MATHEMATICS OF OPERATIONS RESEARCH

Vol. 00, No. 0, Xxxxx 0000, pp. 000–000

INFORMS
doi 10.1287/xxxx.0000.0000
issn 0364-765X | eissn 1526-5471 | 00 | 0000 | 0001 c 0000 INFORMS

Authors are encouraged to submit new papers to INFORMS journals by means of

a style file template, which includes the journal title. However, use of a template
does not certify that the paper has been accepted for publication in the named jour-
nal. INFORMS journal templates are for the exclusive purpose of submitting to an
INFORMS journal and should not be used to distribute the papers in print or online
or to submit the papers to another publication.

A Law of Large Numbers for Limit Order Books

Ulrich Horst, Michael Paulsen
Institut für Mathematik
Humboldt-Universität zu Berlin
Unter den Linden 6
10099 Berlin
[email protected]

We define a stochastic model of a two-sided limit order book in terms of its key quantities best bid [ask]
price and the standing buy [sell] volume density. For a simple scaling of the discreteness parameters, that
keeps the expected volume rate over the considered price interval invariant, we prove a limit theorem. The
limit theorem states that, given regularity conditions on the random order flow, the key quantities converge
in probability to a tractable continuous limiting model. In the limit model the buy and sell volume densities
are given as the unique solution to first-order linear hyperbolic PDEs, specified by the expected order flow
parameters. We calibrate order flow dynamics to market data for selected stocks and show how our model
can be used to derive endogenous shape functions for models of optimal portfolio liquidation under market
impact.
Key words : limit order book, scaling limit, averaging principle, queueing theory
MSC2000 subject classification : Primary: 60B11 ; secondary: 90B22, 91B70
OR/MS subject classification : Primary: Finance, Probability, Queues; secondary: Limit theorems
History :

Introduction A significant part of financial transactions is nowadays carried out through

electronic limit order books (LOBs). A LOB is a record, maintained by an exchange or special-
ist, of unexecuted orders awaiting execution. From a mathematical perspective, LOBs are high-
dimensional complex priority queueing systems. Incoming limit orders can be placed at many differ-
ent price levels while incoming market orders are matched against standing limit orders according
to a set of priority rules. Almost all exchanges give priority to orders submitted at more compet-
itive prices (“price priority”) and displayed orders have priority over hidden orders at the same
level (“display priority”). Orders with the same display status and submission price are usually
served on a first-come-first-serve basis. The inherent complexity of limit order books renders their
mathematical analysis challenging. In this paper, we propose a queueing theoretic LOB model
whose dynamics converges to a coupled ODE:PDE system after suitable scaling that can be solved
in closed form.
There is a significant economic and econometric literature on LOBs including Biais at al. [7],
Easley and O’Hara [13], Foucault et al. [15], Gloston and Milgrom [18], Parlour [25], Rosu [27] and
many others that puts a lot of emphasis on the realistic modeling of the working of the LOB, and
on its interaction with traders’ order submission strategies. There are, however, only few papers
that analyze order flows and the resulting LOB dynamics in a mathematically rigorous manner.
Among the first was the one by Kruk [22]. He studied a queueing theoretic model of a transparent

1
 Horst and Paulsen: A Law of Large Numbers for LOBs
2 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

double auction in continuous time. The microstructure may be interpreted as that of a simple limit
order book, if one considers a buyer as a buy limit order and a seller as a sell limit order. At the
auction, there are N ∈ N possible prices of the security and thus 2N different classes of customers (N
classes of buyers and N classes of sellers). Kruk established diffusion and fluid limits. The diffusion
limit states that for N=2 possible prices, the scaled number of outstanding buy orders at the lower
price and the scaled number of outstanding sell orders at the higher price converge weakly to a
semimartingale reflected two-dimensional Brownian motion in the first quadrant. The fluid limit
is such that various LOB quantities converge weakly to affine functions of time.
Since information on the best bid and ask price and volume at different price levels is available
to all market participants, it is natural to assume that the order arrival dynamics depends on the
current state of the order book. The feature of conditional state-dependence was considered by
Cont et al. [12], who proposed a continuous time stochastic model with a finite number of possible
prices where events (buy/sell market order arrival, buy/sell limit order placement and cancelation)
are modeled using independent Poisson processes. The arrival rates of limit orders depend on the
distance to the best bid and ask price in a power-law fashion. The authors were able to show that
the state of the order book, defined as a vector containing all volumes in the order book at different
prices, is an ergodic Markov process. Using this fact, several key quantities such as the probabilities
of a mid price move, a move in the bid price before a move in the ask price, or the probability
of volume execution before a price move could be computed and benchmarked against real data
without taking scaling limits.
Cont and de Larrard [11] considered a scaling limit in the diffusion sense for a Markovian limit
order market in which the state is represented by the best bid and ask price and the queue length,
i.e. the number of orders at the best bid and ask price, respectively. With this reduction of the
state space, under symmetry conditions on the spread and stationarity assumptions on the queue
lengths, it was shown that the price converges to a Brownian motion with volatility specified
by the model parameters in the diffusion limit. Very recently, Cont and de Larrard [10] studied
the reduced state space under weaker conditions and proved a refined diffusion limit by showing
that under heavy traffic conditions the bid and ask queue lengths are given by a two-dimensional
Brownian motion in the first quadrant with reflection to the interior at the boundaries, similar to
the diffusion limit result for N = 2 prices in [22].
In the framework analyzed by Abergel and Jedidi [1], the volumes of the order book at different
distances to the best bid and ask were modeled as a finite dimensional continuous time Markov
chain and the order flow as independent Poisson processes. Under the assumption that the width
of the spread is constant in time, using Foster-Lyapunov stability criteria for the Markov chain,
the authors proved ergodicity of the order book and a diffusion limit for the mid price. In the
diffusion limit, the mid price is a Brownian motion with constant volatility given by the averaged
price impact of the model events on the order book.
In this paper, we prove a law of large numbers result for the whole book (prices and volumes).
Specifically, we propose a continuous-time model of a two-sided state-dependent order book with
random order flow and cancelation, and countably many submission price levels. The buy and the
sell side volumes are coupled through the best bid and ask price dynamics.1 We model the buy and
sell side volumes as density functions in relative price coordinates, i.e. relative to their distance to
the best bid/ask prices. Volumes at positive distances are limit orders awaiting execution (“visible
book”); volumes at negative distances specify orders that would be placed in the spread if the next

1
The coupling of the buy and sell sides through prices is essential for the limiting volume dynamics to follow a PDE.
It is not essential for obtaining a scaling limit per se. Our mathematical framework is flexible enough to allow for
dependencies of order flows on standing volumes, but that one would lead to a function-valued ODE as the scaling
limit, rather than a PDE. The PDE-scaling is so much more transparent that it justifies, in our view, the restriction
of order flow dependencies on prices.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 3

order is a spread limit order placement (“shadow book”). The state of the book at any point in
time is thus described by a quadruple comprising the best bid price, the best ask price, the relative
buy volume density function and the relative sell volume density function.
The state dynamics is defined in terms of a recursive stochastic process taking values in a function
space. When the analysis of the market is limited to prices as in e.g. Garman [17], Bayraktar et al. [6]
or Horst and Rothe [21], or to the joint dynamics of prices and aggregate volumes (e.g. at the top of
the book) as in Cont and de Larrard [10, 11], then the limiting dynamics can naturally be described
by ordinary differential equations or real-valued diffusion processes, depending on the choice of
scaling. The analysis of the whole book including the distribution of standing volume across many
price levels is much more complex. Osterrieder [24] modeled LOBs using measure-valued diffusions.
Our approach is based on an averaging principle for Banach space-valued processes. The key is a
uniform law of large numbers for Banach space-valued triangular martingale difference arrays. It
allows us to show that the volume densities take values in L2 and that the noise in the order book
models vanishes in the limit with our choice of scaling.
Our scaling limit requires two time scales: a fast time scale for cancelations and limit order
placements outside the spread (events that do not lead to price changes), and a comparably slow
time scale for market order arrivals and limit order placements in the spread (events that lead to
price changes). The choice of time scales captures the fact that in real-world markets significant
proportions of orders are never executed. Mathematically, the different time scales imply that
aggregate cancelations and limit order placements outside the spread in between two consecutive
price changes can be approximated by their expected values.
Our main result states that when the price tick tends to zero, order arrival rates tend infinity,
and the impact of an individual placement/cancelation on the standing volume tends to zero,
then the sequence of scaled order book dynamics converges in probability uniformly over compact
time intervals to a deterministic limit. The limiting model is such that the best bid and ask price
dynamics can be described in terms of two coupled ODEs, while the dynamics of the relative buy
and sell volume density functions can be described in terms of two linear first-order hyperbolic
PDEs with variable coefficients. A similar limiting dynamics was recently obtained in [16] but with
constant limiting price process. The latter is not required in our model.
Our LOB model can be used to obtain shape functions for models of optimal portfolio liquidation
under market impact. In such models the goal is to find optimal strategies for unwinding large
numbers of shares over small time periods. They typically assume that the dynamics of standing
buy (or sell) side volumes can be described in terms of exogenous shape functions. Within our
modeling framework shape function arise as part of the solution to a coupled ODE:PDE system.
Calibrating the model parameters to market data, this allows for a fully endogenous derivation of
shape functions from order arrival and cancellation dynamics. Using LOBSTER data data for Jan 2,
2014 we calibrated order placement dynamics for selected stocks. For the stocks Ebay and Facebook
we found that exponential densities provide good fits; for Apple the placement densities turn out
to be almost constant. For Microsoft most of the placement and cancellation activity concentrate
around the top of the book. For such stocks, our scaling does not seem to be appropriate. We
further report empirical evidence supporting our assumption of multiple time scales. For instance,
for Apple we found that only 1.6% of all orders on Jan 2, 2014 lead to price changes.
The remainder of this paper is organized as follows. In Section 1 we define a sequence of limit
order book models in terms of four scaling parameters: price tick, expected waiting time between
two consecutive orders, volume placed/canceled and the proportion of order arrivals leading to
price changes, and state our main result. In Section 2 we establish convergence of the bid/ask price
dynamics to a 2-dimensional ODE. Section 3 is devoted to the analysis of the limiting volume
dynamics. Section 4 illustrates how our model can potentially be applied in a portfolio liquidation
framework; Section 5 illustrates how selected model parameters can be calibrated to market data.
 Horst and Paulsen: A Law of Large Numbers for LOBs
4 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

The uniform law of large numbers for triangular martingale difference arrays as well as useful
auxiliary results are proved in an appendix.

1. A sequence of discrete order book models In electronic markets orders can be sub-
mitted for prices that are multiples of the price tick, the smallest increment by which the price
can move. In this section, we introduce a sequence of order book models for which we establish a
scaling limit when the price tick and impact of a single order on the state of the book tend to zero,
while the rate of order arrivals tends to infinity.
The sequence of models is indexed by n ∈ N. We assume that the set of price levels at which
(n)
orders can be submitted in the n:th model is {xj }j∈Z where Z denotes the one-dimensional integer
(n)
lattice.2 We put xj := j · ∆x(n) for j ∈ Z where ∆x(n) is the tick size in the n:th model.
The state of the book changes due to incoming orders and cancelations. The state after k ∈ N
(n)
such events will be described by a random variable Sk taking values in a suitable state space E.
(n)
In the n:th model, the k:th event occurs at a random point in time τk . The time between two
consecutive events will be tending to zero sufficiently fast as n → ∞. The state and time dynamics
will be defined, respectively, as
(n) (n) (n) (n) (n) (n)
S0 := s0 , Sk+1 := Sk + Dk (Sk ) (1)
and
(n) (n) (n) (n) (n)
τ0 := 0, τk+1 := τk + Ck (Sk ). (2)
(n) (n) (n) (n) (n)
Here s0 ∈ E is a deterministic initial state, and Dk (Sk ) : E → E and Ck (Sk ) : E → [0, ∞) are
random operators that will be introducedh below. Thei conditional expected increment
h of the
i state
(n) (n) (n) (n)
sequence, given Sk , will be denoted E Dk (Sk ) ; the unconditional increment E Dk .

1.1. The order book models In the sequel we specify the dynamics of our order book
models. Throughout, all random variables will be defined on a common probability space (Ω, F , P).

1.1.1. The initial state The initial state of the book in the n:th model is given by a pair
(n) (n)
(B0 , A0 ) of best bid and ask prices together with the standing buy and sell limit order volumes
at the various price levels. It will be convenient to identify the standing volumes with step functions
∞
X ∞
X
(n) (n),j (n) (n),j
vb,0 (x) := vb,0 1[x(n) ,x(n) ) (x), vs,0 (x) := vs,0 1[x(n) ,x(n) ) (x) (x ≥ 0)
j j+1 j j+1
j=0 j=0

that specify the liquidity available for buying and selling relative to the best bid and ask price.
The liquidity available for buying (sell side of the book) j ∈ N0 ticks above the best ask price at
(n)
the price level x (n) is
A0 +j
(n)
Z xj+1
(n) (n),j
vs,0 (x)dx = vs,0 · ∆x(n) .
(n)
xj

The volume available for selling (buy side of the book3 ) l ∈ N0 ticks below the best bid price at the
(n)
price level x (n) is
B0 −l
Z x(n)
l+1
(n) (n),l
vb,0 (x)dx = vb,0 · ∆x(n) .
(n)
xl
(n) (n)
In order to conveniently model placements of limit orders into the spread, we extend vb,0 and vs,0
2
The assumption that there is no minimum price is made for analytical convenience and can easily be relaxed.
3
Notice that the liquidity available for buying is captured by the sell side of the book and vice versa.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 5

Quantity
Bid side (offer to buy) Ask side (offer to sell)
500

400

300

200

100

44 45 46 47 48 49 50 53 54 55 56 57 58 59    Price  in  €
Bid price Ask price

Figure 1. Snapshot of the order book.

to the negative half-line. The collection of volumes standing at negative distances from the best
bid/ask price is referred to as the shadow book. The shadow book will undergo the same dynamics
as the standing volume (“visible book”). At any point in time it specifies the volumes that will be
placed into the spread should such an event occur next4 . The role of the shadow will be further
illustrated in Section 1.1.4 below; see also Figure 3.
Definition 1. In the n:th model the initial state of the book is given by a quadruple
 
(n) (n) (n) (n) (n)
S0 (·) = B0 , A0 , vb,0 (·), vs,0 (·)

(n) (n) (n) (n)

where B0 ≤ A0 are the best bid/ask price and the step functions vb,0 , vs,0 : R → [0, ∞) are to be
interpreted as follows:



standing buy [sell] limit order volume density at price distance x
i below [above] the best bid [ask] price, for x ≥ 0 (visible book)


h
(n) (n)
vb,0 (x) vs,0 (x) := (3)



potential buy [sell] limit order volume density at price distance x

above [below] the best bid [ask] price, for x < 0 (shadow book).

