On the Computation of Integrals of Bivariate Gaussian

Distribution
Vincent Savaux, Luc Le Magoarou

To cite this version:

Vincent Savaux, Luc Le Magoarou. On the Computation of Integrals of Bivariate Gaussian Distri-
bution. IEEE Symposium on Computers and Communications (ISCC), Jul 2020, Rennes, France.
�hal-02555869�

HAL Id: hal-02555869

https://hal.archives-ouvertes.fr/hal-02555869
Submitted on 27 Apr 2020

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 On the Computation of Integrals of Bivariate
Gaussian Distribution
Vincent Savaux and Luc Le Magoarou
b<>com, Rennes, France
emails: {vincent.savaux,luc.lemagoarou}@b-com.com

Abstract—This paper deals with the computation of integrals f2(x) y

of centred bivariate Gaussian densities over any domain defined
as an angular sector of R2 . Based on an accessible geometrical
approach of the problem, we suggest to transform the double
integral into a single one, leading to a tractable closed-form f1(x)
expression only involving trigonometric functions. This solution
can also be seen as the angular cumulative distribution of
bivariate centered Gaussian variables (X, Y ). We aim to provide
a didactic approach of our results, and we validate them by
comparing with those of the literature.
Index Terms—Bivariate Gaussian distribution, Digital commu-
nications, Signal processing, PSK x

I. I NTRODUCTION
The problems of estimating unknown parameters from noisy
(disturbed) Gaussian observations and characterizing the per-
formance of estimators are ubiquitous in a vast majority of
quantitative fields. They have been thoroughly studied in an
unified way by detection and estimation theory [1]–[3], and Fig. 1. Example of domain of integration Ω.
are frequently met in signal processing and digital commu-
nications [4], [5] as well. The problems involving Gaussian
observations usually lead to the computation of multiple this approach is original. Furthermore, this leads us to define
integrals of multivariate Gaussian distributions. However, these an angular cumulative distribution function (CDF) of zero-
integrals generally do not admit closed-form solutions, and are mean bivariate Gaussian variables. Our geometrical approach
costly to approximate. of the problem is didactically illustrated, and our results are
In this paper, we focus on integrals of dimension two, i.e. compared with those of the literature [6]–[8].
we consider integrals of bivariate Gaussian densities over any The rest of the paper is organized as follows: Section II
angular sector of R2 . This problem has been widely tackled presents the general problem to be solved and gives the exist-
in literature [6]–[8], but leads to simple specific solutions ing special solutions. In Section III, we derive the generalized
or intractable general solutions. Indeed, authors in [6], [7] solution, and Section V concludes this paper.
propose a closed-form expression of the integral in the case Notations: Vectors and matrices are respectively written in
where two correlated Gaussian variables have the same unitary lower and upper case in boldface: x, and X. The scalars are
variance, and the domain of integration is one quadrant of R2 written in normal font: x or X. Moreover, (.)T and |.| stand
(e.g. (R+ , R+ )). In [8], the distributions of the envelope and for the transpose and determinant operators, respectively, and
the phase of two correlated Gaussian variables are derived in a E{.} for the mathematical expectation.
very general case of non-zero mean and non-unitary variance.
However, these distributions have no closed-form nor tractable II. BACKGROUND AND C OMMON R ESULTS
expressions, and the resulting probabilities even require the Let X and Y be two zero-mean Gaussian random variables
integration of these ones. with variances σx2 = E{X 2 } and σy2 = E{Y 2 }, respectively.
In the present paper, we tackle the problem of the integral Moreover, the two variables X and Y are correlated, such
of a bivariate Gaussian centred distribution over any angular that E{XY } = ρσx σy , where ρ ∈ [−1, 1] is the so-called
sector of R2 delimited by two linear functions, and in the correlation coefficient. The joint probability density function
case where two correlated Gaussian variables have different of (X, Y ), denoted by gX,Y , can be expressed as
variances. To do so, we use a geometrical approach to rewrite
the double integral as a single one, leading to an accessible and 1 1
gX,Y (x) = exp(− xT Σ−1 x), (1)
tractable closed-form solution. To the best of our knowledge,
p
2π |Σ| 2
 y
where x = (x, y) ∈ R2 , and Σ is the covariance matrix defined ν f2(x)
as
σν θ2
!
σx2 ρσx σy
Σ = E{(X, Y )T (X, Y )} = , (2)
ρσx σy σy2 A(θ1,θ2) σξ
ξ
such that |Σ| = σx2 σy2 (1 − ρ2 ). Computing any probability θ1
involving both X and Y (e.g. P(X ≶ 0 ∩ Y ≶ 0) or P(X ≶
0 ∪ Y ≶ 0)) requires the integral of (1) over subsets of R2 . In φ
f1(x)
the following, we consider the integral of gX,Y (x) over any x
angular sector in R2 , and we define the integral
ZZ Fig. 2. Sector area of an ellipse in the general case.
M= gX,Y (x)dxdy, (3)
Ω the semi-major or semi-minor axes of the ellipse (5). Using
where Ω is the angular sector which can be defined by the general ellipse equation and straightforward geometrical
developments, they are expressed as
Ω = {f1 (x) ≶ y ≶ f2 (x), x ∈ R}, (4)

