The (fast) component-by-component construction of lattice point sets and polynomial lattice point sets is a powerful method to obtain quadrature rules for approximating integrals over the dimensional unit cube.
In this talk, we present modifications of the component-by-component algorithm and of the more recent successive coordinate search algorithm, which yield savings of the construction cost for lattice rules and polynomial lattice rules in weighted function spaces. The idea is to reduce the size of the search space for coordinates which are associated with small weights and are therefore of less importance to the overall error compared to coordinates associated with large
weights. We analyze tractability conditions of the resulting quasi-Monte Carlo rules, and show some numerical results.
Separation of Lanthanides/ Lanthanides and Actinides
Ā
QMC: Transition Workshop - Reduced Component-by-Component Constructions of (Polynomial) Lattice Points - Peter Kritzer, May 9, 2018
1. Modiļ¬ed component-by-component
constructions of (polynomial) lattice points
Peter Kritzer
Johann Radon Institute for Computational and Applied Mathematics (RICAM)
Austrian Academy of Sciences
Linz, Austria
Joint work with
J. Dick (UNSW Sydney), A. Ebert (KU Leuven),
G. Leobacher (KFU Graz), F. Pillichshammer (JKU Linz)
SAMSI QMC Transition Workshop, May 9, 2018
Research supported by the Austrian Science Fund, Project F5506-N26
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 1
2. Johann Radon Institute for Computational and Applied Mathematics
Introduction and Motivation
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 2
3. Johann Radon Institute for Computational and Applied Mathematics
Consider integration of functions on [0, 1]d
,
Id (f) =
[0,1]d
f(x) dx ,
where f ā H, and H is some Banach space.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 3
4. Johann Radon Institute for Computational and Applied Mathematics
Consider integration of functions on [0, 1]d
,
Id (f) =
[0,1]d
f(x) dx ,
where f ā H, and H is some Banach space.
Approximate Id by a quasi-Monte Carlo (QMC) rule,
Id (f) ā QN,d (f) =
1
N
Nā1
k=0
f(xk ),
where PN = {x0, . . . , xNā1}.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 3
5. Johann Radon Institute for Computational and Applied Mathematics
Consider integration of functions on [0, 1]d
,
Id (f) =
[0,1]d
f(x) dx ,
where f ā H, and H is some Banach space.
Approximate Id by a quasi-Monte Carlo (QMC) rule,
Id (f) ā QN,d (f) =
1
N
Nā1
k=0
f(xk ),
where PN = {x0, . . . , xNā1}.
Both parameters d and N can be (very) large.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 3
6. Johann Radon Institute for Computational and Applied Mathematics
Worst case error in Banach space H with respect to
PN = {x0, . . . , xNā1} :
eN,d (H, PN) := sup
fāH, f ā¤1
|Id (f) ā QN,d (f)| .
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 4
7. Johann Radon Institute for Computational and Applied Mathematics
Worst case error in Banach space H with respect to
PN = {x0, . . . , xNā1} :
eN,d (H, PN) := sup
fāH, f ā¤1
|Id (f) ā QN,d (f)| .
Need PN that makes eN,d (H, PN) small: How can we ļ¬nd it?
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 4
8. Johann Radon Institute for Computational and Applied Mathematics
Example of function space:
Weighted Korobov space (Hd,Ī±,Ī³, Ā· d,Ī±,Ī³): space of continuous
functions f, where
f 2
d,Ī±,Ī³ =
hāZd
ĻĪ±,Ī³(h) |f(h)|2
,
where f(h)is the h-th Fourier coefļ¬cient of f.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 5
9. Johann Radon Institute for Computational and Applied Mathematics
Example of function space:
Weighted Korobov space (Hd,Ī±,Ī³, Ā· d,Ī±,Ī³): space of continuous
functions f, where
f 2
d,Ī±,Ī³ =
hāZd
ĻĪ±,Ī³(h) |f(h)|2
,
where f(h)is the h-th Fourier coefļ¬cient of f.
Here, ĻĪ±,Ī³(h) moderates the decay of the Fourier coefļ¬cients,
depending on:
Ī± > 1: āsmoothness parameterā (higher Ī± ā smoother functions
in Hd,Ī±,Ī³),
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 5
10. Johann Radon Institute for Computational and Applied Mathematics
Example of function space:
Weighted Korobov space (Hd,Ī±,Ī³, Ā· d,Ī±,Ī³): space of continuous
functions f, where
f 2
d,Ī±,Ī³ =
hāZd
ĻĪ±,Ī³(h) |f(h)|2
,
where f(h)is the h-th Fourier coefļ¬cient of f.
Here, ĻĪ±,Ī³(h) moderates the decay of the Fourier coefļ¬cients,
depending on:
Ī± > 1: āsmoothness parameterā (higher Ī± ā smoother functions
in Hd,Ī±,Ī³),
Ī³ = (Ī³u)uā{1,...,d}: coordinate weights.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 5
11. Johann Radon Institute for Computational and Applied Mathematics
Weights?
