Quantile Mechanics

György Steinbrecher∗, William T. Shaw†

July 16, 2007

Abstract
In both modern stochastic analysis and more traditional probability and statistics, one way of charac-
terizing a static or dynamic distribution is through its quantile function. A direct understanding of this
function offers tangible benefits not available directly from the density function. For example, the simplest
way of simulating any non-uniform random variable is by applying its quantile function to uniform deviates.
Modern methods of Monte-Carlo simulation, techniques based on low-discrepancy sequences and copula
methods all call for the use of quantile functions of marginal distributions. We propose a new approach
to quantiles based first, for the static case, on non-linear ordinary differential equations, which we solve by
the classical method of power series with easily computed non-linear recursion relations for the coefficients.
These methods provide accuracy beyond standard machine precision on subsets of the unit interval and
are well adapted to the benchmarking of approximate techniques, and, for some cases, live simulation. We
provide some web links to implementations that may be starting points for the inclusion of these algorithms
in simulation methods. Next, we consider the time evolution of quantiles via the quantilized Fokker-Planck
equation. This is a non-linear partial differential equation and we are able to utilize the static formulation
to provide elementary time-dependent and equilibrium solutions.

Keywords: Inverse CDF, quantile function, normal, Student, beta, T-distribution, simulation, Monte Carlo,
inverse cumulative distribution function, non linear ordinary differential equations, recurrence relations, Fokker-
Planck equation, stochastic differential equation, partial differential equation.

Working Paper: V1.5 (July 2007 update)

∗ Association EURATOM-MEC Department of Theoretical Physics, Physics Faculty, University of Craiova, Str.A.I.Cuza 13,

Craiova-200585, Romania E-mail: [email protected]

† Corresponding author: Department of Mathematics King’s College, The Strand, London WC2R 2LS, England; E-mail:

[email protected]

1
2 G. Steinbrecher & W.T. Shaw

1 Introduction
In both stochastic analysis and traditional probability and statistics, the quantile function offers a useful way
of characterizing a static or dynamic distribution. An understanding of this function offers real benefits not
available directly from the density or distribution function. For example, the simplest way of simulating any
non-uniform random variable is by applying its quantile function to uniform deviates. This paper proposes to
elevate quantile functions to the same level of management as many of the classical special functions of mathe-
matical physics and applied analysis. That is, we form appropriate ordinary and partial differential equations,
and proceed to the development of power series from the appropriate underlying ordinary differential equations.
We shall consider a static analysis based on ordinary differential equations (ODEs), and a dynamic analysis
based on a corresponding partial differential equation (PDE). In this paper the focus of the dynamic analysis
will be a quantilized form of the integrated Fokker-Planck eqation arising from a stochastic differential equation
(SDE), but other dynamics may be considered in principle. We shall call this the QFPE. The management
of the dynamic QFPE will require an understanding of the solutions of the corresponding ODEs, in much the
same way as the management of, for example, the Schroedinger equation requires an understanding of separa-
tion of variables and solutions of the corresponing (linear) ODEs. This will necessitate some new technology
development for the static analysis. However, the static analysis has intrinsic interest, as well shall now discuss.

1.1 Background to the static analysis

Hitherto quantile functions, in the static context, have largely just been treated as inverses of often complicated
distribution functions, and their use has largely been confined to (a) simple exact solutions, e.g. for the
exponential distribution, (b) in terms of efficient rational approximations, usually for the normal case, and
(c) as a root-finding exercise based on the distribution function. A treatment of some approaches to quantile
functions and other sampling methods, such as rejection, is given in the classic text by Devroye [9]. An extensive
discussion of the use of quantile functions in mainstream statistics is given in the book by Gilchrist[11], where
several other examples of known simple forms for quantile functions are given. Several rational and related
approximations have been given for the quantile function for normal case, sometimes called the probit function
after the work of Bliss [4]. Perhaps the most detailed is “algorithm AS241”, developed by Wichura in 1988
[28] and capable of machine precision in a double precision computing environment. Quantile functions are of
course in widespread use in general statistics, and often find representations in terms of lookup tables for key
percentiles.
A detailed investigation of the quantile function for the Student case has been given by Shaw [22]. However,
while power and tail series were developed there for the general case, the method relied on the inversion of
series for the forward CDF using computer algebra and special computational environments. A prototype for
many quantiles was given by the series solution of the non-linear ODE for the inverse error function given by
Steinbrecher [24]. Here we extend this work to the normal distribution itself, the Student distribution, and
gamma and beta distributions. In particular we are able to facilitate the implementation of the methods of [22]
for the Student case in any computing environment to much higher accuracy. We shall argue that this analytic
approach provides useful and practical techniques for the evaluation of such functions.
The value of quantile functions is increasingly appreciated in Monte Carlo simulation. Quantile functions
work very well with both copula methods and low-discrepancy sequences, in contrast to rejection methods
based on sampling from a circle inside a square, e.g for sampling of the Normal. These issues are discussed by
Jäckel [16]. The use of quantile functions has however not taken root as much as it might due to the relative
intractability of the functions, which thus far have been considered as inverses of the often awkward CDFs of
density functions, which are “special functions” of some computational complexity. While such CDFs and their
inverses are sometimes available in high-level mathematical computation languages such as Mathematica, these
representations (e.g. using an inverse incomplete beta function in the case of the beta and Student distributions)
are less helpful in other languages. In the case of the normal distribution the quantile function has been managed
by the use of composite rational and polynomial approximations based on decomposing the half-unit interval
into two or more regions. The three-region approximation developed by Wichura in 1988 [28], and commonly
known as “AS241” is well-suited to machine-precision computations in a double-precision environment. Acklam’s
method [2] is a very useful and pragmatic method based on two levels of approximation. A simple two-region
approximation provides relative accuracy of order 10−9 for the quantile, and a single refinement based on
Newton-Raphson-Halley iteration produces machine-precision. An assessment of the merits of these schemes
and others in common use has been given by one of us [23].
Quantile Mechanics 3