1
and f1 (x) and f2 (x) are linear functions such as illustrated σξ = σx2 cos2 (ϕ) + σy2 sin2 (ϕ) + ρσx σy sin(2ϕ) 2
in Fig. 1. Note that such angular divisions of R2 are usually
1
σν = σx2 sin2 (ϕ) + σy2 cos2 (ϕ) − ρσx σy sin(2ϕ) 2 , (7)
considered in digital communications, for instance to define
phase shift keying (PSK) modulations. where ϕ is the angle between the axes x and ξ, and is defined
It must be reminded that a straightforward closed form of as
M is obtained when Σ is a scaled identity matrix. Moreover,  
2ρσx σy

1
a closed-form solution of the double integral (3) has been  2 arctan σ2 −σ 2 , if σx2 > σy2
x  y


suggested in [6], [7] for any ρ value, but only in the case

ϕ = π + 1 arctan 2ρσ x σy
if σx2 > σy2 . (8)
where σx2 = σy2 = 1 and Ω = (R+ , R+ ). In that case, we  2 2 σx −σy ,
2 2

π
obtain M = 41 + arcsin(ρ) . We hereby generalize the result to 4, if σx2 = σy2
2π
any case where σx2 and σy2 are non-unitary variances, as well Theorem 1. Let p ∈ [0, 1] be defined as the ratio of the semi-
as any domain of integration defined as Ω. Moreover, this minor and semi-major axes of the ellipse in (5), i.e. p = σσνξ .
leads us to define an angular cumulative distribution function Then, for any θ1 , θ2 ∈ [−π, π], M can be expressed as
of bivariate Gaussian random variables.
ZZ
III. A C LOSED -F ORM S OLUTION TO (3) M= gX,Y (x)dxdy
In this section, we derive a general solution of (3) in Ω
Theorem 1. To this end, let us first consider the change of Q(θ1 , θ2 )
basis (x, y) 7→ (ξ, ν) where (ξ, ν) correspond to semi-axes of = , (9)
π
the ellipse defined by the equation where Q(θ1 , θ2 ) is the function such that the area A(θ1 , θ2 )
(x, y)T Σ−1
m,n (x, y) = 1, (5) of the elliptic sector bounded by (θ1 , θ2 ) is defined by
A(θ1 , θ2 ) = σν σξ Q(θ1 , θ2 ). For simplicity purpose, we denote
such as illustrated in Fig. 2. This ellipse is a level set of θM = max(|θ1 |, |θ2 |) and θm = min(|θ1 |, |θ2 |). Then, if
the function gX,Y . Without loss of generality, we assume that θ1 , θ2 ∈ [− π2 , π2 ], we have
ξ and ν correspond to the semi-major and semi-minor axes,
respectively. Moreover, θ1 and θ2 are arbitrarily defined as the (
u(|θM |) − u(|θm |), if sign(θ1 ) = sign(θ2 )
angles between the semi-major axis ξ and the lines f1 (x) and Q(θ1 , θ2 ) = ,
f2 (x), respectively. u(|θM |) + u(|θm |), if sign(θ1 ) 6= sign(θ2 )
After the aforementioned change of basis, M in (3) can be (10)
rewritten as where u(θ) = 21 arctan( p1 tan(θ)). If θ1 ∈ [−π, − π2 ] ∪ [ π2 , π]
and θ2 ∈ [− π2 , π2 ] (θ1 and θ2 are exchangeable), then we get
ZZ
1  1 ξ