Throughout, we shall use the notation f = O(g) and f = o(g) to indicate that the function f
grows asymptotically no faster than g, respectively that |f (x)/g(x)| → 0 as x → ∞. With this, we
are ready to state our conditions on the initial states. In particular, we assume that the initial
volume density functions vanish outside a compact price interval.5

4
One has to specify the volumes placed into the spread somehow. Our choice of shadow books is one such way. The
role of the shadow book will be further clarified in the following subsection when we define the impact of order arrivals
on the state of the book. Its initial state is part of the model; future states will undergo the dynamics analogous to
those of the visible book.
5
This assumption, which may be generalized, considerably simplifies some of the analysis that follows.
 Horst and Paulsen: A Law of Large Numbers for LOBs
6 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

Quantity
Shadow book
500

400

300

200

100

-6 -4 -2 0 +2 +4 +6 Distance from the

Bid price: 50 best bid price

Figure 2. Buy-side volume density function in relative coordinates; green: standing (“visible”) volume; grey: shadow
book.

Assumption 1 (Convergence of initial states). The initial volume density functions van-
ish outside a compact interval [−M, M ], for some M > 0. Moreover, there exists non-negative
bounded and continuously differentiable functions vr,0 ∈ L2 (r ∈ {b, s}) with bounded derivatives
such that
(n)
||vr,0 − vr,0 ||L2 = o(1)
as well as
(n) (n)
|vr,0 (· ± ∆x(n) ) − vr,0 (·)|∞ = O(∆x(n) ).
Here, k · kL2 denotes the L2 -norm on R with respect to Lebesgue measure. Moreover,
(n) (n)
lim (B0 , A0 ) = (B, A).
n→∞

The first condition on the volume density functions is intuitive. The second condition will become
clear later; it will be used to bound the impact of market orders and limit orders placed into the
spread on the state of the book.

1.1.2. Event types There are eight events - labeled A, ..., H - that change the state of the
book. The events A, ..., D affect the buy side of the book:
A := {market sell order} B := {buy limit order placed in the spread}
C := {cancelation of buy volume} D := {buy limit order not placed in spread}
The remaining four events affect the sell side of the book:
E := {market buy order} F := {sell limit order placed in the spread}
G := {cancelation of sell volume} H := {sell limit order not placed in the spread}.

(n)
We will describe the state dynamics of the n:th model by a stochastic process {Sk }k∈N that takes
values in the Hilbert space
E := R × R × L2 × L2 .
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 7

(n)
The first two components of the vector Sk stand for the best bid and ask price after k events; the
third and fourth component refer to the buy and sell volume density functions relative to the best
bid and ask price, respectively (visible and shadow book). We define a norm on E by
kαkE := |α1 | + |α2 | + kα3 kL2 + kα4 kL2 , α = (α1 , α2 , α3 , α4 ) ∈ E. (4)
In the sequel we specify how different events change the state of the book and how order arrival
times and sizes scale with the parameter n ∈ N.

1.1.3. Active orders Market orders and placements of limit orders in the spread lead to
price changes.6 With a slight abuse of terminology we refer to these order types as active orders.
For convenience, we assume that market orders match precisely against the standing volume at the
best prices and that limit orders placed in the spread improve prices by one tick. The assumption
that market orders decrease (increase) the best bid (ask) price by one tick while limit orders placed
in the spread decrease (increase) prices by the same amount has been made in the literature before
and can be generalized without too much effort. However, this would unnecessarily complicate the
analysis that follows.
Remark 1. A market order whose size exceeds the standing volume at the top of the book
and that would hence move the price by more than one tick is split by the exchange into a series of
consecutively executed orders. The size of each such ‘suborder’, except the last, equals the liquidity
at the current best price. Thus, by definition, a single market order cannot move the price by more
than one tick.
If the k:th event is a sell market order (Event A), then the relative buy volume density shifts
one price tick to the left (the liquidity that stood l ticks into the book now stands l − 1 ticks into
the book), the best bid price decreases by one tick and the relative sell volume density and the
best ask price remain unchanged. Since the relative volume density functions are defined on the
whole real line, the transition operators
(n) (n)
T+ (v)(·) = v(· + ∆x(n) ), T− (v)(·) = v(· − ∆x(n) )
are well defined and one has that
 
(n) (n) (n) (n) (n)
vb,k+1 (·) = T+ vb,k (·), vs,k+1 (·) = vs,k (·)

and
(n) (n) (n) (n)
Bk+1 = Bk − ∆x(n) , Ak+1 = Ak .
The placement of orders into the spread will be modeled using the shadow book. If the k:th event
is a buy limit order placement in the spread (Event B), the relative buy volume density shifts one
price tick to the right (the liquidity that stood one tick above the best bid in the shadow book
now stands at the top of the visible book), the best bid price increases by one tick and the relative
buy volume density and the best ask price remain unchanged:
 
(n) (n) (n) (n) (n)
vb,k+1 (·) = T− vb,k (·), vs,k+1 (·) = vs,k (·)

and
(n) (n) (n) (n)
Bk+1 = Bk + ∆x(n) , Ak+1 = Ak .
Remark 2. Notice that market order arrivals and limit order placements in the spread are
“inverse operations”: a market sell order arrival followed by a limit buy order placement in the
spread (or vice versa) leaves the book unchanged.
6
A market order that does not lead to a price change can be viewed as a cancelation of standing volume at the best
bid/ask price.
 Horst and Paulsen: A Law of Large Numbers for LOBs
8 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

1.1.4. Passive orders Limit order placements outside the spread and cancelations of stand-
ing volume do not change prices. With the same minor abuse of terminology as before, we refer
to these order types as passive orders. We assume that cancelations of buy volume (Event C)
occur for random proportions of the standing volume at random price levels while limit buy order
placements outside the spread (Event D) occur for random volumes at random price levels. The
submission and cancelation price levels are chosen r elative to the best bid price.
Assumption 2. For each k ∈ N there exist random variables ωkC , ωkD taking values in (0, 1),
respectively [0, M ] for some M > 0 and random variables πkC , πkD taking values in [−M, M ] such
that, if the k:th event is a limit buy order cancelation/placement, then it occurs at the price level
(n)
x (n) (j ∈ Z) for which
Bk −j
(n) (n)
πkC,D ∈ [xj , xj+1 ).
The volume canceled, respectively, placed is
(n)
ωkC · ∆v (n) · vb,k (πkC ) respectively ωkD · ∆v (n) .
(n)
Here vb,k (πkC ) is the value of the volume density function at the cancelation price level, and ∆v (n)
is a scaling parameter that describes the impact of an incoming limit order (cancelation) on the
state of the order book.
Volume changes take place in the visible or the shadow book, depending on the sign of of πkI . If
πkI ≥ 0, then the visible book changes; if πkI < 0, then the placement/cancelation takes place in the
shadow book. In order to illustrate the working of the shadow book, suppose that (k + 1)-st event
D
is an order placement of size ωk+1 ∆v (n) in the (buy-side) shadow book one tick above the best bid,
D (n)
i.e. πk+1 ∈ [−∆x , 0) and that the (k + 2)nd event is a buy limit order placement in the spread.
Then,
(n) (n) (n)
Bk+2 = Bk+1 + ∆x(n) = Bk + ∆x(n)
and the value of the volume density function at the top of the book is:

(n),0 (n),−1 (n),−1 D ∆v (n)

vb,k+2 = vb,k+1 = vb,k + ωk+1 .
∆x(n)
This is how the buy order previously placed into the shadow is now part of the visible book. The
role of the shadow book is further illustrated by Figure 3. When a cancelation occurs at this price

Quantity Quantity
500 500

400 400

300 300

200 200

100 100

-6 -4 -2 0 +2 +4 +6 Distance from the -6 -4 -2 0 +2 +4 Distance from the

Bid price: 50 best bid price Bid price: 51 best bid price

Figure 3. Left: initial buy side volume density with shadow book; right: new state (visible and shadow book) after
a limit buy-order has been placed in the spread.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 9

(n) (n) (n)

∆v
level, then the new volume is (∆x(n) − ωkC · ∆v (n) ) · vb,k (xnj ) = (1 − ωkC ∆x n
(n) ) · vb,k (xj ) · ∆x
(n)
because
cancellations are proportional to standing volumes. On the level of the volume density functions
this yields:
∞
(n) (n) ∆v (n) (n),C (n) (n),C
X
vb,k+1 (·) = vb,k (·) − · Mk (·) · vb,k (·), where Mk (x) = ωkC 1{πC ∈[x(n) x(n) )} (x).
∆x(n) j=−∞
k j j+1

Volume placements are additive. If the next order is a limit buy order, then the volume density
function changes according to
∞
(n) (n) ∆v (n) (n),D (n),D
X
vb,k+1 (·) = vb,k (·) + (n)
· Mk (·), where Mk (x) := ωkD 1{πD ∈[x(n) ,x(n) )} (x).
∆x j=−∞
k j j+1

In either case, the bid/ask price and standing sell side volume of the book remain unchanged:
(n) (n) (n) (n) (n) (n)
vs,k+1 (·) = vs,k (·), Bk+1 = Bk , Ak+1 = Ak .

Similar considerations apply to the sell side with respective random quantities ωkG , ωkH and πkG , πkH .
Assumption 3. For I ∈ {C, D, G, H} the sequences {ωkI }k∈N and {πkI }k∈N are independent
sequences of i.i.d. random variables. Moreover, the random variables πkI have C 2 -densities f I with
compact support.
Lipschitz continuity of f I implies existence of a constant K < ∞ such that:
(n)
Z xj
(n) (n) (n) (n)
P[πkI ∈ [xj , xj+1 )] − P[πkI ∈ [xj−1 , xj )] ≤ |f I (y + ∆x(n) ) − f I (y)|dy
(n)
xj−1 (5)

2
≤ K ∆x(n) .

Moreover, if we put
∞ (n)
xj+1
E[ω1I ]
X Z
(n),I (n),I (n),I
f (·) := fj 1[x(n) ,x(n) ) (·) with fj := f I (x) dx, (6)
j=−∞
j j+1 ∆x(n) (n)
xj

then
(n)
kf n,I − f I k∞ = o(1) and kT± ◦ f n,I − f (n),I k∞ = O(∆x(n) ). (7)

1.1.5. Event times The dynamics of event times is specified in terms of a sequence of inter-
arrival times whose distributions may depend on the prevailing best bid and ask prices.
Assumption 4. Let {ϕk }k∈N be a sequence of non-negative random variables that are condi-
tionally independent and identically distributed, given the current best bid and ask price:
(n) (n) (n)
P[ϕk ≤ t|Sk ] = P[ϕk ≤ t|Bk , Ak ].

In the sequel we write ϕ(Ak , Bk ) for ϕk to indicate the dependence of the distribution of ϕk on
the best bid and ask price. Similar notation will be applied to other random variables whenever
convenient. In the n:th model, we scale time by a factor ∆t(n) . More precisely, we assume that the
dynamics of the event times in the n:th model is given by:
(n) (n) (n) (n) (n) (n) (n) (n) (n) (n)
τk+1 = τk + Ck (Bk , Ak ), where Ck (Bk , Ak ) := ϕ(Bk , Ak ) · ∆t(n) . (8)
 Horst and Paulsen: A Law of Large Numbers for LOBs
10 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

1.1.6. Event types Event types are described in terms of a sequence of random event indi-
cator variables {φk }k∈N taking values in the set {A, ..., H}. We assume that the random variables

(n) (n)
φk = φ(Bk , Ak ) (k ∈ N0 )

are conditionally independent and identically distributed, given the best bid/ask price, and that
their conditional probabilities

(n) (n) (n)

p(n),I (Bk , Ak ) =: P[φk = I |Sk ]

satisfy the following condition.

Assumption 5. There exist continuous functions with bounded gradients pI : R2 → [0, 1] and a
scaling parameter ∆p(n) → 0 such that

p(n),I (·, ·) = ∆p(n) · pI (·, ·) for I = A, B, E, F

p(n),I (·, ·) = (1 − ∆p(n) ) · pI (·, ·) for I = C, D, G, H
pA + pB + pE + pF = 1
pC + pD + pG + pH = 1

Remark 3. The preceding assumption implies that an event is an active order with probability
∆p(n) and a passive order with probability 1 − ∆p(n) , independently of the state of the book.
Conditioned on the order being active or passive, it is of type I with a probability pI (·, ·) that
depends on the current best bid and ask price. We allow the above probabilities to be zero in order
(n) (n)
to account for the fact that no price improvements can take place when Bk = Ak and to avoid
depletion of the order book.7
The expected impact of each active order on the state of the book will be of the order ∆x(n) ,
and that of a passive order will be of the order ∆v (n) . Because active orders arrive at a rate that
is ∆p(n) -times slower than that of passive orders, the relative average impact of active to passive
(n) (n)
orders on the state of the book will of the order ∆p∆v∆x(n) . Our scaling limit requires to equilibrate
the impact of active and passive orders. In order to guarantee that there will be no fluctuations in
the standing volumes in the limit as n → ∞ we also need a minimum relative frequency of passive
order arrivals. This motivates the following assumption.

