HAL Id: hal-00680725
https://hal.archives-ouvertes.fr/hal-00680725
Preprint submitted on 20 Mar 2012
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Variance-based sensitivity analysis using harmonic analysis
Jean-Yves Tissot, Clémentine Prieur
To cite this version:
Jean-Yves Tissot, Clémentine Prieur. Variance-based sensitivity analysis using harmonic analysis.
2012. �hal-00680725�
analysis
Jean-YvesTissot
∗
and Clémentine Prieur
Université de Grenoble LJK/INRIA
51, rue des Mathématiques
Campus de Saint Martin d'Hères, BP53
38041 Grenoble edex 09 (Frane)
(jean-yves.tissotimag.fr)
(lementine.prieurimag.fr)
Phone: +33 (0)4 76 635447
Fax: +33 (0)4 76 6312 63
Marh20, 2012
Abstrat
FourierAmplitudeSensitivityTest(FAST)andRandomBalaneDesign(RBD)arepopular
methodsofestimatingvariane-basedsensitivityindies. Werevisittheminlightofthedisrete
Fourier transform (DFT) on nite subgroups of the torus and randomized orthogonal array
sampling. Wethen study theestimation errorofboththese methods. This allowsto improve
FAST and to deriveexpliit rates of onvergene of its estimators byusing theframework of
lattie rules. We also give a natural generalization of the lassi RBD by using randomized
orthogonal arrays having any parameters, and we provide a bias orretion method for its
estimators.
Keywords: global sensitivity analysis, random balane design, Fourier amplitude sensitivity test,
orthogonalarrays,lattie rules
∗
Correspondingauthor.
Variane-based sensitivity analysis onsists in omputing indies the so-alled variane-based
sensitivity indies (SI)or Sobol' indies (see [34 ℄) that areessentially multiple integrals. Many
numerial tehniques have been developed to estimate these quantities. This inludes the rude
MonteCarloestimator(see[34℄,and[18℄for areent work),thepolynomialhaos-basedestimators
(see[37 ℄and[2 ℄)andthe FASTmethod(see [9 ℄and[30 ℄)aswellasits derivedapproah,RBD(see
[38 ℄), and their hybrid approah, RBD-FAST (see [38 ℄ and [24 ℄), and many others (see [29℄ for a
review).
The main purpose of this paper is to revisit FAST and RBD by using the disrete harmoni
analysis framework, in order to arry out a theoretial error analysis. In these methods the SI
estimation amountsto omputinga nite numberof the omplex Fourier oeients of themodel
ofinterestdenedontheunithyperube. Intheorytheseomputationsouldbedonebyperforming
a rude Monte Carlo integration or a ubature on a regular grid. But the rate of onvergene of
the Monte Carlo method is low, and ubatures aregenerally unfeasible inhigh dimension beause
ofthe exponential growthof thenumberofnodes, alsoknownastheurse ofdimensionality.
A rst possible starting point to overome these drawbaks is to note that the disrete om-
plex Fourier oeients omputed by using the ubature approah are exatly the oeients in
the representation of the trigonometri interpolation polynomial of the model of interest on the
regular grid. Consequently this approah onsists of a trigonometri interpolation issue and an
be generalized by using Smolyak algorithm on sparse grids (see [12℄). Suh interpolation shemes
are quiteeient as long as the model of interest is suiently smooth (see [3℄). But thematrix
ofthe interpolationoperator insuha methodsuers froman inreaseof its onditionnumber for
both inreasing renement oftheregulargridand inreasingmodeldimension, andthus makesthe
interpolationshemeunstable(see [19 ℄).
Asa onsequene,it turnsout to be obviousthat, inorder to avoidthestabilityissue, one has
to fous onunitary operators. Thus DFT operators onnite subgroups of thetorus (see e.g. [23℄)
i.e. theunithyperubeviewasagroupwhosematrieshaveaperfetonditionnumberequal
to
1
arepartiularly well-suited inthepresent framework. Thisleadsto theuseof lattierules(see [33 ℄ for a review)to whih FAST, asshown inSubsetion 4.1, is loselyrelated. In a seondtime,by viewing nite subgroups of the torus asorthogonal arrays (see [16 ℄ for a review), theprevious
method an be generalized by performing a randomization proess on these arrays. This leads to
to whih RBD,asshowninSubsetion 4.2,is loselyrelated.
Thepaperproeedsasfollows. InSetion2,wesetupthenotation,wegivebakgroundmaterials
relatedtotheANOVAdeompositionandtotheFourierseriesrepresentation,andweintroduethe
lassof estimators ofinterest. InSetion 3, we rst reviewboth FAST and RBD,andthen revisit
them. Setion 4isdevotedtotheerroranalysisbyusingtherevisiteddenitionprovidedinSetion
3. At last, Setion 5 gives numerial illustrations of RBDestimates on an analytial model. Most
ofthe proofsof the propositions aregiveninappendix A.
2 Bakground
2.1 Notation
First,
E[Y ]
,E[Y |X]
andVar[Y ]
denotethe unonditional expetation ofY
,theonditional expe-tation of
Y
givenX
and the variane ofY
,respetively. By onvention, we deneE[Y |∅] = E[Y ]
.Seondly,onsider a parameter
d
inN ∗
the dependeneon whih is omitted for onvenieneanddene for any
u ∈ {1, . . . , d}
,Z u = {k ∈ Z d | ∀i ∈ u , k i ∈ Z
and∀i / ∈ u , k i = 0}
Z ∗ u = {k ∈ Z d | ∀i ∈ u , k i ∈ Z ∗
and∀i / ∈ u , k i = 0}
andfor all
i ∈ N ∗
,Z u (i) = Z u ∩
− i 2 , i
2 i d
Z ∗ u (i) = Z ∗ u ∩
− i 2 , i
2 i d
.
Lastly,adesign ofexperimentsisommonlydenotedby
D
and,fori ∈ N ∗
,thenotationD(i)
refersto the regulargridin
[0, 1) d
D(i) =
0, 1
i , . . . , i − 1 i
d
.
