Multivariate Transiet Price Impact Model

This article was downloaded by: [155.198.30.
43] On: 26 August 2019, At: 10:36

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA
Mathematics of Operations Research

Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
Multivariate Transient Price Impact and Matrix-Valued

Positive Definite Functions
Aurélien Alfonsi, Florian Klöck, Alexander Schied
To cite this article:

Aurélien Alfonsi, Florian Klöck, Alexander Schied (2016) Multivariate Transient Price Impact and Matrix-Valued Positive
Definite Functions. Mathematics of Operations Research 41(3):914-934. https://doi.org/10.1287/moor.2015.0761
Full terms and conditions of use: https://pubsonline.informs.org/page/terms-and-conditions
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
Copyright © 2016, INFORMS
Please scroll down for article—it is on subsequent pages
INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
MATHEMATICS OF OPERATIONS RESEARCH
Vol. 41, No. 3, August 2016, pp. 914–934
ISSN 0364-765X (print) ISSN 1526-5471 (online)
http://dx.doi.org/10.1287/moor.2015.0761
© 2016 INFORMS
Multivariate Transient Price Impact and Matrix-Valued

Positive Definite Functions
Aurélien Alfonsi
Université Paris-Est, CERMICS, Projet MathRisk ENPC-INRIA-UMLV, Ecole des Ponts, 77455 Marne La Vallée,
France, alfonsi@cermics.enpc.fr
Florian Klöck, Alexander Schied

Department of Mathematics, University of Mannheim, A5, 6, 68131 Mannheim, Germany
{wima1@uni-mannheim.de, schied@uni-mannheim.de}
We consider a model for linear transient price impact for multiple assets that takes cross-asset impact into account. Our main
goal is to single out properties that need to be imposed on the decay kernel so that the model admits well-behaved optimal
trade execution strategies. We first show that the existence of such strategies is guaranteed by assuming that the decay kernel
corresponds to a matrix-valued positive definite function. An example illustrates, however, that positive definiteness alone
does not guarantee that optimal strategies are well-behaved. Building on previous results from the one-dimensional case, we
investigate a class of nonincreasing, non-negative, and convex decay kernels with values in a space of symmetric matrices.
We show that these decay kernels are always positive definite and characterize when they are even strictly positive definite, a
result that may be of independent interest. Optimal strategies for kernels from this class are particularly well-behaved if one
requires that the decay kernel is also commuting. We show how such decay kernels can be constructed by means of matrix
functions and provide a number of examples. In particular, we completely solve the case of matrix exponential decay.
Keywords: multivariate price impact; matrix-valued positive definite function; optimal trade execution; optimal portfolio
liquidation; matrix function
MSC2000 subject classification: 42A82, 90C20, 91G80, 91G10
OR/MS subject classification: Primary: Finance: portfolio; secondary: mathematics: functions, matrices
History: Received October 16, 2013; revised December 16, 2014, July 19, 2015. Published online in Articles in Advance
March 1, 2016.
1. Introduction. Price impact refers to the feedback effect of trades on the quoted price of an asset and
it is responsible for the creation of execution costs. It is an empirically established fact that price impact is
predominantly transient; see, e.g., Moro et al. [26]. When trading speed is sufficiently slow, the effects of
transience can be reduced to considering only a temporary and a permanent price impact component (Bertsimas
and Lo [6], Almgren and Chriss [5]). For higher trading speeds, however, one needs a model that explicitly
describes the decay of price impact between trades. First models of this type were proposed by Bouchaud
et al. [9] and Obizhaeva and Wang [28]. These models were later extended into various directions by Alfonsi
et al. [1, 2], Gatheral [16], Alfonsi et al. [4], Gatheral et al. [18], Predoiu et al. [30], Fruth et al. [15], and
Løkka [25], to mention only a few. A more comprehensive list of references can be found in Gatheral and
Schied [17]. We also refer to Guo [22] for an introduction to the microscopic order book picture that is behind
the mesoscopic models mentioned above.
All above-mentioned models for transient price impact deal only with one single risky asset. Although multi-
asset models for temporary and permanent price impact (Schöneborn [33]) or for generic price impact functionals
(Schied et al. [32], Kratz and Schöneborn [24]) were considered earlier, we are not aware of any previous
approaches to analyzing the specific effects of transient cross-asset price impact. Our goal in this paper is to
propose and analyze a simple model for transient price impact between K different risky assets. Following the
one-dimensional ansatz of Gatheral [16], the time-t impact on the price of the ith asset that is generated by
trading one unit of the jth asset at time s < t will be described by the number Gij 4t − s5 for a certain function
Gij 2 601 5 → . The matrix-valued function G4t5 = 4Gij 4t55i1 j=11 : : : 1K will be called the decay kernel of the
multi-asset price impact model.
When setting up such a model in a concrete situation, the first question one encounters is how to choose
the decay kernel. Already in the one-dimensional situation, K = 1, the decay kernel G needs to satisfy certain
conditions so that the resulting price impact model has some minimal regularity properties such as the existence
of optimal trade execution strategies, the absence of price manipulation in the sense of Huberman and Stanzl [23],
or the nonoccurrence of oscillatory strategies. It was shown in Alfonsi et al. [4] that these properties are satisfied
when G is non-negative, nonincreasing, and convex. Here we will continue the corresponding analysis and extend
it to matrix-valued decay kernels G. Our first observation is that G must correspond to a certain matrix-valued
positive definite function. Such functions were previously characterized and analyzed, e.g., by Cramér [12],
914
Alfonsi, Klöck, and Schied: Multivariate Transient Price Impact and Matrix-Valued Positive Definite Functions
Mathematics of Operations Research 41(3), pp. 914–934, © 2016 INFORMS 915
Naimark [27], Falb [14]. An example illustrates, however, that positive definiteness alone does not guarantee
that optimal strategies are well-behaved. We therefore introduce a class of nonincreasing, non-negative, and
convex decay kernels with values in the symmetric K × K matrices. We show that these decay kernels are
always positive definite, and we characterize in Theorem 2 when they are even strictly positive definite. Optimal
strategies for kernels from this class do not admit oscillations if one additionally requires that the decay kernel
is commuting. Based on this result, we will address in §2.5 the problem of optimizing simultaneously over time
grids and strategies and state the solution in terms of a suitable continuous-time limit. We finally show how such
decay kernels can be constructed by means of matrix functions and provide a number of examples. In particular,
we completely solve the case of matrix exponential decay.
Our main general results are stated in §2. Transformation results for decay kernels and their optimal strategies
along with several explicit examples are given in §3. Since the situation K > 1 is considerably more complex
than the one-dimensional case, we have summarized the main conclusions that can be drawn from our results
in §4. These conclusions will focus on our initial question: From which class of functions should decay kernels
for transient price impact be chosen? Most proofs are given in §5.
2. Statement of general results. In this section, we first introduce a linear market impact model with
transient price impact for K different risky assets. We then discuss which properties a decay kernel should
satisfy so that the corresponding market impact model has certain desirable features and properties. Two of
these properties are the existence of optimal strategies and the absence of price manipulation strategies in the
sense of Huberman and Stanzl [23], which we will both characterize by establishing a link to the theory of
positive definite matrix-valued functions. Requiring positive definiteness, however, will typically not be sufficient
to guarantee that optimal strategies are well-behaved. We will thus be led to a more detailed analysis of positive
definite matrix-valued functions and the associated quadratic minimization problems, an analysis that might be
of independent interest.
2.1. Preliminaries. We introduce here a market impact model for an investor trading in K different securi-
ties. When the investor is not active, the unaffected price process of these assets is given by a right-continuous
K-dimensional martingale 4St0 5t∈601 T 7 defined on a filtered probability space 4ì1 F1 4Ft 5t∈601 T 7 1 5. Now suppose
that the investor can trade at the times of a time grid = 8t1 1 : : : 1 tN 9, where N ∈ and 0 = t1 < t2 < · · · < tN
(an extended setup with the possibility of trading in continuous time will be considered in §2.5). The size of
the order in the ith asset at time tk is described by a Ftk -measurable random variable ki , where positive values
denote buys and negative values denote sells. By k = 4k1 1 : : : 1 kK 5> we denote the column vector of all orders
placed at time tk . Our main interest here will be in admissible strategies that -a.s. liquidate a given initial
portfolio X0 ∈ K . Such strategies are needed in practice when the initial portfolio X0 is too big to be liquidated
immediately; see, e.g., Almgren and Chriss [5].
Definition 1. Let = 8t1 1 : : : 1 tN 9 be a time grid. An admissible strategy for is a sequence Î =
41 1 : : : 1 N 5 of bounded1 K-dimensional random variables such that each k is Ftk -measurable; Î is called
deterministic if each ki does not depend on ∈ ì. The set of admissible liquidation strategies for a given initial
portfolio X0 ∈ K and is defined as
N

X
X41 X0 5 2= Î = 41 1 : : : 1 N 5 Î is admissible and X0 + k = 0 -a.s. 0 (1)
k=1
The set of deterministic liquidation strategies in X41 X0 5 is denoted by Xdet 41 X0 5.

We now turn toward the definition of the price impact generated by an admissible strategy. As discussed in
more detail in the introduction, in recent years several models were proposed that take the transience of price
impact into account. All these models, however, consider only one risky asset. In this paper, our goal is to
extend the model from Alfonsi et al. [4], which is itself a linear and discrete-time version of the model from
Gatheral [16], to a situation with K > 1 risky assets. A decay kernel will be a continuous function
G2 601 5 −→ K×K
1
Boundedness is assumed here for simplicity and can easily be relaxed; for instance, it is enough to assume that both k and S 0 are
square-integrable. Since the total number of shares of an asset is always finite, boundedness can be assumed without loss of generality from
an economic point of view.
916 Mathematics of Operations Research 41(3), pp. 914–934, © 2016 INFORMS
taking values in the space K×K of all real K × K-matrices. When Î is an admissible strategy for some time
grid = 8t1 1 : : : 1 tN 9 and t ≥ tk ∈ , the value Gij 4t − tk 5 describes the time-t impact on the price of the ith
asset that was generated by trading one unit of the jth asset at time tk . We therefore define the impacted price
process as
StÎ = St0 +
X
G4t − tk 5k 1 t ≥ 00 (2)
tk <t
Here G4t − tk 5k denotes the application of the K × K matrix G4t − tk 5 to the K-dimensional vector k .
Let us write StÎ1 i for the ith component of the price vector StÎ = 4StÎ1 1 1 : : : 1 StÎ1 K 5> . The execution of the
kth order, k , shifts the price of the ith asset linearly from StÎ1k i to StÎ1k +i . The order ki of shares of the ith
asset is therefore executed at the average price 21 4StÎ1k +i + StÎ1k i 5. The proceeds from executing the amount of ki
shares of the ith asset are therefore given by −ki 21 4StÎ1k +i + StÎ1k i 5. It follows that the total revenues incurred by
the strategy Î are given by
N
1X
R4Î5 = − > 4StÎ + + StÎk 50 (3)
2 k=1 k k
In the sequel, it will be convenient to switch from revenues to costs, which are defined as the amount X0> S00 −
R4Î5 by which the revenues fall short of the book value, X0> S00 , of the initial portfolio.
Remark 1. In the one-dimensional version of our model, a bid-ask spread is often added so as to provide an
interpretation of k as a market order placed in a block-shaped limit order book; see, e.g., Section 2.6 in Alfonsi
and Schied [3]. In practice, however, execution algorithms will use a variety of different order types, and one
should think of price impact and costs as being aggregated over these order types. For instance, although half
the spread has to be paid when placing a market buy order, the same amount can be earned when a limit sell
order is executed. Other order types may yield rebates when executed or may allow execution at mid price. So
ignoring the bid-ask spread is probably more realistic than adding it to each single execution of an order.
In this paper, we will investigate the minimization of the expected costs of a strategy, which in many situations
is an appropriate optimization problem for determining optimal trade execution strategies. Our main interest,
however, is to provide conditions on the decay kernel G under which the model is sufficiently regular. As
discussed at length in Gatheral and Schied [17], the regularity of a market impact model should be measured by
the existence and behavior of execution strategies that minimize the expected costs, because the regularity of a
model should be considered independently from the possible risk aversion that an agent using this model might
have.
To analyze the expected costs of an admissible strategy Î = 41 1 : : : 1 N 5, it will be convenient to identify the
particular realization, Î45 = 41 451 : : : 1 N 455, with an element of the tensor product space N ⊗ K . We
will also write for the cardinality of a time grid.
Lemma 1. The expected costs of a strategy Î ∈ X41 X0 5 for a time grid are given by
Ɛ6X0> S0 − R4Î57 = Ɛ6C 4Î571 (4)
where the cost function C 2 ⊗ K → is given by
N
1 X
C 4Î5 = > G̃4tk − t` 5` (5)
2 k1 `=1 k
for the function G̃2 → K×K defined by


G4t5
 for t > 0,
G̃4t5 2= 21 4G405> + G4055 for t = 0, (6)

G4−t5> for t < 0.