Assumption 6. The scaling constants ∆p(n) , ∆x(n) , ∆v (n) and ∆t(n) are such that:

∆x(n) · ∆p(n) ∆v (n) ∆p(n)

lim = c0 , lim = c1 , and lim  α = c2
n→∞ ∆v (n) n→∞ ∆t(n) n→∞
∆t(n)

1


for some α ∈ 2
,1 and constants c0 , c1 , c2 > 0.8

7
For simplicity we assumed that the the initial volume density functions vanish outside a compact price interval.
Hence there is a positive probability of depletion unless one assumes that no further buy/sell side price improvements
take place if the distance of the current best bid/ask price from the initial state exceeds some threshold.
(n) (n)
∆v (n)
8
For the results that follow, we will assume that ∆x ∆v·∆p
(n) = 1 and ∆t(n)
= 1 as n → ∞. Any other constant would
require further constants in the limiting dynamics.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 11

1.1.7. Active order times The previous two assumptions introduce two different time scales
for order arrivals: a fast time scale for passive order arrivals, and a comparably slow time scale
for active order arrivals. Inter-arrival times between passive orders are of the order ∆t(n) while
inter-arrival times between active orders are of the order ∆x(n) . In order to see this, let us denote
(n) (n)
by σk the arrival time of the k:th active order. The number rk+1 of events until the (k + 1)-st
active order arrival can be viewed as the first success time in a series of Bernoulli experiments with
success probability ∆p(n) and expected value ∆p1(n) . The (k + 1)-st active order arrives at time
(n) (n) (n)
σk+1 = σk + ζk · ∆x(n)

where
(n)
rk+1 −1
(n)
X
ζk := ϕl · ∆p(n) .
(n)
l=σk +1

Since the random variables ϕσ(n) +1 , ..., ϕr(n) −1 are conditionally independent and identically dis-
k k+1
(n) (n) 1
tributed, {rk } and {ϕk } are independent sequences. Because E[rk ] = ∆p(n)
, the conditional
(n) (n) (n)
expected value m(B , A ) of given the prevailing bid and ask prices is independent of n ∈ N.
ζk ,
We assume that the mapping m(·, ·) is Lipschitz continuous.
(n)
Assumption 7. The conditional expected value m(B, A) of ζk depends in a Lipschitz contin-
uous manner on the prevailing pair of bid and ask prices (B, A).

1.1.8. State dynamics We are now ready to describe the full dynamics of the state sequence.
To this end, we put  
(n) (n) (n) (n) (n)
Sk = Bk , Ak , vb,k , vs,k
      0
(n) (n) (n) (n) (n)
In terms of the indicator function 1k Sk := 1A φk (Bk , Ak ) , . . . , 1H φk (Bk , Ak ) the
(n)
dynamics of the state sequence {Sk } is of the form
(n) (n) (n) (n)
Sk+1 = Sk + Dk (Sk )
(n)
if we define the random operator Dk : E → E by
(n) (n) (n) (n) (n)
Dk (Sk ) := Mk (Sk ) · 1k (Sk ) (9)
(n) (n)
where the matrix Mk (Sk ) equals
−∆x(n) ∆x(n)
 
0 0 0 0 0 0
 
∆x(n) −∆x(n)
 
 0 0 0 0 0 0 
 
.
 
 (n),A (n),B (n) (n),C (n) ∆v (n) (n),D
M
 k Mk − ∆v (n) Mk · vb,k Mk 0 0 0 0 
∆x ∆x(n) 
 
 
(n),E (n),F (n) (n),G (n) ∆v (n) (n),H
0 0 0 0 Mk Mk − ∆v (n) Mk · vs,k Mk
∆x ∆x(n)

Here, the entries referring to shifts in the volume density functions, due to best bid and ask price
changes, are given by
   
(n),A (n) (n) (n) (n),E (n) (n) (n)
Mk := T+ vb,k − vb,k , Mk := T+ vs,k − vs,k
   
(n),B (n) (n) (n) (n),F (n) (n) (n)
Mk := T− vb,k − vb,k , Mk := T− vs,k − vs,k (10)
 Horst and Paulsen: A Law of Large Numbers for LOBs
12 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

and the entries referring the volume changes, due to placement and cancelation of volume, are
given by
∞
X
(n),I
Mv,k (x) := ωkI 1{πI ∈[x(n) ,x(n) )} (x) for events I=C, D, G, H. (11)
k j j+1
j=−∞

Observing the dynamics in continuous time, we define

(n) (n) (n) (n)
S (n) (t) := Sk and τ (n) (t) := τk for t ∈ [τk , τk+1 ). (12)

Remark 4. Overall, the state and time dynamics of our models are driven by the random
sequences {ϕk } (event times), {φk } (event types), {πkI } (placement/cancelation price levels) and
{ωkI } (placed/canceled orders). The joint dynamics of all models can be defined in terms of suitable
independent families
κk := (ϕk (B, A), φk (B, A))(A,B)∈R2 , (πkI , ωkI )I=A,...,H

(k ∈ N0 )
of independent random variables. In particular, the process {(S (n) (t), τ (n) (t))}t∈[0,T ] (n ∈ N) is
adapted to the filtration  
(n) k
Fk := σ κs : 0 ≤ s ≤ b (n) c .
∆t
1.2. The main result Our main result is Theorem 1. It states that - with our choice of
scaling - the order book dynamics can be described by a coupled ODE:PDE system when n → ∞:
the dynamics of the best bid and ask prices will be given in terms of an ODE, while that of the
relative buy and sell volume densities will be given by the respective unique classical solution of a
first order linear hyperbolic PDE with variable coefficients.
Theorem 1 (Law of Large Numbers for LOBs). Let {S (n) }n≥1 be the sequence of contin-
uous time processes defined in (12) and suppose that Assumptions 1 -7 hold. Then, for all T > 0
there exists a deterministic process s : [0, T ] → E such that
lim sup kS (n) (t) − s(t)kE = 0 in probability.
n→∞ t∈[0,T ]
   
γ(t) b(t)
The process s is of the form s(t) = , where γ(t) = is the vector of the best bid
v(·, t)   a(t)
vb (x, t)
and ask prices at time t ∈ [0, T ] and v(x, t) = denotes the vector of buy and sell volume
vs (x, t)
densities at t ∈ [0, T ] relative to the best bid and ask price. In terms of the matrices
 B   C 
p (·) − pA (·) 0 −p (·)f C (x) 0
A(·) := , B(·, x) := , (13)
0 pE (·) − pF (·) 0 −pG (·)f G (x)
the vector
 
pD (·)f D (x)
c(·, x) :=  , (14)
H H
p (·)f (x)
and the function m(·, ·) that specifies the expected waiting time between two consecutive active order
arrivals, the function γ is the unique solution to the 2-dimensional ODE system
  
dγ(t) A(γ(t)) 1
 dt = m(γ(t)) , t ∈ [0, T ]


  1 (15)
B0
 γ(0) =


A0
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 13

and (vb , vs ) is the unique non-negative bounded classical solution of the PDE
(  
1
vt (t, x) = m(γ(t)) A (γ(t)) vx (t, x) + B(γ(t), x)v(t, x) + c (γ(t), x) , (t, x) ∈ [0, T ] × R
. (16)
v(0, x) = v0 (0, x), x∈R

Remark 5. It is worth comparing our limiting model to that in [16] where a PDE limit for
the LOB is obtained by modeling the dynamics of unexecuted orders using birth-death processes.
Their limiting dynamics is less general than ours as the price
 process
 is constant in the limit.
 In(n)our

(n) ∆v (n) (n) ∆v
framework this corresponds to the case where ∆p = o ∆x(n) rather than ∆p = O ∆x (n) .
 (n) 
∆v
In this case A(.) ≡ 0, and the limiting PDE simplifies to a family of ODEs. If ∆p(n) = o ∆x (n) ,

then our method easily extends to models where prices depend on volumes in the approximating
models as in [16]. A general model with fully state dependent order dynamics has recently been
established in [20].

The analysis of the limiting dynamics can be simplified by separating the randomness on the
level of order arrival times from that of order types as shown in the following section. Subsequently,
we give an explicit solution to the limiting PDE.

1.2.1. State and time separation For the continuous-time process S (n) we write

S (n) (t) = Sγ(n) (t), Sv(n) (t)

where Sγ(n) ∈ R2 describes the dynamics of bid and ask prices, and Sv(n) (t) ∈ L2 (R) × L2 (R) describes
the dynamics of the buy and sell volume density functions. According to the following proposition
the process can be expressed as the composition of a state process η (n) and a time process µ(n) .
The proof follows from straightforward modifications of arguments given in Anisimov [4, p.108].
Proposition 1 (State and time separation). The process S (n) can be expressed as the
composition of a random state process

η (n) (t) = ηγ(n) (t), ηv(n) (t)

and a random time process µ(n) as

S (n) (t) = η (n) µ(n) (t) − ∆t(n) .

The state and time process is given by

h 
(n) (n) (n) (n)
η (t) := Sk for t∈ tk , tk+1 (17)

(n)
where tk := k · ∆t(n) and
h 
(n) (n) (n)
y (n) (u) := τk for u ∈ τk , τk+1 , (18)

respectively. The time-change µ(n) is then defined in terms of y (n) as

µ(n) (t) := inf {u > 0 : y (n) (u) > t}.

 Horst and Paulsen: A Law of Large Numbers for LOBs
14 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

The advantage of the state and time separation is that the processes η (n) and µ(n) can be analyzed
separately. In fact, we will show convergence in probability

lim sup η (n) (t) − η(t) E

= 0 and lim sup µ(n) (t) − µ(t) = 0
n→∞ t∈[0,T ] n→∞ t∈[0,T ]

to limiting processes η(t) = (ηγ (t), ηv (t)) and µ(t). Since the state sequence takes values in the
Hilbert space E, the time change theorem as proved in, e.g. Billingsley [8, p. 151], then implies
that
lim sup kS (n) (t) − η(µ(t))kE = 0 in probability.
n→∞ t∈[0,T ]

In our model, bid and ask prices are sufficient statistics for the evolution of the order book. In
particular, the limiting behavior of the sequences ηγ(n) and µ(n) can be analyzed without reference
to volumes. In Section 2 we prove the following proposition.
Proposition 2. Let γ̂ be the unique solution to the ODE
 !
 dγ̂(t) 1

 dt
= A(γ̂(t)) , t ∈ (0, T ]
1




. (19)
 !
B0



γ̂(0) = .


A0

Then,
lim sup ηγ(n) (t) − γ̂(t) = 0 in probability.
n→∞ 0≤t≤T

Moreover, the sequence of processes µ(n) satisfies

Z t
lim sup µ(n) (t) − µ(t) = 0 in probability. where µ−1 (t) = m(γ̂u )du.
n→∞ 0≤t≤T 0

Once the limiting time-change process µ has been identified, what remains to finish the proof
of Theorem 1, is to establish convergence of the volume processes ηv(n) to their deterministic limit.
This will be achieved in Section 3 where we prove the following result.
Proposition 3. Let u b be the unique classical solution of the PDE

u
bt (x, t) = A (b
γ (t)) u
bx (x, t) + B(x, γ
b(t))b b(t)) , (x, t) ∈ R × [0, T ]
u(x, t) + c (x, γ
. (20)
u
b(x, 0) = v0 (x), x∈R

Then
lim sup ηv(n) (t; ·) − u
b(t, ·) L2
=0 in probability.
n→∞ 0≤t≤T

1.2.2. Explicit solution The PDE system (20) is coupled only through the limiting price
dynamics. In particular, the equations for the buy and sell side can be solved independently. For
the buy side PDE, we can write
(
∂ub
∂t
= Ab (t) ∂u
∂x
b
+ Bb (t, x)ub + cb (t, x)
(21)
ub (0, x) = vb,0 (x)
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 15

where Ab (t) = pB (γ̂(t)) − pA (γ̂(t)), Bb (t, x) := −pC (γ̂(t))f C (x), cb (t, x) := pD (γ̂(t))f D (x). Using the
method of characteristic curves, the PDE reduces to a family of ODEs; see [14, Chapter 3] for
details. The characteristic equations for our buy side PDE read:
( (
dx dūb
dτ
= − A b (τ ) dτ
= Bb (τ, x(τ ))ūb + cb (τ, x(τ ))
and (22)
x(0) = ξ ūb (0, ξ) = vb,0 (ξ).

The solution to this ODE-system as a function of the state ξ ∈ R can be given in closed form:
Z t
x(t, ξ) = ξ − Ab (t)dt
0Z
 t  Z t  Z s  
ūb (t, ξ) = exp Bb (u, x(u, ξ))du vb,0 (ξ) + exp − Bb (u, x(u, ξ))du cb (s, x(s, ξ))ds .
0 0 0

It describes the surface {(t, ξ) : ub (t, x(t, ξ)) = ūb (t, ξ) given ub (0, ξ) = v0,b (ξ)}. The solution to the
buy side PDE can be recovered from the solution to the ODE-system through
 Z t 
ub (t, y) = ūb t, y + Ab (s)ds .
0

Due to our smoothness assumptions on the volume placement and cancelation functions it is not
hard to verify that the solution is uniformly bounded with uniformly bounded first and second
order derivatives with respect to the time and space variable. Moreover, since the function vb,0
vanishes outside a compact interval (Assumption 1) and no orders are placed or canceled beyond
a distance M from the best bid/ask price (Assumption 2), the function ub (t, ·) vanishes outside
some compact interval I(T ) for all t ∈ [0, T ]. Altogether, one has the following result.
Proposition 4. Under the assumptions of Theorem 1, the PDE (16) has a unique solution
ub . The solution is uniformly bounded, with uniformly bounded first and second order derivatives
with respect to both variables, and there exists an interval I such that ub (t, x) = 0 for all t ∈ [0, T ]
and x ∈/ I.

