ , … , ⟨ j d ( y ) , ⋅ ⟩) , for some suitable j ( y ) : = ( j 1 ( y ) , … , j d ( y ) ) \mathbf{j}(y):=(j_{1}(y),\ldots,j_{d}(y)) j ( y ) := ( j 1 ( y ) , … , j d ( y )) belonging to ( S ∗ ) d (S^{*})^{d} ( S ∗ ) d for all y ∈ E y\in E y ∈ E . One can show that if either of the following conditions, (i) or (ii), is additionally satisfied, (i) λ ( ⋅ ) : = ⟨ λ , ⋅ ⟩ \lambda(\cdot):=\langle\lambda,\cdot\rangle λ ( ⋅ ) := ⟨ λ , ⋅ ⟩ , for some λ ∈ S ∗ \lambda\in S^{*} λ ∈ S ∗ such that ⟨ λ , x ⟩ > 0 \langle\lambda,x\rangle>0 ⟨ λ , x ⟩ > 0 for all x ∈ S x\in S x ∈ S and j ( ⋅ , y ) = j ( y ) j(\cdot,y)=j(y) j ( ⋅ , y ) = j ( y ) ; (ii) λ ( ⋅ ) : = ⟨ λ ~ , ⋅ ⟩ ⟨ j ∗ , ⋅ ⟩ 2 ∨ ⟨ j ∗ , ⋅ ⟩ ≠ 0 \lambda(\cdot):=\frac{\langle\tilde{\lambda},\cdot\rangle}{\langle j^{*},\cdot\rangle^{2}\vee\langle j^{*},\cdot\rangle\neq0} λ ( ⋅ ) := ⟨ j ∗ , ⋅ ⟩ 2 ∨ ⟨ j ∗ , ⋅ ⟩ = 0 ⟨ λ ~ , ⋅ ⟩ , where λ ~ ∈ T 2 ( R d ) \tilde{\lambda}\in T^{2}(\mathbb{R}^{d}) λ ~ ∈ T 2 ( R d ) , j ∗ ∈ T 1 ( R d ) j^{*}\in T^{1}(\mathbb{R}^{d}) j ∗ ∈ T 1 ( R d ) , and j i ( y ) = H ( y ) j ∗ j_{i}(y)=H(y)j^{*} j i ( y ) = H ( y ) j ∗ for all i = 1 , … , d i=1,\ldots,d i = 1 , … , d and y ∈ E y\in E y ∈ E , and the drift and diffusion coefficients b , a b,a b , a are polynomials on R d \mathbb{R}^{d} R d of degree one and two, respectively, then under suitable integrability conditions on K K K , the truncated signature process of the jump diffusion corresponding to such sig-coefficients is a polynomial jump-diffusion on the truncated tensor algebra. Note that the parametrizations (i) and (ii) correspond to a path-dependent intensity, meaning that the resulting processes go beyond the classical framework of polynomial processes.
Finally, we give an example in which we implement the formula in Corollary 3.25. The corresponding code is available at https://github.com/fprimavera/expected_signature_polyjumpdiffusions.git .
Example 3.28. We consider a 3-dimensional process X ^ t = ( t , X t ( 1 ) , X t ( 2 ) ) \widehat{X}_{t}=(t,X_{t}^{(1)},X_{t}^{(2)}) X t = ( t , X t ( 1 ) , X t ( 2 ) ) , for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , whose last two coordinates follow Ornstein–Uhlenbeck dynamics driven by independent Brownian motions and a common Poisson jump clock:
d X t ( i ) = ( α i + β i X t − ( i ) ) d t + σ i d W t ( i ) + d J t ( i ) , i = 1 , 2 , X 0 = ( c 1 , c 2 ) ∈ R 2 , \mathrm{d}X_{t}^{(i)}=(\alpha_{i}+\beta_{i}X_{t-}^{(i)})\,\mathrm{d}t+\sigma_{i}\,\mathrm{d}W_{t}^{(i)}+\mathrm{d}J_{t}^{(i)},\qquad i=1,2,\qquad X_{0}=(c_{1},c_{2})\in\mathbb{R}^{2}, d X t ( i ) = ( α i + β i X t − ( i ) ) d t + σ i d W t ( i ) + d J t ( i ) , i = 1 , 2 , X 0 = ( c 1 , c 2 ) ∈ R 2 , with W ( i ) W^{(i)} W ( i ) independent 1-dimensional Brownian motions and jumps J t ( i ) = ∑ k = 1 N t t Z k ( i ) J_{t}^{(i)}=\sum_{k=1}^{N_{t}^{t}}Z_{k}^{(i)} J t ( i ) = ∑ k = 1 N t t Z k ( i ) , N t ∼ Poisson ( λ t ) N_{t}\sim\operatorname{Poisson}(\lambda t) N t ∼ Poisson ( λ t ) . The jump sizes Z k = ( Z k ( 1 ) , Z k ( 2 ) ) Z_{k}=(Z_{k}^{(1)},Z_{k}^{(2)}) Z k = ( Z k ( 1 ) , Z k ( 2 ) ) are drawn from the 2-dimensional Gaussian distribution F = N ( μ , Σ ) F=\mathcal{N}(\mu,\Sigma) F = N ( μ , Σ ) , with mean μ = ( μ 1 , μ 2 ) \mu=(\mu_{1},\mu_{2}) μ = ( μ 1 , μ 2 ) , and covariance matrix Σ = diag ( Σ 11 , Σ 22 ) \Sigma=\operatorname{diag}(\Sigma_{11},\Sigma_{22}) Σ = diag ( Σ 11 , Σ 22 ) .
Characteristics. The coefficients of the characteristics of X ^ \widehat{X} X are
b ( x 1 , x 2 , x 3 ) = ( 1 α 1 + β 1 x 2 + λ μ 1 α 2 + β 2 x 3 + λ μ 2 ) , a ( x 1 , x 2 , x 3 ) = ( 0 0 0 0 σ 1 2 0 0 0 σ 2 2 ) b(x_1,x_2,x_3)=\begin{pmatrix} 1 \\ \alpha_1+\beta_1 x_2+\lambda\mu_1 \\ \alpha_2+\beta_2 x_3+\lambda\mu_2 \end{pmatrix}, \qquad a(x_1,x_2,x_3)=\begin{pmatrix} 0 & 0 & 0 \\ 0 & \sigma_1^2 & 0 \\ 0 & 0 & \sigma_2^2 \end{pmatrix} b ( x 1 , x 2 , x 3 ) = 1 α 1 + β 1 x 2 + λ μ 1 α 2 + β 2 x 3 + λ μ 2 , a ( x 1 , x 2 , x 3 ) = 0 0 0 0 σ 1 2 0 0 0 σ 2 2
with a state-independent jump kernel affecting only the OU coordinates,
K ( x 1 , x 2 , x 3 , d ξ ) = λ ( j ∗ F ) ( d ξ ) , j ( z 1 , z 2 ) = ( 0 , z 1 , z 2 ) . K(x_1,x_2,x_3,d\xi)=\lambda(j_\ast F)(d\xi), \qquad j(z_1,z_2)=(0,z_1,z_2). K ( x 1 , x 2 , x 3 , d ξ ) = λ ( j ∗ F ) ( d ξ ) , j ( z 1 , z 2 ) = ( 0 , z 1 , z 2 ) .
Tensors Q ( r ) Q^{(r)} Q ( r ) . For x ∈ S x\in\mathcal{S} x ∈ S ,
Q ( 1 ) ( x ) = ( 1 α 1 + β 1 ⟨ ϵ 2 , x ⟩ + λ μ 1 + β 1 c 1 α 2 + β 2 ⟨ ϵ 3 , x ⟩ + λ μ 2 + β 2 c 2 ) Q^{(1)}(x)=\begin{pmatrix} 1 \\ \alpha_1+\beta_1\langle\epsilon_2,x\rangle+\lambda\mu_1+\beta_1c_1 \\ \alpha_2+\beta_2\langle\epsilon_3,x\rangle+\lambda\mu_2+\beta_2c_2 \end{pmatrix} Q ( 1 ) ( x ) = 1 α 1 + β 1 ⟨ ϵ 2 , x ⟩ + λ μ 1 + β 1 c 1 α 2 + β 2 ⟨ ϵ 3 , x ⟩ + λ μ 2 + β 2 c 2
Q ( 2 ) ( x ) = 1 2 ( 0 0 0 0 σ 1 2 + λ ( μ 1 2 + Σ 11 ) λ μ 1 μ 2 0 λ μ 1 μ 2 σ 2 2 + λ ( μ 2 2 + Σ 22 ) ) Q^{(2)}(x)=\frac{1}{2}\begin{pmatrix} 0 & 0 & 0 \\ 0 & \sigma_1^2+\lambda(\mu_1^2+\Sigma_{11}) & \lambda\mu_1\mu_2 \\ 0 & \lambda\mu_1\mu_2 & \sigma_2^2+\lambda(\mu_2^2+\Sigma_{22}) \end{pmatrix} Q ( 2 ) ( x ) = 2 1 0 0 0 0 σ 1 2 + λ ( μ 1 2 + Σ 11 ) λ μ 1 μ 2 0 λ μ 1 μ 2 σ 2 2 + λ ( μ 2 2 + Σ 22 )
Q ( r ) ( x ) = λ r ! E F [ j ( Z ) ⊗ r ] , r ≥ 3 , Z ∼ F . Q^{(r)}(x)=\frac{\lambda}{r!}\mathbb{E}_{F}[j(Z)^{\otimes r}], \qquad r\geq3,\quad Z\sim F. Q ( r ) ( x ) = r ! λ E F [ j ( Z ) ⊗ r ] , r ≥ 3 , Z ∼ F .
Using the notation of (10) ,
α 1 + λ μ 1 + β 1 c 1 = b 2 ( X ^ 0 ) , β 1 = b 2 ( 2 ) ( X ^ 0 ) , \alpha_1+\lambda\mu_1+\beta_1c_1=b_2(\widehat{X}_0), \qquad\beta_1=b_2^{(2)}(\widehat{X}_0), α 1 + λ μ 1 + β 1 c 1 = b 2 ( X 0 ) , β 1 = b 2 ( 2 ) ( X 0 ) ,
α 2 + λ μ 2 + β 2 c 2 = b 3 ( X ^ 0 ) , β 2 = b 3 ( 3 ) ( X ^ 0 ) . \alpha_2+\lambda\mu_2+\beta_2c_2=b_3(\widehat{X}_0), \qquad\beta_2=b_3^{(3)}(\widehat{X}_0). α 2 + λ μ 2 + β 2 c 2 = b 3 ( X 0 ) , β 2 = b 3 ( 3 ) ( X 0 ) .
Truncation level and jump moments. We compute E [ X ^ T ≤ 5 ] \mathbb{E}[\widehat{X}_T^{\leq5}] E [ X T ≤ 5 ] , the expected signature truncated at level N = 5 N=5 N = 5 , by assembling the matrix L L L from the coefficients of Q ( r ) Q^{(r)} Q ( r ) , r ≤ 5 r\leq5 r ≤ 5 . The non-null jump moments of order 3 appearing in Q ( 3 ) Q^{(3)} Q ( 3 ) are
E F [ ξ 1 3 ] = μ 1 3 + 3 μ 1 Σ 11 , E F [ ξ 2 3 ] = μ 2 3 + 3 μ 2 Σ 22 , \mathbb{E}_{F}[\xi_1^3]=\mu_1^3+3\mu_1\Sigma_{11}, \qquad\mathbb{E}_{F}[\xi_2^3]=\mu_2^3+3\mu_2\Sigma_{22}, E F [ ξ 1 3 ] = μ 1 3 + 3 μ 1 Σ 11 , E F [ ξ 2 3 ] = μ 2 3 + 3 μ 2 Σ 22 ,
E F [ ξ 1 2 ξ 2 ] = μ 2 ( μ 1 2 + Σ 11 ) , E F [ ξ 1 ξ 2 2 ] = μ 1 ( μ 2 2 + Σ 22 ) . \mathbb{E}_{F}[\xi_1^2\xi_2]=\mu_2(\mu_1^2+\Sigma_{11}), \qquad\mathbb{E}_{F}[\xi_1\xi_2^2]=\mu_1(\mu_2^2+\Sigma_{22}). E F [ ξ 1 2 ξ 2 ] = μ 2 ( μ 1 2 + Σ 11 ) , E F [ ξ 1 ξ 2 2 ] = μ 1 ( μ 2 2 + Σ 22 ) .
The jump moments of order 4 appearing in Q ( 4 ) Q^{(4)} Q ( 4 ) are
E F [ ξ 1 4 ] = μ 1 4 + 6 μ 1 2 Σ 11 + 3 Σ 11 2 , E F [ ξ 2 4 ] = μ 2 4 + 6 μ 2 2 Σ 22 + 3 Σ 22 2 , \mathbb{E}_{F}[\xi_1^4]=\mu_1^4+6\mu_1^2\Sigma_{11}+3\Sigma_{11}^2, \qquad\mathbb{E}_{F}[\xi_2^4]=\mu_2^4+6\mu_2^2\Sigma_{22}+3\Sigma_{22}^2, E F [ ξ 1 4 ] = μ 1 4 + 6 μ 1 2 Σ 11 + 3 Σ 11 2 , E F [ ξ 2 4 ] = μ 2 4 + 6 μ 2 2 Σ 22 + 3 Σ 22 2 ,
E F [ ξ 1 3 ξ 2 ] = μ 2 ( μ 1 3 + 3 μ 1 Σ 11 ) , E F [ ξ 1 ξ 2 3 ] = μ 1 ( μ 2 3 + 3 μ 2 Σ 22 ) , \mathbb{E}_{F}[\xi_1^3\xi_2]=\mu_2(\mu_1^3+3\mu_1\Sigma_{11}), \qquad\mathbb{E}_{F}[\xi_1\xi_2^3]=\mu_1(\mu_2^3+3\mu_2\Sigma_{22}), E F [ ξ 1 3 ξ 2 ] = μ 2 ( μ 1 3 + 3 μ 1 Σ 11 ) , E F [ ξ 1 ξ 2 3 ] = μ 1 ( μ 2 3 + 3 μ 2 Σ 22 ) ,
E F [ ξ 1 2 ξ 2 2 ] = ( μ 1 2 + Σ 11 ) ( μ 2 2 + Σ 22 ) . \mathbb{E}_{F}[\xi_1^2\xi_2^2]=(\mu_1^2+\Sigma_{11})(\mu_2^2+\Sigma_{22}). E F [ ξ 1 2 ξ 2 2 ] = ( μ 1 2 + Σ 11 ) ( μ 2 2 + Σ 22 ) .
Finally, the jump moments of order 5 appearing in Q ( 5 ) Q^{(5)} Q ( 5 ) are
E F [ ξ 1 5 ] = μ 1 5 + 10 μ 1 3 Σ 11 + 15 μ 1 Σ 11 2 , E F [ ξ 2 5 ] = μ 2 5 + 10 μ 2 3 Σ 22 + 15 μ 2 Σ 22 2 , \mathbb{E}_{F}[\xi_1^5]=\mu_1^5+10\mu_1^3\Sigma_{11}+15\mu_1\Sigma_{11}^2, \qquad\mathbb{E}_{F}[\xi_2^5]=\mu_2^5+10\mu_2^3\Sigma_{22}+15\mu_2\Sigma_{22}^2, E F [ ξ 1 5 ] = μ 1 5 + 10 μ 1 3 Σ 11 + 15 μ 1 Σ 11 2 , E F [ ξ 2 5 ] = μ 2 5 + 10 μ 2 3 Σ 22 + 15 μ 2 Σ 22 2 ,
E F [ ξ 1 4 ξ 2 ] = μ 2 ( μ 1 4 + 6 μ 1 2 Σ 11 + 3 Σ 11 2 ) , E F [ ξ 1 ξ 2 4 ] = μ 1 ( μ 2 4 + 6 μ 2 2 Σ 22 + 3 Σ 22 2 ) , \mathbb{E}_{F}[\xi_1^4\xi_2]=\mu_2(\mu_1^4+6\mu_1^2\Sigma_{11}+3\Sigma_{11}^2), \qquad\mathbb{E}_{F}[\xi_1\xi_2^4]=\mu_1(\mu_2^4+6\mu_2^2\Sigma_{22}+3\Sigma_{22}^2), E F [ ξ 1 4 ξ 2 ] = μ 2 ( μ 1 4 + 6 μ 1 2 Σ 11 + 3 Σ 11 2 ) , E F [ ξ 1 ξ 2 4 ] = μ 1 ( μ 2 4 + 6 μ 2 2 Σ 22 + 3 Σ 22 2 ) ,
E F [ ξ 1 3 ξ 2 2 ] = ( μ 2 2 + Σ 22 ) ( μ 1 3 + 3 μ 1 Σ 11 ) , E F [ ξ 1 2 ξ 2 3 ] = ( μ 1 2 + Σ 11 ) ( μ 2 3 + 3 μ 2 Σ 22 ) . \mathbb{E}_{F}[\xi_1^3\xi_2^2]=(\mu_2^2+\Sigma_{22})(\mu_1^3+3\mu_1\Sigma_{11}), \qquad\mathbb{E}_{F}[\xi_1^2\xi_2^3]=(\mu_1^2+\Sigma_{11})(\mu_2^3+3\mu_2\Sigma_{22}). E F [ ξ 1 3 ξ 2 2 ] = ( μ 2 2 + Σ 22 ) ( μ 1 3 + 3 μ 1 Σ 11 ) , E F [ ξ 1 2 ξ 2 3 ] = ( μ 1 2 + Σ 11 ) ( μ 2 3 + 3 μ 2 Σ 22 ) .
Numerical validation and computational cost. This computation is available in a notebook at https://github.com/fprimavera/expected_signature_polyjumpdiffusions , which also benchmarks the resulting expected signature against a Monte Carlo estimate, obtained by simulating X ^ \hat{X} X ^ on a fine time grid and computing the signature of the resulting piecewise-linear path, as discussed in Remark 2.4. For readability, Figure 1 restricts the componentwise comparison to levels 1 to 3 (39 non-trivial components): all of them agree with the Monte Carlo estimate within 2 standard errors.
Figure 1. Componentwise comparison between the Monte Carlo estimate (± 2 \pm2 ± 2 standard errors) and the polynomial formula for E [ X ^ T ≤ 3 ] \mathbb{E}[\hat{X}_T^{\leq3}] E [ X ^ T ≤ 3 ] in Example 3.28, at T = 1 T = 1 T = 1 , restricted to levels 1–3 for readability. The computational-cost comparison discussed in the text refers instead to the full truncation level N = 5 N = 5 N = 5 .
The comparison of computational cost, on the other hand, is carried out at the full truncation level N = 5 N = 5 N = 5 , where the closed-form approach is most advantageous. Assembling L L L (a 364 × 364 364 \times364 364 × 364 matrix at N = 5 N = 5 N = 5 ) and evaluating exp ( T L ) \exp(TL) exp ( T L ) takes under a second, whereas obtaining a Monte Carlo estimate of comparable accuracy – simulating 10 4 10^4 1 0 4 paths on a grid of 10 4 10^4 1 0 4 time steps and computing their signatures up to level 5 - takes on the order of a minute, roughly two orders of magnitude slower. Moreover, the cost of the closed-form formula grows only mildly with the truncation level N N N (well under a second already at N = 5 N = 5 N = 5 ), unlike the simulation-based estimate, whose signature-computation step alone grows from a few seconds at N = 3 N = 3 N = 3 to about 48 seconds at N = 5 N = 5 N = 5 .
Score-based generative modeling with entire sig-jump-diffusions
The goal of this section is to apply entire sig-jump-diffusions for score-based generative modeling of time-series data in the spirit of Song et al. [67 ] . More precisely, we encode the empirical law of some time series ( z t ) t ∈ [ − S , 0 ] (z_t)_{t \in[-S,0]} ( z t ) t ∈ [ − S , 0 ] over the time horizon [ − S , 0 ] [-S,0] [ − S , 0 ] with S > 0 S > 0 S > 0 via the empirical law of its signature truncated at level N N N and denoted by z 0 z_0 z 0 . In the spirit of Song et al. [67 ] we then perturb the data z 0 z_0 z 0 via a filtered entire sig-jump-diffusion of the form (13) starting at z 0 z_0 z 0 such that it converges to some invariant measure2 . Sampling from this invariant distribution and using a reverse process denoted by Y ≤ N Y^{\leq N} Y ≤ N that generates the same marginals as the forward process (13) then allows us to generate new samples mimicking the law of z 0 z_0 z 0 . The essential ingredient in the reverse process construction is a generalized score function (that reduces up to a factor to the gradient of the marginal logarithmic density in the constant diffusion case) introduced in Definition 4.1 . As in standard generative diffusion models this score function has to be learned from data. Here, the framework of entire sig-jump-diffusions in the sense of § 3.3.2 and the expected signature formula become crucial as this allows us to reduce the matching of the generalized score function to the solution of linear ODEs and a deterministic optimization problem.