Such rational and related approximations provide useful and fast implementations in the case of the normal
distribution, where much effort has been devoted to analysing this single universal distribution with no embedded
parameters, and a one-off composite rational approximation can be developed in detail. However, other cases of
interest do involve other parameters, e.g. the degrees of freedom in the case of the Student, and two parameters
in the case of the beta distribution, and so far as we are aware there is little known about rational or other
approximations that have error properties that are managed uniformly in the distributional parameters. One
outcome of this note is the capability to manage distributional parameters for several key distributions.
Furthermore, the emergence of 64-bit computers and the associated 96- and 128-bit standards of arithmetic
also requires us to reconsider the use of standard existing approximations for quantiles, often based on rational
approximations, as providing appropriate representations for these functions in high-precision environments.
The methods we shall present here are capable of providing both arbitrary-precision benchmarks for new high-
precision approximations, as well as components of real-time simulations.
We regard it as “mathematically pleasing” that this aspect of probability theory can also be discussed in
the same setting as the classical theory of special functions. Power series methods are at the heart of many
aspects of applied mathematics. For example, the simple solutions of the Schroedinger equation rely for their
interpretation on a discussion of their power series, and their termination to a polynomial or a certain type of
convergence behaviour leads to an understanding of energy levels. Traditionally the interest in a power series
approach to special functions has been focused on linear systems, and where one can derive explicit formulae
for coefficients in terms of some combination of factorials. At this stage in our analysis our characterization
of the the power series coefficients will be through a non-linear recursion associated with a non-linear ordinary
differential equation. But this will allow numerical or symbolic computation to proceed.
The power series method is of course local in character. However, we shall show that for the cases of the
Student and beta distributions, just two series will suffice for global applications. This work has the novel feature
that we can treat a non-linear ODE through almost entirely analytical methods, leaving a computer to solve
the iteration as far as is needed and add up the series.
The use of differential equations and series methods for the analysis of quantile functions has its origins
in the earlier work of Hill and Davis [12], and Abernathy and Smith [1]. This earlier work developed differ-
ential recursions with the emphasis on Cornish-Fisher expansions. In contrast our own approach will obtain
direct algebraic recursions for the quantile functions themselves, which are readily adapted for computation.
Applications to Cornish-Fisher expansions do exist and will be given elsewhere.

1.2 Background to the dynamic analysis

There are many situations in modern applied mathematics where one is interested in a time-dependent proba-
bility distribution, and many types of dynamics that might be postulated. This paper will give just one example
based on a one-dimensional SDE. We shall derive the “quantilized Fokker-Planck equation”, or QPFE, by a
transformation of the ordinary Fokker-Planck equation for a one-dimensional stochastic process. Then solutions
will be developed.
The development of direct evolution equations for quantile functions seems to have a somewhat sparse
presence in the literature. The authors are aware of work by Toscani and co-workers, and in particular a paper
by Carrillo and Toscani [5]. Other discussions have been given by Toscani and co-workers in [27], and most
recently, [3]. The motivation in that work is a one-dimensional approach to mass transportation problems.
Dassios has studied temporal quantiles – see [8] and the references therein, following on from earlier work by
Miura [18]. There is also the distinct concept of a quantile process, as discussed in [7]. The focus of our paper
is, rather, to consider the time evolution of the spatial quantile function. The direct calculation of such spatial
quantile functions would afford a new route to Monte Carlo simulation, and given that quantiles may be robustly
estimated from data, there is also the potential for estimating SDE structures from data, via the use of a QFPE.

1.3 Outline of the paper

The plan of this work is as follows. In section two we derive non-linear ordinary differential equations for
some key distributions. In Section three we derive power series solutions, and give a global solution for some
distributions. Matters to do with some practical implementations are deferred to Appendix A, and technical
details on tails are given in Appendix B. Section four gives a brief discussion on classification aspects and the
bivariate form. Section five gives an introduction to the dynamic analysis. Section six presents conclusions and
further directions, and links to web resources.
4 G. Steinbrecher & W.T. Shaw

2 Quantile statics: ODEs for quantile functions

We will treat this very simply by working through different choices of density function. In each case all we do is
first express the derivative of the quantile function w as a the reciprocal of the usual density function expressed in
terms of w, then keep differentiating until we obtain a closed differential relation, which is generically non-linear.
In each case, if f (x) is the probability density function, the first order quantile ODE is just

dw 1
= , (1)
du f (w)

where w(u) is the quantile function considered as a function of u, with 0 ≤ u ≤ 1. We then differentiate again
to find, at least for some key common cases, a simple second order non-linear ODE.

2.1 The Normal distribution

In this case if w(u) denotes the quantile function as a function of u, then

dw √ w2
 
= 2π exp . (2)
du 2

Now we differentiate again, to obtain

2
d2 w √ w2
  
dw dw
= 2π exp w =w , (3)
du2 2 du du

so we end up with and ODE and “centre” conditions,

2
d2 w

dw
=w ,
du2 du (4)
√
w(1/2) = 0 , w0 (1/2) = 2π .

This ODE is equivalent that posted for inverse erf on the web [13].

2.2 The Student distribution - direct approach

In this case if w(u) denotes the quantile function as a function of u, then

dw √ Γ( n2 ) n+1
= nπ n+1 (1 + w2 /n) 2 . (5)
du Γ( 2 )

Now we differentiate again, to obtain

2
d2 w √ Γ( n2 ) n + 1

n−1 dw n+1 1 dw
2
= nπ n+1 (1 + w2 /n) 2 w = w , (6)
du Γ( 2 ) n du n (1 + w2 /n) du

so we end up with and ODE and “centre” conditions,

2
w 2 d2 w

1 dw
(1 + ) = (1 + )w ,
n du2 n du
(7)
√ Γ( n2 )
w(1/2) = 0 , w0 (1/2) = nπ n+1 .
Γ( 2 )

This ODE and the conditions reduce to those for the normal case as n → ∞.
Quantile Mechanics 5

2.3 The beta distribution and indirect Student

In the case of the beta distribution with paramaters a, b, then if w(u) denotes the quantile function as a function
of u, then
dw
= B(a, b)w1−a (1 − w)1−b . (8)
du
A further differentiation gives us
2
d2 w
    
1−a 1−b dw
= − . (9)
du2 w 1−w du

We differentiate again, reorganize, and obtain the ODE

2
d2 w

dw
w(1 − w) − (1 − a + w(a + b − 2)) =0. (10)
du2 du

This ODE is the same that posted for the inverse beta function on the web [14].
The boundary conditions may be stated as a combination of various conditions:

w(0) = 0 , w(u) ∼ [auB(a, b)]1/a as u → 0 ,

(11)
w(1) = 1 , w(u) ∼ 1 − [b(1 − u)B(a, b)]1/b as u → 1 .