1.3. A benchmark model with Poisson arrivals In this section we discus examples where
the limiting dynamics can be given in closed form, provide simulation results that illustrate how
our model could be used to approximate volume dynamics, and illustrate how our model could be
applied to portfolio liquidation problems.
Let us assume that orders arrive according to independent Poisson processes with smooth, price-
dependent rate functions λA (·, ·), ..., λH (·, ·) in the benchmark model n = 1 . Standard arguments
yield:
1
m(·, ·) = A .
λ (·, ·) + · · · + λH (·, ·)
To simplify the analysis we normalize the arrival rates so that m(·, ·) ≡ 1. In the n:th model we
scale the arrival rates of passive orders by ∆t1(n) and those of active orders by ∆x1(n) . The arrival
rates in the n:th model thus satisfy:

λI (·, ·)
λ(n),I (·, ·) = for I = A, B, E, F
∆x(n)
and
λI (·, ·)
λ(n),I (·, ·) = for I = C, D, G, H.
∆t(n)
 Horst and Paulsen: A Law of Large Numbers for LOBs
16 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

∆t(n)
Then, the probability of the next event being an active order is of the order ∆p(n) = ∆x(n)
and:
λI (·, ·)
pI (·, ·) = for I = A, B, E, F
λA (·, ·) + λB (·, ·) + λE (·, ·) + λF (·, ·)
λI (·, ·)
pI (·, ·) = C for I = C, D, G, H.
λ (·, ·) + λD (·, ·) + λG (·, ·) + λH (·, ·)

1.3.1. Price dynamics Of course, the spread

S(t) := a(t) − b(t)
should be non-negative at all times. This can easily be achieved if we require pA (·, ·) = pE (·, ·) = 0
for S(t) = 0. For most applications it would in fact be sufficient to assume that the price dynamics
depends on the best bid/ask price only through the spread. If the stationarity condition
λA − λB = λE − λF = 0
holds, then the spread is constant: S(t) ≡ S(0).
The following is a simple example where the spread is initially positive, then becomes zero, and
eventually opens again.
Example 1. Assume that the order arrival rates take the form
 A
 λ (b(t), a(t)) = λB (b(t), a(t))
p .
 E F
λ (b(t), a(t)) − λ (b(t), a(t)) = 2 a(t) − B0 , t ∈ [0, T ]
 2
T
Then, the bid price is constant, b(t) ≡ B0 , and a(t) = B0 + t − 2 for t ∈ [0, T ]; see Figure 4.

Figure 4. Bid/ask price dynamics of Example 1.

Next, we consider an example where the spread settles to a stationary level.

Example 2. Assume that the active order arrival rates take the form:
1+µ
λA (b(t), a(t)) := exp(−S(t)+ )
2
1−µ
λE (b(t), a(t)) := exp(−S(t)+ )
2 
1−µ 
λB (b(t), a(t)) := 1 − exp(−S(t)+ )
2
F 1 + µ +

λ (b(t), a(t)) := 1 − exp(−S(t) )
2
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 17

for some µ ∈ (0, 1) where x+ := max{0, x} denotes the positive part of x ∈ R. In particular, the
probability of spread placements increases with the spread, no spread placements occur if S(t) = 0,
and the limiting price dynamics is (cf. Figure 5):

db(t) 1 − µ + da(t) + 1+µ

= − e−S(t) and = e−S(t) − . (23)
dt 2 dt 2
The probability of a price increase, respectively, decrease can also be approximated in terms of
the arrival rates. If the next event is an active order arrival on the buy side, then

λA
lim P[B (n) (t) = B n (t−) − ∆x(n) ] = ,
n→∞ λA + λB
B
λ
lim P[B (n) (t) = B n (t−) + ∆x(n) ] = A .
n→∞ λ + λB
The limiting probabilities that the next price change is an increase/decrease of the best bid/ask
price as well as unconditional probabilities of price changes can be computed analogously. Clearly,
these probabilities increase in the respective rates. A further possible application of our model
includes estimations of the expected time-to-fill of a limit order in the original model. This question
has been studied by, e.g. Lo et al [23]. Our model yields an approximation for that time in terms
of the first time the limiting price process hits the placement price level of the limit order.

Figure 5. Bid/ask price dynamics of Example 3 for µ = 0.1; b(0) = 50; a(0) = 53.

1.3.2. Volume dynamics We now turn to the volume dynamics. When the limiting price is
constant, then the explicit solution to the bid-side volume density function given in (22) simplifies
to  
tBb (y) cb (y)  −tBb (y)

ub (t, y) = e vb,0 (y) + 1−e (24)
Bb (y)
where we write Bb (y) and cb (y) for Bb (γ(0), y) and cb (γ(0), y), respectively. From this we see that
the stationary solution is

cb (y) pD (γ(0))f D (y) pD f D (y)

ub (t, y) = − = C ≡ · . (25)
Bb (y) p (γ(0))f C (y) pC f C (y)
 Horst and Paulsen: A Law of Large Numbers for LOBs
18 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

In particular, the volume at the price level y increases in the order arrival rates and the “proba-
bilities” f D (y) with which an order is placed at that level when a placement occurs. Analogously,
the volume decreases in the cancellation rate and f C (y).
The following is a simple example where the expected volume dynamics is non-stationary but
can still be given in closed form.
Example 3. Assume that
λA (b(t), a(t)) − λB (b(t), a(t)) = λE (b(t), a(t)) − λF (b(t), a(t)) = 1.
Then b(t) = b0 + t and a(t) = b(t) + S(0). Assume moreover that b0 = 0 and that the buy side
passive order arrival rates depend only on the spread:
λC (b(t), a(t)) = λD (b(t), a(t)) = a(t) − b(t).
Thus, pC,D (γ(t)) ≡ pC,D . Let us further assume (ignoring the fact that our density functions need
to be defined on compact intervals for simplicity) that
1 e−x
f C (x) ≡ and f D (x) = .
pC pD
Then,
Bb (γ(t), y) ≡ −1 and cb (γ(t), y) = e−y .
We compute from the general solution formula:
ub (t, x) = e−t ub (0, x) + e−y (1 − e−t ).
2. Convergence of bid/ask prices According to Proposition 1, the process S (n) can be
(n)
represented in terms of a composition of a state process η (n) that jumps at deterministic times {tk }
and a time process µ(n) that accounts for the random arrival times. Prices change less frequently
(n)
at times {σk }. This suggests to introduce a second time scale - which will be referred to as active
order time - defined by
(n)
sk := k · ∆x(n)
(n)
along which to scale the price process. In order to make this more precise, let us denote by Dγ,k
(n)
the restriction of the operator Dk to the price component of the state sequence and put
(n)
X−1
σk
b (n)
D :=
(n)
Dγ,l = D
(n)
γ,k (n)
γ,σk
(n)
l=σk−1 +1

(n)
where the second equality follows from the fact that prices do not change between the times σk−1 +1
(n)
and σk − 1. Furthermore, we introduce the family of continuous time stochastic processes ηbγ(n)
defined by
(n) (n) (n)
ηb(n) (t) := ηbk for t ∈ [sk , sk+1 )
where

(n) (n) b (n) (b(n)


 ηbγ,k+1 := ηbγ,k + Dγ,k ηk )


(n)
! (26)
 (n) B0
ηb0 = .

 (n)
A0


(n)
The quantity ηbγ,k describes the state of the price process after the k:th price change. The following
lemma shows that the process ηγ(n) , evolving on the level of event time, and the process ηbγ(n) , evolving
on the level of active order time, are indistinguishable in the limit when n → ∞.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 19

Lemma 1. For any T > 0 and  > 0, it holds that

 
t/∆x(n) t/∆t(n)
X X
lim P  sup b (n) −
D
(n)
Dγ,k >  = 0.
γ,k
n→∞ 0≤t≤T
k=0 k=0

Pbt/∆x(n) c b (n) Pbt/∆t(n) c (n)

Proof. By construction, the two sums k=0 Dγ,k and k=0 Dγ,k have the same expected
value for any t ∈ [0, T ]:
bt/∆x(n) c bt/∆t(n) c
X (n)
X (n)
ED
b =
γ,k EDγ,k .
k=0 k=0

As a result, it is enough to prove that

   
bt/∆x(n) c n bt/∆t(n) c n
b (n) >   = lim P  sup 
X o X o
lim P  sup Db (n) − ED (n) (n)
Dγ,k − EDγ,k >  = 0.
γ,k γ,k
n→∞ 0≤t≤T
k=0
2 n→∞ 0≤t≤T
k=0
2

The random variables

1 n b (n) o
b (n) ,
Dγ,k − ED γ,k k = 0, ..., bT /∆x(n) c, n ∈ N
∆x(n)
and
1 n (n) (n)
o
Dγ,k − EDγ,k , k = 0, ..., bT /∆t(n) c, n ∈ N
∆t(n)
form triangular martingale difference arrays in the sense of Definition 2 (Appendix A) with respect
to the filtrations {Fσ(n) }k∈N and {Fk }k∈N , respectively. A direct computation shows that they are
k
uniformly L2 -bounded. Thus, it follows from Theorem 2 in Appendix A that for all β > 12 :
 
m n  β
1 X
b (n) ≥ 
b (n) − ED
o T
lim P  sup (n)
D γ,k γ,k (n)
=0
n→∞
0≤m≤b T c ∆x k=0
2 ∆x
∆x(n)

as well as  
m n  β
1 X (n) (n)
o  T
lim P  sup (n)
Dγ,k − EDγ,k ≥ (n)
 = 0.
n→∞
0≤m≤b T c ∆t k=0
2 ∆t
∆t(n)

Choosing β ∈ ( 21 , 1) and multiplying the inequalities in the above probabilities by ∆x(n) and ∆t(n) ,
respectively, proves the assertion. 
(n) (n)
Let γ
b be the solution to the ODE (19) and consider the discretization γ bk := γ b(sk ). The next
(n)
lemma shows that the sequence of expected price processes γ e defined by
(n) (n) (n)
e(n) (t) := γ
γ ek for t ∈ [sk , sk+1 )

where
 h i
(n) (n) b (n) (b (n)


γek+1 ek + E D
:= γ γ,k γk )


(n)
! (27)

 (n) B0
γ =
 e0 (n)

A0

converges uniformly to γ
b on compact time intervals. The proof is standard; we give it merely for
completeness.
 Horst and Paulsen: A Law of Large Numbers for LOBs
20 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

Lemma 2. For any T > 0

e(n) (t) − γ
sup |γ b(t)| = O(∆x(n) ).
t∈[0,T ]

Proof. Active orders change prices by one tick. Moreover, by Assumption 5, the conditional
probabilities of an active order being a market buy/sell order or limit buy/sell order placement in
the spread are independent of n ∈ N. Hence
!
B (n) A (n)
p (b
γ ) − p (b
γ )
h i
E D b (n) (b (n) (n)
· k k
γ,k γk ) = ∆x (n) (n)
pE (b
γk ) − pF (b γk )
  
(n) 1
= ∆x(n) · A(b γk ) . (28)
1

b(n) defines a special case of the classical Euler scheme for the ODE (15) and
Thus, the sequence γ
hence converges uniformly to its unique solution, see e.g. Hairer et al. [19, Theorem 7.3], with rate
∆x(n) . 
We are now ready to prove convergence in probability of the bid and ask prices.

Proof of Proposition 2.
a) We first consider the convergence of the state process ηγ(n) and claim that

lim sup |ηγ(n) (t) − γ

b(t)| → 0, in probability. (29)
n→∞ t∈[0,T ]

In view of Lemma 1, we can write

ηγ(n) (t) = ηbγ(n) (t)

bt/∆x(n) c  
X
=γ (n)
(0) + b (n) ηb(n)
Dγ,k γ,k
k=0
bt/∆x(n) c i bt/∆x (n)
X h
(n) (n)
X c  (n) (n) h i
=γ (n)
(0) + E Dγ,k (b
b ηγ,k ) + D
b (bη ) − E b (n) (b
D η
(n)
)
γ,k γ,k γ,k γ,k
k=0 k=0

up to some random additive constant that vanishes almost surely uniformly in t ∈ [0, T ] as n → ∞.
e(n) yields (again up to a vanishing additive constant):
Adding and subtracting the sequence γ

ηγ(n) (t) − γ e(n) (t) − γ

b(t) ≤ γ b(t)
bt/∆x(n) c h i
X
+ E D b (n) (b
η
(n)
) e(n) (t)
−γ
γ,k γ,k
k=0
bt/∆x(n) c  h i
X
+ b (n) (b
D η
(n)
) − E b (n) (b
D η
(n)
) .
γ,k γ,k γ,k γ,k
k=0

For the first term, we deduce from Lemma 2 that

e(n) (t) − γ
sup |γ b(t)| = O(∆t(n) ). (30)
t∈[0,T ]
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 21

For the second term we use the Lipschitz continuity of the event probabilities pI (·, ·) in order to
establish the existence of a constant Lγ > 0 such that:

bt/∆x(n) c h i bt/∆x(n) c h i h i
X X
b (n) (e
E D η
(n)
) e(n) (t) =
−γ E Db (n) (b
η
(n)
) − E b (n) (b
D γ
(n)
)
γ,k γ,k γ,k γ,k γ,k k
k=0 k=0
bt/∆x(n) c h i h i
X
≤ E Db (n) (b
η
(n)
) − E b (n) (b
D γ
(n)
)
γ,k γ,k γ,k k
k=0
bt/∆x(n) c
X (n) (n)
(n)
≤ ∆x · Lγ · ηbγ,k − γ
bk .
k=0