2.2 Variane-based sensitivity indies
Let
X = (X 1 , . . . , X d ) ∈ [0, 1] d
be ad
-dimensional random vetor and let us onsiderY = f (X)
where
f : [0, 1] d → R
isa measurable funtion suh thatE[Y 2 ] < +∞
. Under theassumption thatX
hasindependentomponents,theHoedingdeomposition[17,41℄statesthatY
anbeuniquelydeomposed into summandsof inreasing dimensions
Y − E[Y ] = X d m=1
X
u ⊆{1,...,d}
| u |=m
f u (X i , i ∈ u )
(1)where the
2 d − 1
randomvariables ontheright-hand sideof (1)shouldsatisfytheproperty∀ v u , E
f u (X i , i ∈ u )|X i , i ∈ v
= 0 .
(2)Notethat inthis ase the randomvariables
f u (X i , i ∈ u )
have meanzero and are mutually unor-related. Thereforetakingthevarianeofbothsidesin(1)givesthevariane deomposition [14 ,34℄
of
Y
Var
[Y ] = X d m=1
X
u ⊆{1,...,d}
| u |=m
Var
f u (X i , i ∈ u ) .
Finally,ifVar
[Y ] 6= 0
,we dene theso-alled variane-basedsensitivityindies or Sobol' indies asS
u (f, X) =
Varf u (X i , i ∈ u )
Var
[Y ] .
Inpratie,global sensitivityanalysisfousesonomputingtherst-order (
| u | = 1
)andtheseond-order(
| u | = 2
) terms.2.3 Fourier series representation
From here on let us assume that the
X i
's are independent and uniformly distributed on[0, 1]
.Thereforethejoint probabilitydensity funtionof
X
on[0, 1] d
isequal to1
and, denotingP n (f, X) =
n 1
X
k 1 =−n 1
· · ·
n d
X
k d =−n d
k (f)
exp(2iπk · X)
where
k (f ) = Z
[0,1] d
f (X)
exp(−2iπk · X)dX ,
the Riesz-Fishertheoremyields
P n (f, X) −→ L 2 Y .
Inpartiular, wehave
Y = X
k 1 ∈Z
· · · X
k d ∈Z
k (f)
exp(2iπk · X)
a.s. (3)andasthefollowingpropositionshows,thisFourierseriesrepresentationgivesanharmoniapproah
to handlethevariane-basedsensitivity indies.
Proposition1. Let
X 1 , . . . , X d
be independentrandomvariables uniformlydistributed on[0, 1]
andletus onsider
Y = f (X)
wheref : [0, 1] d → R
isa measurable funtionsuh thatE[Y 2 ] < +∞
andVar
[Y ] 6= 0
. Thenfor any non-emptysubsetu
of{1, . . . , d}
we haveS
u (f, X) = X
k∈Z ∗ u
k (f ) 2 X
k∈(Z d ) ∗
k (f ) 2 .
(4)Proof. Inviewof(3),itiseasytonotiethattheomponentsintheHoedingdeompositionsatisfy
f u (X i , i ∈ u ) = X
k∈Z ∗ u
k (f )
exp(2iπk · X)
a.s.andtheonlusion follows fromParseval'sidentity.
As in (4) the index S
u (f, X)
does no more depend onX
we now simply denote the sensi-tivity indies by S
u (f )
. In the same way, we now denote Vu (f )
and V(f )
the parts of varianeVar
f u (X i , i ∈ u )
and the total variane Var
[Y ]
, respetively. Lastly, whenu = {i 1 , . . . , i s }
isexpliitelygiven,weuse themore ommon notation V
i 1 ...i s (f )
andSi 1 ...i s (f )
.2.4 Estimation
We now dene basi estimatorsbased onProposition 1. For anynon-emptysubset
u
of{1, . . . , d}
,let
K u
be anite subset ofZ ∗ u
andD
anite subset of[0, 1) d
with|D| = n
. Denotingb
k (f, D) = 1
n X
x∈D
f(x)
exp(−2iπk · x),
(5)we dene theestimatorof V
u (f )
asthetrunated seriesb
V
u (f, K u , D) = X
k∈K u
| b
k (f, D)| 2 ,
(6)theestimatorof V
(f )
asthe empirialvarianeb
V
(f, D) = 1 n
X
x∈D
f (x) − 1 n
X
y∈D
f (y) 2
(7)
andtheestimator ofS
u (f )
naturally asb
S
u (f, K u , D) =
Vb u (f, K u , D) b
V
(f, D) .
(8)Example 1. If the design of experiments
D
is a set of independent random pointsuniformly dis- tributed on[0, 1] d
andK = G
u ⊆{1,...,d}
u 6=∅
K u ,
we have
b
V
u (f, K u , D) =
Vu ( ˜ f )
where
f ˜ (X) = X
k∈K∪{0}
b
k (f, D)
e2iπk·X
istheapproximationof
f (X)
usingthequasi-regressionapproah [1℄basedon therandomsampleD
.Note that
| b
k (f, D)| 2
isa biased estimatorof|
k (f, D)| 2
andit isreommended touse the unbiasedestimator
n
n − 1 | b
k (f, D)| 2 − 1 n 2
X
x∈D
f 2 (x)
!