2 ν
2 
M= exp − + dξdν, (6) 
π
2πσξ σν 2 σξ σν 2 − u(π − |θM |) − u(|θm |), if


Ωξ,ν


sign(θ1 ) = sign(θ2 )

where Ωξ,ν is the new domain of integration of ξ and ν, Q(θ1 , θ2 ) = . (11)
π
which corresponds to the rotated angular sector Ω in the new 2 − u(π − |θM |) + u(|θm |), if




basis. The standard deviations σξ and σν are the length of

sign(θ ) 6= sign(θ )
1 2
 allowing us to write ξ ∈ [0, +∞[ in function of z. Such
as aforementioned, M in (6) can then be rewritten as the
C integral of the sector areas A(ξ(z), θ1 , θ2 ) of the ellipse over
z ∈ [0, C = 2πσ1ν σξ ] (see Fig. 3), which yields
A(ξ(z),θ1,θ2) Z C
M= A(ξ(z), θ1 , θ2 )dz. (15)
0
σν
Hence, considering the substitution X = ln( Cz ) and p = σξ ,
z dz we obtain

Z C
C
M= 2pQ(θ1 , θ2 )σξ2 ln( )dz
z
Z0 +∞
= 2pQ(θ1 , θ2 )σξ2 Ce−X dX
0 0
ξ(z) ξ pQ(θ1 , θ2 )σξ2
=
Fig. 3. The integral M can be computed by integrating A(ξ(z), θ1 , θ2 ) over πσν σξ
the axis z. Q(θ1 , θ2 )
= , (16)
π
Finally, if θ1 , θ2 ∈ [−π, − π2 ] ∪ [ π2 , π], we obtain which concludes the proof.

 It is worth mentioning that the suggested closed-form

u(π − |θm |) − u(π − |θM |),


if solution only involves the computation of the well-known

sign(θ1 ) = sign(θ2 )
 trigonometric functions tan and arctan, and avoids the com-
Q(θ1 , θ2 ) = . (12) putation of the integral of the distributions in [8], or the


 π − u(π − |θM |) − u(π − |θm |), if double integral as in (3). In fact, the latter should usually be


sign(θ ) 6= sign(θ ) approximated by truncating discrete sums or by using a Monte-
1 2
Carlo method, both methods being computationally expensive.
Proof. It must be emphasized that, although the expression of In the following, we provide general results related to the
M seems to be more tractable in (6) than in (3), it can still not analytic solution of the integral M, and we show that it
be computed in closed form, due to the complicated domains reduces to known results given in [6], [7] in specific cases.
of integration. However, it can be noticed that the value M of Furthermore, an example of application is given as well.
the double integral corresponds to the volume under the curve
IV. R ESULTS R ELATED TO M
1  1 ξ ν 
g(ξ, ν) = exp − ( )2 + ( )2 , A. Angular Cumulative Distribution Function M
2πσξ σν 2 σξ σν
A first important result is that M in (9) can be interpreted
in the angular sector Ω (or equivalently Ωξ,ν ). This volume
as the angular cumulative distribution (CDF) of bivariate
can also be assessed by summing (integrating) the areas A
Gaussian variables (X, Y ). Thus, Fig. 4 illustrates M in
of the elliptical sectors corresponding to the angular sector
function of (θ1 , θ2 ) ∈ [−π, π]2 , for p = 13 . It can be observed
Ω (see Figs. 2 and 3). For any angles θ1 , θ2 and any length
that the CDF M is symmetric with respect to the line θ2 = θ1 ,
of the semi-major axis ξ, given that p = νξ , the sector area
for which M = 0, and that M = 1 for θ2 = −θ1 = ±π. Note
A(ξ, θ1 , θ2 ) of the ellipse is given by
that this representation of the angular CDF M depends on the
values of ρ, σx , and σy only through the p value, since Fig. 4
A(ξ, θ1 , θ2 ) = ξνQ(θ1 , θ2 ) = ξ 2 pQ(θ1 , θ2 ). (13) holds in the basis (ξ, ν). Different behaviors of M would be
The expression of A is proved in Appendix A. It follows from obtained in the basis (x, y), such as discussed afterward.
(13) that the sector area of the ellipse only depends of the Alternatively to the two-dimensional representation of M,
angular sector defined by (θ1 , θ2 ), and the length ξ. By setting Fig. 5 shows the CDF M versus θ1 ∈ [−π, π] for a fixed
z = g(ξ, ν = 0) and ξ ∈ R+ , then for any z ∈ [0, C = 2πσ1ν σξ ] θ2 = −π. Note that in that case, the expression of M in
the following bijection holds: (9)-(12) simplifies to:

 u(π+θ )
ξ(z) =g −1 (z)  π ,
 1
if θ1 ∈ [−π, − π2 ]