We will now discuss the possible existence and structure of admissible strategies minimizing the expected
costs within the class X41 X0 5. The problem of optimizing simultaneously over time grids and strategies
Î ∈ X41 X0 5 will be addressed in §2.5.
Lemma 2. There exists a strategy in X41 X0 5 that minimizes the expected costs Ɛ6C 4Ç57 among all strate-
gies Ç ∈ X41 X0 5 if and only if there exists a deterministic strategy that minimizes the cost function C 4Î5
over all Î ∈ Xdet 41 X0 5. In this case, any minimizer Ç∗ ∈ X41 X0 5 can be regarded as a function from ì into
Xdet 41 X0 5 that takes -a.s. values in the set of deterministic minimizers of the cost function C 4 · 5.
The condition
Ɛ6C 4Ç57 ≥ 0 for all 1 X0 ∈ k 1 and Ç ∈ X41 X0 5 (7)
can be regarded as a regularity condition for the underlying market impact model. It rules out the possibility
of obtaining positive expected profits through exploiting one’s own price impact; see, e.g., Alfonsi et al. [4] or
Gatheral and Schied [17] for detailed discussions. In particular, it rules out the existence of price manipulation
strategies in the sense of Huberman and Stanzl [23]. In the sequel we will therefore focus on decay kernels that
satisfy (7). It will turn out that (7) can be equivalently characterized by requiring that the function G̃ from (6)
is a positive definite matrix-valued function in the following sense.
Definition 2. A function H 2 → K×K is called a positive definite matrix-valued function if for all N ∈ ,
t1 1 : : : 1 tN ∈ , and z1 1 : : : 1 zN ∈ K ,
N
X
z∗i H 4ti − tj 5zj ≥ 01 (8)
i1 j=1
∗
where a -superscript denotes the usual conjugate transpose of a complex vector or matrix. If moreover equality
in (8) can hold only for z1 = · · · = zN = 0, then H is called strictly positive definite. When K = 1, we say that
H is a 4strictly5 positive definite function.
Note that a positive definite matrix-valued function H is defined on the entire real line and is allowed
to take values in the complex matrices. A decay kernel, G, on the other hand, is defined only on 601 5 and
takes values in the real matrices, K×K . Considering the extended framework of K×K -valued positive definite
functions will turn out to be convenient for our analysis. The next proposition explains the relation between
positive definite functions and decay kernels with non-negative expected costs.
Proposition 1. For a decay kernel G, the following conditions are equivalent.
(a) Ɛ6 C 4Ç5 7 ≥ 0 for all time grids , initial portfolios X0 ∈ K , and Ç ∈ X41 X0 5.
(b) C 4Î5 ≥ 0 for all time grids and Î ∈ ⊗ K .
(c) For all time grids , C 2 ⊗ K → is convex.
(d) G̃ defined in (6) is a positive definite matrix-valued function.
If moreover these equivalent conditions are satisfied, then the equality C 4Î5 = 0 holds for all time grids
only for Î = 0, if and only if G̃ is strictly positive definite. In this case, C 2 ⊗ K → is strictly convex
for all .
Positive definiteness of G̃ not only excludes the existence of price manipulation strategies. The following
proposition states that it also guarantees the existence of strategies that minimize the expected costs within a class
X41 X0 5. Such strategies will be called optimal strategies in the sequel. Once the existence of optimal strategies
has been established, they can be computed by means of standard techniques from quadratic programming (see,
e.g., Boot [8] or Gill et al. [20]).
Proposition 2. Suppose that G̃ is positive definite. Then there exists an optimal strategy in Xdet 4X0 1 5
(and hence in X4X0 1 5) for all X0 ∈ K and each time grid . Moreover, a strategy Î ∈ Xdet 4X0 1 5 is optimal
if and only if there exists ∈ K such that
N
X
G̃4tk − t` 5` = for k = 11 : : : 1 0 (9)
`=1
If G̃ is strictly positive definite, then optimal strategies and the Lagrange multiplier in (9) are unique.
Propositions 1 and 2 suggest that decay kernels G for multivariate price impact should be constructed such
that the corresponding function G̃ from (6) is a positive definite matrix-valued function. Part (a) of the following
elementary lemma implies that this can be achieved by defining G4t5 2= H 4t5 for t ≥ 0 when H 2 → K×K is
a given continuous positive definite matrix-valued function, because we will then automatically have G̃ = H .
Lemma 3. Let H 2 → K×K be a positive definite matrix-valued function. Then:
(a) The matrix H 405 is non-negative definite, and we have H 4−t5 = H 4t5∗ for every t ∈ . In particular,
H 4−t5 = H 4t5> if H takes its values in K×K .
(b) Also t 7→ H 4t5∗ is a positive definite matrix-valued function; it is strictly positive definite if and only H
is strictly positive definite.
Because of the established one-to-one correspondence of decay kernels with non-negative expected costs and
continuous K×K -valued positive definite functions, we will henceforth use the following terminology.
Definition 3. A decay kernel G2 601 5 → K×K is called 4strictly5 positive definite if the corresponding
function G̃ from (6) is a (strictly) positive definite matrix-valued function.
2.2. Integral representation of positive definite decay kernels. We turn now to characterizations of the
positive definiteness of a matrix-valued function. In the one-dimensional situation, K = 1, Bochner’s theorem
(Bochner [7]) characterizes all continuous positive definite functions as the Fourier transforms of non-negative
finite Borel measures. There are several extensions of Bochner’s theorem to the case of matrix- or operator-valued
functions. Some of these results will be combined in Theorem 1 and Corollary 1 below. For the corresponding
statements, we first introduce some terminology.
As usual, a complex matrix N ∈ n×n is called non-negative definite if z∗ Nz ≥ 0 for every z ∈ n . When
even z∗ Nz > 0 for every nonzero z, N is called strictly positive definite. A non-negative definite complex matrix
N ∈ n×n is necessarily Hermitian, i.e., N = N ∗ . In particular, a real matrix N ∈ n×n is non-negative definite
if and only if it belongs to the set + 4n5 of non-negative definite symmetric real n × n-matrices. By 4n5 we
denote the set of all symmetric matrices in n×n . An arbitrary real matrix M ∈ n×n will be called non-negative
if x> Mx ≥ 0 for every x ∈ n and strictly positive if x> Mx > 0 for all nonzero x ∈ n . Note that a real matrix
M ∈ n×n is non-negative if and only if its symmetric part, 21 4M + M > 5, is non-negative definite.
Let B45 be the Borel -algebra on . A mapping M2 B45 → K×K will be called a non-negative definite
matrix-valued measure if every component Mij is a complex measure with finite total variation and the matrix
M4A5 ∈ K×K is non-negative definite for every A ∈ B45.
The following theorem combines results by Cramér [12], Falb [14], and Naimark [27]; we refer to Glöck-
ner [21] for extensions of this result and for a comprehensive historical account.
Theorem 1. For a continuous function H 2 → K×K the following are equivalent.
(a) H is a positive definite matrix-valued function.
(b) For every z ∈ K , the complex function t 7→ z∗ H 4t5z is positive definite.
(c) H is the Fourier transform of a non-negative definite matrix-valued measure M, i.e.,
Z
H 4t5 = eit M4d5 for t ∈ . (10)

Moreover, any matrix-valued measure M with (10) is uniquely determined by H.

Proof. The equivalence of (b) and (c) was proved in Falb [14]. The equivalence of (a) and (c) follows from
two statements in Gihman and Skorohod [19], namely the remark after Theorem 1 in §1 of Chapter IV and
Theorem 5 in §2 of Chapter IV. The uniqueness of M is standard.
The preceding theorem simplifies as follows when considering positive definite functions H taking values in
the space 4K5 of symmetric real K × K-matrices. By Lemma 3(a), such functions H correspond to positive
definite decay kernels G that are symmetric in the sense that G4t5> = G4t5 for all t ≥ 0. In this case, we have
H 4t5 = G̃4t5 = G4t5 for all t ∈ .
Corollary 1. For a continuous function H 2 → K×K the following statements are equivalent.
(a) H 4t5 ∈ 4K5 for all t, and H is a positive definite matrix-valued function.
(b) H 4t5 ∈ 4K5 for all t, and the real function t 7→ x> H 4t5x is positive definite for every x ∈ K .
(c) H admits a representation (10) with a non-negative definite measure M that takes values in K×K (and
hence in + 4K5) and is symmetric on in the sense that M4A5 = M4−A5 for all A ∈ B45.
Remark 2 (Discontinuous Positive Definite Functions and Temporary Price Impact). Let H0 be a
nonzero non-negative definite matrix. Then H 4t5 2= H0 809 4t5 is a positive definite matrix-valued function that
is not continuous and therefore does not admit a representation (10). It is possible, however, to give a similar
integral representation also for discontinuous matrix-valued positive definite functions satisfying a certain bound-
edness condition. To this end, one needs to replace the measure M by a non-negative definite matrix-valued
measure on the larger space of characters for the additive (semi-)groups or + ; see Glöckner [21, Theo-
rem 15.7]. In the context of price impact modeling, the costs (5) associated with a discontinuous decay kernel of
the form G4t5 2= G0 809 4t5 for some non-negative matrix G0 can be viewed as resulting from temporary price
impact that affects only the order that has triggered it and disappears immediately afterward; see Bertsimas and
Lo [6] and Almgren and Chriss [5] for temporary price impact in one-dimensional models. More generally,
to take account the discontinuity G405 − G40+5 ∈ + 4K5, one will have to precise the definition (3) of the
revenues by assuming R4Î5 = − Nk=1 k> 4StÎk + 21 G405k 5 (note that this is G405 and not G40+5). Last, let us
P
mention that the discontinuity at 0 is the only one relevant in practice: other discontinuities would generate a
weird and predictable price impact. Thus, the temporary price impact can be handled separately and assuming
G continuous is not restrictive.
2.3. Convex, nonincreasing, and non-negative decay kernels. As shown and discussed in Alfonsi et al. [4],
not every decay kernel G2 601 5 → with positive definite G̃ is a reasonable model for the decay of price
impact in a single-asset model. Specifically it was shown that for K = 1 it makes sense to require that decay
kernels are non-negative, nonincreasing, and convex. Since similar effects as in Alfonsi et al. [4] can also be
observed in our multivariate setting (see Figure 1), we need to introduce and analyze further conditions to be
satisfied by G. To motivate the following definition, consider two trades 1 and 2 placed at times t1 < t2 . The
quantity 2> G4t2 − t1 51 describes that part of the liquidation costs for the order 2 that was caused by the
order 1 . When 1 = 2 , it is intuitively clear that these costs should be non-negative and nonincreasing in t2 − t1 .
Definition 4. A matrix-valued function G2 601 5 → K×K is called
(a) nonincreasing, if for every x ∈ K the function t 7→ x> G4t5x is nonincreasing;
(b) non-negative, if G4t5 is a non-negative matrix for every t ∈ 601 5;
(c) 4strictly5 convex, if for all x ∈ K the function t 7→ x> G4t5x is (strictly) convex.
Here and in Lemma 4 and Theorem 2 below, we do not assume that G is continuous. Note that the properties
introduced in the preceding definition depend only on the symmetrization, 21 4G> + G5, of G. We have the
following simple result on two properties introduced in Definition 4.
Lemma 4. Suppose that G2 601 5 → K×K is a nonincreasing and positive definite decay kernel. Then G
is non-negative.
If G is nonincreasing, non-negative, and convex, then so is the function g x 4t5 2= x> G4t5x for each x ∈ K .
Hence, t 7→ g x 4t5 is a positive definite function due to a criterion often attributed to Pólya [29], although this
criterion is also an easy consequence of Young [34]. It hence follows from Corollary 1 that also the matrix-valued
function G̃ is positive definite as soon as G is symmetric and continuous. But an even stronger result is possible:
G is even strictly positive definite as soon as g x is nonincreasing, non-negative, convex, and nonconstant for
each nonzero x ∈ K . This is the content of our subsequent theorem, which extends the corresponding result
for K = 1 (see Theorems 3.9.11 and 3.1.6 in Sasvári [31] or Proposition 2 in Alfonsi et al. [4] for two different
proofs) and is of independent interest.
Theorem 2. If G2 601 5 → K×K is symmetric, non-negative, nonincreasing, and convex then G is positive
definite. Moreover, G is even strictly positive definite if and only if t 7→ x> G4t5x is nonconstant for each nonzero
x ∈ K .
We will see in Proposition 9 that in Theorem 2 we can typically not dispense of the requirement that G is
symmetric to conclude positive definiteness.
1,000
500
–500
–1,000
Figure 1. Optimal strategy Î for X0 = 4101 05, N = 23, and the strictly positive definite decay kernel G4t5 = exp4−4tB52 5 for B = 1 1 .