On the level of the original time series data there is also a nice interpretation. Indeed, consider a process given by the data time series in [ − S , 0 ] [-S,0] [ − S , 0 ] and continued by an entire sig-jump-diffusion afterwards, i.e.,
X t = { z t t ∈ [ − S , 0 ] , z 0 + ∫ 0 t ( b ˉ ( X ‾ s − ≤ N ) d s + σ ˉ ( X ‾ s − ≤ N ) d W s + ∫ E γ ˉ ( X ‾ s − ≤ N , z ) ( μ ( d s , d z ) − F ( d z ) d s ) ) t ∈ [ 0 , T ] . X_t = \begin{cases} z_t & t \in[-S,0], \\ z_0 + \int_0^t \bigl(\bar{b}(\overline{X}^{\leq N}_{s-})\,ds + \bar{\sigma}(\overline{X}^{\leq N}_{s-})\,dW_s + \int_E \bar{\gamma}(\overline{X}^{\leq N}_{s-},z)\bigl(\mu(ds,dz)-F(dz)\,ds\bigr)\bigr) & t \in[0,T]. \end{cases} X t = { z t z 0 + ∫ 0 t ( b ˉ ( X s − ≤ N ) d s + σ ˉ ( X s − ≤ N ) d W s + ∫ E γ ˉ ( X s − ≤ N , z ) ( μ ( d s , d z ) − F ( d z ) d s ) ) t ∈ [ − S , 0 ] , t ∈ [ 0 , T ] .
Then, as explained in Remark 3.20, its signature ( X t ≤ N ) t ∈ [ 0 , T ] (X_t^{\leq N})_{t\in[0,T]} ( X t ≤ N ) t ∈ [ 0 , T ] satisfies (13) for g = z 0 \mathbf{g}=\mathbf{z}_0 g = z 0 . This means that if the time-series data in [ − S , 0 ] [-S,0] [ − S , 0 ] are actually samples of the same entire sig-jump-diffusion, Law ( X − S , 0 ≤ N ) = Law ( Y T ≤ N ) \operatorname{Law}(X^{\leq N}_{-S,0}) = \operatorname{Law}(Y^{\leq N}_T) Law ( X − S , 0 ≤ N ) = Law ( Y T ≤ N ) holds true. This consistency relation can be exploited to specify b ˉ , σ ˉ , γ ˉ \bar{b},\bar{\sigma},\bar{\gamma} b ˉ , σ ˉ , γ ˉ by additionally calibrating an entire sig-jump-diffusion to the data on [ − S , 0 ] [-S,0] [ − S , 0 ] .
To introduce the whole generative modeling procedure we start by considering marginal-preserving reverse dynamics for general jump-diffusions and apply this then subsequently to the lifted entire sig-jump-diffusion of the form (13) .
Marginal-preserving reverse dynamics for general jump-diffusions
Let ( Z t ) t ∈ [ 0 , T ] (Z_t)_{t\in[0,T]} ( Z t ) t ∈ [ 0 , T ] be an R n \mathbb{R}^n R n -valued jump-diffusion satisfying
d Z t = b ( Z t − ) d t + σ ( Z t − ) d W t + ∫ E γ ( Z t − , ξ ) ( μ ( d t , d ξ ) − F ( d ξ ) d t ) , Z 0 ∼ p 0 (18) dZ_t = b(Z_{t-})\,dt + \sigma(Z_{t-})\,dW_t + \int_E \gamma(Z_{t-},\xi)\bigl(\mu(dt,d\xi)-F(d\xi)\,dt\bigr), \qquad Z_0 \sim p_0 \tag*{(18)} d Z t = b ( Z t − ) d t + σ ( Z t − ) d W t + ∫ E γ ( Z t − , ξ ) ( μ ( d t , d ξ ) − F ( d ξ ) d t ) , Z 0 ∼ p 0 (18)
where W W W is an R m \mathbb{R}^m R m -valued Brownian motion, ( E , E , F ) (E,\mathcal{E},F) ( E , E , F ) some measure space, μ \mu μ a Poisson random measure on [ 0 , T ] × E [0,T]\times E [ 0 , T ] × E with compensator F ( d ξ ) d t F(d\xi)\,dt F ( d ξ ) d t , and
b : R n → R n , σ : R n → R n × m , γ : R n × E → R n b:\mathbb{R}^n \to\mathbb{R}^n,\qquad\sigma:\mathbb{R}^n \to\mathbb{R}^{n\times m},\qquad\gamma:\mathbb{R}^n\times E \to\mathbb{R}^n b : R n → R n , σ : R n → R n × m , γ : R n × E → R n
are measurable functions. We set a ( x ) = σ ( x ) σ ( x ) T a(x)=\sigma(x)\sigma(x)^{\mathsf{T}} a ( x ) = σ ( x ) σ ( x ) T and assume that
∫ E ∣ γ ( x , ξ ) ∣ 2 F ( d ξ ) < ∞ , for all x ∈ R n , \int_E \lvert\gamma(x,\xi)\rvert^2 F(d\xi)<\infty,\qquad\text{for all }x\in\mathbb{R}^n, ∫ E ∣ γ ( x , ξ ) ∣ 2 F ( d ξ ) < ∞ , for all x ∈ R n ,
and that (18) admits a unique weak solution, in the sense that its martingale problem is well posed. These assumptions can, for example, be ensured by standard Lipschitz and linear-growth conditions on b b b , σ \sigma σ , and γ \gamma γ . We furthermore suppose that its one-time marginal laws admit densities ( p t ) t ∈ [ 0 , T ] (p_t)_{t\in[0,T]} ( p t ) t ∈ [ 0 , T ] and the associated forward Kolmogorov equation admits a unique weak solution.
For f ∈ C c 2 ( R n ) f\in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) , the generator of (18) , with the identity as “truncation” function, is
G f ( x ) = b ( x ) T ∇ f ( x ) + K f ( x ) , \mathcal{G}f(x)=b(x)^{\mathsf{T}}\nabla f(x)+\mathcal{K}f(x), G f ( x ) = b ( x ) T ∇ f ( x ) + K f ( x ) ,
where
K f ( x ) = 1 2 Tr ( a ( x ) ∇ 2 f ( x ) ) + ∫ E ( f ( x + γ ( x , ξ ) ) − f ( x ) − γ ( x , ξ ) T ∇ f ( x ) ) F ( d ξ ) . \mathcal{K}f(x)=\frac{1}{2}\operatorname{Tr}\bigl(a(x)\nabla^2f(x)\bigr)+\int_E\bigl(f(x+\gamma(x,\xi))-f(x)-\gamma(x,\xi)^{\mathsf{T}}\nabla f(x)\bigr)F(d\xi). K f ( x ) = 2 1 Tr ( a ( x ) ∇ 2 f ( x ) ) + ∫ E ( f ( x + γ ( x , ξ )) − f ( x ) − γ ( x , ξ ) T ∇ f ( x ) ) F ( d ξ ) .
Our goal is now to introduce a reverse process whose marginal laws coincide with those of Z Z Z at reversed time. To this end we need the notion of a generalized score function.
Definition 4.1. A measurable function S t : R n → R n S_t:\mathbb{R}^n\to\mathbb{R}^n S t : R n → R n is called a generalized score function associated with K \mathcal{K} K and p t p_t p t if p t S t ∈ L l o c 1 ( R n ; R n ) p_tS_t\in L^1_{\mathrm{loc}}(\mathbb{R}^n;\mathbb{R}^n) p t S t ∈ L loc 1 ( R n ; R n ) and for every f ∈ C c 2 ( R n ) f\in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) ,
∫ R n K f ( x ) p t ( x ) d x = − ∫ R n ⟨ S t ( x ) , ∇ f ( x ) ⟩ p t ( x ) d x . (19) \int_{\mathbb{R}^n}\mathcal{K}f(x)\,p_t(x)\,dx=-\int_{\mathbb{R}^n}\langle S_t(x),\nabla f(x)\rangle p_t(x)\,dx. \tag*{(19)} ∫ R n K f ( x ) p t ( x ) d x = − ∫ R n ⟨ S t ( x ) , ∇ f ( x )⟩ p t ( x ) d x . (19)
This definition can be viewed as an operator-dependent generalization of the classical score identity underlying score matching as considered in Hyvärinen [45 ] . In the specific setup of Lévy noise, the generalized score defined above corresponds to the score considered in Baule [8 ] for the finite jump activity case and to the one in Yoon et al. [71 ] for the isotropic α \alpha α -stable case, see Remark 4.2 for further details.
Remark 4.2. We here show that the generalized score introduced in Definition 4.1 reduces to the usual additional drift component of the reverse process arising in pure diffusions and the score function of isotropic α \alpha α -stable Lévy processes considered in Yoon et al. [71 ] .
First, consider the purely diffusive case with
K f ( x ) = 1 2 Tr ( a ( x ) ∇ 2 f ( x ) ) . \mathcal{K}f(x) = \frac{1}{2}\operatorname{Tr}\left(a(x)\nabla^{2}f(x)\right). K f ( x ) = 2 1 Tr ( a ( x ) ∇ 2 f ( x ) ) .
Assuming sufficient regularity to perform integration by parts, we have
∫ R n K f ( x ) p t ( x ) d x = 1 2 ∑ i , j = 1 n ∫ R n a i j ( x ) ∂ i j f ( x ) p t ( x ) d x = − 1 2 ∑ i , j = 1 n ∫ R n ∂ j ( a i j ( x ) p t ( x ) ) ∂ i f ( x ) d x . \begin{aligned}
\int_{\mathbb{R}^{n}} \mathcal{K}f(x)p_t(x)\,\mathrm{d}x
&= \frac{1}{2}\sum_{i,j=1}^{n}\int_{\mathbb{R}^{n}}a_{ij}(x)\partial_{ij}f(x)\,p_t(x)\,\mathrm{d}x \\
&= -\frac{1}{2}\sum_{i,j=1}^{n}\int_{\mathbb{R}^{n}}\partial_j\left(a_{ij}(x)p_t(x)\right)\partial_i f(x)\,\mathrm{d}x.
\end{aligned} ∫ R n K f ( x ) p t ( x ) d x = 2 1 i , j = 1 ∑ n ∫ R n a ij ( x ) ∂ ij f ( x ) p t ( x ) d x = − 2 1 i , j = 1 ∑ n ∫ R n ∂ j ( a ij ( x ) p t ( x ) ) ∂ i f ( x ) d x .
Comparing with (19) yields
S t ( x ) = 1 2 p t ( x ) div ( a ( x ) p t ( x ) ) = 1 2 div a ( x ) + 1 2 a ( x ) ∇ log p t ( x ) , S_t(x) = \frac{1}{2p_t(x)}\operatorname{div}\left(a(x)p_t(x)\right) = \frac{1}{2}\operatorname{div}a(x) + \frac{1}{2}a(x)\nabla\log p_t(x), S t ( x ) = 2 p t ( x ) 1 div ( a ( x ) p t ( x ) ) = 2 1 div a ( x ) + 2 1 a ( x ) ∇ log p t ( x ) ,
where the i i i -th component of div ( a ( x ) ) \operatorname{div}(a(x)) div ( a ( x )) is given by
( div a ( x ) ) i = ∑ j = 1 n ∂ j a i j ( x ) , i = 1 , … , n . (20) \left(\operatorname{div}a(x)\right)_i = \sum_{j=1}^{n}\partial_j a_{ij}(x), \qquad i=1,\ldots,n. \tag*{(20)} ( div a ( x ) ) i = j = 1 ∑ n ∂ j a ij ( x ) , i = 1 , … , n . (20)
In particular, if a ( x ) = σ 2 I n a(x)=\sigma^2 I_n a ( x ) = σ 2 I n , then S t ( x ) = σ 2 2 ∇ log p t ( x ) S_t(x)=\frac{\sigma^2}{2}\nabla\log p_t(x) S t ( x ) = 2 σ 2 ∇ log p t ( x ) , so that, up to the natural factor σ 2 / 2 \sigma^2/2 σ 2 /2 , the generalized score coincides with the classical score ∇ log p t \nabla\log p_t ∇ log p t .
Second, let K \mathcal{K} K be the generator of an isotropic symmetric α \alpha α -stable Lévy process with
K = − ( − Δ ) α / 2 , 0 < α < 2. \mathcal{K}=-(-\Delta)^{\alpha/2}, \qquad0<\alpha<2. K = − ( − Δ ) α /2 , 0 < α < 2.
The operator K \mathcal{K} K is self-adjoint, and
− ( − Δ ) α / 2 p = ∇ ⋅ ( ( − Δ ) ( α − 2 ) / 2 ∇ p ) . -(-\Delta)^{\alpha/2}p = \nabla\cdot\left((- \Delta)^{(\alpha-2)/2}\nabla p\right). − ( − Δ ) α /2 p = ∇ ⋅ ( ( − Δ ) ( α − 2 ) /2 ∇ p ) .
Hence
∫ R n K f ( x ) p t ( x ) d x = ∫ R n f ( x ) K p t ( x ) d x = ∫ R n f ( x ) ∇ ⋅ ( ( − Δ ) ( α − 2 ) / 2 ∇ p t ( x ) ) d x = − ∫ R n ∇ f ( x ) T ( − Δ ) ( α − 2 ) / 2 ∇ p t ( x ) d x . \begin{aligned}
\int_{\mathbb{R}^{n}}\mathcal{K}f(x)p_t(x)\,\mathrm{d}x
&= \int_{\mathbb{R}^{n}}f(x)\mathcal{K}p_t(x)\,\mathrm{d}x \\
&= \int_{\mathbb{R}^{n}}f(x)\nabla\cdot\left((- \Delta)^{(\alpha-2)/2}\nabla p_t(x)\right)\,\mathrm{d}x \\
&= -\int_{\mathbb{R}^{n}}\nabla f(x)^{\mathsf{T}}(- \Delta)^{(\alpha-2)/2}\nabla p_t(x)\,\mathrm{d}x.
\end{aligned} ∫ R n K f ( x ) p t ( x ) d x = ∫ R n f ( x ) K p t ( x ) d x = ∫ R n f ( x ) ∇ ⋅ ( ( − Δ ) ( α − 2 ) /2 ∇ p t ( x ) ) d x = − ∫ R n ∇ f ( x ) T ( − Δ ) ( α − 2 ) /2 ∇ p t ( x ) d x .
Comparison with (19) therefore gives
S t ( x ) = ( − Δ ) ( α − 2 ) / 2 ∇ p t ( x ) p t ( x ) . S_t(x)=\frac{(-\Delta)^{(\alpha-2)/2}\nabla p_t(x)}{p_t(x)}. S t ( x ) = p t ( x ) ( − Δ ) ( α − 2 ) /2 ∇ p t ( x ) .
This is precisely the fractional score introduced for isotropic α \alpha α -stable Lévy diffusion models in Yoon et al. [71 ] .
We now obtain a family of reverse dynamics analogous to the family introduced in Popov et al. [63 ] for isotropic α \alpha α -stable Lévy processes.
Theorem 4.3. Let ( Z t ) t ∈ [ 0 , T ] (Z_t)_{t\in[0,T]} ( Z t ) t ∈ [ 0 , T ] satisfy (18) , and suppose that a generalized score S t S_t S t satisfying (19) exists for every t ∈ ( 0 , T ] t\in(0,T] t ∈ ( 0 , T ] . Let η : [ 0 , T ] → [ 0 , ∞ ) \eta:[0,T]\to[0,\infty) η : [ 0 , T ] → [ 0 , ∞ ) be a bounded measurable function and set s = T − t s=T-t s = T − t . Let W ‾ \overline{W} W be an R m \mathbb{R}^m R m -valued Brownian motion, and let μ ‾ η \overline{\mu}_\eta μ η be a Poisson random measure with compensator η T − s F ( d ξ ) d s \eta_{T-s}F(d\xi)\,ds η T − s F ( d ξ ) d s . Consider the R n \mathbb{R}^n R n -valued process ( Y s ) s ∈ [ 0 , T ] (Y_s)_{s\in[0,T]} ( Y s ) s ∈ [ 0 , T ] satisfying
d Y s = ( − b ( Y s − ) + ( 1 + η T − s ) S T − s ( Y s − ) ) d s + η T − s σ ( Y s − ) d W ‾ s + ∫ E γ ( Y s − , ξ ) ( μ ‾ η ( d s , d ξ ) − η T − s F ( d ξ ) d s ) . (21) dY_s=\left(-b(Y_{s-})+(1+\eta_{T-s})S_{T-s}(Y_{s-})\right)ds+\sqrt{\eta_{T-s}}\,\sigma(Y_{s-})\,d\overline{W}_s
+\int_E\gamma(Y_{s-},\xi)\left(\overline{\mu}_\eta(ds,d\xi)-\eta_{T-s}F(d\xi)\,ds\right). \tag*{(21)} d Y s = ( − b ( Y s − ) + ( 1 + η T − s ) S T − s ( Y s − ) ) d s + η T − s σ ( Y s − ) d W s + ∫ E γ ( Y s − , ξ ) ( μ η ( d s , d ξ ) − η T − s F ( d ξ ) d s ) . (21) with initial distribution Y 0 ∼ p T ( x ) d x Y_0\sim p_T(x)\,dx Y 0 ∼ p T ( x ) d x . Assume that the martingale problem associated with (21) is well posed and that the associated forward Kolmogorov equation admits a unique weak solution. Then
Law ( Y s ) = Law ( Z T − s ) , 0 ≤ s ≤ T . (22) \operatorname{Law}(Y_s)=\operatorname{Law}(Z_{T-s}),\qquad0\leq s\leq T. \tag*{(22)} Law ( Y s ) = Law ( Z T − s ) , 0 ≤ s ≤ T . (22) In particular, the reverse process has marginal densities q s ( x ) = p T − s ( x ) q_s(x)=p_{T-s}(x) q s ( x ) = p T − s ( x ) .
Proof. For s = T − t s=T-t s = T − t , the generator of the process Y Y Y in (21) is
G s r e v , η f ( x ) = [ − b ( x ) + ( 1 + η T − s ) S T − s ( x ) ] T ∇ f ( x ) + η T − s K f ( x ) . (23) \mathcal{G}^{\mathrm{rev},\eta}_s f(x)=\left[-b(x)+(1+\eta_{T-s})S_{T-s}(x)\right]^{\mathsf{T}}\nabla f(x)+\eta_{T-s}\mathcal{K}f(x). \tag*{(23)} G s rev , η f ( x ) = [ − b ( x ) + ( 1 + η T − s ) S T − s ( x ) ] T ∇ f ( x ) + η T − s K f ( x ) . (23)
Let q s q_s q s denote the density of Y s Y_s Y s . By uniqueness of both Kolmogorov forward equations, it is enough to verify that q s ( x ) = p T − s ( x ) q_s(x)=p_{T-s}(x) q s ( x ) = p T − s ( x ) solves the Kolmogorov forward equation associated with (23) . For f ∈ C c 2 ( R n ) f\in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) , we have by (the weak formulation of) the Kolmogorov forward equation of the forward process given in (18)
d d t ∫ R n f ( x ) p t ( x ) d x = ∫ R n ( b ( x ) T ∇ f ( x ) + K f ( x ) ) p t ( x ) d x . \frac{d}{dt}\int_{\mathbb{R}^n}f(x)p_t(x)\,dx=\int_{\mathbb{R}^n}\left(b(x)^{\mathsf{T}}\nabla f(x)+\mathcal{K}f(x)\right)p_t(x)\,dx. d t d ∫ R n f ( x ) p t ( x ) d x = ∫ R n ( b ( x ) T ∇ f ( x ) + K f ( x ) ) p t ( x ) d x .
By the defining identity of S t S_t S t in (19) , this becomes
d d t ∫ R n f ( x ) p t ( x ) d x = ∫ R n ( b ( x ) − S t ( x ) ) T ∇ f ( x ) p t ( x ) d x . \frac{d}{dt}\int_{\mathbb{R}^n}f(x)p_t(x)\,dx=\int_{\mathbb{R}^n}\left(b(x)-S_t(x)\right)^{\mathsf{T}}\nabla f(x)p_t(x)\,dx. d t d ∫ R n f ( x ) p t ( x ) d x = ∫ R n ( b ( x ) − S t ( x ) ) T ∇ f ( x ) p t ( x ) d x .
Setting q s = p T − s q_s=p_{T-s} q s = p T − s , we thus have
d d s ∫ f ( x ) q s ( x ) d x = d d s ∫ f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x . (24) \frac{d}{ds}\int f(x)q_s(x)\,dx=\frac{d}{ds}\int f(x)p_{T-s}(x)\,dx
=\int\left(-b(x)+S_{T-s}(x)\right)^{\mathsf{T}}\nabla f(x)p_{T-s}(x)\,dx. \tag*{(24)} d s d ∫ f ( x ) q s ( x ) d x = d s d ∫ f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x . (24)
On the other hand, using (23) , we get
∫ G s r e v , η f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + ( 1 + η T − s ) S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x + η T − s ∫ K f ( x ) p T − s ( x ) d x . \int\mathcal{G}^{\mathrm{rev},\eta}_s f(x)p_{T-s}(x)\,dx=\int\left(-b(x)+(1+\eta_{T-s})S_{T-s}(x)\right)^{\mathsf{T}}\nabla f(x)p_{T-s}(x)\,dx
+\eta_{T-s}\int\mathcal{K}f(x)p_{T-s}(x)\,dx. ∫ G s rev , η f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + ( 1 + η T − s ) S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x + η T − s ∫ K f ( x ) p T − s ( x ) d x .