 C  12 M(θ1 ) = 21 + u(θπ1 ) , if θ1 ∈ [− π2 , π2 ] . (17)
= 2σξ2 ln( ) , (14) 
1 − u(π−θ 1)
, if θ1 ∈ [ π2 , π]

z 
π
 1 r 1 + ρ 
M = arctan . (18)
1 π 1−ρ
0.8
Then, by using some relationship between trigonometric func-
0.6
tions (e.g. the half-angle formula), we obtain the equivalent
CDF

0.4
expressions:
0.2

0
1  1+ρ 
3 M= arctan p
2
π 1 − ρ2
1 1 1  p1 − ρ2 
= − arctan
θ2
0
2 π 1+ρ
-1
1 1
-2 3
= − arccos(ρ)
0
1
2 2 2π
-3
-2
-1
θ1
1 1
-3
= + arcsin(ρ). (19)
4 2π
Similarly, it can be straightforwardly shown that if ρ ≥ 0 and
Fig. 4. M interpreted as the angular CDF of bivariate Gaussian variables
(X, Y ), for p = 1/3. Ω = (R+ , R− ) (or equivalently ρ ≤ 0 and Ω = (R+ , R+ ))
1
we have M = 2π arccos(ρ). These results also match
√ those
1−ρ2
obtained by the integration of the PDF fθ (θ) = 2π(1−ρ sin(2θ))
1
given in [8] (see (4) in [8] considering zero-mean variables):
0.9
Z π2
1 1
0.8 fθ (θ)dθ = + arcsin(ρ). (20)
0 4 2π
0.7
θ2 = - π Moreover, it must be noted that these expressions only
0.6
depend on ρ, and we can show that this behavior holds for
CDF

0.5 any variance of the variables X and Y as well. In fact, we

0.4 directly have
p=1/3
0.3 p=1/10
p=10/11
1 1
0.2 P = P(X ≶ 0 ∩ Y ≶ 0) = P( X ≶0∩ Y ≶ 0), (21)
σx σy
0.1
where σ1x X and σ1y Y have unitary variance. In the following,
0
-4 -3 -2 -1 0 1 2 3 4 we show that the suggested general expression is mandatory
θ1
to calculate more elaborate probabilities than (21).
Fig. 5. Angular CDF M versus θ1 ∈ [−π, π], θ2 = −π. Comparison for C. Application to More Elaborate Probabilities
different p values.
Let us consider zero-mean Gaussian variables X and Y with
variances σx2 and σy2 , respectively, and correlation coefficient
Three behaviors of M, taking p ∈ { 31 , 10 1 10
, 11 } are shown in ρ. For any (α, β) ∈ R2+ , the suggested analysis allows us to
Fig. 5. It can be observed that M tends to the linear function calculate probabilities such as
θ1
M(θ1 ) = 12 + 2π when p approaches 1. This highlights the
fact that p = 1 corresponds to an uncorrelated couple of P = P(X ≥ 0 ∩ Y ≥ 0 ∩ αY ≥ βX), (22)
Gaussian variables (X, Y ) with σx2 = σy2 (i.e. the ellipse (5) where P is not independent of the variances of X and Y
reduces to a circle), therefore having an uniformly distributed anymore due to the additional event αY ≥ βX. To reduce
phase. Otherwise when p approaches 0, M approaches a step the number of degrees of freedom of the problem, let us set
function, highlighting that the ellipse becomes a line. r = σσxy , and γ = αβ
. Then, the orientation φ and the ratio p
B. Particular Cases of the ellipse depends on both ρ and r, and the domain of
integration Ω (partially) depends on γ, such as illustrated in
In this section we show that the general expression of M Fig. 6. Thus, in the example given in (22), we have θ2 = π2
in (9)-(12) simplifies to known results given in [6], [7] when and θ1 = arctan(γ).
σx2 = σy2 = 1 and when Ω is one of the four quadrants of R2 . To better illustrate the dependence of P in (22) on the
If we assume that ρ q ≥ 0 without loss of generality, we have variances of X and Y , we consider the following numerical
ϕ = π4 and then p = 1−ρ 1+ρ . Consider the quadrant (R+ , R+ ), application: ρ = 0.5, γ = 1, and r ∈ { 14 , 1, 4}, and we
then θ1 = − π4 , and θ2 = π4 , hence M in (9)-(10) becomes imposed σx σy = 4. Fig. 7 shows the level sets of the three
 y instead of (x, y), leading to tractable closed-form expressions
γx of M, where the parameters σx , σy , and φ inherent to the
basis (x, y) are included in the geometrical parameter p in the
basis (ξ, ν).
To complete this study, non-zero means of X and Y could
be included in a similar geometrical approach of the problem.
Furthermore, this analysis could be extended in dimensions
larger than two (involving ellipsoides), in order to tackle the
σν problem of the integrals of multivariate Gaussian distributions.