Sloan and WoĀ“zniakowski (1998): Assign weights to different
groups of coordinates to model their different inļ¬uence on a
problem:
Ī³ = (Ī³u)uā{1,...,d}
of positive reals: weights.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 6
12. Johann Radon Institute for Computational and Applied Mathematics
Weights?
Sloan and WoĀ“zniakowski (1998): Assign weights to different
groups of coordinates to model their different inļ¬uence on a
problem:
Ī³ = (Ī³u)uā{1,...,d}
of positive reals: weights.
Suitable weights can help to reduce negative inļ¬uence of the
dimension.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 6
13. Johann Radon Institute for Computational and Applied Mathematics
Weights?
Sloan and WoĀ“zniakowski (1998): Assign weights to different
groups of coordinates to model their different inļ¬uence on a
problem:
Ī³ = (Ī³u)uā{1,...,d}
of positive reals: weights.
Suitable weights can help to reduce negative inļ¬uence of the
dimension.
Important class of weights: product weights
Ī³u =
jāu
Ī³j
for positive reals Ī³j . In this case, assume
1 = Ī³1, and Ī³j ā„ Ī³j+1 for all j.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 6
14. Johann Radon Institute for Computational and Applied Mathematics
Here:
PN = {x0, . . . , xNā1} is a lattice point set with generating vector
z = (z1, . . . , zd ) ā {1, . . . , N ā 1}d
.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 7
15. Johann Radon Institute for Computational and Applied Mathematics
Here:
PN = {x0, . . . , xNā1} is a lattice point set with generating vector
z = (z1, . . . , zd ) ā {1, . . . , N ā 1}d
.
Points of PN:
xn = (xn,1, . . . , xn,d )
with
xn,j =
nzj
N
,
where {t} = t ā t .
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 7
16. Johann Radon Institute for Computational and Applied Mathematics
Here:
PN = {x0, . . . , xNā1} is a lattice point set with generating vector
z = (z1, . . . , zd ) ā {1, . . . , N ā 1}d
.
Points of PN:
xn = (xn,1, . . . , xn,d )
with
xn,j =
nzj
N
,
where {t} = t ā t .
Note: Given N and d, z fully determines the lattice point set.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 7
17. Johann Radon Institute for Computational and Applied Mathematics
Lattice point set with d = 2, N = 34, and z = (1, 21):
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 8
18. Johann Radon Institute for Computational and Applied Mathematics
Explicit formula for the (squared) worst-case error,
e2
N,d,Ī±,Ī³(z) = ā1 +
1
N
Nā1
n=0
d
j=1
1 + Ī³j ĻĪ±
nzj
N
,
where ĻĪ±
k
N can be computed for all values of k = 0, . . . , N ā 1.
The error formula is easy to implement.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 9
19. Johann Radon Institute for Computational and Applied Mathematics
Question: is the Korobov space interesting?
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 10
20. Johann Radon Institute for Computational and Applied Mathematics
Question: is the Korobov space interesting?
Yes, ļ¬rst reason: it is a model function space that makes it easier
to understand how lattice rules work.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 10
21. Johann Radon Institute for Computational and Applied Mathematics
Question: is the Korobov space interesting?
Yes, ļ¬rst reason: it is a model function space that makes it easier
to understand how lattice rules work.
Yes, second reason:
Bounds on the worst-case error of lattice rules in the Korobov
space with Ī± = 2 immediately yield bounds on the worst-case
error of slightly modiļ¬ed lattice rules in certain Sobolev spaces.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 10
22. Johann Radon Institute for Computational and Applied Mathematics
All that remains is to ļ¬nd āgoodā z ā {1, . . . , N ā 1}d
.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 11
23. Johann Radon Institute for Computational and Applied Mathematics
All that remains is to ļ¬nd āgoodā z ā {1, . . . , N ā 1}d
.
Huge search space of size (N ā 1)d
. (e.g., N = 10 000 and
d = 100).
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 11
24. Johann Radon Institute for Computational and Applied Mathematics
All that remains is to ļ¬nd āgoodā z ā {1, . . . , N ā 1}d
.
Huge search space of size (N ā 1)d
. (e.g., N = 10 000 and
d = 100).
Component by component (CBC) construction: greedy algorithm
to construct zj one at a time.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 11
25. Johann Radon Institute for Computational and Applied Mathematics
All that remains is to ļ¬nd āgoodā z ā {1, . . . , N ā 1}d
.
Huge search space of size (N ā 1)d
. (e.g., N = 10 000 and
d = 100).
Component by component (CBC) construction: greedy algorithm
to construct zj one at a time.
Size of search space is N ā 1 per component.
(Almost) optimal convergence of error bounds.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 11
26. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 1 (CBC construction)
Let N be given. Construct z = (z1, . . . , zd ) as follows.
Set z1 = 1.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 12
27. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 1 (CBC construction)
Let N be given. Construct z = (z1, . . . , zd ) as follows.
Set z1 = 1.
For s ā {2, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā {1, . . . , N ā 1} such that
e2
N,s,Ī±,Ī³((z1, . . . , zsā1, zs))
is minimized as a function of zs.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 12
28. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 1 (CBC construction)
Let N be given. Construct z = (z1, . . . , zd ) as follows.