Applying (19) once more yields
∫ G s r e v , η f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + ( 1 + η T − s ) S T − s ( x ) − η T − s S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x . (25) \int\mathcal{G}^{\mathrm{rev},\eta}_s f(x)p_{T-s}(x)\,dx=\int\left(-b(x)+(1+\eta_{T-s})S_{T-s}(x)-\eta_{T-s}S_{T-s}(x)\right)^{\mathsf{T}}\nabla f(x)p_{T-s}(x)\,dx
=\int\left(-b(x)+S_{T-s}(x)\right)^{\mathsf{T}}\nabla f(x)p_{T-s}(x)\,dx. \tag*{(25)} ∫ G s rev , η f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + ( 1 + η T − s ) S T − s ( x ) − η T − s S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x = ∫ ( − b ( x ) + S T − s ( x ) ) T ∇ f ( x ) p T − s ( x ) d x . (25)
Since the right-hand sides of (24) and (25) coincide, we conclude that q s = p T − s q_s=p_{T-s} q s = p T − s satisfies the Kolmogorov forward equation of the backward process given in (21) , i.e.,
d d s ∫ f ( x ) q s ( x ) d x = ∫ G s r e v , η f ( x ) q s ( x ) d x . \frac{d}{ds}\int f(x)q_s(x)\,dx=\int\mathcal{G}^{\mathrm{rev},\eta}_s f(x)q_s(x)\,dx. d s d ∫ f ( x ) q s ( x ) d x = ∫ G s rev , η f ( x ) q s ( x ) d x .
The fact that q 0 = p T q_0=p_T q 0 = p T and the uniqueness of the martingale problem of the backward process imply (22) .
Remark 4.4. For the special choice η t ≡ 0 \eta_t\equiv0 η t ≡ 0 , all stochastic terms in (21) disappear and one obtains the following ODE
d Y s = ( − b ( Y s ) + S T − s ( Y s ) ) d s . (26) dY_s=\left(-b(Y_s)+S_{T-s}(Y_s)\right)\,ds. \tag*{(26)} d Y s = ( − b ( Y s ) + S T − s ( Y s ) ) d s . (26)
This deterministic flow (26) has the same one-time marginal distributions p T − s p_{T-s} p T − s as the forward jump-diffusion.
For the other special case, namely η t ≡ 1 \eta_t\equiv1 η t ≡ 1 , only the drift is modified by the score such that the reverse dynamics have the same local diffusion covariance and the same jump intensity as the forward process:
d Y s = ( − b ( Y s − ) + 2 S T − s ( Y s − ) ) d s + σ ( Y s − ) d W ‾ s + ∫ E γ ( Y s − , ξ ) ( μ ‾ ( d s , d ξ ) − F ( d ξ ) d s ) . dY_s=\left(-b(Y_{s-})+2S_{T-s}(Y_{s-})\right)\,ds+\sigma(Y_{s-})\,d\overline{W}_s+\int_E\gamma(Y_{s-},\xi)\left(\overline{\mu}(ds,d\xi)-F(d\xi)\,ds\right). d Y s = ( − b ( Y s − ) + 2 S T − s ( Y s − ) ) d s + σ ( Y s − ) d W s + ∫ E γ ( Y s − , ξ ) ( μ ( d s , d ξ ) − F ( d ξ ) d s ) .
Remark 4.5. Note that the process (21) is in general not the true pathwise time reversal of (18) . Rather, it belongs to a family of processes whose one-time marginal distributions agree exactly with those of the forward process under time reversal. This coincides with the notion of reverse dynamics used in Theorem 1 of Popov et al. [63 ] . In contrast, the pathwise time reversal of a Markov process with jumps generally involves a modification of its jump characteristics depending on the marginal laws; we refer to Comforti and Léonard [19 ] for a characterization of the semimartingale characteristics of the time-reversed process in a pathwise sense.
Marginal-preserving reverse dynamics for entire sig-jump-diffusions
We now apply the preceding construction to entire sig-jump-diffusions with truncated signature-dependent coefficients as considered in § 3.3.2 , more precisely to the lifted equation Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N given in (13) with some initial value z 0 ∼ p 0 = p d a t a z_0\sim p_0=p_{\mathrm{data}} z 0 ∼ p 0 = p data . Note that we are in particular assuming that the process defined in (11) is an entire sig-jump-diffusion in the sense of Remark 3.21. With the definition of b , σ , γ b,\sigma,\gamma b , σ , γ given there and with a ( x ) = σ ( x ) σ ( x ) T a(x)=\sigma(x)\sigma(x)^\mathsf{T} a ( x ) = σ ( x ) σ ( x ) T , the generator associated with (13) is given by
G N f ( x ) = ⟨ b ( x ) , ∇ f ( x ) ⟩ + K N f ( x ) . \mathcal{G}^{N}f(x)=\langle b(x),\nabla f(x)\rangle+\mathcal{K}^{N}f(x). G N f ( x ) = ⟨ b ( x ) , ∇ f ( x )⟩ + K N f ( x ) .
where
K N f ( x ) = 1 2 Tr ( a ( x ) ∇ 2 f ( x ) ) + ∫ E ( f ( x + γ ( x , z ) ) − f ( x ) − ⟨ γ ( x , z ) , ∇ f ( x ) ⟩ ) F ( d z ) . \mathcal{K}^{N}f(x)=\frac{1}{2}\operatorname{Tr}\left(a(x)\nabla^2f(x)\right)+\int_E\left(f(x+\gamma(x,z))-f(x)-\langle\gamma(x,z),\nabla f(x)\rangle\right)F(dz). K N f ( x ) = 2 1 Tr ( a ( x ) ∇ 2 f ( x ) ) + ∫ E ( f ( x + γ ( x , z )) − f ( x ) − ⟨ γ ( x , z ) , ∇ f ( x )⟩ ) F ( d z ) .
We set P t N = Law ( Z t ≤ N ) P_t^N=\operatorname{Law}(\mathbb{Z}_t^{\leq N}) P t N = Law ( Z t ≤ N ) and adapt the definition of the generalized score function to the current setting, accounting for the fact that we do not have a density on T N ( R d ) T^N(\mathbb{R}^d) T N ( R d ) since the law P t N P_t^N P t N is only supported on (possibly a subset of) group-like elements G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) . We additionally assume that the associated Kolmogorov forward equation for P t N P_t^N P t N admits a unique weak solution.
Definition 4.6. A measurable function S t N : T N ( R d ) → T N ( R d ) S_t^N:T^N(\mathbb{R}^d)\to T^N(\mathbb{R}^d) S t N : T N ( R d ) → T N ( R d ) is called a generalized score function of Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N at time t t t if it is integrable with respect to P t N P_t^N P t N and
∫ K N f ( x ) P t N ( d x ) = − ∫ ⟨ S t N ( x ) , ∇ f ( x ) ⟩ P t N ( d x ) (27) \int\mathcal{K}^N f(x)\,P_t^N(dx)=-\int\langle S_t^N(x),\nabla f(x)\rangle P_t^N(dx) \tag*{(27)} ∫ K N f ( x ) P t N ( d x ) = − ∫ ⟨ S t N ( x ) , ∇ f ( x )⟩ P t N ( d x ) (27)
for every f ∈ C c 2 ( T N ( R d ) ) f\in C_c^2(T^N(\mathbb{R}^d)) f ∈ C c 2 ( T N ( R d )) for which the above expressions are finite.
We can now apply the marginal-preserving reverse construction directly to the signature process.
Corollary 4.7. Assume that, for every t ∈ ( 0 , T ] t \in(0,T] t ∈ ( 0 , T ] , there exists a generalized score S t N S_t^N S t N satisfying (27) . Let η : [ 0 , T ] → [ 0 , ∞ ) \eta: [0,T] \to[0,\infty) η : [ 0 , T ] → [ 0 , ∞ ) be bounded and measurable, and introduce the increasing reverse-time variable s = T − t s = T - t s = T − t . Let W ‾ \overline{W} W be a d d d -dimensional Brownian motion and let μ ‾ η \overline{\mu}_{\eta} μ η be a Poisson random measure with compensator η T − s F ( d z ) d s \eta_{T-s}F(\mathrm{d}z)\mathrm{d}s η T − s F ( d z ) d s . Consider
d Y s ≤ N = ( − b ( Y s − ≤ N ) + ( 1 + η T − s ) S T − s N ( Y s − ≤ N ) ) d s + η T − s σ ( Y s − ≤ N ) d W ‾ s + ∫ E γ ( Y s − ≤ N , z ) ( μ ‾ η ( d s , d z ) − η T − s F ( d z ) d s ) , (28) \mathrm{d}Y_s^{\leq N} = \left(-b\left(Y_{s-}^{\leq N}\right) + \left(1+\eta_{T-s}\right)S_{T-s}^N\left(Y_{s-}^{\leq N}\right)\right)\mathrm{d}s + \sqrt{\eta_{T-s}}\sigma\left(Y_{s-}^{\leq N}\right)\mathrm{d}\overline{W}_s
+ \int_E \gamma\left(Y_{s-}^{\leq N},z\right)\left(\overline{\mu}_{\eta}(\mathrm{d}s,\mathrm{d}z)-\eta_{T-s}F(\mathrm{d}z)\mathrm{d}s\right), \tag*{(28)} d Y s ≤ N = ( − b ( Y s − ≤ N ) + ( 1 + η T − s ) S T − s N ( Y s − ≤ N ) ) d s + η T − s σ ( Y s − ≤ N ) d W s + ∫ E γ ( Y s − ≤ N , z ) ( μ η ( d s , d z ) − η T − s F ( d z ) d s ) , (28) with Y 0 ≤ N ∼ P T N Y_0^{\leq N} \sim P_T^N Y 0 ≤ N ∼ P T N . Assume that the martingale problem associated with (28) is well posed on T N ( R d ) T^N(\mathbb{R}^d) T N ( R d ) and that the associated Kolmogorov forward equation admits a unique weak solution. Then
Law ( Y s ≤ N ) = P T − s N = Law ( Z T − s ≤ N ) , 0 ≤ s ≤ T . \operatorname{Law}\left(Y_s^{\leq N}\right) = P_{T-s}^N = \operatorname{Law}\left(Z_{T-s}^{\leq N}\right), \qquad0 \leq s \leq T. Law ( Y s ≤ N ) = P T − s N = Law ( Z T − s ≤ N ) , 0 ≤ s ≤ T .
Remark 4.8. Since the law P t N P_t^N P t N of the forward truncated signature is supported on (subsets of) group-like elements G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) , the marginal identity Law ( Y s ≤ N ) = P T − s N \operatorname{Law}\left(Y_s^{\leq N}\right) = P_{T-s}^N Law ( Y s ≤ N ) = P T − s N implies that P ( Y s ≤ N ∈ G N ( R d ) ) = 1 \mathbb{P}\left(Y_s^{\leq N} \in G^N(\mathbb{R}^d)\right) = 1 P ( Y s ≤ N ∈ G N ( R d ) ) = 1 for every fixed s ∈ [ 0 , T ] s \in[0,T] s ∈ [ 0 , T ] . Since G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) is closed and Y ≤ N Y^{\leq N} Y ≤ N has càdlàg paths, it follows by considering a countable dense set of times and using right-continuity that
P ( Y s ≤ N ∈ G N ( R d ) for all s ∈ [ 0 , T ] ) = 1. \mathbb{P}\left(Y_s^{\leq N} \in G^N(\mathbb{R}^d)\text{ for all }s \in[0,T]\right) = 1. P ( Y s ≤ N ∈ G N ( R d ) for all s ∈ [ 0 , T ] ) = 1.
Thus, the marginal-preserving reverse dynamics remain in G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) . It is, however, not automatic that Y ≤ N Y^{\leq N} Y ≤ N is the truncated signature of its own first-level process, since the generalized score may introduce higher-level drift components that do not coincide with the canonical signature lift of the first-level reverse dynamics. Additional compatibility conditions on the generalized score would therefore be needed.
Learning the generalized score function of entire sig-jump-diffusions via an entire approximation
We now aim to use the above reverse process to generate signatures mimicking the law p d a t a p_{\mathrm{data}} p data of the signature Z 0 Z_0 Z 0 stemming from some time-series data ( z t ) t ∈ [ − S , 0 ] (z_t)_{t\in[-S,0]} ( z t ) t ∈ [ − S , 0 ] and truncated at level N N N . To this end the generalized score function has to be learned. For this purpose we approximate it by an entire function in time and the signature components, and use Corollary 3.15 for the computation of the expected signature to obtain a tractable training objective. Consider a function
h u : R × T N ( R d ) → R h_u : \mathbb{R} \times T^N(\mathbb{R}^d) \to\mathbb{R} h u : R × T N ( R d ) → R
which is jointly entire in ( t , x ) (t,\mathbf{x}) ( t , x ) . We write its power series representation as
h u ( t , x ) = ∑ k ≥ 0 ∑ α ∈ N 0 D N u k , α t k x α h_u(t,\mathbf{x}) = \sum_{k\geq0}\sum_{\alpha\in\mathbb{N}_0^{D_N}} u_{k,\alpha}t^k\mathbf{x}^{\alpha} h u ( t , x ) = k ≥ 0 ∑ α ∈ N 0 D N ∑ u k , α t k x α
with multi-indices α ∈ N 0 D N \alpha\in\mathbb{N}_0^{D_N} α ∈ N 0 D N and D N : = dim T N ( R d ) D_N := \dim T^N(\mathbb{R}^d) D N := dim T N ( R d ) and define the score approximation by
S t N ( x ) ≈ g u ( t , x ) : = ∇ x h u ( t , x ) . S_t^N(\mathbf{x}) \approx g_u(t,\mathbf{x}) := \nabla_{\mathbf{x}}h_u(t,\mathbf{x}). S t N ( x ) ≈ g u ( t , x ) := ∇ x h u ( t , x ) .
For Z 0 ≤ N = z 0 Z_0^{\leq N} = \mathbf{z}_0 Z 0 ≤ N = z 0 , let S t ( ⋅ ∣ z 0 ) S_t(\cdot\mid\mathbf{z}_0) S t ( ⋅ ∣ z 0 ) denote a conditional generalized score, characterized weakly by
E [ K N f ( Z t ≤ N ) ∣ Z 0 ≤ N = z 0 ] = − E [ ⟨ S t ( Z t ≤ N ∣ z 0 ) , ∇ f ( Z t ≤ N ) ⟩ ∣ Z 0 ≤ N = z 0 ] (29) \mathbb{E}\left[\mathcal{K}^Nf\left(Z_t^{\leq N}\right)\mid Z_0^{\leq N}=\mathbf{z}_0\right] = -\mathbb{E}\left[\left\langle S_t\left(Z_t^{\leq N}\mid\mathbf{z}_0\right),\nabla f\left(Z_t^{\leq N}\right)\right\rangle\mid Z_0^{\leq N}=\mathbf{z}_0\right] \tag*{(29)} E [ K N f ( Z t ≤ N ) ∣ Z 0 ≤ N = z 0 ] = − E [ ⟨ S t ( Z t ≤ N ∣ z 0 ) , ∇ f ( Z t ≤ N ) ⟩ ∣ Z 0 ≤ N = z 0 ] (29)
for all admissible test functions f f f . We consider now the following denoising generalized score-matching objective
argmin u ∫ 0 T λ ( t ) E Z 0 ∼ p d a t a [ E [ ∥ g u ( t , Z t ≤ N ) − S t ( Z t ≤ N ∣ Z 0 ) ∥ 2 ∣ Z 0 ≤ N = Z 0 ] ] d t , \operatorname*{argmin}_u \int_0^T \lambda(t)\mathbb{E}_{Z_0\sim p_{\mathrm{data}}}\left[\mathbb{E}\left[\left\|g_u\left(t,Z_t^{\leq N}\right)-S_t\left(Z_t^{\leq N}\mid Z_0\right)\right\|^2\mid Z_0^{\leq N}=Z_0\right]\right]\mathrm{d}t, u argmin ∫ 0 T λ ( t ) E Z 0 ∼ p data [ E [ g u ( t , Z t ≤ N ) − S t ( Z t ≤ N ∣ Z 0 ) 2 ∣ Z 0 ≤ N = Z 0 ] ] d t ,
with λ \lambda λ some positive weight function. By Fubini-Tonelli’s theorem this is equivalent to
arg min u E Z 0 ∼ p d a t a [ E [ ∫ 0 T λ ( t ) ∥ g u ( t , Z t ≤ N ) − S t ( Z t ≤ N ∣ z 0 ) ∥ 2 d t ∣ Z 0 ≤ N = z 0 ] ] . (30) \underset{\mathbf{u}}{\arg\min}\ \mathbb{E}_{Z_0\sim p_{\mathrm{data}}}\left[\mathbb{E}\left[\left.\int_0^T \lambda(t)\left\|g_{\mathbf{u}}(t,Z_t^{\leq N})-S_t(Z_t^{\leq N}\mid z_0)\right\|^2\,\mathrm{d}t\right|Z_0^{\leq N}=z_0\right]\right]. \tag*{(30)} u arg min E Z 0 ∼ p data [ E [ ∫ 0 T λ ( t ) g u ( t , Z t ≤ N ) − S t ( Z t ≤ N ∣ z 0 ) 2 d t Z 0 ≤ N = z 0 ] ] . (30)
and also to
arg min u E Z 0 ∼ p d a t a [ E [ ∫ 0 T λ ( t ) ( ∥ g u ( t , Z t ≤ N ) ∥ 2 − 2 ⟨ g u ( t , Z t ≤ N ) , S t ( Z t ≤ N ∣ z 0 ) ⟩ ) d t ∣ Z 0 ≤ N = z 0 ] ] . \underset{\mathbf{u}}{\arg\min}\ \mathbb{E}_{Z_0\sim p_{\mathrm{data}}}\left[\mathbb{E}\left[\left.\int_0^T \lambda(t)\left(\left\|g_{\mathbf{u}}(t,Z_t^{\leq N})\right\|^2-2\left\langle g_{\mathbf{u}}(t,Z_t^{\leq N}),S_t(Z_t^{\leq N}\mid z_0)\right\rangle\right)\,\mathrm{d}t\right|Z_0^{\leq N}=z_0\right]\right]. u arg min E Z 0 ∼ p data [ E [ ∫ 0 T λ ( t ) ( g u ( t , Z t ≤ N ) 2 − 2 ⟨ g u ( t , Z t ≤ N ) , S t ( Z t ≤ N ∣ z 0 ) ⟩ ) d t Z 0 ≤ N = z 0 ] ] .
Applying (29) with f ( x ) = h u ( t , x ) f(\mathbf{x})=h_{\mathbf{u}}(t,\mathbf{x}) f ( x ) = h u ( t , x ) yields
arg min u E Z 0 ∼ p d a t a [ E [ ∫ 0 T λ ( t ) ( ∥ ∇ x h u ( t , Z t ≤ N ) ∥ 2 + 2 K N h u ( t , Z t ≤ N ) ) d t ∣ Z 0 ≤ N = z 0 ] ] . (31) \underset{\mathbf{u}}{\arg\min}\ \mathbb{E}_{Z_0\sim p_{\mathrm{data}}}\left[\mathbb{E}\left[\left.\int_0^T \lambda(t)\left(\left\|\nabla_{\mathbf{x}}h_{\mathbf{u}}(t,Z_t^{\leq N})\right\|^2+2\mathcal{K}^N h_{\mathbf{u}}(t,Z_t^{\leq N})\right)\,\mathrm{d}t\right|Z_0^{\leq N}=z_0\right]\right]. \tag*{(31)} u arg min E Z 0 ∼ p data [ E [ ∫ 0 T λ ( t ) ( ∇ x h u ( t , Z t ≤ N ) 2 + 2 K N h u ( t , Z t ≤ N ) ) d t Z 0 ≤ N = z 0 ] ] . (31)
Here, K N \mathcal{K}^N K N acts only on the signature variable x \mathbf{x} x . In particular, the unknown generalized score no longer appears in (31) .
The objective (31) can now be computed via Corollary 3.15. Indeed, define
q u ( t , x ) : = ∥ ∇ x h u ( t , x ) ∥ 2 + 2 K N h u ( t , x ) . q_{\mathbf{u}}(t,\mathbf{x}):=\left\|\nabla_{\mathbf{x}}h_{\mathbf{u}}(t,\mathbf{x})\right\|^2+2\mathcal{K}^N h_{\mathbf{u}}(t,\mathbf{x}). q u ( t , x ) := ∥ ∇ x h u ( t , x ) ∥ 2 + 2 K N h u ( t , x ) .
