Electronic Journal of Statistics

Vol. 11 (2017) 3081–3119

ISSN: 1935-7524
DOI: 10.1214/17-EJS1310

Poincaré inequalities on intervals –

application to sensitivity analysis
Olivier Roustant
Mines Saint-Etienne
UMR CNRS 6158, LIMOS, F–42023 Saint-Étienne, France
e-mail: roustant@emse.fr

Franck Barthe
Institut de Mathématiques de Toulouse
Université Paul Sabatier, 31062 Toulouse Cedex 9, France
e-mail: franck.barthe@math.univ-toulouse.fr

and

Bertrand Iooss
Electricité de France R&D
6 quai Watier, Chatou, F–78401, France
&
Institut de Mathématiques de Toulouse
Université Paul Sabatier, 31062 Toulouse Cedex 9, France
e-mail: bertrand.iooss@edf.fr
Abstract: The development of global sensitivity analysis of numerical
model outputs has recently raised new issues on 1-dimensional Poincaré in-
equalities. Typically two kinds of sensitivity indices are linked by a Poincaré
type inequality, which provides upper bounds of the most interpretable in-
dex by using the other one, cheaper to compute. This allows performing
a low-cost screening of unessential variables. The efficiency of this screen-
ing then highly depends on the accuracy of the upper bounds in Poincaré
inequalities.
The novelty in the questions concern the wide range of probability dis-
tributions involved, which are often truncated on intervals. After providing
an overview of the existing knowledge and techniques, we add some theory
about Poincaré constants on intervals, with improvements for symmetric
intervals. Then we exploit the spectral interpretation for computing exact
value of Poincaré constants of any admissible distribution on a given inter-
val. We give semi-analytical results for some frequent distributions (trun-
cated exponential, triangular, truncated normal), and present a numerical
method in the general case.
Finally, an application is made to a hydrological problem, showing the
benefits of the new results in Poincaré inequalities to sensitivity analysis.

Keywords and phrases: Poincaré inequality, spectral gap, truncated dis-

tribution, Kummer’s functions, Sobol-Hoeffding decomposition, Sobol in-
dices, derivative-based global sensitivity measures, finite elements.

Received December 2016.

3081
3082 O. Roustant, F. Barthe and B. Iooss

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3082
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3082
1.2 Aim and plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3084
2 Background on Poincaré inequalities . . . . . . . . . . . . . . . . . . . 3085
2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3085
2.2 A general criterion . . . . . . . . . . . . . . . . . . . . . . . . . . 3088
2.3 Perturbation and transport . . . . . . . . . . . . . . . . . . . . . 3089
2.4 Spectral interpretation . . . . . . . . . . . . . . . . . . . . . . . . 3091
2.5 More techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . 3095
3 Poincaré inequalities on intervals . . . . . . . . . . . . . . . . . . . . . 3096
3.1 General results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3096
3.2 Improvements for symmetric settings . . . . . . . . . . . . . . . . 3097
4 Exact values of Poincaré constants . . . . . . . . . . . . . . . . . . . . 3100
4.1 Semi-analytical results . . . . . . . . . . . . . . . . . . . . . . . . 3100
4.2 A numerical method . . . . . . . . . . . . . . . . . . . . . . . . . 3108
4.3 Illustrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3109
5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3111
5.1 First study on a simplified flood model . . . . . . . . . . . . . . . 3112
5.2 Application on a 1D hydraulic model . . . . . . . . . . . . . . . . 3114
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3116
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3116

1. Introduction

1.1. Motivation

This research is motivated by sensitivity analysis of a costly numerical function

f (x1 , . . . , xd ) of independent random variables xi . We denote by μ = μ1 ⊗· · ·⊗μd
the distribution of x = (x1 , . . . , xd ) and assume that f (x) ∈ L2 (μ). Such prob-
lem occurs for instance in computer experiments, where a physical phenomenon
is studied with a complex numerical code ([15]). One important question is
screening, i.e. to identify unessential variables, which can be done by comput-
ing several criteria, called sensitivity indices. Among them, the variance-based
indices – or Sobol indices, are often preferred by practitioners due to their easy
interpretation ([21]). They are defined on the Sobol-Hoeffding decomposition
([20], [17], [33]),
  
f (x) = fI (xI ) = f0 + fi (xi ) + fi,j (xi , xj ) + . . . ,
I⊆{1,...,d} 1≤i≤d 1≤i<j≤d

where xI is the subvector extracted from x whose coordinates belong to I. The

terms fI satisfy non-simplification conditions E(fI (xI )|xJ ) = 0 for all strict
 Poincaré inequalities on intervals – application to sensitivity analysis 3083

subset J ⊂ I, equivalent to:


fI (xI )dμi (xi ) = 0 (1.1)

for all i ∈ I and all xI . Such conditions imply orthogonality, and lead to the
variance decomposition:

Var(f (x)) = Var(fI (xI )).
I⊆{1,...,d}

The “total Sobol index” SiT is then defined as the ratio of variance of f (x)
explained by xi (potentially with other variables):
 

T
Si = Var fI (xI ) /D, (1.2)
Ii

where D = Var(f (x)). Thus one can decide that xi is not influential if SiT is
less than (say) 5%.
Despite their nice interpretation, Sobol indices require numerous computa-
tions for their estimation. Then another global sensitivity index, called DGSM
(Derivative-based Global Sensitivity Measure), can advantageously be used as
a proxy [34, 23]. Defined by
  2
∂f (x)
νi = dμ(x) (1.3)
∂xi
for i = 1, . . . , d, they are cheaper to compute, especially when the gradient of
f is available (e.g. as output of an evaluation of a complex numerical code).
As shown in [24], Sobol indices and DGSM are connected by a 1-dimensional
Poincaré-type inequality:
SiT ≤ C(μi )νi /D, (1.4)
where C(μi ) is a Poincaré constant for μi . Recall that μi satisfies a Poincaré
inequality if the energy of any centered
 function is controlled by the energy of
its derivative: For all g satisfying gdμi = 0 there exists C(μi ) > 0 s.t.
 
g 2 dμi ≤ C(μi ) (g  )2 dμi . (1.5)

Inequality 1.4 is easily obtained by applying the Poincaré inequality 1.5 to the
functions xi → Ii fI (xI ), which are centered for any choice of xI with i ∈ I
(condition 1.1), and integrating with respect to the other variables xj (j = i).
This allows performing a “low-cost” screening based on the upper bound of
Sobol indices in 1.4 instead of Sobol indices directly. Hence, one can decide that
xi is not influential if C(μi ) νDi is less than (say) 5%.
The efficiency of this low-cost screening strongly depends on the accuracy
of the Poincaré-type inequality 1.4, and this motivates the investigation of 1-
dimensional Poincaré inequalities. Notice that there is a large variety of prob-
ability distributions used in sensitivity analysis. They are often linked to prior
3084 O. Roustant, F. Barthe and B. Iooss

knowledge about the range of variation of physical parameters. As a conse-

quence most of them are continuous with finite support, possibly obtained by
truncation. The most frequent probability density function (pdf) are: Uniform,
(truncated) Gaussian, triangular, (truncated) lognormal, (truncated) exponen-
tial, (truncated) Weibull, (truncated) Gumbel (see for example [15]). Less fre-
quently, it can be found: (inverse) Gamma, Beta, trapezoidal, generalized ex-
treme value.
It is also important to notice that one may not easily reduce the problem
to the case of the uniform distribution, by applying the standard reincreasing
arrangement technique. Indeed by denoting Fi the cdf of μi , an idea would be
to consider the function g(u1 , . . . , ud ) = f F1−1 (u1 ), . . . , Fd−1 (ud ) where the
ui = Fi (xi ) are independent and uniform on [0, 1]. Then the total Sobol indices
of f (x) and g(u) are equal. However, the derivatives of the transformations Fi−1
can be large, and the DGSM computed on g may be large and even infinite.
For instance, if f (x1 , x2 ) = x1 + x2 with x1 , x2 i.i.d. N (0, 1), the DGSM of
g(u1 , u2 ) = Φ−1 (u1 ) + Φ−1 (u2 ) is equal, for each variable ui , to:
 1  1  
1 1
[(Φ−1 ) (s)]2 ds =
2
νg = −1 (s))2
ds = dt = et /2 dt = +∞.
0 0 φ(Φ R φ(t) R

On the other hand, the DGSM νf of f is equal to 1 for each xi , and the corre-
ν
sponding upper bound is CP (N (0, 1)) Df = 12 , which is here exactly equal to the
total Sobol index of xi .
To conclude about motivations in sensitivity analysis, it is worth mentioning
that the idea of low-cost screening can be extended to higher-order interactions,
and also depends on the accuracy of 1-dimensional Poincaré inequalities [31]. It
allows screening out useless interactions and discovering additive structures in f .
Finally notice that measuring the influence of variables has useful applications
in other fields of probability. For instance, assuming a negligibility condition in
terms of variance ratios, [14] proves a central limit theorem for homogeneous
sums in Sobol-Hoeffding decomposition. Similarly, controlling the influence of
variables allows [25] to establish an invariance principle on the probability dis-
tribution of multilinear polynomials.

1.2. Aim and plan

Our aim is to bridge the gap between industrial needs coming from sensitivity
analysis problems and the theory of Poincaré inequalities, by providing an ac-
cessible introduction to Poincaré inequalities for non-specialists, and by using
and developing the theory in order to deal with the specific situations motivated
by low-cost screening.
In Section 2, we present the general background on Poincaré inequalities,
together with the main techniques available to establish them (Muckenhoupt’s
criterion, perturbation and transportation methods, spectral methods). Most
of these techniques provide upper bounds on the Poincaré constant for large
classes of measures.
 Poincaré inequalities on intervals – application to sensitivity analysis 3085

In the literature of 1-dimensional Poincaré inequalities, the exact value of the

Poincaré constant is known for very few measures, such as: Uniform, Gaussian
(which are classical, see e.g. [3], [5]), exponential [9] or logistic [6]. More measures
are needed for sensitivity analysis, as well as their restrictions to intervals. These
situations were not studied specifically so far (apart from the Gaussian pdf, for
which optimal constants are known on the whole real line and on intervals
formed by successive zeros of Hermite polynomials, see e.g. [13]). Section 3
provides further results which allow to deal with the family of restrictions of a
given probability measure to intervals. In the case of symmetric measures and
intervals, we derive new improvements.
Section 4 deals with the exact constant for some distributions required by
sensitivity analysis. We present new semi-analytical results for the optimal con-
stant of the triangular, truncated normal and truncated exponential. The opti-
mal constant is obtained as the first zero of an analytic function. We also provide
a general numerical algorithm that converges to the optimal Poincaré constant.
Finally we come back to our initial motivation and apply this research to two
test-cases (Section 5). It is observed that the new results significantly improve
on the existing ones, and now allow performing a low-cost screening of Sobol
indices based on upper bounds computed by DGSMs.

2. Background on Poincaré inequalities

In this section, we provide a quick survey of the main simple techniques allowing
to derive Poincaré inequalities for probability measures on the real line. We often
make regularity assumptions on the measures. This allows to avoid technicalities,
without reducing the scope for realistic applications. Indeed, all the measures
we are interested in are supported on an open interval (a, b), and have a positive
continuous, piecewise continuously differentiable density. Moreover, when a is
finite, they are monotonic on a neighborhood of a (and similarly for b).

2.1. Definitions

Consider an open interval of the real line Ω = (a, b) with −∞ ≤ a < b ≤ +∞.
A locally integrable function f : Ω → R is weakly differentiable if there exists a
locally integrable function g : Ω → R such that for all functions φ of class C ∞
with compact support in Ω:
 

f (t)φ (t)dt = − g(t)φ(t)dt.
Ω Ω
Then g is a.e. uniquely determined (more precisely two functions with this prop-
erty coincide almost everywhere), it is called the weak derivative of f and de-
noted by f  .
Let μ be a probability measure on Ω, and f : Ω → R be a Borel measurable
function. Recall that the variance of f for μ is defined as

Varμ (f ) = inf (f − a)2 dμ.
a∈R Ω
3086 O. Roustant, F. Barthe and B. Iooss
 
Obviously Varμ (f ) = +∞ if f 2 dμ = +∞. When f 2 dμ < +∞, it holds
  2    2
Varμ (f ) = f dμ −
2
f dμ = f − f dμ dμ.

