Skip to main content Accessibility help
×
Home
Hostname: page-component-55597f9d44-rn2sj Total loading time: 2.423 Render date: 2022-08-08T14:18:58.156Z Has data issue: true Feature Flags: { "shouldUseShareProductTool": true, "shouldUseHypothesis": true, "isUnsiloEnabled": true, "useRatesEcommerce": false, "useNewApi": true } hasContentIssue true

Unified signature cumulants and generalized Magnus expansions

Published online by Cambridge University Press:  09 June 2022

Peter K. Friz
Affiliation:
Institut für Mathematik, Technische Universität Berlin, Str. des 17. Juni 136, Berlin10586,Germany; E-mail: friz@math.tu-berlin.de Weierstraß-Institut für Angewandte Analysis und Stochastik, Mohrenstr. 39, Berlin10117,Germany; E-mail: tapia@wias-berlin.de
Paul P. Hager
Affiliation:
Institut für Mathematik, Humboldt Universität zu Berlin, Unter den Linden 6, Berlin10099,Germany; E-mail: paul.hager@hu-berlin.de
Nikolas Tapia
Affiliation:
Institut für Mathematik, Technische Universität Berlin, Str. des 17. Juni 136, Berlin10586,Germany; E-mail: friz@math.tu-berlin.de Weierstraß-Institut für Angewandte Analysis und Stochastik, Mohrenstr. 39, Berlin10117,Germany; E-mail: tapia@wias-berlin.de

Abstract

The signature of a path can be described as its full non-commutative exponential. Following T. Lyons, we regard its expectation, the expected signature, as a path space analogue of the classical moment generating function. The logarithm thereof, taken in the tensor algebra, defines the signature cumulant. We establish a universal functional relation in a general semimartingale context. Our work exhibits the importance of Magnus expansions in the algorithmic problem of computing expected signature cumulants and further offers a far-reaching generalization of recent results on characteristic exponents dubbed diamond and cumulant expansions with motivations ranging from financial mathematics to statistical physics. From an affine semimartingale perspective, the functional relation may be interpreted as a type of generalized Riccati equation.

Type
Computational Mathematics
Creative Commons
Creative Common License - CCCreative Common License - BY
This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution, and reproduction in any medium, provided the original work is properly cited.
Copyright
© The Author(s), 2022. Published by Cambridge University Press

1 Introduction and main results

Write $\mathcal {T} := T\mathopen {(\mkern -3mu(}\mathbb {R}^d\mathclose {)\mkern -3mu)} = \Pi _{k \ge 0} (\mathbb {R}^d)^{\otimes k}$ for the tensor series over $\mathbb {R}^d$ , equipped with a concatenation product, elements of which are written indifferently as

$$ \begin{align*}\mathbf{x} = (\mathbf{x}^{(0)},\mathbf{x}^{(1)},\mathbf{x}^{(2)},\dotsc) \equiv \mathbf{x}^{(0)}+ \mathbf{x}^{(1)}+ \mathbf{x}^{(2)}+\dotsb. \end{align*} $$

The affine subspace $\mathcal {T}_0$ (respectively, $\mathcal {T}_1$ ) with scalar component $\mathbf {x}^{(0)} = 0$ (respectively, $ = 1$ ) has a natural Lie algebra (respectively, formal Lie group) structure, with the comutator Lie-bracket $[\mathbf {y}, \mathbf {x}] = \mathbf {y}\mathbf {x} - \mathbf {x}\mathbf {y}$ for $\mathbf {x}, \mathbf {y} \in \mathcal {T}_0$ , where $\mathbf {x}\mathbf {y}$ stands for the concatenation product.Footnote 1

Let further $\mathscr {S} = \mathscr {S} (\mathbb {R}^d)$ , respectively, $\mathscr {S}^c= \mathscr {S}^c(\mathbb {R}^d)$ , denote the class of càdlàg,Footnote 2 respectively, continuous, d-dimensional semimartingales on some filtered probability space $(\Omega , (\mathcal {F}_t)_{t \ge 0}, \mathbb {P})$ . We recall that this is the somewhat decisive class of stochastic processes that allows for a reasonable, stochastic integration theory. Classic texts include [Reference Revuz and Yor59, Reference Le Gall43] (continuous semimartingale) and [Reference Jacod and Shiryaev36, Reference Protter57] (càdlàg semimartingales); a concise introduction for readers with no background in stochastic analysis is [Reference Aït-Sahalia and Jacod2], Chapter 1.Footnote 3 (Readers with no background in probability may also focus in a first reading on deterministic semimartingales; these are precisely càdlàg paths of finite variation on compacts.) Following Lyons [Reference Lyons45], for a continuous semimartingale $X\in \mathscr {S}^c$ , the signature given as the formal sum of iterated Stratonovich integrals,

$$\begin{align*}\mathrm{Sig} (X)_{s, t} = 1 + X_{s, t} + \int_s^t X_{s, u}\,{\circ\mathrm d}X_u + \int_s^t \left( \int_s^{u_1} X_{s, u_2}\,{\circ \mathrm d}X_{u_2} \right)\,{\circ\mathrm d} X_{u_1} + \cdots \end{align*}$$

for $0 \le s \le t$ defines a random element in $\mathcal {T}_1$ and, as a process, a formal $\mathcal {T}_1$ -valued semimartingale. By regarding the d-dimensional semimartingale X as a $\mathcal {T}_0$ -valued semimartingale ( $X \leftrightarrow \mathbf {X} = (0,X,0,\dots $ )), we see that the signature of $X\in \mathscr {S}^c$ satisfies the Stratonovich stochastic differential equation

(1.1) $$ \begin{align} \mathrm d S = S\,{\circ\mathrm d} \mathbf{X}. \end{align} $$

In the general case of $\mathbf {X}\in \mathscr {S}(\mathcal {T}_0)$ with possibly nontrivial higher-order semimartingale components $\mathbf {X} = (0, \mathbf {X}^{(1)}, \mathbf {X}^{(2)}, \dots )$ , the solution to equation (1.1) is also known as the Lie group valued stochastic exponential (or development), with classical references [Reference McKean51, Reference Hakim-Dowek and Lépingle31]; the càdlàg case [Reference Estrade20] is consistent with the geometric or Marcus [Reference Marcus49, Reference Marcus50, Reference Kurtz, Pardoux and Protter41, Reference Applebaum4, Reference Friz and Shekhar25] interpretation of equation (1.1)Footnote 4 with jump behavior $S_t = e^{\Delta \mathbf {X}_t} S_{t-}$ . From a stochastic differential geometry point of view, one aims for an intrinsic understanding of equation (1.1) valid for arbitrary Lie groups. For instance, if $\mathbf {X}$ takes values in any sub Lie algebra $\mathcal {L} \subset \mathcal {T}_0$ , then S takes values in the group $\mathcal {G} = \exp \mathcal {L}$ . In case of a d-dimensional semimartingale X, the minimal choice is the free Lie algebra $\mathrm {Lie}\mathopen {(\mkern -3mu(}\mathbb {R}^d\mathclose {)\mkern -3mu)}$ spanned by $\mathbb {R}^d$ (see, for example, [Reference Reutenauer58]), and the resulting Lie algebra structure of iterated integrals (both in the smooth and Stratonovich semimartingale case) is well-known. The extrinsic linear ambient space $\mathcal {T} \supset \exp {\mathcal {L}}$ will be important to us. Indeed, writing $S_t=\mathrm {Sig}(\mathbf {X})_{0,t}$ for the (unique, global) $\mathcal {T}_1$ -valued solution of equation (1.1) driven by $\mathcal {T}_0$ -valued $\mathbf {X}$ , started at $S_0 = 1$ , we define, whenever $\mathrm {Sig} (\mathbf {X})_{0, T}$ is (componentwise) integrable, the expected signature and signature cumulants (SigCum)

$$\begin{align*}\boldsymbol{\mu} (T) := \mathbb{E} (\mathrm{Sig} (\mathbf{X})_{0, T})\in\mathcal{T}_1, \quad \boldsymbol{\kappa} (T) := \log \boldsymbol{\mu} (T) \in \mathcal{T}_0. \end{align*}$$

Already when $\mathbf {X}$ is deterministic, and sufficiently regular to make equation (1.1) meaningful, this leads to an interesting (ordinary differential) equation for $\boldsymbol {\kappa }$ with accompanying (Magnus) expansion, well understood as an effective computational tool [Reference Iserles, Munthe-Kaas, Nørsett and Zanna34, Reference Blanes, Casas, Oteo and Ros6]. The importance of the stochastic case $\mathbf {X} = \mathbf {X}(\omega )$ , with expectation and logarithm thereof, was developed by Lyons and coauthors; see [Reference Lyons45] and references therein, with a variety of applications, ranging from machine learning to numerical algorithms on Wiener space known as cubature [Reference Lyons and Victoir47]; signature cumulants were named and first studied in their own right in [Reference Bonnier and Oberhauser7]. The joint appearance of cumulants and Magnus type-expansion is also seen in non-commutative probability [Reference Celestino, Ebrahimi-Fard, Patras and Perales10], although the methods and aims appear quite different.Footnote 5

In the special case of $d=1$ and $\mathbf {X}=(0,X,0,\dots )$ , where X is a scalar semimartingale, $\boldsymbol {\mu } (T)$ and $\boldsymbol {\kappa } (T)$ are nothing but the sequence of moments and cumulants of the real valued random variable $X_T-X_0$ . When $d> 1$ , the expected signature / cumulants provides an effective way to describe the process X on $[0,T]$ ; see [Reference LeJan and Qian44, Reference Lyons45, Reference Chevyrev and Lyons13]. The question arises how to compute them. If one takes $\mathbf {X}$ as d-dimensional Brownian motion, the signature cumulant $\boldsymbol {\kappa }(T)$ equals $(T/2) \mathbf {I}_d$ , where $\mathbf {I}_d$ is the identity $2$ -tensor over $\mathbb {R}^d$ . This is known as Fawcett’s formula [Reference Lyons and Victoir47, Reference Friz and Hairer24]. Loosely speaking, and postponing precise definitions, our main result is a vast generalization of Fawcett’s formula.

Theorem 1.1 FunctEqu $\mathscr {S}$ -SigCum

For sufficiently integrable $\mathbf {X}\in \mathscr {S}(\mathcal {T}_0)$ , the (time-t) conditional signature cumulants $ \boldsymbol {\kappa }_t (T) \equiv \boldsymbol {\kappa }_t := \log \mathbb {E}_t (\mathrm {Sig} (\mathbf {X})_{t, T})$ is the unique solution of the functional equation

(1.2)

where all integrals are understood in an Itô– and Riemann–Stieltjes sense, respectively,Footnote 6 and $\operatorname {\mathrm {ad}}{\mathbf {x}}=[\mathbf {x}, \cdot ]:\mathcal {T}_0\to \mathcal {T}_0$ denotes the adjoined operator associated to $\mathbf {x} \in \mathcal {T}_0$ . The functions $H,G,Q$ are defined in equation (4.1) below; see also also Section 2 for further notation.

As displayed in Figures 1 and 2, this theorem has an avalanche of consequences on which we now comment.

  • Equation (1.2) allows to compute $\boldsymbol {\kappa }^{(n)} \in (\mathbb {R}^d)^{\otimes n}$ as a function of $\boldsymbol {\kappa }^{(1)},\dotsc ,\boldsymbol {\kappa }^{(n-1)}$ . (This remark applies mutatis mutandis to all special cases seen as vertices in Figure 1.) The resulting expansions, displayed in Figure 2, are of computational interest. In particular, our approach allows us, in some special cases, either to derive closed expressions for the conditional cumulant series $\boldsymbol {\kappa }_t$ or to characterize it a as the unique solution of a certain parabolic PDE (see Section 6). In any case, this provides a means of computation that can in principle be more efficient than the naïve Monte Carlo approach. Even when such a concrete form of $\boldsymbol {\kappa }_t$ is not available, the recursive nature of the expressions for each homogeneous component can be useful in some numerical scenarios.

    Figure 1 FunctEqu $\mathscr {S}$ -SigCum (Theorem 4.1) and implications. $\mathscr {S}$ (respectively, $\mathscr {S}^{c}$ ) stands for general (respectively, continuous) semimartingales and $\mathscr {V}$ (respectively, $\mathscr {V}^{c}$ ) stands for finite variation (respectively, finite variation and continuous) processes.

    Figure 2 Computational consequence: accompanying recursions.

  • The most classical consequence of equation (1.2) appears when $\mathbf {X}$ is a deterministic continuous semimartingale: that is, in particular, the components of $\mathbf {X}$ are continuous paths of finite variation, which also covers the absolutely continuous case with integrable componentwise derivative $\dot {\mathbf {X}}$ . In this case all bracket terms and the final jump-sum disappear. What remains is a classical differential equation due to [Reference Hausdorff32], here in backward form:

    (1.3) $$ \begin{align} - \mathrm{d} \boldsymbol{\kappa}_t (T) = H(\operatorname{\mathrm{ad}}{\boldsymbol{\kappa}_{t}})\mathrm{d}\mathbf{X}_t. \end{align} $$
    The accompanying expansion is then precisely Magnus expansion [Reference Magnus48, Reference Iserles and Nørsett33, Reference Iserles, Munthe-Kaas, Nørsett and Zanna34, Reference Blanes, Casas, Oteo and Ros6]. By taking $\mathbf {X}$ continuous and piecewise linear on two adjacent intervals, say $[0,1)\cup [1,2)$ , one obtains the Baker–Campbell–Hausdorff formula (see, e.g., [Reference Miller52, Theorem 5.5])
    (1.4) $$ \begin{align} \begin{aligned} \boldsymbol{\kappa}_0(2)=\log\bigl(\exp(\mathbf{x}_1)\exp(\mathbf{x}_2) \bigr) &=: \operatorname{BCH}(\mathbf{x}_1,\mathbf{x}_2)\\ &=\mathbf{x}_2+\int_0^1\Psi(\exp(\operatorname{\mathrm{ad}} t\mathbf{x}_1)\circ\exp(\operatorname{\mathrm{ad}}\mathbf{x}_2))(\mathbf{x}_1)\,\mathrm dt, \end{aligned} \end{align} $$
    with
    $$\begin{align*}\Psi(z):=\frac{\ln(z)}{z-1}=\sum_{n\ge 0}\frac{(-1)^n}{n+1}(z-1)^n. \end{align*}$$
    It is also instructive to let $\mathbf {X}$ piecewise constant on these intervals, with $\Delta \mathbf {X}_1 =\mathbf {x}_1, \Delta \mathbf {X}_2 = \mathbf {x}_2$ , in which case equation (1.2) reduces to the first equality in equation (1.4). Such jump variations of the Magnus expansion are discussed in Section 5.1.
  • Write $\pi _{\mathrm {Sym}}: \mathcal {T} \to \mathcal {S}$ for the canonical projection to the extended symmetric algebra $\mathcal {S}$ , the linear space identified with symmetric tensor series, and define the $\mathcal {S}$ -valued semimartingale $\hat {\mathbf {X}} := \pi _{\mathrm {Sym}}(\mathbf {X})$ and symmetric signature cumulants $\hat {\boldsymbol {\kappa }}(T) := \log (\mathbb {E}_{\cdot }(\mathrm {Sig}(\hat {\mathbf {X}})_{\cdot , T})) = \pi _{\mathrm {Sym}}(\boldsymbol {\kappa }(T))$ (see Section 2.3.1 for more detail). Then equation (1.2), in its projected and commutative form becomes (see also Section 5.2)

    (1.5) $$ \begin{align} \begin{aligned} \qquad \text{FunctEqu }\mathscr{S}\text{-Cum:} \quad \hat{\boldsymbol{\kappa}}_t (T) & = \mathbb{E}_t\bigg\{ \hat{\mathbf{X}}_{t,T} + \frac{1}{2} \left\langle (\hat{\mathbf{X}}+ \hat{\boldsymbol{\kappa}})^{c} \right\rangle_{t,T}\\ &\qquad + \sum_{t < u \le T}\bigg(\exp \Big( \Delta \hat{\mathbf{X}}_u + \Delta \hat{\boldsymbol{\kappa}}_u \Big) - 1 - (\Delta \hat{\mathbf{X}}_u + \Delta \hat{\boldsymbol{\kappa}}_u ) \bigg) \bigg\}, \end{aligned} \end{align} $$
    where $\exp \colon \mathcal {S}_0 \mapsto \mathcal {S}_1$ is defined by the usual power series. First-level tensors are trivially symmetric, and therefore equation (1.5) applies to the a $\mathbb {R}^d$ -valued semimartingale X via the canonical embedding $\hat {\mathbf {X}} = (0, X, 0, \dots ) \in \mathscr {S}(\mathcal {S}_0)$ . More interestingly, the case $\hat {\mathbf {X}} = (0,aX,b\langle X \rangle , 0, \dotsc )$ for a d-dimensional continuous martingale X can be seen to underlie the expansions of [Reference Friz, Gatheral and Radoičić23], which improves and unifies previous results [Reference Lacoin, Rhodes and Vargas42, Reference Alos, Gatheral and Radoičić3] treating $(a,b)=(1,0)$ and $(a,b)=(1,-1/2)$ , respectively. Following Gatheral and coworkers, equation (1.5) and subsequent expansions involve ‘diamond’ products of semimartingales, given, whenever well-defined, by
    $$ \begin{align*}(A \diamond B)_t(T) := \mathbb{E}_t \big( \left\langle A^c, B^c \right\rangle_{t,T} \big). \end{align*} $$
    We note that equation (1.5) induces recursive formulae for cumulants, dubbed Diamond expansions in Figure 2, previously discussed in [Reference Lacoin, Rhodes and Vargas42, Reference Alos, Gatheral and Radoičić3, Reference Friz, Gatheral and Radoičić23, Reference Fukasawa and Matsushita28], together with a range of applications, from quantitative finance (including rough volatility models [Reference Abi Jaber, Larsson and Pulido1, Reference Gatheral and Keller-Ressel29]) to statistical physics: in [Reference Lacoin, Rhodes and Vargas42], the authors rely on such formulae to compute the cumulants function of log-correlated Gaussian fields, more precisely approximations thereof, that underlies the Sine-Gordon model, which is a key ingredient in their renormalization procedure.

