PLoSONE16 GeometricPhase
PLoSONE16 GeometricPhase
net/publication/301838514
CITATIONS READS
3 2,304
1 author:
Claudio Altafini
Linköping University
211 PUBLICATIONS 5,969 CITATIONS
SEE PROFILE
All content following this page was uploaded by Claudio Altafini on 08 October 2016.
Abstract
Geometric phases describe how in a continuous-time dynamical system the displacement
of a variable (called phase variable) can be related to other variables (shape variables)
undergoing a cyclic motion, according to an area rule. The aim of this paper is to show that
geometric phases can exist also for discrete-time systems, and even when the cycles in
shape space have zero area. A context in which this principle can be applied is stock trad-
ing. A zero-area cycle in shape space represents the type of trading operations normally
a11111 carried out by high-frequency traders (entering and exiting a position on a fast time-scale),
while the phase variable represents the cash balance of a trader. Under the assumption
that trading impacts stock prices, even zero-area cyclic trading operations can induce geo-
metric phases, i.e., profits or losses, without affecting the stock quote.
OPEN ACCESS
Fig 1. Geometric phase in continuous and discrete time. Panel (a): classical picture for geometric phase of a
continuous-time dynamical system: a cyclic path in shape space (orange curve) induces a net motion on the phase
variable (blue curve), proportional to the area of the cycle in shape space, see Supplementary Notes for an explicit model.
Panel (b): geometric phase for a discrete-time dynamical system. A cyclic trajectory, even of zero-area in shape space
(orange), induces a net nonzero motion in the phase variable (blue). In the stock trading application, the shape variables
are the amount of stocks owned by a trader and the stock quote. The phase variable is the cash balance of the trader.
doi:10.1371/journal.pone.0161538.g001
elements of chronological calculus needed to deal with such formal power series; [14] survey
the controllability of nonlinear discrete-time systems from a Lie-algebraic perspective; [17]
studies invariance of discrete-time distributions of vector fields, [18] their linearization via
feedback transformations, and so on. However, in our knowledge the properties of discrete-
time geometric phases have never been investigated, let alone the existence of phase motions
induced by zero-area shape cycles.
The second aim of this paper is to present an application where discrete-time zero-area geo-
metric phases admit a useful interpretation. The application we want to discuss is stock trading.
In the idealized stock exchange we consider, we can take as shape variables the amount of a cer-
tain stock owned by a trader and the quote of that stock. The phase variable is the cash balance
of our trader, which is a function of how many stocks are bough/sold but also of the quote at
which they are bought/sold. The product of these two quantities gives the nonlinearity required
to have a nonintegrable system of difference equations. The discrete events which constitute
the time instants of interest for our model are the trading events that concern the stock we are
dealing with. Using a discrete-time dynamics for them is natural, especially when fast time
scales are considered. At the most microscopic time scale (microsecond or less), in fact, the
dynamics of stock trading consists of a series of events like emission of orders by the traders
and order execution by the stock exchange engine (other more complex processes like order
redirections among different stock exchanges are not considered).
Assume that our trader is the only actor in our stock market, and that his market orders
(i.e., orders to buy or sell at the current best available price) are executed against a limit order
book, i.e., a book of orders to buy or sell at prespecified stock prices [19, 20]. Assume further
that trading impacts prices, i.e., a buy order drives prices up, while a sell order drives prices
down. When nothing else happens to our stock, a zero-area cyclic trajectory in shape space is a
sequence of a buy order followed by a sell order of equal magnitude, see Fig 1(b). At the end of
the cycle, the stock quote (a shape variable) is unchanged, as the two price impacts compensate
each other. However the cash balance of the trader (the phase variable) is not zero at the end of
the cycle. In fact, the quote at which the stocks are bought does not contain the price variation
due to the action of buying itself. However, this price impact is incorporated in the quote at
selling time, which makes the quote at which stocks are sold higher than that at which the
stocks are bought, as an effect of the act of trading itself. This is the geometric phase of stock
trading.