φ V. C ONCLUSION
In this paper we suggested a geometrical approach to
x solve the problem of integrating centered bivariate Gaussian
densities over any domain defined as an angular sector of R2 .
This led to a tractable solution composed of trigonometric
functions, which can be seen a the angular CDF of a couple
of centered Gaussian variables (X, Y ). The developments
were illustrated to provide a didactic approach of our results.
Furthermore, we compared them to those of the literature in
Fig. 6. Example of application for problems such as (22). order to validate our solution. Future works will investigate the
possible extension of the suggested geometrical approach to
non-centred correlated Gaussian variables, as well as problems
10 of dimension larger than two.
9 r = 0.25
P = 0.29
A PPENDIX
8
A. Proof of the expression of A(ξ, θ)
7
y=x We consider an ellipse with semi-major and semi-minor
6 axes ξ and ν, respectively. The area A of the elliptical sector
5
r=1 between the semi-major axis and the radius r with angle dθ
y

P = 0.167 2
is equal to A = r2 dθ. The parametric equation of the ellipse
4
is expressed as
r=4
3
P = 0.04
x = ξ cos(t)
2
y = ν sin(t).
1
Thus, any radius with length r and angle φ (corresponding to
0
0 1 2 3 4 5 6 7 8 9 10 parameter t = tφ ) with respect to the semi-major axis admits
x
the equations:

Fig. 7. M largely depends on the variances of X and Y .

y ν
= tan(φ) = tan(tφ )
x ξ
considered bivariate Gaussian distributions. The domain of ξ
⇒tφ = arctan( tan(φ)), (23)
integration Ω is above the line y = x. The corresponding ν
values of probability P have been reported on Fig. 7, and it and then the substitution of x and y in r2 = x2 + y 2 yields
can be shown that the lower r, the larger ϕ and therefore the
larger P, since the sector areas A(ξ, θ1 , θ2 ) of the ellipses in  ξ 2  ξ tan(φ) 2
Ω are greater. r2 (φ) = q + q ,
ξ 2 tan2 (φ) ξ 2 tan2 (φ)
Such as mentioned in Section II and stated in [8], such ν2 ν2
results can be applied in the domains of digital communi-
which simplifies into
cations and signal processing. For instance, it allows for an
accurate characterization of the error rate of PSK-modulated 1
r2 (φ) = cos2 (φ) sin2 (φ)
.
signals with inphase/quadrature (I//Q) correlated noise. The +
ξ2 ν2
main difference of the suggested analysis with results given
in [8] is that we carry out our developments in the basis (ξ, ν) Hence we finally obtain
 1 θ 2
Z
A(θ) = r (φ)dφ
2 0
ξν ξ
= arctan( tan(θ)), (24)
2 ν
which concludes the proof.
R EFERENCES
[1] E. L. Lehmann and J. P. Romano, Testing statistical hypotheses. Springer
Science & Business Media, 2006.
[2] E. L. Lehmann and G. Casella, Theory of point estimation. Springer
Science & Business Media, 2006.
[3] H. L. Van Trees, Detection, estimation, and modulation theory, part I:
detection, estimation, and linear modulation theory. John Wiley & Sons,
2004.
[4] S. M. Kay, “Fundamentals of Statistical Signal Processing: Estimation
Theory”. Prentice-Hall, 1998.
[5] ——, “Fundamentals of Statistical Signal Processing: Detection The-
ory”. Prentice-Hall, 1998.
[6] C. Rose and M. D. Smith, Mathematical Statistics with Mathematica.
Springer-Verlag, 2002, ch. 6: Multivariate Distributions, pp. 187 – 249.
[7] A. Stuart and J. K. Ord, Kendall’s Advanced Theory of Statistics, Volume
1, Distribution Theory. Wiley, 2010, ch. 7: Multivariate Distributions.
[8] P. Dharmawansa, N. Rajatheva, and C. Tellambura, “Envelope and Phase
Distribution of Two Correlated Gaussian Variables,” IEEE Transactions
on Communications, vol. 57, no. 4, pp. 915 – 921, April 2009.