(see e.g. [22℄). In the sameway, the empirial variane V
b (f, D)
should be replaed by the unbiasedsample variane
n
n−1
Vb (f, D)
.Example2. If thedesignof experiments
D
istheregular gridD(q)
withn = q d
,q ∈ N ∗
andiffor all non-emptysubsets
u
of{1, . . . , d}
,K u = Z ∗ u (q)
andK = G
u ⊆{1,...,d}
u 6=∅
K u
then by Parseval's identity,it an be easily shown that
b
S
u f, K u , D(q)
=
Su ( ˜ f)
where
f ˜ (x) = X
k∈K
b
k f, D(q)
e
2iπk·x
is the trigonometri interpolation polynomial of
f (x)
(see e.g. [11 ℄) at then = q d
equally spaednodes
x ∈ D(q)
.3 New introdution to FAST and RBD
Inthe sequel,sine the
X i
'sareindependent and uniformlydistributedon[0, 1]
, we haveE
f (X)
= Z
[0,1] d
f (x)dx
sowe use no more probabilisti notation. Moreover, theintegrability assumption on
f
now readsf ∈ L 2 ([0, 1] d )
.3.1 Review of FAST
3.1.1 Numerial integration
FAST is essentially an appliation of the following result due to Weyl [43 ℄ (see also the Weyl's
ergoditheorem [42℄ingerman or [32℄)
Theorem 1. [Weyl℄ Let
g
be a bounded Riemann integrable funtion on[0, 1] d
and for alli = 1, . . . , d
,x i (t) = {ω i t}
where theω i
's are real numberslinearly independent overQ
and{·}
denotesthe frational part,then
Z
[0,1] d
g(x)dx = lim
T →∞
1 2T
Z T
−T
g x 1 (t), . . . , x d (t)
dt.
(9)Inpartiular, for any
k ∈ Z d
andg : x 7→ f (x)
exp− 2iπk · x)
,(9)readsk (f ) = lim
T →∞
1 2T
Z T
−T
f ◦ x(t)
exp− 2iπ(k · ω)t
dt.
(10)ThenFASTonsistsinreplaing
x i (t) = {ω i t}
withsemiparametri funtionsx i (t) = G i sin(ω i t)
(see [8 ℄) where the
ω i
's are positive integers and the transformationsG i
are hosen to preservethemarginal distributions of the
X i
's. If the latterare uniformly distributed asin the present paper , it an be shown (see [9℄ and [30℄) thatG i (·) = π 1 arcsin(·) + 1 2
. Saltelli et al. [30℄ alsopropose to add a random phase-shift
ϕ i ∈ [0, 2π)
, getting the semiparametri funtionsx ∗ i (t) =
1
π arcsin sin(2πω i t + ϕ i )
+ 1 2
. Hene, replaingx
withx ∗
in(10)givesk (f ) ≈ lim
T →∞
1 2T
Z T
−T
f ◦ x ∗ (t)
exp− 2iπ(k · ω)t dt.
Thus,sine the funtions
x ∗ i
are1
-periodi, itomesk (f ) ≈ Z 1
0
f ◦ x ∗ (t)
exp− 2iπ(k · ω)t dt
andapplying the retanglerule to theright-hand sideintegral gives
k (f ) ≈ b
k·ω (f ◦ x ∗ ).
(11)where
b
k·ω (f ◦ x ∗ ) = 1 n
n−1 X
j=0
f ◦ x ∗ j n
exp
− 2iπj k · ω n
istheomplex disreteFourieroeient oftheone-dimensionalfuntion
f ◦ x ∗
. Inthesequel,thedependene on
n
,ω
andϕ
is generallyomitted foronveniene.The estimators of V
u (f )
, V(f )
and onsequently of Su (f )
were introdued by using the approx-imation in (11) (see [8℄ and Appendix C in [9℄). On the one hand, for any non-empty subset
u ⊆ {1, . . . , d}
andanynitesubsetK u ⊆ Z ∗ u
,(11)leadsto thedenitionoftheestimator ofVu (f ) b
V FAST
u (f, K u , x ∗ ) = X
k∈K u
b
k·ω (f ◦ x ∗ ) 2 .
(12)Ontheotherhand, (11)gives
V
(f ) =
0 (f 2 ) −
0 (f) 2
≈ b
0 (f 2 ◦ x ∗ ) − b
0 (f ◦ x ∗ ) 2
andParseval'sidentity leads to the denitionoftheestimator ofV
(f ) b
V
FAST
(f, x ∗ ) =
n−1 X
k=1
b
k (f ◦ x ∗ ) 2 .
Thisnaturally leads to theestimatorof thevariane-based sensitivityindies S
u (f ) b
S FAST
u (f, K u , x ∗ ) = X
k∈K u
b
k·ω (f ◦ x ∗ ) 2
n−1 X
k=1
b
k (f ◦ x ∗ ) 2 .
Asin Example 2,note that by Parseval's identity V
b
FAST(f, x ∗ )
is equal to the empirial varianeb
V
(f, {x ∗ ( n j )} j=0..n−1 )
.3.1.3 Choie of parameters
ω
andn
Asdisussed byShaiblyandShuler[31℄andCukier etal. [10 ℄,
ω
andn
shouldbeorretlyhosensoas to minimize theubature error inthe approximationin (11). In orderto avoid interferenes
i.e.
k · ω − k ′ · ω = 0
fork
,k ′ ∈ Z d
,k 6= k ′
andaliasing i.e.
k · ω − k ′ · ω = jn
fork
,k ′ ∈ Z d
,k 6= k ′
andj ∈ Z ∗
that both lead to
b
k·ω (f ◦ x ∗ ) = b
k ′ ·ω (f ◦ x ∗ )
Shaiblyand Shuler[31℄ proposeto hooseω 1
,...,
ω d
free ofinterferenes up toorderN ∈ N ∗
:(k − k ′ ) · ω 6= 0
for allk, k ′ ∈ Z d , k 6= k ′
,s.t.X d i=1
|k i − k i ′ | ≤ N + 1
(13)and
n
suiently largen ≈ N max(ω 1 , . . . , ω d ).
(14)More reently, referring to the lassial information theory, Saltelli et al. [30 ℄ suggest to replae
(14)withNyquist-Shannon sampling theorem(see e.g. [24 ℄)
n > 2N max(ω 1 , . . . , ω d ).
