Strongly nested interpolatory quadrature and

extension to n-D via Smolyak’s method

J. Dumont–Le Brazidec2 J. Peter1

1
ONERA - The French Aerospace Lab. - Aerodynamics, Aeroelasticity, Aeroacoustics Dept.
2
IRSN - Institut de Radioprotection et de Sécurité Nucléaire

June 20, 2018

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 1 / 26
 Introduction

Outline

1 Introduction

2 Interpolatory quadrature

3 Accuracy of interpolatory rules. Proposed 1D rules

4 Tests for classical 1D functions

5 Extension to nD functions via Smolyak’s sparse grids

6 Test for nD functions

7 Conclusions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 2 / 26
 Introduction

Introduction

ONERA does Uncertainty Quantification in CFD, for missile, profiles,

wings..., in dimension 1 to 10 with geometrical and operational uncertainty
parameters
Study strongly nested quadrature for integration in 1D and nD

Derive n/n + 2 nested sets of points from upper bounds of integration error
of interpolatory quadrature
Compare accuracy for 1D function w.r.t. Leja sequences
Extend results to corresponding Smolyak’s sparse grids

Comprehensive & extensive description of this work is avaible in a

ResearchGate Preprint
Dumont - Le Brazidec, J., Peter, J. Strongly nested 1D interpolatory quadrature and
extension to 1D via Smolyak method. DOI: 10.13140/RG.2.2.22860.51840

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 3 / 26
 Interpolatory quadrature

Outline

1 Introduction

2 Interpolatory quadrature

3 Accuracy of interpolatory rules. Proposed 1D rules

4 Tests for classical 1D functions

5 Extension to nD functions via Smolyak’s sparse grids

6 Test for nD functions

7 Conclusions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 4 / 26
 Interpolatory quadrature

Introduction
1D Interpolatory quadrature

x1 < x2 < ... < xn distinct nodes in [-1,1]

Numerical integration by interpolatory quadrature
Z 1 n
X
I [f ] = f (x)dx In [f ] = wi f (xi )
−1 i=1

w1 , w2 ...wn being the weights

Definition of weights for fixed nodes for exact quadrature of (1, x, x, ...x n−1 )
Classical properties
wi is the exact sum of Lagrangian polynomial Li over [-1,1]
In [f ] is the exact integral of Lagragian/polynomial interpolant in {x1 , x2 , ..., xn }

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 5 / 26
 Accuracy of interpolatory rules. Proposed 1D rules

Outline

1 Introduction

2 Interpolatory quadrature

3 Accuracy of interpolatory rules. Proposed 1D rules

4 Tests for classical 1D functions

5 Extension to nD functions via Smolyak’s sparse grids

6 Test for nD functions

7 Conclusions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 6 / 26
 Accuracy of interpolatory rules. Proposed 1D rules

Accuracy of interpolatory rules C 0 function (1/4)

Denote Z 1 n
X
Rn [f ] = f (u)du − wi f (xi )
−1 i=1

Integration error point of view

Error of an n-point interpolatory rule for a C 0 function

En−1 = Minbn−1 ∈Pn−1 ||f − bn−1 ||∞

i=n
X
|I [f ] − In [f ]| ≤ 2En−1 + |wi |En−1
i=1

|I [f ] − In [f ]| ≤ 4En−1 for positive weights

Does not provide any a clear way to define sets of successive points for
interpolatory quadrature

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 7 / 26
 Accuracy of interpolatory rules. Proposed 1D rules

Accuracy of interpolatory rules C 0 function (2/4)

Polynomial interpolation error point of view

Error of Lagrangian polynomial interpolation in infinity norm
n
X
Leb n (x) = |Lj (x)| Λn = max Leb n (x)
x∈[−1,1]
j=1
kf − Πn (f )k∞ ≤ (1 + Λn ) En−1 |Rn [f ]| ≤ (2 + 2Λn ) En−1 .

Defined sets of 3,5,7,9,11...33 points to get lowest possible values of

Lebesgue constant Λn .
successive optimization LebConst − so sequence of nodes
global optimization LebConst − go sequence of nodes.