The first component of Î is plotted in light gray, the second component in dark gray. Note that the amplitude of the oscillations exceeds the
initial asset position by a factor of more than 110. That G is strictly positive definite follows from Remark 4.
2.4. Commuting decay kernels. We will now introduce another property that one can require from a decay
kernel.
Definition 5. A decay kernel G2 601 5 → K×K is called commuting if G4t5G4s5 = G4s5G4t5 holds for
all s1 t ≥ 0.
If a symmetric decay kernel is commuting, it may be simultaneously diagonalized, and its properties can
be characterized via the resulting collection of one-dimensional decay kernels, as explained in the following
proposition.
Proposition 3. A symmetric decay kernel G is commuting if and only if there exists an orthogonal matrix
O and functions g1 1 : : : 1 gK 2 601 5 → such that
G4t5 = O > diag4g1 4t51 : : : 1 gK 4t55O0 (11)
Moreover, the following assertions hold.
(a) G is (strictly) positive definite if and only if the -valued functions t 7→ gi 4t5 are (strictly) positive
definite for all i.
(b) G is non-negative if and only if gi 4t5 ≥ 0 for all i and t.
(c) G is nonincreasing if and only if gi is nonincreasing for all i.
(d) G is convex if and only if gi is convex for all i.
(e) If G is positive definite, then a strategy Î = 41 1 : : : 1 5 ∈ Xdet 41 X0 5 is optimal if and only if it is of
the form
j = O > 4j1 1 : : : 1 jK 5> 1
i
where Çi = 41i 1 : : : 1 5 ∈ ⊗ is an optimal strategy in Xdet 41 4OX0 5i 5 for the one-dimensional decay
i
kernel gi (here 4OX0 5 denotes the ith component of the vector OX0 ).
For K = 1, we know that a non-negative nonincreasing convex function is positive definite, and even strictly
positive definite when it is nonconstant. Thus, Proposition 3 implies Theorem 2 in the special situation of
commuting decay kernels.
In the case K = 1, it was observed in Alfonsi et al. [4] that there exist nonincreasing, non-negative, and strictly
positive definite decay kernels G for which the optimal strategies exhibit strong oscillations between buy and
sell orders (“transaction-triggered price manipulation”); see Figure 1 for an example in our multivariate setting.
Theorem 1 in Alfonsi et al. [4] gives conditions that exclude such oscillatory strategies for K = 1 and guarantee
that optimal strategies are buy-only or sell-only: G should be non-negative, nonincreasing, and convex. For
K > 1, however, the situation changes and one cannot expect to exclude the coexistence of buy and sell orders
in the same asset. The reason is that liquidating a position in a first asset may create a drift in the price of
a second asset through cross-asset price impact. Exploiting this drift in the second asset via a round trip may
help to mitigate the costs resulting from liquidating the position in the first asset; see Figure 2. Therefore, one
cannot hope to completely rule out all round trips for decay kernels that are not diagonal. Nevertheless, our
next result gives conditions on G under which optimal strategies can be expressed as linear combinations of
K strategies with buy-only/sell-only components which leads to a uniform bound of the total number of shares
traded by the optimal strategy, preventing large oscillations as in Figure 1. This result will also allow us to
construct minimizers on nondiscrete time grids in §2.5.
10
9
8
7
1 2 3 4 5 6 7 8 90 99
6
5
4
3
2
1
1 2 3 4 5 6 7 8 9 10 11
Figure 2. Optimal strategy Î for G as in Corollary 2 with = 1, ˜ = 108, = 003, X0 = 4−501 15> , T = 5, and N = 11. Left: 11 1 : : : 1 11
1
,
right: 12 1 : : : 1 11
2
.
Proposition 4. Let G be a symmetric, non-negative, nonincreasing, convex and commuting decay kernel.
Then there exist an orthonormal basis v1 1 : : : 1 vK of K and, for each time grid , optimal strategies Î4i5 ∈
Xdet 41 vi 5, i = 11 : : : 1 K, such that the following conditions hold.
(a) The components of each Î4i5 consist of buy-only or sell-only strategies. More precisely, for i1 j ∈
811 : : : 1 K9 and n1 m ∈ 811 : : : 1 9 we have m4i51 j n4i51 j ≥ 0.
(b) For X0 = Ki=1 i vi ∈ K given, Î 2= Ki=1 i Î4i5 is an optimal strategy in Xdet 41 X0 5.
P P
Note that for K = 1 every decay kernel is symmetric and commuting. Hence, for K = 1 the preceding
proposition reduces to Theorem 1 in Alfonsi et al. [4]: The optimal strategy for a one-dimensional, nonconstant,
non-negative, nonincreasing, and convex decay kernel is buy-only or sell-only.
Remark 3. Oscillations of trading strategies as those observed in Figure 1 can be prevented by adding
sufficiently high transaction costs to each trade. Such transaction costs arise naturally if only market orders are
permitted; see, e.g., Sections 7.1 and 7.2 in Busseti and Lillo [10]. As discussed in Remark 1, however, actual
trading strategies will often incur much lower transaction costs than strategies that only use market orders and,
if transaction costs are sufficiently small, oscillations may only be dampened, but not be completely eliminated.
As a matter of fact, oscillatory trading strategies of high-frequency traders played a major role in the “Flash
Crash” of May 6, 2010; see CFTC-SEC [11, p. 3].
Propositions 3 and 4 give not only a characterization of nice properties of certain decay kernels. They also
provide a way of constructing decay kernels G that have all desirable properties. One simply needs to start with
an orthogonal matrix O and nonincreasing, convex, and nonconstant functions g1 1 : : : 1 gK 2 601 5 → 601 5 and
then define a decay kernel as G4t5 = O > diag4g1 4t51 : : : 1 gK 4t55O. A special case of this construction is provided
by the so-called matrix functions, which we will explain in the sequel; see also §3.2 for several examples in this
context.
Let g2 601 5 → be a function and B ∈ + 4K5. Then there exists an orthogonal matrix O such that
B = O > diag41 1 : : : 1 K 5O, where 1 1 : : : 1 K ≥ 0 are the eigenvalues of B. The matrix g4B5 ∈ 4K5 is then
defined as
g4B5 2= O > diag4g41 51 : : : 1 g4K 55O3 (12)
see, e.g., Donoghue [13]. We can thus define a decay kernel G2 601 5 → 4K5 by
G4t5 = g4tB5 = O > diag4g4t1 51 : : : 1 g4tK 55O1 t ≥ 00 (13)
We summarize the properties of G in the following remark. In §3.2 we will analyze decay kernels that arise as
matrix exponentials and explicitly compute the corresponding optimal strategies.
Remark 4. The decay kernel G defined in (13) is commuting. Moreover, it is of the form (11) with gi 4t5 =
g4ti 5, and so Proposition 3 characterizes the properties of G. In particular, it is positive definite if and only
if t 7→ g4t5 is a positive definite function. Moreover, it will be non-negative, nonincreasing, or convex if and
only if g has the corresponding properties. In addition, optimal strategies can be computed via Proposition 3(e).
Remark 5. Let G be a non-negative, symmetric and commuting decay kernel, and f 2 601 5 → 601 5 be
a convex nondecreasing function. We define the kernel F by F 4t5 2= f 4G4t55 as in (12). We get easily from
Proposition 3 that F is nonincreasing and convex if G is also nonincreasing and convex. It is therefore positive
definite in this case.
2.5. Strategies on nondiscrete time grids. If and 0 are time grids such that ⊂ 0 , then X41 X0 5 ⊂
X40 1 X0 5 and hence
min Ɛ6C 4Î57 ≥ 0 min0 Ɛ6C0 4Î0 570
Î∈X41 X0 5 Î ∈X4 1 X0 5
It is therefore clear that problem of minimizing Ɛ6C 4Î57 jointly over Î ∈ X41 X0 5 and time grids has in
general no solution within the class of finite time grids. For this reason it is natural to consider an extension of
our framework to nondiscrete time grids. For the one-dimensional case K = 1 a corresponding framework was
developed in Gatheral et al. [18]. Proposition 4(b) will enable us to obtain a similar extension in our present
framework.
Definition 6. Let be an arbitrary compact subset of 601 T 7. An admissible strategy for is a left-
continuous, adapted, and bounded K-dimensional stochastic process 4Xt 5 such that t 7→ Xti is of finite variation
for i = 11 : : : 1 K and satisfies Xt = 0 for all t > T . We assume furthermore that the vector-valued random
measure dXt is supported on and that its components have -a.s. bounded total variation. The class of
strategies with given initial condition X0 will be denoted by X41 X0 5, the subset of deterministic strategies in
X41 X0 5 will be denoted by Xdet 41 X0 5.
If = 8t1 1 : : : 1 tN 9 is a finite time grid and Î ∈ X41 X0 5 is an admissible strategy in the sense of Defini-
tion 1, then
XtÎ 2= X0 −
X
k (14)
tk ∈1 tk <t
is an admissible strategy in the sense of Definition 6. Therefore, Definition 6 is consistent with Definition 1.
Now let be an arbitrary compact subset of 601 T 7 and G be a decay kernel. For X ∈ X41 X0 5 we define the
associated costs as >
1Z Z
C 4X5 2= G̃4t − s5 dXs dXt 0
2
When is a finite time grid, Î ∈ X41 X0 5, and X Î is defined by (14) then we clearly have C 4X Î 5 = C 4Î5,
and so also the definition of the cost functional is consistent with our earlier definition for discrete time grids.
We have the following result.
Theorem 3. Let G be a symmetric, non-negative, nonincreasing, convex, nonconstant, and commuting decay
kernel and be a compact subset of 601 T 7. Then the following assertions hold.
(a) For X0 ∈ K there exists precisely one strategy X ∗ ∈ X41 X0 5 that minimizes the expected costs,
Ɛ6C 4X57, over all strategies X ∈ X41 X0 5. Moreover, X ∗ is deterministic and can be characterized as the
unique strategy in Xdet 41 X0 5 that solves the following generalized Fredholm integral equation for some ∈ K ,
Z
G̃4t − s5 dXs = for all t ∈ . (15)