    With regard to the existing (‘commutative’) literature, our algebraic setup is ideally suited to work under finite moment assumptions; we are able to deal with jumps, not treated in [Reference Lacoin, Rhodes and Vargas42, Reference Alos, Gatheral and Radoičić3]. Equation (1.5), the commutative shadow of equation (1.2), should be compared with Riccati’s ordinary differential equation from affine process theory [Reference Duffie, Filipović and Schachermayer19, Reference Cuchiero, Filipović, Mayerhofer and Teichmann16, Reference Keller-Ressel, Schachermayer and Teichmann40]. A systematic comparison would lead us too far astray from our main object of study; nevertheless, we illustrate the connection in Remark 6.9. Of course, our results, in particular equations (1.2) and (1.5), are not restricted to affine semimartingales. In turn, expected signatures and cumulants - and subsequently all our statements above these - require moments, which is not required for the Riccati evolution of the characteristic function of affine processes. Of recent interest, explicit diamond expansions have been obtained for ‘rough affine’ processes, non-Markov by nature, with a cumulant generating function characterized by Riccati Volterra equations; see [Reference Abi Jaber, Larsson and Pulido1, Reference Gatheral and Keller-Ressel29, Reference Friz, Gatheral and Radoičić23]. It is remarkable that analytic tractability remains intact when one passes to path space and considers signature cumulants, as we illustrate in Section 6.3.

  • Finally, we mention Signature-SDEs [Reference Arribas, Salvi and Szpruch5], tractable classes of stochastic differential equations that can be studied from an infinite dimensional affine and polynomial perspective [Reference Cuchiero, Svaluto-Ferro and Teichmann18]. Calibration of such models hinges on the efficient computation of expected signatures, which is the very purpose of this paper.

We conclude this introduction with some remarks on convergence. As explained, this work contains generalizations of cumulant type recursions, previously studied in [Reference Alos, Gatheral and Radoičić3, Reference Lacoin, Rhodes and Vargas42, Reference Friz, Gatheral and Radoičić23], the interest therein being the algorithmic computation of cumulants. Basic facts of analytic functions show that classical moment- and cumulant-generating functions, for random variables with finite moments of all orders, have a radius of convergence $\rho \ge 0$ , directly related to growth of the corresponding sequence. Convergence, in the sense $\rho> 0$ , implies that the moment problem is well-posed. That is, the moments (equivalently: cumulants) determine the law of the underlying random variable. (See also [Reference Friz, Gatheral and Radoičić23] for a related discussion in the context of diamond expansions.) The point of view taken here is to work directly on this space of sequences, which is even more natural in the non-commutative setting, as already seen in the deterministic setting of [Reference Magnus48]. While convergence of expected signatures or signature cumulants series is not directly an interesting question,Footnote 7 understanding their growth most certainly is: in a celebrated paper [Reference Chevyrev and Lyons13], it was shown that under a growth condition of the expected signature, the ‘expected signature problem’ is well-posed; that is, the expected signature (equivalently: signature cumulants) determines the law of the random signature. With this in mind, it is conceivable that Theorem 1.1 will prove useful toward controlling the growth of signature cumulants (and hence expected signatures).

2 Preliminaries

2.1 The tensor algebra and tensor series

Denote by $T({\mathbb {R}^d})$ the tensor algebra over ${\mathbb {R}^d}$ : that is,

$$ \begin{align*} T({\mathbb{R}^d}):= \bigoplus_{k=0}^\infty ({\mathbb{R}^d})^{\otimes k}, \end{align*} $$

elements of which are finite sums (also known as tensor polynomials) of the form

(2.1) $$ \begin{align} \mathbf{x} = \sum_{k \ge 0} \mathbf{x}^{(k)} = \sum_{w \in \mathcal{W}^d} \mathbf{x}^w e_w \end{align} $$

with $\mathbf {x}^{(k)} \in ({\mathbb {R}^d})^{\otimes k}, \mathbf {x}^w \in \mathbb {R}$ and linear basis vectors $e_w := e_{i_1}\dotsm e_{i_k}\in ({\mathbb {R}^d})^{\otimes k}$ , where w ranges over all words $w=i_1\dotsm i_k\in \mathcal {W}_d$ over the alphabet $\{1,\dots ,d\}$ . Note $\mathbf {x}^{(k)} = \sum _{|w|=k} \mathbf {x}^w e_w$ , where $|w|$ denotes the length a word w. The element $e_\emptyset = 1 \in ({\mathbb {R}^d})^{\otimes 0} \cong \mathbb {R}$ is the neutral element of the concatenation (also known as a tensor) product, which is obtained by linear extension of $e_we_{w'}=e_{ww'}$ , where $ww' \in \mathcal {W}_d$ denotes concatenation of two words. We thus have, for $\mathbf {x},\mathbf {y} \in T({\mathbb {R}^d})$ ,

$$ \begin{align*}\mathbf{x}\mathbf{y} = \sum_{k \ge 0} \sum_{\ell =0}^k \mathbf{x}^{(\ell)} \mathbf{y}^{(k-\ell)} = \sum_{w \in \mathcal{W}^d} \left( \sum_{w_1w_2 = w} \mathbf{x}^{w_1}\mathbf{y}^{w_2} \right) e_w \in T({\mathbb{R}^d}). \end{align*} $$

This extends naturally to infinite sums, also known as tensor series, elements of the ‘completed’ tensor algebra

$$ \begin{align*} \mathcal{T} := T\mathopen{(\mkern-3mu(}\mathbb{R}^d\mathclose{)\mkern-3mu)}:= \prod_{k=0}^\infty ({\mathbb{R}^d})^{\otimes k}, \end{align*} $$

which are written as in equation (2.1), but now as formal infinite sums with identical notation and multiplication rules; the resulting algebra $\mathcal {T}$ obviously extends $T(\mathbb {R}^d)$ . For any $n\in {\mathbb {N}_{\ge 1}}$ , define the projection to tensor levels by

$$ \begin{align*}\pi_n: \mathcal{T} \to ({\mathbb{R}^d})^{\otimes n}, \quad \mathbf{x} \mapsto \mathbf{x}^{(n)}.\end{align*} $$

Denote by $\mathcal {T}_0$ and $\mathcal {T}_1$ the subspaces of tensor series starting with $0$ and $1$ , respectively; that is, $\mathbf {x} \in \mathcal {T}_0$ (respectively, $\mathcal {T}_1$ ) if and only if $\mathbf {x}^\emptyset =0$ (respectively, $\mathbf {x}^\emptyset =1$ ). Restricted to $\mathcal {T}_0$ and $\mathcal {T}_1$ , respectively, the exponential and logarithm in $\mathcal {T}$ , defined by the usual series,

$$ \begin{align*} \exp\colon\mathcal{T}_0 \to \mathcal{T}_1,& \quad \mathbf{x} \mapsto \exp(\mathbf{x}) := 1 + \sum_{k=1}^\infty \frac{1}{k!}(\mathbf{x})^k, \\ \log\colon\mathcal{T}_1 \to \mathcal{T}_0,& \quad 1 + \mathbf{x} \mapsto \log( 1 + \mathbf{x}) := \sum_{k=1}^\infty \frac{(-1)^{k+1}}{k}(\mathbf{x})^k, \end{align*} $$

are globally defined and inverse to each other. The vector space $\mathcal {T}_0$ becomes a Lie algebra with the commutator bracket

$$ \begin{align*}\left[ \mathbf{x}, \mathbf{y} \right] := \mathbf{x}\mathbf{y}-\mathbf{y}\mathbf{x}, \quad \mathbf{x}, \mathbf{y} \in \mathcal{T}_0.\end{align*} $$

Define the adjoined operator associated to an Lie-algebra element $\mathbf {y} \in \mathcal {T}_0$ by

$$ \begin{align*} \operatorname{\mathrm{ad}}{\mathbf{y}}\colon \mathcal{T}_0 \to \mathcal{T}_0, \ \mathbf{x} \mapsto \left[ \mathbf{y}, \mathbf{x} \right]. \end{align*} $$

The exponential image $\mathcal {T}_1=\exp (\mathcal {T}_0)$ is a Lie group, at least formally so. We refrain from equipping the infinite-dimensional $\mathcal {T}_1$ with a differentiable structure, not necessary in view of the ‘locally finite’ nature of the group law $(\mathbf {x},\mathbf {y}) \mapsto \mathbf {x} \mathbf {y}$ .

Let $(a_k)_{k\ge 1}$ be a sequence of real numbers and $\mathbf {x}\in \mathcal {T}_0$ . Then we can always define a linear operator on $\mathcal {T}_0$ by

$$ \begin{align*} \left[ \sum_{k \ge 0}a_k(\operatorname{\mathrm{ad}}\mathbf{x})^{k} \right]: \mathcal{T}_0 \to \mathcal{T}_0, \quad \mathbf{y}\mapsto \sum_{k \ge 0}a_k(\operatorname{\mathrm{ad}}\mathbf{x})^{k}(\mathbf{y}), \end{align*} $$

where $(\operatorname {\mathrm {ad}}\mathbf {x})^{0} = \mathrm {Id}$ is the identity operator and $(\operatorname {\mathrm {ad}} \mathbf {x})^{n} = \operatorname {\mathrm {ad}}\mathbf {x} \circ (\operatorname {\mathrm {ad}} \mathbf {x})^{n-1}$ for any $n\in {\mathbb {N}_{\ge 1}}$ . Indeed, there is no convergence issue due to the graded structure, as can be seen by projecting to some tensor level $n\in {\mathbb {N}_{\ge 1}}$

(2.2)

where the inner summation in the right-hand side is over a finite set of multi-indices $\ell = (l_1, \dotsc , l_{k+1})\in ({\mathbb {N}_{\ge 1}})^{k+1}$ , where and . In the following we will simply write $(\operatorname {\mathrm {ad}}\mathbf {x}\operatorname {\mathrm {ad}}\mathbf {y}) \equiv (\operatorname {\mathrm {ad}}\mathbf {x} \circ \operatorname {\mathrm {ad}}\mathbf {y})$ for the composition of adjoint operators. Further, when $\ell = (l_1)$ is a multi-index of length one, we will use the notation $(\operatorname {\mathrm {ad}}\mathbf {x}^{(l_2)} \cdots \operatorname {\mathrm {ad}}\mathbf {x}^{(l_{k+1})}) \equiv \mathrm {Id}$ . Note also that the iteration of adjoint operations can be explicitly expanded in terms of left- and right-multiplication, as follows:

(2.3)

For a word $w \in \mathcal {W}_d$ with $|w|>0$ , we define the directional derivative for a function $f\colon \mathcal {T} \to \mathbb {R}$ by

$$ \begin{align*} (\partial_w f)(\mathbf{x}):=\partial_t (f(\mathbf{x} + t e_w))\big\vert_{t=0}, \end{align*} $$

for any $\mathbf {x} \in \mathcal {T}$ such that the right-hand derivative exists.

2.2 The outer tensor algebra

Denote by $\mathrm {m}: T({\mathbb {R}^d})\otimes T({\mathbb {R}^d}) \to T({\mathbb {R}^d})$ the multiplication (concatenation) map of the tensor algebra. Note that $\mathrm {m}$ is linear and, due to the non-commutativity of the tensor product, not symmetric. The map can naturally be extended to linear map $\mathrm {m}: T({\mathbb {R}^d}) \overline {\otimes } T({\mathbb {R}^d}) \to \mathcal {T}$ , where $T({\mathbb {R}^d}) \overline {\otimes } T({\mathbb {R}^d})$ is the following graded algebra:

$$ \begin{align*} T({\mathbb{R}^d}) \overline{\otimes} T({\mathbb{R}^d}) := \prod_{n=0}^{\infty} \left(\bigoplus_{i = 0}^{n} ({\mathbb{R}^d})^{\otimes i} \otimes ({\mathbb{R}^d})^{\otimes (n-i)}\right). \end{align*} $$

Note that there is the following natural linear embedding

$$ \begin{align*} \mathcal{T}\otimes\mathcal{T} \hookrightarrow T({\mathbb{R}^d}) \overline{\otimes} T({\mathbb{R}^d}), \quad \mathbf{x}\otimes \mathbf{y} \mapsto \sum_{n=0}^{\infty} \left(\sum_{i=0}^{n} \mathbf{x}^{(i)}\otimes \mathbf{y}^{(n-i)}\right). \end{align*} $$

We will of course refrain from explicitly denoting the embedding and simply regard $\mathbf {x} \otimes \mathbf {y}$ as an element in $T({\mathbb {R}^d}) \overline {\otimes } T({\mathbb {R}^d})$ . We emphasize that here $\otimes $ does not denote the (inner) tensor product in $\mathcal {T}$ , for which we did not reserve a symbol, but it denotes another (outer) tensor product. We can lift two linear maps $g, f\colon \mathcal {T}\to \mathcal {T}$ to a linear map $g \odot f: T({\mathbb {R}^d}) \overline {\otimes } T({\mathbb {R}^d}) \to \mathcal {T}$ defined by

$$ \begin{align*} g \odot f := \mathrm{m}\circ(g \otimes f). \end{align*} $$

In particular, for all $\mathbf {x},\mathbf {y}\in \mathcal {T}$ , it holds that

$$ \begin{align*} (g \odot f) (\mathbf{x} \otimes \mathbf{y}) = g(\mathbf{x})f(\mathbf{y})\in \mathcal{T}. \end{align*} $$

2.3 Some quotients of the tensor algebra

2.3.1 The symmetric tensor algebra

The symmetric algebra over ${\mathbb {R}^d}$ , denoted by $S({\mathbb {R}^d})$ , is the quotient of $T({\mathbb {R}^d})$ by the two-sided ideal I generated by $\{xy-yx:x,y\in {\mathbb {R}^d}\}$ . A linear basis of $S({\mathbb {R}^d})$ is then given by $\{ \hat {e}_v \}$ over non-decreasing words, $v=(i_1,\dotsc ,i_n) \in \widehat {\mathcal {W}}_d$ , with $1 \le i_1 \le \dots \le i_n \le d, n \ge 0$ . Every $\hat {\mathbf {x}} \in S(\mathbb {R}^d)$ can be written as a finite sum,

$$ \begin{align*}\hat{\mathbf{x}} = \sum_{v \in \widehat{\mathcal{W}}_d} \hat{\mathbf{x}}^v \hat{e}_v , \end{align*} $$

and we have an immediate identification with polynomials in d commuting indeterminates. The canonical projection

(2.4) $$ \begin{align} \pi_{\mathrm{Sym}}:T({\mathbb{R}^d})\twoheadrightarrow S({\mathbb{R}^d}), \quad\mathbf{x} \mapsto \sum_{w\in\mathcal{W}_d} \mathbf{x}^{w}\hat{e}_{\hat{w}}, \end{align} $$

where $\hat {w}\in \hat {\mathcal {W}}_d$ denotes the non-decreasing reordering of the letters of the word $w\in \mathcal {W}_d$ , is an algebra epimorphism, which extends to an epimorphism $\pi _{\mathrm {Sym}}: \mathcal {T}\twoheadrightarrow \mathcal {S}$ , where $\mathcal {S} = S \mathopen {(\mkern -3mu(} \mathbb {R}^d\mathclose {)\mkern -3mu)}$ is the algebra completion, identifiable as a formal series in d non-commuting (respectively, commuting) indeterminates. As a vector space, $\mathcal {S}$ can be identified with symmetric formal tensor series. Denote by $ \mathcal {S}_0$ and $\mathcal {S}_1$ the affine space given by those $\hat {\mathbf {x}}\in \mathcal {S}$ with $ \hat {\mathbf {x}}^\emptyset =0$ and $ \hat {\mathbf {x}}^\emptyset =1$ , respectively. The usual power series in $\mathcal {S}$ define $\widehat {\exp {}}\colon \mathcal {S}_0 \to \mathcal {S}_1$ with inverse $\widehat {\log {}}\colon \mathcal {S}_1 \to \mathcal {S}_0$ , and we have

$$ \begin{align*} \pi_{\mathrm{Sym}}\exp{(\mathbf{x} + \mathbf{y})} &= \widehat{\exp{}}(\hat{\mathbf{x}})\widehat{\exp{}}(\hat{\mathbf{y}}), \quad \mathbf{x}, \mathbf{y} \in \mathcal{T}_0\\ \pi_{\mathrm{Sym}}\log{(\mathbf{x} \mathbf{y})} &= \widehat{\log{}}(\hat{\mathbf{x}}) + \widehat{\log{}}(\hat{\mathbf{y}}), \quad \mathbf{x},\mathbf{y} \in \mathcal{T}_1. \end{align*} $$

We shall abuse notation in what follows and write $\exp $ (respectively, $\log $ ), instead of $\widehat {\exp }$ (respectively, $ \widehat {\log }$ ).