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 8 / 26
 Accuracy of interpolatory rules. Proposed 1D rules

Accuracy of interpolatory rules C 0 function (3/4)

successive mimization of Lebesgue constant up to n = 257, LebConst − so

sequence of nodes
LebConst−so
S = { 0, −1, 1,
33
± 0.620911304689912, ± 0.851116906275467, ± 0.366893560557081, ± 0.949647785059444,
± 0.228417726292775, ± 0.726262806499719, ± 0.982331094378673, ± 0.476576613596911,
± 0.796922940675856, ± 0.131045085133918, ± 0.913065535238725, ± 0.555583265445366,
± 0.994175387122204, ± 0.303955266207398, ± 0.683362049812905}

global mimization of Lebesgue constant up to n = 33, LebConst − go

sequence of nodes
LebConst−go
S = { 0, −1, 1,
33
± 0.782948160530396, ± 0.442592038865533, ± 0.952401768588314, ± 0.289462405510965,
± 0.631474499429657, ± 0.900375626425636, ± 0.139818101722143, ± 0.984577322855455,
± 0.723776947906398, ± 0.524915574329834, ± 0.860067776493385, ± 0.204617156721863,
± 0.997188614738299, ± 0.376693743737914, ± 0.817795412857899}

Lebesgue constant, integral of Lebesgue function, infinity norm of nodal

polynomial for the successive sets

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 9 / 26
 Accuracy of interpolatory rules. Proposed 1D rules

Accuracy of interpolatory rules C 0 function (4/4)

Integration error point of view

Error of an n-point interpolatory rule for a C 0 function
 Z 1 
n
|I [f ] − In [f ]| ≤ 2 + Leb (x)dx En−1
−1

Defined sets of 3,5,7,9,11...33 points to get lowest possible values of integral

of Lebesgue function.
successive optimization LebInt − so sequence of nodes
global optimization LebInt − go sequence of nodes

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 10 / 26
 Accuracy of interpolatory rules. Proposed 1D rules

Accuracy of interpolatory rules C n function

Error of an n-point interpolatory rule for a C n function

l=n
2 (n) Y
|I [f ] − In [f ]| ≤ ||f ||∞ || (x − xl )||∞
n!
l=1

Low infinity norm of nodal polynomial defines the xi as cos of regularly

spread angles in [0, π]

Pick the location of the maximum of {x1 , x2 ...xn } nodal polynomial and add
it to the set of current nodes = define a Leja sequence
Classical way of defining strongly nested sets of points for interpolatory
quadrature

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 11 / 26
 Accuracy of interpolatory rules. Proposed 1D rules

1D Interpolatory nested quadrature

Definition : all or some of the nodes of the n-point rule are also included in
some of the further (n + p)-point rules.

Not a property of the basic versions of Gauss quadratures. Nested extension

and restriction exist. Gauss-Patterson

Property of Fejér’s second rule [Fejér 1933] n/2n + 1 nesting.

Property of Clenshaw and Curtis’ rule [Clenshaw Curtis 1960, Imhof 1963] n/2n − 1
nesting.

Leja sequence allow n/n + 2 or n/n + 1 nesting

Test interpolatory quadratures based on symmetric n/n + 2 nested
sequences of points, with odd cardinal, starting from {−1, 0, 1}

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 12 / 26
 Tests for classical 1D functions

Outline

1 Introduction

2 Interpolatory quadrature

3 Accuracy of interpolatory rules. Proposed 1D rules

4 Tests for classical 1D functions

5 Extension to nD functions via Smolyak’s sparse grids

6 Test for nD functions

7 Conclusions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 13 / 26
 Tests for classical 1D functions

Test for entire functions

Figure: Test of 1D rules for entire functions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 14 / 26
 Tests for classical 1D functions

Test for C ∞ functions

Figure: Test of 1D rules for C ∞ functions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 15 / 26
 Tests for classical 1D functions

Test for C 0 ,C 1 ,C 2 functions

Figure: Test of 1D rules for C 0 ,C 1 ,C 2 functions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 16 / 26
 Extension to nD functions via Smolyak’s sparse grids