Set z1 = 1.
For s ā {2, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā {1, . . . , N ā 1} such that
e2
N,s,Ī±,Ī³((z1, . . . , zsā1, zs))
is minimized as a function of zs.
Increase s and repeat the second step until (z1, . . . , zd ) is found.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 12
29. Johann Radon Institute for Computational and Applied Mathematics
Can do fast CBC (Cools, Nuyens, 2006): computation cost of
O(dN log N).
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 13
30. Johann Radon Institute for Computational and Applied Mathematics
Can do fast CBC (Cools, Nuyens, 2006): computation cost of
O(dN log N).
Computation cost of O(dN log N) can still be demanding for big
N, d
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 13
31. Johann Radon Institute for Computational and Applied Mathematics
Can do fast CBC (Cools, Nuyens, 2006): computation cost of
O(dN log N).
Computation cost of O(dN log N) can still be demanding for big
N, d
Might want to have big N, d simultaneously.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 13
32. Johann Radon Institute for Computational and Applied Mathematics
Can do fast CBC (Cools, Nuyens, 2006): computation cost of
O(dN log N).
Computation cost of O(dN log N) can still be demanding for big
N, d
Might want to have big N, d simultaneously. ā Can we speed
up the search?
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 13
33. Johann Radon Institute for Computational and Applied Mathematics
The reduced CBC construction
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 14
34. Johann Radon Institute for Computational and Applied Mathematics
Idea: make search space smaller for later components of z.
Let N be a prime power, N = bm
, b prime, m ā N.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 15
35. Johann Radon Institute for Computational and Applied Mathematics
Idea: make search space smaller for later components of z.
Let N be a prime power, N = bm
, b prime, m ā N.
Let w1, . . . , wd , . . . ā N0 with 0 = w1 ā¤ Ā· Ā· Ā· ā¤ wd ā¤ Ā· Ā· Ā· .
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 15
36. Johann Radon Institute for Computational and Applied Mathematics
Idea: make search space smaller for later components of z.
Let N be a prime power, N = bm
, b prime, m ā N.
Let w1, . . . , wd , . . . ā N0 with 0 = w1 ā¤ Ā· Ā· Ā· ā¤ wd ā¤ Ā· Ā· Ā· .
Consider the sequence of reduced search spaces
ZN,wj
:=
{1 ā¤ z < bmāwj : gcd(z, N) = 1} if wj < m,
{1} if wj ā„ m.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 15
37. Johann Radon Institute for Computational and Applied Mathematics
Idea: make search space smaller for later components of z.
Let N be a prime power, N = bm
, b prime, m ā N.
Let w1, . . . , wd , . . . ā N0 with 0 = w1 ā¤ Ā· Ā· Ā· ā¤ wd ā¤ Ā· Ā· Ā· .
Consider the sequence of reduced search spaces
ZN,wj
:=
{1 ā¤ z < bmāwj : gcd(z, N) = 1} if wj < m,
{1} if wj ā„ m.
Note that
|ZN,wj
| :=
bmāwj ā1
(b ā 1) if wj < m,
1 if wj ā„ m.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 15
38. Johann Radon Institute for Computational and Applied Mathematics
Idea: make search space smaller for later components of z.
Let N be a prime power, N = bm
, b prime, m ā N.
Let w1, . . . , wd , . . . ā N0 with 0 = w1 ā¤ Ā· Ā· Ā· ā¤ wd ā¤ Ā· Ā· Ā· .
Consider the sequence of reduced search spaces
ZN,wj
:=
{1 ā¤ z < bmāwj : gcd(z, N) = 1} if wj < m,
{1} if wj ā„ m.
Note that
|ZN,wj
| :=
bmāwj ā1
(b ā 1) if wj < m,
1 if wj ā„ m.
write Yj := bwj .
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 15
39. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 2 (Reduced CBC construction)
Let N, w1, . . . , wd , and Y1, . . . , Yd be as above. Construct
z = (Y1z1, . . . , Yd zd ) as follows.
Set z1 = 1.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 16
40. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 2 (Reduced CBC construction)
Let N, w1, . . . , wd , and Y1, . . . , Yd be as above. Construct
z = (Y1z1, . . . , Yd zd ) as follows.
Set z1 = 1.
For s ā {2, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā ZN,ws
such that
e2
N,s,Ī±,Ī³((Y1z1, . . . , Ysā1zsā1, Yszs))
is minimized as a function of zs.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 16
41. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 2 (Reduced CBC construction)
Let N, w1, . . . , wd , and Y1, . . . , Yd be as above. Construct
z = (Y1z1, . . . , Yd zd ) as follows.
Set z1 = 1.
For s ā {2, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā ZN,ws
such that
e2
N,s,Ī±,Ī³((Y1z1, . . . , Ysā1zsā1, Yszs))
is minimized as a function of zs.
Increase s and repeat the second step until (Y1z1, . . . , Yd zd ) is
found.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 16
42. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 2 (Reduced CBC construction)
Let N, w1, . . . , wd , and Y1, . . . , Yd be as above. Construct
z = (Y1z1, . . . , Yd zd ) as follows.