2.3.2 The (step-n) truncated tensor algebra

For $n\in {\mathbb {N}}$ , the subspace

$$\begin{align*}\mathcal{I}_n:=\prod_{k=n+1}^\infty({\mathbb{R}^d})^{\otimes k} \end{align*}$$

is a two-sided ideal of $\mathcal {T}$ . Therefore, the quotient space $\mathcal {T}/\mathcal {I}_n$ has a natural algebra structure. We denote the projection map by $\pi _{(0,n)}$ . We can identify $\mathcal {T}/\mathcal {I}_n$ with

$$\begin{align*}\mathcal{T\,}^n := \bigoplus_{k=0}^n({\mathbb{R}^d})^{\otimes k}, \end{align*}$$

equipped with truncated tensor product

$$ \begin{align*}\mathbf{x}\mathbf{y} = \sum_{k=0}^{n} \sum_{\ell_1 + \ell_2=k} \mathbf{x}^{(\ell_1)} \mathbf{y}^{(\ell_2)} = \sum_{w \in \mathcal{W}^d,|w|\le n} \left( \sum_{w_1w_2 = w} \mathbf{x}^{w_1}\mathbf{y}^{w_2} \right) e_w \in \mathcal{T\,}^n. \end{align*} $$

The sequence of algebras $(\mathcal {T}^n:n\ge 0)$ forms an inverse system with limit $\mathcal {T}$ . There are also canonical inclusions $\mathcal {T}^k\hookrightarrow \mathcal {T}^{n}$ for $k\le n$ ; in fact, this forms a direct system with limit $T({\mathbb {R}^d})$ . The usual power series in $\mathcal {T}^{n}$ define $\exp _n\colon \mathcal {T}^n_0 \to \mathcal {T}^n_1$ with inverse $\log _n\colon \mathcal {T}^n_1 \to \mathcal {T}^n_0$ ; we may again abuse notation and write $\exp $ and $\log $ when no confusion arises. (In the proof section Section 7.2, we will stick to the $\exp _n$ notation to emphasis the presence of truncation.) As before, $\mathcal {T}^n_0$ has a natural Lie algebra structure, and $\mathcal {T}^n_1$ (now finite dimensional) is a bona fide Lie group.

We equip $T(\mathbb {R}^d)$ with the norm

$$ \begin{align*} |a|_{T(\mathbb{R}^d)} := \max_{k\in{\mathbb{N}}}|a^{(k)}|_{({\mathbb{R}^d})^{\otimes k}}, \end{align*} $$

where $|\cdot |_{({\mathbb {R}^d})^{\otimes k}}$ is the Euclidean norm on $({\mathbb {R}^d})^{\otimes k}\cong \mathbb {R}^{d^k}$ , which makes it a Banach space. The same norm makes sense in $\mathcal {T}^n$ , and since the definition is consistent in the sense that $|a|_{\mathcal {T}^k} = |a|_{\mathcal {T}^n}$ for any $a \in \mathcal {T}^{n}$ and $k \ge n$ , and $|a|_{\mathcal {T}^n} = |a|_{({\mathbb {R}^d})^{\otimes n}}$ for any $a \in ({\mathbb {R}^d})^{\otimes n}$ . We will drop the index whenever it is possible and write simply $|a|$ .

2.4 Semimartingales

Let $\mathscr {D}$ be the space of adapted càdlàg process $X\colon \Omega \times [0,T) \to \mathbb {R}$ with $T\in (0,\infty ]$ defined on some filtered probability space $(\Omega , (\mathcal {F}_t)_{0 \le t \le T}, \mathbb {P})$ . The space of semimartingales $\mathscr {S}$ is given by the processes $X\in \mathscr {D}$ that can be decomposed as

$$ \begin{align*}X_{t}=X_0+M_{t}+A_{t}, \end{align*} $$

where $M \in \mathscr {M}_{\mathrm {loc}}$ is a càdlàg local martingale and $A\in \mathscr {V}$ is a càdlàg adapted process of locally bounded variation, both started at zero. Recall that every $X \in \mathscr {S}$ has a well-defined continuous local martingale part denoted by $X^c\in \mathscr {M}^c_{\mathrm {loc}}$ . The quadratic variation process of X is then given by

$$ \begin{align*} [X]_t = \left\langle X^{c} \right\rangle_t + \sum_{0 < u \le t} (\Delta X_u)^{2}, \quad 0 \le t \le T, \end{align*} $$

where $\left \langle \cdot \right \rangle $ denotes the (predictable) quadratic variation of a continuous semimartingale. Covariation square, respectively, angle brackets $[X,Y]$ and $\left \langle X^{c}, Y^{c} \right \rangle $ , for another real-valued semimartingale Y, are defined by polarization. For $q \in [1, \infty )$ , write $\mathcal {L}^{q} = L^{q}(\Omega , \mathcal {F}, \mathbb {P})$ ; then a Banach space $\mathscr {H}^q \subset \mathscr {S}$ is given by those $X\in \mathscr {S}$ with $X_0 = 0$ and

$$ \begin{align*}\| X \|_{\mathscr{H}^q} := \inf_{X = M + A} \bigg\Vert \left[ M \right]^{1 / 2}_{T} + \int_0^{T} |\mathrm{d} A_s | \bigg\Vert_{\mathcal{L}^{q}}< \infty. \end{align*} $$

Note that for a local martingale $M \in \mathscr {M}_{\mathrm {loc}}$ , it holds that (see [Reference Protter57, Ch. V, p. 245])

For a process $X\in \mathscr {D}$ , we define

and define the space $\mathscr {S}^q\subset \mathscr {S}$ of semimartingales $X\in \mathscr {S}$ such that $\Vert {X}\Vert _{\mathscr {S}^q}< \infty $ . Note that there exists a constant $c_q>0$ depending on q such that (see [Reference Protter57, Ch. V, Theorem 2])

(2.5) $$ \begin{align} \Vert{X}\Vert_{\mathscr{S}^q} \le c_q \| X \|_{\mathscr{H}^q}. \end{align} $$

We view d-dimensional semimartingales, $X= \sum _{i=1}^d X^i e_i \in \mathscr {S} (\mathbb {R}^d)$ , as special cases of tensor series valued semimartingales $\mathscr {S} (\mathcal {T}\,)$ of the form

$$ \begin{align*}\mathbf{X} = \sum_{w \in \mathcal{W}_d} \mathbf{X}^w e_w\end{align*} $$

with each component $\mathbf {X}^w$ a real-valued semimartingale. This extends mutatis mutandis to the spaces of $\mathcal {T}$ -valued adapted càdlàg processes $\mathscr {D}(\mathcal {T}\,)$ , martingales $\mathscr {M}(\mathcal {T}\,)$ and adapted càdlàg processes with finite variation $\mathscr {V}(\mathcal {T}\,)$ . Note also that we typically deal with $\mathcal {T}_0$ -valued semimartingales, which amounts to having only words with length $|w| \ge 1$ . Standard notions such as continuous local martingale $\mathbf {X}^{c}$ and jump process $\Delta \mathbf {X}_t = \mathbf {X}_t - \mathbf {X}_{t^-}$ are defined componentwise.

Brackets: Now let $\mathbf {X}$ and $\mathbf {Y}$ be $\mathcal {T}$ -valued semimartingales. We define the (non-commutative) outer quadratic covariation bracket of $\mathbf {X}$ and $\mathbf {Y}$ by

Similarly, define the (non-commutative) inner quadratic covariation bracket by

for continuous $\mathcal {T}$ -valued semimartingales $\mathbf {X}, \mathbf {Y}$ , this coincides with the predictable quadratic covariation

$$\begin{align*}\left\langle \mathbf{X}^{}, \mathbf{Y}^{} \right\rangle_t := \sum_{w\in\mathcal{W}_d }\left(\sum_{w_1 w_2 = w} \left\langle \mathbf{X}^{w_1}, \mathbf{Y}^{w_2} \right\rangle_t \right)e_w \in \mathcal{T}. \end{align*}$$

As usual, we may write

and $\left \langle \mathbf {X} \right \rangle \equiv \left \langle \mathbf {X}, \mathbf {X} \right \rangle $ .

$\mathscr {H}$ -spaces: The definition of $\mathscr {H}^{q}$ -norm naturally extends to tensor valued martingales. More precisely, for $\mathbf {X}^{(n)} \in \mathscr {S}(({\mathbb {R}^d})^{\otimes n})$ with $n\in {\mathbb {N}_{\ge 1}}$ and $q\in [1,\infty )$ , we define

where the infimum is taken over all possible decompositions $\mathbf {X}^{(n)} = \mathbf {M} + \mathbf {A}$ with $\mathbf {M}\in \mathscr {M}_{\mathrm {loc}}(({\mathbb {R}^d})^{\otimes n})$ and $\mathbf {A}\in \mathscr {V}(({\mathbb {R}^d})^{\otimes n})$ , where

with the supremum taken over all partitions of the interval $[0,T]$ . One may readily check that

Further define the following subspace $\mathscr {H}^{q,N} \subset \mathscr {S}(\mathcal {T}_0^{N})$ of homogeneously q-integrable semimartingales

$$ \begin{align*} \mathscr{H}^{q,N} :=\left\{ \mathbf{X} \in \mathscr{S}(\mathcal{T}_0^{N}) \;\Big\vert\; \mathbf{X}_0 = 0,\; \vert\mkern-2.5mu\vert\mkern-2.5mu\vert{\mathbf{X}}\vert\mkern-2.5mu\vert\mkern-2.5mu\vert_{\mathscr{H}^{q,N}} < \infty \right\}, \end{align*} $$

where for any $\mathbf {X} \in \mathscr {S}(\mathcal {T\,}^{N})$ , we define

$$ \begin{align*} \vert\mkern-2.5mu\vert\mkern-2.5mu\vert{\mathbf{X}}\vert\mkern-2.5mu\vert\mkern-2.5mu\vert_{\mathscr{H}^{q,N}} := \sum_{n=1}^{N} \big( \left\Vert {\mathbf{X}^{(n)}} \right\Vert _{\mathscr{H}^{qN/n}}\big)^{1/n}. \end{align*} $$

Note that $\vert \mkern -2.5mu\vert \mkern -2.5mu\vert {\cdot }\vert \mkern -2.5mu\vert \mkern -2.5mu\vert _{\mathscr {H}^{q,N}}$ is sub-additive and positive definite on $\mathscr {H}^{q, N}$ , and it is homogeneous under dilation in the sense that

We also introduce the following subspace of $\mathscr {S}(\mathcal {T}\,)$ :

$$ \begin{align*} \mathscr{H}^{\infty-}(\mathcal{T}) := \left\{ \mathbf{X} \in \mathscr{S}(\mathcal{T}):\; \mathbf{X}^w\in\mathscr{H}^q, \;\forall\, 1 \le q < \infty,\;w\in\mathcal{W}_d\right\}. \end{align*} $$

Note that if $\mathbf {X}\in \mathscr {S}(\mathcal {T}\,)$ such that $\vert \mkern -2.5mu\vert \mkern -2.5mu\vert {\mathbf {X}^{(0,N)}}\vert \mkern -2.5mu\vert \mkern -2.5mu\vert _{\mathscr {H}^{1,N}} < \infty $ for all $N\in {\mathbb {N}_{\ge 1}}$ , then it also holds that $\mathbf {X}\in \mathscr {H}^{\infty -}(\mathcal {T}\,)$ .

Stochastic integrals: We are now going to introduce a notation for the stochastic integration with respect to tensor valued semimartingales. Denote by $\mathcal {L}(\mathcal {T}; \mathcal {T}) = \{ f: \mathcal {T} \to \mathcal {T}\;|\; f \text { is linear.}\}$ the space of endomorphism on $\mathcal {T}$ , and let $\mathbf {F}: \Omega \times [0,T] \to \mathcal {L}(\mathcal {T}; \mathcal {T})$ with $(t, \omega ) \mapsto \mathbf {F}_t(\omega; \cdot )$ such that it holds

(2.6) $$ \begin{align} &(\mathbf{F}_t(\mathbf{x}))_{0 \le t \le T} \in \mathscr{D}(\mathcal{T}), \quad \text{for all } \mathbf{x}\in\mathcal{T} \end{align} $$
(2.7) $$ \begin{align} \text{and}\quad &\mathbf{F}_t(\omega; \mathcal{I}_n) \subset \mathcal{I}_n, \quad \text{for all } n\in{\mathbb{N}}, \; (\omega, t)\in\Omega\times[0,T], \end{align} $$

where $\mathcal {I}_n\subset \mathcal {T}$ was introduced in Section 2.3.2, consisting of series with tensors of level $n+1$ and higher. In this case, we can define the stochastic Itô integral (and then analogously the Stratonovich/Marcus integral) of $\mathbf {F}$ with respect to $\mathbf {X}\in \mathscr {S}(\mathcal {T}\,)$ by

(2.8)

For example, let $\mathbf {Y}, \mathbf {Z} \in \mathscr {D}(\mathcal {T}\,)$ , and define $\mathbf {F} := \mathbf {Y}\,\mathrm {Id}\,\mathbf {Z}$ : that is, $\mathbf {F}_t(\mathbf {x}) = \mathbf {Y}_t \, \mathbf {x} \, \mathbf {Z}_t$ , the concatenation product from the left and right, for all $\mathbf {x}\in \mathcal {T}$ . Then we see that $\mathbf {F}$ indeed satisfies the conditions in equations (2.6) and (2.7), and we have

(2.9) $$ \begin{align} \int_{(0, \cdot]} (\mathbf{Y}_{t-}\,\mathrm{Id}\,\mathbf{Z}_{t-})(\mathrm{d} \mathbf{X}_t)= \int_{(0, \cdot]} \mathbf{Y}_{t-}\mathrm{d} \mathbf{X}_t \mathbf{Z}_{t-}= \sum_{w\in\mathcal{W}_d}\left( \sum_{w_1 w_2w_3 = w} \int_{(0, \cdot]} \mathbf Z^{w_1}_{t-} \mathbf{Y}^{w_3}_{t-}\,\mathrm d \mathbf{X}_t^{w_2} \right)e_w. \end{align} $$

Another important example is given by $\mathbf {F} = (\operatorname {\mathrm {ad}} \mathbf {Y})^{k}$ for any $\mathbf {Y}\in \mathscr {D}(\mathcal {T}_0)$ and $k\in {\mathbb {N}}$ . Indeed, we immediately see that $\mathbf {F}$ satisfies the condition in equation (2.7); and recalling from equation (2.3) that the iteration of adjoint operations can be expanded in terms of left- and right-multiplication, we also see that $\mathbf {F}$ satisfies equation (2.6). More generally, let $(a_k)_{k=0}^{\infty }\subset \mathbb {R}$ , and let $\mathbf {X}\in \mathscr {S}(\mathcal {T}_0)$ ; then the following integral

(2.10)

is well defined in the sense of equation (2.9). The definition of the integral with integrands of the form $\mathbf {F}: \Omega \times [0,T] \to \mathcal {L}(T({\mathbb {R}^d}) \overline {\otimes } T({\mathbb {R}^d}); \mathcal {T})$ with respect to processes $\mathbf {X} \in \mathscr {S}(T({\mathbb {R}^d}) \overline {\otimes } T({\mathbb {R}^d}))$ is completely analogous.

