Deep Signature Transforms

Patric Bonnier1,∗ Patrick Kidger1,2,∗ Imanol Perez Arribas1,2,∗

Cristopher Salvi1,2,∗ Terry Lyons1,2

Mathematical Institute, University of Oxford

1

The Alan Turing Institute, British Library

2

{bonnier, kidger, perez, salvi, tlyons}@[Link]

Abstract
The signature is an infinite graded sequence of statistics known to characterise
a stream of data up to a negligible equivalence class. It is a transform which
has previously been treated as a fixed feature transformation, on top of which a
model may be built. We propose a novel approach which combines the advantages
of the signature transform with modern deep learning frameworks. By learning
an augmentation of the stream prior to the signature transform, the terms of the
signature may be selected in a data-dependent way. More generally, we describe
how the signature transform may be used as a layer anywhere within a neural
network. In this context it may be interpreted as a pooling operation. We present
the results of empirical experiments to back up the theoretical justification. Code
available at [Link]/patrick-kidger/Deep-Signature-Transforms.

1 Introduction
1.1 What is the signature transform?

When data is ordered sequentially then it comes with a natural path-like structure: the data may be
thought of as a discretisation of a path X : [0, 1] → V , where V is some Banach space. In practice
we shall always take V = Rd for some d ∈ N. For example the changing air pressure at a particular
location may be thought of as a path in R; the motion of a pen on paper may be thought of as a path
in R2 ; the changes within financial markets may be thought of as a path in Rd , with d potentially
very large.
Given a path, we may define its signature, which is a collection of statistics of the path. The map
from a path to its signature is called the signature transform.
Definition 1.1. Let x = (x1 , . . . , xn ), where xi ∈ Rd . Let f = (f1 , . . . , fd ) : [0, 1] → Rd be
continuous, such that f ( n−1
i−1
) = xi , and linear on the intervals in between. Then the signature of x
is defined as the collection of iterated integrals2
  
Z Z Yk
 df ij 
Sig(x) =  ··· (tj )dt1 · · · dtk   .
j=1
dt
0<t1 <···<tk <1 1≤i1 ,...,ik ≤d k≥0
∗
Equal contribution.
2
For clarity here we have used more widely-understood notation. The definition of the signature transform
is usually written in an equivalent but alternate manner using the notation of stochastic calculus; see Definition
A.1 in Appendix A.

33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.
We shall often use the term signature to refer to both a path’s signature and the signature transform.
Other texts sometimes use the term path signature in a similar manner.
We refer the reader to [1] for a primer on the use of the signature in machine learning. A brief
overview of its key properties may be found in Appendix A, along with associated references.
In short, the signature of a path determines the path essentially uniquely, and does so in an efficient,
computable way. Furthermore, the signature is rich enough that every continuous function of the
path may be approximated arbitrarily well by a linear function of its signature; it may be thought of
as a ‘universal nonlinearity’. Taken together these properties make the signature an attractive tool
for machine learning. The most simple way to use the signature is as feature transformation, as it
may often be simpler to learn a function of the signature than of the original path.
Originally introduced and studied by Chen in [2, 3, 4], the signature has seen use in finance [5, 6, 7,
8, 9], rough path theory [10, 11] and machine learning [12, 13, 14, 15, 16, 17, 18, 19, 20].

1.2 Comparison to the Fourier transform

The signature transform is most closely analogous to the Fourier transform.

The fundamental difference between the signature transform and classical signal transforms such as
Fourier transforms and wavelets is that the latter are used to model a curve as a linear combination
in a functional basis. The signature does not try to model or parameterise the curve itself, but instead
provides a basis for functions on the space of curves.
For example, regularly seeing the sequence: phone call, trade, price movement in the stream
of office data monitoring a trader might be an indication of insider trading. Such occurrences
are straightforward to detect by via a linear regression composed with the signature transform.
Modelling this signal using Fourier series or wavelets would be much more expensive: linearity
of these transforms imply that each channel must be resolved accurately enough to see the order of
events.
From a signal processing perspective, the signature can be thought of as a filter which is invariant to
resampling of the input signal. (See Proposition A.7 in Appendix A).