Set z1 = 1.
For s ā {2, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā ZN,ws
such that
e2
N,s,Ī±,Ī³((Y1z1, . . . , Ysā1zsā1, Yszs))
is minimized as a function of zs.
Increase s and repeat the second step until (Y1z1, . . . , Yd zd ) is
found.
Usual CBC construction: wj = 0 and Yj = 1 for all j.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 16
43. Johann Radon Institute for Computational and Applied Mathematics
Theorem 3 (Dick/K./Leobacher/Pillichshammer, 2015)
Let z = (Y1z1, . . . , Yd zd ) ā Zd
be constructed according to the
reduced CBC algorithm. Then,
eN,d,Ī±,Ī³((Y1z1, . . . , Yd zd )) ā¤
ā¤ NāĪ±/2+Ī“
ļ£«
ļ£2
d
j=1
1 + Ī³
1
Ī±ā2Ī“
j 2Ī¶ Ī±
Ī±ā2Ī“ bwj
ļ£¶
ļ£ø
Ī±/2āĪ“
for all Ī“ ā 0, Ī±ā1
2 , where Ī¶ is the Riemann zeta function.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 17
44. Johann Radon Institute for Computational and Applied Mathematics
Theorem 3 (Dick/K./Leobacher/Pillichshammer, 2015)
Let z = (Y1z1, . . . , Yd zd ) ā Zd
be constructed according to the
reduced CBC algorithm. Then,
eN,d,Ī±,Ī³((Y1z1, . . . , Yd zd )) ā¤
ā¤ NāĪ±/2+Ī“
ļ£«
ļ£2
d
j=1
1 + Ī³
1
Ī±ā2Ī“
j 2Ī¶ Ī±
Ī±ā2Ī“ bwj
ļ£¶
ļ£ø
Ī±/2āĪ“
for all Ī“ ā 0, Ī±ā1
2 , where Ī¶ is the Riemann zeta function.
Theorem formulated for product weights, similar result holds for
general weights.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 17
45. Johann Radon Institute for Computational and Applied Mathematics
If
B :=
ā
j=1
Ī³
1
Ī±ā2Ī“
j bwj
< ā,
then the error can be bounded independently of the dimension.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 18
46. Johann Radon Institute for Computational and Applied Mathematics
The reduced fast CBC construction
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 19
47. Johann Radon Institute for Computational and Applied Mathematics
Fast CBC construction (Nuyens/Cools) for non-reduced case
(wj = 0) has computation cost of O(dN log N).
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 20
48. Johann Radon Institute for Computational and Applied Mathematics
Fast CBC construction (Nuyens/Cools) for non-reduced case
(wj = 0) has computation cost of O(dN log N).
Idea also works for the reduced case (assuming product weights;
POD weights: work in progress), yields reduced cost by
exploiting additional structure of the case wj > 0.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 20
49. Johann Radon Institute for Computational and Applied Mathematics
Fast CBC construction (Nuyens/Cools) for non-reduced case
(wj = 0) has computation cost of O(dN log N).
Idea also works for the reduced case (assuming product weights;
POD weights: work in progress), yields reduced cost by
exploiting additional structure of the case wj > 0.
Bonus: once wj ā„ m the search space contains only one
element, construction of additional components zj incurs no extra
cost.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 20
50. Johann Radon Institute for Computational and Applied Mathematics
Fast CBC construction (Nuyens/Cools) for non-reduced case
(wj = 0) has computation cost of O(dN log N).
Idea also works for the reduced case (assuming product weights;
POD weights: work in progress), yields reduced cost by
exploiting additional structure of the case wj > 0.
Bonus: once wj ā„ m the search space contains only one
element, construction of additional components zj incurs no extra
cost.
Computational cost of the reduced fast CBC construction is
O
ļ£«
ļ£N log N + min{d, dā
}N +
min{d,dā
}
j=1
(m ā wj )Nbāwj
ļ£¶
ļ£ø ,
where dā
:= max{j ā N : wj < m}.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 20
51. Johann Radon Institute for Computational and Applied Mathematics
Example:
Suppose (product) weights Ī³j are Ī³j = jā3
.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 21
52. Johann Radon Institute for Computational and Applied Mathematics
Example:
Suppose (product) weights Ī³j are Ī³j = jā3
.
Fast CBC construction needs O(dmbm
) operations to compute a
generating vector for which the worst-case error is bounded
independently of the dimension.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 21
53. Johann Radon Institute for Computational and Applied Mathematics
Example:
Suppose (product) weights Ī³j are Ī³j = jā3
.
Fast CBC construction needs O(dmbm
) operations to compute a
generating vector for which the worst-case error is bounded
independently of the dimension.
Reduced fast CBC construction: choose, e.g., wj = 3
2 logb j .
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 21
54. Johann Radon Institute for Computational and Applied Mathematics
Example:
Suppose (product) weights Ī³j are Ī³j = jā3
.