Definition 1. Let μ(dt) = ρ(t)dt be an absolutely continuous probability mea-

sure Ω. We say that μ verifies a Poincaré inequality on Ω if there exists a
constant C < +∞ such that for all f weakly differentiable functions f on Ω:

Varμ (f ) ≤ C (f  )2 dμ. (2.1)
Ω

In this case, the smallest possible constant C above is denoted CP (μ), it is re-
ferred to as the Poincaré constant of the measure μ.
Observe that the above integrals are always defined, with values in [0, +∞].
Roughly speaking, a Poincaré inequality expresses in a quantitative way that
a function with a small weak derivative, measured in the sense of μ, has to be
close to a constant function again in the sense of μ.
Remark 1. Weakly differentiable functions are exactly the functions which ad-
mit (in their equivalence class for a.e. equality) a continuous version which
satisfies (see e.g. [2], [11])
 y
∀x, y ∈ Ω, f (y) = f (x) + f  (t) dt.
x

Such functions are also called (locally) absolutely continuous. Their variations
can be recovered by integrating their weak derivatives. It is therefore plain that
they provide a good setting for Poincaré inequalities. On the contrary, it is not
possible to work just with a.e. differentiable functions: for instance the famous
Cantor function (a.k.a. the Devil’s stairs) increases from 0 to 1 but is a.e.
differentiable with zero derivative. However, everywhere differentiable functions
with locally integrable derivative are weakly differentiable, see e.g. [32].
To be very precise, we should have denoted the Poincaré constant as CP (μ, Ω).
Indeed, we could also consider the probability measure μ on Ω as acting on any
larger open interval Ω . The restriction to Ω of weakly differentiable functions
on Ω are specific weakly differentiable function on Ω (they are continuous at
boundary points). Therefore, satisfying a Poincaré inequality on Ω is formally a
more demanding property. From now on, we will assume that Ω is the interior
of the convex hull of the support of μ, which is consistent with the notation
CP (μ).
Obviously, some measures cannot satisfy a Poincaré inequalities, for instance
the uniform measure on (0, 1) ∪ (2, 3) (one can choose a differentiable
 function
f on (0, 3) which is equal to 0 on (0, 1) and to 1 on (2, 3). Then (f  )2 dμ = 0).
Next let us present an equivalent definition of Poincaré inequalities, in a
more convenient analytic setting. Let L2 (μ) be the set of (equivalence classes
for a.e. equality of) measurable functions on Ω such that Ω f 2 dμ < +∞. We
 Poincaré inequalities on intervals – application to sensitivity analysis 3087

 1/2
denote . L2 (μ) , or simply . , its associated norm: f = Ω
f 2 dμ . The
first weighted Sobolev space Hμ1 (Ω) is defined by

Hμ1 (Ω) = {f ∈ L2 (μ) such that f  ∈ L2 (μ)},

where f  is the weak derivative of f . It is well known that Hμ1 (Ω) is an Hilbert
space with the norm f 2H1 (Ω) = f 2 + f  2 . More generally, for an integer
μ

 ≥ 1, the space Hμ (Ω) is defined by:

Hμ (Ω) = {f ∈ L2 (μ) such that for all k ≤ , f (k) ∈ L2 (μ)},

where f (k) is the k-th weak derivative of f . An important particular case is

when Ω is bounded and there exists two positive constants m, M such that

∀t ∈ Ω, 0 < m ≤ ρ(t) ≤ M.

Then if Leb denotes the Lebesgue measure on Ω, it holds that L2 (μ) = L2 (Leb)
and Hμ (Ω) = HLeb

(Ω) for all positive integers , and the norms on these spaces
are equivalent when μ is replaced by Leb. In other words, the weighted Sobolev
spaces are equivalent to the usual Sobolev spaces. This remark will allow us
to use several results which are available for Leb and a bounded Ω. Firstly,
Hμ1 (μ) then contains functions which are continuous on Ω and piecewise C 1 on
Ω. Secondly, the spectral theory of elliptic problems (see e.g. [2], Chap. 7) will
then be valid.
We now come back to the general case.
Definition 2. Let μ(dt) = ρ(t)dt be an absolutely continuous probability mea-
sure on an open interval Ω, such that ρ > 0 almost everywhere on Ω. We say
that μ admits a Poincaré inequality on  Ω if there exists a constant C < +∞
such that for all f in Hμ1 (Ω) verifying Ω f dμ = 0, we have:
 
f 2 dμ ≤ C (f  )2 dμ. (2.2)
Ω Ω

The best possible constant C is denoted CP (μ). If there exists fopt , a centered
function of Hμ1 (Ω), such that (2.2) is an equality for C = CP (μ), we say that
the inequality is saturated by fopt .

This definition means Varμ (f ) ≤ C(μ) Ω (f  )2 dμ for weakly differentiable
functions f such that f and f  are square integrable for μ. Hence Definition 1
is formally stronger as it involves general weakly differentiable functions. Nev-
ertheless, the Poincaré inequality of Definition 2 implies the one of Definition 1:
first it is enough to consider functions f with square integrable weak derivative
f  , then one can apply the Poincaré inequality to certain truncations of f , which
belong to Hμ1 (Ω), in order to show that f is also necessarily square integrable
for μ.
Poincaré inequalities are often stated in terms of the ratio of energies of a
function and its derivative:
3088 O. Roustant, F. Barthe and B. Iooss

Definition 3. Let f in Hμ1 (Ω) with f 2 dμ > 0. The Rayleigh ratio of f is:
 2
f 2 f dμ
J(f ) = 2
= Ω 2 . (2.3)
f Ω
f dμ

Thus μ admits a Poincaré inequality if and only if the Rayleigh ratio admits
 lower bound over the subspace of centered functions Hμ (Ω) = {f ∈
1,c
a positive
Hμ (Ω), Ω f dμ = 0}. In that case,
1

 −1
CP (μ) = inf
1,c
J(f ) = sup J(f )−1 .
f ∈Hμ (Ω)−{0} 1,c
f ∈Hμ (Ω)−{0}

The expression of CP (μ) as a supremum makes it easy to compute lower bounds,

by choosing an appropriate test function f . An example is given below:

Proposition 1. CP (μ) ≥ Varμ , with equality if μ = N (0, 1).

Proof. The lower bound is obtained for f (x) = x − Eμ . The equality case is
well-known (see e.g. [3], [5]).

2.2. A general criterion

The class of probability measures that admit a Poincaré inequality has been
completely characterized in the work of Muckenhoupt [26]. See also [8] for re-
finements.

Theorem 1 ([26]). Let μ(dt) = ρ(t)dt be a probability measure on Ω = (a, b).

Let m be a median of μ and define:
 m  x
1 1
A− = sup μ (a, x) dt, A+ = sup μ (x, b) dt,
a<x<m x ρ(t) m<x<b m ρ(t)

with the convention 0 · ∞ = 0. Then μ admits a Poincaré inequality iff A− and

A+ are finite, and in this case

1
max(A− , A+ ) ≤ CP (μ) ≤ 4 max(A− , A+ ). (2.4)
2

This result explains how to estimate, up to a multiplicative factor, the

Poincaré constant by a simpler quantity. As explained in [7], it can be inter-
preted as a reduction from functions to sets (by decomposition functions ac-
cording to their level sets): a Poincaré inequality is equivalent to a comparison
between the measures of sets, and a notion of μ-capacity. In one dimension, a
further reduction allows to restrict to increasing function vanishing at m, for
which level sets for positive values are of the form (x, b).
 Poincaré inequalities on intervals – application to sensitivity analysis 3089

2.3. Perturbation and transport

Given a measure verifying a Poincaré inequality, one may try to deform it into
another measure still having a finite Poincaré constant. Many such results exist
in the literature. Here we mention two fundamental ones: the bounded pertur-
bation principle, and the Lipschitz transportation principle. They hold in very
general settings, but for the purpose of this article, it is enough to state them
on R.
Lemma 1. Let μ be a probability measure on Ω ⊂ R. Let ψ : Ω → R be
a bounded measurable function and let μ̃ be the probability measure given by
dμ̃ = eψ dμ/Z where Z is the normalizing constant. Then
CP (μ̃) ≤ esup ψ−inf ψ CP (μ).
This fact is easily proved at the level of Poincaré inequalities by using Varμ̃ =
inf a∈R (f −a)2 dμ̃ and applying obvious bounds on ψ. See e.g. Proposition 4.2.7
of [5] and related comments.
Lemma 2. Let μ be an absolutely continuous probability measure on an open
interval Ω ⊂ R. Let T : Ω → R be a Lipschitz map, (i.e. verifying that there
exists L ∈ R such that for all x, y ∈ Ω, |T (x) − T (y)| ≤ L|x − y|). Denote by T μ
the image measure of μ by T , i.e. the measure on R such that for every Borel
set B, T μ(B) = μ(T −1 (B)). Assume that T μ is absolutely continuous. Then
CP (T μ) ≤ T 2
Lip CP (μ),

where T Lip is the smallest possible value of the constant L above.

Proof. It is convenient here to work with locally Lipschitz functions (i.e. func-
tions which are Lipschitz on any compact interval). They are weakly differen-
tiable with locally bounded derivative (see e.g. [11]). However one can also define
pointwise and as a whole their “absolute value of the derivative” as
|f (y) − f (x)|
|f  |(x) := lim sup .
y→x |y − x|
By Rademacher’s theorem, a locally Lipschitz function is differentiable almost
everywhere, and the latter coincides a.e. with the absolute value of the derivative
of f  (x). By hypothesis, |f  |(x) = |f  (x)| will hold almost surely in the sense of
μ and T μ as well.
Set ν = T μ. By a density argument it is enough to prove the Poincaré
inequality for ν for all locally Lipschitz functions. Let g be a locally Lipschitz
function  on R. Then f := g ◦ T is also locally Lipschitz on Ω, and Varμ (f ) ≤
CP (μ) |f  |2 dμ. Since ν is the image measure of μ by T , Varμ (f ) = Varν (g).
Observe that f = g ◦ T is also locally Lipschitz and verifies |f  |(x) ≤ L|g  |(T (x))
for all x. Consequently
  
|f  |2 dμ ≤ L2 (|g  |(T (x)))2 dμ(x) = L2 |g  |2 dν.

Hence we have proved that Varν (g) ≤ CP (μ)L2 |g  |2 dν. The result follows.
3090 O. Roustant, F. Barthe and B. Iooss

In the case of probability measures μ, ν on the real line, and when μ has
no atoms, a natural map which pushes μ forward to ν is the monotonic map
T := Fν−1 ◦ Fμ (here Fν−1 stands for the generalized left inverse). It remains to
estimate the Lipschitz norm of T .
Lemma 3. Let μ and ν be probability measures on R. Assume that μ(dt) =
ρμ (t)dt where ρμ is positive and continuous on (aμ , bμ ) (−∞ ≤ aμ , bμ ≤ +∞)
and vanishes outside. Let us make the same structural assumption for ν. Then
T := Fν−1 ◦ Fμ is well defined and differentiable on (aμ , bμ ), with T  = ρμ /ρν ◦
Fν−1 ◦ Fμ . Consequently its Lipschitz norm is

ρμ ρμ ◦ Fμ−1 ρμ ◦ Fμ−1 ◦ Fν
T Lip = sup −1 = sup −1 = sup .
(aμ ,bμ ) ρν ◦ Fν ◦ Fμ (0,1) ρν ◦ Fν (aν ,bν ) ρν

The previous two lemmas can be applied when μ(dx) = exp(−|x|)dx/2, the
double exponential measure. In this case ρν ◦Fν−1 (t) = min(t, 1−t) and CP (μ) =
4. This is how Bobkov and Houdré [10] established the following estimate:
 2
min(Fν (x), 1 − Fν (x))
CP (ν) ≤ 4 sup . (2.5)
x∈(aν ,bν ) ρν (x)

These authors also deduced from this approach that for log-concave probability
measures ν on R, with median m, CP (ν) ≤ 1/ρν (m)2 .
Actually, we can improve on (2.5) by choosing the logistic measure μ(dx) =
ex
(1+ex )2 dx instead of the double exponential measure in the proof. Indeed, in
this case ρμ ◦ Fμ−1 (t) = t(1 − t) is smaller than min(t, 1 − t), while the Poincaré
constant is the same CP (μ) = 4, as shown in [6]. Consequently
Corollary 1. Let ν be a probability measure on (a, b) ⊂ R, with a positive
continuous density on (a, b). Then
 2
Fν (x)(1 − Fν (x))
CP (ν) ≤ 4 sup .
x∈(a,b) ρν (x)

Remark 2. If we apply the previous two lemmas with μ being the uniform
probability measure on [− 12 ; 12 ], for which CP (μ) = π −2 (see Table 1 below), we
get that
 2
1
CP (ν) ≤ .
π inf x∈(a,b) ρν (x)
This inequality is meaningful only when the density of ν is bounded from below by
a positive constant (which implies in particular that ν is compactly supported).
This is related to the fact (put forward in the introduction) that rewriting sen-
sitivity analysis questions in terms of uniform variables may not provide good
results for our purposes. The present section shows that rewriting the problem
in terms of logistic variables is more effective in general.
 Poincaré inequalities on intervals – application to sensitivity analysis 3091

2.4. Spectral interpretation

When μ has a continuous density that does not vanish on a compact interval
[a, b], then μ admits a Poincaré constant on [a, b], as follows from the Muck-
enhoupt condition (Theorem 1) or from the bounded perturbation principle
(Lemma 1). Furthermore, modulo a small additional regularity condition, the
Poincaré inequality is saturated and the Poincaré constant is related to a spec-
tral problem, as detailed now.
Theorem 2. Let Ω = (a, b) be a bounded open interval of the real line, and
assume that μ(dt) = ρ(t)dt = e−V (t) dt where V is continuous and piecewise C 1
on Ω = [a, b]. Consider the three following problems:
 2 
(P1) Find f ∈ Hμ1 (Ω) s.t. J(f ) = ff 2 is minimum under f dμ = 0.
 