1.3 Use of the signature transform in machine learning

The signature is an infinite sequence, so in practice some finite collection of terms must be selected.
Since the magnitude of the terms exhibit factorial decay, see Proposition A.5 in Appendix A, it is
usual [21] to simply choose the first N terms of this sequence, which will typically be the largest
terms . These first N terms are called the signature of depth N or the truncated signature of depth
N , and the corresponding transform is denoted SigN . But if the function to be learned depended
nontrivially on the higher degree terms, then crucial information has nonetheless been lost.
This may be remedied. Apply a pointwise augmentation to the original stream of data before taking
the signature. Then the first N terms of the signature may better encode the necessary information
[19, 20]. Explicitly, let Φ : Rd → Re be fixed; one could ensure that information is not lost by taking
Φ(x) = (x, ϕ(x)) for some ϕ. Then rather than taking the signature of x = (x1 , . . . , xn ), where
xi ∈ Rd , instead take the signature of Φ(x) = (Φ(x1 ), . . . , Φ(xn )). In this way one may capture
higher order information from the stream in the lower degree terms of the signature.

1.4 Our work

But how should this augmentation Φ be chosen? Previous work has fixed it arbitrarily, or
experimented with several options before choosing one [19, 20]. Observe that in each case the
map x 7→ SigN (Φ(x)) is still ultimately just a feature transformation on top of which a model is
built. Our more general approach is to allow the selection of Φ to be data-dependent, by having
it be learned; in particular it may be a neural network. Furthermore there is no reason it should
necessarily operate pointwise, nor (since it is now learned) need it be of the form (x, ϕ(x)). In this
way we may enjoy the benefits of using signatures while avoiding their main limitation.
But this means that the signature transform is essentially operating as a layer within a neural network.
It consumes a tensor of shape (b, d, n) – corresponding to a batch of size b of paths in Rd that have

2
been sampled n times – and returns a tensor of shape (b, (dN +1 −1)/(d−1)), where N is the number
of terms used in the truncated signature.3 The signature is being used as a pooling operation.
There is no reason to stop here. If the signature layer works well once then it is natural to seek to
use it again. The obvious problem is that the signature transform consumes a stream of data and
returns statistics which have no obvious stream-like qualities. The solution is to lift the input stream
to a stream of streams; for example, the stream of data (x1 , . . . , xn ) may be lifted to the ‘expanding
windows’ of (x2 , . . . , xn ), where xi = (x1 , . . . , xi ). Now apply the signature to each stream to
obtain a stream of signatures (SigN (x2 ), . . . , SigN (xn )), which is essentially a stream in Euclidean
space. And now this new stream may be augmented via a neural network and the process repeated
again, as many times as we wish.
In this way the signature transform has been elevated from a one-time feature transformation to
a first-class layer within a neural network. Thus we may reap the benefits of both the signature
transform, with its strong corpus of mathematical theory, and the benefits of neural networks, with
their great empirical success.
Naturally all of this implies the need for an efficient implementation of the signature transform. Such
concerns have motivated the creation of the spin-off Signatory project [22].
The remainder of the paper is laid out as follows. In Section 2 we briefly discuss some related work,
in Section 3 we detail the specifics of embedding the signature as a layer within a neural network.
Sections 4 covers experiments; we demonstrate positive results for generative, supervised, and
reinforcement learning problems. Section 5 is the conclusion. Appendix A provides an exposition
of the theoretical properties of the signature, and Appendix B specifies implementation details.

2 Related Work
The signature transform is roughly analogous to the use of wavelets or Fourier transforms, and there
are also related models based around these, for example [23, 24, 25, 26]. We do not know of a
detailed comparison between the use of these various transformations in the context of machine
learning.
Some related work using signatures has already been discussed in the previous section. We expand
on their proposed models here.
Definition 2.1. Given a set V , the space of streams of data in V is defined as
S(V ) = {x = (x1 , . . . , xn ) : xi ∈ V, n ∈ N}.
Given x = (x1 , . . . , xn ) ∈ S(V ), the integer n is called the length of x.
Two simple models utilising the signature layer are shown in Figure 1.
In principle the universal nonlinearity property of signatures (see Proposition A.6 in Appendix A)
guarantees that the model shown in Figure 1a, is rich enough to learn any continuous function. (With
the neural network taken to be a single linear layer and the input stream assumed to already be time-
augmented.) In practice, of course, the signature must be truncated. Furthermore, it is not clear
how to appropriately choose the truncation hyperparameter N . Thus a more practical approach is to
remove the restriction that the neural network must be linear, and learn a nonlinear function instead.
This approach has been applied successfully in various tasks [5, 12, 13, 14, 15, 16, 17, 18].
3 PN
As (dN +1 − 1)/(d − 1) = k=0 dk is the number of scalar values in a signature with N terms.