(b) Let T denote the class of all finite time grids in . Then
inf min Ɛ6C0 4Î57 = Ɛ6C 4X ∗ 570

0 ∈T Î∈X40 1 X0 5
3. Examples.
3.1. Constructing decay kernels by transformation. In this section we will now look at some transforma-
tions of decay kernels. The first of these results concerns decay kernels of the simple form G4t5 = g4t5L where
g2 601 5 → is a function and L ∈ K×K is a fixed matrix.
Proposition 5. For L ∈ + 4K5 and a positive definite function g2 → , the decay kernel G4t5 2= g4t5L is
positive definite. If, moreover, g is a strictly positive definite function and L is a strictly positive definite matrix,
then G is also strictly positive.
The simple decay kernels from the preceding proposition provide a class of examples to which also the next
result applies. In particular, by choosing in the subsequent Proposition 6 the decay kernel as G4t5 2= g4t5Id for
g2 → positive definite and Id ∈ K×K denoting the identity matrix, one sees that the optimal strategies for
decay kernels of the form g4t5L with L ∈ + 4K5 do not depend on the cross-asset impact g4t5Lij for i 6= j.
Hence, cross-asset impact will only become relevant when the components of G decay at varying rates.
Proposition 6. Let G be a decay kernel and define GL 4t5 2= LG4t5 for some L ∈ K×K . When both G and
GL are positive definite, then every optimal strategy in Xdet 41 X0 5 for G is also an optimal strategy for GL .
The main message obtained from combining Propositions 5 and 6 is the following: if the price impact between
all pairs of assets decays at the same rate, then cross-asset impact can be ignored and one can simply consider
each asset individually.
We show next that also congruence transforms preserve positive definiteness. This result extends Proposi-
tion 3(e).
Proposition 7. If G is a (strictly) positive definite decay kernel and L and an invertible K × K matrix,
then GL 2= L> G4t5L is (strictly) positive definite. If, moreover, Î is an optimal strategy for G in Xdet 41 LX0 5,
then ÎL 2= 4L−1 1 1 : : : 1 L−1 5 is an optimal strategy for GL in Xdet 41 X0 5.
Example 1 (Permanent Impact). Let G4t5 = G0 , where G0 is any fixed matrix in K×K . For any time grid
, X0 ∈ K , and Î ∈ X41 X0 5 we then have C 4Î5 = X0> G0 X0 . Hence, G is positive definite as soon as G0 is
non-negative. By taking G0 such that X0> G0 X0 ≥ 0 for some nonzero X0 and Y0> G0 Y0 < 0 for some other Y0 one
gets an example illustrating that it is not possible to fix X0 in part (a) of Proposition 1.
3.2. Exponential decay kernels. In this section we will discuss decay kernels with an exponential decay
of price impact. For K = 1 exponential decay was introduced in Obizhaeva and Wang [28] and further studied,
e.g., in Alfonsi et al. [1] and Predoiu et al. [30]. The next example extends the results from Obizhaeva and
Wang [28] and Alfonsi et al. [1] to a multivariate setting in which the decay kernel is defined in terms of matrix
exponentials. The remaining results of this section are stated in a more general but two-dimensional context.
The main message of these examples is that, on the one hand, it is easy to construct decay kernels with all
desirable properties via matrix functions. But, on the other hand, it is typically not easy to establish properties
such as positive definiteness for decay kernels that are defined coordinate-wise.
Example 2 (Matrix Exponentials). For an orthogonal matrix O, 1 1 : : : 1 K ≥ 0, and B =
O > diag41 1 : : : 1 K 5O ∈ + 4K5, the decay kernel G4t5 = exp4−tB5 is of the form (13) with g4t5 = e−t . It fol-
lows that G is non-negative, nonincreasing, and convex. In particular, G is positive definite. When the matrix
B is strictly positive definite, as we will assume from now on, the decay kernel G is even strictly positive
definite. We now compute the optimal strategy Î = 41 1 : : : 1 N 5 for an initial portfolio X0 ∈ K and time grid
= 8t1 1 : : : 1 tN 9. To this end, we will use part (e) of Proposition 3. Let Çi 2= 41i 1 : : : 1 Ni 5 be the optimal
strategy for the initial position yi and for the one-dimensional decay kernel gi 4t5 = e−ti . Let
−yi
ain 2= e−4tn −tn−1 5i and i 2= 0
2/41 + ai2 5 + Nn=3 441 − ain 5/41 + ain 55
P
Theorem 3.1 in Alfonsi et al. [1] implies that the optimal strategy Çi = 41i 1 : : : 1 Ni 5 in Xget 41 y i 5 is given by
ain+1

i i i 1 i
1 = i
1 n = − i
i for n = 21 : : : 1 N − 11 and Ni = 0
1 + a2 i
1 + an 1 + an+1 1 + aiN
Via part (e) of Proposition 3, we can now compute the optimal strategy Î. Consider first the optimal strategy Ç for
the decay kernel D4t5 2= diag4exp4−1 t51 : : : 1 exp4−K t55 and initial position OX0 . Then Ç = 4Ç1 1 : : : 1 ÇK 5>
for y i 2= 4OX0 5i . When defining Qn 2= D4tn − tn−1 5 and
N −1
˜ = − 24Id + Q2 5−1 + 4Id − Qn 54Id + Qn 5−1
X
OX0 1
n=3
Ç = 41 1 : : : 1 N 5 can be conveniently expressed as follows:

˜
1 = 41 + Q2 5−1 1
n = 4Id + Qn 5−1 ˜ − Qn+1 4Id + Qn+1 5−1 ˜ for n = 21 : : : 1 N − 1,
˜
N = 4Id + QN 5−1 0
By part (e) of Proposition 3 the optimal strategy Î for G and X0 is now given by Î = O T . To remove O from
these expressions, define An = e−4tn −tn−1 5B = O > Qn O and
N −1
−1
X −1
2= − 24Id + A2 5 + 4Id − Ai 54Id + Ai 5 X0 0
i=3
˜ we find that the components of the optimal

By observing that 4Id + An 5−1 = O > 4Id + Qn 5−1 O and = O > ,
strategy Î are
1 = 4Id + A2 5−1 1
n = 4Id + An 5−1 − An+1 4Id + An+1 5−1 for n = 21 : : : 1 N − 1,
N = 4Id + AN 5−1 0
Let us finally consider the situation of an equidistant time grid, ti = 4i − 15/4N − 15. In this case, all matrices
Ai are equal to a single matrix A. Our formula for then becomes
= −4Id + A54N Id − 4N − 25A5−1 X0 0
The formula for the optimal strategy thus simplifies to
1 = −4N Id − 4N − 25A5−1 X0 1
i = 4Id − A51 for i = 21 : : : 1 N − 11
N = 1 0
It is not difficult to extend this result to the setting of §2.5 by arguing as in Gatheral et al. [18, Example 2.12].
The details are left to the reader.
When g2 → is an analytic function, the definition of g4B5 is also possible for nonsymmetric matrices by
letting

ak B k 1
X
g4B5 2=
k=0
P k
where g4x5 = k=0 ak x is the power series development of g. In the following example we analyze the prop-
erties of the decay kernel G4t5 2= exp4−tB5 for the particular nonsymmetric but strictly positive 2 × 2-matrix
B = 0b b1 with b > 0. We will see that G may or may not be positive definite, according to the particular choice
of b. Thus, our general results obtained for decay kernels defined as matrix functions of symmetric matrices do
not carry over to the nonsymmetric case.
Example 3 (Nonsymmetric Matrix Exponential Decay). Let B = 0b b1 , where b > 0 and consider the

following decay kernel !

−tB
exp4−tb5 −t exp4−tb5
G4t5 = e = 0
0 exp4−tb5
Applying Lemmas 5 and 6, we easily see that G is not symmetric, not non-negative, not nonincreasing, and
not convex. But G is positive definite
R if and only if b ≥ 1/2. To see this, we observe by calculating the inverse
Fourier transform that G̃4t5 = eitz M4z5 dz with
b
 −1 
1  b + z2
2 24b + iz52 
M4z5 =  0
 −1 b 
24b − iz52 b 2 + z2
From Theorem 1 and Lemma 5, G is positive definite if and only if for all z ∈
2
1 1 b
≤ 1
4 4b 2 + z2 52 b 2 + z2
which is in turn equivalent to 1/2 ≤ b.
For the following results we no longer require that the decay kernel is given in the particular form of a matrix
function.
Proposition 8. Let !
a11 exp4−b11 t5 a12 exp4−b12 t5
G4t5 =
a21 exp4−b21 t5 a22 exp4−b22 t5
with a11 1 a12 1 a21 1 a22 1 b11 1 b12 1 b21 1 b22 > 0.
(a) G is non-negative if and only if min8b12 1 b21 9 ≥ 21 4b11 + b22 5 and 41 4a12 + a21 52 ≤ a11 a22 .
(b) G is nonincreasing if and only if min8b12 1 b21 9 ≥ 21 4b11 + b22 5 and 41 4a12 b12 + a21 b21 52 ≤ a11 b11 a22 b22 .
(c) G is convex if and only if min8b12 1 b21 9 ≥ 21 4b11 + b22 5 and 41 4a12 b12
2 2 2
+ a21 b21 2
5 ≤ a11 b11 2
a22 b22 .
(d) Let G be nonincreasing and a12 = a21 . Then G is positive definite.
(e) G is commuting if and only if either b11 = b12 = b21 = b22 , or b11 = b22 and b12 = b21 and a11 = a22 .
For the following simpler and symmetric decay kernel, the results follow immediately from the preceding
proposition. See Figure 2 for an illustration of a corresponding optimal strategy.
Corollary 2. Let 1 1 ˜ > 0 and
!
exp4−t5 exp4−t5
˜
G4t5 = 0
exp4−t5
˜ exp4−t5
(a) G is non-negative if and only if /˜ ≤ 1 and ≤ 1.

