Introduction to polynomials chaos with NISP

Michael Baudin (EDF)

Jean-Marc Martinez (CEA)
Version 0.4
January 2013

Abstract
This document is an introduction to polynomials chaos with the NISP module for Scilab.

Contents
1 Orthogonal polynomials 4
1.1 Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Orthogonal polynomials for probabilities . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Hermite polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Legendre polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5 Laguerre polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6 Chebyshev polynomials of the first kind . . . . . . . . . . . . . . . . . . . . . . . . 19
1.7 Accuracy of evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.8 Notes and references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2 Multivariate polynomials 24
2.1 Occupancy problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2 Multivariate monomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Multivariate polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Generating multi-indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Multivariate orthogonal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.6 Tensor product of orthogonal polynomials . . . . . . . . . . . . . . . . . . . . . . 32
2.7 Multivariate orthogonal polynomials and probabilities . . . . . . . . . . . . . . . . 36
2.8 Notes and references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3 Polynomial chaos 39
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Truncated decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Univariate decomposition examples . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Generalized polynomial chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5 Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6 Notes and references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

1
4 Thanks 48

A Gaussian integrals 48
A.1 Gaussian integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
A.2 Weighted integral of xn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

Bibliography 51

2
 Copyright c 2013 - Michael Baudin
This file must be used under the terms of the Creative Commons Attribution-ShareAlike 3.0
Unported License:

http://creativecommons.org/licenses/by-sa/3.0

3
1 Orthogonal polynomials
In this section, we define the orthogonal polynomials which are used in the context of polynomial
chaos decomposition. We first present the weight function which defines the orthogonality of the
square integrable functions. Then we present the Hermite, Legendre, Laguerre and Chebyshev
orthogonal polynomials.

1.1 Definitions and examples

We will denote by I an interval in R. The interval [a, b] is the set of real numbers such that x ≥ a
and x ≤ b, for two real numbers a and b. The open interval (−∞, +∞) is also considered as an
interval, although its boundaries are infinite. The half-open intervals (−∞, b) and (a, +∞) are
also valid intervals.
Definition 1.1. ( Weight function of R) Let I be an interval in R. A weight function w is a
nonnegative integrable function of x ∈ I.

Example (Weight function for Hermite polynomials) The weight function for Hermite polyno-
mials is
 2
x
w(x) = exp − , (1)
2
for x ∈ R. This function is presented in the figure 1. Clearly, this function is differentiable and
nonnegative, since exp(−x2 ) > 0, for any finite x. The integral
Z √
exp(−x2 /2)dx = 2π (2)
R

is called the Gaussian integral (see the proposition A.1 for the proof).

Definition 1.2. ( Weighted L2 space in R) Let I be an interval in R and let L2w (I) be the set of
functions g which are square integrable with respect to the weight function w, i.e. such that the
integral
Z
kgk = g(x)2 w(x)dx
2
(3)
I

is finite. In this case, the norm of g is kgk.

Example (Functions in L2w (R)) Consider the Gaussian weight w(x) = exp(−x2 /2), for x ∈ R.
Obviously, the function g(x) = 1 is in L2w (R).
Now consider the function
g(x) = x (4)
for x ∈ R. The proposition A.3 (see in appendix) proves that
Z √
x2 exp(−x2 /2)dx = 2π. (5)
R

Therefore, g ∈ L2w (R).

4
 Figure 1: The Gaussian weight function f (x) = exp(−x2 /2).

Definition 1.3. ( Inner product in L2w (I) space) For any g, h ∈ L2w (I), the inner product of g
and h is
Z
hg, hi = g(x)h(x)w(x)dx. (6)
I

Assume that g ∈ We can combine the equations 3 and 6, which implies that the L2w (I)
L2w (I).
norm of g can be expressed as an inner product:
kgk2 = hg, gi . (7)
Example (Inner product in L2w (R)) Consider the Gaussian weight w(x) = exp(−x2 /2), for x ∈ R.
Then consider the function g(x) = 1 and h(x) = x. We have
Z Z
g(x)h(x)w(x)dx = x exp(−x2 /2)dx (8)
R R
Z Z 
2 2
= x exp(−x /2)dx + x exp(−x /2)dx . (9)
x≤0 x≥0

However,
Z Z
2
x exp(−x /2)dx = − x exp(−x2 /2)dx, (10)
x≤0 x≥0

by symmetry. Therefore,
Z
g(x)h(x)w(x)dx = 0, (11)
R

which implies hg, hi = 0.

5
 The L2w (I) space is a Hilbert space.
Definition 1.4. ( Orthogonality in L2w (I) space) Let I be an interval of R. Let g and h be two
functions in L2w (I). The functions g and h are orthogonal if

hg, hi = 0. (12)

Definition 1.5. ( Polynomials) The function Pn is a polynomial if there is a sequence of real

numbers αi ∈ R, for i = 1, 2, ..., n such that :
n
X
Pn (x) = αi xi , (13)
i=0

where x ∈ I. In this case, the degree of the polynomial is n.

Definition 1.6. ( Orthogonal polynomials) The set of polynomials {Pn }n≥0 are orthogonal poly-
nomials if Pn is a polynomial of degree n and:

hPi , Pj i = 0 (14)

for i 6= j.
We assume that, for all orthogonal polynomials, the degree zero polynomial P0 is equal to one:

P0 (x) = 1, (15)

for any x ∈ I.
Proposition 1.7. ( Integral of orthogonal polynomials) Let {Pn }n≥0 be orthogonal polynomials.
We have
Z Z
P0 (x)w(x)dx = w(x)dx (16)
I I

Moreover, for n ≥ 1, we have

Z
Pn (x)w(x)dx = 0 (17)
I

Proof. The equation 16 is the straightforward consequence of 15. Moreover, for any n ≥ 1, we
have
Z Z
Pn (x)w(x)dx = P0 (x)Pn (x)w(x)dx (18)
I I
= hP0 (x), Pn (x)i (19)
= 0, (20)

by the orthogonality property.

The figure 2 presents the function Hen (x)w(w), for n = 0, 1, ..., 5, where w is the Gaussian
weight w(x) = exp(−x2 /2). The proposition 1.7 states that the only n for which the integral is
nonzero is n = 0. Indeed, for n = 2, 4, i.e. for n even, we can see that the area of the segments
above zero is equal to the area of the segments below zero. This is more obvious for n = 1, 3, 5,
i.e. for n oddd, because the function Hen (x) is antysymmetric, i.e. Hen (−x) = −Hen (x).

6
Figure 2: The function Hen (x)w(x), for n = 0, 1, ..., 5, and the Gaussian weight w(x) =
2
exp(−x /2).

1.2 Orthogonal polynomials for probabilities

In this section, we present the properties of Pn (X), when X is a random variable associated with
the orthogonal polynomials {Pn }n≥0 .
Proposition 1.8. ( Distribution from a weight) Let I be an interval in R and let w be a weight
function on I. The function:
w(x)
f (x) = R , (21)
I
w(x)dx

for any x ∈ I, is a distribution function.

Proof. Indeed, its integral is equal to one:
Z Z
1
f (x)dx = R w(x)dx (22)
I I
w(x)dx I
= 1, (23)

which concludes the proof.

Example (Distribution function for Hermite polynomials) The distribution function for Hermite
polynomials is
 2
1 x
f (x) = √ exp − , (24)
2π 2

for x ∈ R. The function f (x) is presented in the figure 3.

7
 Figure 3: The Gaussian distribution function f (x) = √1 exp(−x2 /2).
2π

Proposition 1.9. ( Expectation of orthogonal polynomials) Let {Pn }n≥0 be orthogonal polyno-
mials. Assume that X is a random variable associated with the probability distribution function
f , derived from the weight function w. We have

E(P0 (X)) = 1 (25)

Moreover, for n ≥ 1, we have

E(Pn (X)) = 0 (26)

Proof. By definition of the expectation,

Z
E(P0 (X)) = P0 (x)f (x)dx (27)
ZI
= f (x)dx (28)
I
= 1, (29)

by the equation 22. Moreover, for any n ≥ 1, we have

Z
E(Pn (X)) = Pn (x)f (x)dx (30)
I Z
1
= R Pn (x)w(x)dx (31)
I
x(x)dx I
= 0, (32)

where the first equation derives from the equation 21, and the second equality is implied by
17.

8
Proposition 1.10. ( Variance of orthogonal polynomials) Let {Pn }n≥0 be orthogonal polynomials.
Assume that x is a random variable associated with the probability distribution function f , derived
from the weight function w. We have

V (P0 (X)) = 0. (33)

Moreover, for n ≥ 1, we have

kPn k2
V (Pn (X)) = R . (34)
I
w(x)dx

Proof. The equation 33 is implied by the fact that P0 is a constant. Moreover, for n ≥ 1, we have:

V (Pn (X)) = E (Pn (X) − E(Pn (X)))2

(35)
2


= E Pn (X) (36)
Z
= Pn (x)2 f (x)dx (37)
I Z
1
= R Pn (x)2 w(x)dx, (38)
I
w(x)dx I

where the first equality is implied by the equation 26.

Figure 4: Histogram of N = 10000 outcomes of Hen (X), where X is a standard normal random
variable, for n = 0, 1, ..., 5.