Input Signature Neural Input Feature Signature Neural

Output Output
stream transform network stream map transform network

x SigN fθ σ x Φ SigN fθ σ

(a) Neural-signature model. (b) Neural-signature-augment model.

Trainable parameters: θ. Trainable parameters: θ.
Figure 1: Two simple architectures with a signature layer.

3
An alternate model is shown in Figure 1b. Following [19, 20], a pointwise transformation
could be applied to the stream before taking the signature transform. That is, applying the
feature map Φ : Rd → Re to the d-dimensional stream of data (x1 , . . . , xn ) ∈ S(Rd ) yields
(Φ(x1 ), . . . , Φ(xn )) ∈ S(Re ); the signature of Φ(x) may then potentially capture properties of
the stream of data that will yield more effective models.

3 The signature transform as a layer in a neural network

However, there is not always a clear candidate for the feature map Φ and a good choice is likely to
be data-dependent. Thus we propose to make Φ learnable by taking Φ = Φθ to be a neural network
with trainable parameters θ. In this case, we again obtain the neural network shown in Figure 1b,
except that Φ is now also learnable.
The signature has now become a layer within a neural network. It consumes a tensor of shape
(b, d, n) – corresponding to a batch of size b of paths in Rd that have been sampled n times – and
returns a tensor of shape (b, (dN +1 − 1)/(d − 1)), where N is the number of terms used in the
truncated signature.
Despite being formed of integrals, the signature is in fact straightforward and efficient to compute
exactly, see Section A.3 in Appendix A. More than that, the computation may in fact be described in
terms of standard tensor operations. As such it may be backpropagated through without difficulty.

3.1 Stream-preserving neural networks

Let x = (x1 , . . . , xn ) ∈ S(Rd ). Whatever the choice of Φθ , it must preserve the stream-like nature
of the data if we are to take a signature afterwards. The simplest way of doing this is to have Φθ
map Rd → Re , so that it operates pointwise. This defines Φ(x) by
Φ(x) = (Φθ (x1 ), . . . , Φθ (xn )) ∈ S(Re ). (1)
Another way to preserve the stream-like nature is to sweep a one dimensional convolution along the
stream; more generally one could sweep a whole feedforward network along the stream. For some
m ∈ N and Φθ : Rd×m → Re this defines Φ(x) by
Φ(x) = (Φθ (x1 , . . . , xm ), . . . , Φθ (xn−m+1 , . . . , xn )) ∈ S(Re ). (2)
More generally still the network could be recurrent, by having memory. Let Φ0 = 0, fix m ∈ N,
and define Φk = Φθ (xk , . . . , xk+m ; Φk−1 ) for k = 1, . . . , n − m + 1. Then define Φ(x) by
Φ(x) = (Φ1 , . . . , Φn−m+1 ) ∈ S(Re ). (3)

3.2 Stream-like data

It is worth taking a moment to think what is really meant by ‘stream-like nature’. The signature
transform is defined on paths; it is applied to a stream of data in S(Rd ) by first interpolating the data
into a path and then taking the signature.
The data is treated as a discretisation or set of observations of some underlying path. Note that there
is nothing wrong with the path itself having a discrete structure to it; for example a sentence.
In principle one could reshape a tensor of shape (b, nd) with no stream-like nature into one of shape
(b, d, n), and then take the signature. However it is not clear what this means mathematically. There
is no underlying path. The signature is at this point an essentially arbitrary transformation, without
the mathematical guarantees normally associated with it.

3.3 Stream-preserving signatures, using lifts

We would like to apply the signature layer multiple times. However applying the signature transform
consumes the stream-like nature of the data, which prevents this. The solution is to construct a
stream of signatures in the following way: given a stream x = (x1 , . . . , xn ) ∈ S(Rd ), let xk =
(x1 , . . . , xk ) for k = 2, . . . , n, and apply the signature to each xk to obtain the stream
N +1
(SigN (x2 ), . . . , SigN (xn )) ∈ S(R(d −1)/(d−1)
). (4)

4
The shortest stream it is meaningful to take the signature of is of length two, which is why there is
no corresponding SigN (x1 ) term.
In this way the stream-like nature of the data is preserved through the signature transform.
This notion may be generalised: let
` = (`1 , `2 , . . . , `v ) : S(Rd ) → S(S(Re )),
which we refer to as a lift into the space of streams of streams (and v will likely depend on the length
of the input to `). Then apply the signature stream-wise to define SigN (`(x)) by