(b) G is nonincreasing if and only if ≤ /˜ ≤ 1. In this case, it is also non-negative.
(c) G is convex if and only if ≤ 2 /˜ 2 ≤ 1.
(d) If G is nonincreasing, G is positive definite.
(e) G is commuting.
The following proposition shows that we cannot drop the assumption of symmetry in Theorem 2 in general.
Proposition 9. Let 1
!
exp4−t ∧ 15 8
exp4−24t ∧ 155
G4t5 = 1
0
8
exp4−34t ∧ 155 exp4−t ∧ 15
G is continuous, convex, nonincreasing, and non-negative, but not positive definite.
3.3. Linear decay. In this section, we analyze linear decay of price impact for K = 2 assets.
Proposition 10. Let !
4a11 − b11 t5+ 4a12 − b12 t5+
G4t5 =
4a21 − b21 t5+ 4a22 − b22 t5+
with a11 1 a12 1 a21 1 a22 1 b11 1 b12 1 b21 1 b22 > 0.
(a) G is non-negative if and only if max8a12 /b12 1 a21 /b21 9 ≤ min8a11 /b11 1 a22 /b22 9 and 41 4a12 + a21 52 ≤ a11 a22 .
(b) G is nonincreasing if and only if max8a12 /b12 1 a21 /b21 9 ≤ min8a11 /b11 1 a22 /b22 9 and 41 4b12 +b21 52 ≤ b11 b22 .
(c) Assume that max8a12 /b12 1 a21 /b21 9 ≤ min8a11 /b11 1 a22 /b22 9 and a12 = a21 . Then, G is positive definite if
2
and only if G is symmetric (i.e., a12 = a21 and b12 = b21 ), a11 /b11 = a12 /b12 = a22 /b22 and b12 ≤ b11 b22 . In this
case, we set = a11 /b11 and have !
+
b11 b12
G4t5 = 4 − t5 1
b12 b22
and G is also nonincreasing, convex, and commuting.
4. Conclusion. Our goal in this paper was to analyze a linear market impact model with transient price
impact for K different risky assets. We were in particular interested in the question which properties a decay
kernel should satisfy so that the corresponding market impact model has certain desirable features and properties.
Let us summarize some of the main messages for the practical application of transient price impact models that
can be drawn from our results.
(a) To exclude price manipulation in the sense of Huberman and Stanzl [23] and to guarantee the existence
of optimal strategies, decay kernels should be positive definite in the sense of Definition 3 (Propositions 1 and 2
and Lemma 3).
(b) Requiring only positive definiteness is typically not sufficient to guarantee that optimal strategies are
well-behaved (Figure 1). In particular, the nonparametric estimation of decay kernels can be problematic.
(c) Assuming that the decay kernel is symmetric, non-negative, nonincreasing, convex, and commuting guar-
antees that optimal strategies have many desirable properties and can easily be computed (Propositions 3 and 4).
The additional assumption that t 7→ x> G4t5x is nonconstant for all x ∈ K guarantees that optimal strategies
are unique (Theorem 2 and Proposition 2). In this setting, one can also optimize jointly over time grids and
strategies and pass to a continuous-time limit.
(d) Matrix functions (13) provide a convenient method for constructing decay kernels satisfying the properties
from (c). Optimal strategies for matrix exponential decay can be computed in closed form (Example 2).
(e) If the price impact between all asset pairs decay at a uniform rate, then cross-asset impact can be ignored
and one can consider each asset individually (Propositions 5 and 6).
5. Proofs.
Proof of Lemma 1. Using the continuity of G and the right-continuity of S 0 , we have
N
1X > Î Î
− Ɛ6R4Î57 = Ɛ 4St + + Stk 5
2 k=1 k k
N N N k−1
X > 0 1X > X X >
=Ɛ k Stk + Ɛ k G405k + k G4tk − t` 5` 0
k=1 2 k=1 k=1 `=1
0
PN
From the martingale property of S and the requirement that k=1 k = −X0 we obtain that
N N
X > 0 X > 0
Ɛ k Stk = Ɛ k ST = −X0> S00 0
k=1 k=1
Furthermore,
N N k−1
1X X X >
k> G405k + k G4tk − t` 5`
2 k=1 k=1 `=1
N N k−1 N k−1
1X 1X X > 1X X >
= k> G̃405k + k G̃4tk − t` 5` + G̃4t` − tk 5k = C 4Î50
2 k=1 2 k=1 `=1 2 k=1 `=1 `
This proves (4).

Proof of Lemma 2. Suppose that a minimizer Ç ∈ X41 X0 5 of Ɛ6C 4Ç57 exists but that, by way of contra-
diction, there is no deterministic minimizer of C 4·5. Then there can be no Î ∈ Xdet 41 X0 5 such that C 4Î5 ≤
Ɛ6C 4Ç57. Since Ç45 ∈ Xdet 41 X0 5 for -a.e. , we must thus have C 4Ç455 > Ɛ6C 4Ç57 for -a.e. ∈ ì.
But this is a contradiction. The proofs of the remaining assertions are also obvious and left to the reader.
Proof of Proposition 1. The equivalence of conditions (a) and (b) follows from Lemma 2.
To prove the equivalence of (b) and (c), it is sufficient to observe that C 4Î5 is a quadratic form on ⊗ K ,
and it is well known that a quadratic form is convex if and only if it is non-negative.
We next prove the equivalence of (b) and (d). Clearly, (d) immediately implies (b) using the representation (5)
of C 4 · 5 and comparing it with (8) with zi ∈ K . For the proof of the converse implication, we fix t1 1 : : : 1 tN ∈ .
Clearly we can assume without loss of generality that = 8t1 1 : : : 1 tN 9 is a time grid in the sense that 0 = t1 <
t2 < · · · < tN . An N -tuple Æ 2= 41 1 : : : 1 N 5 with i ∈ K can be regarded as an element in the tensor product
N ⊗ K . Let us thus define the linear map L2 N ⊗ K → N ⊗ K by
N N N
X X X
LÆ = G̃4t1 − tj 5j 1 G̃4t2 − tj 5j 1 : : : 1 G̃4tN − tj 5j 0 (16)
j=1 j=1 j=1
We claim that L is Hermitian. Indeed, for Ç1 Æ ∈ N ⊗ K , the inner product in N ⊗ K between Ç and LÆ is
given by
N
X N
X N
X
Ç1 LÆ = i∗ G̃4ti − tj 5j = j∗ G̃4ti − tj 5∗ i = j∗ G̃4tj − ti 5i = Æ1 LÇ1
i1 j=1 i1 j=1 i1 j=1
where we have used the fact that G̃4ti − tj 5∗ = G̃4ti − tj 5> = G̃4tj − ti 5. It follows that the restriction of L to
N ⊗ K is symmetric and, due to condition (b), satisfies 0 ≤ C 4Î5 = Î1 LÎ for all Î ∈ N ⊗ K . By the
symmetry of L and since L has only real entries, it follows that Æ1 LÆ ≥ 0 for all Æ ∈ N ⊗ K , which is the
same as (8) and hence yields (d). The remaining assertions are obvious.
Proof of Proposition 2. We first show the existence of optimal strategies when G̃ is positive definite. We
will use the notation introduced in the proof of Proposition 1. For X0 ∈ K and with N = fixed, the
minimization of C 4Î5 over Î ∈ Xdet 41 X0 5 is equivalent to the minimization of the symmetric and positive
semidefinite quadratic form N ⊗ K 3 Î 7→ Î1 LÎ
P under the equality constraint AÎ = X0 , where L is as in (16)
and A2 N ⊗ K → K is the linear map AÎ 2= Nk=1 k . For fixed Ç ∈ Xdet 41 X0 5, every other Î ∈ Xdet 41 X0 5
can be written as Î = Ç + Î0 for some Î0 ∈ Xdet 41 05. Then, due to the symmetry of L,
Î1 LÎ = Ç1 LÇ + 2LÇ1 Î0 + Î0 1 LÎ0 1

and our problem is now equivalent to the unconstraint minimization of the right-hand expression over Î0 ∈
Xdet 41 05. Clearly, LÎ0 = 0 implies that also 2LÇ1 Î0 = 2Ç1 LÎ0 = 0. Therefore the existence of minimizers
follows from Section 2.4.2 in Boot [8].
The uniqueness of optimal strategies for strictly positive definite G̃ follows immediately from the strict con-
vexity of Î 7→ C 4Î5 (see Proposition 1). The characterization of optimal strategies through Lagrange multipliers
as in (9) is standard.
Proof of Lemma 3. (a) That H 405 is non-negative definite follows by taking N = 1 in (8). To show
H 4−t5 = H 4t5∗ for any given t ∈ we take N = 2 in (8) and let t1 = 0 and t2 = t. It follows from the preceding
assertion that z∗1 H 4−t5z2 + z∗2 H 4t5z1 must be a real number for all z1 1 z2 ∈ K . Taking z1 = c1 ei and z2 = c2 ej
with ck ∈ and e` denoting the `th unit vector in K yields that c̄1 c2 Hij 4−t5 + c1 c̄2 Hji 4t5 ∈ , where c̄ denotes
the complex conjugate of c ∈ . Choosing c1 = c2 = 1 gives Im4Hij 4−t55 = − Im4Hji 4t55 and c1 = 11 c2 = i
yields Re4Hij 4−t55 = Re4Hji 4t55.
(b) For t1 1 : : : 1 tN ∈ , we define t̃i = tN − tN +1−i and get from part (a) that for Æ ∈ N ⊗ K
N
X N
X
0≤ N∗ +1−i H 4t̃i − t̃j 5N +1−j = N∗ +1−i H 4−4tN +1−i − tN +1−j 55N +1−j
i1 j=1 i1 j=1
N
X
= i∗ H 4ti − tj 5∗ j 0
i1 j=1
Proof of Corollary 1. For the proof of implication (c)⇒(b), we note first that the matrix M4d5 is
symmetric, as M is non-negative definite and K×K -valued. This implies that the Rmatrix H 4t5 is also sym-
metric
R for all t. Next, the symmetry of M on implies that the imaginary part of eit Mk` 4d5 is equal to
0
4sin4t5 + sin4−t55Mk` 4d5 = 0. Therefore, H takes values in K×K and, in turn, in 4K5. We next define
a finite + -valued measure through 4A5 2= x> M4A5x for A ∈ B45. Then the function t 7→ x> H 4t5x is the
Fourier transform of and hence a positive definite function by Bochner’s theorem.
To prove (b) ⇒ (a), we √ will establish condition (b) of Theorem 1. To this end, write z ∈ K as z = x + iw,
K
where x1 w ∈ and i = −1. Then z∗ H4t5z = x> H 4t5x + w > H 4t5w due to the symmetry of H 4t5. Hence
t 7→ z∗ H 4t5z is the sum of two real-valued positive definite functions and therefore positive definite.
To prove (a) ⇒ (c), note that each component Hk` of H is equal to the Fourier transform of the complex
measure Mk` . Since Mk` is uniquely determined through Hk` and since Hk` = H`k we must have that Mk` = M`k .
But a symmetric matrix can be non-negative definite, and hence Hermitian, only if it is real. Therefore we
must have M4A5 ∈ + 4K5 for all A ∈ B45. Finally, the fact that the symmetric positive definite matrix-valued
function Hk` takes only real values implies via Lemma 3(a) that H 4−t5 = H 4t5. Therefore, H is equal to the
Fourier transform of the measure N 4A5 2= 21 4M4A5 + M4−A55, A ∈ B45. But, since M is uniquely determined
by H according to Theorem 1, we get that N = M, and so M must be symmetric on .
Proof of Lemma 4. We assume by way of contradiction that there exist x ∈ K , t ∗ > 0 and > 0 such that
g 4t5 2= x> G4t5x satisfies g x 4t ∗ 5 = −. We are going to show that the function g x is not positive definite. Set
x
tk = k · t ∗ and xk =
P1n for k ∈ .x Since tk − tl ≥ t ∗ for k 6= l and g x is nonincreasing, we have g x 4tk − tl 5 ≤ −
x
for k 6= l. Thus, k1 l=1 xk xl g 4tk − tl 5 ≤ ng 405 − 4n2 − n5. If n is large enough, the latter expression is
negative. Thus, g x is not positive definite, and so G can not be positive definite.
We now start preparing the proof of Theorem 2 and give a representation of a convex, nonincreasing, non-
negative, and symmetric function G2 601 5 → K×K . To this end, let us first observe that, for such G, the
limit G45 2= limt↑ G4t5 is well defined in the set of non-negative definite matrices. Indeed, for any x ∈ K ,
g x 4t5 = x> G4t5x is a convex, nonincreasing, non-negative function and thus converges to a limit that we denote
by g x 45. Let ei denote the ith unit vector. By polarization, we have Gij 4t5 = 41 4g ei +ej 4t5 − g ei −ej 4t55, and this
expression converges to Gij 45 = 41 4g ei +ej 45 − g ei −ej 455. In particular, we have g x 45 = x> G45x for any
x ∈ K .
Proposition 11. Let G2 601 5 → K×K be convex, nonincreasing, non-negative, symmetric, and continuous.
There exists a non-negative Radon measure on 401 5 and a measurable function å2 401 5 → + 4K5
such that Z
G4t5 = G45 + 4r − t5+ å4r5 4dr50 (17)
401 5
Furthermore, G is the Fourier transform of the non-negative definite matrix-valued measure
M4d5 = G45 0 4d5 + ê45 d1
where ê2 → + 4K5 is the continuous function given by