Example In order to experimentally check the propositions 1.9 and 1.10, we consider Hermite
polynomials. We generate 10000 pseudo random outcomes of a standard normal variable X, and

9
compute Hen (X). The figure 4 presents the empirical histograms of Hen (X), for n = 0, 1, ..., 5.
The histogram for n = 0 is centered on X = 1, since He0 (x) = 1. This confirms that E(He0 (X)) =
1 and V (He0 (X)) = 0. The histograms for n = 1, 3, 5 are symmetric, which is consistent with
the fact that E(Hen (X)) = 0, for n ≥ 1. The following Scilab session presents more detailed
numerical results. The first column prints n, the second column prints the empirical mean, the
third column prints the empirical variance and the last column prints n!, which is the exact value
of the variance.
n mean ( X ) variance ( X ) n!
0. 1. 0. 1.
1. 0.0041817 0.9978115 1.
2. - 0.0021810 2.0144023 2.
3. 0.0179225 6.1480027 6.
4. 0.0231110 24.483042 24.
5. - 0.0114358 112.13277 120.

Proposition 1.11. Let {Pn }n≥0 be orthogonal polynomials. Assume that x is a random variable
associated with the probability distribution function f , derived from the weight function w. For
two integers i, j ≥ 0, we have

E(Pi (X)Pj (X)) = 0 (39)

if i 6= j, and

E(Pi (X)2 ) = V (Pi (X)). (40)

Proof. We have
Z
E(Pi (X)Pj (X)) = Pi (x)Pj (x)f (x)dx (41)
I Z
1
= R Pi (x)Pj (x)w(x)dx (42)
I
w(x)dx I
hPi , Pj i
= R . (43)
I
w(x)dx

If i 6= j, the orthogonality of the polynomials implies 39. If, on the other hand, we have i = j,
then
hPi , Pi i
E(Pi (X)2 ) = R (44)
I
w(x)dx
kPi k2
= R . (45)
I
w(x)dx

We then use the equation 34, which leads to the equation 40.
In the next sections, we review a few orthogonal polynomials which are important in the
context of polynomial chaos. The table 5 summarizes the results.

10
 Distrib. Support Poly. w(x)  f (x)  2  kPn k2 V (Pn )
x2
√
N (0, 1) R Hermite exp − 2 √1 exp − x 2πn! n!
2π 2
1 2 1
U(−1, 1) [−1, 1] Legendre 1 2 2n+1 2n+1
E(1) R+ Laguerre exp(−x) exp(−x) 1 1

Figure 5: Some of properties of orthogonal polynomials

1.3 Hermite polynomials

In this section, we present Hermite polynomials and their properties.
Hermite polynomials are associated with the Gaussian weight:

w(x) = exp(−x2 /2), (46)

for x ∈ R. The integral of this weight is:

Z √
w(x)dx = 2π. (47)
R

The distribution function for Hermite polynomials is the standard Normal distribution:
 2
1 x
f (x) = √ exp − , (48)
2π 2
for x ∈ R.
Here, we consider the probabilist polynomials Hen , as opposed to the physicist polynomials
Hn [17].
The first Hermite polynomials are

He0 (x) = 1, (49)

He1 (x) = x (50)

The remaining Hermite polynomials satisfy the recurrence:

Hen+1 (x) = xHen (x) − nHen−1 (x), (51)

for n = 1, 2, ....

Example (First Hermite polynomials) We have

He2 (x) = xHe1 (x) − He0 (x) (52)

= x·x−1 (53)
= x2 − 1. (54)

Similarily,

He3 (x) = xHe2 (x) − 2He1 (x) (55)

= x(x2 − 1) − 2x (56)
= x3 − 3x. (57)

11
 He0 (x) = 1
He1 (x) = x
He2 (x) = x2 − 1
He3 (x) = x3 − 3x
He4 (x) = x4 − 6x2 + 3
He5 (x) = x5 − 10x3 + 15x

Figure 6: Hermite polynomials

Figure 7: The Hermite polynomials Hen , for n = 0, 1, 2, 3.

12
 n c0 c1 c2 c3 c4 c5 c6 c7 c8 c9
0 1
1 1
2 -1 1
3 -3 1
4 3 -6 1
5 15 -10 1
6 -15 45 -15 1
7 -105 105 -21 1
8 105 -420 210 -28 1
9 945 -1260 378 -36 1

Figure 8: Coefficients of Hermite polynomials Hen , for n = 0, 1, 2, ..., 9

The first Hermite polynomials are presented in the figure 6.

The figure 7 plots the Hermite polynomials Hen , for n = 0, 1, 2, 3.
The figure 8 presents the coefficients ck , for k ≥ 0 of the Hermite polynomials, so that
n
X
Hen (x) = ck x k . (58)
k=0

The Hermite polynomials are orthogonal with respect to the weight w(x). Moreover,
√
kHen k2 = 2πn!. (59)
Hence,
V (Hen (X)) = n!, (60)
for n ≥ 1, where X is a standard normal random variable.
The following HermitePoly Scilab function creates the Hermite polynomial of order n.
function y = HermitePoly ( n )
if ( n ==0) then
y = poly (1 , " x " ," coeff " )
elseif ( n ==1) then
y = poly ([0 1] , " x " ," coeff " )
else
polyx = poly ([0 1] , " x " ," coeff " )
// y (n -2)
yn2 = poly (1 , " x " ," coeff " )
// y (n -1)
yn1 = polyx
for k =2: n
y = polyx * yn1 -( k -1)* yn2
yn2 = yn1
yn1 = y
end
end
endfunction

13
 The script:
for n =0:10
y = HermitePoly (x , n );
disp ( y )
end
produces the following output:
1

2
- 1 + x

3
- 3x + x

2 4
3 - 6x + x

3 5
15 x - 10 x + x
7

1.4 Legendre polynomials

In this section, we present Legendre polynomials and their properties.
Legendre polynomials are associated with the unity weight
w(x) = 1, (61)
for x ∈ [−1, 1]. We have
Z 1
w(x)dx = 2. (62)
−1

The associated probability distribution function is:

1
f (x) = , (63)
2
for x ∈ [−1, 1]. This corresponds to the uniform distribution in probability theory.
The first Legendre polynomials are
P0 (x) = 1, (64)
P1 (x) = x (65)
The remaining Legendre polynomials satisfy the recurrence:
(n + 1)Pn+1 (x) = (2n + 1)xPn (x) − nPn−1 (x), (66)
for n = 1, 2, ....

14
 P0 (x) = 1
P1 (x) = x
P2 (x) = 12 (3x2 − 1)
P3 (x) = 12 (5x3 − 3x)
P4 (x) = 18 (35x4 − 30x2 + 3)
P5 (x) = 18 (63x5 − 70x3 + 15x)

Figure 9: Legendre polynomials

Example (First Legendre polynomials) We have

2P2 (x) = 3xP1 (x) − P0 (x) (67)

= 3x2 − 1, (68)

which implies
1
P2 (x) = (3x2 − 1). (69)
2
Similarily,

3P3 (x) = 5xP2 (x) − 2P1 (x) (70)

1
= 5x (3x2 − 1) − 2x (71)
2
15 3 5
= x − x − 2x (72)
2 2
15 3 9
= x − x (73)
2 2
1
= (15x3 − 9x). (74)
2
Finally, we divide both sides of the equation by 3 and get
1
P3 (x) = (5x3 − 3x). (75)
2
The first Legendre polynomials are presented in the figure 9.
The Legendre polynomials are orthogonal with respect to w(x) = 1. Moreover,
2
kPn k2 = (76)
2n + 1
Furthermore,
1
V (Pn (X)) = , (77)
2n + 1
for n ≥ 1, where X is a random variable uniform in the interval [−1, 1].
The figure 10 plots the Legendre polynomials Pn , for n = 0, 1, 2, 3.
The following LegendrePoly Scilab function creates the Legendre polynomial of order n.

15
 Figure 10: The Legendre polynomials Pn , for n = 0, 1, 2, 3.

function y = LegendrePoly ( n )
if ( n ==0) then
y = poly (1 , " x " ," coeff " )
elseif ( n ==1) then
y = poly ([0 1] , " x " ," coeff " )
else
polyx = poly ([0 1] , " x " ," coeff " )
// y (n -2)
yn2 = poly (1 , " x " ," coeff " )
// y (n -1)
yn1 = polyx
for k =2: n
y =((2* k -1)* polyx * yn1 -( k -1)* yn2 )/ k
yn2 = yn1
yn1 = y
end
end
endfunction

1.5 Laguerre polynomials

In this section, we present Laguerre polynomials and their properties.
Laguerre polynomials are associated with the weight

w(x) = exp(−x), (78)

16
for x ≥ 0.
The figure 11 plots the exponential weight function.

Figure 11: The Exponential weight.

We have:
Z +∞
w(x)dx = 1. (79)
0

Hence, the associated probability distribution function is:

f (x) = exp(−x), (80)

for x ≥ 0. This is the standard exponential distribution in probability theory.

The first Laguerre polynomials are

L0 (x) = 1, (81)
L1 (x) = −x + 1 (82)

The remaining Laguerre polynomials satisfy the recurrence:

(n + 1)Ln+1 (x) = (2n + 1 − x)Ln (x) − nLn−1 (x), (83)

for n = 1, 2, ....
The first Laguerre polynomials are presented in the figure 12.