Under suitable conditions on h u h_{\mathbf{u}} h u (Cuchiero et al., [26 ] , Lemma A.1), the conditions on K N \mathcal{K}^N K N yield that q u q_{\mathbf{u}} q u is a jointly entire function. To incorporate the explicit time dependence, we shall use the time-extended version of (13) , which we denote by Z ^ ≤ N \widehat{\mathbb{Z}}^{\leq N} Z ≤ N . Its starting value is z ^ 0 \widehat{z}_0 z 0 which is the signature of the time-series data. More precisely, z ^ 0 − 1 ⊗ Z ^ ≤ N \widehat{z}_0^{-1}\otimes\widehat{Z}^{\leq N} z 0 − 1 ⊗ Z ≤ N is the signature of X ^ z ^ 0 : = ( t , X t ) z ^ 0 \widehat{X}^{\widehat{z}_0}:=(t,X_t)^{\widehat{z}_0} X z 0 := ( t , X t ) z 0 given by (11) extended to R d + 1 \mathbb{R}^{d+1} R d + 1 with b ^ 0 = 1 \widehat{b}_0=1 b 0 = 1 and σ ^ 0 i = δ ^ 0 = 0 \widehat{\sigma}_{0i}=\widehat{\delta}_{0}=0 σ 0 i = δ 0 = 0 . Here, the 0-index refers to time. Using this we can express
q u ( t , Z ^ t ≤ N ) = ⟨ v ( u ) , Z ^ t ⟩ , q_{\mathbf{u}}(t,\widehat{Z}_t^{\leq N})=\langle\mathbf{v}(\mathbf{u}),\widehat{Z}_t\rangle, q u ( t , Z t ≤ N ) = ⟨ v ( u ) , Z t ⟩ ,
where v ( u ) \mathbf{v}(\mathbf{u}) v ( u ) denotes the corresponding coefficient sequence of q u q_{\mathbf{u}} q u . Note that the right-hand side depends on the non-truncated signature Z ^ \widehat{Z} Z , but just through the terms that can be expressed as a polynomial in Z ^ ≤ N \widehat{Z}^{\leq N} Z ≤ N . In particular, the relevant terms of higher level than N N N of Z ^ 0 = Z ^ 0 ∞ \widehat{Z}_0=\widehat{Z}_0^\infty Z 0 = Z 0 ∞ correspond to powers of the lower ones and are thus uniquely determined by them.
Let L L L be the operator associated to X ^ ∞ \widehat{\mathcal{X}}^\infty X ∞ as defined in (5) and L ‾ \overline{L} L its extension to the full signature as of Definition 3.11. Provided that v ( u ) ∈ D ( L ‾ ) \mathbf{v}(\mathbf{u})\in D(\overline{L}) v ( u ) ∈ D ( L ) and that the corresponding integrability conditions hold and since Z ^ t ≤ N = Z ^ 0 ⊗ X ^ t z ^ 0 \widehat{Z}_t^{\leq N}=\widehat{Z}_0\otimes\widehat{X}_t^{\widehat{z}_0} Z t ≤ N = Z 0 ⊗ X t z 0 , Corollary 3.15 thus yields
E [ ⟨ v ( u ) , Z ^ t ⟩ ∣ Z ^ 0 = z ^ 0 ∞ ] = ⟨ c u ( t ) , z ^ 0 ∞ ⟩ , \mathbb{E}\left[\left.\langle\mathbf{v}(\mathbf{u}),\widehat{Z}_t\rangle\right|\widehat{Z}_0=\widehat{z}_0^\infty\right]=\langle\mathbf{c}_{\mathbf{u}}(t),\widehat{z}_0^\infty\rangle, E [ ⟨ v ( u ) , Z t ⟩ Z 0 = z 0 ∞ ] = ⟨ c u ( t ) , z 0 ∞ ⟩ ,
where
d d s c u ( s ) = L ‾ c u ( s ) , c u ( 0 ) = v ( u ) , (32) \frac{\mathrm{d}}{\mathrm{d}s}\mathbf{c}_{\mathbf{u}}(s)=\overline{L}\mathbf{c}_{\mathbf{u}}(s),\qquad\mathbf{c}_{\mathbf{u}}(0)=\mathbf{v}(\mathbf{u}), \tag*{(32)} d s d c u ( s ) = L c u ( s ) , c u ( 0 ) = v ( u ) , (32)
under the condition that (32) admits a solution. The inner expectation of the loss (31) then becomes ∫ 0 T λ ( t ) ⟨ c u ( t ) , z ^ 0 ∞ ⟩ d t \int_0^T\lambda(t)\langle\mathbf{c}_{\mathbf{u}}(t),\widehat{z}_0^\infty\rangle\,\mathrm{d}t ∫ 0 T λ ( t ) ⟨ c u ( t ) , z 0 ∞ ⟩ d t . If λ ( t ) \lambda(t) λ ( t ) is an entire function the above reasoning can be applied to
⟨ v ~ ( u ) , Z ^ T ⟩ : = ∫ 0 T λ ( t ) q u ( t , Z ^ t ≤ N ) d t (33) \langle\widetilde{\mathbf{v}}(\mathbf{u}),\widehat{Z}_T\rangle:=\int_0^T\lambda(t)q_{\mathbf{u}}(t,\widehat{Z}_t^{\leq N})\,\mathrm{d}t \tag*{(33)} ⟨ v ( u ) , Z T ⟩ := ∫ 0 T λ ( t ) q u ( t , Z t ≤ N ) d t (33)
so that not even an integral has to be computed. Again, provided that v ~ ( u ) ∈ D ( L ‾ ) \widetilde{\mathbf{v}}(\mathbf{u})\in D(\overline{L}) v ( u ) ∈ D ( L ) and the corresponding integrability conditions hold, E [ ⟨ v ~ ( u ) , Z ^ t ⟩ ∣ Z ^ 0 = z ^ 0 ∞ ] = ⟨ c ~ u ( t ) , z ^ 0 ∞ ⟩ \mathbb{E}[\langle\widetilde{\mathbf{v}}(\mathbf{u}),\widehat{Z}_t\rangle\mid\widehat{Z}_0=\widehat{z}_0^\infty]=\langle\widetilde{\mathbf{c}}_{\mathbf{u}}(t),\widehat{z}_0^\infty\rangle E [⟨ v ( u ) , Z t ⟩ ∣ Z 0 = z 0 ∞ ] = ⟨ c u ( t ) , z 0 ∞ ⟩ where c ~ u ( t ) \widetilde{\mathbf{c}}_{\mathbf{u}}(t) c u ( t ) is the solution of (32) with c ~ u ( 0 ) = v ~ ( u ) \widetilde{\mathbf{c}}_{\mathbf{u}}(0)=\widetilde{\mathbf{v}}(\mathbf{u}) c u ( 0 ) = v ( u ) . Then (31) becomes
arg min u E Z 0 ∼ p d a t a [ ⟨ c ~ u ( T ) , z ^ 0 ∞ ⟩ ] . \underset{\mathbf{u}}{\arg\min}\ \mathbb{E}_{Z_0\sim p_{\mathrm{data}}}\left[\langle\widetilde{\mathbf{c}}_{\mathbf{u}}(T),\widehat{z}_0^\infty\rangle\right]. u arg min E Z 0 ∼ p data [ ⟨ c u ( T ) , z 0 ∞ ⟩ ] .
Thus, generalized score matching is reduced to the solution of a linear coefficient equation together with a deterministic optimization over the coefficients u \mathbf{u} u . We now summarize this in the following theorem.
Theorem 4.9. Let λ \lambda λ in (30) be an entire function. Moreover, let L L L be the operator associated to X ^ ⊕ ∞ \widehat{X}_{\oplus}^{\infty} X ⊕ ∞ as defined in (5) and L ‾ \overline{L} L its extension to the full signature as of Definition 3.11. Assume that all the conditions of Corollary 3.15 with u \mathbf{u} u replaced by v ~ ( u ) \widetilde{\mathbf{v}}(\mathbf{u}) v ( u ) defined in (33) are satisfied. Then, (30) is equivalent to
argmin u E z 0 ∼ p d a t a [ ⟨ v ~ ( u ) , Z ^ T ⟩ ] = argmin u E z 0 ∼ p d a t a [ ⟨ c ~ u ( T ) , z ^ 0 ∞ ⟩ ] . \operatorname*{argmin}_{\mathbf{u}}\mathbb{E}_{\mathbf{z}_0\sim p_{\mathrm{data}}}\left[\langle\widetilde{\mathbf{v}}(\mathbf{u}),\widehat{Z}_{T}\rangle\right]=\operatorname*{argmin}_{\mathbf{u}}\mathbb{E}_{\mathbf{z}_0\sim p_{\mathrm{data}}}\left[\langle\widetilde{c}_{\mathbf{u}}(T),\widehat{z}_0^{\infty}\rangle\right]. u argmin E z 0 ∼ p data [ ⟨ v ( u ) , Z T ⟩ ] = u argmin E z 0 ∼ p data [ ⟨ c u ( T ) , z 0 ∞ ⟩ ] . with
d d s c ~ u ( s ) = L ‾ c ~ u ( s ) , c ~ u ( 0 ) = v ~ ( u ) . (34) \frac{d}{ds}\widetilde{c}_{\mathbf{u}}(s)=\overline{L}\widetilde{c}_{\mathbf{u}}(s),\qquad\widetilde{c}_{\mathbf{u}}(0)=\widetilde{\mathbf{v}}(\mathbf{u}). \tag*{(34)} d s d c u ( s ) = L c u ( s ) , c u ( 0 ) = v ( u ) . (34)
Truncating the operator L ‾ \overline{L} L , z ^ 0 ∞ \widehat{z}_0^{\infty} z 0 ∞ and v ~ ( u ) \widetilde{\mathbf{v}}(\mathbf{u}) v ( u ) at some level K K K and denoting them by L ‾ K \overline{L}^{K} L K , z ^ 0 K \widehat{z}_0^{K} z 0 K and v ~ K ( u ) \widetilde{\mathbf{v}}^{K}(\mathbf{u}) v K ( u ) , the (34) becomes
argmin u ( v ~ K ( u ) ) T exp ( L ‾ K T ) E z 0 ∼ p d a t a [ z ^ 0 K ] , \operatorname*{argmin}_{\mathbf{u}}\left(\widetilde{\mathbf{v}}^{K}(\mathbf{u})\right)^{\mathsf{T}}\exp\left(\overline{L}^{K}T\right)\mathbb{E}_{\mathbf{z}_0\sim p_{\mathrm{data}}}\left[\widehat{z}_0^{K}\right], u argmin ( v K ( u ) ) T exp ( L K T ) E z 0 ∼ p data [ z 0 K ] ,
meaning that only one deterministic optimization and one computation of a matrix exponential have to be performed.
Once the score function is learned by this method, we can use g u ( t , x ) g_{\mathbf{u}}(t,\mathbf{x}) g u ( t , x ) as an approximation of the generalized score function in the reverse process Y ^ ≤ N \widehat{Y}^{\leq N} Y ≤ N to generate signature samples whose distribution resembles the empirical law p d a t a p_{\mathrm{data}} p data .
Invariant measures for the Markovian lift of entire sig-diffusions
To illustrate the practical applicability of the above approach, we here give an example of an entire sig-diffusion whose lift admits an invariant measure on G N ( R d ) G^{N}(\mathbb{R}^{d}) G N ( R d ) .
Let m G m_{G} m G be a Haar measure3 on G N ( R d ) G^{N}(\mathbb{R}^{d}) G N ( R d ) , i.e., the unique (up to a positive multiplicative factor) nonzero, locally finite measure on G N ( R d ) G^{N}(\mathbb{R}^{d}) G N ( R d ) such that, for every measurable set A A A and every h ∈ G N ( R d ) h \in G^{N}(\mathbb{R}^{d}) h ∈ G N ( R d ) ,
m G ( h ⊗ A ) = m G ( A ⊗ h ) = m G ( A ) . m_{G}(h \otimes A) = m_{G}(A \otimes h) = m_{G}(A). m G ( h ⊗ A ) = m G ( A ⊗ h ) = m G ( A ) .
The following theorem now shows a construction of entire sig-diffusions with truncated signature-dependent coefficients in spirit of Langevin type dynamics that admit an invariant density. As we show in the proof its lifted sig-diffusion on G ≤ N G^{\leq N} G ≤ N belongs to the class of hypoelliptic gradient diffusions on Lie groups studied in Baudoin et al. [7 ] . Indeed, the free nilpotent group G N ( R d ) G^{N}(\mathbb{R}^{d}) G N ( R d ) is a finite-dimensional Lie group, and the vector fields associated with the first-level signature components are left-invariant and satisfy Hörmander’s bracket-generating condition. The corresponding generator is therefore of the form considered in Baudoin et al. [7 ] , where global existence and invariance of a Gibbs-type measure with a density with respect to the Haar measure are established. Moreover, the strong Feller property and irreducibility of the lifted diffusion yield uniqueness of the invariant measure and convergence in total variation by Doob’s ergodic theorem, which we in apply in the version provided by Gerlach and Nittka [39 ] .
Theorem 4.11. Let U : G N ( R d ) → [ 0 , ∞ ) U:G^{N}(\mathbb{R}^{d}) \to[0,\infty) U : G N ( R d ) → [ 0 , ∞ ) be the restriction of an entire function on T N ( R d ) T^{N}(\mathbb{R}^{d}) T N ( R d ) such that ∫ G N ( R d ) e − U ( g ) m G ( d g ) < ∞ \int_{G^{N}(\mathbb{R}^{d})} e^{-U(\mathbf{g})}m_{G}(d\mathbf{g}) < \infty ∫ G N ( R d ) e − U ( g ) m G ( d g ) < ∞ . With the notation introduced for (11) , define σ ˉ = i d \bar{\sigma} = \mathrm{id} σ ˉ = id , δ ˉ = 0 \bar{\delta} = 0 δ ˉ = 0 , b ˉ : G N ( R d ) → R d \bar{b}:G^{N}(\mathbb{R}^{d}) \to\mathbb{R}^{d} b ˉ : G N ( R d ) → R d via
b ˉ i ( g ) = − 1 2 d d s U ( g ⊗ exp N ( s e i ) ) ∣ s = 0 , \bar{b}_{i}(\mathbf{g}) = -\frac{1}{2}\left.\frac{d}{ds}U(\mathbf{g}\otimes\exp_{N}(s\mathfrak{e}_{i}))\right|_{s=0}, b ˉ i ( g ) = − 2 1 d s d U ( g ⊗ exp N ( s e i )) s = 0 , and consider for every z 0 ∈ G N ( R d ) \mathbf{z}_{0} \in G^{N}(\mathbb{R}^{d}) z 0 ∈ G N ( R d ) the corresponding entire sig-diffusion with truncated signature-dependent coefficients:
d X t z 0 = b ˉ ( z 0 ⊗ X t z 0 ) d t + d W t , X t z 0 = z t ( 1 ) , (35) dX_{t}^{\mathbf{z}_{0}} = \bar{b}(\mathbf{z}_{0}\otimes X_{t}^{\mathbf{z}_{0}})\,dt + dW_{t}, \qquad X_{t}^{\mathbf{z}_{0}} = z_{t}^{(1)}, \tag*{(35)} d X t z 0 = b ˉ ( z 0 ⊗ X t z 0 ) d t + d W t , X t z 0 = z t ( 1 ) , (35) where W W W is a d d d -dimensional Brownian motion, and the lifted equation
Z t ≤ N = z 0 ⊗ X t z 0 . (36) Z_{t}^{\leq N} = \mathbf{z}_{0}\otimes X_{t}^{\mathbf{z}_{0}}. \tag*{(36)} Z t ≤ N = z 0 ⊗ X t z 0 . (36) Then Z ≤ N Z^{\leq N} Z ≤ N admits a unique global strong solution and
μ ( d g ) = e − U ( g ) m G ( d g ) ∫ G N ( R d ) e − U ( g ) m G ( d g ) (37) \mu(d\mathbf{g}) = \frac{e^{-U(\mathbf{g})}m_{G}(d\mathbf{g})}{\int_{G^{N}(\mathbb{R}^{d})} e^{-U(\mathbf{g})}m_{G}(d\mathbf{g})} \tag*{(37)} μ ( d g ) = ∫ G N ( R d ) e − U ( g ) m G ( d g ) e − U ( g ) m G ( d g ) (37) is the unique invariant probability measure such that
lim t → ∞ ∥ p t ( ⋅ ∣ z 0 ) − μ ∥ T V = 0 , for all z 0 ∈ G N ( R d ) , \lim_{t\to\infty} \lVert p_{t}(\cdot\mid\mathbf{z}_{0})-\mu\rVert_{\mathrm{TV}} = 0, \qquad\text{for all } \mathbf{z}_{0} \in G^{N}(\mathbb{R}^{d}), t → ∞ lim ∥ p t ( ⋅ ∣ z 0 ) − μ ∥ TV = 0 , for all z 0 ∈ G N ( R d ) , where p t ( ⋅ ∣ z 0 ) p_{t}(\cdot\mid\mathbf{z}_{0}) p t ( ⋅ ∣ z 0 ) denotes the transition density of Z t ≤ N Z_{t}^{\leq N} Z t ≤ N given z 0 \mathbf{z}_{0} z 0 and TV the total variation norm. In particular, the limiting distribution of Z t ≤ N Z_{t}^{\leq N} Z t ≤ N is independent of its initial condition z 0 \mathbf{z}_{0} z 0 .
Proof. We apply Proposition 3.1 of Baudoin et al. [7 ] to show global strong existence and uniqueness for the lifted diffusion Z ≤ N Z^{\leq N} Z ≤ N .
Denote by V 1 , … , V d V_{1},\ldots,V_{d} V 1 , … , V d the left-invariant vector fields on G N ( R d ) G^{N}(\mathbb{R}^{d}) G N ( R d ) given by
V i ( g ) = d d s ( g ⊗ exp N ( s e i ) ) ∣ s = 0 = g ⊗ e i . V_{i}(\mathbf{g}) = \left.\frac{d}{ds}(\mathbf{g}\otimes\exp_{N}(s\mathfrak{e}_{i}))\right|_{s=0} = \mathbf{g}\otimes\mathfrak{e}_{i}. V i ( g ) = d s d ( g ⊗ exp N ( s e i )) s = 0 = g ⊗ e i .
Consequently, when acting on a smooth function f : G N ( R d ) → R f:G^{N}(\mathbb{R}^{d})\to\mathbb{R} f : G N ( R d ) → R , we have
V i f ( g ) = d d s f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 . V_i f(\mathbf{g})=\left.\frac{d}{ds}f\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_i)\right)\right|_{s=0}. V i f ( g ) = d s d f ( g ⊗ exp N ( s ϵ i ) ) s = 0 .
The lifted diffusion Z t ≤ N Z_t^{\leq N} Z t ≤ N can now be written in terms of these vector fields, i.e.,
d Z t ≤ N = ∑ i = 1 d b ‾ i ( Z t ≤ N ) ( Z t ≤ N ⊗ ϵ i ) d t + ∑ i = 1 d V i ( Z t ≤ N ) ∘ d W t i = − 1 2 ∑ i = 1 d ( V i U ) ( Z t ≤ N ) V i ( Z t ≤ N ) d t + ∑ i = 1 d V i ( Z t ≤ N ) ∘ d W t i . dZ_t^{\leq N}=\sum_{i=1}^{d}\overline{b}_i(Z_t^{\leq N})(Z_t^{\leq N}\otimes\epsilon_i)\,dt+\sum_{i=1}^{d}V_i(Z_t^{\leq N})\circ dW_t^i
=\!-\frac{1}{2}\sum_{i=1}^{d}(V_iU)(Z_t^{\leq N})V_i(Z_t^{\leq N})\,dt+\sum_{i=1}^{d}V_i(Z_t^{\leq N})\circ dW_t^i. d Z t ≤ N = i = 1 ∑ d b i ( Z t ≤ N ) ( Z t ≤ N ⊗ ϵ i ) d t + i = 1 ∑ d V i ( Z t ≤ N ) ∘ d W t i = − 2 1 i = 1 ∑ d ( V i U ) ( Z t ≤ N ) V i ( Z t ≤ N ) d t + i = 1 ∑ d V i ( Z t ≤ N ) ∘ d W t i .
and its generator reads
L U f ( g ) = 1 2 ∑ i = 1 d V i 2 f ( g ) − 1 2 ∑ i = 1 d ( V i U ) V i f ( g ) = 1 2 ∑ i = 1 d d 2 d s 2 f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 − 1 2 ∑ i = 1 d d d s U ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 d d s f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 . (38) \begin{aligned}
\mathcal{L}^{U}f(\mathbf{g})=\frac{1}{2}\sum_{i=1}^{d}V_i^2f(\mathbf{g})-\frac{1}{2}\sum_{i=1}^{d}(V_iU)V_if(\mathbf{g}) \\
&=\frac{1}{2}\sum_{i=1}^{d}\left.\frac{d^2}{ds^2}f\left(\mathbf{g}\otimes\exp_N(s\epsilon_i)\right)\right|_{s=0}-\frac{1}{2}\sum_{i=1}^{d}\left.\frac{d}{ds}U\left(\mathbf{g}\otimes\exp_N(s\epsilon_i)\right)\right|_{s=0}\left.\frac{d}{ds}f\left(\mathbf{g}\otimes\exp_N(s\epsilon_i)\right)\right|_{s=0}. \tag*{(38)}
\end{aligned} L U f ( g ) = 2 1 i = 1 ∑ d V i 2 f ( g ) − 2 1 i = 1 ∑ d ( V i U ) V i f ( g ) = 2 1 i = 1 ∑ d d s 2 d 2 f ( g ⊗ exp N ( s ϵ i ) ) s = 0 − 2 1 i = 1 ∑ d d s d U ( g ⊗ exp N ( s ϵ i ) ) s = 0 d s d f ( g ⊗ exp N ( s ϵ i ) ) s = 0 . (38)
Setting L = 1 2 ∑ i = 1 d V i 2 \mathcal{L}=\frac{1}{2}\sum_{i=1}^{d}V_i^2 L = 2 1 ∑ i = 1 d V i 2 and Γ ( f , g ) = ∑ i = 1 d ( V i f ) ( V i g ) \Gamma(f,g)=\sum_{i=1}^{d}(V_if)(V_ig) Γ ( f , g ) = ∑ i = 1 d ( V i f ) ( V i g ) , we have L U = L − 1 2 Γ ( U , ⋅ ) \mathcal{L}^{U}=\mathcal{L}-\frac{1}{2}\Gamma(U,\cdot) L U = L − 2 1 Γ ( U , ⋅ ) . From this, we see that the lifted diffusion Z ≤ N Z^{\leq N} Z ≤ N is precisely of the form considered in Proposition 3.1 of Baudin et al. (2008). Note that the conditions on U U U are also satisfied, since it is entire and ∫ G N ( R d ) e − U ( g ) m G ( d g ) < ∞ \int_{G^{N}(\mathbb{R}^{d})}e^{-U(\mathbf{g})}\,m_G(d\mathbf{g})<\infty ∫ G N ( R d ) e − U ( g ) m G ( d g ) < ∞ by assumption. Thus it remains to verify that the Hörmander conditions of the diffusion vector fields V 1 , … , V d V_1,\ldots,V_d V 1 , … , V d are satisfied. To this end note that the Lie bracket of two such vector fields satisfies
[ V i , V j ] ( g ) = g ⊗ ϵ i ⊗ ϵ j − g ⊗ ϵ j ⊗ ϵ i = g ⊗ [ ϵ i , ϵ j ] . [V_i,V_j](\mathbf{g})=\mathbf{g}\otimes\epsilon_i\otimes\epsilon_j-\mathbf{g}\otimes\epsilon_j\otimes\epsilon_i=\mathbf{g}\otimes[\epsilon_i,\epsilon_j]. [ V i , V j ] ( g ) = g ⊗ ϵ i ⊗ ϵ j − g ⊗ ϵ j ⊗ ϵ i = g ⊗ [ ϵ i , ϵ j ] .