Fast CBC construction needs O(dmbm
) operations to compute a
generating vector for which the worst-case error is bounded
independently of the dimension.
Reduced fast CBC construction: choose, e.g., wj = 3
2 logb j .
We need O(mbm
+ min{d, dā
}mbm
) operations to compute a
generating vector for which the worst-case error is still bounded
independently of the dimension, as
j
Ī³j bwj
< Ī¶(3/2) < ā.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 21
55. Johann Radon Institute for Computational and Applied Mathematics
Example:
Suppose (product) weights Ī³j are Ī³j = jā3
.
Fast CBC construction needs O(dmbm
) operations to compute a
generating vector for which the worst-case error is bounded
independently of the dimension.
Reduced fast CBC construction: choose, e.g., wj = 3
2 logb j .
We need O(mbm
+ min{d, dā
}mbm
) operations to compute a
generating vector for which the worst-case error is still bounded
independently of the dimension, as
j
Ī³j bwj
< Ī¶(3/2) < ā.
Reduced fast CBC construction signiļ¬cantly reduces
computation cost.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 21
56. Johann Radon Institute for Computational and Applied Mathematics
The successive coordinate search (SCS) algorithm
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 22
57. Johann Radon Institute for Computational and Applied Mathematics
Question: can one improve on the quality of the output vector of the
CBC algorithm?
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 23
58. Johann Radon Institute for Computational and Applied Mathematics
Question: can one improve on the quality of the output vector of the
CBC algorithm?
Ebert/Leƶvey/Nuyens (2016): successive coordinate search (SCS)
algorithm.
Basic idea:
Assume product weights.
Begin with a start vector z(0)
= (z
(0)
1 , . . . , z
(0)
d ).
Update components one after the other by minimizing worst-case
error.
Returned vector is at least as good as initial vector.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 23
59. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 4 (SCS construction)
Let N be a prime. Let z(0)
= (z
(0)
1 , . . . , z
(0)
d ) ā {0, 1, . . . , N ā 1}d
be
given.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 24
60. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 4 (SCS construction)
Let N be a prime. Let z(0)
= (z
(0)
1 , . . . , z
(0)
d ) ā {0, 1, . . . , N ā 1}d
be
given.
For s ā {1, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā {1, . . . , N ā 1} such that
e2
N,d,Ī±,Ī³((z1, . . . , zsā1, zs, z
(0)
s+1, . . . , z
(0)
d ))
is minimized as a function of zs.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 24
61. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 4 (SCS construction)
Let N be a prime. Let z(0)
= (z
(0)
1 , . . . , z
(0)
d ) ā {0, 1, . . . , N ā 1}d
be
given.
For s ā {1, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā {1, . . . , N ā 1} such that
e2
N,d,Ī±,Ī³((z1, . . . , zsā1, zs, z
(0)
s+1, . . . , z
(0)
d ))
is minimized as a function of zs.
Increase s and repeat the second step until z = (z1, . . . , zd ) is
found.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 24
62. Johann Radon Institute for Computational and Applied Mathematics
Theorem 5 (Ebert/Leƶvey/Nuyens, 2016)
Let z(1)
be constructed by the fast CBC algorithm.
Set z(0)
:= z(1)
in the SCS algorithm.
Let z be the vector returned by the SCS algorithm. Then
e2
N,d,Ī±,Ī³(z) ā¤ e2
N,d,Ī±,Ī³(z(1)
).
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 25
63. Johann Radon Institute for Computational and Applied Mathematics
Theorem 5 (Ebert/Leƶvey/Nuyens, 2016)
Let z(1)
be constructed by the fast CBC algorithm.
Set z(0)
:= z(1)
in the SCS algorithm.
Let z be the vector returned by the SCS algorithm. Then
e2
N,d,Ī±,Ī³(z) ā¤ e2
N,d,Ī±,Ī³(z(1)
).
Theorem 6 (Ebert/Leƶvey/Nuyens, 2016)
Set z(0)
:= (0, 0, . . . , 0) in the SCS algorithm.
Then the SCS algorithm yields the same result as the usual CBC
algorithm. The SCS algorithm is thus a generalization of the CBC
algorithm.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 25
64. Johann Radon Institute for Computational and Applied Mathematics
The SCS algorithm yields improvements over the CBC algorithm
for pre-asymptotically moderately decreasing weights, e.g.
Ī³j = 0.95j
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 26
65. Johann Radon Institute for Computational and Applied Mathematics
The SCS algorithm yields improvements over the CBC algorithm
for pre-asymptotically moderately decreasing weights, e.g.
Ī³j = 0.95j
There is a fast implementation of the SCS algorithm:
Pre-computation with cost of O(dN),
matrix-vector multiplication of same speed as in CBC algorithm,
total cost of O(dN log N).