(15)Inour opinion,the riterion stated in(13) shouldbe written
(k − k ′ ) · ω 6= 0
for allk, k ′ ∈ Z d , k 6= k ′
,s.t.X d i=1
|k i | ≤ N ′
andX d i=1
|k ′ i | ≤ N ′
(16)sinethemainobjetiveistoavoidinterfereneswithinanitesubsetof
Z d
outofwhihtheFourieroeients of
f
are a priori negligible in (16), this subset is the losedl 1
-norm ball of radiusN ′
. Thus we may reformulate the wholeriterion stated in (13) and (15) with respet to the setK = ⊔ u K u
where theK u
's are thetrunation sets in the FAST estimator of Vu (f )
given in(12).We proposeto hoose
ω 1
,...,ω d
freeof interferenes withinK
i.e.(k − k ′ ) · ω 6= 0
for allk, k ′ ∈ K, k 6= k ′
andn > max
k,k ′ ∈K (k − k ′ ) · ω
.
(17)Inthe sequel,we refer tothelatter asthe"lassi"riterion of FAST.
3.2 Review of RBD
RBD makes use of the previous framework setting
ϕ = 0
,ω 1 = · · · = ω d = ω ∈ N ∗
usuallyset to
1
and applying random permutations on the oordinates of the resulting pointsx ∗ ( n j )
.Morepreisely, let
σ 1
, ...,σ d
be random permutations on{0, . . . , n − 1}
andS
denote the set of all possibleσ = (σ 1 , . . . , σ d )
. Givenσ ∈ S
, onsider the funtionx × = (x × 1 , . . . , x × d )
dened on{0, n 1 , . . . , n−1 n }
suh thatfor alli ∈ {1, . . . , d}
andj ∈ {0, . . . , n − 1}
,x × i j n
= 1
π arcsin sin
2πω σ i (j) n
+ 1 2 .
Thus denoting
σ i −1
the inverse permutation ofσ i
,denex ×,i j
n
= x × σ −1 i (j) n
.
Finallythrougha heuristi argument Tarantola etal. [38℄ introdue theRBDestimators of V
u (f )
,V
(f )
and Su (f )
for rst-order terms i.e.u = {i}
,i ∈ {1, . . . , d}
. For any nite subsetK {i} ⊆ Z ∗ {i}
,we haveb
V RBD
i (f, K {i} , x × ) = X
k∈K {i}
b
k i ω (f ◦ x ×,i ) 2 ,
b
V
RBD
(f, x × ) =
n−1 X
k=1
b
k (f ◦ x × ) 2 .
and
b
S RBD
i (f, K {i} , x × ) = X
k∈K {i}
b
k i ω (f ◦ x ×,i ) 2
n−1 X
k=1
b
k (f ◦ x × ) 2 .
Asin FAST note that by Parseval's identity, the estimator V
b
RBD(f, x × )
is equal to the empirialvariane V
b (f, {x × ( n j )} j=0..n−1 )
. In thesequel,the dependene onω
andσ
isgenerally omittedforonveniene.
3.3 FAST and RBD revisited
3.3.1 Main result
Firstweintrodue morenotation. For any
p ∈ N ∗
,letr p : [0, 1] −→ [0, 1]
x 7−→
2{px}
if0 ≤ {px} < 1 2 2 − 2{px}
if1 2 ≤ {px} ≤ 1
andfor any
ϕ ∈ [0, 2π)
t ϕ : [0, 1] −→ [0, 1]
x 7−→ {x + ˜ ϕ}
withϕ ˜ = 1 4 + 2π ϕ .
Then we dene the linear operators
R p
andT ϕ
(see Figure 1) onL 2 ([0, 1] d )
suh that for allx ∈ [0, 1] d
,R p f (x) = f r p (x 1 ) . . . , r p (x d )
et
T ϕ f (x) = f t ϕ 1 (x 1 ), . . . , t ϕ d (x d ) .
andnote that
R p = R 1 ◦ · · · ◦ R 1
| {z }
p
times. We also introdue two lassialdesigns ofexperiments. For any
ω ∈ (N ∗ ) d
,we denoteG(ω) = n j n ω 1 o
, . . . , n j n ω d o
, j ∈ {0, . . . , n − 1}
.
theylisubgroupoforder
n/gcd(ω 1 , . . . , ω d , n
ofthetorus
T d = (R/Z) d ≃ [0, 1) d
generatedby
({ ω n 1 }, . . . , { ω n d })
(see e.g. [15 ℄). For anyσ ∈ S
we alsodenoteA(σ) = σ 1 (j)
n , . . . , σ d (j) n
, j ∈ {0, . . . , n − 1}
0 0.5 1 0
2 4 6
(a)Plotof
f : x 7→ x + sin(x)
0 0.5 1
0 2 4 6
(b)Plotof
R 1 f
0 0.5 1
0 2 4 6
()Plotof
T π
10 f
0 0.5 1
0 2 4 6
(d)Plotof
( T π
30 ◦ R 1 )f
Figure1: Examples of operators
R p
andT ϕ
indimension1
.theorthogonal array of strength
1
and index unity with elements taken from{0, 1 n , . . . , n−1 n }
andbasedon thepermutation
σ
(see e.g. [16℄). FAST andRBD methods are now introdued ina newway byusing the basi estimatorin(8).
Proposition 2. Let
f : [0, 1] d → R
be a square-integrable funtion. For any non-empty subsetu ⊆ {1, . . . , d}
, any nite subsetK u ⊆ Z ∗ u
,ϕ ∈ [0, 2π) d
andω ∈ (N ∗ ) d
, we haveb
S FAST
u (f, K u , x ∗ ) = b
Su (T ϕ ◦ R 1 )f, K u , G(ω)
.
(18)For any
i ∈ {1, . . . , d}
, any nite subsetK {i} ⊆ Z ∗ {i}
,σ ∈ S
andω ∈ N ∗
,we haveb
S RBD
i (f, K {i} , x × ) = b
Si (T ω ˜ ◦ R ω )f, ωK {i} , A(σ)
.