By induction, every iterated Lie bracket of V 1 , … , V d V_1,\ldots,V_d V 1 , … , V d is obtained by left multiplication of the corresponding iterated Lie bracket of ϵ 1 , … , ϵ d \epsilon_1,\ldots,\epsilon_d ϵ 1 , … , ϵ d , i.e.,
[ V i 1 , [ V i 2 , … , [ V i k − 1 , V i k ] … ] ] ( g ) = g ⊗ [ ϵ i 1 , [ ϵ i 2 , … , [ ϵ i k − 1 , ϵ i k ] … ] ] . [V_{i_1},[V_{i_2},\ldots,[V_{i_{k-1}},V_{i_k}]\ldots]](\mathbf{g})=\mathbf{g}\otimes[\epsilon_{i_1},[\epsilon_{i_2},\ldots,[\epsilon_{i_{k-1}},\epsilon_{i_k}]\ldots]]. [ V i 1 , [ V i 2 , … , [ V i k − 1 , V i k ] … ]] ( g ) = g ⊗ [ ϵ i 1 , [ ϵ i 2 , … , [ ϵ i k − 1 , ϵ i k ] … ]] .
Let now g N ( R d ) \mathfrak{g}^{N}(\mathbb{R}^{d}) g N ( R d ) be the free step-N N N nilpotent Lie algebra generated by ϵ 1 , … , ϵ d \epsilon_1,\ldots,\epsilon_d ϵ 1 , … , ϵ d , which in particular implies that these iterated Lie brackets of lengths at most N N N span g N ( R d ) \mathfrak{g}^{N}(\mathbb{R}^{d}) g N ( R d ) . Moreover, left multiplication by g \mathbf{g} g identifies the Lie algebra with the tangent space at g \mathbf{g} g , that is, T g G N ( R d ) = g ⊗ g N ( R d ) T_{\mathbf{g}}G^{N}(\mathbb{R}^{d})=\mathbf{g}\otimes\mathfrak{g}^{N}(\mathbb{R}^{d}) T g G N ( R d ) = g ⊗ g N ( R d ) . Consequently, the strong Hörmander bracket condition of the diffusion vector fields
span { V i ( g ) , [ V i , V j ] ( g ) , [ V i , [ V j , V k ] ] ( g ) , … } = T g G N ( R d ) \operatorname{span}\{V_i(\mathbf{g}),\ [V_i,V_j](\mathbf{g}),\ [V_i,[V_j,V_k]](\mathbf{g}),\ldots\}=T_{\mathbf{g}}G^{N}(\mathbb{R}^{d}) span { V i ( g ) , [ V i , V j ] ( g ) , [ V i , [ V j , V k ]] ( g ) , … } = T g G N ( R d )
holds for every g ∈ G N ( R d ) \mathbf{g}\in G^{N}(\mathbb{R}^{d}) g ∈ G N ( R d ) , where brackets of length at most N N N are sufficient. Therefore all assumptions of Proposition 3.1 of Baudin et al. (2008) are satisfied, which together with the Lipschitz property of the vector fields yields a unique global strong solution for Z ≤ N Z^{\leq N} Z ≤ N for every initial condition z 0 ∈ G N ( R d ) z_0\in G^{N}(\mathbb{R}^{d}) z 0 ∈ G N ( R d ) .
Concerning the invariant measure, we here give the full argument even though the first part, namely that μ \mu μ is an invariant measure can be deduced from Section 2 in Baudin et al. (2008). We start by showing that the generator L U \mathcal{L}^{U} L U is formally symmetric in L 2 ( μ ) L^2(\mu) L 2 ( μ ) . Let f , φ ∈ C c ∞ ( G N ( R d ) ) f,\varphi\in C_c^\infty(G^{N}(\mathbb{R}^{d})) f , φ ∈ C c ∞ ( G N ( R d )) . Recall that the Haar measure m G m_G m G is invariant under right ⊗ \otimes ⊗ multiplications. Consequently, for every F ∈ C c ∞ ( G N ( R d ) ) F\in C_c^\infty(G^{N}(\mathbb{R}^{d})) F ∈ C c ∞ ( G N ( R d )) and i = 1 , … , d i=1,\ldots,d i = 1 , … , d ,
∫ G N ( R d ) d d s F ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 d m G ( g ) = 0. (39) \int_{G^{N}(\mathbb{R}^{d})}\left.\frac{d}{ds}F\left(\mathbf{g}\otimes\exp_N(s\epsilon_i)\right)\right|_{s=0}\,dm_G(\mathbf{g})=0. \tag*{(39)} ∫ G N ( R d ) d s d F ( g ⊗ exp N ( s ϵ i ) ) s = 0 d m G ( g ) = 0. (39)
For fixed i i i , define
F i ( g ) = f ( g ) e − U ( g ) d d s φ ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 . F_i(\mathbf{g})=f(\mathbf{g})e^{-U(\mathbf{g})}\left.\frac{d}{ds}\varphi\left(\mathbf{g}\otimes\exp_N(s\varepsilon_i)\right)\right|_{s=0}. F i ( g ) = f ( g ) e − U ( g ) d s d φ ( g ⊗ exp N ( s ε i ) ) s = 0 .
Since f f f has compact support, we have F i ∈ C c ∞ ( G N ( R d ) ) F_i\in C^\infty_c(G^N(\mathbb{R}^d)) F i ∈ C c ∞ ( G N ( R d )) . Applying (39) to F i F_i F i gives
0 = ∫ G N ( R d ) e − U ( g ) [ d d s f ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 d d s φ ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 + f ( g ) d 2 d s 2 φ ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 \begin{aligned}
0=\int_{G^N(\mathbb{R}^d)}e^{-U(\mathbf{g})}[\frac{d}{ds}f(\mathbf{g}\otimes\exp_N(s\varepsilon_i))|_{s=0}\frac{d}{ds}\varphi(\mathbf{g}\otimes\exp_N(s\varepsilon_i))|_{s=0} \\
&+\left.f(\mathbf{g})\frac{d^2}{ds^2}\varphi\left(\mathbf{g}\otimes\exp_N(s\varepsilon_i)\right)\right|_{s=0}
\end{aligned} 0 = ∫ G N ( R d ) e − U ( g ) [ d s d f ( g ⊗ exp N ( s ε i )) ∣ s = 0 d s d φ ( g ⊗ exp N ( s ε i )) ∣ s = 0 + f ( g ) d s 2 d 2 φ ( g ⊗ exp N ( s ε i ) ) s = 0
− f ( g ) d d s U ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 d d s φ ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 ] d m G ( g ) . -f(\mathbf{g})\frac{d}{ds}U(\mathbf{g}\otimes\exp_N(s\varepsilon_i))|_{s=0}\frac{d}{ds}\varphi(\mathbf{g}\otimes\exp_N(s\varepsilon_i))|_{s=0}]\,dm_G(\mathbf{g}). − f ( g ) d s d U ( g ⊗ exp N ( s ε i )) ∣ s = 0 d s d φ ( g ⊗ exp N ( s ε i )) ∣ s = 0 ] d m G ( g ) .
Summing over i = 1 , … , d i=1,\ldots,d i = 1 , … , d , dividing by 2 ∫ G N ( R d ) e − U ( g ) m G ( d g ) 2\int_{G^N(\mathbb{R}^d)}e^{-U(\mathbf{g})}\,m_G(d\mathbf{g}) 2 ∫ G N ( R d ) e − U ( g ) m G ( d g ) , and recalling (37) we obtain, by the expression (38) for the generator,
∫ G N ( R d ) f ( g ) L U φ ( g ) μ ( d g ) = − 1 2 ∑ i = 1 d ∫ G N ( R d ) d d s f ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 d d s φ ( g ⊗ exp N ( s ε i ) ) ∣ s = 0 μ ( d g ) . \int_{G^N(\mathbb{R}^d)}f(\mathbf{g})\mathcal{L}^U\varphi(\mathbf{g})\,\mu(d\mathbf{g})
=-\frac{1}{2}\sum_{i=1}^{d}\int_{G^N(\mathbb{R}^d)}\left.\frac{d}{ds}f\left(\mathbf{g}\otimes\exp_N(s\varepsilon_i)\right)\right|_{s=0}\left.\frac{d}{ds}\varphi\left(\mathbf{g}\otimes\exp_N(s\varepsilon_i)\right)\right|_{s=0}\,\mu(d\mathbf{g}). ∫ G N ( R d ) f ( g ) L U φ ( g ) μ ( d g ) = − 2 1 i = 1 ∑ d ∫ G N ( R d ) d s d f ( g ⊗ exp N ( s ε i ) ) s = 0 d s d φ ( g ⊗ exp N ( s ε i ) ) s = 0 μ ( d g ) .
Since the right-hand side is symmetric in f f f and φ \varphi φ , it follows that
E ( f , φ ) : = − ∫ G N ( R d ) f ( g ) L U φ ( g ) μ ( d g ) = − ∫ G N ( R d ) φ ( g ) L U f ( g ) μ ( d g ) . \mathcal{E}(f,\varphi):=-\int_{G^N(\mathbb{R}^d)}f(\mathbf{g})\mathcal{L}^U\varphi(\mathbf{g})\,\mu(d\mathbf{g})=-\int_{G^N(\mathbb{R}^d)}\varphi(\mathbf{g})\mathcal{L}^Uf(\mathbf{g})\,\mu(d\mathbf{g}). E ( f , φ ) := − ∫ G N ( R d ) f ( g ) L U φ ( g ) μ ( d g ) = − ∫ G N ( R d ) φ ( g ) L U f ( g ) μ ( d g ) .
Thus, L U \mathcal{L}^U L U is formally symmetric in L 2 ( μ ) L^2(\mu) L 2 ( μ ) . Since μ \mu μ has a smooth, strictly positive density with respect to Haar measure, the bilinear form E ( f , φ ) \mathcal{E}(f,\varphi) E ( f , φ ) is closable in L 2 ( μ ) L^2(\mu) L 2 ( μ ) , and its closure is a regular, strongly local Dirichlet form. By local uniqueness of the martingale problem, the diffusion associated to this Dirichlet form coincides with Z N \mathbb{Z}^N Z N up to its lifetime. Since we have already established the existence of a global solution, it is thus conservative. Consequently, its transition semigroup ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 is symmetric in L 2 ( μ ) L^2(\mu) L 2 ( μ ) and satisfies P t 1 = 1 P_t1=1 P t 1 = 1 (which is also proved in Proposition 3.1 of Baudoin et al. [7 ] ). Hence
∫ P t f d μ = ⟨ P t f , 1 ⟩ L 2 ( μ ) = ⟨ f , P t 1 ⟩ L 2 ( μ ) = ∫ f d μ . \int P_tf\,d\mu=\langle P_tf,1\rangle_{L^2(\mu)}=\langle f,P_t1\rangle_{L^2(\mu)}=\int f\,d\mu. ∫ P t f d μ = ⟨ P t f , 1 ⟩ L 2 ( μ ) = ⟨ f , P t 1 ⟩ L 2 ( μ ) = ∫ f d μ .
Consequently, μ \mu μ is an invariant probability measure for Z N \mathbb{Z}^N Z N .
We now prove that μ \mu μ is unique and that
lim t → ∞ ∥ p t ( ⋅ ∣ z 0 ) − μ ∥ T V = 0 , \lim_{t\to\infty}\|p_t(\cdot\mid z_0)-\mu\|_{\mathrm{TV}}=0, t → ∞ lim ∥ p t ( ⋅ ∣ z 0 ) − μ ∥ TV = 0 ,
for every z 0 ∈ G N ( R d ) z_0\in G^N(\mathbb{R}^d) z 0 ∈ G N ( R d ) . To this end we shall apply a version of Doob’s ergodic theorem Doob [29 ] as proved in Gerlach and Nittka [39 ] . To verify the conditions of Proposition 4.3(ii) therein, we first show the strong Feller property of the transition semigroup ( P t ) t ≥ 0 i . e . , t h a t i t m a p s b o u n d e d m e a s u r a b l e f u n c t i o n s o n (P_t)_{t\geq0} i.e., that it maps bounded measurable functions on ( P t ) t ≥ 0 i . e . , t ha t i t ma p s b o u n d e d m e a s u r ab l e f u n c t i o n so n G^N(\mathbb{R}^d)$ to bounded continuous functions on G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) . Since the diffusion vector fields satisfy the strong Hörmander bracket condition as proved above, Theorem 7.1 of Arnaudon and Thalmaier [4 ] (see also p. 1286 therein) then implies that the transition semigroup is strong Feller.
Proposition 4.3(ii) in Gerlach and Nittka [39 ] now allows us to conclude L 1 ( μ ) L^1(\mu) L 1 ( μ ) irreducibility in the sense of Definition 2.4 therein.4 First, G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) is a connected, finite-dimensional Lie group and hence a second-countable Polish space. Moreover, the invariant probability measure μ ( d g ) \mu(\mathrm{d}\mathbf{g}) μ ( d g ) is strictly positive in the sense that μ ( O ) > 0 \mu(O)>0 μ ( O ) > 0 for every nonempty open subset O ⊂ G N ( R d ) O\subset G^N(\mathbb{R}^d) O ⊂ G N ( R d ) , since e − U ( g ) > 0 e^{-U(\mathbf{g})}>0 e − U ( g ) > 0 for every g \mathbf{g} g , and the Haar measure has full support. Since ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 is conservative, P t 1 = 1 P_t1=1 P t 1 = 1 for all t ≥ 0 t\geq0 t ≥ 0 . Thus, the constant function 1 1 1 is a strictly positive fixed point of ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 . Consequently, all assumptions of [39 Proposition 4.3(ii] are satisfied and it follows that the extension of ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 to L 1 ( μ ) L^1(\mu) L 1 ( μ ) is irreducible.
Uniqueness of the invariant probability measure and convergence to equilibrium now follow from Theorems 4.6 and 4.10 in Gerlach and Nittka [39 ] . Note that the semigroup is stochastically continuous because Z ≤ N Z^{\leq N} Z ≤ N has continuous sample paths. In particular, for every f ∈ C b ( G N ( R d ) ) f\in C_b(G^N(\mathbb{R}^d)) f ∈ C b ( G N ( R d )) ,
lim t ↓ 0 P t f ( g ) = f ( g ) \lim_{t\downarrow0} P_t f(\mathbf{g})=f(\mathbf{g}) t ↓ 0 lim P t f ( g ) = f ( g )
for every g ∈ G N ( R d ) \mathbf{g}\in G^N(\mathbb{R}^d) g ∈ G N ( R d ) . By dominated convergence,
lim t ↓ 0 ∫ ∣ P t f − f ∣ d μ = 0. \lim_{t\downarrow0}\int\lvert P_t f-f\rvert\,\mathrm{d}\mu=0. t ↓ 0 lim ∫ ∣ P t f − f ∣ d μ = 0.
Using the density of C b ( G N ( R d ) ) C_b(G^N(\mathbb{R}^d)) C b ( G N ( R d )) in L 1 ( μ ) L^1(\mu) L 1 ( μ ) and the fact that ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 acts as a contraction semigroup on L 1 ( μ ) L^1(\mu) L 1 ( μ ) , the same convergence holds for every f ∈ L 1 ( μ ) f\in L^1(\mu) f ∈ L 1 ( μ ) . Thus, the continuity assumption of Theorem 4.6 in Gerlach and Nittka [39 ] is also satisfied. Moreover, its transition probabilities are Borel probability measures so that in the notation of Gerlach and Nittka [39 ] P t ′ δ z 0 = p t ( ⋅ ∣ z 0 ) P'_t\delta_{\mathbf{z}_0}=p_t(\cdot\mid\mathbf{z}_0) P t ′ δ z 0 = p t ( ⋅ ∣ z 0 ) is a Borel measure. With e = 1 e=1 e = 1 , Theorem 4.6 in Gerlach and Nittka [39 ] therefore yields
lim t → ∞ P t f ( g ) = ∫ G N ( R d ) f ( h ) μ ( d h ) \lim_{t\to\infty}P_t f(\mathbf{g})=\int_{G^N(\mathbb{R}^d)}f(\mathbf{h})\,\mu(\mathrm{d}\mathbf{h}) t → ∞ lim P t f ( g ) = ∫ G N ( R d ) f ( h ) μ ( d h )
for every bounded measurable function f f f , uniformly on compact sets of G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) . Finally, by Theorem 4.10 in Gerlach and Nittka [39 ] , this convergence can be strengthened to convergence in total variation:
lim t → ∞ ∥ p t ( ⋅ ∣ z 0 ) − μ ∥ T V = 0 \lim_{t\to\infty}\lVert p_t(\cdot\mid\mathbf{z}_0)-\mu\rVert_{\mathrm{TV}}=0 t → ∞ lim ∥ p t ( ⋅ ∣ z 0 ) − μ ∥ TV = 0
for every z 0 ∈ G N ( R d ) \mathbf{z}_0\in G^N(\mathbb{R}^d) z 0 ∈ G N ( R d ) . In particular, μ \mu μ is the unique invariant probability measure, and the limiting distribution of Z t ≤ N Z_t^{\leq N} Z t ≤ N is independent of its initial condition z 0 \mathbf{z}_0 z 0 . □ \square □
In the following, we provide an example of a quadratic potential U U U that gives rise to a Gaussian-type Gibbs measure on G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) as the invariant measure.
Example 4.12. Fix z 0 , m ∈ G N ( R d ) \mathbf{z}_0,\mathbf{m}\in G^N(\mathbb{R}^d) z 0 , m ∈ G N ( R d ) and set U ( g ) = 1 2 ∥ g − m ∥ 2 U(\mathbf{g})=\frac{1}{2}\lVert\mathbf{g}-\mathbf{m}\rVert^2 U ( g ) = 2 1 ∥ g − m ∥ 2 . Observe that the corresponding entire sig-diffusion (35) reads
d X t z 0 , i = − 1 2 ( z 0 ⊗ X t z 0 − m ) π ≤ N ( ( z 0 ⊗ X t z 0 ) ⊗ ϵ i ) d t + d W t i , i = 1 , … , d , \mathrm{d}X_t^{\mathbf{z}_0,i}=-\frac{1}{2}\left(\mathbf{z}_0\otimes X_t^{\mathbf{z}_0}-\mathbf{m}\right)_{\pi\leq N}\left((\mathbf{z}_0\otimes X_t^{\mathbf{z}_0})\otimes\epsilon_i\right)\mathrm{d}t+\mathrm{d}W_t^i,\qquad i=1,\ldots,d, d X t z 0 , i = − 2 1 ( z 0 ⊗ X t z 0 − m ) π ≤ N ( ( z 0 ⊗ X t z 0 ) ⊗ ϵ i ) d t + d W t i , i = 1 , … , d ,
where X z 0 X^{\mathbf{z}_0} X z 0 denotes the signature of X z 0 X^{\mathbf{z}_0} X z 0 truncated at level N N N . Then, by Theorem 4.11 the process z 0 ⊗ X t z 0 \mathbf{z}_0\otimes X_t^{\mathbf{z}_0} z 0 ⊗ X t z 0 admits the unique invariant law μ \mu μ given by (37) which reads
μ ( d g ) = e − ∥ g − m ∥ 2 / 2 m G ( d g ) ∫ G N ( R d ) e − ∥ g − m ∥ 2 / 2 m G ( d g ) . \mu(\mathrm{d}\mathbf{g})=\frac{e^{-\lVert\mathbf{g}-\mathbf{m}\rVert^2/2}m_G(\mathrm{d}\mathbf{g})}{\int_{G^N(\mathbb{R}^d)}e^{-\lVert\mathbf{g}-\mathbf{m}\rVert^2/2}m_G(\mathrm{d}\mathbf{g})}. μ ( d g ) = ∫ G N ( R d ) e − ∥ g − m ∥ 2 /2 m G ( d g ) e − ∥ g − m ∥ 2 /2 m G ( d g ) .