The third term corresponds to the noise-term of the price process. For each n ∈ N, the sequence
h i
b (n) (b
ykn := D η
(n)
) − E b (n) (b
D η
(n)
) , k = 0, ..., bT /∆x(n) c
γ,k γ,k γ,k γ,k

is a martingale difference sequence. A direct computation shows that

2
sup E|ykn |2 ≤ C · ∆x(n) .
n,k

Hence, the law of large numbers for triangular martingale difference arrays (Theorem 2 and Corol-
lary 1) implies " #
m
X
lim P sup | ykn | > 0 = 0,
n→∞
0≤m≤T /∆x(n) k=0

just as in the proof of Lemma 1. Thus, using Lemma 1 again, we see that
bt/∆x(n) c
X (n) (n)
ηγ(n) (t) − γ
b(t) = ηbγ(n) (t) − γ
b(t) ≤ ∆x (n)
· Lγ · ηbγ,k − γ
bk + o(1) in probability
k=0

for some additive term of order o(1) uniform in t ∈ [0, T ]. As a result, (29) follows from an appli-
cation of Gronwall’s lemma along with Lemma 1.
b) Let us now consider the cumulative “active order time process”

bt/∆x(n) c 
X (n)
(n)
y (t) = ζ (n) ηbγ,k · ∆x(n)
k=0  
(n) (n)
bt/∆x
X c h  i bt/∆x
X c   h  i
(n) (n) (n)
= ∆x(n) · E ζ (n) ηbγ,k + ζ (n) ηbγ,k − E ζ (n) ηbγ,k
 
k=0 k=0
(n)
bt/∆x c   bt/∆x(n) c    h  i
X (n)
X (n) (n)
(n) (n)
= ∆x m ηbγ,k + ∆x · ζ (n) ηbγ,k − E ζ (n) ηbγ,k .
k=0 k=0
(31)

By the above established uniform convergence of ηbγ(n) to γ

b in probability and because the function
m is Lipschitz continuous, the first sum converges to the function
Z t
y(t) = m(b
γ (u))du. (32)
0
 Horst and Paulsen: A Law of Large Numbers for LOBs
22 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

Applying the same arguments as above to the martingale difference sequences


(n)
 h
(n)

(n)
i t
ζ (n) ηbγ,k − E ζk ηbγ,k , k = 0, ..., b c
∆x(n)

we see that the second term vanishes uniformly in t ∈ [0, T ] in probability. Thus,

lim sup |y (n) (t) − y(t)| = 0 in probability.

n→∞ t∈[0,T ]

Since y (n) and y are increasing functions, their inverses µ(n) and µ exist. By continuity

sup |µ(n) (t) − µ(t)| → 0 in probability as n → ∞

t∈[0,T ]

and

0 1 1 1
µ0 (t) = y −1 (t) = = =
. (33)
y 0 (y −1 (t)) m(µ(t)) m γ
b(µ(t))

Since both the state and the time process converge, we conclude from the time change theorem
that
sup ηγ(n) (t) − γ(t) → 0 in probability n → ∞,
t∈[0,T ]

where γ(t) = γ
b(µ(t)) and
   
0 0 A(b
0 γ (µ(t))) 1 A(γ(t)) 1
b (µ(t)) · µ (t) =
γ (t) = γ = .
m(b
γ (µ(t))) 1 m(γ(t)) 1

3. Convergence of volume densities In this section we prove Proposition 3. To this end,

(n) (n) (n)
we denote by Dv,k (·, ·) the restriction of the operator Dk to L2 × L2 , i.e the restriction of Dk
to the volume components of the state process. We need to show that the sequence {ηv(n) }n∈N of
L2 × L2 -valued step-functions defined recursively by

(n) (n) (n)

ηv(n) (t, ·) := ηv,k for t ∈ [tk , tk+1 ) (34)

where
  
(n) (n) (n) (n) (n)
ηv,k+1

 := ηk + Dv,k ηγ,k , ηv,k
(35)

η (n)

:=
(n)
v0
v,0

converges in probability in L2 to the unique solution of the PDE (20). We will show convergence in
several steps. In a first step, we find a convergent discretization scheme of the PDE that is coherent
with the order book dynamics. Subsequently, we link this scheme to the expected dynamics of the
volume densities.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 23

3.1. A numerical scheme for the limiting PDE For any n ∈ N, the scaling parameters
(n) (n) (n)
∆x(n) and ∆t(n) define a grid {(tk , xk )} on [0, T ] × R through tk = k · ∆t(n) (k ∈ N0 ) and
(n)
xj = j · ∆x(n) (j ∈ Z). In a first step, we approximate the unique solution u b : [0, T ] × R → R2 to
(20) by a sequence of grid-point functions u b(n) : [0, T ] × R → R2 . To this end, we put
 B   A 
p (γ̂(t)) 0 p (γ̂(t)) 0
AR (t) := , AL (t) :=
0 pE (γ̂(t)) 0 pF (γ̂(t))

and
F (t, x) := B(γ̂(t), x), g(t, x) := c(γ̂(t), x).
(n)
Furthermore, we introduce operators Ht that act on v ∈ L2 according to
(n)
Ht (v) := v + ∆p(n) · AR (t) v(· + ∆x(n) ) − v(·) 
 

+ ∆p(n) · AL (t) v(· − ∆x(n) ) − v(·)

+ ∆v (n) · (1 − ∆p(n) ) · [F (t, ·) · v(·) + g(t, ·)] .

The sequence of grid-point approximations is then defined recursively by

(n) (n (n)
b(n) (t, ·) := u
u bk for t ∈ [tk , tk+1 ) (36)

where

(n) (n) (n)
u

 bk+1 = H (n) (b
uk )
tk
(37)


u (n) (n)
b0 = v0 .

The sequence of step-functions {u b(n) } essentially describes a discretized limiting volume dynamics
of the order book. We benchmark this dynamics against the expected pre-limit volume dynamics
when prices are replaced by their limiting dynamics. More precisely, we introduce another sequence
of step functions u(n) : [0, T ] × R → R by
(n) (n (n)
u(n) (t, ·) := uk for t ∈ [tk , tk+1 ) (38)

where
 h  i
(n) (n) (n) (n) (n)
uk+1

 = uk + E Dv,k γ̂(tk ), uk
(39)

u(n)

:=
(n)
v0 .
0

b(n) approximate the solution

In a first step, we are now going to show that the grid-point functions u
u
b of our PDE. Subsequently, we show that the PDE can as well be approximated by the functions
u(n) .
Proposition 5 (Convergence of the numerical scheme). Assume that the assumptions
of Theorem 1 hold. Then, the processes ub(n) define a convergent finite difference scheme of the PDE
(20), i.e.
sup kub(n) (t, ·) − u
b(t, ·)kL2 → 0 as n → ∞. (40)
t∈[0,T ]
 Horst and Paulsen: A Law of Large Numbers for LOBs
24 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

Proof. For t ∈ [0, T ] the local truncation error associated with the grid-point approximations u b(n)
is defined as
1  (n)

L(n) (t, x) := u (t + ∆t (n)
, x) − Ht (b
u (t, · ))(x) . (41)
∆t(n)
b

Smoothness of the solution u b (bounded with uniformly bounded first and second order derivatives)
along with Assumption 6 implies that the following estimate holds uniformly in t ∈ [0, T ] and x ∈ R:
1 
L(n) (t, x) = ub(t + ∆t(n) , x) − u b(t, x) − ∆p(n) · AR (t)(b u(t, x + ∆x(n) ) − u b(t, x))
∆t(n) (n) (n)
− ∆p · AL (t)(b u(t, x − ∆x ) − u b(t, x)) 
(n) (n)
− ∆v · (1 − ∆p ) · [F (t, x)b u(t, x) − g(t, x)]
1 
= (n) u bt (t, x)∆t(n) + o(∆t(n) ) − ∆p(n) · AR (b ux (t, x)∆x(n) + o(∆x(n) ))
∆t
− ∆p(n) · AL (−u bx (t, x))∆x(n) + o(∆x(n) )) 
− ∆v (n) · (1 − ∆p(n) ) · [F (t, x)b u(t, x) − g(t, x)]
=ubt (x, t) − (AL (t) − AR (t)) u bx (x, t) − F (x, t)bu(x, t) − g(x, t) + o(1)
=ubt (t, x) − A(t)b ux (t, x) − B(t, x)bu(t, x) − c(t, x) + o(1)
= o(1) (42)

b(t; ·) to the PDE vanishes outside a compact interval for all t ∈ [0, T ] (Proposition
The solution u
4) and hence so does L(n) (t, ·). As a result, we also have that

lim sup kL(n) (t, ·)kL2 = 0. (43)

n→∞ t∈[0,T ]

(n)
b(t + ∆t(n) , ·) = Ht (b
From (41), one has that u u(t, ·)) + ∆t(n) L(n) (t, ·). In terms of the error-
function
u(n) (t, x) := u
δb b(n) (t, x) − u
b(t, x) (44)
this yields
(n) (n) (n)
u(n) (tk+1 ; ·) = u
δb b(n) (tk+1 ; ·) − ub(tk + ∆t(n) ; ·)
(n) (n) (n) (n) (n)
= Hk (b u(n) (tk , ·)) − Hk (bu(tk , ·)) − ∆t(n) L(n) (tk , ·)
(n) (n) (n) (n) (n) (n)
= Hk (δb u (tk , ·)) − ∆t L (tk , ·),
(n) (n) (n)
due to the linearity of Hk . Using this property iteratively, and putting (H(n) )k := H1 ◦ · · · ◦ Hk
one finds:
k
X
(n) (n)
δbu(tk+1 , ·) = (H(n) )k (δb
u(n) (0, ·)) − ∆t(n) (H(n) )k−i (L(n) (ti ), ·). (45)
i=0
(n)
From the definition of Ht together with the fact that the functions F and g are uniformly bounded
by assumption, one finds (for large enough n ∈ N):
(n) (n)
kHt (v)kL2 ≤ kv kL2 1 − ∆p(n) (AR (t) + AS (t)) + kT+ (v)kL2 · ∆p(n) · AR (t)



(n)
+kT− (v)kL2 · ∆p(n) · AL (t) + ∆v (n) (kF k∞ kv kL2 + kg k∞ ) .

(n)
Thus, by the isometry property kT± (v)kL2 = kv kL2 of the translation operators, there exists a
constant C > 0 that is independent of t ∈ [0, T ] such that:
(n)
sup kHt (v)kL2 ≤ 1 + C∆t(n) .
kvkL2 =1
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 25

In particular, since k ≤ b ∆tT(n) c:

(n)
sup (H(n) )k (δb
u(n) (0, ·)) L2
≤ (1 + C∆t(n) )bT /∆t c
u(n) (0, ·)kL2
kδb
k
≤ eCT kδb
u(n) (0, ·)kL2
= o(1),

where the last equality follows from Assumption 1. Similarly, from (43):
(n)
sup (H(n) )k−i (L(n) (ti−1 ), ·) = o(1).
k,i,n L2

Using the same arguments as before, we conclude that the sum in (45) vanishes uniformly in time.
Hence,

b(n) (t, ·) − u
lim sup ku b(t, ·)kL2 = 0.
n→∞ t∈[0,T ]


Next, we show that the functions u(n) also approximate the PDE. More precisely, the following
holds.
Proposition 6. Under the assumptions of Theorem 1

b(n) (t; ·) − u(n) (t; ·)kL2 = 0.

lim sup ku
n→∞ t∈[0,T ]

(n)
Proof. The proof is similar to that of Proposition 5. By analogy to the operator Ht we introduce
bt(n) on L2 by
an operator H
bt(n) (v) := v + ∆p(n) · AR (t) v(· + ∆x(n) ) − v(·)
 
H
+ ∆p(n) · AL (t) v(· −∆x(n) ) − v(·)
 

+ ∆v (n) · (1 − ∆p(n) ) F (n) (t, ·) · v(·) + g (n) (t, ·)



(n) (n)
where for t ∈ [tk , tk+1 ):
!
(n)
−f (n),C (·) · pC (bγ (tk )) 0
F (n) (t, ·) := (n),G (n)
0 −f (·) · pG (b
γ (tk ))
!
(n)
f (n),D (·) · pD (b
γ (tk ))
g (n) (t, ·) := (n) .
f (n),H (·) · pH (b
γ (tk ))
(n) (n) (n)
For the error function δbuk (·) := u bk (·) − uk (·) we then obtain:
 
(n) bt(n) (δb(n) (n)
δbuk+1 (·) = H vk (·)) + ∆v (n) · (1 − ∆p(n) ) · δF (n) (·) · u
bk (·) + δg (n) (·)
bt(n) (δb
=: H
(n) (n)
u (·)) + ∆v (n) · (1 − ∆p(n) ) · Lb(t ; ·)
k k

where
!
(n)
−(f (n),C (·) − f C (·)) · pC (b
γ (tk )) 0
δF (n) (t, ·) := (n),G (n) ,
0 −(f (·) − f (·)) · pG (b
G
γ (tk ))
!

(n)
f (n),D (·) − f D (·) · pD (b
γ (tk ))
δg (n) (t, ·) :=
(n) .
f (n),H (·) − f H (·) · pH (bγ (tk ))
 Horst and Paulsen: A Law of Large Numbers for LOBs
26 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

b(n) are uniformly bounded. As a result, it follows from

By construction, the grid-point functions u
Assumption 2 that
(n)
lim sup kLb(tk ; ·)kL2 = 0.
n→∞
k=0,...,bT /∆t(n) c

One can now proceed as in the proof of Proposition 5, to conclude that

b(n) (t, ·) − u(n) (t; ·)kL2 = 0.

lim sup ku
n→∞ t∈[0,T ]

3.2. Expected volume dynamics and discretized PDEs To show the convergence of the
volume density functions we compare the random states ηv(n) with the deterministic approximations
of the limiting PDE obtained in the previous subsection. For this, we introduce the deterministic
e(n) :
step function valued processes u
(n) (n (n)
e(n) (t, ·) := u
u ek for t ∈ [tk , tk+1 ) (46)

where
 h  i
(n) (n) (n) (n) (n)
u

 ek+1 ek + E Dv,k ηγ,k , u
:= u ek
. (47)


u (n) (n)
e0 := v0 .

It describes the expected dynamics of the volume density functions for the actual price process; in
e(n) is a stochastic process. By contrast, the process u(n) describes the dynamics of the
particular, u
expected volume density functions when the random evolution of bid and ask prices is replaced by
its deterministic limit. We have:

b(t, ·)kL2 ≤ kηv(n) (t, ·) − u

kηv(n) (t, ·) − u e(n) (t, ·) − u(n) (t, ·)kL2
e(n) (t, ·)kL2 + ku
(n) (n)
+ku (t, ·) − u b (t, ·)k + ku b(n) (t, ·) − u
b(t, ·)kL2 .

The last two terms are deterministic and converges uniformly to zero by Propositions 5 and 6. It
remains to show convergence of the first two (random) terms. This will be achieved in the following
two subsections.