(P2) Find f ∈ Hμ (Ω) s.t. f , g  = λf, g ∀g ∈ Hμ (Ω).
1 1

(P3) Find f ∈ Hμ2 (Ω) s.t. f  − V  f  = −λf and f  (a) = f  (b) = 0.

Then the eigenvalue problems (P2) and (P3) are equivalent, and their eigen-
values form an increasing sequence (λk )k≥0 of non-negative real numbers that
tends to infinity. Moreover 0 is an eigenvalue, and all eigenvalues are simple.
The eigenvectors (uk )k≥0 form a Hilbert basis of L2 (μ) and verify:
 b
λk
uk (x) = uk (t)ρ(t)dt. (2.6)
ρ(x) x

In particular, up to a modification with Lebesgue measure zero, the uk ’s are C 1

on [a, b] and if ρ is of class C k , uk is of class C k+1 .
Furthermore when λ = λ1 , the first positive eigenvalue, (P 2) and (P 3) are
equivalent to (P 1) and the minimum of (P 1) is attained for f = u1 . In other
words, the optimal Poincaré constant is CP (μ) = 1/λ1 and the inequality is
saturated by a non-zero solution of (P3). Finally, u1 is strictly monotonic.
Although the equivalence between (P1) and (P3) is well-known, it was not
easy to find a self-contained proof. Several arguments can be found in the lit-
erature of functional inequalities, and especially in [8] for mixed Dirichlet and
Neumann boundary conditions f (a) = f  (b) = 0. We also found complementary
arguments in the literature of numerical analysis, for instance in [13], for the
uniform distribution. Since neither of them are fully corresponding to our set-
ting, we have inserted below a proof of Theorem 2, making a synthesis of the
ideas of the two literature fields. In particular, we adopt the Hilbertian point of
view chosen in [13], which is central for estimating the optimal Poincaré con-
stant, as seen in Section 4.2, and adapt the short and elementary proofs of [8]
coming under more general principles (maximum principle, regularity theorem).
Proof of Theorem 2. Since Ω is bounded an ρ is continuous on Ω, there exists
two positive real numbers m, M such that:

∀t ∈ Ω, 0 < m ≤ ρ(t) ≤ M.
3092 O. Roustant, F. Barthe and B. Iooss

Then, as mentioned in § 2.1, L2 (μ) and all weighted Sobolev spaces Hμ (Ω) are
topologically unchanged when we replace μ by the Lebesgue measure on Ω. This
allows using the spectral theory of elliptic problems (see e.g. [2], Chap. 7) on
usual Sobolev spaces with bounded Ω.
More precisely, let us consider problem (P2), and denote H = L2 (μ) and
V = Hμ1 (Ω). Then the injection V ⊂ H is compact, and V is dense in H. (P2)
can be written as an eigenvalue problem of the form:
a(f, g) = λf, g ∀g ∈ V, (2.7)
where a(f, g) = f  , g   and ., . denotes the scalar product on H. In order to
apply spectral theory, a should be coercive with respect to V, i.e. ∃C > 0 s.t.
a(f, f ) ≥ Cf, f H1μ (Ω) for all f , which is not true here. A convenient way to
overcome this issue is to consider the equivalent shifted problem (see [13], §8.2.):
α(f, g) = (λ + 1)f, g ∀g ∈ V, (2.8)
where α(f, g) = f  , g   + f, g is the usual scalar product on V. It is coercive
on V. Then we can apply Theorem 7.3.2 in [2]: The possible eigenvalues form
an increasing sequence (λk + 1)k≥0 of non-negative values that tends to infinity
and the eigenvectors (uk )k≥0 form a Hilbert basis of L2 (μ). Further, from (2.7)
with f = g, we have λk ≥ 0. Now remark that 0 is an eigenvalue and the
corresponding eigenspace is spanned by the constant function u0 = 1. Indeed,
b
taking g = f in (2.2) leads to a f  (x)2 e−V (x) dx = 0 and f is a constant function.
Now let us prove the equivalence between (P2) and (P3). We restrict the
presentation to λ > 0 since the case λ = 0 is direct, using a Sturm-Liouville
form of (P3), i.e. (f  V ) = −λf . Formally the link comes from an integration
by part of the left hand side of (P2):
 b  b
   
f g ρ = f (b)ρ(b)g(b) − f (a)ρ(a)g(a) − (f  ρ) g. (2.9)
a a
b
Since this must be equal to λ a f gρ for all g ∈ Hμ1 (Ω), we should have f  (a) =
f  (b) = 0 and (f  ρ) = −λf ρ which is problem (P3). Actually this method, read
in the reverse sense, shows that (P3) implies (P2). Indeed if f ∈ Hμ2 (Ω) then
f  ρ ∈ Hμ1 (Ω) and the integration by part is valid. However, to see that (P2)
implies (P3), an argument of regularity must be used since f is only assumed
to belong to Hμ1 (Ω). This is achieved by proving that, μ-almost surely,
 b
λ
f  (x) = f (t)ρ(t)dt. (2.10)
ρ(x) x
Indeed, this implies that f  ρ is C 1 on [a, b] and (2.9) is valid. Furthermore, (2.10)
also implies f  (a) = f  (b) = 0 and this gives (P3). Let us now come  xback to the
proof of (2.10). Following [8], Lemma 4.3., by using g(x) = g(a) + a g  (t)dt and
Fubini’s theorem, it holds:
 b  b  b  b 
f (x)g(x)ρ(x)dx = g(a) f (x)ρ(x)dx + f (t)ρ(t)dt g  (x)dx.
a a a x
 Poincaré inequalities on intervals – application to sensitivity analysis 3093

b
Recall that for λ > 0, a
f ρ = 0. Thus (P2) can be rewritten:
 b    b

b
f  (x)g  (x)ρ(x)dx = λ f (t)ρ(t)dt g  (x)dx.
a a x

Since it is true for all g  in L2 (μ) this gives (2.10).

Now let us prove that all eigenvalues are simple. Notice that, from (2.10), all
solutions of (P3) are locally Lipschitz, up to a modification of Lebesgue measure
0. Thus when λ is an eigenvalue, the vector space of solutions of the second order
linear differential equation f  − V  f  = −λf has dimension 2. Adding Neumann
conditions, the dimension is equal to 1. Indeed, let f and g be two solutions of
(P 3). Then f (b) = 0 otherwise f would be identically null by uniqueness of the
Cauchy problem (since f  (b) = 0). Hence there exists a real number c such that
g(b) = cf (b). Then consider h = g − cf . We have h(b) = h (b) = 0 implying, by
uniqueness of the Cauchy problem, that h = 0, i.e. g = cf .
In this Hilbertian framework, the  equivalence between (P2) and (P1) is easily
seen. Indeed, if f ∈ Hμ1 (Ω) verifies f dμ = 0 then f is orthogonal to 1 and hence
+∞
f = k=1 fk uk . Then, it holds:

f 2 +∞ 2
α(f, f ) k=1 λk fk
J(f ) = 2
= −1= +∞ 2 ≥ λ1 ,
f f 2 k=1 fk

with equality iff f is proportional to u1 . This also proves that the minimum in

(P1) is attained precisely on the 1-dimensional space spanned by u1 .
It remains to show that f , a minimizer of the Rayleigh ratio, is strictly
monotonic. The main arguments can be found e.g. in [8]. Define
 x
g(x) = |f  |dμ.
a


 
= |f | and thus (g  )2 dμ = (f  )2 dμ. Secondly, using the
Firstly, we have g  

formula Varμ g = 12 [g(y) − g(x)]2 dμ(x)dμ(y), we have:

   y 2
1 
Varμ g = |f (t)|dμ(t) dμ(x)dμ(y)
2 x
  y 2
1
≥ f  (t)dμ(t) dμ(x)dμ(y) = Varμ f.
2 x

Since f is a minimizer  y of the Rayleigh ratio,

y the above inequality must be an
equality, leading to x |f  (t)|dμ(t) = | x f  (t)dμ(t)| for all x, y in [a, b]. By
continuity of f  , there exists ∈ {−1, +1} ( = sgn[f (b) − f (a)]) such that
|f  | = f  everywhere, hence f is monotonic.
To see that f is strictly monotonic, consider again (2.10). Assume for in-
stance that f is increasing (f  ≥ 0), and let us prove that f is strictly increasing
(f  > 0). Then, as it is centered, the increasing function f is first ≤ 0 and then
≥ 0. From (2.10), it implies that f  is first increasing from f  (a) = 0 and then
3094 O. Roustant, F. Barthe and B. Iooss

decreasing to f  (b) = 0. Furthermore, as f is increasing and centered, we must

have f (a) < 0 (otherwise f ≥ 0 on [a, b] and hence identically zero). By conti-
nuity, f < 0 in a neighborhood of a. Hence from (2.10) f  is strictly increasing
in a neighborhood of a. Similarly f  is strictly decreasing in a neighborhood of
b. Finally f  > 0 on (a, b).
For completeness, we sketch the standard way to derive (P2) from (P1) by
using calculus of variation. Without loss of generality we assume that f 2 dμ =
1. (P1) is then equivalent to:
 b  b  b
1  2 2
min (f ) dμ s.t. f dμ = 1, f dμ = 0.
f a 2 a a

Define L(t, f, f  ) = ( 12 f 2 − 12 λ(f 2 − 1) − βf )e−V (t) . The Lagrangian is:

 b
J(f, λ, β) = L(t, f (t), f  (t))dt.
a

Considering a small deviation function g, we have

J(f + g, λ, β) = J(f, λ, β) +
 b 
∂L   ∂L 
g(t) (t, f (t), f (t)) + g (t) (t, f (t), f (t)) dt + O( 2 ).
a ∂f ∂f
The first order condition leads to cancel the integral, which gives here:
 b  b
(−λf g + f  g  ) dμ = β gdμ.
a a
b
Now, with g = 1 and using a
f dμ = 0, we get β = 0. This gives (P2).
Remark 3. Another solution to make coercive the bilinear form a(f, g) =
f  , g   is to project onto the space of centered functions Hμ1,c (Ω), as done in
[13]. Indeed, on that space, coercivity is equivalent to the existence of a Poincaré
constant, which can be proved independently by the bounded perturbation prin-
ciple as mentioned above. However, this seems a less convenient setting for the
numerical method developed in Section 4.2 since the ‘hat’ functions used do not
belong to Hμ1,c (Ω).
Definition 4 (Spectral gap). The smallest strictly positive eigenvalue λ1 of
Lf = f  − V  f  with Neumann boundary conditions f  (a) = f  (b) = 0 is called
Neumann spectral gap, or simply spectral gap, and denoted by λ(μ).
We conclude this section by a proposition showing that, under regularity
conditions on V , the Neumann problem (P3) can be written as a Dirichlet
problem.
Proposition 2. If V is of class C 2 , then (P3) is equivalent to:
(P4) Find h ∈ Hμ2 (Ω) s.t. h − V  h − V  h = −λh and h(a) = h(b) = 0.
 Poincaré inequalities on intervals – application to sensitivity analysis 3095

Proof. Define L and K the operators

Lf = f  − V  f  , Kf = f  − V  f  − V  f.
Let f be a solution of (P3). Since V is of class C 2 then by Theorem 2, f ∈ H3 (μ).
Now define h = f  . We have: K(f  ) = (Lf ) = −λf  and h is solution of (P4).
Conversely,
 let h be a solution of (P4) and let f be the primitive function of
h such that f dμ = 0. Then (Lf ) = −λf  and there exists c ∈ R such that
b b
Lf = −λf + c. Remarking that a L(f )e−V = a (f  e−V ) = 0, the centering
condition implies c = 0 which proves that f is a solution of (P3).