N +1
SigN (`(x)) = SigN (`1 (x)), . . . , SigN (`v (x)) ∈ S(R(e −1)/(e−1) ). (5)

In the example of equation (4), ` is given by

`(x) = (x2 , . . . , xn ). (6)

Other plausible choices for ` are to cut up x into multiple pieces, for example
`(x) = ((x1 , x2 ), (x3 , x4 ), . . . , (x2bn/2c−1 , x2bn/2c )), (7)
or to take a sliding window
`(x) = ((x1 , x2 , x3 ), (x2 , x3 , x4 ), . . . , (xn−2 , xn−1 , xn )). (8)

3.4 Multiple signature layers

By inserting lifts, the signature transform may be composed as many times as desired. That is,
suppose we wish to learn a map from S(Rd ) to X , where X is some set. (Which may be finite for
a classification problem or infinite for a regression problem.) Let ci , di , ei , Ni ∈ N be such that
d1 = d and di+1 = (cN i
i +1
− 1)/(ci − 1), for i = 1, . . . , k.
Let
Nk +1
Φθi i : Rdi ×mi → Rei , `i : S(Rei ) → S(S(Rci )), f θk+1 : S(R(ck −1)/(ck −1)
) → X,
where Φθi i and `i are defined in the manner of equations (1)–(3) and (6)–(8), and θ1 , . . . , θk+1 are
some trainable parameters. Then defining compositions in the manner of equations (1)–(5), let
 
σ = f θk+1 ◦ SigNk ◦ `k ◦ Φθkk ◦ · · · ◦ Φθ22 ◦ SigN1 ◦ `1 ◦ Φθ11 (x).

This defines the deep signature model, summarised in Figure 2.

An important special case is when V = S(Re ), so that the final network f θk+1 is stream-preserving.
Then the overall model x 7→ σ is also stream-preserving. See for example Section 4.1.
Note that in principle it is acceptable to take the trivial lift to a sequence of a single element,
`(x) = (x). (9)
Taking the signature of this will then essentially remove the stream-like nature, however, so it is
suitable only for the final lift of a deep signature model. We observe in particular that this is what is
done in the models described in Figure 1, which we identify as special cases of the deep signature
model, lacking also any learned transformation before the signature.
It is easy to see that the deep signature model exhibits the universal approximation property. This fact
follows from the universal approximation theorem for neural networks [27] and from the universal
nonlinearity property of signatures (see Proposition A.6 in Appendix A).

Input Neural Signature Neural Signature Neural Output

Lift Lift
stream network transform network transform network

x Φθ11 `1 SigN1 Φθ22 ... `k SigNk f θk+1 σ

Figure 2: Deep signature model. Trainable parameters: θ1 , . . . , θk+1 .

5
3.5 Implementation

When using the signature transform as a feature transformation, then it suffices to just pre-process
and save the entire dataset before training. However when the signature transform is placed within
a neural network then the signature transform must be evaluated and backpropagated through for
each step of training; this is much more computationally intensive. This has motivated the creation
of the separate spin-off Signatory project [22], to efficiently perform and backpropagate through the
signature transform.

3.6 Inverting the truncated signature

How well does a truncated signature encode the original stream of data? A simple experiment is to
attempt to recover the original stream of data given its truncated signature. We remark that finding
a mathematical description of this inversion is a challenging task [28, 29, 30].
Fix a stream of data x = (x1 , . . . , xn ) ∈ S(Rd ). Assume that the truncated signature SigN (x) and
the number of steps n ∈ N are known. Now apply gradient descent to minimise
2
L(y; x) = SigN (y) − SigN (x) 2
for y = (y1 , . . . , yn ) ∈ S(Rd ).

Figure 3 shows four handwritten digits from the PenDigits dataset [31]. The solid blue path is
the original path x, whilst the dashed orange path is the reconstructed path y minimising L(y; x).
Truncated signatures of order N = 12 were used for this task. We see that the truncated signatures
have managed to encode the input paths x almost perfectly.

Figure 3: Original path (blue) and path reconstructed from its signature (dashed orange) for four
handwritten digits in the PenDigits dataset [31].