Example (First Laguerre polynomials) We have

2L2 (x) = (3 − x)L1 (x) − L0 (x) (84)

= (3 − x)(1 − x) − 1 (85)
= 3 − x − 3x + x2 − 1 (86)
= x2 − 4x + 2. (87)

17
 L0 (x) = 1
L1 (x) = −x + 1
L2 (x) = 2!1 (x2 − 4x + 2)
L3 (x) = 3!1 (−x3 + 9x2 − 18x + 6)
L4 (x) = 4!1 (x4 − 16x3 + 72x2 − 96x + 24)
L5 (x) = 5!1 (−x5 + 25x4 − 200x3 + 600x2 − 600x + 120)

Figure 12: Laguerre polynomials

Hence,
1 2
L2 (x) = (x − 4x + 2). (88)
2
Similarily,

3L3 (x) = (5 − x)L2 (x) − 2L1 (x) (89)

1
= (5 − x) (x2 − 4x + 2) − 2(1 − x). (90)
2
Therefore,

6L3 (x) = (5 − x)(x2 − 4x + 2) − 4(1 − x) (91)

= 5x2 − 20x + 10 − x3 + 4x2 − 2x − 4 + 4x (92)
= −x3 + 9x2 − 18x + 6, (93)

which implies
1
L3 (x) = (−x3 + 9x2 − 18x + 6). (94)
6
The figure 13 plots the Laguerre polynomials Ln , for n = 0, 1, 2, 3.
The Laguerre polynomials are orthogonal with respect to w(x) = exp(−x). Moreover,

kLn k2 = 1. (95)

Furthermore,

V (Ln (X)) = 1, (96)

for n ≥ 1, where X is a standard exponential random variable.

The following LaguerrePoly Scilab function creates the Laguerre polynomial of order n.
function y = LaguerrePoly ( n )
if ( n ==0) then
y = poly (1 , " x " ," coeff " )
elseif ( n ==1) then
y = poly ([1 -1] , " x " ," coeff " )
else
polyx = poly ([0 1] , " x " ," coeff " )

18
 Figure 13: The Laguerre polynomials Ln , for n = 0, 1, 2, 3.

// y (n -2)
yn2 = poly (1 , " x " ," coeff " )
// y (n -1)
yn1 =1 - polyx
for k =2: n
y =((2* k -1 - polyx )* yn1 -( k -1)* yn2 )/ k
yn2 = yn1
yn1 = y
end
end
endfunction

1.6 Chebyshev polynomials of the first kind

In this section, we present Chebyshev polynomials of the first kind and their properties.
It is important to mention that Chebyshev polynomials of the second kind are not presented
in this document.
Chebyshev polynomials are associated with the weight
1
w(x) = √ , (97)
1 − x2
for x ∈ [−1, 1]. This weight does not correspond to a particular distribution in probability theory.
The first Chebyshev polynomials are
T0 (x) = 1, (98)
T1 (x) = x (99)

19
 T0 (x) = 1
T1 (x) = x
T2 (x) = 2x2 − 1
T3 (x) = 4x3 − 3x
T4 (x) = 8x4 − 8x2 + 1
T5 (x) = 16x5 − 20x3 + 5x

Figure 14: Chebyshev polynomials of the first kind

The remaining Chebyshev polynomials satisfy the recurrence:

Tn+1 (x) = 2xTn (x) − Tn−1 (x), (100)
for n = 1, 2, ....
The first Chebyshev polynomials are presented in the figure 14.
Example (First Chebyshev polynomials) We have
T2 (x) = 2xT1 (x) − T0 (x) (101)
= 2x2 − 1. (102)
Similarily,
T3 (x) = 2xT2 (x) − T1 (x) (103)
= 2x(2x2 − 1) − x (104)
= 4x3 − 2x − x (105)
= 4x3 − 3x. (106)
The figure 15 plots the Chebyshev polynomials Tn , for n = 0, 1, 2, 3.
The following ChebyshevPoly Scilab function creates the Chebyshev polynomial of order n.
function y = ChebyshevPoly ( n )
if ( n ==0) then
y = poly (1 , " x " ," coeff " )
elseif ( n ==1) then
y = poly ([0 1] , " x " ," coeff " )
else
polyx = poly ([0 1] , " x " ," coeff " )
// y (n -2)
yn2 = poly (1 , " x " ," coeff " )
// y (n -1)
yn1 = polyx
for k =2: n
y =2* polyx * yn1 - yn2
yn2 = yn1
yn1 = y
end
end
endfunction

20
 Figure 15: The Chebyshev polynomials Tn , for n = 0, 1, 2, 3.

1.7 Accuracy of evaluation

In this section, we present the accuracy issues which appear when we evaluate the orthogonal
polynomials with large values of the polynomial degree n. More precisely, we present a funda-
mental limitation in the accuracy that can be achieved by the polynomial representation of the
orthogonal polynomials, while a sufficient accuracy can be achieved with a different evaluation
algorithm.
The following LaguerreEval function evaluates the Laguerre polynomial of degree n at point
x. One one hand, the LaguerrePoly function creates a polynomial which can be evaluated with
the horner function, while, on the other hand, the function LaguerreEval directly evaluates the
polynomial. As we are going to see soon, although the recurrence is the same, the numerical
performances of these two functions is rather different.
function y = LaguerreEval (x , n )
if ( n ==0) then
nrows = size (x , " r " )
y = ones ( nrows ,1)
elseif ( n ==1) then
y =1 - x
else
// y (n -2)
yn2 =1
// y (n -1)
yn1 =1 - x
for k =2: n
y =((2* k -1 - x ).* yn1 -( k -1)* yn2 )/ k

21
 yn2 = yn1
yn1 = y
end
end
endfunction
Consider the value of the polynomial L100 at the point x = 10. From Wolfram Alpha, which
uses an arbitrary precision computer system, the exact value of L100 (10) is

13.277662844303454137789892644070318985922997400185621...

In the following session, we compare the accuracies of LaguerrePoly and LaguerreEval.

-->n =100;
-->x =10;
-->L = LaguerrePoly ( n );
--> y1 = horner (L , x )
y1 =
3.695 D +08
--> y2 = LaguerreEval (x , n )
y2 =
13.277663
We see that the polynomial created by LaguerrePoly does not produce a sufficient accuracy,
while LaguerreEval seems to be accurate. In fact, the value produced by LaguerreEval has
more than 15 correct decimal digits.
In order to see the trend when x and n increases, we compute the number of common digits,
in base 2, in the evaluation of Ln (x), computed from LaguerrePoly and LaguerreEval. Since
Scilab uses double precision floating point numbers with 64 bits and 53 bits of precision, this
number varies from 53 (when the two functions return exactly the same value) to 0 (when the
two functions return two value which have no digit in common). The figure 16 plots the number
of common digits, when n ranges from 1 to 100 and x ranges from 1 to 21.
We see that the number of correct digits produced by LaguerrePoly decreases when x and n
increase.
The explanation for this fact is that the polynomial representation of the Laguerre polynomial
forces to use n = 100 additions of positive and negative powers of x. This computation is ill-
conditionned, because it involves large numbers with different signs.
For example, the following session presents some of the coefficients of L100 .
-->L
L =
2 3 4 5 6
1 - 100 x + 2475 x - 26950 x + 163384.37 x - 627396 x + 1655628.3 x
7 8 9 10
- 3176103.3 x + 4615275.2 x - 5242040.9 x + 4770257.2 x
[...]
95 96 97 98
- 7.28 D -141 x + 3.95 D -144 x - 1.68 D -147 x + 5.25 D -151 x
99 100
- 1.07 D -154 x + 1.07 D -158 x

22
Figure 16: The number of common digits in the evaluation of Laguerre polynomials, for increasing
n and x.

We see that the coefficients signs alternate.

Consider n real numbers yi , for i = 1, 2, ..., n, representing the monomial values which appear
in the polynomial representation. We are interested in the sum

S = y1 + y2 + ... + yn .

We know from numerical analysis [8] that the condition number of a sum is

|y1 | + |y2 | + ...|yn |

.
|y1 + y2 + ...yn |
Hence, when the sum S is small in magnitude, but involves large terms yi , the condition number
is large. In other words, small relative errors in yi are converted into large relative errors in S,
and this is what happens for polynomials.
Hence, the polynomial representation should be avoided in practice to evaluate the orthogonal
polynomial, although they are a convenient way of using them in Scilab. On the other hand, a
straightforward evaluation of the polynomial value based on the recurrence formula can produce
a good accuracy.

1.8 Notes and references

The classic [1] gives a brief overview of the main results on orthogonal polynomials in the chap-
ter 22 ”Orthogonal polynomials”. More details on orthogonal polynomials are presented in [3],
especially the chapter 12 ”Legendre functions” and chapter 13 ”More special functions”.
The book [2] has three chapters on orthogonal polynomials, including quadrature, and other
advanced topics.
One of the references on the topic is [15], which covers the topic in depth.

23
 Cell 1 Cell 2 Cell 3
1 | *** | - | - |
2 | - | *** | - |
3 | - | - | *** |
4 | ** | * | - |
5 | ** | - | * |
6 | * | ** | - |
7 | - | ** | * |
8 | * | - | ** |
9 | - | * | ** |
10 | * | * | * |

Figure 17: Placing d = 3 balls into p = 3 cells. The balls are represented by stars ∗.

In the appendices of [11], we can find a presentation of the most common orthogonal polyno-
mials used in uncertainty analysis.

2 Multivariate polynomials
In this section, we define the multivariate polynomials which are used in the context of polynomial
chaos decomposition. We consider the regular polynomials as well as multivariate orthogonal
polynomials.