In what follows an idealized scenario is presented first, as a way to shown that nontrivial
geometric phases for zero-area shape trajectories indeed exist. The scenario is then rendered
more realistic by incorporating in the model a bid/ask spread for the stock quote, commission
fees, a stock quote drift (representing the operations of all other traders of the same stock), and
general models of price impact, without altering the nature of the geometric phase phenome-
non just described.
Entering and liquidating a position, possibly very quickly, is the typical pattern of high-fre-
quency trading (i.e., fast computerized trading with short holding periods and no inventory
[21–24]). The small margins (“pennies”) of gains or losses resulting from such “round trip”
operations resemble very much the tiny values that the discrete-time geometric phase accumu-
lates at the end of a cycle. In order for such cyclic trade operations to yield a penny profit, there
must be an increase of the stock quote between the event of buy and that of sell for a buy-then-
sell cycle, or a decrease for a sell-then-buy cycle. Clearly many forces are normally at work in a
stock market to induce such price changes, most of them out of control of a single trader. How-
ever, it is tempting to speculate that on the very short time scale also the “endogenous” price
impact induced by the operations carried out by the trader performing the round trip contrib-
utes to move the prices in the right direction. When a stock quote has a big spread between its
bid and ask prices, then such an endogenous quote movement is clearly not enough to com-
pensate for the spread gap. However, when the spread is small and the limit order book has suf-
ficient depth and a high rate of limit order arrival on both bid and ask sides (situation which is
normally referred to as “locked” market [20] and which is not so uncommon for highly traded
stocks), then geometric phases induced by endogenous price impacts may in principle be able
to compensate for the quote spread, and may contribute actively to the penny profit of cyclic
trade operations. Models are developed in the paper to try to quantify this contribution.
A situation in which the endogenous price impact is potentially playing a major role is
when high frequency traders can anticipate (“front-run”) the operations of slower “classical”
traders. We show in the paper that such interlacing of fast cycles and slow classical operations
is essentially a way of increasing the price impact (and hence the geometric phase) of the fast
trader performing a cycle in shape space, at the expense of the slow trader who looses part of
his geometric phase (recall that the geometric phase corresponds to the cash balance of a
trader). Practices like front-running are sometimes attributed to high-frequency trading [25].
Models
Consider an idealized “stock exchange” scenario consisting of a single trader trading on a single
stock through market orders, executed against a limit order book ([19], more details are given
below). Denote y(t) the quantity of stocks owned by the trader at time t and with s(t) the stock
quote. When the trader buys or sell stocks, his cash balance z(t) changes accordingly.
We make the assumption that trades impact prices: selling stocks drives prices down, while
buying stocks drives prices up. In discrete time an elementary model for the scenario being
considered is the following:
yðtiþ1 Þ ¼ yðti Þ þ uðti Þ ð1Þ
Here u(ti) = y(ti+1) − y(ti) is the amount of stock being bought (u(ti)>0) or sold (u(ti)<0) at
time ti by the trader. At time of execution, a buy order (u(ti)>0) raises the quote of the stock,
i.e., s(ti+1)>s(ti), while the cash balance decreases z(ti+1)<z(ti). The events represented by eqs
(1)–(3) all have the same time stamp ti, meaning that for all 3 equations the variables are
updated at the next time instant, ti+1. The length of the time interval Δti = ti+1 − ti can be very
small, provided that Δti > 0. The coefficient r quantifies the price impact, i.e., how influential
an order is for a certain stock. Its value may depend on the total volume of the stock, its liquid-
ity, volatility, frequency of buy vs sell orders, number of traders, etc. A basic dynamical model
is described for instance in [26]. For the purposes of this paper, and given that we are interested
in tracking the activity of a single trader, assuming a constant r is enough, although a more gen-
2 3
yðtÞ
6 7
eral formula will be provided later on. In vector form, calling xðtÞ ¼ 4 sðtÞ 5 the state of the
zðtÞ
system, the eqs (1)–(3) can be written as