Quotient algebras: All of this extends in a straightforward way to the case of semimartingales in the quotient algebra of Section 2.3: that is, symmetric and truncated algebra, in particular given $\mathbf {X}$ and $\mathbf {Y}$ in $\mathscr {S}(\mathcal {S})$ have well-defined continuous local martingale parts denoted by $\mathbf {X}^c,\mathbf {Y}^c$ , respectively, with inner (predictable) quadratic covariation given by

$$\begin{align*}\langle\mathbf{X}^c,\mathbf{Y}^c\rangle =\sum_{w_1,w_2\in \widehat{\mathcal{W}}_d}\langle \mathbf{X}^{w_1,c},\mathbf{Y}^{w_2,c}\rangle\hat{e}_{w_1}\hat{e}_{w_2}. \end{align*}$$

Write $\mathcal {S}^N$ for the truncated symmetric algebra, linearly spanned by $\{ \hat {e}_{w}: w \in \widehat {\mathcal {W}}_d, |w| \le N\}$ and $\mathcal {S}^N_0$ for those elements with zero scalar entry. In complete analogy with non-commutative setting discussed above, we then write $\widehat {\mathscr {H}}^{q,N} \subset \mathscr {S}(\mathcal {S}^N_0)$ for the corresponding space of homogeneously q-integrable semimartingales.

2.5 Diamond products

We extend the notion of the diamond product introduced in [Reference Alos, Gatheral and Radoičić3] for continuous scalar semimartingales to our setting. Denote by $\mathbb {E}_t$ the conditional expectation with respect to the sigma algebra $\mathcal {F}_t$ . Then we have the following:

Definition 2.1. For $\mathbf {X}$ and $\mathbf {Y}$ in $\mathscr {S}(\mathcal {T}\,)$ , define

$$ \begin{align*}(\mathbf{X} \diamond \mathbf{Y})_t(T) := \mathbb{E}_t \big( \left\langle \mathbf{X}^c, \mathbf{Y}^c \right\rangle_{t,T} \big)=\sum_{w\in\mathcal{W}_d}\left( \sum_{w_1w_2=w}(\mathbf{X}^{w_1}\diamond\mathbf{Y}^{w_2})_t(T) \right)e_w \in \mathcal{T} \end{align*} $$

whenever the $\mathcal {T}$ -valued quadratic covariation that appears on the right-hand side is integrable. Similar to the previous section, we also define an outer diamond, for $\mathbf {X},\mathbf {Y}\in \mathcal {T}$ , by

This definition extends immediately to semimartingales with values in the quotient algebras of Section 2.3. In particular, given $\hat {\mathbf {X}}$ and $\hat {\mathbf {Y}}$ in $\mathscr {S}(\mathcal {S})$ , we have

$$ \begin{align*}( \hat{\mathbf{X}} \diamond \hat{\mathbf{Y}})_t(T) := \mathbb{E}_t \big( \langle \hat{\mathbf{X}}^c, \hat{\mathbf{Y}}^c \rangle_{t,T} \big) = \sum_{w_1,w_2\in \widehat{\mathcal{W}}_d} (\hat{\mathbf{X}}^{w_1} \diamond \hat{\mathbf{Y}}^{w_2})_t(T) \hat{e}_{w_1}\hat{e}_{w_2} \in \mathcal{S}, \end{align*} $$

where the last expression is given in terms of diamond products of scalar semimartingales.

Lemma 2.2. Let $p,q,r\in [1,\infty )$ such that $1/p + 1/q + 1/r < 1$ , and let $X\in \mathscr {M}_{\mathrm {loc}}^{c}(({\mathbb {R}^d})^{\otimes l})$ , $Y\in \mathscr {M}_{\mathrm {loc}}^{c}(({\mathbb {R}^d})^{\otimes m})$ , and $Z\in \mathscr {D}(({\mathbb {R}^d})^{\otimes n})$ with $l,m,n\in {\mathbb {N}}$ , such that . Then it holds that for all $0 \le t \le T$

$$ \begin{align*} \mathbb{E}_t\left(\int_t^{T}Z_{u-}\mathrm{d}(X \diamond Y)_u(T)\right) = -\mathbb{E}_t\left(\int_{t}^T Z_{u-}\mathrm{d}\left\langle X, Y \right\rangle_u\right). \end{align*} $$

Proof. Using the Kunita-Watanabe inequality (Lemma 7.1) we see that the expectation on the right-hand side is well defined. Further note that it follows from Emery’s inequality (Lemma 7.3) and Doob’s maximal inequality that the local martingale

$$ \begin{align*} \int_0^{\cdot}Z_{u-}\mathrm{d}(\mathbb{E}_u\left\langle X, Y \right\rangle_T) \end{align*} $$

is a true martingale. Recall the definition of the diamond product, and observe that the difference of the left- and right-hand sides of the above equation is a conditional expectation of a martingale interment and is hence zero.

2.6 Generalized signatures

We now give the precise meaning of equation (1.1): that is, $\mathrm d\mathbf {S}=\mathbf {S}\,{\circ \mathrm d}\mathbf {X}$ , or component-wise, for every word $w\in \mathcal {W}_d$ ,

$$\begin{align*}\mathrm d\mathbf{S}^w=\sum_{w_1w_2=w}\mathbf{S}^{w_1}\,{\circ\mathrm d}\mathbf{X}^{w_2}, \end{align*}$$

where the driving noise $\mathbf {X}$ is a $\mathcal {T}_0$ -valued semimartingale, so that $\mathbf {X}^{\emptyset } \equiv 0$ . Following [Reference Marcus49, Reference Marcus50, Reference Estrade20, Reference Kurtz, Pardoux and Protter41, Reference Friz and Shekhar25, Reference Bruned, Curry and Ebrahimi-Fard8], the integral meaning of this equation, started at time s from $\mathbf {s} \in \mathcal {T}_1$ , for times $t \ge s$ , is given by

(2.11) $$ \begin{equation} \mathbf{S}_t = \mathbf{s} + \int_{(s,t]} \mathbf{S}_{u-}\,\mathrm d \mathbf{X}_u + \frac{1}{2}\int_s^{t} \mathbf{S}_{u-}\,\mathrm d\left\langle \mathbf{X}^{c} \right\rangle_u + \sum_{s< u \le t} \mathbf{S}_{u-}\big(\exp(\Delta \mathbf{X}_u)-1-\Delta \mathbf{X}_u\big), \end{equation} $$

leaving the component-wise version to the reader. We have

Proposition 2.3. Let $\mathbf {s}\in \mathcal {T}_1$ and suppose $\mathbf {X}$ takes values in $\mathcal {T}_0$ . For every $s \ge 0$ and $\mathbf {s} \in \mathcal {T}_1$ , equation (2.11) has a unique global solution on $\mathcal {T}_1$ starting from $\mathbf {S}_s=\mathbf {s}$ .

Proof. Note that $\mathbf {S}$ solves equation (2.11) iff $\mathbf {s}^{-1} \mathbf {S}$ solves the same equation started from $1 \in \mathcal {T}_1$ . We may thus take $\mathbf {s} = 1$ without loss of generality. The graded structure of our problem, and more precisely that $\mathbf {X} = (0,X,\mathbb {X},\dots )$ in equation (2.11) has no scalar component, shows that the (necessarily) unique solution is given explicitly by iterated integration, as may be seen explicitly when writing out $\mathbf {S}^{(0)} \equiv 1$ , $\mathbf {S}^{(1)}_t = \int _s^t \mathrm d X = X_{s,t} \in \mathbb {R}^d$ ,

$$ \begin{align*}\mathbf{S}^{(2)}_t = \int_{(s,t]} \mathbf{S}^{(1)}_{u-}\,\mathrm d X_u +\mathbb{X}_{t} -\mathbb{X}_{s} + \frac{1}{2} \left\langle X^{c} \right\rangle_{s,t} + \frac{1}{2} \sum_{s< u \le t} (\Delta X_u)^2 \in (\mathbb{R}^d)^{\otimes 2}, \end{align*} $$

and so on. (In particular, we do not need to rely on abstract existence, uniqueness results for Marcus SDEs [Reference Kurtz, Pardoux and Protter41] or Lie group stochastic exponentials [Reference Hakim-Dowek and Lépingle31].)

Definition 2.4. Let $\mathbf {X}$ be a $\mathcal {T}_0$ -valued semimartingale defined on some interval $[s,t]$ . Then

$$ \begin{align*}\mathrm{Sig} (\mathbf{X} \vert_{[s,t]}) \equiv \mathrm{Sig}(\mathbf{X})_{s,t}\end{align*} $$

is defined to be the unique solution to equation (2.11) on $[s,t]$ , such that $\mathrm {Sig}(\mathbf {X})_{s,s}=1$ .

The following can be seen as a (generalized) Chen relation.

Lemma 2.5. Let $\mathbf {X}$ be a $\mathcal {T}_0$ -valued semimartingale on $[0,T]$ and $0 \le s \le t \le u \le T$ . Then the following identity holds with probability one, for all such $s,t,u$ :

(2.12) $$ \begin{align} \mathrm{Sig}(\mathbf{X})_{s,t}\mathrm{Sig}(\mathbf{X})_{t,u}=\mathrm{Sig}(\mathbf{X})_{s,u}. \end{align} $$

Proof. Call $\Phi _{t \leftarrow s} \mathbf {s} := \mathbf {S}_t$ the solution to equation (2.11) at time $t \ge s$ , started from $\mathbf {S}_s = \mathbf {s}$ . By uniqueness of the solution flow, we have $ \Phi _{u \leftarrow t} \circ \Phi _{t \leftarrow s} = \Phi _{u \leftarrow s}. $ It now suffices to remark that, thanks to the multiplicative structure of equation (2.11), we have $ \Phi _{t \leftarrow s} \mathbf {s} = \mathbf {s} \mathrm {Sig}(\mathbf {X})_{s,t}$ .

3 Expected signatures and signature cumulants

3.1 Definitions and existence

Throughout this section, let $\mathbf {X} \in \mathscr {S}(\mathcal {T}_0)$ be defined on a filtered probability space $(\Omega , \mathcal {F}, (\mathcal {F}_t)_{0 \le t \le T}, \mathbb {P})$ . Recall that $\mathbb {E}_t$ denotes the conditional expectation with respect to the sigma algebra $\mathcal {F}_t$ . When $\mathbb {E}(|\mathrm {Sig}(\mathbf {X})^w_{0,t}|)<\infty $ for all $0 \le t \le T$ and all words $w\in \mathcal {W}_d$ , then the (conditional) expected signature

$$ \begin{align*} \boldsymbol{\mu}_t(T) := \mathbb{E}_t\left(\mathrm{Sig}(\mathbf{X})_{t,T}\right) = \sum_{w\in\mathcal{W}_d}\mathbb{E}_t(\mathrm{Sig}(\mathbf{X})^w_{t,T})e_w \in \mathcal{T}_1, \quad 0 \le t \le T, \end{align*} $$

is well defined. In this case, we can also define the (conditional) signature cumulant of $\mathbf {X}$ by

$$ \begin{align*} \boldsymbol{\kappa}_{t}(T):=\log\left(\boldsymbol{\mu}_t(T)\right) \in \mathcal{T}_0, \quad 0 \le t \le T. \end{align*} $$

An important observation is the following:

Lemma 3.1. Given $\mathbb {E}(|\mathrm {Sig}(\mathbf {X})^w_{0,t}|)<\infty $ for all $0 \le t \le T$ and words $w\in \mathcal {W}_d$ , then $\boldsymbol {\mu }(T) \in \mathscr {S}(\mathcal {T}_1)$ and $\boldsymbol {\kappa }(T)\in \mathscr {S}(\mathcal {T}_0)$ .

Proof. It follows from the relation in equation (2.12) that

$$ \begin{align*} \boldsymbol{\mu}_t(T) = \mathbb{E}_t\left(\mathrm{Sig}(\mathbf{X})_{t,T}\right) = \mathbb{E}_t\left(\mathrm{Sig}(\mathbf{X})_{0,t}^{-1}\mathrm{Sig}(\mathbf{X})_{0,T}\right) = \mathrm{Sig}(\mathbf{X})_{0,t}^{-1}\mathbb{E}_t\left(\mathrm{Sig}(\mathbf{X})_{0,T}\right). \end{align*} $$

Therefore, projecting to the tensor components, we have

$$ \begin{align*} \boldsymbol{\mu}_t(T)^w = \sum_{w_1w_2 = w}(-1)^{|w_1|}\mathrm{Sig}(\mathbf{X})^{w_1}_{0,t}\mathbb{E}_t\left(\mathrm{Sig}(\mathbf{X})^{w_2}_{0,T}\right), \quad 0 \le t \le T, \quad w \in \mathcal{W}_d. \end{align*} $$

Since $(\mathrm {Sig}(\mathbf {X})^w_{0, t})_{0 \le t \le T}$ and $(\mathbb {E}_t(\mathrm {Sig}(\mathbf {X})^w_{0,T})_{0 \le t \le T}$ are semimartingales (the latter in fact a martingale), it follows from Itô’s product rule that $\boldsymbol {\mu }^w(T)$ is also a semimartingale for all words $w\in \mathcal {W}_d$ , hence $\boldsymbol {\mu }(T)\in \mathscr {S}(\mathcal {T}_1)$ . Further recall that $\boldsymbol {\kappa }(T) = \log (\boldsymbol {\mu }(T))$ , and therefore it follows from the definition of the logarithm on $\mathcal {T}_1$ that each component $\boldsymbol {\kappa }(T)^w$ with $w\in \mathcal {W}_d$ is a polynomial of $(\boldsymbol {\mu }(T)^{v})_{v\in \mathcal {W}_d, |v|\le |w|}$ . Hence it follows again by Itô’s product rule that $\boldsymbol {\kappa }(T)\in \mathscr {S}(\mathcal {T}_0)$ .

It is of strong interest to have a more explicit necessary condition for the existence of the expected signature. The following theorem, the proof of which can be found in Section 7.1, yields such a criterion.

Theorem 3.2. Let $q\in [1, \infty )$ and $N\in {\mathbb {N}_{\ge 1}}$ ; then there exist two constants $c,C>0$ depending only on d, N and q, such that for all $\mathbf {X} \in \mathscr {H}^{q,N}$

$$ \begin{align*} c\vert\mkern-2.5mu\vert\mkern-2.5mu\vert{\mathbf{X}}\vert\mkern-2.5mu\vert\mkern-2.5mu\vert_{\mathscr{H}^{q,N}} \le \vert\mkern-2.5mu\vert\mkern-2.5mu\vert{\mathrm{Sig}(\mathbf{X})_{0,\cdot}}\vert\mkern-2.5mu\vert\mkern-2.5mu\vert_{\mathscr{H}^{q,N}} \le C\vert\mkern-2.5mu\vert\mkern-2.5mu\vert{\mathbf{X}}\vert\mkern-2.5mu\vert\mkern-2.5mu\vert_{\mathscr{H}^{q,N}}. \end{align*} $$

In particular, if $\mathbf {X}\in \mathscr {H}^{\infty -}(\mathcal {T}_0)$ , then $\mathrm {Sig}(\mathbf {X})_{0,\cdot }\in \mathscr {H}^{\infty -}(\mathcal {T}_1)$ , and the expected signature exists.