The choice a = n/2, b = 1/2 is special in that by a change of variables we may recover the Student distribution.
The details of this will be given in the next section.

2.4 The Gamma distribution

In this case if w(u) denotes the quantile function as a function of u, then, for p > 0,

dw
= Γ(p)w1−p exp(w) (12)
du
Now we differentiate again, reorganize as before, and obtain the ODE
2
d2 w

dw
w − (w + 1 − p) =0. (13)
du2 du

The boundary conditions at the origin take the form:

w(0) = 0 , w(u) ∼ [uΓ(p + 1)]1/p as u → 0 . (14)

3 Power Series Solutions of the ODEs

In this section we explore the solution of the non-linear ODEs by the standard method of power series. First
we review the argument for the normal case, which appears to be special in that the recursion is reducible to a
quadratic relation.

3.1 The power series for the normal quantile

There are two ways of approaching this. First we give the cubic recursion algorithm. Then we give the simplified
quadratic method of [24]. First we set √
v = 2π(u − 1/2) , (15)
and note the scale invariance of the independent variable in the ODE allows us to write the problem as:
2
d2 w

dw
=w ,
dv 2 dv (16)
0
w(0) = 0 , w (0) = 1 .
6 G. Steinbrecher & W.T. Shaw

We now assume that

∞
X
w= cp v 2p+1 , (17)
p=0

and substitute into the differential equation. After some careful but essentially easy manipulations we obtain
an explicit forward cubic recurrence
(p−1) (p−k−1)
1 X X
cp = (2l + 1)(2p − 2k − 2l − 1)ck cl cp−k−l−1 . (18)
(2p)(2p + 1)
k=0 l=0

Next we review the optimized approach due to one of us [GS] noted in [24] for the error function. Here we give
the argugment normalized for the quantile function. The ODE may be recast as
 
d 1
+ w(v) = 0 . (19)
dv w0 (v)
So now let Z v
W (v) = w(t)dt . (20)
0
Integrating and using the boundary conditions gives
1
+ W (v) = 1 , (21)
w0 (v)
and so
w0 (v) − w0 (v)W (v) = 1 . (22)
So now we have a quadratic identity. We now assume that
∞
X wp
w= v 2p+1 (23)
p=0
(2p + 1)

Simplification of these last two relations gives the quadratic recurrence

p
1X wj wp−j
wp+1 = . (24)
2 j=0 (j + 1)(2j + 1)

This is analogous to the original expression in [24], where some trivial relative factors of 2 arise from the
consideration here of the normal quantile, whereas in [24] the inverse error function was considered.

3.2 The power series for the Student quantile

First of all we set
√ Γ( n2 )
v= nπ (u − 1/2) , (25)
Γ( n+1
2 )
and exploit the scale invariance of the ODE to write the problem as:
2
w 2 d2 w

1 dw
(1 + ) = (1 + )w ,
n dv 2 n dv (26)
w(0) = 0 , w0 (0) = 1 .

We now assume
∞
X
w= cp v 2p+1 , (27)
p=0

and substitute into the differential equation. After some algebraic manipulations we obtain an explicit forward
cubic recurrence:
(2p)(2p + 1)cp =
Quantile Mechanics 7

(p−1)(p−k−1)   
X X 1 2k(2k + 1)
ck cl cp−k−l−1 1+ ((2l + 1)(2p − 2k − 2l − 1)) − . (28)
n n
k=0 l=0

It is readily verified that the first few terms of the series obtained by this recursion correspond to those obtained
by direct inversion of the CDF power series (see e.g. [22]). Now it is possible to generate many more terms with
ease. Similarly, cases with closed-form solutions, n = 1, 2, 4 also confirm the recursion.

3.3 The power series for the beta quantile

One can consider working from either end. But we shall work from u = 0 and then use symmetry to infer the
other case.
2
d2 w

dw
w(1 − w) 2 = (1 − a + w(a + b − 2)) ,
du du (29)
w(0) = 0 , w(u) ∼ [auB(a, b)]1/a as u → 0 .

As before, rescaling invariance of the ODE allows us to set

v = auB(a, b) , (30)

so that the system is now

2
d2 w

dw
w(1 − w) 2 = (1 − a + w(a + b − 2)) ,
dv dv (31)
1/a
w(0) = 0 , w(u) ∼ v as v → 0 .

Some experimentation suggests that a power series in x = v 1/a is usefu, so we change independent variable to
x. We then find that, with now w = w(x),
2
d2 w (1 − a) dw
  
dw
w(1 − w) + = (1 − a + w(a + b − 2)) ,
dx2 x dx dx (32)
w(0) = 0 , w0 (0) = 1 .

Our experimentation suggests that we may now substitute

∞
X
w(x) = cp xp , (33)
p=1

with c1 = 1, and after some manipulations we find that there is a cubic recursion of the form, for p ≥ 2:

cp [p2 + (a − 2)p + (1 − a)] =

p−1
X
(1 − δp,2 ) ck cp+1−k [(1 − a)k(p − k) − k(k − 1)]
k=2 (34)
p−1 p−k
X X
+ ck cn cp+1−k−n [k(k − a) + (a + b − 2)n(p + 1 − k − n)] .
k=1 n=1

The quadratic term only contributes for p ≥ 3. If we work out the first few we find that

b−1
c1 = 1 , c2 = ,
a+1

(35)
(b − 1) a2 + (3b − 1)a + 5b − 4
c3 = .
2(a + 1)2 (a + 2)

which is in accordance with known values [15].