3.2.1. Estimating the price impact of expected volume dynamics The term
e(n) (t, ·) − u(n) (t, ·)kL2 measures the impact of the noise in the price process on the expected stand-
ku
ing volume. The following proposition shows that it converges to zero in probability, uniformly
over compact time intervals.
Proposition 7. Under the assumptions of Theorem 1 it holds that:

e(n) (t, ·) − u(n) (t, ·)kL2 → 0

sup ku in probability as n → ∞.
t∈[0,T ]

Proof. We argue again as in the proof of Proposition 5. Analogously to the operator H b (n) defined
(n) (n)
in the proof of Proposition 6 we define for t ∈ [tk , tk+1 ) the operator

e (n) (v) := v + ∆p(n) · Ae(n) (t) v(· + ∆x(n) ) − v(·)

 
H k R
+ ∆p(n) · A eL (t) v(· − ∆x(n) ) − v(·)
 
h i
+ ∆v (n) · (1 − ∆p(n) ) Fe(n) (t, ·) · v(·) + ge(n) (t, ·)
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 27

(n) (n)
where for t ∈ [tk , tk+1 ) we put
!
(n)
e(n) (t) := pA (ηγ,k ) 0
AR E (n) .
0 p (ηγ,k )

The functions A e(n) (t), Fe(n) (t, ·) and ge(n) (t, ·) are defined analogously. Let us further put δp(n),I :=
L k
(n) (n)
pI (γ̂(tk )) − pI (ηγ,k ) for I = A, B, E, F and
! !
(n),A (n),B
(n) δpk 0 (n) δpk 0
δAR,k := (n),E , δAL.k := (n),F
0 δpk 0 δpk

(n) (n)
and denote by δfk and δgk the corresponding quantities for cancelations and limit order place-
(n) (n) (n) (n)
uk (·) := u
ments. Then, the error function δe ek (·) − uk (·) satisfies δe
u0 = 0 and can be represented
(n)
in terms of the operator H
e as follows:
k

(n) e (n) (δe(n) (n) (n)

uk+1 (·) = H
δe k uk (·)) + ∆t(n) · Lek (tk , ·)

where

(n) (n) ∆p(n) (n)

h
(n) (n) (n)
i
Lek (tk , ·) := · δA u ( · + ∆x ) − u (· )
∆t(n)(n) R,k hk k
∆p (n) (n) (n) (n)
i
+ · δA L,k u ( · − ∆x ) − u ( ·)
∆t(n) h k k
i
(n) (n) (n) (n)
+ (1 − ∆p ) · δf · uk (·) + δgk .

Corollary 2 establishes
(n) (n) (n) (n)
kuk kL2 ≤ L and kT± uk − uk kL2 ≤ L · ∆x(n)

for some constant L < ∞ that is independent of (n, k). Using our assumptions on the placement,
cancelation and event probability functions along with the fact that the functions pI have bounded
gradients and that
(n) (n)
lim sup |ηγ,k − γ
b(tk )| → 0 in probability,
n→∞
k=0,...,bT /∆t(n) c

this implies
(n) (n)
lim sup kLek (tk , ·)kL2 = 0 in probability.
n→∞
k=0,...,bT /∆t(n) c

We can, therefore, argue as in the proof of Proposition 5 to conclude. 

3.2.2. Convergence of volumes to their expected values In this subsection we apply

a law of large number for Hilbert space-valued triangular martingale difference arrays (TMDAs)
in order to establish the missing convergence to zero of the distance between ηv(n) and u
e(n) . More
precisely, our goal is to prove the following result.
Proposition 8. Suppose the assumption of Theorem 1 hold. Then,

sup kηv(n) (t, ·) − u

e(n) (t, ·)kL2 → 0 in probability as n → ∞.
t∈[0,T ]
 Horst and Paulsen: A Law of Large Numbers for LOBs
28 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

(n) (n)
Proof. Using the definition of ηv,k in (35) and u
ek in (46) we see that
(n) (n)
u
ek = Eηv,k ,

conditioned on the price process. As a result,

bt/∆t(n) c  
X (n) (n) (n) (n)
kηv(n) (t, ·) − u
e(n) (t, ·)kL2 =k Dv,k (·, ·) − E[Dv,k (ηγ,k , ηv,k )] kL2 .
k=0

In order to establish convergence of the sum to zero uniformly in time we introduce the L2 -valued
triangular martingale-difference-array
(n) (n) (n) (n)
Ykn := Dv,k (·, ·) − E[Dv,k (ηγ,k , ηv,k )]. (48)
1
If we can show that there exists β > 2
such that
 2β
1
sup E[kYkn k2L2 ] < ∞, (49)
n,k ∆t(n)

then Theorem 2 and Corollary 1 would guarantee that

" m
#
X
n
lim P sup k Yk kL2 >  = 0, (50)
n→∞
0≤m≤bT /∆t(n) c k=0

and the proposition would be proved. To establish (49), we need to bound the following terms:
"  2 #
∆v (n) (n),D,H ∆v (n) (n),D,H

sup E 1Dk ,Hk M − E 1Dk ,Hk M
∆x(n) v,k ∆x(n) v,k
"n,k L2
 2 #
(n) (n)

∆v (n),C,G (n) ∆v (n),C,G (n)
sup E 1Ck ,Gk (n)
Mv,k ηv,k − E 1Ck ,Gk (n)
Mv,k ηv,k
∆x ∆x
n,k     h   
L2
i 2 
(n) (n) (n) (n) (n) (n)
sup E 1Ak ,Bk T± ηvb ,k − ηvb ,k − E 1Ak ,Bk T± ηvb ,k − ηvb ,k
n,k L2
     h    i 2 
(n) (n) (n) (n) (n) (n)
sup E 1Ek ,Fk T± ηvs ,k − ηvs ,k − E 1Ek ,Fk T± ηvs ,k − ηvs ,k .
n,k L2

This is done in Lemma 7 in the appendix. In particular, this lemma shows that the place-
ment/cancellation and shift terms are of the order

O (∆v (n) )2 and O (∆v (n) )2α + (∆v (n) )2−α ,

respectively. Since α ∈ (1/2, 1) the assertion follows for β := min{α, 1 − α/2} . 

4. Application to portfolio liquidation We now discuss how our LOB model could be used
to obtain endogenous shape functions for models of optimal portfolio liquidation under market
impact. In such models the goal is to find trading strategies that unwind a large number X > 0 of
shares within a pre-specified time window [0, T ] at minimal cost. It is typically assumed that prices
are continuous (as in our limiting model), and that the expected distribution of the standing buy
(or sell) side volume can be described in terms of a shape function

f : [0, T ] × R → R+ .
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 29

The benchmark case studied in the seminal paper of Almgren and Chriss [3] corresponds to a
block-shaped order book where f (t, x) ≡ δ, for some δ > 0; discrete-time liquidation problems with
more general shape functions have been studied in, e.g. [2]. For a given shape function a sell order
of size Et submitted at time t ∈ [0, T ] moves the best bid price by an amount Dt defined through
Z Dt
Et = f (t, x) dx.
0

Let us denote by F −1 (t, ·) the inverse of the anti-derivative of the shape function and assume for
simplicity that there is no permanent price impact. If the cost of trading is benchmarked against
the mid quote, then the cost c(t, Et ) of trading Et shares is half the spread plus the market impact
cost (see [2] and references therein for details):
Z F −1 (t,Et )
1
c(t, Et ) = S(t)Et + xf (t, x)dx. (51)
2 0

If orders can be submitted at discrete points in time tn (n = 1, ..., N ), the resulting optimization
problem is given by:
N
X N
X
min c(tn , Etn ) s.t. Etn = X. (52)
(Etn )N
n=1 i=1 n=1

The goal is now to obtain dynamic shape functions from order book data. Empirical placement
(n),I
and cancelation densities fi (i = 0, 1, 2, ...) for the visible book, average volumes placed/cancelled
and active/passive order arrival rates can be estimated from flow data; examples are given in the
next section. Smooth approximations f I and pI of the empirical density and probability functions
can then be obtained by interpolation. The case of exponential densities is particularly transparent
as illustrated by the following example.
(n),C (n),D
Example 4. Assume that we are given empirical Poisson arrival rates ηi and ηi for the
price levels i · ∆x(n) (i = 0, 1, 2, ...) that satisfy
(n),C (n),D
ηi+1 −κC ∆x(n) ηi+1 D
∆x(n)
(n),C
=e and (n),D
= e−κ
ηi ηi
for some constants κC,D > 0. Then, the functions f C,D satisfy:
C,D
f C,D (x) ∝ e−κ x
.

Assuming that κD > κC (as it is the case for, e.g. Ebay and Facebook; see Table 2 below), the
stationary solution is of the form ub (t, x) = κ1 e−κ2 x for κ1 , κ2 > 0 and
Z ∞
κ1
ub (t, x) dx = =: κ.
0 κ2
Calibrating the shadow book is more difficult. While an approximation of the shadow book
should in principle be possible from spread placements, there are several challenges. For instance,
one would have to identify orders that genuinely provide liquidity, that is, to clean the data of
spread placements that are cancelled after very short periods of time. In particular, one would have
to identify and eliminate “ping-orders”, i.e. orders sent to detect hidden liquidity in the spread.
Subsequently, one would have to estimate average spread placements (as a function of the best
bid/ask price or spread). One possibility to bypass this problem is to work with the stationary
solution as given in (25). In that case the density functions are only required on the positive half
 Horst and Paulsen: A Law of Large Numbers for LOBs
30 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

line. This is in fact often done indirectly in portfolio liquidation models when one assumes that
the (benchmark) price is a martingale and can hence be treated as a constant in the optimization
problem, and the shape function is independent of time. Another possibility is to consider short
periods of time over which no price changes occur as in [16]. In any case, assuming that the density
functions f I have been constructed along with the probabilities pI , the resulting optimization
problem reads:
N Z U −1 (tn ,Etn ) ! N
X 1 b X
min S(tn )Etn + xub (tn , x)dx s.t. Etn = X
(Etn )
n=1
2 0 n=1

where Ub−1 (t; ·) denotes the inverse of the anti-derivative of the volume density function ub (t; ·). In
the framework of Example 4 it is given by:

1 y
Ub−1 (t, y) = − ln(1 − ).
κ2 κ

5. Some calibration results In this section we illustrate how our model parameters, espe-
cially the densities f I and the probabilities pI can be estimated from flow data. For the reasons
outlined above we restrict ourselves to the visible book, i.e. we consider short periods of time over
which no price changes occurred. Using LOBSTER9 data for Jan 2, 2014 we computed empirical
buy-side placement and cancellation probabilities in between two consecutive price changes over
that day for the first 5 ticks (best bid price and four ticks below the best bid) for Apple (AAPL),
Ebay (EBAY), Facebook (FB), Kraft Foods (KRFT) and Microsoft (MSFT).

5.1. Placement and cancellation densities Table 1 reports average numbers of passive
orders in between two consecutive price changes, the empirical probabilities p(n),D of a passive
(n),D
buy-side order being a placement, and the empirical placement and cancellation densities fi
(n),C
and fi for i = 0, 1, 2, 3, 4 for AAPL, EBAY, FB, KRFT and MSFT. For instance, there were
on average 459 passive orders for MFST (corresponding to ∆p = 0.002) and 62 passive orders in
between two consecutive active orders for AAPL (corresponding to ∆p = 0.016). The empirical
probability of a passive APPL order to be a placement was 0.52; the conditional probability that
a placement (at the first five ticks) took place at the top of the book was 0.23. In fact, for APPL
the empirical placement and cancellation densities are essentially constant. By contrast, for MSFT
almost all of the activity concentrates at the top of the book. For stocks such as MSFT a continuous
approximation does not seem appropriate to us.

Stock Events Bid 0 Bid 1 Bid 2 Bid 3 Bid 4 Bid 0 Bid 1 Bid 2 Bid 3 Bid 4 p(n),D
AAPL 62.0 0.23 0.19 0.2 0.19 0.19 0.21 0.2 0.2 0.20 0.19 0.52
EBAY 33.29 0.56 0.15 0.1 0.1 0.09 0.52 0.2 0.1 0.09 0.09 0.53
FB 21.51 0.51 0.22 0.11 0.09 0.07 0.47 0.28 0.1 0.08 0.07 0.52
MFST 458.98 0.81 0.07 0.04 0.05 0.03 0.79 0.1 0.04 0.04 0.03 0.52
KRFT 16.39 0.32 0.17 0.13 0.16 0.22 0.31 0.21 0.15 0.16 0.17 0.56
Table 1. Empirical buy-side placement (left) and cancellation (right) probabilities between price changes

9
LOBSTER is an online limit order book data tool, giving access to flow and reconstructed limit order book data for
the entire universe of NASDAQ traded stocks. We thank Nikolaus Hautsch for data provision and Gökhan Cebiroglu
for assistance with the estimation results.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 31

We approximated f (n),C and f (n),D for EBAY, FB and KRFT using smooth functions f C,D . For
EBAY and FB exponential densities of the form

f C,D (x) = aebx + c

provided very good fits. For KRFT we chose quadratic polynomials:

f C,D (x) = ax2 + bx + c.

The fitted parameter values are reported in Table 2. Figure 6 displays the empirical and the
theoretical cancellation densities; read bars correspond to empirical densities; blue bars correspond
to the values of the integrals of the theoretical (fitted) densities over the respective price bins as
in (6) (for E[ω1I ] = 1).

a b c
EBAY submission 1.0921 -2.0285 0.092923
EBAY cancellation 0.81613 -1.3455 0.078672
FB submission 0.72488 -1.0779 0.065857
FB cancellation 0.64554 -0.72558 0.025527
KRFT submission 0.033273 -0.19045 0.39952
KRFT cancellation 0.020006 0.13456 0.37005
Table 2. Theoretical placement and cancellation densities

5.2. Volume-weighted placement densities and cancellation ratios Our model

assumes that order sizes are random but do not depend on the submission level. Such a dependence
can easily be incorporated into the model if we assume that f D models volume-weighted placement
densities. Table 3 reports average buy-side submission and cancellation volumes. Empirical and
theoretical volume-weighted placement density functions for KRFT are displayed in Figure 7.