2.1 Occupancy problems

In the next section, we will count the number of monomials of degree d with p variables. This
problems is the same as placing d balls into p cells, which is the topic of the current section.
Consider the problem of placing d = 3 balls into p = 3 cells. The figure 17 presents the 10
ways to place the balls.
Let us denote by αi ≥ 0 the number of balls in the i-th cell, for i = 1, 2, ..., p. Let α =
(α1 , α2 , ..., αp ) be the associated vector. Each particular configuration is so that the total number
of balls is d, i.e. we have

α1 + α2 + ... + αp = d. (107)

Example (Case where d = 8 and p = 6) Consider the case where there are d = 8 balls to
distribute into p = 6 cells. We represent the balls with stars ”*” and the cells are represented by
the spaces between the p + 1 bars ”|”. For example, the string ”|***|*| | | |****|” represents the
configuration where α = (3, 1, 0, 0, 0, 4).

Proposition 2.1. ( Occupancy problem) The number of ways to distribute d balls into p cells is:
   
p p+d−1
= . (108)
d d

The left hand side of the previous equation is the multiset coefficient.

24
Proof. Within the string, there is a total of p+1 bars : the first and last bars, and the intermediate
p − 1 bars. We are free to move only the p − 1 bars and the d stars, for a total of p + d − 1 symbols.
Therefore, the problem reduces to finding the place of the d stars in a string of p + d − 1 symbols.
From combinatorics, we know that the number of ways to choose k items in a set of n items is
 
n n!
= , (109)
k k!(n − k)!
for any nonnegative integers n and k, and k ≤ n. Hence, the number of ways to choose the place
of the d stars in a string of p + d − 1 symbols is defined in the equation 108.

2.2 Multivariate monomials

The four one-dimensional monomials 1, x, x2 can be used to create a multivariate monomial, just
by multiplying the appropriate components. Consider, for example, p = 3 and let x ∈ R3 be a
point in the three-dimensional space. The function
x2 x23 (110)
is the monomial associated to the vector α = (0, 1, 2), which entries are the exponents of x =
(x1 , x2 , x3 ).
Definition 2.2. ( Multivariate monomial) Let x ∈ Rp . The function Mα (x) is a multivariate
monomial if there is a vector α = (α1 , α2 , ..., αp ) of integers exponents, where αi ∈ {0, 1, ..., p},
for i = 1, 2, ..., p such that
Mα (x) = xα1 1 xα2 2 ...xαp p , (111)
for any x ∈ Rp .
The vector α is called a multi-index.
The previous equation can be rewritten in a more general way:
p
Y
Mα (x) = xαi i , (112)
i=1

for any x ∈ Rp .
Definition 2.3. ( Degree of a multivariate monomial) Let M (x) be a monomial associated with
the exponents α, for any x ∈ Rp . The degree of M is
d = α1 + α2 + ... + αp . (113)
The degree d of a multivariate monomial is also denoted by |α|.
The figure 18 presents the 10 monomials with degree d = 2 and p = 4 variables.
Proposition 2.4. ( Number of multivariate monomials) The number of degree d multivariate
monomials of p variables is:
   
p p+d−1
= . (114)
d d
Proof. The problem can be reduced to distributing d balls into p cells: the i-th cell represents
the variable xi where we have to put αi balls. It is then straightforward to use the proposition
2.1.
The figure 19 presents the number of monomials for various d and p.

25
 α1 α2 α3 α4
2 0 0 0
1 1 0 0
1 0 1 0
1 0 0 1
0 2 0 0
0 1 1 0
0 1 0 1
0 0 2 0
0 0 1 1
0 0 0 2

Figure 18: The monomials with degree d = 2 and p = 4 variables.

p\d 0 1 2 3 4 5
1 1 1 1 1 1 1
2 1 2 3 4 5 6
3 1 3 6 10 15 21
4 1 4 10 20 35 56
5 1 5 15 35 70 126

Figure 19: Number of degree d monomials with p variables.

2.3 Multivariate polynomials

Definition 2.5. ( Multivariate polynomial) Let x ∈ Rp . The function P (x) is a degree d multi-
variate polynomial if there is a set of multivariate monomials Mα (x) and a set of real numbers
βα such that
X
P (x) = βα Mα (x), (115)
|α|≤d

for any x ∈ Rp .
We can plug the equation 112 into 115 and get
p
X Y
P (x) = βα xαi i , (116)
|α|≤d i=1

for any x ∈ Rp .
Example (Case where d = 2 and p = 3) Consider the degree d = 2 multivariate polynomial with
p = 3 variables:
P (x1 , x2 , x3 ) = 4 + 2x3 + 5x21 − 3x1 x2 x3 (117)
for any x ∈ R3 . The relevant multivariate monomials exponents α are:
(0, 0, 0), (0, 0, 1), (118)
(1, 0, 0), (1, 1, 1). (119)

26
The only nonzero coefficients β are:

β(0,0,0) = 4, β(0,0,1) = 2, (120)

β(1,0,0) = 5, β(1,1,1) = −3. (121)

Proposition 2.6. ( Number of multivariate polynomials) The number of degree d multivariate

polynomials of p variables is:
 
p p+d
Pd = . (122)
d

Proof. The proof is by recurrence on d. The only polynomial of degree d = 0 is:

P0 (x) = 1. (123)

Hence, there are:

 
p
1= (124)
0

multivariate polynomials of degree d = 0. This proves that the equation 122 is true for d = 0.
The polynomials of degree d = 1 are:

Pi (x) = xi , (125)

for i = 1, 2, ..., p. Hence, there is a total of 1 + p polynomials of degree d = 1, which writes

 
p+1
1+p = . (126)
1

This proves that the equation 122 is true for d = 1.

Now, assume that the equation 122 is true for d, and let us prove that it is true for d + 1. For
a multivariate polynomial P (x) of degree d + 1, the equation 115 implies
X
P (x) = βα Mα (x), (127)
|α|≤d+1
X X
= βα Mα (x) + βα Mα (x), (128)
|α|≤d |α|=d+1

for any x ∈ Rp . Hence, the number of multivariate polynomials of degree d + 1 is the sum of
the number of multivariate polynomials of degree d and the number of multivariate monomials
of degree d + 1. From the equation 114, the number of such monomials is
   
p p+d
= . (129)
d+1 d+1

Hence, the number of multivative polynomials of degree d + 1 is:

   
p+d p+d
+ . (130)
d+1 d

27
 p\d 0 1 2 3 4 5 6
1 1 2 3 4 5 6 7
2 1 3 6 10 15 21 28
3 1 4 10 20 35 56 84
4 1 5 15 35 70 126 210
5 1 6 21 56 126 252 462
6 1 7 28 84 210 462 924

Figure 20: Number of degree d polynomials with p variables.

However, we know from combinatorics that

     
n n n+1
+ = , (131)
k k−1 k

for any nonnegative integers n and k. We apply the previous equality with n = p+d and k = d+1,
and get
     
p+d p+d p+d+1
+ = , (132)
d+1 d d+1

which proves that the equation 122 is also true for d + 1, and concludes the proof.
The figures 20 and 21 present the number of polynomials for various d and p.
One possible issue with the equation 115 is that it does not specify a way of ordering the Pdp
multivariate polynomials of degree d. However, we will present in the next section a constructive
way of ordering the multi-indices is such a way that there is a one-to-one mapping from the single
index k, in the range from 1 to Pdp , to the corresponding multi-index α(k) . With this ordering,
the equation 115 becomes
p
Pd
X
P (x) = βk Mα(k) (x), (133)
k=1

for any x ∈ Rp .

Example (Case where d = 2 and p = 2) Consider the degree d = 2 multivariate polynomial with
p = 2 variables:

P (x1 , x2 ) = β(0,0) + β(1,0) x1 + β(0,1) x2 + β(2,0) x21 + β(1,1) x1 x2 + β(0,2) x22 (134)

for any x1 , x2 ∈ R2 . This corresponds to the equation

P (x1 , x2 ) = β1 + β2 x1 + β3 x2 + β4 x21 + β5 x1 x2 + β6 x22 (135)

for any x1 , x2 ∈ R2 .

28
 Figure 21: Number of degree d polynomials with p variables.

2.4 Generating multi-indices

In this section, we present an algorithm to generate all the multi-indices α associated with a
degree d multivariate polynomial of p variables.
The following polymultiindex function implements the algorithm suggested by Lemaı̂tre and
Knio in the appendix of [11]. The function returns a matrix a with p columns and Pdp rows. Each
row a(i,:) represents the i-th multivariate polynomial, for i = 1, 2, ..., Pdp . For each row i, the
entry a(i,j) is the exponent of the j-th variable xj , for i = 1, 2, ..., Pdp and j = 1, 2, ..., p. The
sum of exponents for each row is lower or equal to d. The first row is the zero-degree polynomial,
so that all entries in a(1,1:p) are zero. The rows 2 to p+1 are the first degree polynomials: all
columns are zero, except one entry which is equal to 1.
function a = polymultiindex (p , d )
// Zero - th order polynomial
a (1 ,1: p )= zeros (1 , p )
// First order polynomials
a (2: p +1 ,1: p )= eye (p , p )
P = p +1
pmat =[]
pmat (1: p ,1)=1
for k =2: d
L=P
for i =1: p
pmat (i , k )= sum ( pmat ( i :p ,k -1))

29
 end
for j =1: p
for m =L - pmat (j , k )+1: L
P = P +1
a (P ,1: p )= a (m ,1: p )
a (P , j )= a (P , j )+1
end
end
end
endfunction
For example, in the following session, we compute the list of multi-indices corresponding to
multivariate polynomials of degree d = 3 with p = 3 variables.
-->p =3;
-->d =3;
-->a = polymultiindex (p , d )
a =
0. 0. 0.
1. 0. 0.
0. 1. 0.
0. 0. 1.
2. 0. 0.
1. 1. 0.
1. 0. 1.
0. 2. 0.
0. 1. 1.
0. 0. 2.
3. 0. 0.
2. 1. 0.
2. 0. 1.
1. 2. 0.
1. 1. 1.
1. 0. 2.
0. 3. 0.
0. 2. 1.
0. 1. 2.
0. 0. 3.
There are 20 rows in the previous matrix a, which is consistent with the table 20. In the
previous session, the first row correspond to the zero-th degree polynomial, the rows 2 to 4
corresponds to the degree one monomials, the rows from 5 to 10 corresponds to the degree 2
monomials, whereas the rows from 11 to 20 corresponds to the degree 3 monomials.