8 G. Steinbrecher & W.T. Shaw

3.4 The tail power series for the Student quantile

Now consider the beta distribution with a = n/2 and b = 1/2. This is an important case as it allows us to
generate the Student distribution with n degrees of freedom. The quantile function for the Student distribution
may be written as v  
 u
1 u  1
u
wTn (u) = sgn u −
− 1 , (36)
IIf−1u< 1 ,2u,2(1−u) n2 , 12
tn
2
[ 2 ]
where the beta quantile function w(u) with parameters a, b is just

w(u) = Iu−1 (a, b) . (37)

So the beta distribution covers the Student T in two pieces. If we focus on the region u > 1/2, then we have
v !
u
u 1
wTn (u) = tn −1 n 1

−1 , (38)
I2(1−u) 2, 2

so we see that the tail series for the Student is given by the expansion of the beta around u = 0, but with
u → 2(1 − u). So we take
 2/n
n 1
x = n(1 − u)B( , ) , (39)
2 2
and we form the series
∞
X
wβ = dp xp , (40)
p=1

where d1 = 1, and for p ≥ 2:

n n
dp [p2 + ( − 2)p + (1 − )] =
2 2
p−1
X h n i
(1 − δp,2 ) dk dp+1−k (1 − )k(p − k) − k(k − 1)
2 (41)
k=2
p−1 p−k  
X X n n 3
+ dk dm dp+1−k−m k(k − ) + ( − )m(p + 1 − k − m) .
m=1
2 2 2
k=1

The quadratic term only contributes for p ≥ 3. Then we finally calculate

q
wT (u) = n(1/wβ − 1) . (42)

3.5 The power series for the gamma quantile

In this case the boundary conditions at the origin indicate the change of variable
1/p
v = (uΓ(p + 1)) , (43)

and this maps the ODE to

2
d2 w (1 − p) dw
  
dw
w + − (w + 1 − p) =0. (44)
dv 2 v dv dv
The boundary conditions at the origin take the form

w(0) = 0 , w0 (0) = 1 . (45)

We now try a series of the form

∞
X
w(v) = gn v n , (46)
n=1
Quantile Mechanics 9

with g1 = 1. The non-linear recurrence is found to be:

n n−k+1
X X
n(n + p)gn+1 = gk gl gn−k−l+2 l(n − k − l + 2)
k=1 l=1
n
(47)
X
− ∆(n) gk gn−k+2 k[k − p − (1 − p)(n + 2 − k)] ,
k=2

where ∆(n) = 0 if n < 2 and ∆(n) = 1 if n ≥ 2. . The first few terms are
1 3p + 5
g1 = 1 , g2 = , g3 = , (48)
p+1 2(p + 1)2 (p + 2)

and some experimentation with many terms confirms that if p = 1 then gn = n1 in accordance with the exact
solution, w(u) = − log(1 − u), for this case. Another important case is that of p = 1/2, when we have the χ21
distribution, with v = π4 u2 , and the coefficent sequence
 
2 26 176 8138 286816 28631908 480498944
gn = 1, , , , , , , ,... . (49)
3 45 315 14175 467775 42567525 638512875

4 Classification and generalization

All of the cases considered here arise from ODEs of the form
2
d2 w

dw
= R(w) , (50)
du2 du

where R(w) is a simple rational function. We shall say that the quantile is rational and of type (p, q) if R = P/Q
where P, Q are polynomials with no common factors of degree p, q. We can classify as follows:
• p = q = 0: exponential distribution;
• p = 1, q = 0: normal distribution;
• p = 1, q = 1: gamma distribution;
• p = 1, q = 2: beta and Student distribution.
More generally, given a density f (w), the associated quantile function satisfies a pair of first and second order
differential equations
2
d2 w

dw 1 dw
= , = H(w) , (51)
du f (w) du2 du
where
d
H(w) = −
log{f (w)} . (52)
dw
Notably1 , the cases where H is a rational function of type up to (1, 2) are closely related to Pearson’s system
of distributions [29, 19]. Pearson’s system, in our notation, is indeed characterized by the choices of a, b, c, m in
the function
w−m
H(w) = . (53)
a + bw + cw2
The distributions arising from this, Pearson’s “Types”, are summarized on p. 480-481 of [9] and discussed
in detail by Johnson and Kotz [17]. The function H is not rational in general. An algebraic case is easily
realized as Bagnold’s hyperbolic distribution, and more general functions appear if we consider, for example,
Barndorff-Nielsen’s generalized hyperbolic distribution and special cases such as the variance gamma. Further
work is needed to explore the optimal routes for solving such cases, and to establish when a power series method
remains viable. These representations also open up avenues based on numerical differential equation solution
techniques that are complementary to numerical root-finding methods based on the CDF.
1 We are grateful to Prof. Luc. Devroye for this observation.
10 G. Steinbrecher & W.T. Shaw

4.1 Multivariate quantiles

We shall only make brief comment on the multivariate case. Chaper 13 of the book by Gilchrist contains an
excellent summary of bivariate quantile distributions. Some clues as to how to proceed are also given in texts
on copula theory in the context of copula simulation, for example Chapter 6 of the book by Cherubini et al [6].
Working in two dimensions, the most natural approach would appear to be to consider the bivariate extension of
Eqn. (72), in the form suggested by Gilchrist. In our notation we proceed as follows. We form a vector quantile
(w1 (u1 , u2 ), w2 (u1 , u2 )). Then given a probability density function f (w1 , w2 ), the basic first order quantile ODE
is then given by
∂(w1 , w2 ) 1
= , (54)
∂(u1 , u2 ) f (w1 , w2 )
subject to suitable boundary conditions. This of course is just the standard Jacobian transformation for a
change of variables. It is an interesting question to consider the usefulness of taking further derivatives of this
relationship. It is also interesting to note that there potentially many interesting solutions to Eqn. (75). One
elementary illustration of this, the details of which we leave to the reader, is to contrast the bivariate normal
based on a pair of univariate quantiles with the Box-Muller simulation method, as discussed in Section 7.2 of
[21], where a check of Eqn. (75) is exhibited for the case of two independent normal variables. One can check
that it is also the case that
p
w1 (u1 , u2 ) = w0 (u1 ) , w2 (u1 , u2 ) = w0 (u2 ) 1 − ρ2 + ρw0 (u1 ) , (55)

is also a solution of Eqn. (75), as would be expected, where w0 satisfies Eqn. (2) and f is given by
 
1 1 2 2
f (w1 , w2 ) = exp − (w + w − 2ρw w
1 2 ) . (56)
2(1 − ρ2 ) 1 2
p
2π 1 − ρ2

So we can build two quite different mappings from the unit square to the bivariate normal. In general, multi-
variate and non-normal case, quantiles cannot of course be built up by such elementary superposition.