(19)where
ω ˜ = (1−ω)π
2ω , · · · , (1−ω)π 2ω
and
ωK {i} = {(ωk 1 , . . . , ωk d ), k ∈ K {i} }
.Proof. Itessentiallyonsistsinshowing thatfor all
j ∈ {0, . . . n − 1}
f ◦ x ∗ j n
= (T ϕ ◦ R 1 )f n j n ω 1 o
, . . . , n j
n ω d o
and
f ◦ x × j n
= (T ω ˜ ◦ R ω )f σ 1 (j)
n , . . . , σ d (j) n
.
See detailsinAppendix A.1.
Remark1.IntheRBDmethod,theparameter
ω
isusuallysetto1
butitsroleisnotwellunderstoodupto now. In our opinion there isno reason toset
ω 6= 1
sine ifgcd(ω, n) = 1
then it leads to thease
ω = 1
, and otherwise the estimatorin (19) is potentially less eient than in the aseω = 1
(see details in Appendix A.2.).
3.3.2 What FAST and RBDare
It is lear from Proposition 2 that FAST and RBD only onsists in applying the basi estimator
introdued in(8)to a partiular transform
(T ϕ ◦ R p )f
ofthe funtionf
anda partiular designofexperiments
G(ω)
orA(σ)
. Now it is also lear that the basi estimator generates an error termdue to trunations in(6) and an other one due to numerial integrations in(5) and (7).
Moreover, theuse of
(T ϕ ◦ R p )f
instead off
ould also have an impat on the sensitivity indies estimationerror. Wenowinvestigatethis latterissuebyintroduingthenotionof invariane ofthevariane deomposition.
Denition 1. Let
L
be a linear operator onL 2 ([0, 1] d )
. The variane deomposition is said to beL
-invariant onL 2 ([0, 1] d )
if for any non-empty setu ⊆ {1, . . . , d}
and any funtionf ∈ L 2 ([0, 1] d )
we have
V
u (Lf ) =
Vu (f ).
Thisleads to thefollowing result
Lemma1. Forany
p ∈ N ∗
andanyϕ ∈ [0, 2π) d
,thevarianedeompositionisR p
andT ϕ
-invarianton
L 2 ([0, 1] d )
.Proof. See Appendix A.3.
Asa onsequene,for anynon-emptysubset
u ⊆ {1, . . . , d}
,we haveS
u (T ϕ ◦ R p )f
=
Su (f )
andthisassertsthevalidityofFASTandRBDmethods. Notethatthelinearoperator
R p
"regular- ize"thefuntionf
inthesense thatifx 7→ f (x)
isontinuouson[0, 1] d
andx → f ({x 1 }, . . . , {x d })
is disontinuous on
R d
thenx → R p f ({x 1 }, . . . , {x d })
is ontinuous onR d
. This is an impor-tant property sine by Riemann-Lebesgue lemma
|
k (f )|
onverges to0
as||k||
tends to∞
, andthe smoother the funtion
f
, the faster the onvergene (see e.g. [45℄). The other operatorT ϕ
essentially allows to dene randomized estimatorsin FAST.
3.3.3 Potential generalizations
To endwith, we listthree natural generalizationsthatarefurther disussed inthenextsetion:
- theestimator
b
Su (T ϕ ◦ R 1 )f, K u , G(ω)
an alsobe dened fora group
G
of any rankr ≤ d
- the estimator
b
Si (T ω ˜ ◦ R ω )f, ωK {i} , A(σ)
an also be dened for a sensitivity index of any
order:
b
Su (T ω ˜ ◦ R ω )f, ωK u , A(σ)
,notethatithasbeen alreadyapplied in[44℄
- thelatterestimator
b
Su (T ω ˜ ◦ R ω )f, ωK u , A(σ)
analsobedened foranorthogonalarray
A
having anyparameters.
4 Error analysis
For onveniene, operators
T ϕ
andR p
are now omitted. Moreover, we assumethat thefuntionf
hasanabsolutely onvergent Fourier representation,i.e.
X
k∈Z d
|
k (f )| < +∞ .
4.1 Cubature error in FAST
4.1.1 Two points of view
Inthis setionwe mainly fous onthe error term
e k (f, G) = b
k (f, G) −
k (f )
(20)where
G
is a subgroup ofT d
of ordern
andk ∈ Z d
. By its denition, the termb
k (f, G)
onsistsof an equal weight ubature rule at the
n
nodes of the groupG
, also known as a lattie rule (see[33 ℄forasurvey). MoreoverbythegeneralizedPoissonsummationformula(seee.g. [23℄),theerror
term in(20)ispreisely
e k (f, G) = X
h∈G ⊥ \{0}
k+h (f )
(21)where
G ⊥ = {h ∈ Z d | ∀x ∈ G, h · x ≡ 0 (
mod1)}
is the subgroup ofZ d
orthogonal toG
, alsoknownastheduallattie of
G
.In the lattie rules eld,
e 0 (f, G)
is the only term of interest, and there exist two main points ofviewtoontrolit. Oneonsistsinlookingfor"good"groups
G
suhthattheubatureruleisexatfora set oftrigonometri polynomials,i.e. for a nitesubset
K
ofZ d
,e 0 (f, G) = 0
for allf
suh that∀k ∈ / K,
k (f ) = 0 .
Theother point ofview aimsto nd"good"groups
G
suh thattheubature rulehasan absoluteerror
|e 0 (f, G)|
dominated byan expliitbound forallf
inapartiular spae ofsmooth funtions.Notethatthese approahes areompatible to eahother (see e.g. [7℄ andthereferenes therein).
NowonerningthestudyoferrorinFAST,therstpointofview,whihessentiallyorresponds
to the lassi FAST, onsists of a trigonometri interpolation issue and leads to a metamodel ap-
proah of the estimation of thesensitivityindies. The seond one, whih ismore original, allows
to derive error bounds for V
b u (f, K u , G)
and Vb (f, G)
in spaes of smooth funtions. Both thesemethods are disussedbelow.