2 3
1
6 7
xðtiþ1 Þ ¼ xðti Þ þ 4 r 5uðti Þ: ð4Þ
sðti Þ
If no other trading operation happens during the time interval, the evolution of the state vector
is
8
>
> xð0Þ t 2 ½0; tb
>
> 2 3
>
> k
>
>
>
> 6 7
>
> 6 7
>
> xð0Þ þ 6 rk 7 t 2 t ; t
>
> 4 5 bþ1 s
<
xðtÞ ¼ sð0Þk ð6Þ
>
> 2 3
>
> 0
>
>
>
> 6 7
>
> 6 7
>
> xð0Þ þ 6 0 7 t 2 tsþ1 ; tend :
>
> 4 5
>
>
: 2
rk
In other words, the net amount of stock owned by the trader at the end of the cycle is equal to
that at the start: y(tend) = y(0). When no other operation happens during the time interval then
also s(tend) = s(0), i.e., the stock quote is unchanged. However, if r > 0, the cash balance is posi-
tive, i.e, z(tend) = z(0)+rk2 > z(0). Such a positive cash balance is solely due to the price impact
of the transactions being made by our trader. It is proportional to the square of the number of
stocks traded. A similar result occurs for a sell-then-buy cycle, which also leads to z(tend) > z(0),
see Fig 3. In both cases, while y(t) and s(t) accomplishes a cyclic trajectory returning to their
Fig 2. Effect of cyclic trading operations. Panel (a): time profiles of the variables for a single buy-then-sell
cycle. Panel (b): corresponding shape and phase space profiles. At time tb = 20 the trader buys a unit of
stock. The stock quote s increases consequently because of the price impact of the operation, while the cash
balance z becomes negative. At time ts = 80 the stock is sold and the stock quote returns to the level it had at
t = 0. However, the cash balance z becomes positive because the stock is sold at a price higher than it was
bought. The shape trajectory of the variables (y, s) is cyclic and has zero area (orange curve in panel (b)). The
phase variable z experiences a net displacement because of the cycle, i.e., a geometric phase is produced
(blue curve in panel (b); the starting point is the green dot and the final point is the red dot). Panels (c) and (d):
When the stock quote is allowed to drift because of the operations of other traders, then periodic operations of
a trader no longer result in cyclic trajectories in shape space (y, s). However a geometric phase in the z
variable is still present and keeps building up whenever the trader completes a cycle of operations. In panel
(c), the plot for s shows in black the stock quote in absence of trading cycles, and in red the deviations due to
these trading cycles, visible only during the intervals in which stocks are hold (or shorted, for sell-then-buy
cycles). Such red deviations are a direct consequence of our price impact assumption, and induce the
accumulation of positive cash balance (green trace in (c)).
doi:10.1371/journal.pone.0161538.g002
starting point, z(t) does not. The key in understanding why a nonzero cash flow is generated by
cyclic operations lies in the nonlinearity of the difference eq (3), namely s(t)u(t) is a bilinear
term, i.e., linear, but simultaneously so, in both state and input. This nonlinearity introduces a
behavior which is typical of nonintegrable vector fields [7, 8].
The classical description of geometric phase one finds in the literature [3, 8] deals with non-
linear continuous-time vector fields, in which 2 or more of the variables (corresponding to our
y(t) and s(t)) evolve on a shape space, and the remaining variables (here z(t)) on a phase space
above the shape space, see Fig 1(a). Cyclic trajectories in shape space yield a net motion on the
phase variable which is proportional to the area of the cyclic shape trajectory and it is zero
when this area is zero, see Fig A of S1 File.
In discrete-time, however, the picture is different. As shown in Fig 2 and as computed
explicitly in eq (6), even a zero-area cyclic trajectory in shape space yields a non-zero motion
on the phase variable. In our specific example, the origin of the positive cash flow induced by a
cyclic trajectory (entering then exiting a position) is due to the fact that stocks are bought at a
Fig 3. A sell-then-buy cycle. Panel (a): time profiles of the variables. Panel (b): shape and phase space
profiles. Also in this case the trader obtains a profit from the cyclic operation (red dot in panel (b) is above the
green dot), because the stock is bought at a lower price than it was sold.
doi:10.1371/journal.pone.0161538.g003
lower price s(tb) and are sold at a higher price s(ts), see Fig 2. The price increase is due to our
assumption that stock trades impact stock quotes. The argument is analogous if the cycle con-
sists of a sell-then-buy: in this case the quote at selling time is higher than that at buying time,
see Fig 3.