5 Quantile dynamics
There are many different ways in which the evolution of a time-dependent quantile function might be controlled.
Given the importance of stochastic analysis in modern applied mathematics, we shall consider the case where the
governing dynamics is defined by an elementary stochastic differential equation, or SDE. Let xt be a stochastic
process governed by the SDE
dxt = µ(xt , t)dt + Σ(xt , t)dWt , (57)
where µ, Σ are deterministic functions of x and t, and Wt is a standard Brownian motion. Let f (x, t) denote
the probability density function consistent with this SDE. The quantile function associated with this system is
Q(u, t), where 0 ≤ u ≤ 1. It is defined by the integral condition:
Z Q(u,t)
f (x, t)dx = u . (58)
−∞

First we differentiate this relation with respect to u, and obtain the partial differential constraint

∂Q(u, t)
f (Q(u, t), t) =1 . (59)
∂u
We differentiate this relation again, and obtain the following second order PDE,
2
∂ 2 Q(u, t)

∂f (Q(u, t), t) ∂Q(u, t)
+ f (Q(u, t), t) =0. (60)
∂Q ∂u ∂u2

This may be reorganized into the equation

2
∂ 2 Q(u, t)

∂ log(f (Q(u, t), t)) ∂Q(u, t)
=− . (61)
∂u2 ∂Q ∂u
Quantile Mechanics 11

In the case of no time dependence, this is readily recognized as the second order non-linear ODE derived and
solved previously. In the time-dependent case we go back to the condition of Eqn. (58) and now differentiate
with respect to time. The fundamental theorem of calculus gives, first,
Z Q(u,t)
∂Q ∂f (x, t)
f (Q(u, t), t) + dx = 0 . (62)
∂t −∞ ∂t
We use Eqn. (59) to write this as:
Z Q(u,t)
∂Q ∂Q ∂f (x, t)
=− dx . (63)
∂t ∂u −∞ ∂t
Next we use the Fokker-Planck (forward Kolmogorov) equation in the form [10]
 
∂f (x, t) ∂ 1 ∂
= −µf + (Σ2 f ) , (64)
∂t ∂x 2 ∂x
and observe that we can do the integration directly, and obtain the following:
∂Σ2 (Q, t)
  
∂Q ∂Q 1 ∂f
=− −µ(Q, t)f (Q, t) + Σ2 (Q, t) +f . (65)
∂t ∂u 2 ∂Q ∂Q
Finally we use our previous Equations (59)-(60) to eliminate f and its derivative completely, leaving us with
−2 2
1 ∂Σ2 Σ2 (Q, t) ∂Q

∂Q ∂ Q
= µ(Q, t) − + . (66)
∂t 2 ∂Q 2 ∂u ∂u2
This is the quantilized Fokker-Planck equation, or QFPE for short, associated with the SDE of Eqn. (57). It is
a non-linear PDE, and is the form of the Carrillo-Toscani equation [5] appropriate to solutions of an SDE.

5.1 Elementary solutions of the QFPE

Despite its non-linearities, we known that the QFPE of Eqn. (66) must have some elementary Gaussian solutions
that will now be verified. First we explore the QFPE for arithmetical Brownian motion. In this case we have
µ and Σ constant and it is easy to check that
−2 2
Σ2 ∂Q

∂Q ∂ Q
=µ+ (67)
∂t 2 ∂u ∂u2
has an obvious solution √
Q(u, t) = c + µt + Σ tw(u) , (68)
where w(u) is the static normal quantile function we constructed in Section 3.1. Now consider Σ(t) as indepen-
dent of Q but time-dependent, and the class of quantiles of the form
Q(u, t) = a(t) + b(t)w(u) . (69)
Substitution into the QFPE and using Eqn. (4) gives
da db Σ2 (t)
+ w(u) = µ(a(t) + b(t)w(u), t) + w(u) . (70)
dt dt 2b(t)
It is clear that this equation can be separated into two ODEs if µ is at most linear in Q. So we assume
µ = m1 (t) − m2 (t)Q , (71)
and have the two ordinary differential equations:
da db2
+ m2 (t)a = m1 (t) , + 2m2 (t)b2 = Σ2 (t) , (72)
dt dt
that are trivially solved by the use of integrating factors. If for example m1 and m2 are constant, we obtain the
anticipated quantile solution for the OU process for time-dependent volatility in the form
s
Z t
m 1
Q = Q0 e−m2 t + (1 − e−m2 t ) + w(u)e−m2 t dsΣ2 (s)e2m2 s . (73)
m2 0
12 G. Steinbrecher & W.T. Shaw

5.2 Geometric Brownian motion

This follows from the analysis of arithmetical Brownian motion by taking logarithms. If we consider a typical
financial representation, and allow for deterministic time dependence, we have

µ = θ(t)Q , Σ = σ(t)Q , (74)

and so the QFPE Eqn. (66) becomes

−2
σ 2 (t)Q2 ∂2Q

∂Q ∂Q
= (θ(t) − σ 2 (t))Q + . (75)
∂t 2 ∂u ∂u2

We make the assumption

Q(t, u) = Q0 exp{a(t) + b(t)w(u)} , (76)
and are lead quickly to the condition

σ 2 (t) w(u) σ 2 (t) σ 2 (t)

 
da db
+ w(u) = θ(t) − σ 2 (t) + + 1 = θ(t) − + w(u) , (77)
dt dt 2 b 2 2b

with solution s
Z t Z t
σ 2 (s)

a(t) = θ(s) − ds , b(t) = dsσ 2 (s) . (78)
0 2 0

This solution, and the other Gaussian examples, are clearly what might have been expected from a knowledge
of the nature of the solution of the associated SDEs. Our purpose here was simply to illustrate the direct
solution of the QFPE using classical separation of variables methods utilising a knowledge of the non-linear
ODEs satisfied by the Gaussian quantile.