Remark 4.13. It is important to consider the Markovian lifts of genuine sig-(jump)-diffusions (with, in particular, coefficients depending on the whole truncated signature) rather than restricting attention to lifts of classical Markovian ones.
Indeed, consider a classical Markovian SDE
d X t = b ( X t ) d t + σ ( X t ) d W t , X 0 = z 0 ( 1 ) , \mathrm{d}X_t=b(X_t)\,\mathrm{d}t+\sigma(X_t)\,\mathrm{d}W_t,\qquad X_0=\mathbf{z}_0^{(1)}, d X t = b ( X t ) d t + σ ( X t ) d W t , X 0 = z 0 ( 1 ) ,
with z 0 ∈ G N ( R d ) \mathbf{z}_0 \in G^N(\mathbb{R}^d) z 0 ∈ G N ( R d ) , and note that its dynamics can equivalently be written as
d X t z 0 = b ( ( z 0 ⊗ X t z 0 ) ( 1 ) ) d t + σ ( ( z 0 ⊗ X t z 0 ) ( 1 ) ) d W t , X 0 z 0 = z 0 ( 1 ) . (40) dX_t^{\mathbf{z}_0}=b((\mathbf{z}_0\otimes X_t^{\mathbf{z}_0})^{(1)})\,dt+\sigma((\mathbf{z}_0\otimes X_t^{\mathbf{z}_0})^{(1)})\,dW_t,\qquad X_0^{\mathbf{z}_0}=z_0^{(1)}. \tag*{(40)} d X t z 0 = b (( z 0 ⊗ X t z 0 ) ( 1 ) ) d t + σ (( z 0 ⊗ X t z 0 ) ( 1 ) ) d W t , X 0 z 0 = z 0 ( 1 ) . (40)
Consider then its signature lift Z t z 0 : = z 0 ⊗ X t z 0 \mathbb{Z}_t^{\mathbf{z}_0}:=\mathbf{z}_0\otimes X_t^{\mathbf{z}_0} Z t z 0 := z 0 ⊗ X t z 0 , truncated at level N N N . Recall that we omit ≤ N \leq N ≤ N for readability. Note that for every g ∈ G N ( R d ) \mathbf{g}\in G^N(\mathbb{R}^d) g ∈ G N ( R d ) with g ( 1 ) = 0 \mathbf{g}^{(1)}=0 g ( 1 ) = 0 we get Z t g ⊗ z 0 = g ⊗ Z t z 0 \mathbb{Z}_t^{\mathbf{g}\otimes\mathbf{z}_0}=\mathbf{g}\otimes\mathbb{Z}_t^{\mathbf{z}_0} Z t g ⊗ z 0 = g ⊗ Z t z 0 . This follows because ( g ⊗ z 0 ) ( 1 ) = z 0 ( 1 ) (\mathbf{g}\otimes\mathbf{z}_0)^{(1)}=z_0^{(1)} ( g ⊗ z 0 ) ( 1 ) = z 0 ( 1 ) , so that the underlying Markovian SDE (40) has the same initial condition in both cases.
This implies that if the law of Z t z 0 \mathbb{Z}_t^{\mathbf{z}_0} Z t z 0 converges in law to an invariant probability distribution μ \mu μ , then the law of Z t g ⊗ z 0 \mathbb{Z}_t^{\mathbf{g}\otimes\mathbf{z}_0} Z t g ⊗ z 0 converges in law to
μ g ( A ) : = ∫ 1 { g ⊗ ξ ∈ A } μ ( d ξ ) \mu^{\mathbf{g}}(A):=\int\mathbf{1}_{\{\mathbf{g}\otimes\xi\in A\}}\,\mu(d\xi) μ g ( A ) := ∫ 1 { g ⊗ ξ ∈ A } μ ( d ξ )
for all measurable sets A ⊆ G N ( R d ) A\subseteq G^N(\mathbb{R}^d) A ⊆ G N ( R d ) . Since the noncompactness of G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) prevents μ g \mu^{\mathbf{g}} μ g from corresponding to μ \mu μ for each g \mathbf{g} g , we can conclude that the law of Z t g ⊗ z 0 \mathbb{Z}_t^{\mathbf{g}\otimes\mathbf{z}_0} Z t g ⊗ z 0 depends on g \mathbf{g} g also for t t t converging to infinity. Hence, forgetting the initial condition, which is necessary for generative diffusion modeling, is in this case not possible.
In contrast, signature-dependent coefficients generally break this equivariance, allowing for mean reversion in higher signature levels and the existence of a unique invariant probability measure independent of the initial condition, as demonstrated by the preceding theorem.
If one wishes to consider signature-like lifts of classical Markovian SDE, an alternative is to work with the process Z ~ t z 0 \widetilde{\mathbb{Z}}_t^{\mathbf{z}_0} Z t z 0 defined by
Z ~ t z 0 = D t λ ( z 0 ⊗ X ~ t z 0 ) , \widetilde{\mathbb{Z}}_t^{\mathbf{z}_0}=D_t^\lambda(\mathbf{z}_0\otimes\widetilde{X}_t^{\mathbf{z}_0}), Z t z 0 = D t λ ( z 0 ⊗ X t z 0 ) ,
where X ~ z 0 \widetilde{X}^{\mathbf{z}_0} X z 0 is the signature of ( ∫ 0 t e λ u d X u z 0 ) t ∈ [ 0 , T ] \left(\int_0^t e^{\lambda u}dX_u^{\mathbf{z}_0}\right)_{t\in[0,T]} ( ∫ 0 t e λ u d X u z 0 ) t ∈ [ 0 , T ] and D t λ D_t^\lambda D t λ is the dilation operator defined by ⟨ ϵ I , D t λ u ⟩ = e − λ t ∣ I ∣ u I \langle\epsilon_I,D_t^\lambda u\rangle=e^{-\lambda t|I|}u_I ⟨ ϵ I , D t λ u ⟩ = e − λ t ∣ I ∣ u I . Replacing z 0 \mathbf{z}_0 z 0 by the truncated signature z 0 λ \mathbf{z}_0^\lambda z 0 λ of the path
( ∫ 0 r e − λ ( S − u ) d X − S + u ) r ∈ [ 0 , S ] ′ , \left(\int_0^r e^{-\lambda(S-u)}dX_{-S+u}\right)_{r\in[0,S]}', ( ∫ 0 r e − λ ( S − u ) d X − S + u ) r ∈ [ 0 , S ] ′ ,
the process Z ~ t z 0 λ \widetilde{\mathbb{Z}}_t^{\mathbf{z}_0^\lambda} Z t z 0 λ corresponds to the exponentially fading memory signature of ( X t ) t ∈ [ − S , T ] (X_t)_{t\in[-S,T]} ( X t ) t ∈ [ − S , T ] studied in Abi Jaber and Sotnikov [3 ] , where the relevant ergodic properties are also established for X t = ( t , W t ) X_t=(t,W_t) X t = ( t , W t ) . Since D t λ D_t^\lambda D t λ is a deterministic linear operator on T N ( R d ) T^N(\mathbb{R}^d) T N ( R d ) that preserves G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) , the analytical methods developed in this section can still be applied.
Next, we turn to the computation of the generalized score function of Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N in this specific setting. Consider Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N given by (36) . Our aim here is to identify the generalized score function of Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N in the sense of Definition 4.6, that is in the case when K N \mathcal{K}^N K N is the second-order operator defined with respect to the ambient truncated tensor algebra T N ( R d ) T^N(\mathbb{R}^d) T N ( R d ) .
Recall that the dynamics of Z t ≤ N \mathbb{Z}_t^{\leq N} Z t ≤ N are given by
d Z t ≤ N = ( Z t ≤ N ⊗ b ˉ ( Z t ≤ N ) + 1 2 ∑ i = 1 d Z t ≤ N ⊗ ε i ⊗ 2 ) d t + ∑ i = 1 d Z t ≤ N ⊗ ε i d W t i . d\mathbb{Z}_t^{\leq N}=\left(\mathbb{Z}_t^{\leq N}\otimes\bar{b}(\mathbb{Z}_t^{\leq N})+\frac{1}{2}\sum_{i=1}^{d}\mathbb{Z}_t^{\leq N}\otimes\varepsilon_i^{\otimes2}\right)dt+\sum_{i=1}^{d}\mathbb{Z}_t^{\leq N}\otimes\varepsilon_i\,dW_t^i. d Z t ≤ N = ( Z t ≤ N ⊗ b ˉ ( Z t ≤ N ) + 2 1 i = 1 ∑ d Z t ≤ N ⊗ ε i ⊗ 2 ) d t + i = 1 ∑ d Z t ≤ N ⊗ ε i d W t i .
Hence, its drift and diffusion coefficients are
b ( x ) = x ⊗ b ˉ ( x ) + 1 2 ∑ i = 1 d x ⊗ ε i ⊗ 2 and σ i ( x ) = x ⊗ ε i , b(\mathbf{x})=\mathbf{x}\otimes\bar{b}(\mathbf{x})+\frac{1}{2}\sum_{i=1}^{d}\mathbf{x}\otimes\varepsilon_i^{\otimes2}\qquad\text{and}\qquad\sigma_i(\mathbf{x})=\mathbf{x}\otimes\varepsilon_i, b ( x ) = x ⊗ b ˉ ( x ) + 2 1 i = 1 ∑ d x ⊗ ε i ⊗ 2 and σ i ( x ) = x ⊗ ε i ,
respectively. Consequently, the generator decomposes as L f ( x ) = ⟨ b ( x ) , ∇ f ( x ) ⟩ + K N f ( x ) Lf(\mathbf{x})=\langle b(\mathbf{x}),\nabla f(\mathbf{x})\rangle+\mathcal{K}^Nf(\mathbf{x}) L f ( x ) = ⟨ b ( x ) , ∇ f ( x )⟩ + K N f ( x ) , for
K N f ( x ) = 1 2 ∑ i = 1 d ∇ 2 f ( x ) [ x ⊗ ε i , x ⊗ ε i ] . \mathcal{K}^Nf(\mathbf{x})=\frac{1}{2}\sum_{i=1}^{d}\nabla^2f(\mathbf{x})[\mathbf{x}\otimes\varepsilon_i,\mathbf{x}\otimes\varepsilon_i]. K N f ( x ) = 2 1 i = 1 ∑ d ∇ 2 f ( x ) [ x ⊗ ε i , x ⊗ ε i ] .
Here ∇ f \nabla f ∇ f and ∇ 2 f \nabla^2 f ∇ 2 f denote the gradient and Hessian in the ambient tensor algebra T N ( R d ) T^N(\mathbb{R}^d) T N ( R d ) and in particular
∇ 2 f ( x ) [ h , g ] = ∑ ∣ I ∣ , ∣ J ∣ ≤ N ∂ x I x J 2 f ( x ) h I g J . \nabla^2 f(\mathbf{x})[\mathbf{h},\mathbf{g}] = \sum_{|I|,|J|\leq N} \partial^2_{\mathbf{x}_I\mathbf{x}_J}f(\mathbf{x})\mathbf{h}_I\mathbf{g}_J. ∇ 2 f ( x ) [ h , g ] = ∣ I ∣ , ∣ J ∣ ≤ N ∑ ∂ x I x J 2 f ( x ) h I g J .
Remark 4.14. It is important to distinguish K N \mathcal{K}^N K N from the second-order operator defined by differentiation along the signature group. Indeed, for fixed g ∈ G N ( R d ) \mathbf{g}\in G^N(\mathbb{R}^d) g ∈ G N ( R d ) and i = 1 , … , d i=1,\ldots,d i = 1 , … , d , consider the curve
h i ( s ) = g ⊗ exp N ( s ϵ i ) . h_i(s)=\mathbf{g}\otimes\exp_N(s\epsilon_i). h i ( s ) = g ⊗ exp N ( s ϵ i ) .
We have
h i ′ ( 0 ) = g ⊗ ϵ i , h i ′ ′ ( 0 ) = g ⊗ ϵ i ⊗ 2 . (41) h_i'(0)=\mathbf{g}\otimes\epsilon_i,\qquad h_i''(0)=\mathbf{g}\otimes\epsilon_i^{\otimes2}. \tag*{(41)} h i ′ ( 0 ) = g ⊗ ϵ i , h i ′′ ( 0 ) = g ⊗ ϵ i ⊗ 2 . (41)
By the second-order chain rule,
d 2 d s 2 f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 = ∇ 2 f ( g ) [ g ⊗ ϵ i , g ⊗ ϵ i ] + ⟨ ∇ f ( g ) , g ⊗ ϵ i ⊗ 2 ⟩ . \left.\frac{d^2}{ds^2}f\bigl(\mathbf{g}\otimes\exp_N(s\epsilon_i)\bigr)\right|_{s=0}
=\nabla^2 f(\mathbf{g})[\mathbf{g}\otimes\epsilon_i,\mathbf{g}\otimes\epsilon_i]
+\left\langle\nabla f(\mathbf{g}),\mathbf{g}\otimes\epsilon_i^{\otimes2}\right\rangle. d s 2 d 2 f ( g ⊗ exp N ( s ϵ i ) ) s = 0 = ∇ 2 f ( g ) [ g ⊗ ϵ i , g ⊗ ϵ i ] + ⟨ ∇ f ( g ) , g ⊗ ϵ i ⊗ 2 ⟩ .
Consequently,
K N f ( g ) = 1 2 ∑ i = 1 d d 2 d s 2 f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 − 1 2 ∑ i = 1 d ⟨ ∇ f ( g ) , g ⊗ ϵ i ⊗ 2 ⟩ . (42) \mathcal{K}^Nf(\mathbf{g})=\frac{1}{2}\sum_{i=1}^{d}\left.\frac{d^2}{ds^2}f\bigl(\mathbf{g}\otimes\exp_N(s\epsilon_i)\bigr)\right|_{s=0}
-\frac{1}{2}\sum_{i=1}^{d}\left\langle\nabla f(\mathbf{g}),\mathbf{g}\otimes\epsilon_i^{\otimes2}\right\rangle. \tag*{(42)} K N f ( g ) = 2 1 i = 1 ∑ d d s 2 d 2 f ( g ⊗ exp N ( s ϵ i ) ) s = 0 − 2 1 i = 1 ∑ d ⟨ ∇ f ( g ) , g ⊗ ϵ i ⊗ 2 ⟩ . (42)
Thus, the intrinsic second-order operator and the second-order part of the ambient Itô generator differ by a first-order correction term.
Let P t N = L ( Z t ≤ N ) P_t^N=\mathcal{L}(\mathbb{Z}_t^{\leq N}) P t N = L ( Z t ≤ N ) denote the law of the lifted process at time t t t . Since the diffusion vector fields D 1 , … , D d D_1,\ldots,D_d D 1 , … , D d satisfy Hörmander’s bracket-generating condition as proved in Theorem 4.11 , the transition law admits a smooth density with respect to Haar measure due to the standard hypoellipticity assumptions. Moreover, they define a bracket-generating sub-Riemannian structure on G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) . Furthermore, G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) is connected and complete with respect to the associated Carnot–Carathéodory distance. Since the generator can be written as
L f = 1 2 e U ∑ i = 1 d D i ( e − U D i f ) , Lf=\frac{1}{2}e^U\sum_{i=1}^{d}D_i(e^{-U}D_if), L f = 2 1 e U i = 1 ∑ d D i ( e − U D i f ) ,
it is a weighted sub-Laplacian associated with the smooth, strictly positive invariant measure μ \mu μ . By the strict positivity of heat kernels for such sub-Laplacians on connected, complete sub-Riemannian manifolds, Strichartz [68 ] and Barilari et al. [6 ] , the transition density satisfies
p t ( g ∣ z 0 ) > 0 , ∀ t > 0 , z 0 , g ∈ G N ( R d ) . p_t(\mathbf{g}\mid\mathbf{z}_0)>0,\qquad\forall t>0,\quad\mathbf{z}_0,\mathbf{g}\in G^N(\mathbb{R}^d). p t ( g ∣ z 0 ) > 0 , ∀ t > 0 , z 0 , g ∈ G N ( R d ) .
Hence, we can write
P t N ( d g ) = p t ( g ) m G ( d g ) , P_t^N(d\mathbf{g})=p_t(\mathbf{g})\,m_G(d\mathbf{g}), P t N ( d g ) = p t ( g ) m G ( d g ) ,
with p t ( g ) = ∫ G N ( R d ) p t ( g ∣ z 0 ) p d a t a ( d z 0 ) p_t(\mathbf{g})=\int_{G^N(\mathbb{R}^d)}p_t(\mathbf{g}\mid\mathbf{z}_0)p_{\mathrm{data}}(d\mathbf{z}_0) p t ( g ) = ∫ G N ( R d ) p t ( g ∣ z 0 ) p data ( d z 0 ) smooth and strictly positive. Using this, the following result now establishes the form of the generalized score function of Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N given by (36) .
Proposition 4.15. Let t > 0 t>0 t > 0 and suppose that the map S t N : G N ( R d ) → T N ( R d ) S_t^N:G^N(\mathbb{R}^d)\to T^N(\mathbb{R}^d) S t N : G N ( R d ) → T N ( R d ) , defined by
S t N ( g ) = 1 2 ∑ i = 1 d [ d d s log p t ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 ( g ⊗ ϵ i ) + g ⊗ ϵ i ⊗ 2 ] , (43) S_t^N(\mathbf{g})=\frac{1}{2}\sum_{i=1}^{d}\left[\left.\frac{d}{ds}\log p_t\bigl(\mathbf{g}\otimes\exp_N(s\epsilon_i)\bigr)\right|_{s=0}\bigl(\mathbf{g}\otimes\epsilon_i\bigr)+\mathbf{g}\otimes\epsilon_i^{\otimes2}\right], \tag*{(43)} S t N ( g ) = 2 1 i = 1 ∑ d [ d s d log p t ( g ⊗ exp N ( s ϵ i ) ) s = 0 ( g ⊗ ϵ i ) + g ⊗ ϵ i ⊗ 2 ] , (43)
is integrable with respect to P t N P_t^N P t N . Then S t N S_t^N S t N is a generalized score function of Z ≤ N \mathbb{Z}^{\leq N} Z ≤ N , i.e.,
∫ K N f ( g ) P t N ( d g ) = − ∫ ⟨ S t N ( g ) , ∇ f ( g ) ⟩ P t N ( d g ) (44) \int\mathcal{K}^Nf(\mathbf{g})\,P_t^N(d\mathbf{g})=-\int\left\langle S_t^N(\mathbf{g}),\nabla f(\mathbf{g})\right\rangle P_t^N(d\mathbf{g}) \tag*{(44)} ∫ K N f ( g ) P t N ( d g ) = − ∫ ⟨ S t N ( g ) , ∇ f ( g ) ⟩ P t N ( d g ) (44)
for every f ∈ C c 2 ( T N ( R d ) ) f\in C_c^2(T^N(\mathbb{R}^d)) f ∈ C c 2 ( T N ( R d )) .
Proof. By (42) ,
∫ K N f d P t N = 1 2 ∑ i = 1 d ∫ d 2 d s 2 f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 p t ( g ) m G ( d g ) − 1 2 ∑ i = 1 d ∫ ⟨ ∇ f ( g ) , g ⊗ ϵ i ⊗ 2 ⟩ p t ( g ) m G ( d g ) . \int\mathcal{K}^{N} f\,dP_{t}^{N}=\frac{1}{2}\sum_{i=1}^{d}\int\left.\frac{d^{2}}{ds^{2}}f\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}p_{t}(\mathbf{g})\,m_{G}(d\mathbf{g})-\frac{1}{2}\sum_{i=1}^{d}\int\left\langle\nabla f(\mathbf{g}),\mathbf{g}\otimes\epsilon_{i}^{\otimes2}\right\rangle p_{t}(\mathbf{g})\,m_{G}(d\mathbf{g}). ∫ K N f d P t N = 2 1 i = 1 ∑ d ∫ d s 2 d 2 f ( g ⊗ exp N ( s ϵ i ) ) s = 0 p t ( g ) m G ( d g ) − 2 1 i = 1 ∑ d ∫ ⟨ ∇ f ( g ) , g ⊗ ϵ i ⊗ 2 ⟩ p t ( g ) m G ( d g ) .
Since the Haar measure is invariant under right ⊗ \otimes ⊗ multiplications, integration by parts gives
∫ d 2 d s 2 f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 p t ( g ) m G ( d g ) = − ∫ d d s f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 d d s p t ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 m G ( d g ) . \int\left.\frac{d^{2}}{ds^{2}}f\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}p_{t}(\mathbf{g})\,m_{G}(d\mathbf{g})
=-\int\left.\frac{d}{ds}f\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}\left.\frac{d}{ds}p_{t}\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}m_{G}(d\mathbf{g}). ∫ d s 2 d 2 f ( g ⊗ exp N ( s ϵ i ) ) s = 0 p t ( g ) m G ( d g ) = − ∫ d s d f ( g ⊗ exp N ( s ϵ i ) ) s = 0 d s d p t ( g ⊗ exp N ( s ϵ i ) ) s = 0 m G ( d g ) .