where we still assume that operations of our trader are cyclic and follow eq (5), but Eq (8) has
an extra drift term. At the end of a buy-then-sell cycle, instead of eq (6) we have:
2 3
yð0Þ
6 Pend 7
xðtend Þ ¼ 4 sð0Þ þ i¼0 wðti Þ 5 ð10Þ
Ps
zð0Þ þ i¼b wðti Þk þ rk 2
i.e., the stock trend in the hold interval [tb, ts] also affects the cash flow of our trader, but only
as a first order term of the number of stocks held during the cycle, as opposed to the second
order term of the noncommutative effect corresponding to the geometric phase. Iterating over
many cyclic operations of our trader, initiated at time instants {tb} drawn from a Poissson pro-
cess, and drawing also {ts − tb} from a Poisson distribution of shorter mean arrival time, then
the stochastic difference eqs (7)–(9) is such that at tend, y(tend) = y(0), while instead now the
stock quote has a nontrivial trend s(tend) 6¼ s(0). From eq (10), the influence of the stock trend
on the cash balance can be made small by reducing the length of the time interval [tb, ts], so
that regardless of the value of the drift term {w(ti)}, it is possible to obtain z(tend) > z(0) = 0.
Hence, provided that a cycle is completed at sufficiently high frequency, the price impact
induced by the cycle itself can still dominate the trend imposed by the market. A sample trajec-
tory depicting this situation is shown in Fig 2(c) and 2(d). In panel (c), in the plot of s, the
stock quote driven only by {w(ti)} is shown in black. The red deviations appearing in the plot
correspond to the price impact of the operations carried out by our trader which in the model
(8) are superimposed to {w(ti)}. In fact, combining eqs (6) and (10) for a single buy-then-sell
cycle the value of s in the holding interval [tb, ts] is given by
X
i1
sðti Þ ¼ sð0Þ þ wðtj Þk þ rk; ti 2 tbþ1 ; ts
j¼b
At the end of the cycle, the cash balance from eq (11) is z(tend) = z(0) + rk2 − qk. If we include
also a fee for each trading operation (assumed constant and equal to c), then
zðtend Þ ¼ zð0Þ þ rk2 qk 2c: ð12Þ
Fig 4. Effect of a wide bid-ask spread. The buy order empties the shaded red region of the ask side of the
limit order book. The stock quote (midprice) consequently increases, while the best bid price (blue) remains
unchanged. The sell order needed to complete the cycle settles at a price below the original buy price, hence
the cycle results in a net loss for the trader.
doi:10.1371/journal.pone.0161538.g004
Fig 5. Effect of a small bid-ask spread. In a deep market with many buy and sell orders arriving
continuously, the spread remains small around the midprice, even after the execution of the buy order has
raised the stock quote. Hence the price impact of the buy order can lead to a new best bid price which is
above the quote at which the buy order was executed. In this case, completing the cycle indeed results in a
net gain.
doi:10.1371/journal.pone.0161538.g005
Hence a cycle yields a net profit if z(tend) − z(0) = rk2 − qk − 2c > 0. See Fig 6 for an example
of how spread can reduce gains (or induce losses) in z(t). In eq (12), if the market is deep
enough and the spread limited enough, then z(tend) − z(0) can be made positive by choosing a
sufficiently high value for k. Regardless of its sign, the quantity z(tend) − z(0) is still a geometric
phase, representing profit when z(tend) − z(0) > 0 or losses when z(tend) − z(0) < 0.
In presence of a drift term {w(ti)} for the stock price as in eqs (7)–(9), the cash balance can
be modified according to Eq (10) as
!