2.5. More techniques

We conclude this quick overview by briefly quoting two other methods.

The first one plays a major role in the study of multidimensional functional
inequalities. It has its roots in early works by Hörmander and Brascamp-Lieb,
and was developed by Bakry and his collaborators who introduced and exploited
a notion of curvature and dimension for diffusion generators, see e.g. [3], [5]. The
simplest instance of these results is as follows:
Theorem 3. Let μ be a probability measure on an open interval Ω with dμ(x) =
e−V (x) 1Ω (x) dx where V is twice continuously differentiable. If there exists r > 0
such that for all x ∈ Ω, V  (x) ≥ r then CP (μ) ≤ 1/r.
The estimate is exact for the standard Gaussian measure.
The second result is known as Chen’s variational formula [12]. Its main ad-
vantage is to express the Poincaré constant as an infimum (while, as already
mentioned, its definition is given as a supremum). This allows to obtain upper
bounds rather easily. We state it in a different form, which appeared in [16].
Theorem 4. Let dμ(x) = e−V (x) 1[a,b] (x)dx be a probability measure on a seg-
ment, with V twice continuously differentiable on [a, b]. Then
1 (−Lg)
λ(μ) = = sup inf ,
CP (μ) g >0 (a,b) g 
where Lf = f  − V  f  and the supremum is taken over three times differentiable
functions on (a, b), having positive first derivative.
As an example of application, Chen’s formula recovers the following classical
statement, which often allows to calculate the spectral gap. It is the recipro-
cal of the fact that a solution of the spectral problem (P3) corresponding to
the smallest positive eigenvalue is necessarily strictly monotonic (one part of
Theorem 2).
Corollary 2. Under the framework and notations of Section 2.4, if f is a
strictly monotonic solution of problem (P3)
Lf = −λf, f  (a) = f  (b) = 0
for some λ > 0, then λ is the spectral gap.
The same is true for problem (P4) replacing ‘monotonic’ by ‘non vanishing’.
3096 O. Roustant, F. Barthe and B. Iooss

Proof. By definition of the spectral gap, λ(μ) ≤ λ. Conversely, up to a change

of sign we can assume that f  > 0 on (a, b). Then Chen’s formula immediately
)
gives λ(μ) ≥ inf (a,b) (−Lf
f = λ.
The last part comes from the equivalence between (P3) and (P4): Ifh is a non
vanishing solution of (P4), then the primitive function of h such that f dμ = 0
is a strictly monotonic solution of (P3) (see the proof of Prop. 2).

3. Poincaré inequalities on intervals

This section gathers tools for the study of the Poincaré constant of the restric-
tions of a given measure to subintervals.

3.1. General results

The Poincaré constant cannot increase by restriction to a subinterval:

Lemma 4. Let μ be a probability measure on R, of the form dμ(t) = 1Ω (t)ρ(t) dt
where Ω is an open interval on which ρ is continuous and positive. Let I ⊂ Ω be
an open interval, and μ|I denote the probability measure obtained by restriction
of μ to I (defined by μ|I (A) = μ(I ∩ A)/μ(I)). Then
CP (μ|I ) ≤ CP (μ).
Consequently, the map defined on intervals by I → CP (μ|I ) is non-decreasing.
Proof. Let f ∈ H1 (μ|I ). Set Ω = (α, β) and I = (α , β  ). If α < α then by
hypothesis, ρ is bounded from below by a positive constant in a neighborhood
of α , which ensures that f has a limit at α , and can be continuously extended
by a constant on (α, α ]. Indeed f verifies the functional Cauchy criterion thanks
to the following bound: for ε > 0 small enough and for all x, y ∈ (α , α + ε),
 y  y  12  y  12
1
|f (y) − f (x)| = f  (t) dt ≤ f  (t)2 ρ(t)dt dt
x x x ρ(t)
  12   12
 2 1
≤ (f ) dμ sup |y − x|.
I (α ,α +ε) ρ

The same applies if β  < β. This allows to build a weakly differentiable function
f˜ on Ω which coincides with f on I and has zero derivative outside I. The result
easily follows:
 
dμ dμ
Varμ|I (f ) = inf (f − a)2 ≤ inf (f˜ − a)2
a μ(I) a μ(I)
I
 Ω

 2 dμ dμ
≤ CP (μ) (f )˜ = CP (μ) (f  )2
μ(I) μ(I)
Ω I

= CP (μ) (f  )2 dμ|I .
I
 Poincaré inequalities on intervals – application to sensitivity analysis 3097

The previous result helps proving a continuity type property with respect to
the support:
Lemma 5. Consider again the framework of Lemma 4, and let I be a family
and increasing subintervals such that I ↑ Ω when → 0. Typically, I = (a +
, b − ) for Ω = (a, b), I = (a, a + 1/ ) for Ω = (a, +∞) and so on. Then the
Poincaré constant on Ω is the limit of the Poincaré constant on subintervals:
CP (μ) = lim CP (μ|I ).
→0

Proof. Firstly, by Lemma 4, the sequence CP (μ|I ) is increasing and bounded

by CP (μ). Hence CP (μ) ≥ sup CP (μ|I ) = lim CP (μ|I ).
>0 →0
Conversely, let f ∈ Hμ1 (Ω). Observe that, by Lebesgue’s theorem,

μ(I ) = ρ(t)dt → μ(Ω) = 1,
I →0

and for all μ-integrable functions g:

  
ρ(t)
Eμ (g) = g(t)ρ(t)dt = lim g(t)ρ(t)dt = lim g(t) dt = lim Eμ|I (g).
Ω →0 I →0 I μ(I ) →0

Applying this result to f , f 2 and f 2 , it holds:

Var (f ) Varμ|I (f )
 μ = lim  ≤ lim CP (μ|I ).
2
f dμ →0 f 2 dμ|I →0
Ω I

Since it is true for all f ∈ Hμ1 (Ω), we get CP (μ) ≤ lim CP (μ|I ).
→0

3.2. Improvements for symmetric settings

In general, the Poincaré constant of a restriction to an interval is at most the

one of the initial measure. A better result is available in symmetric settings:
Lemma 6. Let μ be a probability measure on R. Assume μ(dx) = ρμ (x)dx
where ρμ is a unimodal function with a maximum at x = 0. Let I ⊂ R be a non-
empty interval and ν := μ(· ∩ I)/μ(I). Assume in addition that μ(R− ) = ν(R− ).
Then the monotonic transport from μ to ν is μ(I)-Lipschitz. Consequently
CP (ν) ≤ μ(I)2 CP (μ).
Proof. By hypothesis, Fν (0) = Fμ (0). Hence T (0) = 0 and since T is increasing,
T (x) and x have the same sign.
We claim that for all x (in the interior of the support of μ), |T (x)| ≤ |x|.
Indeed for x ≥ 0, this inequality amounts to Fμ (x) ≤ Fν (x). The latter is
obvious when x > sup(I) since in this case Fν (x) = 1. For 0 ≤ x ≤ sup(I), the
claim boils down to
 x  x
− − ρμ
μ(R ) + ρμ ≤ ν(R ) + ,
0 0 μ(I)
3098 O. Roustant, F. Barthe and B. Iooss

which is obvious using μ(R− ) = ν(R− ).

For x ≤ 0, the claim amounts to x ≤ T (x), that is Fν (x) ≤ Fμ (x). This is
obvious if x < inf(I). If inf(I) ≤ x ≤ 0, the inequality amounts to
 0  0
− ρμ
ν(R ) − ≤ ν(R− ) − ρμ ,
x μ(I) x

which is easily verified.

Eventually, for x in the interior of the support of μ,

ρμ (x) ρμ (x)
T  (x) = = μ(I) ≤ μ(I),
ρν (T (x)) ρμ (T (x))

since on R+ , 0 ≤ T (x) ≤ x and ρμ is non-increasing, and on R− , x ≤ T (x) ≤ 0

and ρμ is non-decreasing.
Our goal here is to get finer results when the underlying measure is even. A
first natural observation is that extremal (or almost extremal) functions for the
Poincaré inequality can be found among odd functions.
Lemma 7. Let μ be an even probability measure on R. Assume that it is sup-
ported on a symmetric interval, and is absolutely continuous, with a positive
continuous density in the interior of this interval. Then for any non constant
function f ∈ H1 (μ), there exists an odd function g ∈ H1 (μ) with
  2   2
(f ) dμ (g ) dμ
≥ ·
Varμ (f ) Varμ (g)

Proof. Let (−b, b) denote the interior of the support of μ. By hypothesis, f (0)
is well defined and by the variational definition of the variance
   2 0  2 b
(f  )2 dμ (f ) dμ −b
(f ) dμ + 0 (f  )2 dμ
≥ = 0 b .
Varμ (f ) (f − f (0))2 dμ (f − f (0))2 dμ + 0 (f − f (0))2 dμ
−b

Using the inequality u+vx+y ≥ min( x , y ), which is valid for non-negative numbers
u v

u, v, x, y such that x + y > 0 (and with the convention that v/0 = +∞), we get
that  0  2 b  2 
  2
(f ) dμ −b
(f ) dμ (f ) dμ
≥ min  0 , b 0 .
Varμ (f ) (f − f (0))2 dμ
−b (f − f (0))2 dμ
0

In order to conclude, we consider the odd functions g, h defined on (−b, b), such
that for all x ∈ [0, b), g(x) = f (x) − f (0), and for all x ∈ (−b, 0], h(x) =
f (x) − f (0). By symmetry of μ,
0  b 
−b
(f  )2 dμ (h )2 dμ (f  )2 dμ (g  )2 dμ
0 = , and b
0
= ·
−b
(f − f (0))2 dμ Varμ (h)
0
(f − f (0))2 dμ Varμ (g)
 Poincaré inequalities on intervals – application to sensitivity analysis 3099

Next, we derive from the above proposition a perturbation principle for the
Neumann spectral gap. It should be compared with the “bounded perturbation
principle” that we recalled in the previous section.
Proposition 3. Let dμ(x) = 1(−b,b) e−V (x) dx be an even probability measure,
with continuous potential V . Let ϕ : (−b, b) → R+ be an even function. Assume
that ϕ is non-increasing on [0, b) and that μ̃ defined by dμ̃ = ϕ dμ is a probability
measure. Then
CP (μ̃) ≤ CP (μ).

Proof. By the previous lemma, it is enough to establish a Poincaré inequality

for μ̃ for odd functions only. Let f be an odd weakly differentiable function. By
Lemma 4, for any a ∈ (0, b), CP (μ|(−a,a) ) ≤ CP (μ). Since f is odd  2
and μ is even,
(−a,a)
f dμ = 0, and the inequality Var μ |(−a,a)
(f ) ≤ C P (μ) (f ) dμ|(−a,a) can
be rewritten as
 
f 1(−a,a) dμ ≤ CP (μ) (f  )2 1(−a,a) dμ.
2

Next, for all x ∈ (−b, b),

 +∞  +∞
ϕ(x) = 1t≤ϕ(x) dt = 1It (x) dt,
0 0

where It = ϕ−1 [t, +∞) . Our hypotheses on ϕ ensure that It is empty, or is a

symmetric interval included in (−b, b). Hence, if it is not empty, it is equal, up
to a negligible set to (−a, a) for some a. Therefore we know that
 
f 2 1It dμ ≤ CP (μ) (f  )2 1It dμ,


is true for all t ≥ 0. But, by definition, pointwise, ϕ = R+ 1It . Thus integrating
the latter inequality over t yields
 
f 2 ϕ dμ ≤ CP (μ) (f  )2 ϕ dμ.


This can be rewritten as Varμ̃ f ≤ CP (μ) (f  )2 dμ̃.