1Z 1 − cos xy
ê4x5 = å4y5 4dy50
401 5 x2
Proof. By Lemmas 4.1 and 4.2 in Gatheral et al. [18], we find that for every x ∈ K there is a Radon
measure x on 401 5 such that
Z
g x 4t5 = g x 45 + 4r − t5+ x 4dr51 t > 00 (18)
401 5
x
Moreover, g 4t5 is the Fourier transform of the following non-negative Radon measure on
x 4dt5 = x> G45x 0 4dt5 + x 4t5 dt1 (19)
where 0 is the Dirac measure concentrated in 0 and

1Z 1 − cos ty
x 4t5 = x 4dy50 (20)
401 5 t2
P
We consider the finite set Z 2= 8ei ± ej i1 j = 11 : : : 1 K9 and define = z∈Z z . Then each z with z ∈ Z is
absolutely continuous with respect to and has the Radon-Nikodym derivative z = dz /d. We set
1
åij 4r5 2= 4ei +ej 4r5 − ei −ej 4r551 r > 00 (21)
4
Clearly, åij 4r5 = åji 4r5, and it remains to prove that å is -a.s. non-negative definite. Let x ∈ K . Since
g x 4t5 = 41 Ki1 j=1 xi xj 4g ei +ej 4t5 − g ei −ej 4t55, we necessarily have from (18):
P
Z Z K
g x 4t5 − g x 45 =
X
4r − t5+ x 4dr5 = 4r − t5+ xi xj åij 4r54dr51 t > 00
401 5 401 5 i1 j=1
R
Writing 4r − t5+ as 0
8t≤s<r9 ds, integrating by parts, and taking derivatives with respect to t gives
Z K
X
0 ≤ x 44t1 +55 = xi xj åij 4r5 4dr5
4t1 +5 i1 j=1
for any t ≥ 0 and so Ki1 j=1 xi xj åij 4r5 ≥ 0 for r 6∈ Nx , where Nx is such that 4Nx 5 = 0. We define N =
P
S PK K
x∈QK Nx . Then 4N 5 = 0 and, by continuity,
PK i1 j=1 xi xj åij 4r5 ≥ 0 for all x ∈ and r 6∈ N .
Now that we have shown x 4dr5 = i1 j=1 xi xj åij 4r54dr5, we get (17) from (18). Next, we obtain
from (20) that
K
1Z 1 − cos ty X
x 4t5 = xi xj åij 4y5 4dy50
401 5 t2 i1 j=1
Again, we define by polarization ê4t5ij = 41 4ei +ej 4t5 − ei −ej 4t55. We then have
1Z 1 − cos ty
ê4t5 = å4y5 4dy51
401 5 t2
and x 4t5 = x> ê4t5x. Together with (19), this gives the claim.
Proof of Theorem 2. From Theorem 1 and the fact that g x is positive definite for each x, we already know
that G is positive definite. Note also that G cannot be strictly positive definite if there P exists some nonzero
x ∈ K such that t 7→ x> G4t5x is constant, for then the choice z1 = x and z2 = −x gives 2i1 j=1 z∗i G4ti − tj 5zj = 0
for all t1 1 t2 ∈ .
It thus remains to show that G strictly definite positive if > G4t5 is nonconstant for any ∈ K . We argue
first that, in proving this assertion, we can assume without loss of generality that G is continuous. To this end,
consider again the functions g x 4t5 2= x> G4t5x for x ∈ K . As these functions are convex and nonincreasing, they
are continuous on 401 5 and admit right-hand limits, g x 40+5 2= limt↓0 g x 4t5 ≤ g x 405. Using polarization as in the
paragraph preceding Proposition 11, we thus conclude that G is continuous on 401 5, admits a right-hand limit
G40+5, and that ãG405 2= G405 − G40+5 is non-negative definite. On the other hand, the continuous matrix-
valued function Gcont 4t5 2= G4t+5 also satisfies our assumptions and so will be strictly positive definite when the
assertion has been established for continuous matrix-valued functions. But then G4t5 = Gcont 4t5 + 809 4t5ãG405
will also be strictly positive definite, because ãG405 is non-negative definite.
Now, let M, ê, å, and be as in Proposition 11. It follows from this proposition that for Æ =
41 1 2 1 : : : 1 N 5 ∈ N ⊗ K and t1 1 t2 1 : : : 1 tN ∈
N N Z
i4tk −t` 5
X X
k∗ G̃4tk − t` 5` = k∗ e M4d5 `
k1 `=1 k1 `=1
N ∗ N Z N ∗ N
X X X −it X −it
= k G45 k + e k k ê45 e k k d
k=1 k=1 k=1 k=1
Z
= v405∗ G45v405 + v45∗ ê45v45 d1
where v45 2= Nk=1 e−itk k . We are now going to show that v45∗ ê45v45 d is strictly positive unless
P R
Æ = 0. To this end, we note first that the components of the vector field v4·5 are holomorphic functions of ∈ .
When Æ 6= 0, at least one of these components is nonconstant and hence vanishes only for at most countably
many ∈ . It follows that v45 6= 0 for all but countably many ∈ . Moreover, we are going to argue next
that the matrix ê45 is strictly positive definite for all but countably many ∈ . Thus, v45∗ ê45v45 > 0
for Lebesgue-almost every ∈ , and it will follow that k∗ G̃4tk − t` 5` > 0.
P
So let us show now that the matrix ê45 is strictly positive definite for all but countably many ∈ . To this
end, we first note that for z ∈ K
Z
g z 4t5 = z∗ G45z + 4r − t5+ z∗ å4r5z 4dr50
401 5
Since the matrix å4r5 is non-negative definite for all r, the fact that g z is nonconstant for z 6= 0 implies that
Z
z∗ å4r5z 4dr5 > 0 for z 6= 0. (22)
401 5
Now let D be the set of all y > 0 such that 48y95 > 0, and let
d 4E5 2= 4D ∩ E5 and c 4E5 2= 4Dc ∩ E5
be the discrete and continuous parts of , respectively. Clearly,

[
N 2= 8x ∈ cos xy = 1 for some y ∈ D9 = 8x ∈ cos xy = 19
y∈D
is at most countable. Moreover, the set 8y > 0 cos xy = 19 is a c -nullset for all x 6= 0. It thus follows that the
measure 441 − cos xy5/x2 5 4dy5 is equivalent to for all x y N ∪ 809. Therefore (22) implies that
1Z 1 − cos xy ∗
z∗ ê4x5z = z å4r5z 4dy5 > 0
401 5 x2
for all z 6= 0 as long as x y N ∪ 809. This concludes the proof.
Proof of Proposition 3. We first prove (11). To this end, give a constructive proof for the existence of O.
For t ∈ , we write K = Et 1 4t5 ⊕ · · · ⊕ Et `4t5 4t5 for the orthogonal direct sum of the eigenspaces of G4t5
corresponding to the distinct eigenvalues 1 4t51 : : : 1 `4t5 4t5 of G4t5. It follows from the commuting property
that the eigenspaces of G4t5 are stable under the map G4s5, because i 4t5G4s5v = G4t5G4s5v if v ∈ Ei 4t5 .
P`4t 5 t
Let t1 ≥ 0 and define D1 = i=11 4dim4E1i 4t1 5 5 − 15+ ≤ K − 1. If for any s ≥ 0 and 1 ≤ i ≤ `4t1 5 there is i 4s5
t1 d
such that G4s5v = i 4s5v for any v ∈ Ei 4t1 5 , we are done by considering an orthonormal basis 4vi1 1 : : : 1 vi i 5 of
t d d
each eigenspaces E1i 4t1 5 and by setting O = 4v11 1 : : : 1 v1 1 1 : : : 1 v`1 1 : : : 1 v` ` 5 for ` 2= `4t1 5. This is necessarily the
case if D1 = 0. Otherwise, there is t2 ≥ 0 such that, for at least one i ∈ 811 : : : 1 `4t1 59, the decomposition
t t t t t
E1i 4t1 5 = 4E21 4t2 5 ∩ E1i 4t1 5 5 ⊕ · · · ⊕ 4E2`4t 4t2 5 ∩ E1i 4t1 5 5
25
is such that
`4t2 5
t X t t
dim4E1i 4t1 5 5 − 1 > 4dim4E2k 4t2 5 ∩ E1i 4t1 5 5 − 15+ 0
k=0
We write
t t
K =
M
E1i 4t1 5 ∩ E2i 4t2 5
1 2
1≤i1 ≤`4t1 51 1≤i2 ≤`4t2 5
P`4t1 5 P`4t2 5 t1 t2 +
and have D2 = i1 =1 i2 =1 4dim4Ei 4t1 5 ∩ Ei 4t2 5 5 − 15 < D1 . Once again, we are done if there is for any s ≥ 0,
1 2
t t
1 ≤ i1 ≤ `4t1 5, and 1 ≤ i2 ≤ `4t2 5, i1 1 i2 4s5 such that G4s5v = i1 1 i2 4s5v for any v ∈ E1i 4t1 5 ∩ E2i 4t2 5 . This is the
1 2
case when D2 = 0. Otherwise, there is t3 such that
`4t1 5 `4t2 5 `4t3 5
XXX t t t +
D3 = 4dim4E1i 4t1 5 ∩ E2i 4t2 5 ∩ E3i 4t3 5 5 − 15 < D2
1 2 3
i1 =1 i2 =1 i3 =1
and we repeat this procedure at most K times to get (11).

We now prove properties (a)–(e). Let vi be the ith column PK of O. Then vi isPthe eigenvector of Gi 4t5 for the
K
eigenvalue gi 4t5. A given x ∈ K can be written as x = v
i=1 i i . Then Ox = i=1 i ei , where ei is the ith unit

vector. It follows from (11) that g x 4t5 = Ki=1 2i gi 4t5. From here, the assertions (a)–(d) are obvious. Part (e)
P
follows from Proposition 7.
Proof of Proposition 4. Let O and g1 1 : : : 1 gK be as in Proposition 3. We let v1 1 : : : 1 vK be the columns
i
of O. By Theorem 1 from Alfonsi et al. [4] there is a one-dimensional optimal strategy Çi = 41i 1 : : : 1 5∈
Xdet 41 15 for the one-dimensional, nonincreasing, non-negative, and convex decay kernel gi , and Çi has only
non-negative components. By part (e) of P Proposition 3, Î4i5 2= Çi vi is an optimal strategy for G in Xdet 41 vi 5
that satisfies condition (a). When X0 = Ki=1 i vi ∈ K is given, the strategy with components i Çi vi is an
optimal strategy in Xdet 41 X0 5 by Proposition 3(e).
Proof of Theorem 3. The proof of part (a) can be performed along the lines of the proof of Theorem 2.20
from Gatheral et al. [18] by noting that Proposition 4(b) yields an upper bound on the number of shares traded
by an optimal strategy Î ∈ X41 X0 5 uniformly over finite time grids ⊂ :
K K
nj ≤ vij 0
X X X
i
1≤n≤1 1≤j≤K i=1 j=1
The details are left to the reader.