Outline

1 Introduction

2 Interpolatory quadrature

3 Accuracy of interpolatory rules. Proposed 1D rules

4 Tests for classical 1D functions

5 Extension to nD functions via Smolyak’s sparse grids

6 Test for nD functions

7 Conclusions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 17 / 26
 Extension to nD functions via Smolyak’s sparse grids

Smolyak’s sparse grids

Hierarchy of 1D quadratures Ql
Differences of quadrature Q0 [f ] = 0. ∆k [f ] = Qk [f ] − Qk−1 [f ]
Note that X
Ql1 ⊗ ... ⊗ Qld [f ] = (∆k1 ⊗ ... ⊗ ∆kd )f
k/ 1≤kj ≤lj

Reduce the set of indices to a simplex = Smolyak’s sparse grids

X
Qld [f ] = (∆k1 ⊗ ... ⊗ ∆kd )[f ]
|k|1 ≤l+d−1

Comprehensive, extensive presentation illustrated for 2D in ResearchGate

preprint

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 18 / 26
 Extension to nD functions via Smolyak’s sparse grids

Smolyak sparse grids – 2D illustration

Figure: Visualization of Smolyak’s sparse grids in 2D U12 , U22 , U32 for a nested
1 − 3 − 5-points univariate base rule

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 19 / 26
 Test for nD functions

Outline

1 Introduction

2 Interpolatory quadrature

3 Accuracy of interpolatory rules. Proposed 1D rules

4 Tests for classical 1D functions

5 Extension to nD functions via Smolyak’s sparse grids

6 Test for nD functions

7 Conclusions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 20 / 26
 Test for nD functions

Test for nD functions

Using “Smobol” Package by R.P. Dwight TU Delft

Four of Genz functions
i=d
X 1
F1 = cos (2Πu1 + ad xi ) F2 =
i=d
i=1
Y
(1/ad2 + (xi − ui )2
i=1

i=d
X i=d
X
F4 = exp(− ad (xi − ui )2 ) F5 = exp(− ad |xi − 0.5|).
i=1 i=1

Define 1D rule LebConst − so, LebConst − go, LebInt − so, LebInt − go, Leja,
LejaSym in new python classes. Smobol buids corresponding sparse grids
Clenshaw-Curtis and Gauss-Patterson already available in Smobol

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 21 / 26
 Test for nD functions

Test for cosine function

Figure: Test of sparse grid rules for C ∞ oscillatory functions F1

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 22 / 26
 Test for nD functions

Test for exponential of absolute value

Figure: Test of sparse grid rules based for C 0 function F5

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 23 / 26
 Test for nD functions

Test for nD functions

Equivalent results for classical Leja ,LejaSym, LebConst − so, LebConst − go,
LebInt − so, LebInt − go

This types of Smolyak’s sparse grids are better than those based on
Clenshaw-Curtis for regular functions and small dimensions
Need theorems about number of nodes and accuracy for these n/n + 2 nested
based sparse grids, counterpart of well-known results for Clenshaw-Curtis
sparse grids
1
n(Q d ) ' 2l−1 (l − 1)d−1
l (d − 1)! 2d−1
|R(Qlq )[f ]| ≤ cr ,d n−rd log (n(Q d ) )(d−1)(r +1) ||f ||∞ ∀f ∈ Fdr
(Ql ) l
−r /d
|R(Qlq )[f ]| ≤ cr ,d n log (n(Q d ) )(d−1)(r /d+1) ||f ||∞ ∀f ∈ Cdr
(Qld ) l

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 24 / 26
 Conclusions

Outline

1 Introduction

2 Interpolatory quadrature

3 Accuracy of interpolatory rules. Proposed 1D rules

4 Tests for classical 1D functions

5 Extension to nD functions via Smolyak’s sparse grids

6 Test for nD functions

7 Conclusions

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 25 / 26
 Conclusions

Conclusions

New ways to define n/n + 2 sequences of points for interpolatory quadrature

1D results as good as those obtained with classical Leja sequences

Extension to nD through Smolyak’s sparse grids

Superiority of n/n + 1 n/n + 2 nested quadratures over Clenshaw-Curtis for
low dimensions and regular functions. To be analysed

J. Dumont– Le Brazidec, J. Peter (ONERA) 1D and nD strongly nested interpolatory quadrature June 20, 2018 26 / 26