As an example of application, using V ≡ 0 and φ(x) = e−x /2 , one can

2

bound the Poincaré constant of the Gaussian measure restricted to [−b, b] by

2
the one of the uniform measure on that interval: CP (N (0, 1)|[−b, b]) ≤ 4b
π 2 . The
bound is quite good when b tends to 0 since the restriction of the Gaussian
density to a small interval is almost constant then. For larger values the bound
is less interesting and Lemma 6 provides a much better result, always below 1,
consistent with the fact that the Poincaré constant cannot increase by restriction
to a subinterval.
3100 O. Roustant, F. Barthe and B. Iooss

4. Exact values of Poincaré constants

In this section, we show how to derive a Poincaré constant on an interval Ω either

as a first zero of a function expressed analytically (semi-analytical approach), or
by a full numerical method based on finite elements. We consider the setting of
the spectral theorem (Thm. 2): Ω bounded and ρ > 0 on Ω (with ρ continuous
on Ω and piecewise C 1 on Ω). The general case can be derived from it by the
continuity type property of Lemma 5. Hence, if Ω is unbounded or ρ vanishes at
the boundary of Ω, then the Poincaré constant is the limit of Poincaré constants
on bounded subintervals on which ρ > 0.

4.1. Semi-analytical results

This section presents some examples where the spectral gap is obtained in a
closed-form expression or by numerically searching a zero of a function ex-
pressed analytically. They are summarized in Table 1. The most famous one
is about the uniform distribution on [a, b], corresponding to a spectral prob-
lem for the Laplacian operator, f  = −λf . Another well-known result is about
the truncated normal, for the intervals formed by consecutive zeros of Hermite
polynomials. After recalling this result, we show how to extend it for a general
interval and develop the case of truncated (double) exponential and triangular
distributions. Without loss of generality, we will consider only one value for the
parameters, remarking that the general result is obtained by rescaling (Lemma
2 applied to T and T −1 when T is an affine function).

Prob. Dist. μ Support

 1 1 CP (μ) Form of fopt (x)
Uniform −2, 2 1/π 2 sin(πx)
N (0, 1) R 1 x
[rn,i , rn,i+1 ] 1/(n + 1) Hn+1 (x)
[a, b] see Prop. 8 related to Kummer’s functions
Double exp. R 4 ×
   2  −1  
[a, b] ⊆ R+ 1
4
+ π
b−a
e x/2 cos b−a π
x+φ
General case See Prop. 5
Triangular [−1, 1] r1−2 ≈ 0.1729 sign(x)J0 (r1 (1 − |x|))
Table 1
Examples of analytical results of the spectral problem with Neumann conditions. We have
used the following notations: Hn is the Hermite polynomial of degree n and rn,i its ith zero,
J0 is the Bessel function of the first kind with ν = 0, and r1 is its first root.

In all these examples, solutions of the second-order differential equation can

be expressed analytically as a linear combination of odd and even known func-
tions. The whole spectrum then corresponds to the zeros of one determinant
obtained with boundary conditions, as detailed in the following proposition.
Proposition 4. Consider the spectral problem with Neumann conditions (P3),
u − V  u = −λu. Let fλ , gλ a basis of solutions of the second-order differential
 Poincaré inequalities on intervals – application to sensitivity analysis 3101

equation u − V  u = −λu. Then the spectral gap on [a, b] is the first zero of
  
fλ (a) gλ (a)
d(λ) = det = fλ (a)gλ (b) − fλ (b)gλ (a).
fλ (b) gλ (b)

Furthermore if V is even and a = −b, the spectral gap is the first zero of λ →
gλ (b).
Remark 4. This result is immediately adapted to the equivalent spectral problem
with Dirichlet conditions (P4), u −V  u −V  u = −λu. In that case, with similar
notations, the spectral gap is the first zero of
 
fλ (a) gλ (a)
d(λ) = det = fλ (a)gλ (b) − fλ (b)gλ (a)
fλ (b) gλ (b)

and if a = −b with even V , it is the first zero of λ → fλ (b).

Proof of Proposition 4. If the interval is symmetric and V even, then by
Lemma 7 a solution can be found among odd functions, and we can assume
that u = gλ . The Neumann conditions are then equivalent to gλ (b) = 0, which
proves the result.
In the general case, let u be a non-zero solution of the spectral problem. Then
u can be written as u = Af + Bg, for some A,B ∈R. The boundary   conditions 
A fλ (a) gλ (a)
lead to the linear system Xc = 0, with c = and X = .
B fλ (b) gλ (b)
Now there exists a non zero solution if and only if the system is degenerated,
i.e. d(λ) = 0. This gives the whole spectrum and, in particular, the spectral gap
is the first zero.

Truncated exponential distribution

Proposition 5. Let μ be the (double) exponential distribution on the real line,

μ(dt) = 12 e−|t| dt, and let a < b. The Poincaré constant of the restriction of μ
−1
to [a, b] is CP (μ|[a,b] ) = 14 + ω 2 , where:
• If a, b have the same sign, ω = π
|b−a| . The inequality is saturated by f (x) =
|x|/2
e (−ω cos(ω(x − a)) + sin(ω(x − a) .
1
2
• If a < 0 < b, then ω > b−a π
. More precisely ω is the zero of x →
cotan(|a|x) + cotan(|b|x) + 1/x in ]0, min(π/|a|, π/|b|)[. The inequality is
saturated by
   
sgn(x) + 1 A
f (x) = e|x|/2 A cos(ωx) + B − sin(ωx) ,
2 ω

with A = ωcotan(ωa) − 12 and B = 12 cotan(ωa) + ω.

In particular if a = −b then ω is the zero of x → 2cotan(bx) + 1/x in
]0, π/b[, and the inequality is saturated by f (x) = e|x|/2 sin(ωx).
3102 O. Roustant, F. Barthe and B. Iooss

Proof. a) Let us start by the case 0 ≤ a < b.

The spectral problem is given by:

f  − f  = −λf.

The characteristic equation r2 − r = −λ can be written as:

 2
1 1
r− = −ω 2 , with ω 2 = λ − .
2 4
It is known that the spectral gap of the double exponential distribution on the
whole real line is 14 [9]. By monotonicity of the spectral gap with respect to
inclusion, we can deduce that λ ≥ 14 . It seems clear that this inequality must
be strict, and this can be proved directly. Indeed, if λ = 14 , the general form
of the solution is f (t) = (At + B)et/2 . Thus f  (t) = A B
2t+ 2 +A e
t/2
and the
Neumann conditions lead to the linear system M u = 0 with:
   
1 + a2 12 A
M= and u = .
1 + 2b 12 B

Since a = b, this implies A = B = 0 which is impossible.

Hence λ > 14 , and the solution has the form f (t) = et/2 (A cos(ωt)+B sin(ωt)).
By applying Proposition 4, we deduce that the spectral gap is the first zero of

d(λ) = cos(ωa) sin(ωb) − cos(ωb) sin(ωa) = sin(ω(b − a)).

 2
π 1 π
Hence ω = b−a and λ = 4 + b−a .

b) The same proof is immediately adapted to the case a < b ≤ 0 by using

the change of variables x → −x.
c) Let us now assume that a < 0 < b. By the same argument as in a), we
know that λ ≥ 14 . Let us temporarily admit that the case λ = 14 is impossible.
Then for λ > 14 , the solution has the form:
 |x|/2
e (A+ cos(ωx) + B+ sin(ωx)) if x ≥ 0
f (x) = ,
e|x|/2 (A− cos(ωx) + B− sin(ωx)) if x ≤ 0

where ω 2 = λ − 14 . From Theorem 2 we know that f , searched in Hμ1 ((a, b)),

is actually of class C 1 , and in particular f and f  must be continuous at 0.
Conversely, if this condition is satisfied, f will be of class C 1 since f is C ∞ on
[a, 0[ and ]0, b]. Hence f will belong to Hμ1 ((a, b)), as ρ is bounded on [a, b]. Thus
a necessary and sufficient condition for f to belong to Hμ1 ((a, b)) is that f and
f  are continuous at 0.
The first condition immediately implies A+ = A− . For the second one, we
have:

 −|x|/2 (A/2 + B+ ω) cos(ωx) + (B+ /2 − Aω) sin(ωx) if x ≥ 0
f (x)e =
(−A/2 + B− ω) cos(ωx) + (−B− /2 − Aω) sin(ωx) if x ≤ 0
 Poincaré inequalities on intervals – application to sensitivity analysis 3103

and continuity of f  at 0 leads to B+ − B− = −A/ω. Denoting B = B− , we

have B+ = B − A/ω, and f  is rewritten:

 −|x|/2 −B/2 + A ω + 2ω 1
−f (x)e = (A/2 − Bω) cos(ωx) + sin(ωx) × .
B/2 + Aω

The Neumann conditions f  (a) = f  (b) = 0 then provide a linear system in A, B

which must be singular, leading to the condition det(M ) = 0 with:
1 
2 cos(ωb) + ω + 2ω 1
sin(ωb) −ω cos(ωb) − 12 sin(ωb)
M= .
1
2 cos(ωa) + ω sin(ωa) −ω cos(ωa) + 12 sin(ωa)
Using the properties of the determinant, one can replace the first column C1 by
1
C1 + 2ω C2 :
 
1
ω + 4ω sin(ωb) −ω cos(ωb) − 12 sin(ωb)
det(M ) = det 1
ω + 4ω sin(ωa) −ω cos(ωa) + 12 sin(ωa)
   
1 1 −ωcotan(ωb) − 12
= ω+ sin(ωa) sin(ωb) det
4ω 1 −ωcotan(ωa) + 12
 
1
= ω ω+ sin(ωa) sin(ωb) cotan(ω|a|) + cotan(ω|b|) + 1/ω .
4ω
The last factor in the above product is a decreasing function of ω with un-
bounded limits in the interval ]0, c[, where c = min (π/|a|, π/|b|). It has a unique
zero on this interval, and it is clearly the first non-negative zero of det(M ) as a
function of ω. This proves the result for the Poincaré constant. The inequality
π
ω > b−a is obtained by observing that g : ω → cotan(|a|ω) + cotan(|b|ω) cancels
at b−a < c (for a < 0 < b) since −a b−a
π π
= π − b b−a
π
. Thus if ω ≤ b−a π
then
g(ω) + ω ≥ ω > 0.
1 1

Finally the expression of f is obtained by using the relation between B+ and

B− , and the relation M (A B)T = 0, which is equivalent (since det(M ) = 0)
to AM21 + BM22 = 0. Recalling that sin(ωa) does not vanish for the selected
ω < π/|a|, one can choose A = −M22 / sin(ωa) = ωcotan(ωa) − 12 and B =
M21 / sin(ωa) = 12 cotan(ωa) + ω.
Finally, let us see that λ = 14 is impossible. Then the solution has the form:

A+ x + B+ if x ≥ 0
f (x) = e|x|/2 × .
Ax + B if x ≤ 0

As above, we express that f and f  are continuous at 0. The first condition

implies B+ = B. For the second one, we have:

 −|x|/2 (A+ /2)x + A+ + B/2 if x ≥ 0
f (x)e =
(−A/2)x + A − B/2 if x ≤ 0

and continuity of f  at 0 leads to A+ = A − B. Hence:


 −|x|/2 (A/2 − B/2)x if x ≥ 0
f (x)e = A − B/2 + .
(−A/2)x if x ≤ 0
3104 O. Roustant, F. Barthe and B. Iooss

Now the Neumann conditions f  (a) = f  (b) = 0 provide the linear system
M (A B)T = 0, with:
 
1 − a/2 −1/2
M= .
1 + b/2 −1/2 − b/2

This system must be singular otherwise f would be identically zero, hence

det(M ) = 0. However, det(M ) = 14 (a − b + ab) < 0 since a < 0 < b.