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 26
66. Johann Radon Institute for Computational and Applied Mathematics
The reduced fast SCS algorithm
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 27
67. Johann Radon Institute for Computational and Applied Mathematics
Joint work with A. Ebert (2017/18):
Combine advantages of reduced fast CBC construction and fast SCS
construction:
reduced fast SCS construction.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 28
68. Johann Radon Institute for Computational and Applied Mathematics
Same setting as for reduced fast CBC construction, i.e.,
Let N be a prime power, N = bm
, b prime, m ā N.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 29
69. Johann Radon Institute for Computational and Applied Mathematics
Same setting as for reduced fast CBC construction, i.e.,
Let N be a prime power, N = bm
, b prime, m ā N.
Let w1, . . . , wd , . . . ā N0 with 0 = w1 ā¤ Ā· Ā· Ā· ā¤ wd ā¤ Ā· Ā· Ā· .
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 29
70. Johann Radon Institute for Computational and Applied Mathematics
Same setting as for reduced fast CBC construction, i.e.,
Let N be a prime power, N = bm
, b prime, m ā N.
Let w1, . . . , wd , . . . ā N0 with 0 = w1 ā¤ Ā· Ā· Ā· ā¤ wd ā¤ Ā· Ā· Ā· .
Consider the sequence of reduced search spaces
ZN,wj
:=
{1 ā¤ z < bmāwj : gcd(z, N) = 1} if wj < m,
{1} if wj ā„ m.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 29
71. Johann Radon Institute for Computational and Applied Mathematics
Same setting as for reduced fast CBC construction, i.e.,
Let N be a prime power, N = bm
, b prime, m ā N.
Let w1, . . . , wd , . . . ā N0 with 0 = w1 ā¤ Ā· Ā· Ā· ā¤ wd ā¤ Ā· Ā· Ā· .
Consider the sequence of reduced search spaces
ZN,wj
:=
{1 ā¤ z < bmāwj : gcd(z, N) = 1} if wj < m,
{1} if wj ā„ m.
write Yj := bwj .
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 29
72. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 7 (Reduced SCS construction)
Let N be a prime power. Let
z(0)
= (z
(0)
1 , . . . , z
(0)
d ) ā {0, 1, . . . , N ā 1}d
be given.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 30
73. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 7 (Reduced SCS construction)
Let N be a prime power. Let
z(0)
= (z
(0)
1 , . . . , z
(0)
d ) ā {0, 1, . . . , N ā 1}d
be given.
For s ā {1, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā ZN,ws
such that
e2
N,d,Ī±,Ī³((Y1z1, . . . , Ysā1zsā1, Yszs, z
(0)
s+1, . . . , z
(0)
d ))
is minimized as a function of zs.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 30
74. Johann Radon Institute for Computational and Applied Mathematics
Algorithm 7 (Reduced SCS construction)
Let N be a prime power. Let
z(0)
= (z
(0)
1 , . . . , z
(0)
d ) ā {0, 1, . . . , N ā 1}d
be given.
For s ā {1, . . . , d} assume that z1, . . . , zsā1 have already been
found. Now choose zs ā ZN,ws
such that
e2
N,d,Ī±,Ī³((Y1z1, . . . , Ysā1zsā1, Yszs, z
(0)
s+1, . . . , z
(0)
d ))
is minimized as a function of zs.
Increase s and repeat the second step until z = (Y1z1, . . . , Yd zd )
is found.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 30
75. Johann Radon Institute for Computational and Applied Mathematics
Theorem 8 (Ebert/K., 2018)
Assume product weights and let z = (Y1z1, . . . , Yd zd ) be constructed
by Algorithm 7. Then we have for all Ī“ ā (0, Ī±ā1
2 ] that
eN,d,Ī±,Ī³(z) ā¤ Cd,Ī±,Ī³,Ī“ NāĪ±/2+Ī“
,
where Cd,Ī±,Ī³,Ī“ is bounded independently of d if
ā
j=1
Ī³
1
Ī±ā2Ī“
j bwj
< ā.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 31
76. Johann Radon Institute for Computational and Applied Mathematics
Theorem 8 (Ebert/K., 2018)
Assume product weights and let z = (Y1z1, . . . , Yd zd ) be constructed
by Algorithm 7. Then we have for all Ī“ ā (0, Ī±ā1
2 ] that
eN,d,Ī±,Ī³(z) ā¤ Cd,Ī±,Ī³,Ī“ NāĪ±/2+Ī“
,
where Cd,Ī±,Ī³,Ī“ is bounded independently of d if
ā
j=1
Ī³
1
Ī±ā2Ī“
j bwj
< ā.
Theorem formulated for product weights, similar result holds for
general weights.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 31
77. Johann Radon Institute for Computational and Applied Mathematics
Special case of reduced SCS construction:
Set w1 = w2 = Ā· Ā· Ā· = wd = 0. Then the new result generalizes the
previous SCS construction with respect to
General weights instead of product weights,
prime power N instead of prime N.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 32
78. Johann Radon Institute for Computational and Applied Mathematics
Reduced fast SCS construction:
Assume product weights,
Assume initial vector of the form z(0)
= (Y1z
(0)
1 , . . . , Yd z
(0)
d ) with
z
(0)
j ā ZN,wj
.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 33
79. Johann Radon Institute for Computational and Applied Mathematics
Reduced fast SCS construction:
Assume product weights,
Assume initial vector of the form z(0)
= (Y1z
(0)
1 , . . . , Yd z
(0)
d ) with
z
(0)
j ā ZN,wj
.