Remark 3.3. Let $\mathbf {X} = (0, M, 0, \dotsc , 0)$ , where $M \in \mathscr {M}({\mathbb {R}^d})$ is a martingale; then

and we see that the above estimate implies that

$$ \begin{align*} \max_{n=1, \dotsc, N} \left\Vert {\mathrm{Sig}(\mathbf{X})^{(n)}_{0, \cdot}} \right\Vert _{\mathscr{S}^{qN/n}}^{1/n} \le C \left\Vert {M} \right\Vert _{\mathscr{H}^{qN}}. \end{align*} $$

This estimate is already known and follows from the Burkholder-Davis-Gundy inequality for enhanced martingales, which was first proved in the continuous case in [Reference Friz and Victoir26] and for the general case in [Reference Chevyrev and Friz12].

Remark 3.4. When $q>1$ , the above estimate also holds true when the signature $\mathrm {Sig}(\mathbf {X})_{0,\cdot }$ is replaced by the conditional expected signature $\boldsymbol {\mu }(T)$ or the conditional signature cumulant $\boldsymbol {\kappa }(T)$ . This will be seen in the proof of Theorem 4.1 below (more precisely in Claim 7.12).

3.2 Moments and cumulants

We quickly discuss the development of a symmetric algebra valued semimartingale, more precisely, $\hat {\mathbf {X}} \in \mathscr {S}(\mathcal {S}_0)$ , in the group $\mathcal {S}_1$ . That is, we consider

(3.1) $$ \begin{align} \mathrm{d} \hat{\mathbf{S}} = \hat{\mathbf{S}}\,\circ \mathrm{d} \hat{\mathbf{X}}. \end{align} $$

It is immediate (validity of chain rule) that the unique solution to this equation, at time $t \ge s$ , started at $\hat {\mathbf {S}}_s = \hat {\mathbf {s}} \in \mathcal {S}_1$ is given by

$$\begin{align*}\hat{\mathbf{S}}_{t}:= \exp\left( \hat{\mathbf{X}}_t- \hat{\mathbf{X}}_s \right)\hat{\mathbf{s}} \in \mathcal{S}_1, \end{align*}$$

and we also write $\hat {\mathbf {S}}_{s,t} = \exp ( \hat {\mathbf {X}}_t-\hat {\mathbf {X}}_s )$ for this solution started at time s from $1\in \mathcal {S}_1$ . The relation to signatures is as follows. Recall that the $\pi _{\mathrm {Sym}}$ denotes the canonical projection from $\mathcal {T}$ to $\mathcal {S}$ .

Proposition 3.5. Let $\mathbf {X},\mathbf Y\in \mathscr {S}(\mathcal {T}\,)$ , and define $\hat {\mathbf {X}} := \pi _{\mathrm {Sym}}(\mathbf {X})$ and $\hat {\mathbf {Y}} := \pi _{\mathrm {Sym}}(\mathbf {Y})$ . Then it holds that

  1. (i) $\hat {\mathbf {X}}, \hat {\mathbf {Y}}\in \mathscr {S}(\mathcal {S})$ , and for the indefinite Itô integral, we have in the sense of indistinguishable processes

    (3.2) $$ \begin{align} \pi_{\mathrm{Sym}} \int \mathbf{X} \mathrm d\mathbf Y= \int \hat{\mathbf{X}}\,\mathrm d\hat{\mathbf Y}, \end{align} $$
  2. (ii) $\hat {\mathbf {S}} := \pi _{\mathrm {Sym}}{\mathrm {Sig}(\mathbf {X})_{s,\cdot }}$ solves equation (3.1) started at time s from $1\in \mathcal {S}_1$ and driven by $\hat {\mathbf {X}}$ . In particular,

    $$ \begin{align*}\hat{\mathbf{S}}_{s,t}=\exp(\hat{\mathbf{X}}_t-\hat{\mathbf{X}}_s)\end{align*} $$
    for all $0 \le s \le t \le T$ .

Proof. (i) That the projections $\hat {\mathbf {X}},\hat {\mathbf Y}$ define $\mathcal {S}$ -valued semimartingales follows from the componentwise definition and the fact that the canonical projection is linear. In particular, the right-hand side of equation (3.2) is well defined. To show equation (3.2), we apply the canonical projection $\pi _{\mathrm {Sym}}$ to both sides of equation (2.9) after choosing $Z_t\equiv \mathbf 1$ ; and using the explicit action of $\pi _{\mathrm {Sym}}$ on basis tensors, we obtain the identity

$$\begin{align*}\pi_{\mathrm{Sym}}\int\mathbf{X}\,\mathrm d\mathbf{Y}=\sum_{w\in\mathcal{W}^d}\left( \sum_{uv=w}\int \mathbf{X}^u\,\mathrm d\mathbf{Y}^v \right)\hat{e}_{\hat{w}}=\int\hat{\mathbf{X}}\,\mathrm d\hat{\mathbf{Y}} \end{align*}$$

by equation (2.4). Part (ii) is then immediate.

Assuming componentwise integrability, we then define symmetric moments and cumulants of the $\mathcal {S}$ -valued semimartingale $\hat {\mathbf {X}}$ by

$$ \begin{align*} \hat{\boldsymbol{\mu}}_t(T) & := \mathbb{E}_t\left(\exp\left( \hat{\mathbf{X}}_T- \hat{\mathbf{X}}_t\right)\right) = \sum_{v\in\hat{\mathcal{W}}_d} \mathbb{E}_t \left( \exp\left( \hat{\mathbf{X}}_T-\hat{\mathbf{X}}_t\right)^v_{t,T}\right )\hat{e}_v \in \mathcal{S}_1,\\ \hat{\boldsymbol{\kappa}}_t(T) & := \log\left( \hat{\boldsymbol{\mu}}_t(T) \right)\in \mathcal{S}_0, \end{align*} $$

for $0\le t \le T$ . If $\hat {\mathbf {X}} = \pi _{\mathrm {Sym}}(\mathbf {X})$ , for $ \mathbf {X} \in \mathscr {S}(\mathcal {T}\,)$ , with expected signature and signature cumulants $\boldsymbol {\mu }(T)$ and $\boldsymbol {\kappa }(T)$ , it is then clear that the symmetric moments and cumulants of $\hat {\mathbf {X}}$ are obtained by projection,

$$ \begin{align*}\hat{\boldsymbol{\mu}}(T) = \pi_{\mathrm{Sym}}( \boldsymbol{\mu}(T)), \quad \hat{\boldsymbol{\kappa}}(T) = \pi_{\mathrm{Sym}}(\boldsymbol{\kappa}(T)).\end{align*} $$

Example 3.6. Let $ X$ be an $\mathbb {R}^d$ -valued martingale in $\mathscr H^{\infty -}$ , and $\hat {\mathbf {X}}_t:=\sum _{i=1}^dX^i_t\hat {e}_i$ . Then

$$\begin{align*}\hat{\boldsymbol{\mu}}_t(T)=\sum_{n=0}^\infty\frac{1}{n!}\mathbb{E}_t(X_T-X_t)^n=1+\sum_{n=1}^\infty\frac{1}{n!}\sum_{i_1,\dotsc,i_n=1}^d\mathbb{E}_t\left[ (X_T^{i_1}-X_t^{i_1})\dotsm(X_T^{i_n}-X_t^{i_n}) \right]\hat{e}_{\widehat{i_1\dotsm i_n}} \end{align*}$$

consists of the (time-t conditional) multivariate moments of $X_T-X_t \in \mathbb {R}^d$ . Here, the series on the right-hand side is understood in the formal sense. It readily follows, also noted in [Reference Bonnier and Oberhauser7, Example 3.3], that $\hat {\boldsymbol {\kappa }}_t (T) = \log (\hat {\boldsymbol {\mu }}_t(T))$ consists precisely of the multivariate cumulants of $X_T-X_t$ . Note that the symmetric moments and cumulants of the scaled process $a X$ , $a \in \mathbb {R}$ , are precisely given by $\delta _a \hat {\boldsymbol {\mu }}$ and $\delta _a \hat {\boldsymbol {\kappa }}$ , where the linear dilation map is defined by $\delta _a\colon \hat {e}_w \mapsto a^{|w|} \hat {e}_w$ . The situation is similar for $a \cdot X=(a_1X^1,\dotsc ,a_dX^d)$ , $a \in \mathbb {R}^d$ , but now with $\delta _a\colon \hat {e}_w \mapsto a^w \hat {e}_1^{|w|}$ with $a^w = a_1^{n_1} \cdots a_d^{n_d}$ , where $n_i$ denotes the multiplicity of the letter $i \in \{1,\dots , d\}$ in the word w.

We next consider linear combinations, $\hat {\mathbf {X}} = a X + b \langle X \rangle $ , for general pairs $a,b \in \mathbb {R}$ , having already dealt with $b=0$ . The special case $b = - a^2/2$ , by scaling there is no loss in generality to take $(a,b) = (1,-1/2)$ , yields a (at least formally) familiar exponential martingale identity.

Example 3.7. Let $ X$ be an $\mathbb {R}^d$ -valued martingale in $\mathscr H^{\infty -}$ , and define

$$\begin{align*}\hat{\mathbf{X}}_t:=\sum_{i=1}^dX^i_t\hat{e}_i-\frac12\sum_{1\le i\le j\le d}\langle X^i,X^j\rangle_t\hat{e}_{ij}. \end{align*}$$

In this case, we have trivial symmetric cumulants $\hat {\boldsymbol {\kappa }}_t(T)=0$ for all $0\le t\le T$ . Indeed, Itô’s formula shows that $t\mapsto \exp (\hat {\mathbf {X}}_t)$ is an $\mathcal {S}_1$ -valued martingale, so that

$$\begin{align*}\hat{\boldsymbol{\mu}}_t(T)=\mathbb{E}_t\exp(\hat{\mathbf{X}}_T-\hat{\mathbf{X}}_t)=\exp(-\hat{\mathbf{X}}_t)\mathbb{E}_t\exp(\hat{\mathbf{X}}_T)=1. \end{align*}$$

While the symmetric cumulants of the last example carries no information, it suffices to work with

$$ \begin{align*}\hat{\mathbf{X}} = \sum_{i=1}^d a^i X^{i}\hat{e}_i + \sum_{1\le i\le j\le d} b_{jk} \langle X^j,X^k \rangle \hat{e}_{ij}, \end{align*} $$

in which case $\hat {\boldsymbol {\mu }} = \hat {\boldsymbol {\mu }} (a,b), \hat {\boldsymbol {\kappa }} = \hat {\boldsymbol {\kappa }}(a,b)$ contains full information of the joint moments of X and its quadratic variation process. A recursion of these was constructed as a diamond expansion in [Reference Friz, Gatheral and Radoičić23].

4 Main results

4.1 Functional equation for signature cumulants

Let $\mathbf {X}\in \mathscr {S}(\mathcal {T}_0)$ defined on a filtered probability space $(\Omega , \mathcal {F}, (\mathcal {F}_t)_{0\le t\le T<\infty },\mathbb {P})$ satisfying the usual conditions. For all $\mathbf {x}\in \mathcal {T}_0$ (or $\mathcal {T}_0^N$ ), define the following operators, with Bernoulli numbers $(B_k)_{k\ge 0} = (1, -\frac {1}{2}, \frac {1}{6}\dotsc )$ ,

(4.1) $$ \begin{align} \begin{aligned} G(\operatorname{\mathrm{ad}}{\mathbf{x}}) = \sum_{k = 0}^{\infty} \frac{(\operatorname{\mathrm{ad}}{\mathbf{x}})^k}{(k + 1) !}, \quad &Q(\operatorname{\mathrm{ad}}{\mathbf{x}}) = \sum_{m, n = 0}^{\infty}2\frac{(\operatorname{\mathrm{ad}}{\mathbf{x}})^n\odot (\operatorname{\mathrm{ad}}{\mathbf{x}})^m}{(n + 1) ! (m) ! (n + m + 2)},\\ H(\operatorname{\mathrm{ad}}{\mathbf{x}}) &:= \sum_{k=0}^{\infty}\frac{B_k}{k!}(\operatorname{\mathrm{ad}}{\mathbf{x}})^{k}, \end{aligned} \end{align} $$

noting $G(z) = (\exp (z)-1)/z$ , $H(z) = G^{-1}(z) = z/(\exp (z)-1)$ . Our main result is the following.

Theorem 4.1. Let $\mathbf {X} \in \mathscr {H}^{\infty -}(\mathcal {T}_0)$ ; then the signature cumulant $\boldsymbol {\kappa } =\boldsymbol {\kappa } (T) = (\log \mathbb {E}_t(\mathrm {Sig}(\mathbf {X})_{t,T}))_{0 \le t \le T}$ is the unique solution (up to indistinguishably) of the following functional equation: for all $0 \le t \le T$

(4.2)

Equivalently, $\boldsymbol {\kappa } =\boldsymbol {\kappa } (T) $ is the unique solution to

(4.3)

Furthermore, if $\mathbf {X}\in \mathscr {H}^{1,N}$ for some $N\in {\mathbb {N}_{\ge 1}}$ , then the identities in equations (4.2) and (4.3) still hold true for the truncated signature cumulant $\boldsymbol {\kappa } := (\log \mathbb {E}_t(\mathrm {Sig}(\mathbf {X}^{(0,N)})_{t,T}))_{0\le t \le T}$ .

Proof. We postpone the proof of the fact that $\boldsymbol {\kappa }$ satisfies equations (4.2) and (4.3) to section Section 7.2. The uniqueness part of the statement can be easily seen as follows. Regarding equation (4.2), we first note that it holds

$$ \begin{align*} \mathbb{E}_t \left\{ \int_{(t,T]} G(\operatorname{\mathrm{ad}} \boldsymbol{\kappa}_{u-})(\mathrm{d} \boldsymbol{\kappa}_u)\right\} = \mathbb{E}_t \left\{ \int_{(t,T]} (G(\operatorname{\mathrm{ad}} \boldsymbol{\kappa}_{u-})-\mathrm{Id})(\mathrm{d} \boldsymbol{\kappa}_u) \right\} - \boldsymbol{\kappa}_t, \quad 0 \le t \le T, \end{align*} $$

where we have used that $\boldsymbol {\kappa }_T \equiv 0$ (and the fact that the conditional expectation is well defined, which is shown in the first part of the proof). Hence, after subtracting the identity from G, we can bring $\boldsymbol {\kappa }_t$ to the left-hand side in equation (4.2). This identity is an equality of tensor series in $\mathcal {T}_0$ and can be projected to yield an equality for each tensor level of the series. As presented in more detail in the following subsection, we see that projecting the latter equation to tensor level, say $n\in {\mathbb {N}_{\ge 1}}$ , the right-hand side only depends on $\boldsymbol {\kappa }^{(k)}$ for $k < n$ , hence giving an explicit representation $\boldsymbol {\kappa }^{(n)}$ in terms of $\mathbf {X}$ and strictly lower tensor levels of $\boldsymbol {\kappa }$ . Therefore equation (4.2) characterizes $\boldsymbol {\kappa }$ up to a modification and then due to right-continuity up to indistinguishably. The same argument applies to equation (4.3), referring to the following subsections for details on the recursion.