Triangular distribution

Proposition 6. Let us consider the triangular distribution T supported by

[−1, 1], μ(dt) = (1 − |t|)1[−1,1] (t)dt. Denote by J0 the Bessel function of the
first kind with parameter ν = 0 ([1], Chapter 9), and r1 its first root. Then
the Poincaré constant for T is r1−2 , approximately equal to 0.1729, saturated by
f (x) = sign(x)J0 (r1 (1 − |x|)).
Remark 5. The result, involving Bessel functions, shows a connection with the
spectrum of the Laplacian on the unit disk D. Actually, using the symmetry of the
problem, we can see that the spectral problem for the triangular distribution is the
same than for the Laplacian on D restricted to radial solutions, as investigated
e.g. in [13], § 8.1.1d.
Proof of Proposition 6. For the triangular distribution T on Ω = (−1, 1), we
have: μ(dt) = ρ(t)dt = e−V (t) dt with V (t) = − ln(1−|t|). A simple application of
Muckenhoupt criterion (2.4) shows that μ admits a Poincaré constant. However,
we cannot directly apply Theorem 2 since ρ vanishes at the boundary of Ω. The
idea is to consider an increasing sequence of subintervals I such that I ↑ Ω
when → 0, for instance I = (−1+ , 1− ). Since T is a symmetric distribution
on a symmetric support, a minimizer of the Rayleigh ratio can be found among
odd functions (Lemma 7). Thus we can consider the spectral problem on I+ =
(0, 1 − ):
1
f  (t) − f  (t) + λf (t) = 0,
1−t
with boundary conditions f (0) = f  (1 − ) = 0. By applying the change of
variable f (t) = y(x) with x = 1 − t, we obtain the equivalent problem on ( , 1):

1 
y  (x) + y (x) + λy(x) = 0,
x
with initial conditions y(1) = y  ( ) = 0. This is a Bessel-type differential equa-
tion, which solution has the form (see e.g. [1], Chapter 9):
 √   √ 
y(x) = AJ0 x λ + BY0 x λ ,

where J0 and Y0 are the Bessel functions of first and second type. J0 is an even
function of class C ∞ on R, whereas Y0 is C ∞ on R∗+ with infinite limit at 0+ .
 Poincaré inequalities on intervals – application to sensitivity analysis 3105

By using the same argument as in Proposition 4, λ(μ|I ) is then the first zero of
⎛ √  √  ⎞
J0 λ Y0 λ
˜ = det ⎝  √ 
d(λ)  √ ⎠ .
 
J0 λ Y0 λ

Now J  (0) = 0, but Y0 is not defined at 0 and xY0 (x) → π2 as x tends to zero
([1], Chapter 9, Eq. 9.1.9 and 9.1.28). So it is convenient to consider:
⎛ √  √  ⎞
J0 λ Y0 λ
˜ = det ⎝
d(λ, ) = d(λ)  √   √ ⎠ .
J0 λ Y0 λ

By lemma 5, λ(μ) = lim→0 λ(μ|I ) and d is continuous at (λ(μ), 0). Thus, it

holds:
0 = d(λ(μ|I ), ) → d(λ(μ), 0).
→0

This proves that λ(μ) is a zero of

 √  √ 
J0 λ Y0 λ 2 √ 
d(λ, 0) = det 2
= √ J0 λ .
0 √ π λ
π λ

Hence λ(μ) ≥ r1 , where r1 is the first zero of J0 .

In order to prove the converse inequality, we compute the Rayleigh quotient
of the following odd function defined on [−1, 1]:

f (t) := sign(t)J0 (r1 (1 − |t|)) .

The expression of f has been deduced from y(x) = J0 (r1 x) by applying the
transformation x = 1 − t, followed by a symmetrization. By construction, f and
f  are continuous at 0. Thus f is C 1 on Ω and piecewise C 2 on Ω. Furthermore,
it verifies the following equation on [0, 1)

f  − V  f  + r12 f = 0, (4.1)

with boundary conditions f (0) = f  (1) = 0. Since μ is even and f is odd, the
1 1
Rayleigh quotient of f is simply 0 (f  )2 ρ/ 0 f 2 ρ. In order to calculate this
quotient, we multiply Eq. 4.1 by f ρ and integrate by parts on [0, 1). However,
since V = − ln(ρ) has a singularity at 1, we rather fix > 0 and work on
I = [0, 1 − ]. Observe that f and f  ρ are C 1 on I . Starting from Eq. 4.1 and
then integrating by parts, we obtain:
 1−  1−  1−  1−
    
f 2 ρ−[f  f ρ]0 .
1−
2
r1 f ρ=−
2
(f −V f )ρf = − (ρf ) f =
0 0 0 0

The 1last term

 1 tends to 0, when tends to 0. Thus by monotone convergence,
r12 0 f 2 ρ = 0 f 2 ρ. We have shown that the Rayleigh ratio of f is equal to r12 .
Therefore λ(μ) ≤ r12 and the proof is complete.
3106 O. Roustant, F. Barthe and B. Iooss

Truncated normal distribution

Let us consider the standard Gaussian distribution dγ(x) = √1 e−V (x) with
2π
2
V (x) = x2 . The spectral problem (P3) associated to Poincaré inequality is
relative to the operator Lf = f  − V  f  :

Lf = f  − xf  .

The spectral gap is known on every interval formed by successive roots of Her-
mite polynomials (see e.g. [13]), as detailed in Prop. 7. Recall that Hermite
polynomials Hn are the orthonormal basis of polynomials for L2 (dγ) with uni-
∞
tary highest coefficient: −∞ Hn (x)Hm (x)dγ(x) = δn,m . The first ones are:

H0 (x) = 1, H1 (x) = x, H2 (x) = x2 − 1, H3 (x) = x3 − 3x,

and the following can be obtained by the recursive relation: Hn+1 (x) = xHn (x)−
Hn (x). They are also shown to verify the identity Hn+1 
(x) = (n + 1)Hn (x)
which proves, together with the recursion formula, that Hn is solution of the
differential equation Lf = −nf .
Proposition 7 (Spectral gap of the normal distribution, specific case). Let
n ≥ 1, x1 < · · · < xn the roots of the n-th Hermite polynomial Hn . Let I
be one of the intervals: (−∞, x1 ], [x1 , x2 ], . . . , [xn−1 , xn ], [xn , +∞). Then the
(Neumann) spectral gap on I is n + 1.
Proof. On each I, the function Hn is of constant sign and vanishes on the

boundary. Since Hn+1 = (n+1)Hn then inside I, Hn+1 is strictly monotonic and

Hn+1 = 0 on the boundary. Furthermore, Hn+1 is solution of Lf = −(n + 1)f .
Hence, by Corollary 2, n + 1 is the spectral gap.
In the general case, the spectral gap of the truncated normal distribution is
related to hypergeometric series, and more specifically to confluent hypergeo-
metric series or Kummer’s function ([1], Section 13). The Kummer’s function
Ma1 ,b1 (z) = 1 F1 (a1 ; b1 ; z) is a series p≥0 xp with x0 = 1 satisfying

xp+1 (p + a1 )z
= . (4.2)
xp (p + b1 )(p + 1)

Proposition 8 (Spectral gap of the truncated normal distribution). Let a < b,

and define:
 2  2
t t
h0 (λ, t) = M 1−λ ; 1 , h1 (λ, t) = M 2−λ ; 3 .
2 2 2 2 2 2

Notice that h0 and h1 generalize Hermite polynomials: When λ is an odd (resp.

even) positive integer, then t → h0 (λ, t) (resp. t → th1 (λ, t)) is equal, up to a
multiplicative constant, to the Hermite polynomial of degree λ − 1; In particular
they have the same zeros ([1], formulas 13.6.17 & 13.6.18).
 Poincaré inequalities on intervals – application to sensitivity analysis 3107

Denote N|[a,b] the standard normal distribution N (0, 1) truncated on [a, b].
Then its spectral gap λ(N|[a,b] ) is the first zero of the function

d(.) = bh0 (., a)h1 (., b) − ah0 (., b)h1 (., a).

Furthermore:
• If there exists λ such that a, b are two successive zeros of h0 (λ, .) or two
successive zeros of h1 (λ, .) then λ(N|[a,b] ) = λ.
• If a = −b then λ(N|[a,b] ) is the first zero of h0 (., a).
• If a = 0 (resp. b = 0), then λ(N|[a,b] ) is the first zero of h1 (., b) (resp.
h1 (., a)).
Proof. Here it is easier to consider the spectral problem with Dirichlet conditions
(Problem (P4), see Prop. 2). It is given by:

f  (t) − tf  (t) = −(λ − 1)f (t) s.t. f (a) = f (b) = 0. (4.3)

n
Let us look for a series expansion of f , i.e. f (t) = n≥0 cn t . Then,
 
f  (t) = ncn tn−1 ⇒ tf  (t) = ncn tn ,
n≥1 n≥0
 

f (t) = n(n − 1)cn t n−2
= (n + 2)(n + 1)cn+2 tn ,
n≥2 n≥0

and solving (4.3) leads to finding a sequence (cn ) satisfying:

n − (λ − 1)
cn+2 = cn . (4.4)
(n + 1)(n + 2)

Let us split f into even an odd part, f (t) = f0 (t)+tf1 (t) with f0 (t) = p≥0 up t2p
2p
and f1 (t) = p≥0 vp t , where up = c2p and vp = c2p+1 . Then by defini-

t2
tion of Kummer series (4.2), we recognize f0 (t) = c0 M 1−λ ; 1 2 and f1 (t) =
 2 2 2

c1 M 2−λ ; 3 t2 . Thus, with the notations of the Proposition,

2 2

f (t) = c0 h0 (λ, t) + c1 th1 (λ, t).

By applying Proposition 4 and Remark 4, we obtain that the spectral gap is

the first zero of d(λ) and hence the first part of the Proposition. The symmetric
case a = −b is also deduced.
Furthermore:
• If there exists λ such that a, b are two successive zeros of h0 (λ, .), then up
to a change of sign, f = h0 (λ, .) is a positive solution satisfying f (a) =
f (b) = 0. Hence by Corollary 2, λ(N|[a,b] ) = λ. Analogous conclusion
stands with h1 .
• The last item is direct, using that, by construction, h0 (., 0) =
h1 (., 0) = 1.
3108 O. Roustant, F. Barthe and B. Iooss

Fig 1. Basis of finite elements P1 on [0, 1]. The gi ’s are hat functions for i = 1, . . . , n − 1,
truncated at the boundaries (i = 0 and i = n).

4.2. A numerical method

In this section we present a numerical method to estimate the spectral gap. The
technique is well-known in the literature of elliptic problems (see e.g. [30] or
[2]), and we adapt it to our case. We consider the framework and notations of
Section 2.4. The spectral problem (P3),

f  − V  f  = −λf, f  (a) = f  (b) = 0

can be solved numerically by using finite element methods.

The first step is to consider the variational problem (P2), rewritten under
the form (see the Proof of Theorem 2):

α(f, g) = (λ + 1)f, g ∀g ∈ Hμ1 (Ω), (4.5)

where α(f, g) = f  , g  +f, g is coercive on Hμ1 (Ω), and to show its equivalence
to (P3). This was done in Theorem 2.
The second step is to observe that (4.5) can be solved algebraically in a
finite-dimensional space. Denote Gh a finite-dimensional space of Hμ1 (Ω), where
h > 0 is a discretization parameter, and let (gi )0≤i≤n be a basis of Gh . Typically,
Gh = P1 , the space of Lagrange finite elements, composed of piecewise linear
functions on [a, b]. A basis is formed by the ‘hat’ functions gi represented in
Figure 1. A solution in Gh of (4.5) is given by


n
fh = fh,i gi
i=0

and the weak formulation in Gh can be written in the matricial form:

Kh fh = (λ + 1)Mh fh , (4.6)

where fh is the vector of (fh,i )0≤i≤n , and Kh and Mh are called respectively
‘rigidity matrix’ and ‘mass matrix’, defined by:

Kh = (α(gi , gj ))0≤i,j≤n , Mh = (gi , gj )0≤i,j≤n .

 Poincaré inequalities on intervals – application to sensitivity analysis 3109

Notice that Mh and Kh are symmetric and positive definite. Then the problem
can be reexpressed in a standard form by using the Choleski decomposition of
h =
Mh = Lh LTh , where Lh is a lower triangular matrix. Indeed, denoting K
−1 T −1  T
Lh Kh (Lh ) and fh = L fh , (4.6) is written:
h fh = (λ + 1)fh .
K
Thus λ and fh are obtained by performing the eigen decomposition of the sym-
metric matrix K h . Finally fh is deduced from the relation fh = (LT )−1 fh .
h
The last step is to observe that the solutions of the finite-dimensional weak
formulation (4.6) converge to the solutions of (P2), and equivalently (P3), when
h → 0, as stated in [30] (Theorem 6.5.1.), [2] (§ 6.2.2.), or [4] (§ 8, Eq. 8.43 and
§ 10.1.2. Eq. 10.22), with a speed of convergence linked to the regularity of the
solutions. These results are expressed for the Lebesgue measure on [a, b], but are
still valid under the assumptions of Theorem 2 since for all integer , Hμ (Ω) is
then equal to HLeb 
(Ω) with equivalent norms (see the proof of Theorem 2). In
our situation, the space spanned by the first two eigenfunctions lies in Hμ+1 (Ω),
with  ≥ 1 (see Theorem 2). Hence the smallest strictly positive eigenvalue of
(4.6), i.e. the second one, converges to the spectral gap at the speed O(h2 ), and
– since it is a simple eigenvalue – the corresponding eigenvector converges to a
function saturating the Poincaré inequality at the speed O(h ).
Remark 6 (Numerical improvements). We briefly mention two well-known
improvements for finite elements algorithms, which apply to our case (see e.g.
[2]). Firstly, it is possible to replace the computations of integrals in Mh by
quadrature formulas (‘mass lumping’ speed-up technique). This allows obtain-
ing a diagonal matrix and Lh is simply the square root of its diagonal terms.
The convergence result is still valid. Secondly, if the solution is regular enough
(typically f ∈ Hμ3 (Ω)), more regular basis functions can be used (such as P2
finite elements). The speed of convergence is higher, at the price of a higher
computational cost.