Use fast pre-computation (due to structure of z(0)
) and fast
matrix-vector multiplication (as in reduced fast CBC
construction).
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 33
80. Johann Radon Institute for Computational and Applied Mathematics
Reduced fast SCS construction:
Assume product weights,
Assume initial vector of the form z(0)
= (Y1z
(0)
1 , . . . , Yd z
(0)
d ) with
z
(0)
j ā ZN,wj
.
Use fast pre-computation (due to structure of z(0)
) and fast
matrix-vector multiplication (as in reduced fast CBC
construction).
Implementation with overall cost
O
ļ£«
ļ£N log N + min{d, dā
}N +
min{d,dā
}
j=1
(m ā wj )Nbāwj
ļ£¶
ļ£ø ,
where dā
:= max{j ā N : wj < m}.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 33
81. Johann Radon Institute for Computational and Applied Mathematics
Numerical results
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 34
82. Johann Radon Institute for Computational and Applied Mathematics
Computation times
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 35
83. Johann Radon Institute for Computational and Applied Mathematics
Computation times (in seconds):
unreduced (normal font) and reduced CBC (bold font)
Ī± = 2, b = 2, Ī³j = (0.7)j
, wj = 3 logb j .
d = 50 d = 100 d = 500 d = 1000 d = 2000
m = 10
0.0173 0.0329 0.16 0.323 0.636
0.00298 0.00206 0.00218 0.00222 0.00241
m = 12
0.0256 0.0481 0.241 0.48 0.953
0.00358 0.00365 0.0037 0.00354 0.00439
m = 14
0.0469 0.0851 0.438 0.856 1.88
0.00803 0.00761 0.0105 0.00712 0.00747
m = 16
0.14 0.239 1.33 2.49 5.05
0.0237 0.0233 0.0233 0.0227 0.0251
m = 18
0.443 0.832 4.44 8.54 17.1
0.0798 0.0897 0.0915 0.091 0.09
m = 20
2.17 4.17 21.5 42.4 84.3
0.38 0.623 0.643 0.636 0.628
Intel Core i5-2400S CPU with 2.5GHz using Matlab.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 36
84. Johann Radon Institute for Computational and Applied Mathematics
Computation times (in seconds):
unreduced (normal font) and reduced SCS (bold font)
Ī± = 2, b = 2, Ī³j = (0.7)j
, wj = 3 logb j
d = 50 d = 100 d = 500 d = 1000 d = 2000
m = 10
0.0275 0.0516 0.256 0.516 1.03
0.00408 0.00327 0.00354 0.00347 0.00329
m = 12
0.0418 0.0751 0.383 0.756 1.56
0.00592 0.00504 0.00612 0.00516 0.00794
m = 14
0.0792 0.14 0.767 1.39 2.82
0.014 0.0136 0.0163 0.0138 0.0138
m = 16
0.204 0.388 2.09 4.05 8.04
0.0441 0.0434 0.0434 0.0423 0.0462
m = 18
0.686 1.35 6.89 13.7 26.8
0.16 0.177 0.182 0.183 0.187
m = 20
3.28 6.71 34.4 67.4 132
0.843 1.4 1.51 1.37 1.36
Intel Core i5-2400S CPU with 2.5GHz using Matlab.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 37
85. Johann Radon Institute for Computational and Applied Mathematics
Errors
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 38
86. Johann Radon Institute for Computational and Applied Mathematics
Logarithmic errors for
d = 100, b = 2, Ī± = 2, Ī³j = 0.7j
, wj = 2 logb j :
zCBC zSCS zred
CBC zred
SCS
m = 6 -0.9858 -0.9874 -0.7760 -0.7817
m = 7 -1.6268 -1.6255 -1.4292 -1.4425
m = 8 -2.2860 -2.2810 -2.0997 -2.1021
m = 9 -2.9535 -2.9482 -2.7784 -2.7745
m = 10 -3.6389 -3.6271 -3.4536 -3.4631
For SCS errors, we consider the average of 100 runs of the
algorithms with random starting vectors.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 39
87. Johann Radon Institute for Computational and Applied Mathematics
Convergence rates
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 40
88. Johann Radon Institute for Computational and Applied Mathematics
101
102
103
104
105
106
10-6
10-5
10-4
10-3
10-2
10-1
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 41
89. Johann Radon Institute for Computational and Applied Mathematics
101
102
103
104
105
106
10-6
10-5
10-4
10-3
10-2
10-1
100
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 42
90. Johann Radon Institute for Computational and Applied Mathematics
Polynomial lattice rules:
Polynomial lattice rules: similar to lattice rules, but integer
arithmetic replaced by polynomial arithmetic over ļ¬nite ļ¬elds.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 43
91. Johann Radon Institute for Computational and Applied Mathematics
Polynomial lattice rules:
Polynomial lattice rules: similar to lattice rules, but integer
arithmetic replaced by polynomial arithmetic over ļ¬nite ļ¬elds.