4 Numerical experiments
4.1 A generative model for a stochastic process

Generative models are typically trained to learn to transform random noise to a target distribution.
One common approach are Generative Adversarial Networks [32]. An alternative approach is to
define a distance on the space of distributions by embedding them into a Reproducing Kernel
Hilbert Space. The discriminator is then a fixed two-sample test based on a kernel maximum mean
discrepancy. This is known as a Generative Moment Matching Network [33, 34, 35].
With this framework we propose a deep signature model to generate sequential data. The
discriminator is as in [19, 20]. The natural choice for random noise is Brownian motion Bt .
Define the kernel k : S(Rd ) × S(Rd ) → R by


k(x, y) = SigN (λx x) , SigN (λy y) ,
where λx ∈ R is a certain normalising constant which guarantees that k is the kernel of a
Reproducing Kernel Hilbert Space, and ( · , · ) denotes the dot product.
Given n samples {x(i) }ni=1 ⊆ S(Rd ) from the generator and m samples {y(i) }m i=1 ⊆ S(R ) from
d

the target distribution, define the loss T by

  1 X 2 X 1 X
T {x(i) }ni=1 , {y(i) }m
i=1 = 2 k(x(i) , x(j) ) − k(x(i) , y(j) ) + 2 k(y(i) , y(j) ).
n i,j nm i,j m i,j

Let the input to the network be time-augmented Brownian motion

B = ((t1 , Bt1 ), . . . , (tn , Btn )) ∈ S(R2 ).

6
 ps
ste
s of
ep
St
B Φθ1 ` SigN f θ2 x SigM
Steps

Steps

Steps

Steps

Steps

Steps

Steps
Channels Channels Channels Signature Channels Channels Signature Two-sample
terms terms test

Generator Discriminator y SigM

Steps

Steps
Channels Signature
terms

Figure 4: Generative model architecture. Trainable parameters: θ1 , θ2 . There is an implicit batch

dimension throughout.

Given two stream-preserving neural networks Φθ1 and f θ2 , and a lift `, then the generative model is
defined by
x = (f θ2 ◦ SigN ◦ ` ◦ Φθ1 )(B).
The overall model is shown in Figure 4. In a nice twist, both the generator and the discriminator
involve the signature.
Observe how the generative part is a particular
case of the deep signature model, and that
furthermore the whole generator-discriminator
pair is also a particular case of the deep
signature model, with the trivial lift of equation
(9) before the second signature layer.
We applied the proposed model to a dataset
of 1024 realisations of an Ornstein–Uhlenbeck
process [36]. The loss was minimised at 6.6 ×
10−4 , which implies that the generated paths
are statistically almost indistinguishable from
the real Ornstein–Uhlenbeck process. Figure
5 shows the generated paths alongside the
original ones. Further implementation details Figure 5: Generated paths alongside the original
are in Appendix B. paths.

4.2 Supervised learning with fractional Brownian motion

Fractional Brownian motion [37] is a Gaussian process B H : [0, ∞) → R that generalises Brownian
motion. It is self-similar and exhibits fractal-like behaviour. Fractional Brownian motion depends
upon a parameter H ∈ (0, 1), known as the Hurst parameter. Lower Hurst parameters result in
noticeably rougher paths. The case of H = 1/2 corresponds to usual Brownian motion. Fractional
Brownian motion has been successfully used to model phenomena in diverse fields. For example,
empirical evidence from financial markets [38] suggests that log-volatility is well modelled by
fractional Brownian motion with Hurst parameter H ≈ 0.1.
Estimating the Hurst parameter of a fractional Brownian motion path is considered a nontrivial
task because of the paths’ non-stationarity and long range dependencies [39]. We train a variety of
models to perform this estimation. That is, to learn the map xH 7→ H, where
xH = ((t0 , BtH0 ), . . . , (tn , BtHn )) ∈ S(R2 )
for some realisation of B H .