4.3. Illustrations

We illustrate the results of the previous paragraphs on the truncated dou-

ble exponential and truncated normal distributions. For the former, the semi-
analytical result has been used. The numerical method was also tested, and gives
a so good approximation that the difference between the two methods could not
be seen in Figure 2. For the latter, however, the numerical method based on
finite elements has been preferred, rather than searching for the first zero of
Kummer’s functions. Indeed, in practice, it is not easy to find numerically the
first zero of a given function. In particular, it was not possible to justify the-
oretically the empirical observation that the first zero is the only zero lying in
the interval ]0, σI−2 [, where σI2 is the variance of the truncated distribution on I.
On the other hand, the numerical method of Section 4.2 is theoretically sound.
Implementations have been done with the R software [35] and the R package
‘orthopolynom’ [27].
3110 O. Roustant, F. Barthe and B. Iooss

Truncated exponential distribution

Denote by μ the double exponential distribution. Figure 2 shows the Poincaré

constant of μ truncated on I = [a, b]. In the general case, we plotted the contour
lines of CP (μ) as a function of Fμ (a) = P (X ≤ a) and 1 − Fμ (b) = P (X ≥ b),
where X ∼ μ. In this representation, symmetric intervals correspond to the
diagonal y = x, intervals (−∞, 0] to y = 12 , and intervals [0, +∞) to x = 12 . We
also added the upper bound of Prop. 5, which is an exact bound in the regions
b ≤ 0 and a ≥ 0, i.e. 1 − Fμ (b) ≥ 12 and Fμ (a) ≥ 12 .
For symmetric intervals, CP (μ) is represented as a function of the mass μ(I),
as well as two upper bounds and the lower bound corresponding to the variance
−b b2
of μ on I, σI = 2 1 − e
2
1+b+ 2 /Fμ (b) . We can see that the upper
bound obtained by symmetric transport 4μ(I)2 (Lemma 6) is quite large here,
except for strong truncations. The upper bound of Prop. 5 is much sharper. The
lower bound is also very sharp, except for small truncations, since it tends to 2
when μ(I) tends to 1 whereas the Poincaré constant tends to 4.

Fig 2. Poincaré constant of the double exponential distribution μ truncated on [a, b]. Left:
π 2 −1
Contour plot as a function of Fμ (a) and 1 − Fμ (b), with the upper bound ( 14 + ( b−a ) ) (red
dashed lines). Right: Representation in the symmetric case for I = [−b, b], with two upper
bounds and the lower bound given by the variance.

Truncated normal distribution

Similar plots have been produced for the normal distribution N (0, 1), truncated
on I = [a, b], gathered in Figure 3. In the symmetric case, the upper bound
obtained by transport μ(I)2 (Lemma 6) looks globally accurate, especially for
strong truncations. The lower bound given by the variance of the truncated
distribution σI2 = 1 − 2 bφ(b)
μ(I) is even sharper. In order to visualize the link to
Hermite polynomials, we have added the points corresponding to intervals [−b, b]
where −b, b are two successive zeros of Hermite polynomials of degree 2n, up to
degree 100. Recall that in that case CP (μ) = 1/(2n + 1).
 Poincaré inequalities on intervals – application to sensitivity analysis 3111

In the general case, we have added the intervals [rn,i−1 , rn,i ] corresponding to
consecutive zeros of Hermite polynomials of degree n (up to n = 100), associated
to Poincaré constants equal to 1/(n + 1). We can observe that the whole set of
Hermite consecutive zeros poorly fill the space of possible intervals, except for
very small symmetric intervals (located around the diagonal x + y = 1). The
Poincaré constants obtained with their extensions, namely Kummer’s functions,
thus provide a useful improvement in many cases.

Fig 3. Poincaré constant of the normal distribution N (0, 1) truncated on [a, b]. Left: Contour
plot as a function of Φ(a) and 1 − Φ(b). Right: Representation in the symmetric case for
I = [−b, b], with the upper bound μ(I)2 and the lower bound given by the variance. Colored
points correspond to Hermite polynomials: For a given degree, the same color is used.

5. Applications

In this section, we perform some global sensitivity analysis by computing the

DGSM-based upper bounds of the total Sobol indices (see Equation 1.4), on
two hydraulic applications based on different models. For these applications,
the common problem under investigation is the assessment of the water level in
the terminal section of a watercourse in case of flood. The water level is evalu-
ated as a function of the discharge and other physical parameters which will be
detailed hereinafter. Uncertainties on these physical parameters are modelled by
probability distributions, chosen from expert knowledge and in accordance with
the empirical distributions obtained during measurement campaigns. Flood phe-
nomenon is governed by the Saint Venant shallow water equations, connecting
the water level localized in space and time, the discharge, the water section, the
lateral inflows, the slope and the head losses due to the friction between the wa-
ter body and the riverbed. The first application relies on a coarse simplification
of these equations while the second application will be based on a 1D solving
model.
As explained in the motivation section of the introduction (see Section 1.1),
global sensitivity analysis is applied on computer codes (here the hydraulic
3112 O. Roustant, F. Barthe and B. Iooss

numerical models) simulating physical phenomena in order to understand the

effects of the model inputs on the variability of the model outputs. To this end,
the variance-based sensitivity measures, called Sobol indices are most often used.
Equation (1.4) shows that the total Sobol index (Eq. (1.2)) of any random input
is upper bounded by its Poincaré constant multiplied by the DGSM (Eq. (1.3)),
which is another sensitivity index. The interest of this inequality relies on the
fact that the computational cost of DGSM is often much lower than the one of
Sobol indices.
In this section, the computations of Poincaré constants have been done us-
ing the R package ‘sensitivity’ [29] of the R software [35]. In this package, the
function ‘PoincareConstant()’ allows to compute the upper bounds given by the
double exponential transport and logistic transport, while the function ‘Poincar-
eOptimal()’ provides the Poincaré constant by solving numerically the spectral
problem.

5.1. First study on a simplified flood model

This application aims at simulating the height of a river and compares it to

the height of a dyke that protects industrial facilities. This academic model has
been introduced for a pedagogical purpose in several methodological papers (see
for example [21]), and used for illustrating DGSM-based sensitivity analysis in
[24] and [31].
In the case of steady flow, with no inflows, and large rectangular section,
assuming that the classical Manning-Strickler formulation is used for the head
losses, we obtain the closed-form solution:
⎛ ⎞0.6
Q
S = H + Zv − Hd − Cb with H = ⎝  ⎠ , (5.1)
Zm −Zv
BKs L

where the output variable S is the maximal annual overflow (in meters), H is
the maximal annual height of the river (in meters), Q is the river flowrate, Ks
the Strickler’s friction coefficient, Zm and Zv the upstream and downstream
riverbed levels (w.r.t. a fixed reference), L and B are the length and width of
the water section, Hd is the dyke height and Cb is the bank level. Table 2 gives
the probability distributions of the model input variables which are supposed to
be independent. Notice that, as in the rest of this paper, we have used here the
notation N (μ, σ 2 ) where σ is the standard deviation, contrarily to [24] which
uses N (μ, σ). Hence the standard deviation of Ks is equal to 8.
Table 3 gives the Poincaré constants and two upper bounds for the different
probability distributions used in this test case (the uniform distribution is not
represented as the result is well known). For the sake of interpretation, we have
considered scaled distributions. Indeed, a simple change of variable (which can
also be viewed as a linear transport) shows that:
  
a−μ b−μ
CP N (μ, σ )|[a, b] = σ CP N (0, 1)
2 2
, ,
σ σ
 Poincaré inequalities on intervals – application to sensitivity analysis 3113

Input Description Unit Probability distribution

X1 = Q Maximal annual flowrate m3 /s Gumbel G(1013, 558)
truncated on [500, 3000]
X 2 = Ks Strickler coefficient - Normal N (30, 82 )
truncated on [15, +∞[
X3 = Zv River downstream level m Triangular T (49, 51)
X4 = Zm River upstream level m Triangular T (54, 56)
X5 = Hd Dyke height m Uniform U[7, 9]
X6 = Cb Bank level m Triangular T (55, 56)
X7 =L River stretch m Triangular T (4990, 5010)
X8 =B River width m Triangular T (295, 305)
Table 2
Input variables of the flood model and their probability distributions.
  
a−μ b−μ
CP (G(μ, β)|[a, b]) = 2
β CP G(0, 1) , ,
β β
 2
b−a
CP (T (a, b)) = CP (T (−1, 1)),
2
where the notation |I means that the distribution is truncated on the interval
I. We can see the strong decrease factor between the Poincaré constant and the
upper bounds based on double exponential transport (gain factor around 6) or
logistic transport (gain factor around 2).

Prob. dist. Upper bound Upper bound CP (μ) Lower bound

μ (db. exp. transp.) (logis. transp.) Var(μ)
T (−1, 1) 1 0.296 0.173 0.167
N (0, 1 | [−1.87, +∞)) 5.912 1.484 0.892 0.862
G (0, 1 | [−0.92, +3.56]) 6.956 2.418 1.257 1.012
Table 3
Poincaré constants and bounds for the scaled laws used in the simplified flood model.

Figure 4 gives the final global sensitivity analysis results for this test case. It
is based on the numerical values obtained in [24]. In particular, the DGSM νi
(i = 1, . . . , 8) have been computed via a sample of the output derivatives com-
ing from a low discrepancy sequence of the inputs of size 10 000. The previous
results of [24] (given here in Fig. 4, left) have shown that the DGSM-based up-
per bounds can be used for a screening purpose (i.e. identifying non influential
inputs which have a total Sobol index close to zero), but are useless for quanti-
fying the effects of the influential inputs (because of their non-informative high
values). In contrary, our new results (Fig. 4, right) show that the DGSM-based
upper bounds give a reliable information on the real influence of the inputs (in
terms of contribution to the model output variance) thanks to their closeness to
total Sobol indices. The hierarchy of influence given by the DGSM-based upper
bounds is the same than for total Sobol indices: the flowrate Q is the most in-
fluential input, followed by Hd , Zv and Ks . The remaining four inputs are not
active in the model. We can also note that screening is now fully efficient as Cb
can be judged as non influential (by choosing for example a threshold of 5% of
the total variance for the sensitivity index).
3114 O. Roustant, F. Barthe and B. Iooss

Fig 4. Comparison between total Sobol indices and DGSM-based upper bounds for the simpli-
fied flood model. Left: Results with double exponential transport; Right: New optimal bounds,
obtained with Poincaré constants.

Remark 7. For the DGSM-based upper bound of Ks , our value (Table 3, left)
is different from the one of [24] (Table 5). Indeed, an error is present in [24]
where the multiplicative factor 2π has been omitted (see also the erratum in [31],
Remark 3). Hence, in [24], 0.198 has to be replaced by 1.244.