5.3 Equilibrium quantiles

If a solution to the SDE has reached an equilibrium distribution then the QFPE reduces to
−2 2
1 ∂Σ2 Σ2 (Q, t) ∂Q

∂ Q
− µ(Q, t) = . (79)
2 ∂Q 2 ∂u ∂u2

We can consider Pearson’s differential classification in relation to this equilibrium structure. We leave the reader
to explore some of the better-known special cases, and consider here the general Pearson category and a special
case leading to a Student t-distribution as the equilibrium. We consider an SDE containing independent sources
of additive and multiplicative noise, in the form

dxt = (µ1 − µ2 xt )dt + σ1 dW1t + σ2 xt dW2t , (80)

where the Wi are correlated such that E[dW1t dW2t ] = ρdt. Then Eqn. (80) is equivalent to a one-dimensional
system in the form
1
dxt = (µ1 − µ2 xt )dt + (σ12 + σ22 x2t + 2ρσ1 σ2 xt ) 2 dWt . (81)
From this we can read off the functions in the QFPE as

µ = µ1 − µ2 Q , Σ2 = σ12 + σ22 Q2 + 2ρσ1 σ2 Q . (82)

The equilibrium QPFE then gives the non-linear ODE

−2
∂ 2 Q ∂Q 2[(ρσ1 σ2 − µ1 ) + (σ22 + µ2 )Q]

= , (83)
∂u 2 ∂u σ12 + σ22 Q2 + 2ρσ1 σ2 Q

which corresponds to an interesting subset of the Pearson system embodied by Eqns. (51) and (53). A particular
case of interest is obtained by considering µ1 = 0 = ρ, so that we obtain the SDE

dxt = −µ2 xt dt + σ1 dW1t + σ2 xt dW2t , (84)

Quantile Mechanics 13

where the Wit are independent, and the quantile ODE reduces to
−2
∂ 2 Q ∂Q 2(σ 2 + µ2 )Q

2
= 22 2 2 . (85)
∂u ∂u σ1 + σ2 Q

If we compare this with Eqn. (7) we see that we have a Student quantile with
σ1
Q= p w(u) , (86)
σ22 + 2µ2

where w(u) is the standard Student quantile satisfying Eqn. (7) with degrees of freedom
µ2
n=1+2 (87)
σ22

So it is clear that we need µ2 > 0 for this to be a Student distribution. This of course corresponds to the
requirement that the underlying SDE mean-revert to the origin, and this mean-reversion condition in turn
allows an equilibrium to establish. This equilibrium origination of the standard Student distribution arises
naturally in the study of the simplest linear and stochastic approximation of very complex physical systems,
with manifest turbulence and self-organized criticality character. In particular the Student distribution arises as
the equilibrium distribution, in special limiting cases of the stochastic processes studied in [25, 26]. The faster
the mean-reversion rate is compared to the multiplicative volatility, the closer the system is to the normally
distributed limit. The Student distribution also arises naturally in the modelling of asset returns [20, 22].

6 Conclusions and Resources

We have shown how to characterize the key quantile functions of mathematical statistics through non-linear
ordinary differential equations and shown how they may be solved via a power series whose coefficients are
determined by a non-linear, usually cubic, recurrence relation. In the case of the Student and beta distributions,
we are able to give both central and tail series, and for these distributions the pair of relevant series provide a good
basis for numerical implementation. In the case of the normal and gamma, we have the start of a representation
valid in arbitrary precision environments. Any of these methods may be used for precision benchmarking and
for the test of rational approximations, and indeed to augment rational methods with polynomials in the central
zone.
In contrast with rational approximation theory, these power series methods allow a meaningful mathematical
extension of the concept of the quantile to the complex plane. Whether the “complex quantile function” has a
useful application is an open question.
We have also considered the non-linear integrated form of the Fokker-Planck equation, as introduced by
Carrillo and Toscani [5], and demonstrated how the resulting non-linear PDE may be solved in some cases of
interest.

6.1 Internet resources

Our intention is to provide a growing list of links to on-line computational resources for the implementation of
these and related methods. Some links are provided in the references. Specific discussions are also available as
follows in Mathematica, C++ and Fortran 95. These will evolve as our knowledge about optimal implementa-
tions grows.

• A Mathematica file with investigations of some these methods is at:

http://www.mth.kcl.ac.uk/~shaww/web_page/papers/quantiles/QuantileDemos.nb

• C++ code for the central normal power series is at:

http://www.mth.kcl.ac.uk/~shaww/web_page/papers/quantiles/normalquantile.cpp

• Quad precision F95 code for the central normal power series is at:
14 G. Steinbrecher & W.T. Shaw

http://www.mth.kcl.ac.uk/~shaww/web_page/papers/quantiles/normalquantile.f95

• Quad precision F95 code for the Student case (central & tail) is at:

http://www.mth.kcl.ac.uk/~shaww/web_page/papers/quantiles/studentquantile.f95

In addition a web page is being maintained with further examples and related article, with the entry point at

http://www.mth.kcl.ac.uk/~shaww/web_page/papers/quantiles/Quantiles.htm

Acknowledgments
We wish to thank Prof. L Devroye for his comment on the Pearson classification, and Prof. W. Gilchrist for
several useful remarks. Our understanding of the history of the approaches to quantiles has benefited from
comments by Dr. A. Dassios and Prof. G. Toscani. WS wishes to thank Absoft corporation for their help
on quadruple precision Fortran. The Fortran 95 associated with this work was developed and tested with the
Absoft Fortran for MacOS on Intel compiler. The C++ code was developed with the Bloodshed Dev C++
compiler.

References
[1] R.W. Abernathy and R.P. Smith, Applying Series Expansion to the Inverse Beta Distribution to Find
Percentiles of the F-Distribution, ACM Transactions on Mathematical Software 19, 4, 474-480, 1993.

[2] P. J. Acklam An algorithm for computing the inverse normal cumulative distribution function,
http://home.online.no/~pjacklam/notes/invnorm/

[3] G. Aletti, G. Naldi, G. Toscani, First-order continuous models of opinion formation, SIAM J. Appl.
Math. 67, 3, 827-853, 2007.

[4] C.I. Bliss, The method of probits. Science 39, 38-39, 1934.

[5] J.A Carrillo, and G. Toscani Wasserstein metric and large-time asymptotics of nonlinear diffusion
equations, in New trends in mathematical physics, pp. 234244 World Sci. Publ., Hackensack, NJ , 2004.

[6] U. Cherubini, E. Luciano, W. Vecchiato, Copula Methods in Finance, Wiley, 2004.

[7] M. Csorgo Quantile Processes with Statistical Applications SIAM, 1983.

[8] A. Dassios On the quantiles of Brownian motion and their hitting times Bernoulli 11, 1, pp. 29-36, 2005.

[9] L. Devroye Non-uniform random variate generation, Springer 1986. Out of print - now available on-line
from the author’s web site at
http://cg.scs.carleton.ca/~luc/rnbookindex.html

[10] A. Etheridge, A Course in Financial Calculus, Cambridge University Press, 2002.

[11] W. Gilchrist, Statistical Modelling with Quantile Functions, CRC Press, 2000.