X
s
zðtend Þ ¼ zð0Þ þ rk2 þ wðti Þ q k 2c:
i¼b
As already mentioned, when the hold interval [tb, ts] is short, so is the effect of the drift term
Ps
i¼b wðti Þk on the final value of the cash balance. Fig 7 shows the value of z(tend) as we vary
the number of trading cycles and the average duration of the hold intervals for a large number
of realizations {w(ti)}. When in a formula like eq (12) the geometric phase is positive (z(tend)>z
(0)), then even in presence of drift the cash flow remains positive at least as long as [tb, ts] is
Fig 6. Spread reduces profits (or induces losses). Panel (a): In Eq (12), the presence of a quote spread
decreases the value of z(tend) − z(0) i.e., the geometric phase at the end of a cycle. In the plot for z, if the green
curve corresponds to q = c = 0, the magenta and brown curves correspond to different choices of these
parameters. Depending on the values of r, q and c, the geometric phase can correspond to a profit (green and
magenta curves, z(tend) > z(0)), or to a loss (brown, z(tend) < z(0)). Panel (b): The geometric phase (and the
corresponding profits or losses) accumulates over repeated trading cycles. For short trading cycles the
presence of a stock drift {w(ti)} does not alter the result, i.e., what happens in a single cycle (panel (a)) to the
cash balance z(t) happens also when many cycles are performed in presence of stock drift {w(ti)} (panel (b)).
doi:10.1371/journal.pone.0161538.g006
Fig 7. Endpoint of a series of cyclic trading operations, in presence of stock quote spread and quote
drift. Each dot is the result of 1000 simulations, under the assumption that z(tend) − z(0) 0 when the stock
drift is absent (formula (12)). The number of cyclic operations is considered itself a random variable (denoted
“n. of trade cycles”), and so is the duration of the holding intervals [tb, ts] (denoted “mean duration of cycle”).
For short cycles (with respect to the standard deviation σ of the stock drift term), the trader always obtains a
profit (panel (a)). The profit is proportional to the number of trade cycles performed by the trader (panel (b)).
The correlation tends to get lost when the mean duration of the trade cycles becomes large (red dots in
panels (b) and (c)). For short cycles, the profit happens independently of the stock trend (blue dots in panel
(d)). Also when the trade cycles become long and profit may be lost, there is no significant correlation
between z and s (panel (d)).
doi:10.1371/journal.pone.0161538.g007
sufficiently short compared to σ. Short-lived deviations in the stock quote are enough to mani-
fest the noncommutativity effect we are describing here. In Fig 7, notice how rather than to the
quote drift, z(tend) is correlated to the number of trading cycles carried out in [0, tend] (which
follows from r being a constant).
If r() is an odd function, i.e., r(−u) = −r(u), then the neutrality of the stock quote to cyclic
trades is preserved, that is, in absence of drift s(tend) = s(0) and our arguments are still valid.
For instance, combining eqs (13) and (14) with the spread model (11) and the constant trading
fee c yields
zðtend Þ ¼ zð0Þ þ ðrðkÞ qÞk 2c
which becomes
!
X
s
zðtend Þ ¼ zð0Þ þ wðti Þ þ rðkÞ q k 2c
i¼b
Fig 8. Front-running an order. A high-frequency trader that can anticipate the operations of a slower
classical trader can benefit of the price impact created by the latter, in addition to the one induced by his own
trades. In the top plot of both panels, the high-frequency trader input uh (green) always precedes the moves of
the classical trader (uc, blue). In the plot for s of both panels the black trace (denoted so) represents the stock
quote in absence of high frequency trader, while the red trace (s) is the stock quote in presence of a front-
running high-frequency trader. The geometric phase (and hence any possible profit) tends to swap from the
classical trader to the high-frequency trader. For the classical trader, the profit in presence of a front-runner is
reduced (blue curve in the lower plot of both panels, denoted zc) with respect to the same curve in absence of
front-run (magenta curve, denoted zco ).
doi:10.1371/journal.pone.0161538.g008
the system:
ya ðtiþ1 Þ ¼ ya ðti Þ þ ua ðti Þ; a ¼ fc; hg ð15Þ
X
sðtiþ1 Þ ¼ sðti Þ þ rua ðti Þ ð16Þ
a¼fc; hg
sðtend Þ ¼ sð0Þ
where kc and kh are the amount of stocks traded by the classical and high-frequency trader dur-
ing the cycle. Comparing eqs (18) and (19) with eq (6), it can be seen that kc ðkc kh Þ < k2c ,
meaning that for the classical trader the “benefit” of the geometric phase in terms of cash bal-
ance is eroded when a high-frequency trader front-runs his orders. On the contrary, for the
high-frequency trader kh ðkc þ kh Þ > k2h , meaning that front-running the order leads to an extra
gain. In particular, if the quantities traded by the two actors are equal, kh = kc, then
zc ðtend Þ ¼ zc ð0Þ
zh ðtend Þ ¼ zh ð0Þ þ 2rk2h
i.e., the price increase of the classical trader in cyclic operations is completely lost, to the benefit
of the high-frequency trader. If instead kh > kc, then the cycle always results in a loss for the
classical trader: zc(tend) < zc(0). Analogous considerations apply to the more complex models
discussed in this paper, incorporating quote drift, quote spread and alternative price impact
formulas.