2.5 Multivariate orthogonal functions

In this section, we introduce the weighted space of square integrable functions in Rp , so that we
can define the orthogonality of multivariate functions.
Definition 2.7. ( Interval of Rp ) An interval I of Rp is a subspace of Rp , such that x ∈ I if
x1 ∈ I1 , x2 ∈ I2 , ..., xp ∈ Ip , (136)

30
where I1 , I2 ,..., Ip are intervals of R. We denote this tensor product by:

I = I1 ⊗ I2 ⊗ ... ⊗ Ip . (137)

Now that we have multivariate intervals, we can consider a multivariate weight function on
such an interval.

Definition 2.8. ( Multivariate weight function of Rp ) Let I be an interval in Rp . A weight

function w on I is a nonnegative integrable function of x ∈ I.

Such a weight can be created by making the tensor product of univariate weights.

Proposition 2.9. ( Tensor product of univariate weights) Let I1 , I2 , ..., Ip be intervals of R. As-
sume that w1 , w2 , ..., wp are univariate weights on I1 , I2 , ..., Ip . Therefore, the tensor product func-
tion

w(x) = w1 (x1 )w2 (x2 )...wp (xp ), (138)

for any x ∈ R, is a weight function of x ∈ I.

Proof. Indeed, its integral is

Z Z
w(x)dx = w1 (x1 )w2 (x2 )...wp (xp )dx1 dx2 ...dxp (139)
I I
Z  Z  Z !
= w1 (x1 )dx1 w2 (x2 )dx2 ... wp (xp )dxp . (140)
I1 I2 Ip

However, all the individual integrals are finite, since each function wi is, by hypothesis, a weight
on Ii , for i = 1, 2, ..., p. Hence, the product is finite, so that the function w is, indeed, a weight
on I.

Example (Multivariate Gaussian weight) In the particular case where all weights are the same,
the expression simplifies further. Consider, for example, the Gaussian weight

exp(−x2 /2), (141)

for x ∈ R. We can then consider the tensor product:

w(x) = w(x1 )w(x2 )...w(xp ), (142)

for x1 , x2 , ..., xp ∈ R.

Example (Multivariate Hermite-Legendre weight) Consider the weight for Hermite polynomials
w1 (x1 ) = exp(−x21 /2) for x1 ∈ R, and consider the weight for Legendre polynomials w2 (x2 ) = 1
for x2 ∈ [−1, 1]. Therefore,

w(x) = w1 (x1 )w2 (x2 ) (143)

for x ∈ I is a weight on I = R ⊗ [−1, 1].

31
 In the remaining of this document, we will not make a notation difference between the uni-
variate weight function w(x) and the multivariate weight w(x), assuming that it is based on a
tensor product if needed.
Definition 2.10. ( Multivariate weighted L2 space in Rp ) Let I be an interval in Rp . Let w be a
multivariate weight function on I. Let L2w (I) be the set of functions g which are square integrable
with respect to the weight function w, i.e. such that the integral
Z
kgk = g(x)2 w(x)dx
2
(144)
I

is finite. In this case, the norm of g is kgk.

Definition 2.11. ( Multivariate inner product in L2w (I) space) Let I be an interval of Rp . Let w
be a multivariate weight function on I. For any g, h ∈ L2w (I), the inner product of g and h is
Z
hg, hi = g(x)h(x)w(x)dx. (145)
I

Let I be an interval of Rp and assume that g ∈ L2w (I). We can combine the equations 144
and 145, which implies that the L2w (I) norm of g can be expressed as an inner product:
kgk2 = hg, gi . (146)
Proposition 2.12. ( Multivariate probability distribution function) Let I be an interval of Rp .
Let w be a multivariate weight function on I. Assume {Xi }i=1,2,...,p are independent random
variables associated with the probability distribution functions fi , derived from the weight functions
wi , for i = 1, 2, ..., p. Therefore, the function
f (x) = f1 (x1 )f2 (x2 )...fp (xp ), (147)
for x ∈ I is a probability distribution function.
Proof. We must prove that the integral of f is equal to one. Indeed,
Z Z
f (x)dx = f1 (x1 )f2 (x2 )...fp (xp )dx1 dx2 ...dxp (148)
I I
Z  Z  Z !
= f1 (x1 )dx1 f2 (x2 )dx2 ... fp (xp )dxp (149)
I1 I2 Ip

= 1, (150)
where each integral is equal to one since, by hypothesis, fi is a probability distribution function
for i = 1, 2, ..., p.

2.6 Tensor product of orthogonal polynomials

In this section, we present a method to create multivariate orthogonal polynomials.
Consider the cas where p = 2 and let’s try to create a bivariate Hermite orthogonal polyno-
mials. Therefore, we consider the inner product:
Z
hg, hi = g(x1 , x2 )h(x1 , x2 )w(x1 )w(x2 )dx1 dx2 , (151)
R2

32
associated with the multivariate Gaussian weight w(x).
We can make the tensor product of some Hermite polynomials, which leads, for example, to:

Ψ1 (x1 , x2 ) = He1 (x1 ) (152)

Ψ2 (x1 , x2 ) = He1 (x1 )He2 (x2 ), (153)

associated with the multivariate Gaussian weight w(x). These polynomials are of degree 2. We
may wonder if these two polynomials are orthogonal. We have

hΨ1 , Ψ2 i (154)
Z
= Ψ1 (x1 , x2 )Ψ2 (x1 , x2 )w(x1 )w(x2 )dx1 dx2 (155)
R 2
Z
= He1 (x1 )2 He2 (x2 )w(x1 )w(x2 )dx1 dx2 (156)
2
RZ  Z 
2
= He1 (x1 ) w(x1 )dx1 He2 (x2 )w(x2 )dx2 . (157)
R R

By the orthogonality of Hermite polynomials, we have

Z Z
He2 (x2 )w(x2 )dx2 = He0 (x2 )He2 (x2 )w(x2 )dx2 (158)
R R
= 0. (159)

Hence,

hΨ1 , Ψ2 i = 0, (160)

which implies that Ψ1 and Ψ2 are orthogonal. However,

Z
2
kΨ1 k = He1 (x1 )2 w(x1 )w(x2 )dx1 dx2 (161)
2
RZ  Z 
2
= He1 (x1 ) w(x1 )dx1 w(x2 )dx2 (162)
R R
√
= 2πkHe1 k2 . (163)

Definition 2.13. ( Tensor product of orthogonal polynomials) Let I be a tensor product interval
of Rp and let w the associated multivariate tensor product weight function. Let φα(k) (xi ) be a family
i
of univariate orthogonal polynomials. The associated multivariate tensor product polynomials are
p
Y
Ψk (x) = φα(k) (xi ) (164)
i
i=1

for k = 1, 2, ..., Pdp , where the degree of Ψk is

p
X (k)
d= αi . (165)
i=1

33
 d Multi-Index Polynomial
0 α(1) = [0, 0] Ψ1 (x) = He0 (x1 )He0 (x2 ) = 1
1 α(2) = [1, 0] Ψ2 (x) = He1 (x1 )He0 (x2 ) = x1
1 α(3) = [0, 1] Ψ3 (x) = He0 (x1 )He1 (x2 ) = x2
2 α(4) = [2, 0] Ψ4 (x) = He2 (x1 )He0 (x2 ) = x21 − 1
2 α(5) = [1, 1] Ψ5 (x) = He1 (x1 )He1 (x2 ) = x1 x2
2 α(6) = [0, 2] Ψ6 (x) = He0 (x1 )He2 (x2 ) = x22 − 1
3 α(7) = [3, 0] Ψ7 (x) = He3 (x1 )He0 (x2 ) = x31 − 3x1
3 α(8) = [2, 1] Ψ8 (x) = He2 (x1 )He1 (x2 ) = (x21 − 1)x2
3 α(9) = [1, 2] Ψ9 (x) = He1 (x1 )He2 (x2 ) = x1 (x22 − 1)
3 α(10) = [0, 3] Ψ10 (x) = He0 (x1 )He3 (x2 ) = x32 − 3x2

Figure 22: Multivariate Hermite polynomials of p = 2 variables and degree d = 3.

Figure 23: The multivariate Hermite polynomials, with degree d = 2 and p = 2 variables.

34
Example (Multivariate Hermite polynomials) The figure 22 presents the multivariate Hermite
polynomials of p = 2 variables and degree d = 3. The figure 23 presents the multivariate Hermite
polynomials, with degree d = 2 and p = 2 variables.