[12] G.W. Hill, A.W. Davis, Generalized Asymptotic Expansions of Cornish-Fisher Type, Annals of Math-
ematical Statistics, 39, 4, 1264-1273, 1968.

[13] http://functions.wolfram.com/GammaBetaErf/InverseErf/13/01/
[14] http://functions.wolfram.com/GammaBetaErf/InverseBetaRegularized/13/01/
[15] http://functions.wolfram.com/GammaBetaErf/InverseBetaRegularized/06/01/
[16] P. Jäckel, Monte Carlo methods in finance, Wiley, 2002.
Quantile Mechanics 15

[17] N.L. Johnson and S. Kotz. Distributions in Statistics. Continuous Univariate Distributions, Wiley,
1970.
[18] R. Miura. A note on a look-back option based on order statistics, Hitosubashi Journal of Commerce and
Management, 27, 15-28, 1992.
[19] K. Pearson Second Supplement to a Memoir on Skew Variation. Phil. Trans. A 216, 429-457, 1916.
[20] K. Fergusson, E. Platen, On the Distributional Characterization of daily Log-returns of a World Stock
Index, Applied Mathematical Finance, 13 (1), 19-38, March 2006.
[21] W.H. Press, S.A.Teukolsky, W.T. Vetterling, B.P. Flannery, Numerical Recipes in C++, Cam-
bridge University Press, 2005 edition.
[22] W. T. Shaw, Sampling Student’s T distribution - use of the inverse cumulative distribution function.
Journal of Computational Finance, Vol. 9, No. 4, 2006
[23] W. T. Shaw, Refinement of the Normal quantile, King’s College working paper,
http://www.mth.kcl.ac.uk/~shaww/web_page/papers/NormalQuantile1.pdf
[24] G. Steinbrecher, Taylor expansion for inverse error function around origin, Univ. of Craiova working
paper, Jan. 2002.
[25] G. Steinbrecher and B. Weyssow, Generalized Randomly Amplified Linear System Driven by Gaus-
sian Noises: Extreme Heavy Tail and Algebraic Correlation Decay in Plasma Turbulence. Physical Review
Letters, 92, 12, 125003-125006, 2004.
[26] G. Steinbrecher and B. Weyssow, Extreme anomalous particle transport at the plasma edge, Univ.
of Craiova/Univ Libre de Bruxells working paper, May. 2007.
[27] G. Toscani ad H. Li, Long-time asymptotics of kinetic models of granular flows, Archiv. Ration. Mech.
Anal. 172, 407-428, 2004.
[28] Wichura, M.J., Algorithm AS 241: The Percentage Points of the Normal Distribution. Applied Statistics,
37, 477-484, 1988.
[29] On-line discussion of Pearson’ s system of distributions defined via differential equations.
http://mathworld.wolfram.com/PearsonSystem.html

Appendix A: practicalities and implementations

In this section we discuss some issues related to computer implementation. We can consider both symbolic
and exact implementations, and implementations in traditional floating-point arithmetic with various levels of
precision.

The normal case

Here the raw recursions given here and in [24] will overflow or underflow floating point number bounds eventually.
Fortunately there are simple rescalings that may be used to cure this problem. If we work in terms of the variable
v̂ = 2u − 1 , (88)
then the power series will exhibit a unit radius of convergence. We know it will blow up at u = 0 and u = 1.
Then we form the power series r ∞
πX dp
w= v̂ 2p+1 , (89)
2 p=0 (2p + 1)
with d0 = 1. In terms of these variables the quadratic recurrence becomes
p
πX dj dp−j
dp+1 = . (90)
4 j=0 (j + 1)(2j + 1)
16 G. Steinbrecher & W.T. Shaw

With such a scheme the ratio dp+1 /dp → 1 as p → ∞, and numerical implementations confirm this. The limits
of how far the series can be taken are then just a matter of computation time. The second observation is that
the sum involves the product dj dp−j twice, so that one can optimize by summing half-way and taking care to
note whether p is odd or even. The recurrence can be written
p−1
b 2 c  
π X 1 1
dp+1 = dj dp−j +
4 j=0 (j + 1)(2j + 1) (p − j + 1)(2p − 2j + 1)
(91)
π d2p
2
+ If(p is even, , else 0) .
4 ( p2 + 1)(p + 1)

In our initial implementations of the normal case in C++ and F95, attention has been confined to the central
power series, and with very simple series termination criteria, involving termination when the next term to be
computed falls below a certain level. While this is appropriate in a rigorous sense for alternating series, it is
a temporary solution for monotone series such as those that are being considered here. The resulting error
characteristics are nevertheless encouraging, and we can cover a large proportion of the unit interval with a
precision that is more dependent on hardware and compiler choices than on the algorithm itself. In C++ the
level of precision achievable depends on the quality of the implementation of the “long double” data type. This
is not, regrettably, rigorously specified in the C++ standard. The demonstration C++ code cited in Section 6.1
works on 0.0005 < u < 0.9995 with maximum relative error less than 1.7 × 10−15 and on 0.0035 < u < 0.9965
with maximum relative error less than 2.5 × 10−16 ., when using the Bloodshed Dev C++ under Windows XP
on a Pentium M. Results with other hardware and/or compilers will vary, especially if a poor long double
implementation is used. For a more controlled implementation, we turn to quadruple precision Fortran 95.
Using the Absoft compiler for MacOS on Intel, we obtained a relative error less than 10−29 on the interval
0.0007 < u < 0.9993. All precision calculations were assessed by reference to the inverse error function built in
to Mathematica 5.2, using 40 significant figures for the calculations. We stress that such code is not necessarily
fast enough to use in live Monte Carlo simulation, but, at this stage in the development, is more appropriately
viewed as a means for benchmarking a new class of fast rational approximations.