Moreover, the first-order chain rule yields
d d s f ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 = ⟨ ∇ f ( g ) , g ⊗ ϵ i ⟩ . (45) \left.\frac{d}{ds}f\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}=\left\langle\nabla f(\mathbf{g}),\mathbf{g}\otimes\epsilon_{i}\right\rangle. \tag*{(45)} d s d f ( g ⊗ exp N ( s ϵ i ) ) s = 0 = ⟨ ∇ f ( g ) , g ⊗ ϵ i ⟩ . (45)
Since p t > 0 p_{t}>0 p t > 0 , we also have
d d s p t ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 = p t ( g ) d d s log p t ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 . \left.\frac{d}{ds}p_{t}\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}=p_{t}(\mathbf{g})\left.\frac{d}{ds}\log p_{t}\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}. d s d p t ( g ⊗ exp N ( s ϵ i ) ) s = 0 = p t ( g ) d s d log p t ( g ⊗ exp N ( s ϵ i ) ) s = 0 .
Combining these identities, we obtain
∫ K N f d P t N = − 1 2 ∑ i = 1 d ∫ ⟨ ∇ f ( g ) , d d s log p t ( g ⊗ exp N ( s ϵ i ) ) ∣ s = 0 ( g ⊗ ϵ i ) + g ⊗ ϵ i ⊗ 2 ⟩ P t N ( d g ) . \int\mathcal{K}^{N}f\,dP_{t}^{N}=-\frac{1}{2}\sum_{i=1}^{d}\int\left\langle\nabla f(\mathbf{g}),\left.\frac{d}{ds}\log p_{t}\left(\mathbf{g}\otimes\exp_{N}(s\epsilon_{i})\right)\right|_{s=0}\left(\mathbf{g}\otimes\epsilon_{i}\right)+\mathbf{g}\otimes\epsilon_{i}^{\otimes2}\right\rangle P_{t}^{N}(d\mathbf{g}). ∫ K N f d P t N = − 2 1 i = 1 ∑ d ∫ ⟨ ∇ f ( g ) , d s d log p t ( g ⊗ exp N ( s ϵ i ) ) s = 0 ( g ⊗ ϵ i ) + g ⊗ ϵ i ⊗ 2 ⟩ P t N ( d g ) .
By the definition of the score in (43) , this is precisely ∫ K N f d P t N = − ∫ ⟨ S t N , ∇ f ⟩ d P t N \int\mathcal{K}^{N}f\,dP_{t}^{N}=-\int\langle S_{t}^{N},\nabla f\rangle dP_{t}^{N} ∫ K N f d P t N = − ∫ ⟨ S t N , ∇ f ⟩ d P t N , which proves the claim. □
Remark 4.16. The generalized score in (43) is defined P t N P_{t}^{N} P t N -almost everywhere on G N ( R d ) G^{N}(\mathbb{R}^{d}) G N ( R d ) . Since P t N P_{t}^{N} P t N is supported on the group, it can be extended measurably to T N ( R d ) T^{N}(\mathbb{R}^{d}) T N ( R d ) without changing (44) . The integrability condition in the proposition is an additional assumption and does not follow immediately from the existence of an invariant measure.
An alternative approach for polynomial diffusions based on inverse signature matching
In this section we provide an alternative to score matching based on what we call “inverse signature matching”, tailored to polynomial diffusion processes for which we can exploit that the expected signature is obtained by solving a finite-dimensional linear ODE. Here, we focus on generating new data in line with some law p d a t a p_{\mathrm{data}} p data on R d \mathbb{R}^{d} R d and not on time-series generation via their signature as in the previous sections.
The goal is to use a polynomial diffusion process to perturb the data z 0 ∼ p d a t a z_{0}\sim p_{\mathrm{data}} z 0 ∼ p data on R d \mathbb{R}^{d} R d and to generate new data by the corresponding reverse process. In contrast to the approach provided in Remark 4.10 (ii), we here do not directly learn the score by a polynomial or an entire function but, exploiting that, under suitable conditions, the expected signature characterizes the law of a stochastic process Chevalier and Oberhauser (2022), we rather match the inverse signature process of the forward polynomial diffusion.
More precisely, we assume that the forward process is given by a polynomial diffusion in the sense of Section 3.3, i.e.,
d X t = b ˉ ( X t ) d t + a ˉ ( X t ) d W t , X 0 = z 0 ∼ p d a t a , (46) dX_{t}=\bar{b}(X_{t})dt+\sqrt{\bar{a}(X_{t})}dW_{t},\qquad X_{0}=z_{0}\sim p_{\mathrm{data}}, \tag*{(46)} d X t = b ˉ ( X t ) d t + a ˉ ( X t ) d W t , X 0 = z 0 ∼ p data , (46)
where b ˉ : R d → R d \bar{b}: \mathbb{R}^{d} \to\mathbb{R}^{d} b ˉ : R d → R d and a ˉ : R d → S + d \bar{a}: \mathbb{R}^{d} \to\mathbb{S}_{+}^{d} a ˉ : R d → S + d have polynomial components of degree at most one and two, respectively, and W W W is a Brownian motion with respect to a filtration ( F t ) (\mathcal{F}_{t}) ( F t ) to which X X X is adapted. We assume that a ˉ \sqrt{\bar{a}} a ˉ have polynomial components of degree at most one, and that, for every t > 0 t>0 t > 0 , X t X_t X t admits a strictly positive density p t p_t p t , with p ∈ C 1 , 2 ( ( 0 , ∞ ) × R d ) p \in C^{1,2}((0,\infty) \times\mathbb{R}^{d}) p ∈ C 1 , 2 (( 0 , ∞ ) × R d ) . We further assume that X X X has a unique invariant density p ∞ p_{\infty} p ∞ and that its law converges weakly to p ∞ ( x ) d x p_{\infty}(x)dx p ∞ ( x ) d x from every deterministic initial state.
Let Y t : = X T − t Y_t := X_{T-t} Y t := X T − t , 0 ≤ t ≤ T 0 \le t \le T 0 ≤ t ≤ T , be the reversed process. By Theorem 2.1 in Haussmann and Pardoux [43 ] , Y Y Y satisfies the following SDE for 0 ≤ t < T 0 \le t < T 0 ≤ t < T :
d Y t = ( − b ˉ ( Y t ) + div ( a ˉ ( Y t ) ) + a ˉ ( Y t ) ∇ log p T − t ( Y t ) ) d t + a ˉ ( Y t ) d B t , Y 0 = X T ∼ p T , (47) dY_t = \left(-\bar{b}(Y_t) + \operatorname{div}(\bar{a}(Y_t)) + \bar{a}(Y_t)\nabla\log p_{T-t}(Y_t)\right)dt + \sqrt{\bar{a}(Y_t)}dB_t,\qquad Y_0 = X_T \sim p_T, \tag*{(47)} d Y t = ( − b ˉ ( Y t ) + div ( a ˉ ( Y t )) + a ˉ ( Y t ) ∇ log p T − t ( Y t ) ) d t + a ˉ ( Y t ) d B t , Y 0 = X T ∼ p T , (47)
where div \operatorname{div} div is defined in (20) and B B B is a Brownian motion with respect to a filtration ( F ^ t ) (\hat{\mathcal{F}}_t) ( F ^ t ) to which Y Y Y is adapted, on an enlargement of the probability space if necessary. Under the stated assumptions, the coefficients of the reverse SDE are locally Lipschitz in space, uniformly on compact time intervals contained in [ 0 , T ) [0,T) [ 0 , T ) . Consequently, pathwise uniqueness holds and Y Y Y is a strong solution on [ 0 , T ) [0,T) [ 0 , T ) . Moreover, Y Y Y extends continuously to T T T , with Y T = X 0 ∼ p d a t a Y_T = X_0 \sim p_{\mathrm{data}} Y T = X 0 ∼ p data .
Note that the drift of (47) corresponds to choosing η ≡ 1 \eta\equiv1 η ≡ 1 in (21) since 1 2 ( div ( a ˉ ) + a ˉ ∇ log p T − t ) \frac{1}{2}(\operatorname{div}(\bar{a})+\bar{a}\nabla\log p_{T-t}) 2 1 ( div ( a ˉ ) + a ˉ ∇ log p T − t ) is exactly the generalized score as shown in Remark 4.2.
As above, the goal here is to learn the map a ˉ ( ⋅ ) ∇ log p ( ⋅ ) \bar{a}(\cdot)\nabla\log p(\cdot) a ˉ ( ⋅ ) ∇ log p ( ⋅ ) from the available data distributed according to p d a t a p_{\mathrm{data}} p data and then to generate new instances of x 0 x_0 x 0 by sampling Y 0 Y_0 Y 0 from the invariant measure and computing Y T Y_T Y T by numerically solving the SDE (47) . Here the main difference is that, relying on Proposition 4.19 below and the results in Chevyrev and Oberhauser [18 ] , the unknown term a ˉ ∇ log p ( ⋅ ) \bar{a}\nabla\log p(\cdot) a ˉ ∇ log p ( ⋅ ) is learned by matching the expected signature of the reverse process to that of the inverse signature of the forward polynomial process, obtained by solving a finite-dimensional linear ODE. The approach is made precise in the following steps.
(i) Signature parametrization of the score. We assume that each component of ∇ log p ( ⋅ ) \nabla\log p(\cdot) ∇ log p ( ⋅ ) can be approximated with a polynomial (in both variables). Moreover, we use that Y t ≈ u ~ ( Y 0 , B ^ t ) Y_t \approx\tilde{u}(Y_0,\hat{\mathbb{B}}_t) Y t ≈ u ~ ( Y 0 , B ^ t ) , for some u ~ ( Y 0 ) ∈ T N ( R d + 1 ) d \tilde{u}(Y_0) \in T^N(\mathbb{R}^{d+1})^d u ~ ( Y 0 ) ∈ T N ( R d + 1 ) d and some N ∈ N N \in\mathbb{N} N ∈ N , where B ^ \hat{\mathbb{B}} B ^ is the Stratonovich signature of the time-extended Brownian motion B ^ \hat{B} B ^ , whose index corresponding to the time component is denoted by 0 0 0 . The proposed approximation is motivated by Lyons’ universal limit theorem Lyons [54 ] , which ensures, under suitable regularity assumptions, that almost surely the solution map ( B ^ t ≤ 2 ) t ∈ [ 0 , T ] ↦ ( Y t ) t ∈ [ 0 , T ] (\hat{\mathbb{B}}_t^{\leq2})_{t\in[0,T]} \mapsto(Y_t)_{t\in[0,T]} ( B ^ t ≤ 2 ) t ∈ [ 0 , T ] ↦ ( Y t ) t ∈ [ 0 , T ] is continuous with respect to the appropriate rough-path topologies. Hence, by the universal approximation theorem for signatures Levin et al. [51 ] ; ); Cuchiero et al. [25 ] ; ); Cuchiero and Möller [23 ] , the whole trajectory ( Y t ) t ∈ [ 0 , T ] (Y_t)_{t\in[0,T]} ( Y t ) t ∈ [ 0 , T ] can be uniformly approximated on compact sets of paths by a linear combination of the time-extended signature of B B B . Under these assumptions and as a ˉ \bar{a} a ˉ is a polynomial of degree at most 2, we get by the shuffle property
a ˉ ( Y t ) ∇ log p T − t ( Y t ) ≈ ⟨ u ( Y 0 ) , B ^ t ⟩ , (48) \bar{a}(Y_t)\nabla\log p_{T-t}(Y_t) \approx\langle\mathbf{u}(Y_0),\hat{\mathbb{B}}_t\rangle, \tag*{(48)} a ˉ ( Y t ) ∇ log p T − t ( Y t ) ≈ ⟨ u ( Y 0 ) , B ^ t ⟩ , (48)
for some u ( Y 0 ) ∈ T N ( R d + 1 ) d \mathbf{u}(Y_0) \in T^N(\mathbb{R}^{d+1})^d u ( Y 0 ) ∈ T N ( R d + 1 ) d . Indeed, after substituting the signature representation of Y t Y_t Y t , every polynomial in Y t Y_t Y t becomes a polynomial in the components of B ^ t \hat{\mathbb{B}}_t B ^ t . By repeated application of the shuffle product, such a polynomial can be rewritten as a linear functional of the signature, possibly at a higher truncation level.
(ii) Parametrized reverse process and its expected signature. By assumption (i), one can now define
d Y t u = ( − b ˉ ( Y t u ) + div ( a ˉ ( Y t u ) ) + ⟨ u ( Y 0 u ) , B ^ t ⟩ ) d t + a ˉ ( Y t u ) d B t , Y 0 u = X T . dY_t^{\mathbf{u}} = \left(-\bar{b}(Y_t^{\mathbf{u}}) + \operatorname{div}(\bar{a}(Y_t^{\mathbf{u}})) + \langle\mathbf{u}(Y_0^{\mathbf{u}}),\hat{\mathbb{B}}_t\rangle\right)dt + \sqrt{\bar{a}(Y_t^{\mathbf{u}})}dB_t,\qquad Y_0^{\mathbf{u}} = X_T. d Y t u = ( − b ˉ ( Y t u ) + div ( a ˉ ( Y t u )) + ⟨ u ( Y 0 u ) , B ^ t ⟩ ) d t + a ˉ ( Y t u ) d B t , Y 0 u = X T .
For every fixed y 0 y_0 y 0 with Y 0 u = y 0 Y_0^{\mathbf{u}} = y_0 Y 0 u = y 0 , the process Z : = ( Y u , B ^ ≤ N ) Z := (Y^{\mathbf{u}},\hat{\mathbb{B}}^{\leq N}) Z := ( Y u , B ^ ≤ N ) is a finite-dimensional polynomial process in the sense of Subsection 3.3.3 whose coefficients depend parametrically on y 0 y_0 y 0 . As explained there, this implies that its expected signature can be computed analytically, i.e.,
E [ Z T M ∣ Y 0 u ] = exp ( T L u ( Y 0 u ) ⊤ ) e ∅ \mathbb{E}[Z_T^M\mid Y_0^{\mathrm{u}}]=\exp\left(TL^{\mathrm{u}}(Y_0^{\mathrm{u}})^\top\right)e_{\varnothing} E [ Z T M ∣ Y 0 u ] = exp ( T L u ( Y 0 u ) ⊤ ) e ∅
where L u ( y ) L^{\mathrm{u}}(y) L u ( y ) denotes the generator matrix of the truncated signature at level M M M of the process Z Z Z starting at z 0 : = ( y , e ∅ ) z_0 := (y,e_{\varnothing}) z 0 := ( y , e ∅ ) (Corollary 3.25). In particular, this implies that
E [ ( Y T u ) ≤ M ∣ Y 0 u ] = f ( Y 0 u , u ( Y 0 u ) ) , \mathbb{E}[(Y_T^{\mathrm{u}})^{\le M}\mid Y_0^{\mathrm{u}}]=f(Y_0^{\mathrm{u}},\mathbf{u}(Y_0^{\mathrm{u}})), E [( Y T u ) ≤ M ∣ Y 0 u ] = f ( Y 0 u , u ( Y 0 u )) ,
for an explicit deterministic vector-valued function f f f given by
f ( y , u ) = ( ⟨ ϵ I , exp ( T L u ( y ) ⊤ ) e ∅ ⟩ ) I ∈ C , (49) f(y,\mathbf{u})=\left(\left\langle\epsilon_I,\exp\left(TL^{\mathrm{u}}(y)^\top\right)e_{\varnothing}\right\rangle\right)_{I\in\mathcal{C}}, \tag*{(49)} f ( y , u ) = ( ⟨ ϵ I , exp ( T L u ( y ) ⊤ ) e ∅ ⟩ ) I ∈ C , (49)
where C \mathcal{C} C is the set of all multiindices I I I such that ⟨ ϵ I , Z T ⟩ = ⟨ ϵ I , ( Y T u ) ≤ M ⟩ \langle\epsilon_I,Z_T\rangle=\langle\epsilon_I,(Y_T^{\mathrm{u}})^{\le M}\rangle ⟨ ϵ I , Z T ⟩ = ⟨ ϵ I , ( Y T u ) ≤ M ⟩ . Note that the columns of the matrix L u ( y ) L^{\mathrm{u}}(y) L u ( y ) are given in (3.14), with b u \mathbf{b}^{\mathrm{u}} b u and a \mathbf{a} a the sig-coefficients of the polynomial diffusion Z Z Z computed in Lemma 3.22 via the coefficients of its characteristics ((3.12) ( b u , a , 0 ) (\mathbf{b}^{\mathrm{u}},\mathbf{a},0) ( b u , a , 0 ) ), which read
b i u ( z ) = − b ˉ i ( π Y ( z ) ) + ∑ k = 1 d ∂ x k a ˉ i k ( π Y ( z ) ) + ⟨ u i , π B ^ ( z ) ⟩ , b_i^{\mathrm{u}}(z)=-\bar b_i(\pi_Y(z))+\sum_{k=1}^{d}\partial_{x_k}\bar a_{ik}(\pi_Y(z))+\left\langle\mathbf{u}_i,\pi_{\widehat{\mathbb{B}}}(z)\right\rangle, b i u ( z ) = − b ˉ i ( π Y ( z )) + k = 1 ∑ d ∂ x k a ˉ ik ( π Y ( z )) + ⟨ u i , π B ( z ) ⟩ ,
b I u ( z ) = 1 { i ∣ I ∣ = 0 } ⟨ ϵ I , π B ^ ( z ) ⟩ + 1 2 1 { i ∣ I ∣ − 1 = i ∣ I ∣ ≠ 0 } ⟨ ϵ I ′ , π B ^ ( z ) ⟩ , b_I^{\mathrm{u}}(z)=\mathbb{1}_{\{i_{|I|}=0\}}\left\langle\epsilon_I,\pi_{\widehat{\mathbb{B}}}(z)\right\rangle+\frac{1}{2}\mathbb{1}_{\{i_{|I|-1}=i_{|I|}\ne0\}}\left\langle\epsilon_{I'},\pi_{\widehat{\mathbb{B}}}(z)\right\rangle, b I u ( z ) = 1 { i ∣ I ∣ = 0 } ⟨ ϵ I , π B ( z ) ⟩ + 2 1 1 { i ∣ I ∣ − 1 = i ∣ I ∣ = 0 } ⟨ ϵ I ′ , π B ( z ) ⟩ ,
a i k ( z ) = a ˉ i k ( π Y ( z ) ) , a I I ( z ) = a I I ( z ) = 1 { i ∣ I ∣ ≠ 0 } σ ˉ i ∣ I ∣ , i ∣ I ∣ ( π Y ( z ) ) ⟨ ϵ I , π B ^ ( z ) ⟩ , a_{ik}(z)=\bar a_{ik}(\pi_Y(z)),\qquad a_{II}(z)=a_{II}(z)=\mathbb{1}_{\{i_{|I|}\ne0\}}\bar\sigma_{i_{|I|},i_{|I|}}(\pi_Y(z))\left\langle\epsilon_I,\pi_{\widehat{\mathbb{B}}}(z)\right\rangle, a ik ( z ) = a ˉ ik ( π Y ( z )) , a I I ( z ) = a I I ( z ) = 1 { i ∣ I ∣ = 0 } σ ˉ i ∣ I ∣ , i ∣ I ∣ ( π Y ( z )) ⟨ ϵ I , π B ( z ) ⟩ ,
a I J ( z ) = 1 { i ∣ I ∣ = i ∣ J ∣ ≠ 0 } ⟨ ϵ I ′ , π B ^ ( z ) ⟩ ⟨ ϵ J ′ , π B ^ ( z ) ⟩ , a_{IJ}(z)=\mathbb{1}_{\{i_{|I|}=i_{|J|}\ne0\}}\left\langle\epsilon_{I'},\pi_{\widehat{\mathbb{B}}}(z)\right\rangle\left\langle\epsilon_{J'},\pi_{\widehat{\mathbb{B}}}(z)\right\rangle, a I J ( z ) = 1 { i ∣ I ∣ = i ∣ J ∣ = 0 } ⟨ ϵ I ′ , π B ( z ) ⟩ ⟨ ϵ J ′ , π B ( z ) ⟩ ,
where π Y ( z ) \pi_Y(z) π Y ( z ) and π B ^ ( z ) \pi_{\widehat{\mathbb{B}}}(z) π B ( z ) denote the components of z z z corresponding to Y Y Y and B ^ ≤ N \widehat{\mathbb{B}}^{\le N} B ≤ N in Z Z Z , respectively, and σ ˉ = a ˉ \bar\sigma=\sqrt{\bar a} σ ˉ = a ˉ . Since by Lemma 3.22 the components of b u \mathbf{b}^{\mathrm{u}} b u and a \mathbf{a} a are equal to some partial derivative of b u b^{\mathrm{u}} b u and a \mathbf{a} a evaluated at z 0 = ( y , e ∅ ) z_0=(y,e_{\varnothing}) z 0 = ( y , e ∅ ) , respectively, and since b ˉ \bar b b ˉ and σ ˉ \bar\sigma σ ˉ are polynomials, we can conclude that the map y ↦ L u ( y ) y\mapsto L^{\mathrm{u}}(y) y ↦ L u ( y ) has polynomial components and thus that the components of f f f in (4.32) are entire in ( y , u ) (y,\mathbf{u}) ( y , u ) .