Proposition 2.14. ( Multivariate orthogonal polynomials) The multivariate tensor product poly-
nomials defined in 2.13 are orthogonal.
Proof. We must prove that, for two different integers k and `, we have
hΨk , Ψ` i = 0. (166)
By the definition of the inner product on L2w (I), we have
Z
hΨk , Ψ` i = Ψk (x)Ψ` (x)w(x)dx (167)
I

By assumption, the multivariate weight function w on I is the tensor product of univariate weight
functions wi on Ii . Hence,
Z Yp p p
Y Y
hΨk , Ψ` i = φα(k) (xi ) φα(`) (xi ) wi (xi )dx (168)
i i
I i=1 i=1 i=1
p Z
Y
= φα(k) (xi )φα(`) (xi )wi (xi )dxi . (169)
i i
i=1 Ii

In other words,
p D E
Y
hΨk , Ψ` i = φα(k) , φα(`) . (170)
i i
i=1

However, the multi-indice ordering implies that, if k 6= `, then there exists an integer i ∈
{1, 2, ..., p}, such that
(k) (`)
αi 6= αi . (171)
By assumption, the φα(k) (xi ) polynomials are orthogonal. This implies
i
D E
φα(k) , φα(`) = 0, (172)
i i

which concludes the proof.

We can use the equation 170 to compute the L2w (I) norm of Ψk . Indeed,
kΨk k2 = hΨk , Ψk i (173)
p D E
Y
= φα(k) , φα(k) (174)
i i
i=1
p
Y
= kφα(k) k2 . (175)
i
i=1

This proves the following proposition.

35
Proposition 2.15. ( Norm of multivariate orthogonal polynomials) The L2w (I) norm of the mul-
tivariate orthogonal polynomials defined in 2.13 is:
p
Y
2
kΨk k = kφα(k) k2 . (176)
i
i=1

Example (Norm of the multivariate Hermite polynomials) Consider the multivariate Hermite
polynomials in the case where p = 2. The figure 22 indicates that

Ψ8 (x) = He2 (x1 )He1 (x2 ), (177)

for x1 , x2 ∈ R. Hence, the equation 176 implies:

kΨ8 k2 = kHe2 k2 kHe1 k2 . (178)

The norm of the univariate Hermite polynomial is given by the equation 59. Hence,
√ √
kΨ8 k2 = 2π · 2! · 2π · 1! (179)
= 4π. (180)

For any p ≥ 1, we have

p
2
Y √ (k)
kΨk k = 2παi ! (181)
i=1
p
Y (k)
2/p
= (2π) αi !. (182)
i=1

2.7 Multivariate orthogonal polynomials and probabilities

In this section, we present the properties of Ψk (X), where X is a multivariate random variable
associated with multivariate orthogonal polynomials {Ψk }k≥1 .

Proposition 2.16. ( Expectation of multivariate orthogonal polynomials) Let I be a tensor prod-

uct interval of Rp and let w the associated multivariate tensor product weight function. Let Ψk
be the family of tensor product orthogonal multivariate polynomials, associated with the weight w.
Let X be a multivariate random variable associated with the f multivariate probability distribution
function, where {Xi }i=1,2,...,p are independent random variables. Therefore,

E(Ψ1 (X)) = 1 (183)

and

E(Ψk (X)) = 0, (184)

for k > 1.

36
Proof. Indeed,
Z
E(Ψ1 (X)) = Ψ1 (x)f (x)dx (185)
ZI
= f (x)dx (186)
I
= 1, (187)

since Ψ1 (x) = 1, for any x ∈ I. Moreover, for any k > 1, we have

Z
E(Ψk (X)) = Ψk (x)f (x)dx (188)
I
p
Z Y p
Y
= φα(k) (xi ) fi (xi )dx1 dx2 ...dxp (189)
i
I i=1 i=1
p Z
Y
= φα(k) (xi )fi (xi )dxi (190)
i
i=1 Ii
p  
Y
= E φα(k) (xi ) . (191)
i
i=1

However, the proposition 1.9 implies:

 
E φα(k) (X) = 1 (192)
i

(k)
if αi = 0 and
 
E φα(k) (X) = 0 (193)
i

(k)
if αi ≥ 1.
(k)
Since k > 1, there is at least one integer i ∈ {1, 2, ..., p} such that αi ≥ 1. If this was not
(k)
true, this would imply that αi = 0, for i = 1, 2, ..., p. This implies that k = 1, which contradicts
the hypothesis.
Therefore, there is at least one integer i such that the expectation is zero, so that the product
in the equation 191 is also zero.

Proposition 2.17. ( Variance of multivariate orthogonal polynomials) Under the same hypotheses
as in proposition 2.16, we have

V (Ψ1 (X)) = 0 (194)

and
p  
Y
V (Ψk (X)) = E φα(k) (Xi )2 , (195)
i
i=1

for k > 1.

37
Proof. Obviously, the random variable Ψ1 (X) = 1 has a zero variance. Furthermore, for k > 1,
we have
V (Ψk (X)) = E (Ψk (X) − E(Ψk (X)))2


(196)
2


= E Ψk (X) , (197)
since E(Ψk (X)) = 0. We then use the tensor product definition of both Ψk and f , and finally get
to the equation 195.
For k > 1, we have
     2
E φα(k) (Xi )2 = V φα(k) (Xi ) + E φα(k) (Xi ) . (198)
i i i

Therefore, based on the propositions 1.9 and 1.10, we can use the equation 195 to evaluate the
variance of Ψk .
Proposition 2.18. ( Covariance of multivariate orthogonal polynomials) Under the same hy-
potheses as in proposition 2.16, for any two integers k and `, we have
Cov(Ψk (X), Ψ` (X)) = 0, (199)
if k 6= `.
Proof. Assume that k > 1. Therefore,
Cov(Ψ1 (X), Ψk (X)) = E ((Ψ1 (X) − µ1 ) (Ψk (X) − µk )) , (200)
where µ1 = E(Ψ1 (X)) and µk = E(Ψk (X)). However, we have Ψ1 (X) = E(Ψ1 (X)) = 1. Hence,
Cov(Ψ1 (X), Ψk (X)) = 0. (201)
The same is true if we consider
Cov(Ψk (X), Ψ1 (X)) = Cov(Ψ1 (X), Ψk (X)) (202)
= 0, (203)
by the symmetry property of the covariance.
Assume now that k and ` are two integers such that k, ` > 1. We have,
Cov(Ψk (X), Ψ` (X)) = E ((Ψk (X) − µk ) (Ψ` (X) − µ` )) (204)
= E (Ψk (X)Ψ` (X)) , (205)
since, by the proposition 2.16, we have µk = µ` = 0. Hence,
Z
Cov(Ψk (X), Ψ` (X)) = Ψk (x)Ψ` (x)f (x)dx (206)
I Z
1
= R Ψk (x)Ψ` (x)w(x)dx (207)
I
w(x)dx I
1
= R hΨk , Ψ` i (208)
I
w(x)dx
by the definition of the weight function w. By the proposition 2.14, the polynomials Ψk and Ψ`
are orthogonal, which concludes the proof.

38
2.8 Notes and references
The stars and bars proof used in section 2.1 is presented in Feller’s book [7], in the section
”Application to occupancy problems” of the chapter 2 ”Elements of combinatorial analysis”. The
book [11] and the thesis [10] present spectral methods, including the multivariate orthogonal
polynomials involved in polynomial chaos. The figures 22 and 23 are presented in several papers
and slides related to multivariate orthogonal polynomials, including [12], for example.

3 Polynomial chaos
3.1 Introduction
The polynomial chaos, introduced by Wiener [16], uses Hermite polynomials as the basis and
involves independent Gaussian random variables.
Denote [4, 13] the set of multi-indices with finite number of nonzero components as

J = {α = (αi )i≥1 , αi ∈ {0, 1, 2, ...}, |α| < ∞} , (209)

where
∞
X
|α| = αi . (210)
i=1

If α ∈ J , then there is only a finite number of nonzero components (otherwise |α| would be
infinite).
Let {Xi }i≥0 be an infinite set of independent standard normal random variables and let f be
the associated multivariate tensor product normal distribution function. Assume that g(X) is a
random variable with finite variance. We are interested in the decomposition of g(X) onto Heαi ,
the Hermite polynomial of degree αi ≥ 0.
The expectation of g(X) is
Z
E(g(X)) = g(x)f (x)dx, (211)

and its variance is:

V (g(X)) = E (g(X) − µ)2 ,

(212)

where µ = E(g(X)).
For any α ∈ J , the Wick polynomial is
∞
Y
Ψα (X) = Heαi (Xi ). (213)
i=1

The degree of the polynomial Tα is |α|. Notice that, since there is only a finite number of nonzero
components in α, the right hand side of the expression 213 has a finite number of factors.
We consider the inner product
Z
hg, Ψα i = g(x)Ψα (x)f (x)dx. (214)

39
We use the norm
Z
2
kΨα k = Ψα (x)2 f (x)dx. (215)

The following theorem is due to Cameron and Martin [5].

Theorem 3.1. ( Cameron-Martin) Let {xi }∞i=1 be an infinite set of independent standard nor-
mal random variables. Let g(X) be a random variable with finite variance. Then g(X) has the
decomposition
X
g(X) = aα Ψα (X) (216)
α∈J

where X = (ξ1 , ξ2 , ...) and

hg, Ψα i
aα = . (217)
kΨα k2
Moreover,

E(g(X)) = aα0 , (218)

X
V (g(X)) = a2α kΨα k2 , (219)
α∈J

where α0 = (0, 0, ...).

The previous theorem is due to [5] and will not be proved here.
Moreover, the decomposition 216 converges in the L2w sense.