Diamond formulation: The functional equations given in Theorem 4.1 above can be phrased in terms of the diamond product between $\mathcal {T}_0$ -valued semimartingales. Writing $\mathbf {J}_t(T) = \sum _{t < u \le T} (\dots )$ for the last (jump) sum in equation (4.2), this equation can be written, thanks to Lemma 2.2, which applies just the same with outer diamonds,

and a similar form may be given for equation (4.3). While one may or may not prefer this equation to equation (4.2), diamonds become very natural in $d=1$ (or upon projection to the symmetric algebra; see also Section 5.2). In this case, $G = \mathrm {Id}$ , $Q = \mathrm {Id} \odot \mathrm {Id}$ ; and with identities of the form

some simple rearrangement, using bilinearity of the diamond product, gives

(4.4) $$ \begin{align} \boldsymbol{\kappa}_t (T) = \mathbb{E}_t \{ \mathbf{X}_{t,T} \} + \frac{1}{2} ((\mathbf{X} + \boldsymbol{\kappa}) \diamond (\mathbf{X} + \boldsymbol{\kappa}))_t(T) + \mathbb{E}_t \{ \mathbf{J}_t(T) \}. \end{align} $$

If we further impose martingality and continuity, we arrive at

$$ \begin{align*}\boldsymbol{\kappa}_t (T) = \frac{1}{2} ((\mathbf{X} + \boldsymbol{\kappa}) \diamond (\mathbf{X} + \boldsymbol{\kappa}))_t(T). \end{align*} $$

4.2 Recursive formulas for signature cumulants

Theorem 4.1 allows for an iterative computation of signature cumulants, trivially started from

$$ \begin{align*} \boldsymbol{\kappa}^{(1)}_t & = \boldsymbol{\mu}^{(1)}_t = \mathbb{E}_t\left(\mathbf{X}^{(1)}_{t, T}\right). \end{align*} $$

The second signature cumulant, obtained from Theorem 4.1 or first principles, reads

$$ \begin{align*} \boldsymbol{\kappa}^{(2)}_t & = \mathbb{E}_t \bigg\{\mathbf{X}^{(2)}_{t,T} + \frac{1}{2}\left\langle \mathbf{X}^{(1)c} \right\rangle_{t,T} +\frac12 \int_{(t,T]}\left[ \boldsymbol{\kappa}^{(1)}_{u-}, \mathrm{d}\boldsymbol{\kappa}^{(1)}_u \right] + \frac{1}{2}\left\langle \boldsymbol{\kappa}^{(1)c} \right\rangle_{t,T} + \left\langle \mathbf{X}^{(1)c}, \boldsymbol{\kappa}^{(1)c} \right\rangle_{t,T} \\ &\hspace{3em}+ \sum_{t<u\le T}\bigg(\frac{1}{2}\left(\Delta \mathbf{X}^{(1)}_u\right)^{2} + \Delta \mathbf{X}^{(1)}_u\Delta\boldsymbol{\kappa}^{(1)}_u + \frac{1}{2}\left(\Delta\boldsymbol{\kappa}^{(1)}_u\right)^{2} \bigg)\bigg\}. \end{align*} $$

For instance, consider the special case with vanishing higher-order components, $\mathbf {X}^{(i)} \equiv 0$ , for $i \ne 1$ , and $\mathbf {X} = \mathbf {X}^{(1)} \equiv M$ , a d-dimensional continuous square-integrable martingale. In this case, $\boldsymbol {\kappa }^{(1)} = \boldsymbol {\mu }^{(1)} \equiv 0$ , and from the very definition of the logarithm relating $\boldsymbol {\kappa }$ and $\boldsymbol {\mu }$ , we have $\boldsymbol {\kappa }^{(2)} = \boldsymbol {\mu }^{(2)} - \frac 12 \boldsymbol {\mu }^{(1)} \boldsymbol {\mu }^{(1)} = \boldsymbol {\mu }^{(2)}$ . It then follows from Stratonovich-Ito correction that

$$ \begin{align*}\boldsymbol{\kappa}^{(2)}_t = \mathbb{E}_t \int_t^T (M_u - M_t) \circ \mathrm{d} M_u = \frac12 \mathbb{E}_t \left\langle M \right\rangle_{t,T} = \frac12 \mathbb{E}_t \left\langle \mathbf{X}^{(1)} \right\rangle_{t,T}, \end{align*} $$

which is indeed a (very) special case of the general expression for $\boldsymbol {\kappa }^{(2)}$ . We now treat general higher-order signature cumulants.

Corollary 4.2. Let $\mathbf {X}\in \mathscr {H}^{1,N}$ for some $N\in \mathbb {N}_{\ge 1}$ ; then we have

$$ \begin{align*} \boldsymbol{\kappa}^{(1)}_t & = \mathbb{E}_t\left(\mathbf{X}^{(1)}_{t, T}\right), \end{align*} $$

for all $0 \le t \le T$ and for $n \in \{2, \dotsc , N\}$ , we have recursively (the right-hand side only depends on $\boldsymbol {\kappa }^{(j)},j<n$ )

(4.5) $$ \begin{align} \boldsymbol{\kappa}^{(n)}_t &= \mathbb{E}_t\left(\mathbf{X}^{(n)}_{t,T }\right) + \frac12\sum_{k = 1}^{n-1}\mathbb{E}_t\left( \left\langle \mathbf{X}^{(k)c}, \mathbf{X}^{(n-k)c} \right\rangle_{t,T}\right) \nonumber\\ & \quad +\sum_{|\ell|\ge 2, \; \|\ell\|=n}\mathbb{E}_t\Big(\mathrm{Mag}(\boldsymbol{\kappa}; \ell)_{t,T} + \mathrm{Qua}(\boldsymbol{\kappa}; \ell)_{t,T} + \mathrm{Cov}(\mathbf{X}, \boldsymbol{\kappa}; \ell)_{t,T} + \mathrm{Jmp}(\mathbf{X},\boldsymbol{\kappa}; \ell)_{t,T}\Big) \end{align} $$

with $\ell = (l_1, \dotsc , l_k)$ , $l_i \in {\mathbb {N}_{\ge 1}}$ , $|\ell |:= k\in {\mathbb {N}_{\ge 1}}$ , $\|\ell \|:= l_1 + \dotsb + l_k$ and

Proof. As in the proof of Theorem 4.1 above, in equation (4.2), we can separate the identity from G and bring the resulting $\boldsymbol {\kappa }_t$ to the left-hand side. Projecting to the tensor level n, and using that the projection can be interchanged with taking the expectation, we obtain the following equation

$$ \begin{align*} \boldsymbol{\kappa}^{(n)}_t = \mathbb{E}_t\left\{\mathbf{X}^{(n)}_{t,T} + \frac{1}{2} \left\langle \mathbf{X}^{c} \right\rangle^{(n)}_{t,T} + \mathbf{Y}^{(n)}_{t,T} + \mathbf{V}^{(n)}_{t,T} + \mathbf{C}^{(n)}_{t,T} + \mathbf{J}^{(n)}_{t,T}\right\}, \end{align*} $$

for all $0 \le t \le T$ , where $\mathbf {Y}\in \mathscr {S}(\mathcal {T}_0^{N})$ and $\mathbf {V}, \mathbf {C}, \mathbf {J} \in \mathscr {V}(\mathcal {T}_0^{N})$ are defined in equation (7.9). Note that we can take the conditional expectation of each term separately, as $\mathbf {X}^{(n)}$ has sufficient integrability by the assumption, and the integrability of the remaining terms is shown in the Proof of Theorem 4.1 below, more precisely in equation (7.27). The recursion then follows by spelling out the explicit composition for each of the terms appearing in the above equation. For the quadratic variation term, we may easily verify from the bilinearity of the covariation bracket that it holds

$$ \begin{align*} \left\langle \mathbf{X}^{c} \right\rangle^{(n)}_{t,T} = \pi_n\sum_{k,j = 1}^{n}\left\langle \mathbf{X}^{(k)c}, \mathbf{X}^{(j)c} \right\rangle_{t,T} = \sum_{k=1}^{n}\left\langle \mathbf{X}^{(k)c}, \mathbf{X}^{(n-k)c} \right\rangle_{t,T}. \end{align*} $$

We can also take the conditional expectation of each term in the above right-hand side separately, as the integrability of these terms follows from the assumptions on $\mathbf {X}$ and Lemma 7.1. The composition of the terms $\mathbf {Y}^{(n)}, \mathbf {V}^{(n)}$ and $\mathbf {C}^{(n)}$ follows from the explicit form of the stochastic Itô integral of a power series of adjoint operations with respect to a tensor valued semimartingale in Section 2.4, more specifically in equation (2.10). We will demonstrate this for the term $\mathbf {Y}^{(n)}$ in more detail:

for all $0 \le t \le T$ . The composition of the terms $\mathbf {V}^{(n)}$ and $\mathbf {C}^{(n)}$ , respectively, in terms of $\mathrm {Qua}(\mathbf {X}, \boldsymbol {\kappa })$ and $\mathrm {Cov}(\mathbf {X}, \boldsymbol {\kappa })$ follows analogously. It then remains to show that the term $\mathbf {J}^{(n)}$ can be composed in terms of $\mathrm {Jmp}(\mathbf {X}, \boldsymbol {\kappa })$ , which is, however, a simple combinatorial exercise.

We obtain another recursion for the signature cumulants from projecting the functional equation (4.3). Note that, apart from the first two levels, it is far from trivial to see that the following recursion is equivalent to the recursion in Corollary 4.2.

Corollary 4.3. Let $\mathbf {X}\in \mathscr {H}^{1,N}$ for some $N\in \mathbb {N}_{\ge 1}$ ; then we have

(4.6) $$ \begin{align} \boldsymbol{\kappa}^{(n)}_t = \mathbb{E}_t\left(\mathbf{X}^{(n)}_{t,T}\right) + \sum_{ |\ell|\ge 2,\; ||\ell||=n} \mathbb{E}_t\bigg( \mathrm{HMag}^{1}(\mathbf{X}, \boldsymbol{\kappa}; \ell)_{t,T} + \frac{1}{2}\mathrm{HMag}^{2}(\mathbf{X}, \boldsymbol{\kappa}; \ell)_{t,T} + \mathrm{HQua}(\boldsymbol{\kappa}; \ell)_{t,T}\nonumber\\ {} + \mathrm{HCov}(\mathbf{X}, \boldsymbol{\kappa}; \ell)_{t,T}+ \mathrm{HJmp}(\mathbf{X}, \boldsymbol{\kappa}; \ell)_{t,T}\bigg) \end{align} $$

with $\ell = (l_1, \dotsc , l_k)$ , $l_i \ge 1$ , $|\ell |=k$ , $||\ell ||=l_1 + \dotsb + l_k$ and

Proof. The recursion follows from projecting equation (4.3) to each tensor level, analogously to the way that the recursion of Corollary 4.2 follows from equation (4.2) (see the proof of Corollary 4.2).

Diamonds. All recursions here can be rewritten in terms of diamonds. In a first step, by definition, the second term in Corollary 4.2 can be rewritten as

$$\begin{align*}\frac12\sum_{k=1}^n(\mathbf{X}^{(k)}\diamond\mathbf{X}^{(n-k)})_t(T). \end{align*}$$

Thanks to Lemma 2.2, we may also write

Similarly,

Inserting these expressions into Equation (4.6), we may obtain a ‘diamond’ form of the recursions in H form.

When $d=1$ (or in the projection onto the symmetric algebra; see also Section 5.2), the recursions take a particularly simple form, since $\operatorname {\mathrm {ad}}\mathbf {x}\equiv 0$ for all $\mathbf {x}\in \mathcal {T}_0$ , for $d=1$ a commutative algebra. Equation (4.5) then becomes

$$ \begin{align*}\boldsymbol{\kappa}^{(n)}_t (T) = \mathbb{E}_t\left(\mathbf{X}_{t,T}^{(n)}\right)+\frac12\sum_{k=1}^{n-1} ( (\mathbf{X}^{(k)} + \boldsymbol{\kappa}^{(k)} ) \diamond (\mathbf{X}^{(n-k)} + \boldsymbol{\kappa}^{(n-k)}))_t(T) + \mathbb{E}_t\left( \mathbf{J}^{(n)}_t (T) \right), \end{align*} $$

where $\mathbf {J}^{(n)}_t(T) = \sum _{|\ell |\ge 2, \; \|\ell \|=n} \mathrm {Jmp}(X,\boldsymbol {\kappa }; \ell )_{t,T}$ contains the nth tensor component of the jump contribution. The above diamond recursion can also be obtained by projecting the functional relation (4.4) to the nth tensor level. We shall revisit this in a multivariate setting and comment on related works in Section 5.2.

5 Two special cases

5.1 Variations on Hausdorff, Magnus and Baker–Campbell–Hausdorff

We now consider a deterministic driver $\mathbf {X}$ of finite variation. This includes the case when $\mathbf {X}$ is absolutely continuous, in which case we recover, up to a harmless time reversal, $t \leftrightarrow T-t$ , Hausdorff’s ODE and the classical Magnus expansion for the solution to a linear ODE in a Lie group [Reference Hausdorff32, Reference Magnus48, Reference Chen11, Reference Iserles and Nørsett33]. Our extension with regard to discontinuities seems to be new and somewhat unifies Hausdorff’s equation with multivariate Baker–Campbell–Hausdorff integral formulas.

Theorem 5.1. Let $\mathbf {X} \in \mathscr {V} (\mathcal {T}_0)$ , and more specifically $\mathbf {X}\colon [0,T] \to \mathcal {T}_0$ deterministic, càdlàg of bounded variation. The log-signature $\Omega _t=\Omega _{t}(T):= \log (\mathrm {Sig}(\mathbf {X})_{t,T})$ satisfies the integral equation

(5.1) $$ \begin{align} \Omega_{t}(T) &= \int_t^{T}H(\operatorname{\mathrm{ad}}{\Omega_{u-}})(\mathrm{d}\mathbf{X}^c_u) + \sum_{t<u\le T} \int_0^1\Psi(\exp(\operatorname{\mathrm{ad}} \theta \Delta \mathbf{X}_u)\circ\exp(\operatorname{\mathrm{ad}}\Omega_u))(\Delta \mathbf{X}_u)\,\mathrm d\theta, \end{align} $$

with $\Psi (z):= H(\log z)={\log z}/{(z-1)}$ as in the introduction. The sum in equation (5.1) is absolutely convergent, over (at most countably many) jump times of $\mathbf {X}$ , and vanishes when $\mathbf {X} \equiv \mathbf {X}^c$ , in which case equation (1.4) reduces to Hausdorff’s ODE.

(i) The accompanying Jump Magnus expansion becomes $\Omega ^{(1)}_{t}(T) = \mathbf {X}^{(1)}_{t,T}$ followed by

$$ \begin{align*}\Omega^{(n)}_{t}(T) = \mathbf{X}^{(n)}_{t,T} + \sum_{|\ell|\ge 2, \Vert\ell\Vert=n}\left(\mathrm{HMag}^{1}(\mathbf{X}, \Omega; \ell)_{t,T} + \mathrm{HJmp}(\mathbf{X}, \Omega; \ell)_{t,T}\right) \end{align*} $$

where the right-hand side only depends on $\Omega ^{(k)}, k<n$ .

(ii) If $\mathbf {X} \in \mathscr {V} (V)$ for some linear subspace $V \subset \mathcal {T}_0 = T_0\mathopen {(\mkern -3mu(}\mathbb {R}^d\mathclose {)\mkern -3mu)}$ , it follows that, for all $t\in [0,T]$ ,

$$ \begin{align*}\Omega_{t}(T) \in \mathcal{L} := \mathrm{Lie}\mathopen{(\mkern-3mu(} V\mathclose{)\mkern-3mu)} \subset \mathcal{T}_0, \qquad \mathrm{Sig}(\mathbf{X})_{t,T} \in \exp (\mathcal{L}) \subset \mathcal{T}_1, \end{align*} $$

and we say that $\Omega _{t}(T)$ is Lie in V. In case $V=\mathbb {R}^d$ , one speaks of (free) Lie series; see also [Reference Lyons45, Def. 6.2].

Proof. Since we are in a purely deterministic setting, the signature cumulant coincides with the log-signature $\boldsymbol {\kappa }_t(T) = \Omega _{t}(T)$ , and Theorem 4.1 applies without any expectation and angle brackets.