As for part (b), the argument from the proof of Theorem 2.20 in Gatheral et al. [18]Syields in particular, that
Ɛ6Cn 4Î4n5 57 decreases to Ɛ6C 4X ∗ 57 if 1 ⊂ 2 ⊂ · · · are finite time grids such that n n is dense in and
Î4n5 is an optimal strategy in X4n 1 X0 5. This proves (b).
Proof of Proposition 5. Let A ∈ K×K be a symmetric square root of the non-negative definite matrix L
so that L = A2 = A> A. For t1 1 : : : 1 tN ∈ and 1 1 : : : 1 N ∈ K let k 2= Ak . It follows that
N N N N
K X
¯ ki `i g4tk − t` 51
X X X X
k∗ G̃4tk − t` 5` = k∗ A> g4tk − t` 5A` = k∗ g4tk − t` 5` =
k1 `=1 k1 `=1 k1 `=1 i=1 k1 `=1
which is non-negative since the function g is positive definite. Now let g and L be even strictly positive definite.
Then the matrix A is nonsingular and so we have 1 = · · · = N = 0 if and only if 1 = · · · = N = 0. It follows
that in all other cases the right-hand side above is strictly positive.
Proof of Proposition 6. Let Î ∈ Xdet 41 X0 5 be an optimal strategy for the decay kernel G. By Proposi-
tion 2 there exists a Lagrange multiplier ∈ K such that
N
X
G̃4tk − t` 5` = for k = 11 : : : 1 .
`=1
By multiplying both sides of this equation with L we obtain

N
X
G̃L 4tk − t` 5` = L for k = 11 : : : 1 ,
`=1
which, again by Proposition 2, implies that Î is also optimal for GL .

Proof of Proposition 7. Since L is invertible, the transformation
Î 7−→ ÎL 2= 4L−1 1 1 : : : 1 L−1 5
is a one-to-one map from Xdet 41 LX0 5 to Xdet 41 X0 5. We also have
0 ≤ k> G̃4tk − t` 5` = 4kL 5> G̃L 4tk − t` 5`L

X X
k1 ` k1 `
for all and Î. Minimizing the two sums over the respective classes of strategies yields the result.
To study the examples for K = 2 assets, we will frequently use the following simple lemma.
Lemma 5. For a1 d ≥ 0 and b1 c ∈ , the matrix M 2= ca db is non-negative if and only if 41 4b + c52 ≤ ad.

When b ∈ , the Hermitian matrix N 2= ba db is non-negative definite if and only if b2 ≤ ad.
Proof. The matrix M is non-negative if and only if its symmetrization, M̃ 2= 21 4M + M > 5, is positive
definite. Since a symmetric matrix is non-negative definite if and only if all its leading principle minors are
non-negative and since det M̃ = ad − 41 4b + c52 , the result follows. In the Hermitian case, the same condition on
the minors holds.
Rt
Lemma 6. Let G2 601 5 → K×K and assume that G4t5 = 0 å4s5 ds for t ≥ 0. Then, G is nonincreasing if
and only if −å4s5 is non-negative for a.e. s. If in addition å is piecewise continuous, then G is convex if and
only if −å is nonincreasing.
Rt
Proof. The function G is nonincreasing if and only if for any ∈ K , 0 > å4s5ds is nonincreasing. This
gives > å4s5 ≥ 0 for s 6∈ N , where N is a set with zero Lebesgue measure. We define N = ∈K N and have
S
> K
by continuity å4s5 ≥ 0 for any s 6∈ N , ∈ . The converse implication as well as the other equivalence are
obvious.
Proof of Proposition 8. (a): By Lemma 5, G is non-negative if and only if for every t ≥ 0
1
4
4a12 exp4−b12 t5 + a21 exp4−b21 t552 ≤ a11 exp4−b11 t5a22 exp4−b22 t50
That is, if and only if

1
4
4a212 exp4−2b12 t5 + 2a12 a21 exp4−4b12 + b21 5t5 + a221 exp4−2b21 t55 ≤ a11 a22 exp4−4b11 + b22 5t50
If G is non-negative, taking t = 0 shows 41 4a12 + a21 52 ≤ a11 a22 , whereas sending t → shows min8b12 1 b21 9 ≥
1
2
4b11 + b22 5. Conversely, if these inequalities hold, G is non-negative.
(b): G is continuously differentiable. By Lemma 6, G is hence nonincreasing if and only if for every t ≥ 0
!
a11 b11 exp4−b11 t5 a 12 b12 exp4−b12 t5
−G0 4t5 =
a21 b21 exp4−b21 t5 a22 b22 exp4−b22 t5
is non-negative. Analogously to (a), the result follows.

(c): Analogously to (b), by Lemma 6 G is convex if and only if for every ≥ 0 its second derivative
2 2
!
00
a11 b11 exp4−b11 t5 a12 b12 exp4−b12 t5
G 4t5 =
2 2
a21 b21 exp4−b21 t5 a22 b22 exp4−b22 t5
is non-negative. The result follows analogously to (a). R

(d): The assumption a12 = a21 gives the continuity of G̃. We have that G̃4t5 = eit M4d5, where M4d5 =
41/4255å45d with the Hermitian matrix
 a11 b11 a12 a12 
2 2
+
 b11 + 2 b21 − i b12 + i 
å45 =  0
 a12 a12 a22 b22 
+ 2 2
b12 − i b21 + i b22 + 2
From Theorem 1, G is positive definite if and only if the matrix å45 is non-negative for almost all ∈ .
According to Lemma 5, this is equivalent to
a212 4b12 + b21 52 a b a b
2 2 2 2
≤ 4 2 11 11 2 2 22 22 2 0
4b12 + 54b21 + 5 b11 + b22 +
This condition is in turn equivalent to
a212 4b12 + b21 52 4b11

2
+ 2 54b22
2
+ 2 5 ≤ 4a11 b11 a22 b22 4b12
2
+ 2 54b21
2
+ 2 50
Comparing the coefficients for 0 , 2 , and 4 , we see that it is sufficient to have
a212 4b12 + b21 52 b11 b22 ≤ 4a11 a22 b12

2 2
b21 (23)
a212 4b12 + b21 52 4b11
2 2
+ b22 2
5 ≤ 4a11 b11 a22 b22 4b12 2
+ b21 5 (24)
a212 4b12 2
+ b21 5 ≤ 4a11 b11 a22 b22 0 (25)
Notepthat (25) follows immediately from (b), since G is nonincreasing and a12 = a21 . To show (23), note
that b11 b22 ≤ 21 4b11 + b22 5 ≤ min8b12 1 b21 9, so b11 b22 ≤ 4min8b12 1 b21 952 ≤ b12 b21 . Together with (25) the result
2 2 2 2
follows. Now, we claim that b11 + b22 ≤ b12 + b21 , which together with (25) gives (24). To see this, we define
m = min8b12 1 b21 9 and assume without loss of generality that b11 ≤ b22 . Since 21 4b11 + b22 5 ≤ m, we have b11 ∈
2 2
401 m7 and b11 + b22 ≤ 42m − b11 52 + b11 2
≤ 2m2 because the polynomial function x ∈ 601 m7 7→ 42m − x52 + x2
reaches its maximum for x ∈ 801 m9.
(e): We find that the left upper entry of G405G4t5 − G4t5G405 is a12 a21 4e−b21 t − e−b12 t 5, so G405G4t5 =
G4t5G405 implies b12 = b21 . Given that, a direct calculation shows that G405G4t5 = G4t5G405 is equivalent to
a11 4e−b11 t − e−b12 t 5 + a22 4e−b12 t − e−b22 t 5 = 0. If a11 = a22 , this implies b11 = b22 . If a11 6= a22 , by the equivalent
equation a22 − a11 = a22 e−4b22 −b12 5t − a11 e−4b11 −b12 5t we see that b11 = b22 = b12 .
Conversely, if either a11 = a22 and b12 = b21 and b11 = b22 , or b11 = b12 = b21 = b22 , a direct calculation shows
that G4s5G4t5 = G4t5G4s5 for all s1 t ≥ 0.
Proof of Proposition 9. G is obviously continuous and Proposition 8 yields that G is non-negative, non-
increasing, and convex since −G0 is nonincreasing. R
To show that G is not positive definite, using Mathematica we find that G4t5 = eit M4d5, where M4d5 =
Cå45d + D0 4d5 with a constant C > 0, a matrix D ∈ 2×2 , the Dirac measure 0 at 0 and å45 given by
2e2 4−cos45 +e −sin455 5e3 −443+2e5 +6i4−1+e55cos45+4i4−3+2e5 −641+e55sin45
 
 3
+ 844 +i5+65 
0
 
 3
 5e −434 +2i5+2e4 −3i55cos45+4−2ie +3i −6e −65sin45 2e2 4−cos45 +e −sin455 
84 +2i54 −3i5 3 +
If G was positive definite, then all eigenvalues of å45 would be positive for 6= 0. But using Mathematica
we find that one eigenvalue of å45 is
1
416e3 4 2 + 154 2 + 454 2 + 95 2
84 2 + 454 2 + 954 3 + 52
− 16e2 4 2 + 154 2 + 454 2 + 954sin45 + cos455
− 4 2 4 2 + 154 4 2 + 454 2 + 95449 + 4e2 + 25e6 5 2
− 10e3 43 + 2e5 2 cos45 + 12e4 2 − 65 cos425
− 60e sin454−2 cos45 + e3 + e2 5 + 3641 + e2 5551/2 51
which is negative for all with 0 < < 0002. So G is not positive definite.
Proof of Proposition 10. (a): By Lemma 5, G is non-negative if and only if for every t ≥ 0
1
4
44a12 − b12 t5+ + 4a21 − b21 t5+ 52 ≤ 4a11 − b11 t5+ 4a22 − b22 t5+ 0
Assume that G is non-negative. Choosing t = 0 yields 41 4a12 +a21 52 ≤ a11 a22 . Choosing t = min8a11 /b11 1 a22 /b22 9
yields that the right-hand side of the preceding equation is zero. So the left-hand side has to be zero which
implies that max8a12 /b12 1 a21 /b21 9 ≤ t.
Conversely, assume that 41 4a12 + a21 52 ≤ a11 a22 and max8a12 /b12 1 a21 /b21 9 ≤ min8a11 /b11 1 a22 /b22 9. So for any
t ≥ 0, we have that max841 − 4b12 /a12 5t5+ 1 41 − b21 /a21 5t5+ 9 ≤ min841 − 4b11 /a11 5t5+ 1 41 − 4b22 /a22 5t5+ 9. Thus,
+ + 2
1 1 b b
44a12 − b12 t5+ + 4a21 − b21 t5+ 52 = a12 1 − 12 t + a21 1 − 21 t
4 4 a12 a21
+ + 2
1 b b
≤ 4a12 + a21 5 max 1 − 12 t 1 1 − 21 t
4 a12 a21
+ + 2
b b
≤ a11 a22 min 1 − 11 t 1 1 − 22 t
a11 a22
+ +
b b
≤ a11 a22 1 − 11 t 1 − 22 t
a11 a22
= 4a11 − b11 t5+ 4a22 − b22 t5+ 0
So G is non-negative.
(b): G is absolutely continuous with derivative
!
−b11 8t<a11 /b11 9 −b12 8t<a12 /b12 9
0
G 4t5 =
−b21 8t<a21 /b21 9 −b22 8t<a22 /b22 9
By Lemmas 5 and 6, G is nonincreasing if and only if for almost all t > 0
1
4
4b12 8t<a12 /b12 9 + b21 8t<a21 /b21 9 52 ≤ b11 8t<a11 /b11 9 b22 8t<a22 /b22 9 0
Assume G is nonincreasing. Then choosing t small enough shows 41 4b12 + b21 52 ≤ b11 b22 . Choosing any t ≥
min8a11 /b11 1 a22 /b22 9 yields that the right-hand side of the preceding equation is zero. So the left-hand-side has
to be zero which implies max8a12 /b12 1 a21 /b21 9 ≤ min8a11 /b11 1 a22 /b22 9.
Conversely, if 41 4b12 + b21 52 ≤ b11 b22 and max8a12 /b12 1 a21 /b21 9 ≤ min8a11 /b11 1 a22 /b22 9, it is obvious that G
is nonincreasing.
(c): By computing the inverse Fourier transform, we easily get that for a1 b+ 1 b− > 0,