Stock Bid 0 Bid 1 Bid 2 Bid 3 Bid 4 Bid 0 Bid1 Bid 2 Bid 3 Bid 4
AAPL 167 161 178 190 188 121 123 112 158 179
EBAY 2341 566 455 557 354 2322 566 465 557 354
FB 2122 1894 682 360 240 2236 1435 527 346 207
KRFT 573 420 354 490 558 504 267 176 307 500
Table 3. Average buy-side cancellation (left) and submission (right) volumes.

Estimating cancellation ratios is more difficult. They can only be estimated by either recon-
structing the full book from flow data or by tracking each individual order until cancellation or
execution. While LOBSTER provides flow data, fully reconstructed books are not readily avail-
able. Using NASDAQ ITCH order-message data for the period ranging from January 2011 to April
2011, Cebiroglu and Horst [9] estimated cancellation ratios at the top of the book for a random
selection of 31 stocks from the S&P 500. They report cross-sectional top-of-the-book cancella-
tion ratios ranging from 0.16 to 0.28, depending on the stock’s liquidity. For APPL, EBAY and
MSFT they estimated the cancellation ratios at 0.46, 0.12 and 0.11, respectively. Gao et al. [16]
report price-dependent cancellation ratios for the stock Bank of America (BAC). In that paper
the authors model the liquidity at a particular price tick as a birth-death process: orders arrive
at price-dependent rates relative to the best bid/ask price, and each order is cancelled after an
 Horst and Paulsen: A Law of Large Numbers for LOBs
32 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

Figure 6. Empirical and theoretical cancellation densities for EBAY (top), FB (center), and KRFT (bottom). Left:
discrete empirical (red) and theoretical (blue) density; right: distribution function associated with the continuous
density f C

exponentially distributed waiting time. Average cancellation rates per second, ΘnA (i), at the possi-
ble ticks i = 0, 1, 2, ..., are estimated by tracking individual orders until cancellation or execution.
In distribution, the approach of canceling each order independently after an exponential waiting
time is equivalent to proportional cancellation. Hence, the limiting function ΘA in [16] corresponds
C C
to our function p mf . The calibrations in [16] are based on an empirical analysis of message-level
order book data from NYSE Arca in August 2010. For BAC the authors present general summary
statistics including the number of buy-side and sell-side events, market order arrival rates, average
limit order sizes, and order arrival rates and cancellation ratios for time windows of 5, 10, 15 and
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 33

Figure 7. Empirical and theoretical volume-weighted placement densities for KRFT; left: discrete empirical (red)
and theoretical (blue) volume-weighted density; right: continuous volume-weighted PDF.

20 minutes starting at 12:45pm on August 5, 2010. Their theoretical buy-side cancellation ratios
at the first 10 price ticks are given in Table 4.

Bid 0 Bid 1 Bid 2 Bid 3 Bid 4 Bid 5 Bid 6 Bid 7 Bid 8 Bid 9
0.562 0.1 0.151 0.207 0.217 0.11 0 0.004 0.004 0.003
Table 4. Empirical cancellation ratios for BAC from [16]

6. Conclusion In this work a law of large numbers for limit order books was established.
Starting from order arrival and cancelation rates for all price levels, we showed that the LOB
dynamics can be described by a coupled PDE:ODE system when tick and order sizes tend to
zero while arrival rates tend to infinity in a particular way. A key insight is that the scaling limit
requires two time scales: a fast time scale for passive order arrivals and a comparably slow time
scale for active order arrivals. The proof of convergence of volume densities was carried out in
three steps: We first showed that the expected LOB dynamics resembles a numerical scheme for
hyperbolic PDEs plus noise, provided the random price dynamics is replaced by its deterministic
limit. Subsequently, we showed that the impact of the noise in the price process on the volume
dynamics vanishes in the limit. Finally, we used a law of large numbers for triangular martingale
difference arrays to prove that the LOB model converges to its expected value.
Our model allows for approximation of key order book statistics such as expected price incre-
ments, expected standing volumes at future times and expected times to fill. We calibrated place-
ment and cancellation densities to market data for selected stocks and provided numerical simu-
lations that suggest that our model can indeed be used to forecast order book shapes over short
periods of time.
Several questions remain open. First, it would be interesting to establish a CLT or, more gener-
ally, a diffusion approximation for LOBs. Based on the idea of having different time scales for active
and passive order arrivals, Bayer, Horst and Qiu [5] have recently established a first SPDE scaling
limit for order books. However, they assume that cancellations are subject to additive (rather than
multiplicative) noise so volumes may become negative. Second, it would be interesting to solve
models of optimal portfolio liquidation based on our limiting model.
 Horst and Paulsen: A Law of Large Numbers for LOBs
34 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

Appendix A: A Law of Large Numbers for Banach-Space-Valued TMDAs In this

appendix we prove a law of large numbers for triangular martingale difference arrays taking values
in real separable p-uniformly smooth Banach spaces. A Banach space E is called p-uniformly
smooth, where p ∈ (1, 2], if
 
x+y x−y
ρE (τ ) = sup k k+k k − 1 : kxk = 1, ky k = τ = O(τ p ).
2 2

All Hilbert spaces are 2-uniformly smooth by the parallelogram identity. The spaces C, l1 and L1
are not uniformly smooth.
Definition 2. A family of random variables ykn , k = 1, ..., n, n = 1, 2, ... defined on some prob-
ability space (Ω, F , P) is called a triangular martingale difference array (TMDA) with respect to
a family {F n }n=1,2,... of filtrations, F n = {Fkn }nk=0 , if for all n = 1, 2, ... the sequence y1n , ..., ynn is a
F n -martingale difference sequence (MDS), i.e.

E ykn |Fk−1
n
 
= 0.

If {ykn } is a TMDA, then for all n = 1, 2, ... one has

Xk
E[ ykn |Fk−1
n
] = ykn ,
j=1

that is, partial sums are martingale. For such martingales, Pisier [26] proved the following moment
estimate.
Lemma 3. Let E be a real separable p-uniformly smooth Banach space (1 ≤ p ≤ 2). Then, for
all r ≥ 1 there exists a constant C > 0 such that for all martingales
( n !)
X
Xi , Gn
i=1 n≥1

with values in E, we have

 r ∞
!r/p
X
p
E sup |Xn | ≤ CE |Xi − Xi−1 | .
n≥1
n=1

The previous lemma allows us to prove the following weak law of large numbers for triangular
martingale difference arrays.
Theorem 2. Let ykn , k = 1, ..., n, n = 1, 2, ... be a TMDA taking values in a real separable p-
uniformly smooth Banach space E for 1 ≤ p ≤ 2 such that

sup E|ykn |p < ∞.

n,k

Then, for all β > 0 such that β · p > 1 one has for all  > 0 that:
" m
#
X
lim P sup | ykn | ≥  · nβ = 0.
n→∞ 1≤m≤n
k=1
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 35

Proof. By Markov’s inequality

" m
# " m
#q
X
n β 1 X
P sup | yk | ≥  · n ≤ E sup | ykn |
1≤m≤n
k=1
nq·β 1≤m≤n
k=1

for all q ≥ 1. Thus, it follows from Lemma 3 that

" m
# " n #q/p
X X
n β −β·q n p
P sup | yk | ≥  · n ≤ Cn E |yk |
1≤m≤n
k=1 k=1
q
≤ Cn−β·q+ p

for a generic constant C > 0 since the random variables ykn have a uniformly bounded p:th moment.
Hence, the assertion follows as soon as −β · q + pq < 0. This holds for all q > 0 as β · p > 1. 
As an immediate corollary from the preceding theorem one obtains the following law of large
numbers for TMDAs.
Corollary 1. Let ykn , k = 1, ..., n, n = 1, 2, ... be a TMDA taking values in a real separable
2-uniformly smooth Banach space E such that
 
sup n2β E|ykn |2 < ∞
n,k

for some β > 21 . Then,

m
X
lim sup | ykn | = 0 in probability.
n→∞ 1≤m≤n
k=1

Proof. We apply Theorem 2 to the TMDA

ŷkn := nβ ykn .

Then " #
m
X
lim P sup | ŷkn | ≥·n β
= 0.
n→∞ 1≤m≤n
k=1

Hence the assertion follows from:

" m
# " m
#
X X
P sup | ŷkn | ≥  · nβ = P sup | ykn | ≥  = 0.
1≤m≤n 1≤m≤n
k=1 k=1

Appendix B: Properties of volume density functions In this appendix we prove some

(n)
properties of the volume density functions. In particular, we show that the sequences {ηv,k } take
values in L2 almost surely. We first use an induction argument to establish a useful representation
of the volume density function.
Lemma 4. The buy side volume density function ηv(n)
b
satisfies:
!
k−1 (n),A
Pi=0 Pk−1 (n),B
 1i  1
i=0 i
 
(n) (n) (n) (n)
ηvb ,k = T+ ◦ T− vb,0
k−1
( (n),A (n),B
! )
∆v (n) X  Pk−1 1
j=i+1 j
 Pk−1 1
j=i+1 j
 
(n) (n) (n),C (n) (n),C (n),D (n),D
+ · T+ ◦ T− −Mv,i ηvb ,i 1i + Mv,i 1i .
∆x(n) i=0
 Horst and Paulsen: A Law of Large Numbers for LOBs
36 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

Proof. For k = 0, the equation holds by definition. Let us therefore assume it holds for all k ≤ p.
For k = p + 1 one then obtains
   
(n) (n) (n)
ηvb ,p+1 = ηv(nb ,p
)
+ T+ (ηv(nb ,p
)
) − ηv(nb ,p
)
1(pn),A + T− (ηv(nb ,p
)
) − ηv(nb ,p
)
1(pn),B
∆v (n) (n),C (n) (n),C ∆v (n) (n),D (n),D
− · M v,p ηv ,p 1 p + · Mv,p 1p
∆x(n) b
! ∆x(n)
 1(n),A
p
  1pB
(n) (n)
ηv(nb ,p
)
− Mp(n),C ηv(nb ,p
) (n),C
+ Mp(n),D 1(pn),D


= T+ ◦ T− 1p
!
 Ppi=0 1(n),A
i
 Ppi=0 1(n),Bi
 
(n) (n) (n)
= T+ ◦ T− vb,0
p
( ! )
∆v (n) X  Ppj=i+1 1(n),A
j
 Ppj=i+1 1(n),B
j
 
(n) (n) (n),C (n) (n),C (n),D (n),D
+ · T+ ◦ T− −Mv,i ηvb ,i 1i + Mv,i 1i .
∆x(n) i=0


B.1. Boundedness of volume densities Using the isometry property of the translation
operator we deduce that the L2 norm of the volume density function can be estimated from above
by considering a model with only passive order placements. In a similar way we can estimate the
expected order book hight at any given price tick. More precisely, we have the following result.

Lemma 5. The expected L2 -norm of the volume density function is uniformly bounded:
(n)
sup Ekηvb ,k k2L2 ≤ C (53)
n∈N, k=0,...,bT /∆t(n) c

for some constant C < ∞. Likewise, the expected order book hight is uniformly bounded, i.e. if we
(n) (n),j
put ηvb ,k = (ηvb ,k )j∈Z , then :
(n),j
sup E|ηvb ,k |2 ≤ C. (54)
j∈Z, n∈N, k=0,...,bT /∆t(n) c

Proof. It is enough to consider a model with only order placements where ωkD = 1 a.s. W.l.o.g. we
(n)
may also assume that |πkD | ≤ 1 a.s. and ηvb ,0 ≡ 0. Furthermore we may as well use a representation
(n)
of the volume densities in absolute rather than relative coordinates. In such a model, Ekηvb ,k k2L2 is
of the form !2
 (n) 2 X X k
(n) 2 ∆v (n)
Ekηvb ,k kL2 = E ai,j · ∆x(n)
∆x(n) j∈Z i=1
(n)
where ai,j := 1{πD ∈[x(n) ,x(n) )} with the distribution of the random variables πiD properly adjusted
i j j+1
to account for the representation in absolute coordinates. Since the random variables πiD have
compact support, only finitely many summands are non-zero and we may rearrange terms. Using
conditional independence of the placement variables though time, this yields:

2 k
2 k
(n) 2 ∆v (n) X X  (n) 2 ∆v (n) X X (n) (n)
Ekηvb ,k kL2 = E a i,j + Eai,j Eai0 ,j .
∆x(n) i=1 j∈Z ∆x(n) 0 0 j∈Z i,i =1,i6=i

Using the fact that no placements take place at price levels with a distance of more than 1 from
the prevailing best bid/ask price:
 2
(n)
E ai,j ≤ kf D k∞ ∆x(n) 1{|j−η(n) |≤1}
(n) (n) D 2

γ,i
(n) 2
Eai,j Eai0 ,j ≤ kf k∞ ∆x 1{|j−η(n) |≤1} 1{|j−η(n) |≤1} .
γ,i γ,i0
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 37

In particular, the inner sums extend over at most ∆x2(n) + 1 terms. As a result, our scaling assump-
tions guarantee that
(n)
sup Ekηvb ,k k2L2 < ∞.
n,k=1,...,bT /∆t(n) c

The second assertion follows analogously as

!2
(n) 2 k


(n),j ∆v X (n)
E|ηvb ,k |2 = E ai,j .
∆x(n) i=1

B.2. Norm estimates The next result will be used to prove a Lipschitz continuity property
of the grid-point approximation of the limiting PDE.
Lemma 6. There exists a constant C < 0 such that for all n ∈ N and k = 0, ..., bT /∆t(n) c:
h i
(n) (n) (n)
E T± (ηvb ,k ) − ηvb ,k ≤ C · ∆x(n) .
L2

(n)
Proof. Using Lemma 4 and the linearity of the translation operator T+ it follows that a.s.
 