Using $\Delta \Omega _u = \Omega _{u} - \Omega _{u-} = \Omega _u - \log (\mathrm e^{\Delta \mathbf {X}_u}\mathrm e^{\Omega _u})$ , we see that

$$ \begin{align*} \Omega_{t}(T) &= \int_t^{T}H(\operatorname{\mathrm{ad}}{\Omega_{u-}})(\mathrm{d}\mathbf{X}^c_u) - \sum_{t<u\le T} \Delta\Omega_u \\ &= \int_t^{T}H(\operatorname{\mathrm{ad}}{\Omega_{u-}})(\mathrm{d}\mathbf{X}^c_u) - \sum_{t<u\le T} \left( \Omega_u - \operatorname{BCH}(\Delta \mathbf{X}_u,\Omega_u) \right) \\ & = \int_t^{T}H(\operatorname{\mathrm{ad}}{\Omega_{u-}})(\mathrm{d}\mathbf{X}^c_u) + \sum_{t<u\le T} \int_0^1\Psi(\exp(\theta \operatorname{\mathrm{ad}}\Delta \mathbf{X}_u)\circ\exp(\operatorname{\mathrm{ad}}\Omega_u))(\Delta\mathbf{X}_u)\,\mathrm d\theta, \end{align*} $$

where we used the identity

(5.2) $$ \begin{align} \operatorname{BCH}(\mathbf{x}_1,\mathbf{x}_2) \!-\! \mathbf{x}_2 \!=\! \log(\exp(\mathbf{x}_1)\exp(\mathbf{x}_2))\!-\! \mathbf{x}_2 \!=\! \int_0^1\Psi(\exp(\theta\operatorname{\mathrm{ad}}\mathbf{x}_1)\circ\exp(\operatorname{\mathrm{ad}}\mathbf{x}_2))(\mathbf{x}_1)\,\mathrm d\theta. \end{align} $$

Remark 5.2 (Baker–Campbell–Hausdorff)

The identity equation (5.2) is well-known but also easy to obtain en passant, thereby rendering the above proof self-contained. We treat directly the n-fold case. Given $\mathbf {x}_1,\dotsc ,\mathbf {x}_n\in \mathcal {T}_0$ , one defines a continuous piecewise affine linear path $(\mathbf {X}_t: 0 \le t \le n)$ with $\mathbf {X}_i - \mathbf {X}_{i-1} = \mathbf {x}_i$ . Then $\mathrm {Sig} ( \mathbf {X} |_{[i-1,i]})=\mathrm {Sig} ( \mathbf {X})_{i-1,i}=\exp (\mathbf {x}_i)$ and by Lemma 2.5 have

$$\begin{align*}\Omega_{0}=\log\left( \exp(\mathbf{x}_1)\dotsm\exp(\mathbf{x}_n) \right)=:\operatorname{BCH}(\mathbf{x}_1,\dotsc,\mathbf{x}_n). \end{align*}$$

A computation based on Theorem 5.1, but now applied without jumps, reveals the general form

$$ \begin{align*} \operatorname{BCH}(\mathbf{x}_1,\dotsc,\mathbf{x}_n) &= \mathbf{x}_n + \sum_{k=1}^{n-1}\int_0^1\Psi(\exp(\theta\operatorname{\mathrm{ad}}\mathbf{x}_k)\circ\exp(\operatorname{\mathrm{ad}}\mathbf{x}_{k+1})\circ\dotsm\circ\exp(\operatorname{\mathrm{ad}}\mathbf{x}_n))(\mathbf{x}_k)\,\mathrm d\theta \\ &=\sum_i\mathbf{x}_i+\frac12\sum_{i<j}[\mathbf{x}_i,\mathbf{x}_j]+\frac{1}{12}\sum_{i<j}([\mathbf{x}_i,[\mathbf{x}_i,\mathbf{x}_j]]+[\mathbf{x}_j,[\mathbf{x}_j,\mathbf{x}_i]])\\&\quad+\frac16\sum_{i<j<k}([\mathbf{x}_i,[\mathbf{x}_j,\mathbf{x}_k]]-[\mathbf{x}_k,[\mathbf{x}_i,\mathbf{x}_j]])-\frac1{24}\sum_{i<j}[\mathbf{x}_i,[\mathbf{x}_j,[\mathbf{x}_i,\mathbf{x}_j]]]\dotsb \end{align*} $$

The flexibility of our Theorem 5.1 is then nicely illustrated by the fact that this n-fold BCH formula is an immediate consequence of equation (5.1), applied to a piecewise constant càdlàg path $(\mathbf {X}_t: 0 \le t \le n)$ with $\mathbf {X}_\cdot - \mathbf {X}_{i-1} \equiv \mathbf {x}_i$ on $[i-1,i)$ .

5.2 Diamond relations for multivariate cumulants

As in Section 2.3, we write $\mathcal {S}$ for the symmetric algebra over $\mathbb {R}^d$ , and $\mathcal {S}_0,\mathcal {S}_1$ for those elements with scalar component $0,1$ , respectively. Recall the exponential map $\exp : \mathcal {S}_0\to \mathcal {S}_1$ with global defined inverse $\log $ . Following Definition 2.1, the diamond product for $\mathcal {S}_0$ -valued semimartingales $\hat {\mathbf {X}}, \hat {\mathbf {Y}}$ is another $\mathcal {S}_0$ -valued semimartingale given by

$$ \begin{align*}(\hat{\mathbf{X}} \diamond \hat{\mathbf{Y}})_t(T) = \mathbb{E}_t \big( \langle \hat{\mathbf{X}}^c, \hat{\mathbf{Y}}^c \rangle_{t,T} \big) = \sum ( \mathbb{E}_t \langle \hat{\mathbf{X}}^{w_1}, \hat{\mathbf{Y}}^{w_2} \rangle_{t,T}) \hat{e}_{w_1} \hat{e}_{w_2}, \end{align*} $$

with summation over all $w_1,w_2 \in \widehat {\mathcal {W}}_d$ , provided all brackets are integrable. This trivially adapts to $\mathcal {S}^N$ -valued semimartingales, $N\in \mathbb {N}_{\ge 1}$ , in which case all words have length less equal N; the summation is restricted accordingly to $|w_1|+|w_2| \le N$ .

Theorem 5.3. (i) Let $\Xi = (0, \Xi ^{(1)},\Xi ^{(2)},\ldots )$ be an $\mathcal {F}_T$ -measurable random variable with values in $\mathcal {S}_0 (\mathbb {R}^d)$ , componentwise in $\mathcal {L}^{\infty -}$ . Then

$$\begin{align*}\mathbb{K}_t(T) := \log \mathbb{E}_t \exp (\Xi) \end{align*}$$

satisfies the following functional equation, for all $0 \le t \le T$ ,

(5.3) $$ \begin{align} \mathbb{K}_t(T) = \mathbb{E}_t \Xi + \frac{1}{2} (\mathbb{K} \diamond \mathbb{K})_t(T) + \mathbb{J}_t(T) \end{align} $$

with jump component,

$$ \begin{align*} \mathbb{J}_t(T) = \mathbb{E}_t \left( \sum_{t < u \le T} \left( e^{\Delta \mathbb{K}_u} - 1 - \Delta \mathbb{K}_u \right)\right) =\mathbb{E}_t\left( \sum_{t < u \le T} \left( \frac{1}{2!}(\Delta \mathbb{K}_u)^2 + \frac{1}{3!} (\Delta\mathbb{K}_u)^3 + \dotsb \right)\right). \end{align*} $$

Furthermore, if $N\in \mathbb {N}_{\ge 1}$ , and $\Xi =(\Xi ^{(1)},\ldots ,\Xi ^{(N)})$ is $\mathcal {F}_T$ -measurable with graded integrability condition

(5.4) $$ \begin{align} \left\Vert {\Xi^{(n)}} \right\Vert _{\mathcal{L}^{N/n}} < \infty, \qquad n=1,\ldots,N, \end{align} $$

then the identity equation (5.3) holds for the cumulants up to level N: that is, for $\mathbb {K}^{(0,N)} := \log (\mathbb {E}_t\exp (\Xi ^{(0,N)}))$ with values in $\mathcal {S}^{(N)}_0 (\mathbb {R}^d)$ .

Remark 5.4. Identity equation (5.3) is reminiscent to the quadratic form of the generalized Riccati equations for affine jump diffusions. The relation will be presented more explicitly in Remark 6.9 of Section 6.2.2, when the involved processes are assumed to have a Markov structure and the functional signature cumulant equation reduces to a PIDE system. The framework described here, however, requires neither Markov nor affine structure. We will show in Section 6.3 that such computations are also possible in the fully non-commutative setting: that is, to obtain signature cumulants of affine Volterra processes.

Proof. We first observe that since $\Xi \in \mathcal L^{\infty -}$ , by Doob’s maximal inequality and the BDG inequality, we have that $\hat {\mathbf {X}}_t:=\mathbb {E}_t\Xi $ is a martingale in $\mathscr {H}^{\infty -}(\mathcal {S}_0)$ . In particular, thanks to Theorem 3.2, the signature moments are well defined. According to Section 3.2, the signature is then given by

$$\begin{align*}\mathrm{Sig}(\hat{\mathbf{X}})_{t,T}=\exp(\Xi-\mathbb{E}_t\Xi), \end{align*}$$

hence $\hat {\boldsymbol {\kappa }}_t(T)=\mathbb {K}_t(T)-\hat {\mathbf {X}}_t$ .

Projecting equation (4.3) onto the symmetric algebra yields

$$ \begin{align*} \hat{\boldsymbol{\kappa}}_t(T)&=\mathbb{E}_t\Bigg\{\hat{\mathbf{X}}_{t,T}+\frac12\langle\hat{\mathbf{X}}^c\rangle_{t,T}+\frac12\langle\boldsymbol{\kappa}(T)^c\rangle_{t,T}+\langle\hat{\mathbf{X}}^c,\boldsymbol{\kappa}(T)^c\rangle_{t,T}\Bigg.\\ \Bigg.&\quad+\sum_{t<u\le T}\left( e^{\Delta\hat{\mathbf{X}}_u+\Delta\boldsymbol{\kappa}_u(T)}-1-\Delta\hat{\mathbf{X}}_u-\Delta\boldsymbol{\kappa}_u(T) \right)\Bigg\} \\&= \mathbb{E}_t\left\{\Xi+ \frac12\langle\mathbb{K}(T)^c\rangle_{t,T} +\sum_{t<u\le T}\left( e^{\Delta\mathbb{K}_u(T)}-1-\Delta\mathbb{K}_u(T) \right) \right\}-\hat{\mathbf{X}}_t, \end{align*} $$

and equation (5.3) follows upon recalling that $(\mathbb {K}\diamond \mathbb {K})_t(T)=\mathbb {E}_t\langle \mathbb {K}(T)^c\rangle _{t,T}$ . The proof of the truncated version is left to the reader.

As a corollary, we provide a general view on recent results of [Reference Alos, Gatheral and Radoičić3, Reference Lacoin, Rhodes and Vargas42, Reference Friz, Gatheral and Radoičić23]. Note that we also include jump terms in our recursion.

Corollary 5.5. The conditional multivariate cumulants $(\mathbb {K}_t)_{0\le t\le T}$ of a random variable $\Xi $ with values in $\mathcal {S}_0(\mathbb {R}^d)$ , componentwise in $\mathcal L^{\infty -}$ satisfy the recursion

(5.5) $$ \begin{align} \mathbb{K}^{(1)}_t = \mathbb{E}_t(\Xi^{(1)}) \quad \text{and} \quad \mathbb{K}^{(n)}_t = \mathbb{E}_t(\Xi^{(n)})+\frac{1}{2}\sum_{k=1}^{n}\left( \mathbb{K}^{(k)} \diamond \mathbb{K}^{(n-k)}\right)_t(T)+\mathbb J^{(n)}_t(T) \quad \text{ for } \quad n \ge 2, \end{align} $$

with

$$\begin{align*}\mathbb J^{(n)}_t(T)=\mathbb{E}_t\left( \sum_{t<u\le T}\sum_{k=2}^n\frac{1}{k!}\sum_{\|\ell\|=n,|\ell|=k}\Delta\mathbb{K}^{(\ell_1)}_u(T)\dotsm\Delta\mathbb{K}_u^{(\ell_k)}(T) \right). \end{align*}$$

The analogous statement holds true in the N-truncated setting: that is, as a recursion for $n=1,..,N$ under the condition in equation (5.4).

Example 5.6 (Continuous setting)

In case of an absence of jumps and higher-order information (i.e., $\mathbb {J} \equiv 0,\Xi ^{(2)} = \Xi ^{(3)} = \ldots \equiv 0$ ), this type of cumulant recursion appears in [Reference Lacoin, Rhodes and Vargas42] and under optimal integrability conditions $\Xi ^{(1)}$ with finite N.th moments [Reference Friz, Gatheral and Radoičić23]. (This requires a localization argument that is avoided here by directly working in the correct algebraic structure.)

Example 5.7 (Discrete filtration)

As the opposite of the previous continuous example, we consider a purely discrete situation, starting from a discretely filtered probability space with filtration $(\mathcal {F}_t\colon t = 0,1,\dotsc ,T \in {\mathbb {N}})$ . For $\Xi $ as in Corollary 5.5, a discrete martingale is defined by $\mathbb {E}_t \exp (\Xi )$ , which we may regard as a càdlàg semimartingale with respect to $\mathcal {F}_t := \mathcal {F}_{[t]}$ , and similar for $\mathbb {K}_t(T) = {\log \mathbb {E}_t \exp (\Xi ) \in \mathcal {S}_0}$ : that is, the conditional cumulants of $\Xi $ . Clearly, the continuous martingale part of $\mathbb {K}(T)$ vanishes, as does any diamond product with $\mathbb {K}(T)$ . What remains is the functional equation

$$ \begin{align*}\mathbb{K}_t(T) = \mathbb{E}_t (\Xi) + \mathbb{J}_t(T) = \mathbb{E}_t (\Xi) + \mathbb{E}_t \bigg( \sum_{u=t+1}^T \big( \exp(\Delta \mathbb{K}_u) - 1 - \Delta \mathbb{K}_u \big)\bigg). \end{align*} $$

As before, the resulting expansions are of interest. On the first level, trivially, $\mathbb {K}^{(1)}_t = \mathbb {E}_t(\Xi ^{(1)})$ , whereas on the second level we see

$$ \begin{align*}\mathbb{K}_t^{(2)}(T) = \mathbb{E}_t(\Xi^{(2)})+ \mathbb{E}_t \bigg( \sum_{u=t+1}^T (\mathbb{E}_u (\Xi^{(1)})-\mathbb{E}_{u-1} (\Xi^{(1)}))^2 \bigg), \end{align*} $$

which one can recognize, in case $\Xi ^{(2)} = 0$ as an energy identity for the discrete square-integrable martingale $\ell _u := \mathbb {E}_u \Xi ^{(1)}$ . Going further in the recursion yields increasingly non-obvious relations. Taking $\Xi ^{(2)} = \Xi ^{(3)} = \ldots \equiv 0$ for notational simplicity gives

$$ \begin{align*}\mathbb{K}_t^{(3)}(T) = \mathbb{E}_t \left( \sum_{u = t + 1}^T (\ell_u - \ell_{u - 1})^3 + 3 (\ell_u - \ell_{u - 1}) \{ \mathbb{E}_u \kappa (\ell, \ell)_{u, T} -\mathbb{E}_{u - 1} \kappa (\ell, \ell)_{u - 1, T} \} \right). \end{align*} $$

It is interesting to note that related identities have appeared in the statistics literature under the name Bartlett identities; see also Mykland [Reference Mykland53] and the references therein.

5.3 Remark on tree representation

As illustrated in the previous section, in the case where $d=1$ , or when projecting onto the symmetric algebra, our functional equation takes a particularly simple form (see Theorem 5.3). If one further specializes the situation, in particular discards all jumps, we are from an algebraic perspective in the setting of Friz, Gatheral and Radoiçić [Reference Friz, Gatheral and Radoičić23], which give a tree series expansion of cumulants using binary trees. This representation follows from the fact that the diamond product of semimartingales is commutative but not associative. As an example (with notations taken from Section 5.2), in case of a one-dimensional continuous martingale, the first terms are

This expansion is organized (graded) in terms of the number of leaves in each tree, and each leaf represents the underlying martingale.

In the deterministic case, tree expansions are also known for the Magnus expansion [Reference Iserles and Nørsett33] and the BCH formula [Reference Casas and Murua9]. These expansions are also, in terms of binary trees, different from the ones above as they are required to be non-planar to account for the non-commutativity of the Lie algebra. As an example (with the notations of Section 5.1), we have

In this expansion, the nodes represent the underlying vector field and edges represent integration and application of the Lie bracket, coming from the $\operatorname {\mathrm {ad}}$ operator.

Since our functional equation and the associated recursion puts both contexts into a single common framework. We suspect that our general recursion, Corollary 4.2 and thereafter, allows for a sophisticated tree representation, at least in absence of jumps, and propose to return to this question in future work.