5.2. Application on a 1D hydraulic model

The Mascaret computer code ([18]) is a computer code based on a 1D solver of

the Saint Venant equations, allowing engineers to calculate water height for river
flood events. The studied case, taken from [28] and illustrated in Figure 5, is
the French Vienne river in permanent regime whose input data are an upstream
flowrate, a downstream water level, physical parameters (Strickler coefficients)
and geometrical data. The geometrical data consist of 12 transverse river pro-
files. In summary, the model includes the following random input variables,
assumed independent:
• 12 Strickler’s friction coefficients of the main channel Ks,c (noted Ks,c 1
,
12
. . . , Ks,c ) whose distributions are uniform on [20, 40],
• 12 Strickler’s friction coefficients of the flood plain Ks,p (noted Ks,p 1
, ...,
12
Ks,p ) whose distributions are uniform on [10, 30],
• 12 slope perturbations dZ (noted dZ 1 , . . . , dZ 12 ) whose distributions are
standard Gaussian, truncated on [−3, 3],
• and one discharge value Q whose distribution is Gaussian with zero mean,
standard deviation 50 and truncated on [−150, 150].
Our goal in this test case is to update with the new DGSM bounds the work
of [28] who compare DGSM bounds and Sobol indices. In this previous work, the
derivatives of the model output with respect to the 37 inputs can be efficiently
(with a cost independent of the number of inputs) computed by using the adjoint
model of Mascaret (obtained by automatic differentiation, [19]). From a Monte
Carlo sample of the inputs of size n = 20 000, n output values (water height)
 Poincaré inequalities on intervals – application to sensitivity analysis 3115

Fig 5. Representation of a river cross section, showing the main channel (low flow channel)
and the flood plain (high flow channel). Source: [28].

and n output derivatives are obtained. This very large n has been used for
a demonstrative purpose, while in other industrial studies n ranges from 100
to 1 000 ([22, 36]). From the sample of derivatives, the 37 DGSM νi are then
computed. [28] has shown that 32 inputs have DGSM-based upper bounds close
to zero. Then, amongst the 37 inputs, only 5 are potentially active.
Table 4 gives our new results for these 5 remaining inputs. To compute the
DGSM bound, we have also computed Var(f (x)) = 0.369 whose standard devia-
tion (sd = 3.4e−3) is obtained by bootstrap, i.e. by resampling with replacement
11 12
the Monte Carlo sample of output values. As inputs Ks,c and Ks,c follow a uni-

Inputs 11
Ks,c 12
Ks,c dZ 11 dZ 12 Q
ST 0.456 0.0159 0.293 0.015 0.239
(2e−3) (1e−4) (1e−3) (1e−4) (1e−3)
ν 5.695e−3 2.728e−4 1.089e−1 6.592e−3 3.553e−5
(3e−5) (4e−6) (3e−4) (9e−5) (6e−8)
By double exponential transport
Upper bound for CP (μ) - - 6.249 6.249 15623.26
Upper bound for S T - - 1.844 0.116 1.504
- - (2e−3) (2e−3) (1.5e−2)
By logistic transport
Upper bound for CP (μ) - - 1.562 1.562 3905.815
Upper bound for S T - - 0.461 0.028 0.376
- - (4e−3) (5e−4) (4e−3)
Optimal, with the Poincaré constant
CP (μ) 40.528 40.528 0.976 0.976 2441.071
Optimal bound for S T 0.625 0.029 0.288 0.017 0.235
(2e−4) (1e−5) (3e−3) (3e−4) (2e−3)
Table 4
Sensitivity indices for the Mascaret test case. S T gives the total Sobol index, ν the DGSM.
Standard deviations of all these estimates are obtained by bootstrap and given in
parentheses. Partial results are not shown for Ks,c11 , K 12 , which are uniformly distributed.
s,c
3116 O. Roustant, F. Barthe and B. Iooss

form distribution, we report only the optimal DGSM bounds, already used in
[28]. For the 3 other inputs (dZ 11 , dZ 12 and Q), as in the previous application, a
strong decrease factor is obtained for the bounds by using the logistic constant
(gain factor around 4) and the optimal constant (gain factor around 6) instead
of the double exponential constant. The new bounds now allow to identify the
variable dZ 12 as inactive. Moreover, they provide the same ranking than to-
tal Sobol indices for the influential inputs. Notice that the optimal bounds are
nearly equal to the total Sobol indices. The numerical errors, assessed by the
standard deviations of the estimates, explain that the bound values are some-
times slightly smaller than the Sobol values.
In terms of interpretation, the sensitivity analysis results show that the hy-
draulic engineers have to concentrate the research efforts on the knowledge of
the physical parameters Ks and dZ at the 11th river profile, in order to be able
to reduce the prediction uncertainty of the water level when a flood occurs.
From a methodological point of view, the numerical model users are now able
to perform a global sensitivity analysis, in the sense of variance decomposition,
at a lower cost than before by using the DGSM-based technique.

Acknowledgements

We acknowledge an associate editor and a reviewer for their relevant comments

on the first version of this article. We are grateful to Fabrice Gamboa, who
initiated this research. We thank EDF R&D/LNHE for providing the Mas-
caret test case and Sébastien Petit who has performed the computations on
this model. We also thank Laurence Grammont and the members of the team
‘Génie Mathématique & Industriel’ for useful discussions. We acknowledge the
participants of MascotNum 2016 & Mexico MascotNum conferences for their
feedback. Part of this research was conducted within the frame of the Chair in
Applied Mathematics OQUAIDO, gathering partners in technological research
(BRGM, CEA, IFPEN, IRSN, Safran, Storengy) and academia (CNRS, Ecole
Centrale de Lyon, Mines Saint-Etienne, University of Grenoble, University of
Nice, University of Toulouse) around advanced methods for Computer Experi-
ments.

References

[1] Abramowitz, M. and Stegun, I. (1972). Handbook of Mathematical

Functions. Dover Publications, Tenth printing. MR0208797
[2] Allaire, G. (2007). Numerical Analysis and Optimization: An Intro-
duction to Mathematical Modelling and Numerical Simulation. Numeri-
cal Mathematics and Scientific Computation. Oxford Science Publications.
MR2326223
[3] Ané, C., Blachère, S., Chafaı̈, D., Fougères, P., Gentil, I., Mal-
rieu, F., Roberto, C. and Scheffer, G. (2000). Sur les inégalités de
 Poincaré inequalities on intervals – application to sensitivity analysis 3117

Sobolev logarithmiques. Panoramas et Synthèses 10. Société Mathématique

de France, Paris. MR1845806
[4] Babuska, I. and Osborn, J. (1991). Eigenvalue Problems In Handbook
of Numerical Analysis II 641–787. Elsevier Science. MR1115240
[5] Bakry, D., Gentil, I. and Ledoux, M. (2014). Analysis and geometry
of Markov diffusion operators, volume 348 of Grundlehren der Mathematis-
chen Wissenschaften [Fundamental Principles of Mathematical Sciences].
Springer, Cham. MR3155209
[6] Barthe, F., Bianchini, C. and Colesanti, A. (2013). Isoperimetry and
stability of hyperplanes for product probability measures. Annali di Matem-
atica 192 165–190. MR3035134
[7] Barthe, F., Cattiaux, P. and Roberto, C. (2006). Interpolated in-
equalities between exponential and Gaussian, Orlicz hypercontractivity and
isoperimetry. Rev. Mat. Iberoam. 22 993–1067. MR2320410
[8] Bobkov, S. and Götze, F. (2009). Hardy Type Inequalities via Riccati
and Sturm-Liouville Equations In Sobolev Spaces In Mathematics I: Sobolev
Type Inequalities 69–86. Springer New York. MR2508839
[9] Bobkov, S. and Ledoux, M. (1997). Poincare’s inequalities and Tala-
grand’s concentration phenomenon for the exponential distribution. Prob-
ability Theory and Related Fields 107 383–400. MR1440138
[10] Bobkov, S. G. and Houdré, C. (1997). Isoperimetric constants for
product probability measures. The Annals of Probability 25 184–205.
MR1428505
[11] Brézis, H. (2010). Functional analysis, Sobolev spaces and partial differ-
ential equations. Springer. MR2759829
[12] Chen, M. (1999). Analytic proof of dual variational formula for the first
eigenvalue in dimension one. Science in China Series A: Mathematics 42
805–815. MR1738551
[13] Dautray, R. and Lions, J. L. (1990). Mathematical Analysis and Nu-
merical Methods for Science and Technology: Volume 3 Spectral Theory and
Applications. Springer. MR1064315
[14] de Jong, P. (1990). A central limit theorem for generalized multilinear
forms. Journal of Multivariate Analysis 34 275–289. MR1073110
[15] de Rocquigny, E., Devictor, N. and Tarantola, S., eds. (2008).
Uncertainty in industrial practice. Wiley.
[16] Djellout, H. and Wu, L. (2011). Lipschitzian norm estimate of one-
dimensional Poisson equations and applications. Ann. Inst. Henri Poincaré
Probab. Stat. 47 450–465. MR2814418 MR2814418
[17] Efron, B. and Stein, C. (1981). The jackknife estimate of variance. The
Annals of Statistics 9 586–596. MR0615434
[18] Goutal, N., Lacombe, J. M., Zaoui, F. and El-Kadi-
Abderrezak, K. (2012). MASCARET: a 1-D open-source software
for flow hydrodynamic and water quality in open channel networks. In
River Flow 2012: Proceedings of the International Conference on Fluvial
Hydraulics (R. Murillo Muñoz, ed.) 2 1169–1174. CRC Press, San José,
Costa Rica.
3118 O. Roustant, F. Barthe and B. Iooss

[19] Griewank, A. and Walther, A. (2008). Evaluating derivatives: Prin-

ciples and techniques of automatic differentiation. SIAM Philadelphia.
MR2454953
[20] Hoeffding, W. F. (1948). A class of statistics with asymptotically normal
distributions. Annals of Mathematical Statistics 19 293–325. MR0026294
[21] Iooss, B. and Lemaitre, P. (2015). A review on global sensitivity analysis
methods. In Uncertainty management in Simulation-Optimization of Com-
plex Systems: Algorithms and Applications (C. Meloni and G. Dellino,
eds.) 101–122. Springer.
[22] Iooss, B., Popelin, A.-L., Blatman, G., Ciric, C., Gamboa, F., La-
caze, S. and Lamboni, M. (2012). Some new insights in derivative-based
global sensitivity measures. In Proceedings of the PSAM11 ESREL 2012
Conference 1094–1104.
[23] Kucherenko, S., Rodriguez-Fernandez, M., Pantelides, C. and
Shah, N. (2009). Monte Carlo evaluation of derivative-based global sensi-
tivity measures. Reliability Engineering and System Safety 94 1135–1148.
[24] Lamboni, M., Iooss, B., Popelin, A. L. and Gamboa, F. (2013).
Derivative-based global sensitivity measures: General links with Sobol’ in-
dices and numerical tests. Mathematics and Computers in Simulation 87
45–54. MR3046876
[25] Mossel, E., O’Donnell, R. and Oleszkiewicz, K. (2010). Noise sta-
bility of functions with low influences: invariance and optimality. Annals of
Mathematics (2) 171 295–341. MR2630040
[26] Muckenhoupt, B. (1972). Hardy’s inequality with weights. Studia Math-
ematica 44 31–38. MR0311856
[27] Novomestky, F. (2013). orthopolynom: Collection of functions for or-
thogonal and orthonormal polynomials, R package version 1.0-5.
[28] Petit, S., Zaoui, F., Popelin, A. L., Goeury, C. and Goutal, N.
(2016). Couplage entre indices à base de dérivées et mode adjoint pour
l’analyse de sensibilité globale. Application sur le code Mascaret. Preprint,
https://hal.archives-ouvertes.fr/hal-01373535.
[29] Pujol, G., Iooss, B. and Janon, A. (2016). sensitivity: Global Sensitivity
Analysis of Model Outputs, R package version 1.13.
[30] Raviart, P. A. and Thomas, J. M. (1988). Introduction à l’analyse
numérique des équations aux dérivés partielles. Masson. MR0773854
[31] Roustant, O., Fruth, J., Iooss, B. and Kuhnt, S. (2014). Crossed-
derivative based sensitivity measures for interaction screening. Mathematics
and Computers in Simulation 105 105–118. MR3240497
[32] Rudin, W. (1987). Real and complex analysis, Third ed. McGraw-Hill Book
Co., New York. MR924157 MR0924157
[33] Sobol, I. (1993). Sensitivity estimates for non linear mathematical mod-
els. Mathematical Modelling and Computational Experiments 1 407–414.
MR1335161
[34] Sobol, I. and Kucherenko, S. (2009). Derivative-based global sensitivity
measures and the link with global sensitivity indices. Mathematics and
Computers in Simulation 79 3009–3017. MR2541313
 Poincaré inequalities on intervals – application to sensitivity analysis 3119

[35] R Core Team (2015). R: A Language and Environment for Statistical

Computing R Foundation for Statistical Computing, Vienna, Austria.
[36] Touzany, S. and Busby, D. (2014). Screening Method Using the
Derivative-based Global Sensitivity Indices with Application to Reservoir
Simulator. Oil & Gas Science and Technology – Rev. IFP Energies nou-
velles 69 619–632.