The Student case

In this case we have a pair of power series for central and tail use and we may proceed directly to an imple-
mentation. One issue that has to be dealt with is the computation of the rescaling factor in the formula of
Eqn. (25), which we write as √
nπ Γ( n2 )
v= (2u − 1) = γn (2u − 1) . (92)
2 Γ( n+12 )
We use the identity p
n(n + 2)
γn+2 = γn (93)
n+1
to simplify the high-precision computation of this quantity. When n is an integer with n ≤ 10 we use a
precomputed value of γn . When n is a larger integer we recurse down to 9 or 10. q
When n is non-integer we
2
recurse up to a value larger than 1000 and then apply the result that as n → ∞, γn π is asymptotic to

1 1 5 21 399 869 39325 334477

1+ + 2
− 3
− 4
+ 5
+ 6
− 7
−
4n 32n 128n 2048n 8192n 65536n 262144n 8388608n8 (94)
28717403 59697183 8400372435 34429291905
+ + − − .
33554432n9 268435456n10 1073741824n11 17179869184n12

Appendix B: Tail issues for the normal and gamma

It is evident that we have found power series about central points (normal and Student distributions) and about
left end points (gamma and beta) and right end points (beta). The left end-point series for the beta also gives
us the tails for the Student case. This leaves the issue of tail series for the normal and gamma. While we can use
pure computational force to figure out many terms of the power series, a better treatment of the tails is needed.
The most important case is that of the normal distribution, or, by taking its square, the gamma distribution
Quantile Mechanics 17

with p = 1/2. There are various ways of thinking about this, based on what we have so far, but the neatest
seems to be to exploit the fact that the gamma quantile ODE is already set up for an asymptotic analysis. The
ODE for the gamma quantile with p = 1/2 is
2
d2 w

1 dw
= (1 + ) , (95)
du2 2w du
so we can see right away that as u → 1, w → +∞ and the asymptotic behaviour must be such that
2
d2 w

dw
∼ , (96)
du2 du
with a solution that we can infer from the exact solution for the exponential case, i.e.,

w ∼ − log[k(1 − u)], k > 0 . (97)

Before proceeding further we look at the known asymptotic properties of the forward gamma CDF, for further
insight. For large w the forward gamma CDF with p = 1/2 gives us

e−w
u∼1− √ , (98)
πw
or, in other words,
√ e−w
π(1 − u) ∼ √ , (99)
w
and taking logs we find that
√ 1
− log( π(1 − u)) ∼ w + log(w) , (100)
2
√
and we see that k = π. We now change independent variables to
√
v = − log[ π(1 − u)] (101)

and establish the differential equation

2
d2 w dw

1 dw
+ = (1 + ) , (102)
dv 2 dv 2w dv
with the asymptotic condition w ∼ v as v → ∞, i.e. as u → 1. It is possible to find asymptotic solutions of
these relations. After some experimentation we find that asymptotic solutions exist in the form
kX
max k+1
1 X
w ∼ v − log(v) + akl (log(v))l−1 v −k , (103)
2
k=1 l=1

with a sequentially solvable set of conditions relating the coefficients akl . While it is difficult to give an explicit
formula for the recursion, the coefficients are easily determined via symbolic computation. For example, with
kmax = 3 we obtain the first few terms as

1 log2 (v) 3 log(v) 7

   
log(v) 1 log(v) 1
w∼v− + − + 2 − +
2 v 4 2 v 16 8 8

1 log3 (v) 7 log2 (v) 17 log(v) 107

 
+ 3 − + − , (104)
v 48 32 16 48
√
with v = − log[ π(1 − u)]. This is the asymptotic form of the quantile function for the gamma distribution
with p = 1/2. If we want the normal quantile for the positive tail, a small change of variables is needed. In the
postive normal tail, the normal quantiles wΦ is given in terms of the gamma quantile wΓ1/2 , that we have just
computed, by q
wΦ (u) = 2wΓ1/2 (2u − 1) . (105)
So to summarize for the normal case in the positive tail, the calculation is:
18 G. Steinbrecher & W.T. Shaw

√
1. let v = − log[2 π(1 − u)];
2. Work about w as above;
√
3. calculate 2w for the normal quantile.

We stress again that these are asymptotic formulae. They will turn out to be extraordinarily accurate in the
deep tail, but less practical in the near tail zone. In the following table we look at the deep tail in more detail.
Consider the quantiles associated with the points u = 1 − 10−k , where k = 3, 4, . . . . We consider the results
from the simple third order formula above, the result from Mathematica’s internal high precision error function
(Mathematica V5.2), Wichura’s algorithm AS241 [28], and Microsoft Excel (Excel 2004 for Mac). Note that
in the last case Excel did not accept numbers as close to unity as supplied to the other algorithms. It is not
designed to treat such numbers. AS241 was implemented in Mathematica in uncompiled form. The “deep
tail” quality of the simple asymptote becomes clear from this table. It does better than the simple spreadsheet
implementation for values of u closer to unity that about 10−11 .

k Order 3 asymp Mathematica AS241 Excel

3 3.089631766 3.090232306 3.090232306 3.090232306
4 3.718896182 3.719016485 3.719016485 3.719016485
5 4.264854901 4.264890794 4.264890794 4.264890794
6 4.753410708 4.753424309 4.753424309 4.753424341
7 5.199331535 5.199337582 5.199337582 5.199337618
8 5.611998229 5.612001244 5.612001244 5.612001258
9 5.997805376 5.997807015 5.997807015 5.997807018
10 6.361339949 6.361340902 6.361340902 6.361340808
11 6.706022570 6.706023155 6.706023155 6.706022335
12 7.034483450 7.034483825 7.034483825 7.034479171
13 7.348795853 7.348796103 7.348796103 7.348680374
14 7.650627921 7.650628093 7.650628093 7.650018522
15 7.941345205 7.941345326 7.941345326 7.934736689
16 8.222082128 8.222082216 8.222082216 NA
17 8.493793159 8.493793224 8.493793224 NA
18 8.757290300 8.757290349 8.757290349 NA
19 9.013271116 9.013271153 9.013271153 NA
20 9.262340061 9.262340090 9.262340090 NA
21 9.505024960 9.505024983 9.505024983 NA
22 9.741789925 9.741789943 9.741789943 NA
23 9.973045605 9.973045620 9.973045620 NA
24 10.19915741 10.19915742 10.19915742 NA
25 10.42045219 10.42045220 10.42045220 NA
26 10.63722367 10.63722368 10.63722368 NA
27 10.84973699 10.84973700 10.84973700 NA
28 11.05823241 11.05823241 11.05823241 NA
29 11.26292848 11.26292848 11.26292848 NA
30 11.46402468 11.46402469 11.46402469 NA
31 11.66170368 11.66170368 11.66170368 NA
32 11.85613321 11.85613322 11.85613322 NA
33 12.04746778 12.04746779 12.04746779 NA
34 12.23585004 12.23585005 12.23585005 NA
35 12.42141204 12.42141204 12.42141204 NA