6 Applications

6.1 Brownian and stopped Brownian signature cumulants

6.1.1 Time dependent Brownian motion

Let B be a m-dimensional standard Brownian motion defined on a portability space $(\Omega , \mathcal {F}, \mathbb {P})$ with the canonical filtration $(\mathcal {F}_t)_{t\ge 0}$ , and define the continuous (Gaussian) martingale $X = (X_t)_{0\le t\le T}$ by

$$ \begin{align*} X_t = \int_0^{t} \sigma(u)\,\mathrm{d} B_u, \quad 0 \le t \le T, \end{align*} $$

with $\sigma \in L^2 ([0, T],\mathbb {R}^{m\times d})$ . The quadratic variation of X is finite and deterministic, and therefore we immediately see that the integrability condition $\mathbf {X} = (0, X, 0, \dots )\in \mathscr {H}^{\infty -}$ is trivially satisfied, and thus Theorem 4.1 applies. The Brownian signature cumulants $\boldsymbol {\kappa }_t(T) = \log (\mathbb {E}_t(\mathrm {Sig}(\mathbf {X})_{t,T}))$ satisfies the functional equation, with $\mathbf {a}(t) := \sigma (t)\sigma (t)^T \in \mathrm {Sym}({\mathbb {R}^d} \otimes {\mathbb {R}^d}),$

(6.1) $$ \begin{align} \boldsymbol{\kappa}_t(T) = \frac{1}{2}\int_t^{T} H(\operatorname{\mathrm{ad}}{\boldsymbol{\kappa}_u(T)})(\mathbf{a}(u)) \mathrm{d} u, \quad 0 \le t \le T. \end{align} $$

Therefore the tensor levels are precisely given by the Magnus expansion, starting with

$$ \begin{align*}\boldsymbol{\kappa}^{(1)}_t(T) = 0,\quad \boldsymbol{\kappa}^{(2)}_t(T) = \frac{1}{2}\int_t^T \mathbf{a}(u) \mathrm{d} u,\end{align*} $$

and the general term

$$ \begin{align*} \boldsymbol{\kappa}^{(2n-1)}_t(T) \equiv 0, \quad \boldsymbol{\kappa}^{(2n)}_t(T) &= \frac{1}{2} \sum_{|\ell|\ge2, \Vert\ell\Vert=2n} \mathrm{HMag}^{2}(\mathbf{X}, \boldsymbol{\kappa}; \ell)_{t,T} \\ &= \frac{1}{2}\sum_{\Vert\ell\Vert=n-1} \frac{B_{k}}{k!} \int_t^{T} \operatorname{\mathrm{ad}}{\boldsymbol{\kappa}^{(2\cdot l_1)}_{u}} \cdots \operatorname{\mathrm{ad}}{\boldsymbol{\kappa}^{(2\cdot l_{k})}_{u}} \left(\mathbf{a}(u)\right)\mathrm{d} u. \end{align*} $$

Note that $\boldsymbol {\kappa }_t(T)$ is Lie in $\mathrm {Sym}({\mathbb {R}^d} \otimes {\mathbb {R}^d}) \subset \mathcal {T}_0$ , but, in general, not a Lie series. In the special case $X=B$ , that is, $m=d$ and identity matrix $\sigma = \mathbf {I}_d= \sum _{i=1}^{d}e_{ii}\in \mathrm {Sym}({\mathbb {R}^d}\otimes {\mathbb {R}^d})$ , all commutators vanish, and we obtain what is known as Fawcett’s formula [Reference Fawcett21, Reference Friz and Hairer24]:

$$ \begin{align*} \boldsymbol{\kappa}_t(T) = \tfrac{1}{2} (T-t)\mathbf{I}_d \,. \end{align*} $$

Example 6.1. Consider $B^1,B^2$ two Brownian motions on the filtered space $(\Omega ,\mathcal F,\mathbb P)$ , with correlation $\mathrm {d}\langle B^1,B^2\rangle _t=\rho \,\mathrm {d} t$ for some fixed constant $\rho \in [-1,1]$ . Suppose that $K^1,K^2\colon [0,\infty )^2\to \mathbb {R}$ are two kernels such that $K^i(t,\cdot )\in L^2([0,t])$ for all $t\in [0,T]$ , and set

$$\begin{align*}X^i_t:= X_0^i+\int_0^tK^i(t,s)\,\mathrm{d} B^i_s,\quad i=1,2 \end{align*}$$

for some fixed initial values $X^1_0,X^2_0$ . Note that neither process is a semimartingale in general. However, for each $T>0$ , the process $\xi ^i_t(T):=\mathbb {E}_t[X^i_T]$ is a martingale, and we have

$$\begin{align*}\xi^i_t(T)=X^i_0+\int_0^tK^i(T,s)\,\mathrm{d} B^i_s, \end{align*}$$

that is, $(\xi ^1,\xi ^2)$ is a time-dependent Brownian motion as defined above. In particular, one sees that

$$\begin{align*}\mathbf{a}(t)=\begin{pmatrix}\int_0^tK^1(T,u)^2\,\mathrm{d} u&\rho\int_0^tK^1(T,u)K^2(T,u)\,\mathrm{d} u\\\rho\int_0^tK^1(T,u)K^2(T,u)\,\mathrm{d} u&\int_0^tK^2(T,u)^2\,\mathrm{d} u\end{pmatrix}. \end{align*}$$

Equation (6.1) and the paragraph below it then give an explicit recursive formula for the signature cumulants, the first of which are given by

$$ \begin{align*} \boldsymbol{\kappa}_t^{(1)}(T)&= 0,\\ \boldsymbol{\kappa}_t^{(2)}(T)&= \frac12\begin{pmatrix}\int_t^T\int_0^uK^1(T,r)^2\,\mathrm{d} r\mathrm{d} u&\rho\int_t^T\int_0^uK^1(T,r)K^2(T,r)\,\mathrm{d} r\mathrm{d} u\\[1ex]\rho\int_t^T\int_0^uK^1(T,r)K^2(T,r)\,\mathrm{d} r\mathrm{d} u&\int_t^T\int_0^uK^2(T,r)^2\,\mathrm{d} r\mathrm{d} u\end{pmatrix},\\ \boldsymbol{\kappa}_t^{(3)}(T)&= 0,\\ \boldsymbol{\kappa}_t^{(4)}(T)&= \frac1{2}\sum_{i,j,i',j'=1}^2\left[\int_t^T\int_u^T\left( \mathbf{a}^{ij}(u)\mathbf{a}^{i'j'}(r)-\mathbf{a}^{i'j'}(u)\mathbf{a}^{ij}(r)\right)\,\mathrm{d} r\mathrm{d} u\right]e_{iji'j'}. \end{align*} $$

We notice that in the particular case when $K^1=K^2\equiv K$ , the matrix $\mathbf {a}$ has the form

$$\begin{align*}\mathbf{a}(t)=\int_0^tK(T,u)^2\,\mathrm{d} u\times\begin{pmatrix}1&\rho\\\rho&1\end{pmatrix}. \end{align*}$$

Therefore, we have $ \mathbf {a}(t)\otimes \mathbf {a}(t')- \mathbf {a}(t')\otimes \mathbf {a}(t)=0$ for any $t,t'\in [0,T]$ . Hence, in this case, our recursion shows that for any $\rho \in [-1,1]$ ,

$$\begin{align*}\boldsymbol{\kappa}_t^{(1)}(T)=0,\quad\boldsymbol{\kappa}_t^{(2)}(T)=\frac12\int_t^T\int_0^uK(T,r)^2\,\mathrm{d} r\,\mathrm{d} u\times\begin{pmatrix}1&\rho\\\rho&1\end{pmatrix}, \end{align*}$$

and $\boldsymbol {\kappa }_t^{(n)}(T)=0$ for all $0\le t\le T$ and $n\ge 3$ .

6.1.2 Brownian motion up to the first exit time from a domain

Let $B=(B_t)_{t\ge 0}$ be a d-dimensional Brownian motion defined on a probability space $(\Omega , \mathcal {F}, \mathbb {P})$ with a possibly random starting value $B_0$ . Assume also that there is a family of probability measures $\{\mathbb {P}^x\}_{x\in {\mathbb {R}^d}}$ on $(\Omega , \mathcal {F})$ such that $\mathbb {P}^x(B_0 = x) = 1$ , and denote by $\mathbb {E}^x$ the expectation with respect to $\mathbb {P}^x$ . We define the canonical Brownian filtrationFootnote 8 by $(\mathcal {F}_t)_{t\ge 0} = (\mathcal {F}^{B}_t)_{t\ge 0}$ . Further let $\Gamma \subset {\mathbb {R}^d}$ be a bounded domain, and define the stopping time $\tau _\Gamma $ of the first exit of B from the domain $\Gamma $ : that is,

$$ \begin{align*} {\tau_\Gamma} = \inf\{t\ge0 \;\vert\; B_t \in \Gamma^{c}\}. \end{align*} $$

In [Reference Lyons and Ni46], Lyons–Ni exhibit an infinite system of partial differential equations for the expected signature of the Brownian motion until the exit time as a functional of the starting point. The following result can be seen as the corresponding result for the signature cumulant, which follows directly from the expansion in Theorem 1.1. Recall that a boundary point $x \in \partial \Gamma $ is called regular if and only if

(6.2) $$ \begin{align} \mathbb{P}^x\big( \inf\{t> 0 \;\vert\; B_t \in \Gamma^{c}\} = 0\big) = 1. \end{align} $$

The domain $\Gamma $ is called regular if all points on the boundary are regular. For example, domains with smooth boundary are regular; and see [Reference Karatzas and Shreve38, Section 4.2.C] for a further characterization of regularity.

Corollary 6.2. Let $\Gamma \subset {\mathbb {R}^d}$ be a regular domain, such that

(6.3) $$ \begin{align} \sup_{x\in\Gamma}\mathbb{E}^x(\tau_\Gamma^n)<\infty, \quad n\in{\mathbb{N}_{\ge1}}. \end{align} $$

The signature cumulant $\boldsymbol {\kappa }_t = \log (\mathbb {E}(\mathrm {Sig}(B)_{t\wedge {\tau _\Gamma }, {\tau _\Gamma }}))$ of the Brownian motion B up to the first exit from the domain $\Gamma $ has the following form

$$ \begin{align*} \boldsymbol{\kappa}_t = \mathbf{1}_{\{t<\tau_\Gamma\}} \mathbf{F}(B_t), \quad t\ge0, \end{align*} $$

where $\mathbf {F} = \sum _{|w|\ge 2} e_w F^{w}$ with $F^{w}\in C^{0}(\overline {\Gamma },\mathbb {R})\cap C^{2}(\Gamma ,\mathbb {R})$ is the unique bounded classical solution to the elliptic PDE

(6.4) $$ \begin{align} -\Delta \mathbf{F} (x) &= \sum_{i=1}^{d}H(\operatorname{\mathrm{ad}}{\mathbf{F} (x)})\Big(e_{ii} + Q(\operatorname{\mathrm{ad}}{\mathbf{F} (x)})(\partial_i \mathbf{F} (x)^{\otimes 2}) + 2e_i G(\operatorname{\mathrm{ad}}{\mathbf{F} (x)})(\partial_i \mathbf{F} (x)) \Big), \end{align} $$

for all $x\in \Gamma $ with the boundary condition $\mathbf {F}\vert _{\partial \Gamma } \equiv 0$ .

Proof. Define the martingale $\mathbf {X} = ((0, B_{t\wedge {\tau _\Gamma }}, 0, \dotsc ) )_{t\ge 0}\in \mathscr {S}(\mathcal {T}_0)$ , and note that . It then follows from the integrability of ${\tau _\Gamma }$ that $\mathbf {X} \in \mathscr {H}^{\infty -}(\mathcal {T}_0)$ and thus by Theorem 3.2 that $(\mathrm {Sig}(\mathbf {X})_{0,t})_{t\ge 0} \in \mathscr {H}(\mathcal {T}_1)^{\infty -}$ . This implies that the signature cumulant $\boldsymbol {\kappa }_t(T):= \log (\mathbb {E}_t(\mathrm {Sig}(\mathbf {X})_{t,T}))$ is well defined for all $0 \le t \le T < \infty $ , and furthermore under (component-wise) application of the dominated convergence theorem that it holds

$$ \begin{align*} \boldsymbol{\kappa}_t = \lim_{T\to\infty} \boldsymbol{\kappa}_t(T) = \lim_{T\to\infty}\log(\mathbb{E}_t(\mathrm{Sig}(\mathbf{X})_{t,T})) = \log(\mathbb{E}_t (\mathrm{Sig}(B)_{t\wedge {\tau_\Gamma}, {\tau_\Gamma}})), \quad t\ge 0. \end{align*} $$

Again by $\mathbf {X} \in \mathscr {H}^{\infty -}(\mathcal {T}_0)$ , it follows that Theorem 1.1 applies to the martingale $(\mathbf {X}_t)_{0\le t \le T}$ for any $T>0$ and therefore $\boldsymbol {\kappa }(T)$ satisfies the functional equation (4.3). It follows from the Itô’s representation theorem [Reference Revuz and Yor59, Theorem 3.4] that all local martingales with respect to the Brownian filtration $(\mathcal {F}_t)_{0 \le t \le T}$ are continuous, and therefore it is easy to see that also $\boldsymbol {\kappa }(T)\in \mathscr {S}^c(\mathcal {T}_0)$ . Therefore equation (4.3) simplifies to the following equation

(6.5)

where we have already used the martingality of $\mathbf {X}$ and the explicit form of the quadratic variation $\left \langle \mathbf {X} \right \rangle _t = \mathbf {I}_d(t \wedge {\tau _\Gamma })$ with $\mathbf {I}_d = \sum _{i=1}^{d}e_{ii} \in ({\mathbb {R}^d})^{\otimes 2}$ . It follows that $\boldsymbol {\kappa }^{(1)} \equiv \boldsymbol {\kappa }(T)^{(1)} \equiv 0$ , and for the second level, we have from the integrability of ${\tau _\Gamma }$ and the strong Markov property of Brownian motion that

$$ \begin{align*} \boldsymbol{\kappa}_t^{(2)} =\frac{1}{2}\mathbf{I}_d \lim_{T\to\infty}\mathbb{E}_t\left( \mathbf{1}_{\{t<\tau_\Gamma\}}({\tau_\Gamma}\wedge T -t) \right)= \frac{1}{2}\mathbf{I}_d \mathbf{1}_{\{t<\tau_\Gamma\}}\left.\mathbb{E}^{x}({\tau_\Gamma})\right\vert_{x=B_t}, \quad t\ge0. \end{align*} $$

Now note that the function $u(x) := \mathbb {E}^x({\tau _\Gamma })$ for $x \in \Gamma $ is in $C^{0}(\overline {\Gamma }, \mathbb {R})\cap C^{2}(\Gamma , \mathbb {R})$ and solves the Poisson equation $ -(1/2)\Delta u = g$ with boundary condition $u\vert _{\partial \Gamma } = 0$ and data $g \equiv 1$ . Indeed, since $\Gamma $ is regular and g is bounded and differentiable, this follows from Theorem 9.3.3 (and the remark thereafter) in [Reference Øksendal55]. Moreover, from the assumption in equation (6.3), we immediately see that u is bounded on $\overline {\Gamma }$ , and it follows from Theorem 9.3.2 in [Reference Øksendal55] that u is the unique bounded classical such solution. Thus we have shown that the statement holds true up to the second tensor level with $\mathbf {F}^{(1)}\equiv 0$ and $\mathbf {F}^{(2)}(u) =\frac {1}{2} \mathbf {I}_d u$ under the usual notation $\mathbf {F}^{(n)} = \sum _{|w|=n}e_wF^{w}$ .

Now assume that the statement of the corollary holds true up to the tensor level $(N-1)$ for some $N \ge 3$ . Then, for any $n, k < N$ , we have by applying Itô’s formula

and

Further define the function $\mathbf {G}^{(N)}$ by the projection under $\pi _N$ of the right-hand side of equation (6.4) multiplied by the factor $1/2$ . Then applying Theorem 4.1 to $\mathbf {X}^{(0,N)}$ on the probability space $(\Omega , \mathcal {F}, (\mathcal {F}_t),\mathbb {P}^{x})$ , we see that it follows from the estimate in equation (7.30) that there exists a constant $c>0$ such that

$$ \begin{align*} \sup_{x\in\Gamma}\mathbb{E}^x \left\{ \int_0^{\tau_\Gamma} \big\vert \mathbf{G}^{(N)}(B_u) \big\vert\,\mathrm{d} u\right\} \le c \sup_{x\in\Gamma} \vert\mkern-2.4mu\vert\mkern-2.5mu\vert {\mathbf{X}^{(0,N)}}