4.1.2 Metamodel approah
Let
K
be a nitesubset ofZ d
. Thenan immediate onsequene of(21) isthat agroupG
satisestheproperty
e k (f, G) = 0
for allk ∈ K
andfor allf
suh that∀k ∈ / K,
k (f ) = 0
ifandonly if
∀k, k ′ ∈ K, k 6= k ′ , ∃x ∈ G, (k − k ′ ) · x 6≡ 0 (
mod1) .
(22)Morefundamentally,for any
E ⊆ Z d
,onsiderthe trigonometri polynomialf ˜ E (x) = X
k∈E
b
k (f, G)
exp(2iπk · x) ,
(23)thentheequivalene above leads tothefollowing result
Proposition 3. Let
G
be a subgroup of the torusT d
of order|G| = n
andK = ∪ u 6=∅ K u
satisfyingthe riterion (22) where for all non-emptysubsets
u
of{1, . . . , d}
,K u ⊆ Z ∗ u
i) if
|K| = n
,thenf ˜ K
isatrigonometriinterpolationpolynomialoff
atthen
nodesx ∈ G
andwe have
b
S
u (f, K u , G) =
Su ( ˜ f K ).
ii) if
|K| < n
, letH
be any subset ofZ d
suh thatK ⊆ H
,H
satises the riterion (22) and|H| = n
. Thenf ˜ H
isa trigonometriinterpolationpolynomial off
atthen
nodesx ∈ G
andwe have
b
V
u (f, K u , G) =
Vu ( ˜ f K )
and Vb (f, G) =
V( ˜ f H ).
Proof. Theonlydiultyistoprove thatthetrigonometri polynomials
f ˜ K
intheassertion i)andf ˜ H
in the assertion ii) are interpolation polynomials at the pointsx ∈ G
. We demonstrate it forf ˜ K
,theproofforf ˜ H
isexatlythesame.Sinethefuntion
f
hasabsolutely onvergent Fourier representation, we an writef (x) = X
k∈Z d
k (f )
exp(2iπk · x) = X
k∈K
X
h∈G ⊥
k+h (f )
exp2iπ(k + h) · x
(24)
(seedetails inAppendixA.4) and bydenition of
G ⊥
,we have thatfor anyx ∈ G
,f (x) = X
k∈K
X
h∈G ⊥
k+h (f )
exp(2iπk · x).
Theonlusion followsfrom thedenitionin(23)sine (20) and(21) give
X
h∈G ⊥
k+h (f ) = b
k (f, G).
From this pointof view,FAST returnsanalytial valuesfrom trigonometri metamodels of the
funtion
(T ϕ ◦ R 1 )f
and theerroranalysis shouldbeperformedon themetamodelitself.In pratie, a set of a priori non-negligible frequenies
K = ∪ u 6=∅ K u
is given and a groupG
satisfyingtheriterion (22)andwiththesmallest order
|G| = n
hastobefound. Searhingforthisgroup
G
isomputationaly expensiveandmayrapidlybeomeunfeasible. Oneoftheheapestway isto look for yli groupsG = G(ω)
, oming bakto thelassi FAST. In this ase, theriterion(22)simplyreads
∀k, k ′ ∈ K, k 6= k ′ , (k − k ′ ) · ω 6≡ 0 (
modn) .
(25)Note that this new riterion plays the same role as the lassi riterion of FAST given in (17).
Themain dierene between these two approahes is thatoptimization on
n
is performed in(25),onsequently this new riterion allows to nd group
G
with smaller ordern
. We illustrate theeieny of both riterions by using basi exhaustive algorithms with omputational omplexity
O(n d )
. TheresultsaregatheredinTable1andshowthatthenewriterionleadstoanon-negligible improvement.d = 2 d = 3 d = 4 d = 5 N1 N 2 |K| n old n new |K| n old n new |K| n old n new |K| n old n new
4 2 20 41 29 36 65 50 56 105 63 80 177 111
5 3 32 61 48 66 141 102 112 241 173 170 471 302
6 4 48 85 65 108 241 155 192 541 323 300 997 613
7 5 68 113 89 162 421 284 296 1177 586 470 1891 1279
8 6 92 145 120 228 625 429 424 1985 1033 680 3457 2222
9 7 120 181 149 306 937 645 576 3007 1706 930
10 8 152 221 185 396 1281 933 752 4501 2529 1220
11 9 188 265 228 498 1805 1284 952 7261 3684 1550
Table 1: Comparison in dimension
d = 2
,3
,4
and5
between the minimum sample sizen
givenby thelassi riterion of FAST(denoted
n old
) and thenew one proposed in(25) (denotedn new
).Here, the
K {i}
'sare equaltoZ ∗ {i} ∩ {|k i | ≤ N 1 }
,theK {i,j}
'sareequal toZ ∗ {i,j} ∩ {|k i | + |k j | ≤ N 2 }
andfor all
u
suhthat| u | > 2
,K u = ∅
. SuhsetsK
arepartiularlywell-suitedtoanalysefuntions whose eetive dimensionis lessthan2
seeDenition 4 inSetion 4.2.2.Remark 2. Even if yli groups seem to be suitable in the previous issue, the omputational ost
of the researh of a generator
ω
an beome prohibitive in high-dimensionalproblems. In this ase,alternativealgorithmsan beused insteadofa systematiresearhtehnique (for a reent referene,
see e.g. [20℄).