7
Table 1: Final test mean squared error (MSE) for the different models, averaged over 3 training runs,
ordered from largest to smallest.
Test MSE
Mean Variance # Params
Rescaled Range 7.2 × 10−2 3.7 × 10−3 N/A
LSTM 4.3 × 10−2 8.0 × 10−3 12961
Feedforward 2.8 × 10−2 3.0 × 10−3 10209
Neural-Sig 1.1 × 10−2 8.2 × 10−4 10097
GRU 3.3 × 10−3 1.3 × 10−3 9729
RNN 1.7 × 10−3 4.9 × 10−4 10091
DeepSigNet 2.1 × 10−4 8.7 × 10−5 9261
DeeperSigNet 1.6 × 10−4 2.1 × 10−5 9686

The results are shown in Figure 6 and Table

1. Also shown in Table 1 are the results of
the rescaled range method [40], which is a
mathematically derived method rather than a
learned method.
RNN, GRU and LSTM models provide
baselines in the context of recurrent neural
networks. The simple Neural-Sig model
outlined previously in Figure 1a provides a
baseline from the context of signatures.
DeepSigNet and DeeperSigNet are both deep
signature models of the form given by Figure
2. DeepSigNet has a single large Neural-Lift-
Signature block, whilst DeeperSigNet has three
smaller ones.
We observe that traditional signature based Figure 6: Performance at estimating the
models perform slightly worse than traditional Hurst parameter for various models, with and
recurrent models, but that deep signature without signatures, for a particular (typical)
models outperform all other models by at least training run.
an order of magnitude. Further implementation
details are found in Appendix B.

4.3 Non-Markovian deep reinforcement learning

Finally we show how these ideas may be extended, by demonstrating a model that adds a residual
connection to the deep signature model; it may also be interpreted as using signatures as the memory
of a recurrent neural network.
As an example, we apply this architecture to tackle a non-Markovian reinforcement learning
problem. This means that the optimal action depends not just on the current state of the environment,
but upon the history of past states, so that the agent must maintain a memory.
N +1
Let Φθ1 : Rd → Re and f θ2 : Rd+(e −1)/(e−1) → {actions} be functions depending on learnable
parameters θ1 , θ2 . Given input xi ∈ Rd at time i, let
yi = Φθ1 (xi ), σi = σi−1 ⊗ SigN ((yi−1 , yi )), ai = f θ2 (xi , σi ),
where ai is the action proposed by the network at time i, and yi and σi are the memory at time i,
and ⊗ denotes the tensor product as in A.13 in Appendix A.

8
The model is summarised in Figure 7 as a recurrent neural network with signature-based memory.
Note that yi is preserved in memory only to compute the signature at the next time step, as the
shortest path it is meaningful to compute the signature of is of length two.
However, note that by Proposition A.15 in Appendix A,
N +1
σi = SigN (Φθ1 (x1 ), . . . , Φθ1 (xi )) ∈ R(e −1)/(e−1)
.

Furthermore the xi , yi , σi and ai may be

collected into streams
(xi )i ∈ S(Rd ), yi−1 σi−1
(yi )i ∈ S(Re ),
N +1
(σi )i ∈ S(R(e −1)/(e−1) ),
(ai )i ∈ S({actions}). xi Φθ 1 yi SigN ⊗

In this way we may interpret this model as

a generalisation of deep signature model:
it has a single Neural-Lift-Signature block, f θ2 σi
with a skip connection across the whole
block. The neural component is given by
the neural network Φθ1 , which is stream-
ai
preserving as it operates pointwise, in the
manner of equation (1). The lift is the
‘expanding window’ lift given by equation Figure 7: Agent architecture as a recurrent network.
(6). Finally f θ2 is another neural network, Trainable parameters: θ , θ .
1 2
which is again pointwise and thus stream-
preserving.
This interpretation of the model is demonstrated in Figure 8.
We test this model on a non-Markovian modification to the classical Mountain Car problem [41],
in which the agent receives only partial information: it is only given the car’s position, and not
its [Link] find that it is capable of learning how to solve the problem within a set number of
episodes, whilst a comparable RNN architecture fails to do so. The reinforcement learning technique
used was Deep Q Learning [42] with the specified models performing function approximation on
Q. Both models were chosen to have comparable numbers of parameters. Further implementation
details can be found in Appendix B.

(xi )i Φθ 1 (yi )i ` SigN (σi )i f θ2 (ai )i

Figure 8: Agent architecture as a residual network. Trainable parameters: θ1 , θ2 . The lift ` is the
‘expanding window’ lift of equation (6).

5 Conclusion
There is a strong corpus of theory motivating the use of the signature transform as a tool to
understand streams of data. Meanwhile neural networks have enjoyed great empirical success. It
is thus desirable to bring them together; in this paper we have described how this may be done in
a general fashion, and have provided examples of how this principle may be used in a variety of
domains.
There are two key contributions. First, we discuss stream-preserving neural networks, which are
what allow for using signature transforms deeper within a network, rather than as just a feature
transformation. Second, we discuss lifts, which is what allows for the use of multiple signature
transforms. In this way we have significantly extended the use of the signature transform in machine
learning: rather than limiting its usage to data preprocessing, we demonstrate how the signature
transform, as a univeral nonlinearity, may be used as a pooling layer within a neural network.