(n) (n) (n)
T+ ηvb ,k − ηvb ,k
!
 Pk−1
i=0 1A
i
 Pk−1
i=0 1B
i
   
(n) (n) (n) (n) (n)
= T+ ◦ T− T+ vb,0 − vb,0 (55)
k−1 Pk−1 Pk−1 !
A B
∆v (n) X  (n)  j=i+1 1j  (n)  j=i+1 1j h (n)  (n),D  (n),D
i 
+ · T + ◦ T − T + M i − M i 1D
i (56)
∆x(n) i=0
k−1 Pk−1 Pk−1 !
A B
∆v (n) X  (n)  j=i+1 1j  (n)  j=i+1 1j h (n)  (n),C (n)  (n),C (n)
i 
− · T+ ◦ T − T+ Mi ηv ,i − M i η v ,i 1C
i
∆x(n) i=0 b b

(57)
Taking the expected value and norms in (57) we find:
h   i
(n) (n) (n)
E T+ ηvb ,k − ηvb ,k
L2
k−1
(n)

(n)

(n) ∆v (n) X (n)  h (n),D i h
(n),D
i
≤ T+ vb,0 − vb,0 + T + E M i − E M i .
L2 ∆x(n) i=0 L2
k−1
∆v (n) X (n)  h (n),C (n) i h
(n),C (n)
i
+ T E M η − E M η . (58)
∆x(n) i=0 + v,i vb ,i v,i vb ,i
L2

 
(n) (n) (n)
By Assumptions 1 and 3 there exists a constant K < ∞ such that T+ vb,0 − vb,0 ≤ K∆x(n)
L2
and
1 (n)
 h
(n),D
i h
(n),D
i
(n) (n),D
− f (n),D


T+ E Mi − E M i = T + f
∆x(n) L2 L2
≤ K∆x(n) .

As for the cancellation terms, independence of the event dynamics from the standing volumes yield:
1 (n)
 h
(n),C (n)
i
(n)

(n),C (n)

T E M η = T f Eη vb ,i ,
∆x(n) + i vb ,i +
1 h
(n),C (n)
i
(n)
E Mi ηvb ,i = f (n),C Eηvb ,i .
∆x(n)
 Horst and Paulsen: A Law of Large Numbers for LOBs
38 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

In view of the second assertion of Lemma 5 and using the fact that f (n),C is bounded along with
Assumption 3 we find a constant K < ∞ such that:
1 (n)
 h
(n),C (n)
i h
(n),C (n)
i
T + E M i η vb ,i − E M η
∆x(n)
h (n)  (n)  i i vb ,i
L2
(n) (n),C (n)
≤ T+ f E T+ ηvb ,i − ηvb ,i
L2
(n) (n),C

(n),C

(n)
+ T f −f Eηvb ,i
   L2 
(n) (n) (n) (n)
≤ K EkT+ ηvb ,i − ηvb ,i kL2 + ∆x .

Altogether, we arrive at the following estimate:

h   i bT /∆t(n) c h   i
(n) (n) (n)
X (n) (n) (n)
E T+ ηvb ,k − ηvb ,k ≤ K∆x(n) + K∆v (n) E T+ ηvb ,i − ηvb ,i . (59)
L2 L2
i=0

Hence, it follows from Gronwall’s lemma that

h   i
(n) (n) (n)
= O ∆x(n) .


sup E T+ ηvb ,k − ηvb ,k
k=0,...,T /∆t(n) L2


Corollary 2. There exists a constant C > 0 such that

ku(n) (t, · + ∆x(n) ) − u(n) (t, ·)kL2 ≤ C · ∆x(n) .

Moreover,
sup ku(n) (t; ·)kL2 < ∞.
n∈N,t∈[0,T ]

Proof. In order to establish the first assertion we represent the functions u(n) as
(n) (n)
uk = Eζk (60)

where
  
(n) (n) (n) (n) (n)
ζk+1

 := ζk + Dv,k γ(tk ), ζk
. (61)

ζ (n),j

:=
(n),j
v0
0

For t ∈ [k · ∆t(n) , (k + 1) · ∆(n) ) the preceding lemma then implies:

h   i
(n) (n) (n)
ku(n) (t, · ± ∆x(n) ) − u(n) (t, ·)kL2 = E T± ζk − ζk ≤ C∆x(n) .
L2

The second assertion follows from (60) together with Lemma 5:

(n) (n) (n)
ku(n) (tk ; ·)kL2 = kEζk kL2 ≤ Ekζk kL2 ≤ C.


Using Lipschitz continuity of f I along with the point wise shift estimate of the initial volume
densities of Assumption 1 the following result can be established by analogy to Corollary 6.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 39

Corollary 3.
h i
(n),j+1 (n),j
sup E ηvb ,k − ηvb ,k = O(∆x(n) ). (62)
n∈N, j∈Z, k=0,...,bT /∆t(n) c

We close this appendix with norm estimates which are key to the proof of Proposition 8.
Lemma 7. The following norm estimates hold:
" 2
#
∆v (n) (n),D,H ∆v (n) (n),D,H
 
= O (∆v (n) )2


sup E 1Dk ,Hk M − E 1Dk ,Hk M
∆x(n) v,k ∆x(n) v,k
"n,k L2
2
#
∆v (n) (n),C,G (n) ∆v (n) (n),C,G (n)
 
= O (∆v (n) )2


sup E 1Ck ,Gk M ηv,k − E 1Ck ,Gk M ηv,k
∆x(n) v,k ∆x(n) v,k
n,k L2 
    h    i 2
(n) (n) (n) (n) (n) (n)
= O (∆v (n) )2α + (∆v (n) )2−α


sup E 1Ak ,Bk T± ηvb ,k − ηvb ,k − E 1Ak ,Bk T± ηvb ,k − ηvb ,k
n,k L2
 
    h    i 2
(n) (n) (n) (n) (n) (n)
= O (∆v (n) )2α + (∆v (n) )2−α .


sup E 1Ek ,Fk T± ηvs ,k − ηvs ,k − E 1Ek ,Fk T± ηvs ,k − ηvs ,k
n,k L2

Proof. The first two estimates follow from boundedness of the density functions f I along with
independence of the event dynamics from volumes and (53). In order to establish the third and
fourth estimate we have to prove that
n h   io
(n) (n) (n)
sup ∆p(n) · E kT+ ηvb ,k − ηvb ,k k2L2 = O (∆v (n) )2α + (∆v (n) )2−α .


n,k

(n) (n) (n)

To this end we use a representation of T+ ηvb ,k − ηvb ,k as in Lemma 6 but in absolute rather
than relative coordinates. This means that the shift terms drop out of the representation but the
probabilities of placements and cancellations need to be properly adjusted.
Assumption 1 allows us to bound the impact of the initial condition (55) by a term of the order
(∆x(n) )2 . To compute the norm of the sum in (56) we need to compute a term of the form
 (n) 2 X X k
!2
∆v (n)
E ai,j · ∆x(n)
∆x(n) j∈Z i=1

where 
 1 if πiD ∈ [(j − 1) · ∆x(n) , j · ∆x(n) )
(n)
ai,j = −1 if πiD ∈ [j · ∆x(n) , (j + 1) · ∆x(n) ) .
0 else


In particular there exists a constant K < ∞ such that

 2
(n)
E ai,j ≤ K∆x(n) 1{|j−η(n) |≤1}
(n)
2 γ,i
Eai,j ≤ K ∆x(n) 1{|j−η(n) |≤1}
γ,i

where the second inequality follows from (5); the indicator functions account for the representation
of volumes in absolute coordinates. Using the fact that the random variables πkD have compact
support and that events are conditionally independent through time, we can now argue as in the
proof of Lemma 5 to deduce that:
 (n) 2 X X k
!2
(n) 2
 
(n) ∆v (n) (n) (n) (∆v ) 1 1 1 (n) 4
∆p E ai,j · ∆x ≤ K∆p + (∆x )
∆x(n) j∈Z i=1
∆x(n) ∆v (n) (∆v (n) )2 ∆x(n)
= K (∆v (n) )2α + (∆v (n) )2−α .


 Horst and Paulsen: A Law of Large Numbers for LOBs
40 Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS

(n)
To compute the norm of the sum in (57) we need to compute a similar term, but with ai,j
replaced by  (n),j
 ηvb ,k if πiD ∈ [(j − 1) · ∆x(n) , j · ∆x(n) )
(n)
bi,j = −ηv(n),j−1 if πiD ∈ [j · ∆x(n) , (j + 1) · ∆x(n) ) .
 b ,k
0 else
 2
(n)
Using (54) we have again that E ai,j ≤ K∆x(n) 1{|j−η(n) |≤1} . Using independence of the event
γ,i
dynamics from volumes along with Lipschitz continuity of f C we also obtain a constant K < ∞
such that:
(n) (n),j (n),j−1
Eai,j ≤ P[πiD ∈ [j · ∆x(n) , (j + 1) · ∆x(n) )] · Eηvb ,i − P[πiD ∈ [(j − 1) · ∆x(n) , j · ∆x(n) )] · Eηvb ,i
n
2 o
(n),j (n),j−1
≤ K ∆x(n) + |Eηvb ,i − Eηvb ,i | · P[πiD ∈ [(j − 1) · ∆x(n) , j · ∆x(n) )] .

Hence it follows from Corollary 3 that

(n)
2
Eai,j ≤ K ∆x(n)

and the assertion follows as in the case of placements.

Acknowledgments. Most of this work was done while the first author was visiting CERE-
MADE; grateful acknowledgment is made for hospitality and financial support. Both authors
acknowledge financial support through the SFB 649 “Economic Risk” and Deutsche Bank AG. We
thank Dörte Kreher for pointing out a gap in an earlier version, Nikolaus Hautsch for providing
LOBSTER data, Gökhan Cebiroglu for assistance with the calibrations, and seminar participants
at various institutions and conferences for valuable comments and suggestions. We are grateful for
comments and suggestions from two anonymous referees, an associate editor and Adam Schwarz,
the Area Editor, that greatly helped to improve the presentation on the results.

References
[1] F. Abergel and A. Jedidi. A mathematical approach to order book modeling. International Journal of
Theoretical and Applied Finance, 2013.
[2] A. Alfonsi, A. Fruth, and A. Schied. Optimal execution strategies in limit order books with general
shape functions. Quantitative Finance, 10(2):143–157, 2010.
[3] R. Almgren and N. Chriss. Optimal execution of portfolio transactions. Journal of Risk, 3:5–39, 2001.
[4] V.V. Anisimov. Switching Processes: Averaging Principle, Diffusion Approximation and Applications.
Acta Applicandae Mathematicae, 40:95–141, 1995.
[5] C. Bayer, U. Horst, and J. Qiu. A diffusion approximation for the limit order book. Preprint, 2014.
[6] E. Bayraktar, U. Horst, and R. Sircar. Queueing theoretic approaches to financial price fluctuations.
Handbook of Financial Engineering, 15:637–677, 2007.
[7] B. Biais, P. Hillion, and C. Spatt. An empirical analysis of the limit order book and the order flow in
the Paris Bourse. The Journal of Finance, 50(5):1655–1689, 1995.
[8] P. Billingsley. Convergence of Probability Measures. Wiley, 1999.
[9] G. Cebiroglu and U. Horst. Optimal order display in limit order markets with liquidity competition.
Journal of Economic Dynamics and Control, 58:81–100, 2015.
[10] R. Cont and A. De Larrard. Order book dynamics in liquid markets: Limit theorems and diffusion
approximations. Available at SSRN: http://ssrn.com/abstract=1757861, 2012.
[11] R. Cont and A. De Larrard. Price dynamics in a Markovian limit order market. SIAM Journal of
Financial Mathematics, 4:1–25, 2013.
Horst and Paulsen: A Law of Large Numbers for LOBs
Mathematics of Operations Research 00(0), pp. 000–000, c 0000 INFORMS 41

[12] R. Cont, S. Stoikov, and R. Talreja. A stochastic model for order book dynamics. Operations Research,
58:3:549–563, 2010.
[13] D. Easley and M. O’Hara. Price, trade size, and information in securities markets. Journal of Financial
Economics, 19(1):69 – 90, 1987.
[14] L.C. Evans. Partial Differential Equations. American Mathematical Society, 2nd edition, 2010.
[15] T. Foucault, O. Kadan, and E. Kandel. Limit order book as a market for liquidity. Review of Financial
Studies, 18(4):1171–1217, 2005.
[16] X. Gao, A.B. Dai, and S.J. Deng. Hydrodynamic limit of order book dynamics. ArXiv e-print 1411.7502,
2014.
[17] M.B. Garman. Market microstructure. Journal of Financial Economics, 3(3):257 – 275, 1976.
[18] L.R. Glosten and P.R. Milgrom. Bid, ask and transaction prices in a specialist market with heteroge-
neously informed traders. Journal of Financial Economics, 14(1):71 – 100, 1985.
[19] E. Hairer, S.P. Nørsett, and G. Wanner. Solving Ordinary Differential Equations I: Nonstiff Problems.
Springer, 2000.
[20] U. Horst and D. Kreher. A weak law of large numbers for a limit order book model with fully state
dependent order dynamics. ArXiv e-print 1502.04359, 2015.
[21] U. Horst and C. Rothe. Queuing, social interactions and the microstructure of financial markets.
Macroeconomic Dynamics, 12:211–233, 2008.
[22] L. Kruk. Functional limit theorems for a simple auction. Mathematics of Operations Research, 28
No.4:716–751, 2003.
[23] A. Lo, A.C. MacKinlay, and J. Zhang. Econometric models of limit-order executions. Journal of
Financial Economics, 65:31–71, 2002.
[24] J. Osterrieder. Arbitrage, Market Microstructure and the Limit Order Book. Ph.D. thesis, ETH Zurich.
DISS. ETH Nr 17121, 2007.
[25] C.A. Parlour. Price dynamics in limit order markets. Review of Financial Studies, 11(4):789–816, 1998.
[26] G. Pisier. Probabilistic methods in the geometry of banach spaces. in: Probability and analysis (Varenna,
1985), Lecture Notes in Mathematics, 1206, 1986.
[27] I. Rosu. A dynamic model of the limit order book. Review of Financial Studies, 22(11):4601–4641, 2009.