4.1.3 Error bounds
Searhing for a nite subgroup
G
of the torusT d
suh thate 0 (f, G)
has an expliit bound in apartiular funtionspaeisa problemknownastheonstrution ofgoodlattie rules (forasurvey
see [33℄ or more reently [25℄). Most of the results inthis eld are established in Korobov spaes
whih are suitable to handle lattie methods; so we derive error bounds for sensitivity indies in
thesespaes. For
α > 1
andγ = (γ u ) u ⊆{1,...,d}
withnon-negativeγ u
's, dene theweightedKorobovspae
H α,γ
to bethe Hilbertspae withreproduing kernelRK α,γ (x, y) = 1 + X
k∈(Z d ) ∗
r(k, α, γ) −1
exp2iπk · (x − y)
where for any
k 6= 0
,r(k, α, γ) = γ u −1 k
Q
i∈ u k |k i | α
,whereu k
is suh thatk ∈ Z ∗ u k
. Fork
suh thatγ u k = 0
,we set byonventionr(k, α, γ) = ∞
. Thus thekernel an be rewrittenRK α,γ (x, y) = 1 + X
k∈(Z d ) ∗ γ u
k 6=0
r(k, α, γ) −1
exp2iπk · (x − y)
andwe deduethat thenormof
f ∈ H α,γ
satises||f|| 2 H α,γ =
0 (f ) 2 + X
k∈(Z d ) ∗ γ u k 6=0
r(k, α, γ)|
k (f )| 2 < +∞
andonsequently
∀k ∈ (Z d ) ∗ such that γ u k 6= 0, |
k (f )| 2 ≤ γ u k ||f || 2 H α,γ Y
i∈ u k
|k i | α .
Note that for any
k ∈ (Z d ) ∗
suh thatγ u k = 0
,f ∈ H α,γ
impliesk (f ) = 0
. We also make arestritiononthesetsoffrequenies
K u
's. Hereweassumethatforanynon-emptysetu ⊆ {1, . . . , d}
,K u
isof Zaremba ross-type (seeFigure 2)K u = Z u ,β u = (
k ∈ Z ∗ u , Y
i∈ u
|k i | ≤ β u
)
where
β u ≥ 1
. Thiskindofsparsegridsispartiularlywell-suitedfortheanalysisofhigh-dimensional smooth funtions. We now give theresulton error boundsfor Vb u (f, K u , G)
andVb (f, G)
inH α
.−5 −4 −3 −2 −1 0 1 2 3 4 5
−5
−4
−3
−2
−1 0 1 2 3 4 5
(a)Plotof
Z {1,2},4
−10 −5 0 5 10
−10
−5 0 5 10
(b)Plotof
Z {1,3},9
Figure2: Illustrationof rosses
Z u ,β u
.Proposition 4. Let
f ∈ H α,γ
withα > 2
andγ = (γ u ) u ⊆{1,...,d}
with non-negative omponents.Let
G
be a subgroup ofT d
of ordern
suh that the ubature error related toG
is dominated by theexpliit bound
B (α, n, d, γ)
on the unit ball ofH α,γ
i.e. for allf
inH α,γ , | b
0 (f, G) −
0 (f )| ≤ B(α, n, d, γ)||f || H α,γ
. Theni) if there exists
α ′ > 2
andγ ′ = (γ u ′ ) u ⊆{1,...,d}
withnon-negative omponents suhthatf 2 ∈ H α ′ ,γ ′
,b
V(f, G) −
V(f ) ≤ ||f || 2 H α B(α, n, d, γ) 2 + B(α, n, d, γ)
+ ||f 2 || H α ′ B (α ′ , n, d, γ ′ )
ii) for any non-empty set
u ⊆ {1, . . . , d}
andK u = Z u ,β u
, we haveb
Vu (f, K u , G) −
Vu (f ) ≤ ||f || 2 H α,γ h
C(α, γ, β u , | u |) + B (α, n, d, γ) 2 S 1 (α, γ, β u , u ) + B(α, n, d, γ)S 2 (α, γ, β u , u ) i
where
S 1 (α, γ, β u , u ) = γ f rac X
k∈K u
Y
i∈ u
|k i | + 1 α
, γ f rac = max
u , v ⊂{1,...,d}
γ v 6=0
γ u /γ v
S 2 (α, γ, β u , u ) = γ f rac γ u 1/2 2 α| u |/2 |K u |
andfor
| u | ≤ 2
, the trunation error termC(α, β u , | u |)
areC(α, γ, β u , 1) = 2γ max ζ (α)
β u α−1
, γ max = max
u ⊂{1,...,d} γ u
(26)C(α, γ, β u , 2) = 4γ max
ζ(α) 2 + ζ(α)
log(β u ) + 2 β u α−1
.
(27)Proof. See Appendix A.5.
It is also possible to derive expliit formulas of the trunation error term for
| u | > 2
, but thisis more ompliated and of seond interest. Seondly, it has to be noted that, inthe seond item
ofProposition4,the funtions
S 1
andS 2
areinreasing withrespettotheparameterβ u
whilethefuntion
C
isdereasing. Asaonsequene,eient boundsonsistofatrade-obetweenβ u
andn
suhthat
B(α, n, d, γ) 2 S 1 (α, γ, β u , u )
,B(α, n, d, γ)S 2 (α, γ, β u , u )
andC(α, γ, β u , | u |)
have thesameorder. For example,
i) if
| u | = 1
andα > 2
,notethat|K u | = 2β u
anddedueS 1 (α, γ, β u , u ) ≤ 2 α| u |+1 β u 1+α
,andreallthat
C(α, γ, β u , 1) = O(β u 1−α )
. Thus thetrade-ogivesb
Vu (f, K u , G) −
Vu (f ) = O
B (α, n, d, γ) 1− α 1 .
ii) if
| u | = 2
andα > 2
, note that|K u | ≤ 4β u (
log(β u ) + 1)
see argument for (A.21)in Appendix A.5 and dedue
S 1 (α, γ, β u , u ) ≤ 2 α| u |+2 β u 1+α (
log(β u ) + 1)
and reall thatC(α, γ, β u , 1) = O β u 1−α
log(β u )
. Thus thetrade-o gives
b
Vu (f, K u , G) −
Vu (f ) = O
log
B (α, n, d, γ) −1/α
B(α, n, d, γ) 1− α 1
.
Remark 3. In unweighted Korobov spaes i.e.
γ = 1
, it is knownthat the optimal rate of onver-gene of a rank-1 lattie rule is
B(α, n, d, γ) = O
(log n) dα/2 n α/2
(see e.g. [33 ℄). For unweighted Korobov spaes, there exist better rates of onvergene for produt
weights i.e.