9
Acknowledgements
PB was supported by the EPSRC grant EP/R513295/1. PK was supported by the EPSRC grant
EP/L015811/1. PK, IPA, CS, TL were supported by the Alan Turing Institute under the EPSRC
grant EP/N510129/1.

References
[1] I. Chevyrev and A. Kormilitzin, “A primer on the signature method in machine learning,” arXiv preprint
arXiv:1603.03788, 2016.
[2] K. T. Chen, “Iterated integrals and exponential homomorphisms,” Proc. London Math. Soc, 4, 502–512,
1954.
[3] K. T. Chen, “Integration of paths, geometric invariants and a generalized Baker-Hausdorff formula,” Ann.
of Math. (2), 65:163–178, 1957.
[4] K. T. Chen, “Integration of paths - a faithful representation of paths by non-commutative formal power
series,” Trans. Amer. Math. Soc. 89 (1958), 395–407, 1958.
[5] T. Lyons, H. Ni, and H. Oberhauser, “A feature set for streams and an application to high-frequency
financial tick data,” ICBDC, 2014.
[6] L. G. Gyurkó, T. Lyons, M. Kontkowski, and J. Field, “Extracting information from the signature of a
financial data stream,” arXiv preprint arXiv:1307.7244, 2014.
[7] T. Lyons, S. Nejad, and I. P. Arribas, “Nonparametric pricing and hedging of exotic derivatives,” arXiv
preprint arXiv:1905.00711, 2019.
[8] T. Lyons, S. Nejad, and I. P. Arribas, “Model-free pricing and hedging in discrete time using rough path
signatures,” arXiv preprint arXiv:1905.01720, 2019.
[9] J. Kalsi, T. Lyons, and I. P. Arribas, “Optimal execution with rough path signatures,” arXiv preprint
arXiv:1905.00728, 2019.
[10] T. J. Lyons, “Differential equations driven by rough signals,” Revista Matemática Iberoamericana, vol. 14,
no. 2, pp. 215–310, 1998.
[11] P. K. Friz and N. B. Victoir, “Multidimensional stochastic processes as rough paths: theory and
applications,” Cambridge University Press, 2010.
[12] W. Yang, L. Jin, and M. Liu, “Chinese character-level writer identification using path signature
feature, DropStroke and deep CNN,” in 2015 13th International Conference on Document Analysis and
Recognition (ICDAR), pp. 546–550, IEEE, 2015.
[13] Z. Xie, Z. Sun, L. Jin, H. Ni, and T. Lyons, “Learning spatial-semantic context with fully convolutional
recurrent network for online handwritten Chinese text recognition,” IEEE transactions on pattern analysis
and machine intelligence, vol. 40, no. 8, pp. 1903–1917, 2018.
[14] W. Yang, L. Jin, D. Tao, Z. Xie, and Z. Feng, “DropSample: A new training method to enhance deep
convolutional neural networks for large-scale unconstrained handwritten Chinese character recognition,”
Pattern Recognition, vol. 58, pp. 190–203, 2016.
[15] W. Yang, L. Jin, and M. Liu, “Deepwriterid: An end-to-end online text-independent writer identification
system,” IEEE Intelligent Systems, vol. 31, no. 2, pp. 45–53, 2016.
[16] C. Li, X. Zhang, and L. Jin, “LPSNet: a novel log path signature feature based hand gesture recognition
framework,” in Proceedings of the IEEE International Conference on Computer Vision, pp. 631–639,
2017.
[17] W. Yang, T. Lyons, H. Ni, C. Schmid, L. Jin, and J. Chang, “Leveraging the path signature for skeleton-
based human action recognition,” arXiv preprint arXiv:1707.03993, 2017.
[18] W. Yang, L. Jin, H. Ni, and T. Lyons, “Rotation-free online handwritten character recognition using dyadic
path signature features, hanging normalization, and deep neural network,” in 2016 23rd International
Conference on Pattern Recognition (ICPR), pp. 4083–4088, IEEE, 2016.
[19] F. J. Király and H. Oberhauser, “Kernels for sequentially ordered data,” Journal of Machine Learning
Research, 2019.
[20] I. Chevyrev and H. Oberhauser, “Signature moments to characterize laws of stochastic processes,” arXiv
preprint arXiv:1810.10971, 2018.
[21] T. Lyons, “Rough paths, signatures and the modelling of functions on streams,” arXiv preprint
arXiv:1405.4537, 2014.