Discussion
A continuous-time equivalent of the models (1)–(3) is presented in the Supplementary Notes.
For it, a periodic input trajectory induces a zero-area cyclic path in shape space, which however
results in no neat motion of the phase variable, see Fig A of S1 File. Only when the area of the
cyclic path in shape space becomes nonzero a geometric phase can appear, but for that to hap-
pen we have to modify the continuous-time model (for instance adding a time delay in the
update law for the stock quote, see Fig B of S1 File). This confirms that the geometric phase
described in this paper is intrinsically a discrete-time phenomenon, with no continuous-time
counterpart.
Repeating a multitude of cycles on a fast time scale is a hallmark of high-frequency trading,
which nowadays in certain equity markets like the US stock market constitutes around 50% of
the total trading volume [21, 22, 24]. There is no consensus in the literature on what are the
sources of profit of high frequency trading operations of this type [20, 23, 24, 27, 28], other
than “penny” profits must accumulate from each trading cycle. What the existence of a geo-
metric phase suggests is that in principle tiny profits (or losses, when z(tend) − z(0) < 0) can be
made through pure speculation, without holding any inventory position except for impercepti-
ble amounts of time.
If information on future market operations is available (like when speed is used to front-run
upcoming orders), then this can be used to increase the chances of making profits by incorpo-
rating the geometric phase induced by the operations of the trader being front-run into that of
the front-runner doing a cycle.
Another conceptual consideration that follows from our models is that, as long as one con-
siders cyclic operations, in principle profits or losses can be made with stock trading in a way
that does not affect the main observable of a stock market, namely the quote of a stock. Making
such profits (or losses) does not require any knowledge of market fundamentals, as the geomet-
ric phase just uses the price momentum induced by the act of trading itself, a bit like a self-ful-
filling prophecy. And this notwithstanding any “efficient market” theory.
Supporting Information
S1 File. Contains supplementary text (continuous-time model) and supplementary figures.
(PDF)
Author Contributions
Conceptualization: CA.
Data curation: CA.
Formal analysis: CA.
Funding acquisition: CA.
Investigation: CA.
Methodology: CA.
Project administration: CA.
Resources: CA.
Software: CA.
Supervision: CA.
Visualization: CA.
Writing – original draft: CA.
Writing – review & editing: CA.
References
1. Brockett RW. Cycles That Effect Change. In: National Research Council, editor. Motion, Control and
Geometry. Proc. of a Symposium. http://www.nap.edu/catalog/5772.html: National Academy Press;
1997.
2. Berry M. Anticipations of the geometric phase. Physics Today. 1990; 43(12):34–40. doi: 10.1063/1.
881219
3. Marsden JE, Ratiu TS. Introduction to Mechanics and Symmetry. vol. 17 of Texts in Applied Mathemat-
ics. 2nd ed. Springer-Verlag; 1999.
4. Wilczek F, Shapere A. Geometric Phases in Physics. Advanced Series on Directions in High Energy
Physics. World Scientific; 1989.
5. Mead CA. The geometric phase in molecular systems. Rev Mod Phys. 1992; 64:51–85. doi: 10.1103/
RevModPhys.64.51
6. Laumond JP, editor. Robot Motion Planning and Control. Lecture notes in control and information sci-
ences. London, UK: Springer-Verlag; 1998. doi: 10.1007/BFb0036069
7. Murray RM, Li Z, Sastry SS, Sastry SS. A Mathematical Introduction to Robotic Manipulation. Taylor &
Francis; 1994.