γ u = Q
i∈ u γ i
(see [21 ℄) or for nite-orderweights i.e∀ u
with| u | > d ∗ (d ∗ ≤ d)
,γ u = 0
(see [13℄). The latter are essentially related toan assumptionon theeetivedimension of
f
in thetrunation sense and in the superposition sense, respetively (see [5℄ for the denition of eetive
dimension).
4.2 Bias in RBD
We now give some results onthewell-known issuerelatedto thebias oftheestimates inRBD.
4.2.1 Preliminaries
We begin withthe denitionsof an orthogonalarrayand the"oinidene defet"
Denition 2. An orthogonal array in dimension
d
, withq
levels, strengtht ≤ d
and indexλ
isamatrix with
n = λq t
rowsandd
olumns suh that in everyn
-by-t
submatrix eah of theq t
possiblerows i.e. the distint
t
-uples(l 1 , . . . , l t )
where thel i
's take their values in the set of theq
levelsours exatly the samenumber
λ
of times.Denition 3. Let
A
be an orthogonal array in dimensiond
, withq
levels, strengtht
and indexλ
. We say thatA
has the oinidene defet when there exist two rows ofA
that do agree int + 1
olumns; otherwise we say that
A
is defet-free.Let
Π(q)
be the set of permutations on{0, 1 q , . . . , q−1 q }
,Π = Π(q, d)
the artesian produt(Π(q)) d
andµ = µ(q, d)
thenormalized ountingmeasureonΠ(q, d)
. LetA
bean orthogonalarrayindimension
d
,withq
levels{0, 1 q . . . , q−1 q }
,strengtht
andindexλ
,anddenoten = λq t
itsnumberofrows. Foranypermutation
π = (π 1 , . . . π d ) ∈ Π
,denoteA(π)
theorthogonalarrayobtainedfromA
after applyingeah permutationπ j
on thelevelsof theorrespondingj
-thfator i.e.for all
1 ≤ i ≤ n
and1 ≤ j ≤ d, A(π)
ij = π j (A ij ) .
Notethatthe
A(π)
'sandA
areorthogonalarrayswiththesameparameters (see[16 ℄). Conversely, itisalsoeasy toshowthatifA
hasstrengthandindex equalto1
i.e. asinthelassiRBDwithanodd integer
1
n
;anyother orthogonalarrayA ′
withthesame paramaters asA
isof theformA(π)
for apermutationπ ∈ Π
. We arenowinterested inthequantitiesE µ
h
Vb (f, A(π)) i
and
E µ
h
Vb u (f, K u , A(π)) i ,
where
K u
isa nite subsetofZ ∗ u
.4.2.2 Bias of the estimator in RBD
Let
b
k (f ) = b
k (f, D(q))
denote thek
-th omplex disrete Fourier oeient; we begin with thefollowing important lemma
Theorem 2. [Owen℄ Followingthe previous notation, we have
Var
µ
h b
0 f, A(π) i
= 1 n 2
X
| u |>t
X | u |
r=0
B( u , r)(1 − q) r−| u | X
k∈Z ∗ u (q)
| b
k (f )| 2
where
B ( u , r) = X n i=1
X n j=1
1 |{l∈ u , A il =A jl }|=r
onsists of the number of pairs of rows
(A i , A j )
that mathon exatlyr
of the axes inu
.Proof. This is exatly Theorem 1 given by Owen in [26℄. Just note that, the embedded ANOVA
termsona
q d
regulargrid denotedβ u
byOwen areβ u (x) = X
k∈Z ∗ u (q)
b
k (f )
exp(2iπk · x).
Indeed,for all
x
inthe regulargrid{0, 1 q , . . . , q−1 q } d
,f (x) = X
u ⊆{1,...,d}
β u (x)
by a trigonometri interpolation argument, and it is also easy to show that the random variables
β u (X i , i ∈ u )
satisfy theproperty (2)for independent random variablesX i
uniformly distributed on{0, 1 q . . . , q−1 q }
.Thenwehavethe followingpropositioninwhihthebiasofthevarianeestimateisinvestigated
inunweightedKorobovspaes
H α = H α,1
(see Setion 4.1.3.)1
If
n
iseven,thedesignof experimentsinRBDonsistsof anorthogonalarraywithn/2
levels,strength1andindex2,andmaybefaedwiththeoinidenedefet.
Proposition5. Let
A
be a defet-free orthogonal array in dimensiond
withparametersq
,t
andλ
in
N ∗
witht < d
. If there existsα > 2t + 1
suh thatf
andf 2
are inH α
, we haveE µ
h
Vb (f, A(π)) i
=
V(f ) − 1 n
X
1≤| u |>t
V
u (f ) + O n −(1+ 1 t ) .
Proof. See Appendix A.6.
As a onsequene, onsidering thelassi denition of eetive dimension inthe superposition
sense(see e.g. [5 ℄)
Denition 4. The eetive dimension of
f
, in the superposition sense, is the smallestd S (f )
suhthat
X
1≤| u |≤d S (f)
V
u (f ) ≥ l S (f )
V(f )
where
l S (f )
isan arbitrary onstant generally set at0.99
.we have theorollary
Corollary1. UndertheassumptionsofProposition5,let
d S (f )
andl S (f )
bedenedasinDenition4. If
t ≥ d S
, we haveE µ
h
Vb (f, A(π)) i
=
1 − ε n
V
(f ) + O n −(1+ 1 t ) ,
where
0 ≤ ε ≤ 1 − l S (f)
.Proof. Straightforward fromProposition 5.
Ina seond time,sine
E µ
h
Vb u (f, K u , A(π)) i
= X
k∈K u
E µ
h b
k f, A(π) 2 i
the analysisofthe biasof the partsof varianeestimates restson thefollowing result
Proposition 6. Let
A
be a defet-free orthogonal array in dimensiond
withparametersq
,t
andλ
in
N ∗
witht < d
. Letu
be a non-empty subset of{1, . . . , d}
andk ∈ Z ∗ u
. If there existsα > 2t + 1
suh that