3.2 Truncated decomposition

In the decomposition 216, there is an infinite number of random variables xi and an unrestricted
polynomial degree |α|. In order to use such an expansion in practice, we have to perform a double
truncation. This is why we keep only p independent normal random variables and only the order
d multivariate polynomials. Define the truncated index set [9]

Jp,d = {α = (α1 , α2 , ..., αp ), |α| ≤ d} . (220)

The finite decomposition is

X
g(X) ≈ aα Ψα (X), (221)
α∈Jp,d

where X = (X1 , X2 , ..., Xp ) are independent standard normal random variables. There is a one-
to-one mapping from the multi-indices of α of the set Jp,d and the indices k = 1, 2, ..., Pdp defined
in the section 2.4. Therefore, the decomposition 216 can be written:
p
Pd
X
g(X) ≈ ak Ψk (X), (222)
k=1

where Ψk is defined by the equation 164.

40
Proposition 3.2. ( Expectation and variance of the truncated PC expansion) The truncated
expansion 222 is such that

E(g(X)) = a1 (223)
Pdp
X
V (g(X)) = a2k V (Ψk (X)). (224)
k=2

Since the expansion 222 involves only a finite number of terms, it is relatively easy to prove
the previous proposition.
Proof. The expectation of g(X) is
p
 
Pd
X
E(g(X)) = E  ak Ψk (X) (225)
k=1
p
Pd
X
= E(ak Ψk (X)) (226)
k=1
p
Pd
X
= ak E(Ψk (X)), (227)
k=1

since the expectation of a sum is the sum of expectations. Then, the equation 223 is a straight-
forward consequence of the proposition 2.16.
The variance of g(X) is
 p 
Pd
X
V (g(X)) = V  ak Ψk (X) (228)
k=1

Pdp Pd p
X X
= V (ak Ψk (X)) + Cov(ak Ψk (X), a` Ψ` (X)) (229)
k=1 k,`=1
k6=`
p p
Pd Pd
X X
= a2k V (Ψk (X)) + ak a` Cov(Ψk (X), Ψ` (X)). (230)
k=1 k,`=1
k6=`

However, the proposition 2.18 states that the covariances are zero. Moreover, the variance of Ψ1
is zero, since this is a constant. This leads to the equation 224.
In the following proposition, we prove the equation 217 in the case of the truncated decom-
position.

Proposition 3.3. ( Coefficients of the truncated decomposition) The truncated expansion 222 is
such that
hg, Ψk i
ak = . (231)
kΨk k2

41
Proof. Indeed,
Z
hg, Ψk i = g(x)Ψk (x)w(x)dx (232)
I
Pdp Z
X
= a` Ψ` (x)Ψk (x)w(x)dx (233)
`=1 I
p
Pd
X
= a` hΨ` , Ψk i (234)
`=1
= ak hΨk , Ψk i (235)
= ak kΨk k2 , (236)
where the equation 235 is implied by the orthogonality of the functions {Ψk }k≥1 .
An immediate consequence of the equation 3.3 is that the decomposition 222 is unique. Indeed,
assume that ak and bk are real numbers and consider the decompositions
p
Pd
X
g(X) ≈ ak Ψk (X), (237)
k=1
and
p
Pd
X
g(X) ≈ bk Ψk (X). (238)
k=1
The equation 231 is satisfied both by the coefficients ak and bk , so that
ak = b k . (239)
The following decomposition shows how the coefficients be be expressed in termes of expecta-
tions and variances.
Proposition 3.4. ( Coefficients of the truncated decomposition (2)) The truncated expansion 222
is such that
a1 = E(g(X)) (240)
and
E (g(X)Ψk (X))
ak = (241)
V (Ψk (X))
for k > 1.
Proof. By definition of the expectation, we have
Z
E (g(X)Ψk (X)) = g(x)Ψk (x)f (x)dx (242)
I Z
1
= R g(x)Ψk (x)w(x)dx (243)
I
w(x)dx I
hg, Ψk i
= R (244)
I
w(x)dx
ak kΨk k2
= R , (245)
I
w(x)dx

42
where the last equation is implied by 231. Hence,
Z
ak
E (g(X)Ψk (X)) = R Ψk (x)2 w(x)dx (246)
w(x)dx
IZ I

= ak Ψk (x)2 f (x)dx (247)

I
= ak E(Ψk (X)2 ). (248)

This implies:
E (g(X)Ψk (X))
ak = . (249)
E(Ψk (X)2 )
For k = 1, we have Ψ1 (X) = 1, so that the previous equation implies 240. For k > 1, we have
E (g(X)Ψk (X))
ak = . (250)
V (Ψk (X)) + E(Ψk (X))2
However, we know from proposition 2.16 that, for k > 1, we have E(Ψk (X)) = 0, which concludes
the proof.

3.3 Univariate decomposition examples

In this section, we present several examples of decomposition of univariate random variables on
Hermite polynomials.
Assume that X is a standard univariate normal random variable (i.e. p = 1) and consider the
random variable g(X) where g is a square integrable function. Its truncated polynomial chaos
decomposition is
Pd1
X
g(X) ≈ ak Ψk (X), (251)
k=1

where Ψk are the univariate Hermite polynomials. The number of univariate polynomials is given
by the proposition 2.6, which states that, for p = 1, we have
 
1 1+d
Pd = (252)
d
(1 + d)!
= (253)
d!1!
= 1 + d. (254)

Moreover, the equation 164 states that the functions Ψk are defined in terms of the Hermite
polynomials as

Ψk (x) = Heα(k) (x), (255)

for any x ∈ R and k ≥ 1, where

α(k) = k − 1. (256)

43
This implies :
Ψk (x) = Hek−1 (x), (257)
for any x ∈ R and k ≥ 1. The equation 251 simplifies to:
g(X) ≈ a1 He0 (X) + a2 He1 (X) + ... + ad+1 Hed (X), (258)
where d ≥ 1 is an integer and {Hei }i=0,1,...,d are Hermite polynomials.
The proposition 3.4 states that the coefficients are
a1 = E(g(X)) (259)
and
E (g(X)Hek−1 (X))
ak = (260)
V (Hek−1 (X))
E (g(X)Hek−1 (X))
= (261)
(k − 1)!
for k > 1. In general, this requires to compute the integral
Z
E (g(X)Hek−1 (X)) = g(x)Hek−1 (x)f (x)dx, (262)
R

for k ≥ 1, where f is the Gaussian probability distribution function defined in the equation 48.
In this section, we consider several specific examples of functions g and present the associated
coefficients ai , for i = 1, 2, ..., d + 1.
Example (Constant) Assume that
g(x) = c, (263)
for some real constant c. A possible exact decomposition is:
g(X) = cHe0 (X). (264)
Since the polynomial chaos decomposition is unique, this is the decomposition. But it is interesting
to see how the coefficients can be computed in this particular case. We have
a1 = E(c) (265)
= c. (266)
Moreover,
E (cHek−1 (X))
ak = (267)
(k − 1)!
E (Hek−1 (X))
= c (268)
(k − 1)!
However, the proposition 1.9 states that E (Hek−1 (X)) = 0 for k > 1. This implies
ak = 0 (269)
for k > 1.

44
Example (Standard normal random variable) Assume that

g(x) = x, (270)

for any x ∈ R. The exact polynomial chaos decomposition is

g(X) = He1 (X), (271)

since He1 (X) = X. Again, it is interesting to see how the coefficients can be computed in this
particular case. We have

a1 = E(X) (272)
= 0, (273)

since, by assumption, X has a mean equal to zero. Moreover,

E (XHek−1 (X))
ak = (274)
(k − 1)!
E (He1 (X)Hek−1 (X))
= , (275)
(k − 1)!

since the first Hermite polynomial is He1 (x) = x. The equation 39 then implies ak = 0, for k > 2.
Moreover,

E (He1 (X)2 )
a2 = (276)
V (He1 (X))
V (He1 (X))
= (277)
0!
0!
= (278)
1
= 1, (279)

where the previous equation comes from the equation 40. This immediately leads to the equation
271.

Example (The square function) Consider the function

g(x) = x2 , (280)

for any x ∈ R. Consider the random variable g(X) = X, where X is a standard normal random
variable. From the table 6, we see that

x2 = x2 − 1 + 1 (281)
= He2 (x) + He0 (x). (282)

Therefore, the exact polynomial chaos decomposition of g is

g(x) = He0 (x) + He2 (x). (283)

45
 k a0 a1 a2 a3 a4 a5 a6 a7 a8 a9
x0 1
x1 1
x2 1 1
x3 3 1
x4 3 6 1
x5 15 10 1
x6 15 45 15 1
x7 105 105 21 1
x8 105 420 210 28 1
x9 945 1260 378 36 1

Figure 24: Coefficients ak in the decomposition of xn into Hermite polynomials, for n = 0, 1, ..., 9.

Example (Decomposition of xn ) We can compute the decomposition of xn , where n is a positive

integer. For example, we have

x5 = 15He1 (x) + 10He3 (x) + He5 (x). (284)

In the figure 24, we present the coefficients ak in the decomposition:

X
xn = ak Hek (x). (285)
k≥0

Example (Transformation) TODO : X normal µ, σ and the standardization of X.

Example (Some approximate decompositions) We compute the approximate polynomial chaos

decomposition of several functions, where X is a standard normal random variable. In order to
perform the numerical integration involved in the integral 262, we have used the intg function,
with an absolute tolerance equal to 1.e-8, and numerical bounds equal to -10 and +10. We
compute the coefficients up to the degree 14.
These numerical experiments are consistent with the following exact integrals, computed from
Wolfram Alpha [14].

Z ∞
cos(x)Hen (x)w(x)dx = { (286)
−∞

3.4 Generalized polynomial chaos

In 2002, Xiu and Karniadakis [18, 19] introduced the generalized polynomial chaos. They notice
that the Hermite polynomials are quite effective to solve some stochastic differential equations
involving Gaussian as well as non Gaussian random variables, e.g. log-normal random variables.
However, they emphasize that, for general non-Gaussian random inputs, the convergence rate is
not fast. This is why they introduce the generalized polynomial chaos, where each probability
distribution function is associated to a family of orthogonal polynomials.