Special cases of Niederreiterās (t, m, d)-nets, powerful QMC
methods.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 43
92. Johann Radon Institute for Computational and Applied Mathematics
Polynomial lattice rules:
Polynomial lattice rules: similar to lattice rules, but integer
arithmetic replaced by polynomial arithmetic over ļ¬nite ļ¬elds.
Special cases of Niederreiterās (t, m, d)-nets, powerful QMC
methods.
All results shown above (CBC, fast CBC, reduced fast CBC,
SCS, fast SCS, reduced fast SCS) work analogously for
polynomial lattice rules.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 43
93. Johann Radon Institute for Computational and Applied Mathematics
Conclusion
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 44
94. Johann Radon Institute for Computational and Applied Mathematics
Component-by-component (CBC) construction of (polynomial)
lattice points is a standard method in modern QMC theory.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 45
95. Johann Radon Institute for Computational and Applied Mathematics
Component-by-component (CBC) construction of (polynomial)
lattice points is a standard method in modern QMC theory.
Fast CBC construction by Nuyens/Cools needs O(dN log N)
operations.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 45
96. Johann Radon Institute for Computational and Applied Mathematics
Component-by-component (CBC) construction of (polynomial)
lattice points is a standard method in modern QMC theory.
Fast CBC construction by Nuyens/Cools needs O(dN log N)
operations.
We can reduce the construction cost depending on the
coordinate weights, sometimes obtain independence of the
dimension.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 45
97. Johann Radon Institute for Computational and Applied Mathematics
Component-by-component (CBC) construction of (polynomial)
lattice points is a standard method in modern QMC theory.
Fast CBC construction by Nuyens/Cools needs O(dN log N)
operations.
We can reduce the construction cost depending on the
coordinate weights, sometimes obtain independence of the
dimension.
The (fast) SCS algorithm can improve on the quality of
generating vectors, as compared to the CBC construction.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 45
98. Johann Radon Institute for Computational and Applied Mathematics
Component-by-component (CBC) construction of (polynomial)
lattice points is a standard method in modern QMC theory.
Fast CBC construction by Nuyens/Cools needs O(dN log N)
operations.
We can reduce the construction cost depending on the
coordinate weights, sometimes obtain independence of the
dimension.
The (fast) SCS algorithm can improve on the quality of
generating vectors, as compared to the CBC construction.
We can reduce the run-time of the SCS algorithm, similarly to
that of the CBC construction.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 45
99. Johann Radon Institute for Computational and Applied Mathematics
Component-by-component (CBC) construction of (polynomial)
lattice points is a standard method in modern QMC theory.
Fast CBC construction by Nuyens/Cools needs O(dN log N)
operations.
We can reduce the construction cost depending on the
coordinate weights, sometimes obtain independence of the
dimension.
The (fast) SCS algorithm can improve on the quality of
generating vectors, as compared to the CBC construction.
We can reduce the run-time of the SCS algorithm, similarly to
that of the CBC construction.
Numerical results demonstrate the effect of the reduced
approach.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 45
100. Johann Radon Institute for Computational and Applied Mathematics
Component-by-component (CBC) construction of (polynomial)
lattice points is a standard method in modern QMC theory.
Fast CBC construction by Nuyens/Cools needs O(dN log N)
operations.
We can reduce the construction cost depending on the
coordinate weights, sometimes obtain independence of the
dimension.
The (fast) SCS algorithm can improve on the quality of
generating vectors, as compared to the CBC construction.
We can reduce the run-time of the SCS algorithm, similarly to
that of the CBC construction.
Numerical results demonstrate the effect of the reduced
approach.
Decision which algorithm to use when depends on parameter
settings.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 45
101. Johann Radon Institute for Computational and Applied Mathematics
Thanks for your attention.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 46
102. Johann Radon Institute for Computational and Applied Mathematics
J. Dick, P. Kritzer. On a projection-corrected component-by-component
construction. J. Complexity 32, 74ā80, 2016.
J. Dick, P. Kritzer, G. Leobacher, F. Pillichshammer. A reduced fast
component-by-component construction of lattice points for integration in weighted
spaces with fast decreasing weights. J. Comput. Appl. Math. 276, 1ā15, 2015.
A. Ebert, P. Kritzer. Constructing lattice points for numerical integration by a
reduced fast successive coordinate search algorithm. Submitted, 2018.
A. Ebert, H. Leƶvey, D. Nuyens. Successive Coordinate Search and
Component-by-Component Construction of Rank-1 Lattice Rules. To appear in:
P. Glynn, A. Owen (eds.), Monte Carlo and Quasi-Monte Carlo Methods 2016,
Springer, 2018.
H. Laimer. On combined component-by-component constructions of lattice point
sets. J. Complexity 38, 22ā30, 2017.
D. Nuyens, R. Cools. Fast algorithms for component-by-component construction
of rank-1 lattice rules in shift-invariant reproducing kernel Hilbert spaces. Math.
Comp. 75, 903ā920, 2006.
Peter Kritzer Modiļ¬ed component-by-component constructions of (polynomial) lattice points 47