8. Bloch AM. Nonholonomic Mechanics and Control. vol. 24 of Interdisciplinary Applied Mathematics.
Springer-Verlag; 2003.
9. Montgomery R. Gauge theory of the falling cat. In: Enos MJ, editor. Dynamics and control of mechanical
systems: The falling cat and related problems. Providence, RI: American Mathematical Society; 1993.
p. 193–218.
10. Shapere A, Wilczek F. Geometry of self-propulsion at low Reynolds number. Journal of Fluid Mechan-
ics. 1989; 198:557–585.
11. Murray R, Sastry S. Nonholonomic motion planning: steering with sinusoids. IEEE Trans on Automatic
Control. 1993; 38:700–716. doi: 10.1109/9.277235
12. Batterman RW. Falling cats, parallel parking, and polarized light. Studies in History and Philosophy of
Science Part B: Studies in History and Philosophy of Modern Physics. 2003; 34(4):527–557. doi: 10.
1016/S1355-2198(03)00062-5
13. Berry MV. Quantal Phase Factors Accompanying Adiabatic Changes. Proceedings of the Royal Soci-
ety of London A: Mathematical, Physical and Engineering Sciences. 1984; 392(1802):45–57. doi: 10.
1098/rspa.1984.0023
14. Jakubczyk B, Sontag ED. Controllability of Nonlinear Discrete-Time Systems: A Lie-Algebraic
Approach. SIAM Journal on Control and Optimization. 1990; 28(1):1–33. doi: 10.1137/0328001
15. Monaco S, Normand-Cyrot D. Nonlinear Systems in Discrete Time. In: Fliess M, Hazewinkel M, editors.
Algebraic and Geometric Methods in Nonlinear Control Theory. Dordrecht: Springer Netherlands;
1986. p. 411–430. doi: 10.1007/978-94-009-4706-1_21
16. Monaco S, Normand-Cyrot D, Califano C. From Chronological Calculus to Exponential Representa-
tions of Continuous and Discrete-Time Dynamics: A Lie-Algebraic Approach. IEEE Transactions on
Automatic Control. 2007; 52(12):2227–2241. doi: 10.1109/TAC.2007.902734
17. Monaco S, Normand-Cyrot D. Invariant distributions for discrete-time nonlinear systems. Systems &
Control Letters. 1984; 5(3):191–196. doi: 10.1016/S0167-6911(84)80102-4
18. Grizzle JW, Kokotovic PV. Feedback linearization of sampled-data systems. IEEE Transactions on
Automatic Control. 1988; 33(9):857–859. doi: 10.1109/9.1316
19. Gould MD, Porter MA, Williams S, McDonald M, Fenn DJ, Howison SD. Limit order books. Quantitative
Finance. 2013; 13(11):1709–1742. doi: 10.1080/14697688.2013.803148
20. Cartea Á, Jaimungal S, Penalva J. Algorithmic and High-Frequency Trading. Mathematics, Finance
and Risk. Cambridge University Press; 2015.
21. Duhigg C. Stock Traders Find Speed Pays, in Milliseconds. New York Times. 2009;.
22. Aldridge I. High-Frequency Trading: A Practical Guide to Algorithmic Strategies and Trading Systems.
Wiley Trading. Wiley; 2013.
23. Brogaard J, Hendershott T, Riordan R. High-Frequency Trading and Price Discovery. Review of Finan-
cial Studies. 2014; doi: 10.1093/rfs/hhu032
24. Carrion A. Very fast money: High-frequency trading on the NASDAQ. Journal of Financial Markets.
2013; 16(4):680–711. doi: 10.1016/j.finmar.2013.06.005
25. Lewis M. Flash Boys: A Wall Street Revolt. W. W. Norton; 2014.
26. Cont R, Kukanov A, Stoikov S. The Price Impact of Order Book Events. Journal of Financial Economet-
rics. 2013;.
27. Hasbrouck J, Saar G. Low-latency trading. Journal of Financial Markets. 2013; 16(4):646–679. doi: 10.
1016/j.finmar.2013.05.003
28. Menkveld AJ. High frequency trading and the new market makers. Journal of Financial Markets. 2013;
16(4):712–740. doi: 10.1016/j.finmar.2013.06.006