46
 cos(x) sin(x)
a0 0.6065307
a1 0.6065307
a2 -0.3032653
a3 -0.1010884
a4 0.0252721
a5 0.0050544
a6 -0.0008424
a7 -0.0001203
a8 0.0000150
a9 0.0000017
a10 -0.0000002
a11 -1.519D-08
a12 1.266D-09
a13 9.740D-11
a14 -6.957D-12

Figure 25: Coefficients ak in the decomposition of several functions into Hermite polynomials.
Distribution Polynomial Support
Continuous Normal Hermite (−∞, ∞)
Distribution Gamma (Exponential) Laguerre [0, ∞)
Beta Jacobi [a, b]
Uniform Legendre [a, b]
Discrete Poisson Chalier {0, 1, 2, ...}
Distribution Binomial Krawtchouk {0, 1, 2, ..., N }
Negative Binomial Meixner {0, 1, 2, ...}
Hypergeometric Hahn {0, 1, 2, ..., N }

Figure 26: Map from a distribution function to the associated orthogonal polynomials [12].

The involved orthogonal polynomials are constructed from limit relationships from the hyper-
geometric orthogonal polynomials. This leads to a classification which allows to create a tree of
relations between the polynomials, called the Askey scheme. The mapping from the distribution
to the corresponding orthogonal polynomial is presented in the figure 26.
The finite decomposition is
X
g(X) ≈ aα Ψα (X), (287)
α∈Jp,d

where X = (X1 , X2 , ..., Xp ) are independent random variables and Ψk is the tensor product of
orthogonal polynomials φ, so that
p
Y
Ψk (x) = φα(k) (xi ) (288)
i
i=1

for k = 1, 2, ..., Pdp .

47
Example Assume that X1 is a standard normal random variable, X2 is a uniform random vari-
able in the interval [−1, 1], and assume that X1 and X2 are independent. Therefore, the gener-
alized polynomial chaos associated with the random variable X = (X1 , X2 ) involves the Hermite
orthogonal polynomials (for X1 ) and the Legendre orthogonal polynomials (for X2 ).

Example (Transformation) TODO : X U(0,1) and G(X)=exp(X).

3.5 Transformations
TODO

3.6 Notes and references

The book [11] and the thesis [10] present spectral methods, including polynomial chaos.
In the figure 24, we present the coefficients ak in the decomposition of xn into Hermite poly-
nomials, for n = 0, 1, ..., 9. This table is similar to the table 22.12 in the chapter 22 of [1]. But
our table presents the coefficients for the probabilist’s Hermite polynomials Hen , while the table
22.12 presents the coefficients of the physicist’s Hermite polynomials Hn .

4 Thanks
TODO

A Gaussian integrals
A.1 Gaussian integral
Proposition A.1. ( Gaussian integral)
Z +∞ √
exp(−x2 /2)dx = 2π. (289)
−∞

Proof. We have,
Z +∞ 2 Z +∞  Z +∞ 
2 2 2
exp(−x )dx = exp(−x )dx exp(−x )dx (290)
−∞ −∞ −∞
Z +∞  Z +∞ 
2 2
= exp(−x )dx exp(−y )dy (291)
−∞ −∞
Z +∞ Z +∞
exp −x2 − y 2 dxdy.


= (292)
−∞ −∞

Let us consider the polar change of variable:

x = r cos(θ), (293)
y = r sin(θ) (294)

48
where r ∈ R and θ ∈ [0, 2π]. Its Jacobian is J(r, θ) = r (see, e.g. [6]). Hence,
Z +∞ 2 Z +∞ Z 2π
2
exp −r2 rdrdθ


exp(−x )dx = (295)
−∞ 0 0
Z +∞
r exp −r2 dr.


= 2π (296)
0

Consider the change of variable s = −r2 . We have ds = 2rdr, which implies

Z +∞ 2 Z 0
2
exp(−x )dx = π exp (−s) ds (297)
−∞ −∞
= π(exp(0) − 0) (298)
= π. (299)
Hence,
√
Z
exp(−x2 )dx = π. (300)
R
√ √
Now consider the change of variable y = 2x. We have dx = dy/ 2, which implies
Z Z
2 1
exp(−x )dx = √ exp(−y 2 /2)dy. (301)
R 2 R

We plug the equation 300 into the previous equality and get 289.
The previous proof does not take into account the improper integrals which appear during the
computation.

A.2 Weighted integral of xn

Proposition A.2. For any odd n,
Z +∞
xn exp(−x2 /2)dx = 0, (302)
−∞

Proof. If n is odd, the function xn is antysymmetric, since (−x)n = −xn for any x ∈ R. On
the other hand, the function exp(−x2 /2) is symmetric, since exp(−(−x)2 /2) = exp(−x2 /2). We
consider the change of variable y = −x and get:
Z 0 Z +∞
n 2
x exp(−x /2)dx = (−y)n exp(−(−y)2 /2)dy (303)
−∞ 0
Z +∞
=− y n exp(−y 2 /2)dy. (304)
0

This implies
Z +∞ Z 0 Z +∞
n 2 n 2
x exp(−x /2)dx = x exp(−x /2)dx + xn exp(−x2 /2)dx (305)
−∞ −∞ 0
= 0. (306)
and concludes the proof.

49
Proposition A.3. For any even n, we have
Z +∞ √
xn exp(−x2 /2)dx = 1 · 3 · 5 · ... · (n − 1) 2π. (307)
−∞

Proof. Let us denote by In the integral

Z +∞
In = xn exp(−x2 /2)dx. (308)
−∞

We are going to prove the equality

In+2 = (n + 1)In . (309)

Integrating by part, we get

 n+1 +∞ Z +∞ n+1
x 2 x
In = exp(−x /2) + x exp(−x2 /2)dx. (310)
n+1 −∞ −∞ n + 1

On the other hand,

lim xn exp(−x2 /2) = 0, (311)

x→+∞

for any integer n. Hence,

+∞
xn+1

exp(−x2 /2) = 0. (312)
n+1 −∞

This implies
+∞
xn+2
Z
In = exp(−x2 /2)dx (313)
−∞ n+1
1
= In+2 . (314)
n+1
√
Moreover, we from from proposition A.1 that I0 = 2π, which concludes the proof.
Consider the Gaussian weight function w(x) = exp(−x2 /2). The consequence of the proposi-
tions A.2 and A.3 is that a monomial xn is in L2w (R), for any integer n.

50
References
[1] Milton Abramowitz and Irene A. Stegun. Handbook of Mathematical Functions with Formu-
las, Graphs, and Mathematical Tables. Dover, New York, ninth dover printing, tenth gpo
printing edition, 1964.

[2] George E. Andrews, Richard Askey, and Ranjan Roy. Special functions, volume 71 of En-
cyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge,
1999.

[3] G. Arfken. Mathematical Methods for Physicists, 3rd ed. Elsevier Academic Press, 2005.

[4] M. Branicki and A.J Majda. Fundamental limitations of polynomial chaos for uncertainty
quantification in systems with intermittent instabilities, 2012.

[5] R. H. Cameron and W. T. Martin. The orthogonal development of non-linear functionals in

series of fourier-hermite functionals. Annals of Mathematics, 48(2):385–392, April 1947.

[6] Philip J. Davis and Philip Rabinowitz. Methods of Numerical Integration. Academic Press,
New York, 1984.

[7] William Feller. An Introduction to Probability Theory and Its Applications - Volume I. John
Wiley and Sons, third edition edition, 1968.

[8] Nicholas J. Higham. Accuracy and Stability of Numerical Algorithms. Society for Industrial
and Applied Mathematics, Philadelphia, PA, USA, second edition, 2002.

[9] Thomas Y. Hou, Wuan Luo, Boris Rozovskii, and Hao min Zhou. Wiener chaos expansions
and numerical solutions of randomly forced equations of fluid mechanics. J. Comput. Phys,
216:687–706, 2006.

[10] O.P. Le Maı̂tre. Méthodes spectrales pour la propagation d’incertitudes paramétriques dans
les modèles numériques, mémoire d’habilitation à diriger des recherches. Springer, 2005.

[11] O.P. Le Maı̂tre and O.M. Knio. Spectral methods for uncertainty quantification. Springer,
2010.

[12] Didier Lucor. Introduction to stochastic spectral methods, November 2008.

[13] Wuan Luo. Wiener Chaos Expansion and Numerical Solutions of Stochastic Stochastic Par-
tial Differential Equations. PhD thesis, California Institute of Technology, May 2006.

[14] Wolfram Research. Wolfram alpha. http://www.wolframalpha.com.

[15] Gabor Szegö. Orthogonal polynomials. American Mathematical Society, 1975.

[16] Norbert Wiener. The homogeneous chaos. Amer. J. Math., 60(4):897–936, 1938.

[17] Wikipedia. Hermite polynomials — wikipedia, the free encyclopedia, 2013. [Online; accessed
26-January-2013].

51
[18] Dongbin Xiu and George Em Karniadakis. The wiener–askey polynomial chaos for stochastic
differential equations. SIAM J. Sci. Comput., 24(2):619–644, February 2002.

[19] Dongbin Xiu and George Em Karniadakis. Modeling uncertainty in flow simulations via
generalized polynomial chaos. J. Comput. Phys., 187(1):137–167, May 2003.

52