Alternatively, the assumptions of point (i) also justify a parametrization of the whole drift directly as a linear combination of the components of B ^ t \widehat{\mathbf{B}}_t B t , thus yielding
d Y ‾ t u = ⟨ u ‾ ( Y ‾ t u ) , B ^ t ⟩ d t + a ˉ ( Y ‾ t u ) d B t , Y ‾ 0 u = X T , d\overline{Y}_t^{\mathrm{u}}=\left\langle\overline{\mathbf{u}}(\overline{Y}_t^{\mathrm{u}}),\widehat{\mathbf{B}}_t\right\rangle dt+\sqrt{\bar a(\overline{Y}_t^{\mathrm{u}})}dB_t,\qquad\overline{Y}_0^{\mathrm{u}}=X_T, d Y t u = ⟨ u ( Y t u ) , B t ⟩ d t + a ˉ ( Y t u ) d B t , Y 0 u = X T ,
which is an even more tractable representation E [ ( Y ‾ T u ) ≤ M ∣ Y ‾ 0 u ] = f ~ ( Y ‾ 0 u , u ‾ ( Y ‾ 0 u ) ) \mathbb{E}[(\overline{Y}_T^{\mathrm{u}})^{\le M}\mid\overline{Y}_0^{\mathrm{u}}]=\widetilde f(\overline{Y}_0^{\mathrm{u}},\overline{\mathbf{u}}(\overline{Y}_0^{\mathrm{u}})) E [( Y T u ) ≤ M ∣ Y 0 u ] = f ( Y 0 u , u ( Y 0 u )) , but losing the specific structure of the drift of the reverse process.
(iii) Learning via inverse signature matching. We replace the map Y 0 ↦ u ( Y 0 ) Y_0\mapsto\mathbf{u}(Y_0) Y 0 ↦ u ( Y 0 ) in (4.31) with some parametric map u ϕ \mathbf{u}_{\phi} u ϕ . Our goal is now to learn ϕ \phi ϕ by matching the conditional expectations of Y T u ϕ Y_T^{\mathbf{u}_{\phi}} Y T u ϕ and Y T Y_T Y T given X T X_T X T , i.e.,
E [ ∥ E [ ( Y T u ϕ ) ≤ M ∣ X T ] − E [ Y T ≤ M ∣ X T ] ∥ 2 ] . (50) \mathbb{E}\left[\left\|\mathbb{E}[(Y_T^{\mathbf{u}_{\phi}})^{\le M}\mid X_T]-\mathbb{E}[Y_T^{\le M}\mid X_T]\right\|^2\right]. \tag*{(50)} E [ E [( Y T u ϕ ) ≤ M ∣ X T ] − E [ Y T ≤ M ∣ X T ] 2 ] . (50)
By the previous point, we know that E [ ( Y T u ϕ ) ≤ M ∣ X T ] = f ( X T , u ϕ ( X T ) ) \mathbb{E}[(Y_T^{\mathbf{u}_{\phi}})^{\le M}\mid X_T]=f(X_T,\mathbf{u}_{\phi}(X_T)) E [( Y T u ϕ ) ≤ M ∣ X T ] = f ( X T , u ϕ ( X T )) , while from Proposition 4.19 below, we get that Y T = X T − 1 Y_T=X_T^{-1} Y T = X T − 1 almost surely and thus
E [ Y T ≤ M ∣ X T ] = E [ ( X T − 1 ) ≤ M ∣ X T ] . \mathbb{E}[Y_T^{\le M}\mid X_T]=\mathbb{E}[(X_T^{-1})^{\le M}\mid X_T]. E [ Y T ≤ M ∣ X T ] = E [( X T − 1 ) ≤ M ∣ X T ] .
Here X T − 1 X_T^{-1} X T − 1 is defined as the unique element of T ( ( R d ) ) T((\mathbb{R}^d)) T (( R d )) such that X T − 1 ⊗ X T = e ∅ X_T^{-1}\otimes X_T=e_{\varnothing} X T − 1 ⊗ X T = e ∅ , namely
⟨ ϵ ( i 1 , … , i m ) , ( X T ) − 1 ⟩ = ( − 1 ) m ⟨ ϵ ( i m , … , i 1 ) , X T ⟩ . (51) \left\langle\epsilon_{(i_1,\ldots,i_m)},(X_T)^{-1}\right\rangle=(-1)^m\left\langle\epsilon_{(i_m,\ldots,i_1)},X_T\right\rangle. \tag*{(51)} ⟨ ϵ ( i 1 , … , i m ) , ( X T ) − 1 ⟩ = ( − 1 ) m ⟨ ϵ ( i m , … , i 1 ) , X T ⟩ . (51)
Observe that
E [ ∥ f ( X T , u ϕ ( X T ) ) − E [ ( X T − 1 ) ≤ M ∣ X T ] ∥ 2 ] = E [ ∥ f ( X T , u ϕ ( X T ) ) − ( X T − 1 ) ≤ M ∥ 2 ] − E [ ∥ ( X T − 1 ) ≤ M − E [ ( X T − 1 ) ≤ M ∣ X T ] ∥ 2 ] , \mathbb{E}\left[\left\|f(X_T,\mathbf{u}_{\phi}(X_T))-\mathbb{E}[(X_T^{-1})^{\le M}\mid X_T]\right\|^2\right]=\mathbb{E}\left[\left\|f(X_T,\mathbf{u}_{\phi}(X_T))-(X_T^{-1})^{\le M}\right\|^2\right]
-\mathbb{E}\left[\left\|(X_T^{-1})^{\le M}-\mathbb{E}[(X_T^{-1})^{\le M}\mid X_T]\right\|^2\right], E [ f ( X T , u ϕ ( X T )) − E [( X T − 1 ) ≤ M ∣ X T ] 2 ] = E [ f ( X T , u ϕ ( X T )) − ( X T − 1 ) ≤ M 2 ] − E [ ( X T − 1 ) ≤ M − E [( X T − 1 ) ≤ M ∣ X T ] 2 ] ,
showing that minimizing (50) with respect to ϕ \phi ϕ coincides with minimizing
E [ ∥ f ( X T , u ϕ ( X T ) ) − ( X T − 1 ) ≤ M ∥ 2 ] . \mathbb{E}\left[\left\|f\left(X_T,\mathbf{u}_{\phi}\left(X_T\right)\right)-\left(\mathbf{X}_T^{-1}\right)^{\leq M}\right\|^2\right]. E [ f ( X T , u ϕ ( X T ) ) − ( X T − 1 ) ≤ M 2 ] .
Letting x 0 ( 1 ) , … , x 0 ( K ) x_0^{(1)},\ldots,x_0^{(K)} x 0 ( 1 ) , … , x 0 ( K ) denote the available training instances and moving to the empirical expectation, we get
min ϕ 1 K ∑ m = 1 K E [ ∥ f ( X T , u ϕ ( X T ) ) − ( X T − 1 ) ≤ M ∥ 2 | X 0 = x 0 ( m ) ] . \min_{\phi}\frac{1}{K}\sum_{m=1}^{K}\mathbb{E}\left[\left\|f\left(X_T,\mathbf{u}_{\phi}\left(X_T\right)\right)-\left(\mathbf{X}_T^{-1}\right)^{\leq M}\right\|^2\middle|X_0=x_0^{(m)}\right]. ϕ min K 1 m = 1 ∑ K E [ f ( X T , u ϕ ( X T ) ) − ( X T − 1 ) ≤ M 2 X 0 = x 0 ( m ) ] .
By Jensen’s inequality, this condition can be relaxed to
min ϕ 1 K ∑ m = 1 K ∥ E [ f ( X T , u ϕ ( X T ) ) | X 0 = x 0 ( m ) ] − E [ ( X T − 1 ) ≤ M | X 0 = x 0 ( m ) ] ∥ 2 , (52) \min_{\phi}\frac{1}{K}\sum_{m=1}^{K}\left\|\mathbb{E}\left[f\left(X_T,\mathbf{u}_{\phi}\left(X_T\right)\right)\middle|X_0=x_0^{(m)}\right]-\mathbb{E}\left[\left(\mathbf{X}_T^{-1}\right)^{\leq M}\middle|X_0=x_0^{(m)}\right]\right\|^2, \tag*{(52)} ϕ min K 1 m = 1 ∑ K E [ f ( X T , u ϕ ( X T ) ) X 0 = x 0 ( m ) ] − E [ ( X T − 1 ) ≤ M X 0 = x 0 ( m ) ] 2 , (52)
which, as explained in (iv) below, has more tractable properties. When the full expected signature determines the conditional law of the reverse process Chevyrev and Oberhauser [18 ] and if the densities are regular enough, equality of the reverse path laws identifies the reverse drift and therefore the score.
(iv) Analytic expected inverse signature and reduction to a deterministic optimization problem. The second conditional expectation in (52) can be computed analytically. Indeed, by equation (51) , the (algebraic) inverse signature ( X T − 1 ) ≤ M (\mathbf{X}_T^{-1})^{\leq M} ( X T − 1 ) ≤ M is a linear transformation of X T ≤ M \mathbf{X}_T^{\leq M} X T ≤ M . Since the forward process X X X in (46) is a polynomial diffusion, the conditional expectation of its signature can be computed via the matrix-exponential formula for expected signatures established in Section 3.3.3 (Corollary 3.25). Note that the initial condition x 0 ( m ) x_0^{(m)} x 0 ( m ) enters in the computation of the generator matrix L L L of X ≤ M \mathbf{X}^{\leq M} X ≤ M via the sig-coefficients of X X X computed in Lemma 3.22 (equation (16) ).
The first conditional expectation in (52) can be approximated by Monte Carlo sampling from p T ( ⋅ ∣ x 0 ) p_T(\cdot\mid x_0) p T ( ⋅ ∣ x 0 ) . Alternatively, it admits an analytic expression whenever the map y ↦ u ϕ ( y ) y\mapsto\mathbf{u}_{\phi}(y) y ↦ u ϕ ( y ) is entire. Indeed, by the definition of f f f in (49) , the map ( y , u ) ↦ f ( y , u ) (y,\mathbf{u})\mapsto f(y,\mathbf{u}) ( y , u ) ↦ f ( y , u ) has entire components. Hence, the composition y ↦ f ( y , u ϕ ( y ) ) y\mapsto f(y,\mathbf{u}_{\phi}(y)) y ↦ f ( y , u ϕ ( y )) has entire components too. Therefore, provided the assumptions of Theorem 3.20 in Cuchiero et al. [26 ] are satisfied, the holomorphic formula established therein applies to X X X , which is a polynomial (and hence holomorphic) diffusion.
It follows that, if y ↦ u ϕ ( y ) y\mapsto\mathbf{u}_{\phi}(y) y ↦ u ϕ ( y ) is entire and Theorem 3.20 in Cuchiero et al. [26 ] applies, both conditional expectations in (52) admit analytic expressions, and the optimization problem (52) reduces to a deterministic one.
(v) Sample generation. Once ϕ \phi ϕ is learned, the map Y 0 ↦ u ϕ ( Y 0 ) Y_0\mapsto\mathbf{u}_{\phi}(Y_0) Y 0 ↦ u ϕ ( Y 0 ) and in turn (48) is fully specified. New samples can now be generated by drawing Y 0 Y_0 Y 0 from the limiting invariant distribution, evaluating u ϕ ( Y 0 ) \mathbf{u}_{\phi}(Y_0) u ϕ ( Y 0 ) , and numerically solving
d Y t u = ( − b ˉ ( Y t u ) + div ( a ˉ ( Y t u ) ) + ⟨ u ϕ ( Y 0 u ) , B ˙ ^ t ⟩ ) d t + a ˉ ( Y t u ) d B t . dY_t^{\mathbf{u}}=\left(-\bar{b}(Y_t^{\mathbf{u}})+\operatorname{div}\left(\bar{a}(Y_t^{\mathbf{u}})\right)+\left\langle\mathbf{u}_{\phi}(Y_0^{\mathbf{u}}),\widehat{\dot{B}}_t\right\rangle\right)dt+\sqrt{\bar{a}(Y_t^{\mathbf{u}})}dB_t. d Y t u = ( − b ˉ ( Y t u ) + div ( a ˉ ( Y t u ) ) + ⟨ u ϕ ( Y 0 u ) , B ˙ t ⟩ ) d t + a ˉ ( Y t u ) d B t .
The terminal value Y T u Y_T^{\mathbf{u}} Y T u is then taken as a generated approximation of a sample from p d a t a p_{\mathrm{data}} p data .
Remark 4.17. Note that if X X X is a polynomial process with constant diffusion coefficient, a ˉ ≡ c o n s t \bar{a}\equiv\mathrm{const} a ˉ ≡ const , as is the case for standard Ornstein–Uhlenbeck processes, then, recalling (48) , the above approach reduces to learning, via inverse signature matching, only the score map, parametrized as a linear function of the signature of the Brownian motion depending parametrically on Y 0 Y_0 Y 0 ,
∇ log p T − t ( Y t ) ≈ ⟨ u ϕ ( Y 0 ) , B ˙ ^ t ⟩ . \nabla\log p_{T-t}(Y_t)\approx\left\langle\mathbf{u}_{\phi}(Y_0),\widehat{\dot{B}}_t\right\rangle. ∇ log p T − t ( Y t ) ≈ ⟨ u ϕ ( Y 0 ) , B ˙ t ⟩ .
New samples are then generated by numerically simulating the reverse-time dynamics
d Y t u = ( − b ˉ ( Y t u ) + a ˉ ⟨ u ϕ ( Y 0 u ) , B ˙ ^ t ⟩ ) d t + a ˉ d B t . dY_t^{\mathbf{u}}=\left(-\bar{b}(Y_t^{\mathbf{u}})+\bar{a}\left\langle\mathbf{u}_{\phi}(Y_0^{\mathbf{u}}),\widehat{\dot{B}}_t\right\rangle\right)dt+\sqrt{\bar{a}}dB_t. d Y t u = ( − b ˉ ( Y t u ) + a ˉ ⟨ u ϕ ( Y 0 u ) , B ˙ t ⟩ ) d t + a ˉ d B t .
Remark 4.18. It is important to notice that the law of X T \mathbb{X}_T X T contains information about X 0 X_0 X 0 . Indeed, even if at time 0 it holds X 0 = e ∅ X_0 = e_{\varnothing} X 0 = e ∅ , we have that the law of X T \mathbb{X}_T X T corresponds to the law of the signature of the polynomial process X ( 1 ) = X − X 0 \mathbb{X}^{(1)} = X - X_0 X ( 1 ) = X − X 0 whose dynamics satisfy
d X t ( 1 ) = b ˉ ( X t ( 1 ) + X 0 ) d t + a ˉ ( X t ( 1 ) + X 0 ) d W t , X 0 = x 0 ∼ p d a t a , X 0 ( 1 ) = 0. d\mathbb{X}^{(1)}_t = \bar{b}(\mathbb{X}^{(1)}_t + X_0)dt + \sqrt{\bar{a}(\mathbb{X}^{(1)}_t + X_0)}dW_t,\quad X_0 = x_0 \sim p_{\mathrm{data}},\quad\mathbb{X}^{(1)}_0 = 0. d X t ( 1 ) = b ˉ ( X t ( 1 ) + X 0 ) d t + a ˉ ( X t ( 1 ) + X 0 ) d W t , X 0 = x 0 ∼ p data , X 0 ( 1 ) = 0.
One can see this also by noticing that (46) corresponds to (11) with truncation level 1 and some g \mathbf{g} g such that g ( 1 ) = X 0 \mathbf{g}^{(1)} = X_0 g ( 1 ) = X 0 . Indeed, with the notation of (11) and b ˉ \bar{b} b ˉ as in (46) , we have
b ˉ g ( x ( 1 ) ) = b ˉ g ( x − X 0 ) = b ˉ ( x ) = b ˉ ( x ( 1 ) + X 0 ) \bar{b}^{\mathbf{g}}(x^{(1)}) = \bar{b}^{\mathbf{g}}(x - X_0) = \bar{b}(x) = \bar{b}(x^{(1)} + X_0) b ˉ g ( x ( 1 ) ) = b ˉ g ( x − X 0 ) = b ˉ ( x ) = b ˉ ( x ( 1 ) + X 0 )
and analogously for the diffusion part. Therefore b ˉ ( x ( 1 ) + X 0 ) \bar{b}(x^{(1)} + X_0) b ˉ ( x ( 1 ) + X 0 ) and a ˉ ( x ( 1 ) + X 0 ) \bar{a}(x^{(1)} + X_0) a ˉ ( x ( 1 ) + X 0 ) enter in the dynamics of the signature process given in (12) .
The following proposition is needed to justify point (iii) in the above approach. Recall the definition of X T − 1 \mathbb{X}^{-1}_T X T − 1 given by (51) .
Proposition 4.19. Let Y t : = X T − t Y_t := X_{T-t} Y t := X T − t , t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] , and denote by X \mathbb{X} X and Y \mathbb{Y} Y the Stratonovich signatures of X X X and Y Y Y , respectively. Then Y T = X T − 1 \mathbb{Y}_T = \mathbb{X}^{-1}_T Y T = X T − 1 almost surely.
Proof. We proceed by verifying the identity on approximating processes and then moving to the limit. Let D n : = { k T 2 n : k = 0 , … , 2 n } D_n := \left\{\frac{kT}{2^n}: k = 0,\ldots,2^n\right\} D n := { 2 n k T : k = 0 , … , 2 n } , and let X n X^n X n be the piecewise linear interpolation of X X X along D n D_n D n . Define
Y t n : = X T − t n , t ∈ [ 0 , T ] . Y^n_t := X^n_{T-t},\qquad t \in[0,T]. Y t n := X T − t n , t ∈ [ 0 , T ] .
Since D n D_n D n is invariant under the map t ↦ T − t t \mapsto T-t t ↦ T − t , Y n Y^n Y n is the piecewise linear interpolation of Y Y Y along the same partition D n D_n D n . Let X n \mathbb{X}^n X n and Y n \mathbb{Y}^n Y n denote the signatures of X n X^n X n and Y n Y^n Y n , respectively, which are well defined as both of them are semimartingales with respect to their own filtration.
Since X n X^n X n is a continuous path of bounded variation and Y n Y^n Y n is the same path run backwards, Theorem 1.25 of Chevyrev and Kormilitzin [16 ] gives
Y T n = ( X T n ) − 1 a.s. (53) \mathbb{Y}^n_T = \left(\mathbb{X}^n_T\right)^{-1}\quad\text{a.s.} \tag*{(53)} Y T n = ( X T n ) − 1 a.s. (53)
Fix now N ∈ N N \in\mathbb{N} N ∈ N . By the Wong–Zakai approximation for continuous semimartingales, applied to the equations defining the truncated signatures, we get that ( X T n ) ≤ N (\mathbb{X}^n_T)^{\leq N} ( X T n ) ≤ N converges to X T ≤ N \mathbb{X}^{\leq N}_T X T ≤ N and ( Y T n ) ≤ N (\mathbb{Y}^n_T)^{\leq N} ( Y T n ) ≤ N to Y T ≤ N \mathbb{Y}^{\leq N}_T Y T ≤ N in probability (Chevyrev and Friz, [15 ] , Example 4.20). Since inversion is continuous on the finite-dimensional set G N ( R d ) G^N(\mathbb{R}^d) G N ( R d ) , we also have that ( ( X T n ) ≤ N ) − 1 \left((\mathbb{X}^n_T)^{\leq N}\right)^{-1} ( ( X T n ) ≤ N ) − 1 converges to ( X T ≤ N ) − 1 (\mathbb{X}^{\leq N}_T)^{-1} ( X T ≤ N ) − 1 in probability. As ( ( X T n ) − 1 ) ≤ N = ( ( X T n ) ≤ N ) − 1 \left((\mathbb{X}^n_T)^{-1}\right)^{\leq N} = \left((\mathbb{X}^n_T)^{\leq N}\right)^{-1} ( ( X T n ) − 1 ) ≤ N = ( ( X T n ) ≤ N ) − 1 , by (53) and uniqueness of the limit in probability, Y T ≤ N = ( X T ≤ N ) − 1 \mathbb{Y}^{\leq N}_T = (\mathbb{X}^{\leq N}_T)^{-1} Y T ≤ N = ( X T ≤ N ) − 1 almost surely. Since N N N is arbitrary, this identity holds at every tensor level. Consequently, Y T = X T − 1 \mathbb{Y}_T = \mathbb{X}^{-1}_T Y T = X T − 1 almost surely, where the equality is understood coordinatewise. □ \square □
Example 4.20. Note that in the case of a one-dimensional process X X X , we have that each term of the signature of Y Y Y at time T T T corresponds to an element of the form ( Y T − Y 0 ) k / k ! (Y_T - Y_0)^k/k! ( Y T − Y 0 ) k / k ! which, by definition of Y Y Y , corresponds to ( X 0 − X T ) k / k ! (X_0 - X_T)^k/k! ( X 0 − X T ) k / k ! . The correspondence between laws then follows directly.