Z 1
8t≥09 4a − b+ t5+ + 8t<09 4a + b− t5+ = eit b 41 − e −4a5/b+
5 + b 41 − e 4a5/b−
5 d0
2 2 + −
R
Thanks to the assumption a12 = a21 , G̃ is continuous and G̃4t5 = eit M4d5 with M4d5 =
41/42 2 55å45d, with the Hermitian matrix
 
a11 4−ia12 5/b12 4ia12 5/b21
2b 11 1 − cos b 12 41 − e 5 + b21 41 − e 5
 b11 
å45 =  0
 

 4−ia12 5/b21 4ia12 5/b12
a22

b21 41 − e 5 + b12 41 − e 5 2b22 1 − cos
b22
From Theorem 1, G is positive definite if and only if å45 is positive definite for every ∈ . Using Lemma 5,
å45 is positive definite if and only if

4−ia12 5/b12 4ia12 5/b21 2 a11 a22
b12 41 − e 5 + b21 41 − e 5 ≤ b11 1 − cos b22 1 − cos 1
b11 b22
i.e., if and only if
2 2
a12 a12 a12 a12
b12 1 − cos + b21 1 − cos + b12 sin − b21 sin
b12 b21 b12 b21

a a
≤ 4b11 1 − cos 11 b22 1 − cos 22 1
b11 b22
which is equivalent to
a11 a12 a21 a22 2
= = = 1 b12 ≤ b11 b22 0
b11 b12 b21 b22
One implication is obvious. To see the other one, we apply the condition to = 24b11 /a11 5, which gives
4a12 b11 5/4b12 a11 5 ∈ and 4a21 b11 5/4b21 a11 5 ∈ , and thus 4a12 b11 5/4b12 a11 5 = 4a21 b11 5/4b21 a11 5 = 1 since
max8a12 /b12 1 a21 /b21 9 ≤ min8a11 /b11 1 a22 /b22 9 by assumption. Similarly, considering = 24b22 /a22 5 gives
4a12 b22 5/4b12 a22 5 = 4a21 b22 5/4b21 a22 5 = 1. In particular, b12 = b21 and the condition for = 0 gives the inequality
on b’s. The remainder is obvious.
Acknowledgments. The authors thank an anonymous referee for comments that helped to substantially improve a
previous version of the manuscript. A.A. is grateful for the support of the “Chaire Risques Financiers” of Fondation du
Risque. F.K. and A.S. thank Martin Schlather and Marco Oesting for discussions and gratefully acknowledge financial
support by Deutsche Forschungsgemeinschaft DFG through Research Grant SCHI 500/3-1.
References
[1] Alfonsi A, Fruth A, Schied A (2008) Constrained portfolio liquidation in a limit order book model. Banach Center Publications
83:9–25.
[2] Alfonsi A, Fruth A, Schied A (2010) Optimal execution strategies in limit order books with general shape functions. Quant. Finance
10(2):143–157.
[3] Alfonsi A, Schied A (2010) Optimal trade execution and absence of price manipulations in limit order book models. SIAM J. Financial
Math. 1(1):490–522.
[4] Alfonsi A, Schied A, Slynko A (2012) Order book resilience, price manipulation, and the positive portfolio problem. SIAM J. Finan.
Math. 3(1):511–533.
[5] Almgren R, Chriss N (2001) Optimal execution of portfolio transactions. J. Risk 3(2):5–39.
[6] Bertsimas D, Lo A (1998) Optimal control of execution costs. J. Financial Markets 1(1998):1–50.
[7] Bochner S (1932) Vorlesungen über Fouriersche Integrale (Akademische Verlagsgesellschaft, Leipzig, Germany).
[8] Boot JCG (1964) Quadratic Programming. Algorithms, Anomalies, Applications, Studies in Mathematical and Managerial Economics,
Vol. 2, (North-Holland Publishing Co., Amsterdam).
[9] Bouchaud J-P, Gefen Y, Potters M, Wyart M (2004) Fluctuations and response in financial markets: The subtle nature of “random”
price changes. Quant. Finance 4(4):176–190.
[10] Busseti E, Lillo F (2012) Calibration of optimal execution of financial transactions in the presence of transient market impact. J. Statist.
Mechanics: Theory Experiment. 2012(09):P09010.
[11] CFTC-SEC (2010) Findings regarding the market events of May 6, 2010. Report. Accessed July 24, 2011, https://www.sec.gov/news/
studies/2010/marketevents-report.pdf.
[12] Cramér H (1940) On the theory of stationary random processes. Ann. Math. 41(1):215–230.
[13] Donoghue WF Jr (1974) Monotone Matrix Functions and Analytic Continuation (Springer, New York). Die Grundlehren der mathe-
matischen Wissenschaften, Band 207.
[14] Falb PL (1969) On a theorem of Bochner. Publications Mathématiques de l’IHÉS 36(1):59–67.
[15] Fruth A, Schöneborn T, Urusov M (2014) Optimal trade execution and price manipulation in order books with time-varying liquidity.
Math. Finance 24(4):651–695.
[16] Gatheral J (2010) No-dynamic-arbitrage and market impact. Quant. Finance 10(7):749–759.
[17] Gatheral J, Schied A (2013) Dynamical models of market impact and algorithms for order execution. Fouque J-P, Langsam J, eds.,
Handbook on Systemic Risk (Cambridge University Press, Cambridge, UK), 579–602.
[18] Gatheral J, Schied A, Slynko A (2012) Transient linear price impact and Fredholm integral equations. Math. Finance 22(3):445–474.
[19] Gihman I, Skorohod A (1974) The Theory of Stochastic Processes I (Springer, Berlin).
[20] Gill PE, Murray W, Wright MH (1981) Practical Optimization (Academic Press Inc. [Harcourt Brace Jovanovich Publishers], London).
[21] Glöckner H (2003) Positive definite functions on infinite-dimensional convex cones. Mem. Amer. Math. Soc. 166(789):xiv+128.
[22] Guo X (2013) Optimal placement in a limit order book. Topaloglu H, ed. TUTORIALS in Operations Research, Vol. 10, 191–200.
[23] Huberman G, Stanzl W (2004) Price manipulation and quasi-arbitrage. Econometrica 74(4):1247–1275.
[24] Kratz P, Schöneborn T (2013) Optimal liquidation in dark pools. Quant. Finance 14(9):1519–1539.
[25] Løkka A (2012) Optimal execution in a multiplicative limit order book. Preprint.
[26] Moro E, Vicente J, Moyano LG, Gerig A, Farmer JD, Vaglica G, Lillo F, Mantegna RN (2009) Market impact and trading profile of
hidden orders in stock markets. Physical Rev. E 80(6):066–102.
[27] Naimark M (1943) Positive definite operator functions on a commutative group. Bull. Acad. Sci. URSS Sér. Math. 6Izvestia Akad. Nauk
SSSR7 7(5):237–244.
[28] Obizhaeva A, Wang J (2013) Optimal trading strategy and supply/demand dynamics. J. Financial Markets 16(1):1–32.
[29] Pólya G (1949) Remarks on characteristic functions. Neyman J, ed. Proc. Berkeley Sympos. Math. Statist. Probab. (University of
California Press, Oakland, CA), 115–123.
[30] Predoiu S, Shaikhet G, Shreve S (2011) Optimal execution in a general one-sided limit-order book. SIAM J. Financial Math. 2(1):
183–212.
[31] Sasvári Z (2013) Multivariate Characteristic and Correlation Functions, de Gruyter Studies in Mathematics, Vol. 50 (Walter de Gruyter
& Co., Berlin).
[32] Schied A, Schöneborn T, Tehranchi M (2010) Optimal basket liquidation for CARA investors is deterministic. Appl. Math. Finance
17(6):471–489.
[33] Schöneborn T (2011) Adaptive basket liquidation. Preprint.
[34] Young WH (1913) On the Fourier series of bounded functions. Proc. London Math. Soc. 12(2):41–70.

Multivariate Transiet Price Impact Model

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Multivariate Transiet Price Impact Model

Uploaded by

Copyright:

Available Formats

This article was downloaded by: [155.198.30.

43] On: 26 August 2019, At: 10:36

Mathematics of Operations Research

Multivariate Transient Price Impact and Matrix-Valued

To cite this article:

Full terms and conditions of use: https://pubsonline.informs.org/page/terms-and-conditions

Copyright © 2016, INFORMS

Please scroll down for article—it is on subsequent pages

Multivariate Transient Price Impact and Matrix-Valued

Florian Klöck, Alexander Schied

The set of deterministic liquidation strategies in X41 X0 5 is denoted by Xdet 41 X0 5.

for the function G̃2  → K×K defined by

Ɛ6C 4Ç57 ≥ 0 for all 1 X0 ∈ k 1 and Ç ∈ X41 X0 5 (7)

Moreover, any matrix-valued measure M with (10) is uniquely determined by H.

G4t5 = g4tB5 = O > diag4g4t1 51 : : : 1 g4tK 55O1 t ≥ 00 (13)

inf min Ɛ6C0 4Î57 = Ɛ6C 4X ∗ 570

Ç = 41 1 : : : 1 N 5 can be conveniently expressed as follows:

˜ we find that the components of the optimal

 = −4Id + A54N Id − 4N − 25A5−1 X0 0

The formula for the optimal strategy thus simplifies to

following decay kernel !

(a) G is non-negative if and only if /˜ ≤ 1 and  ≤ 1.

This proves (4).

Î1 LÎ = Ç1 LÇ + 2LÇ1 Î0  + Î0 1 LÎ0 1

where ê2  → + 4K5 is the continuous function given by

x 4dt5 = x> G45x 0 4dt5 + x 4t5 dt1 (19)

where 0 is the Dirac measure concentrated in 0 and

d 4E5 2= 4D ∩ E5 and c 4E5 2= 4Dc ∩ E5

be the discrete and continuous parts of , respectively. Clearly,

and we repeat this procedure at most K times to get (11).

The details are left to the reader.

By multiplying both sides of this equation with L we obtain

which, again by Proposition 2, implies that Î is also optimal for GL .

Proof of Proposition 7. Since L is invertible, the transformation

Î 7−→ ÎL 2= 4L−1 1 1 : : : 1 L−1  5

0 ≤ k> G̃4tk − t` 5` = 4kL 5> G̃L 4tk − t` 5`L

That is, if and only if

is non-negative. Analogously to (a), the result follows.

is non-negative. The result follows analogously to (a). R

a212 4b12 + b21 52 4b11

Comparing the coefficients for  0 ,  2 , and  4 , we see that it is sufficient to have

a212 4b12 + b21 52 b11 b22 ≤ 4a11 a22 b12

You might also like

The set of deterministic liquidation strategies in X41 X0 5 is denoted by Xdet 41 X0 5.

for the function G̃2 → K×K defined by

Ɛ6C 4Ç57 ≥ 0 for all 1 X0 ∈ k 1 and Ç ∈ X41 X0 5 (7)

G4t5 = g4tB5 = O > diag4g4t1 51 : : : 1 g4tK 55O1 t ≥ 00 (13)

inf min Ɛ6C0 4Î57 = Ɛ6C 4X ∗ 570

Ç = 41 1 : : : 1 N 5 can be conveniently expressed as follows:

= −4Id + A54N Id − 4N − 25A5−1 X0 0

(a) G is non-negative if and only if /˜ ≤ 1 and ≤ 1.

Î1 LÎ = Ç1 LÇ + 2LÇ1 Î0 + Î0 1 LÎ0 1

where ê2 → + 4K5 is the continuous function given by

x 4dt5 = x> G45x 0 4dt5 + x 4t5 dt1 (19)

where 0 is the Dirac measure concentrated in 0 and

d 4E5 2= 4D ∩ E5 and c 4E5 2= 4Dc ∩ E5

be the discrete and continuous parts of , respectively. Clearly,

Î 7−→ ÎL 2= 4L−1 1 1 : : : 1 L−1 5

0 ≤ k> G̃4tk − t` 5` = 4kL 5> G̃L 4tk − t` 5`L

Comparing the coefficients for 0 , 2 , and 4 , we see that it is sufficient to have