10
[22] P. Kidger, “Signatory: differentiable computations of the signature and logsignature transforms, on both
CPU and GPU,” 2019. [Link]
[23] A. Silvescu, “Fourier neural networks,” Proceedings of the International Joint Conference On Neural
Networks, IEEE, 1999.
[24] L. Mingo, L. Aslanyan, J. Castellanos, M. Diaz, and V. Riazanov, “Fourier neural networks: an approach
with sinusoidal activation functions,” Int. J. Inf. Theory Appl., 11, 2004.
[25] M. Gashler and S. Ashmore, “Modeling time series data with deep fourier neural networks,”
Neurocomputing, 2016.
[26] Q. Zhang and A. Benveniste, “Wavelet networks,” IEEE Trans. Neural Netw., 1992.
[27] A. Pinkus, “Approximation theory of the MLP model in neural networks,” Acta Numer., vol. 8, pp. 143–
195, 1999.
[28] T. J. Lyons and W. Xu, “Inverting the signature of a path,” Journal of the European Mathematical Society,
vol. 20, no. 7, pp. 1655–1687, 2018.
[29] J. Chang, N. Duffield, H. Ni, W. Xu, et al., “Signature inversion for monotone paths,” Electronic
Communications in Probability, vol. 22, 2017.
[30] J. Chang, Effective algorithms for inverting the signature of a path. PhD thesis, University of Oxford,
2018.
[31] D. Dua and C. Graff, “UCI Machine Learning Repository,” 2017.
[32] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and
Y. Bengio, “Generative adversarial nets,” in Advances in neural information processing systems, pp. 2672–
2680, 2014.
[33] Y. Li, K. Swersky, and R. Zemel, “Generative moment matching networks,” ICML, 2015.
[34] G. K. Dziugaite, D. M. Roy, and Z. Ghahramani, “Training generative neural networks via maximum
mean discrepancy optimization,” UAI, 2015.
[35] A. Gretton, K. M. Borgwardt, M. Rasch, B. Scholkopf, and A. J. Smola, “A kernel method for the two-
sample problem,” Advances in Neural Information Processing Systems, 2007.
[36] G. E. Uhlenbeck and L. S. Ornstein, “On the theory of the Brownian motion,” Physical review, vol. 36,
no. 5, p. 823, 1930.
[37] Y. Mishura, Stochastic calculus for fractional Brownian motion and related processes, vol. 1929. Springer
Science & Business Media, 2008.
[38] J. Gatheral, T. Jaisson, and M. Rosenbaum, “Volatility is Rough,” Quantitative Finance, 18:6, 933-949,
2018.
[39] L. Lacasa, B. Luque, J. Luque, and J. C. Nuno, “The visibility graph: A new method for estimating the
Hurst exponent of fractional Brownian motion,” EPL (Europhysics Letters), vol. 86, no. 3, p. 30001, 2009.
[40] H. Hurst, “The Long-Term Storage Capacity of Reservoirs,” Transactions of the American Society of Civil
Engineers, 1951.
[41] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba, “Openai
gym,” arXiv preprint arXiv:1606.01540, 2016.
[42] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller,
A. K. Fidjeland, G. Ostrovski, et al., “Human-level control through deep reinforcement learning,” Nature,
vol. 518, no. 7540, p. 529, 2015.
[43] B. M. Hambly and T. J. Lyons, “Uniqueness for the signature of a path of bounded variation and the
reduced path group,” Annals of Mathematics, vol. 171, no. 1, pp. 109–167, 2010.
[44] I. Perez Arribas, “Derivatives pricing using signature payoffs,” arXiv preprint arXiv:1809.09466, 2018.
[45] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” ICLR, 2015.
[46] A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and
A. Lerer, “Automatic differentiation in PyTorch,” 2017.
[47] J. Reizenstein and B. Graham, “The iisignature library: efficient calculation of iterated-integral signatures
and log signatures,” arXiv preprint arXiv:1802.08252, 2018.
[48] P. Embrechts, Selfsimilar processes, vol. 21. Princeton University Press, 2009.
[49] C. J. C. H. Watkins, “Learning from delayed rewards,” 1989.

11