orthogonal array sampling for monte carlo rendering...
TRANSCRIPT
Eurographics Symposium on Rendering 2019T. Boubekeur and P. Sen(Guest Editors)
Volume 38 (2019), Number 4
Orthogonal Array Sampling for Monte Carlo Rendering:supplemental document
Wojciech Jarosz Afnan Enayet Andrew Kensler Charlie Kilpatrick Per Christensen
Dartmouth College Pixar Animation Studios
1. Further reading list
Below is our suggested sequence through a subset of the OA litera-ture for graphics researchers to most quickly get ramped up on thesetopics.
Graphics
1. First become familiar with Chiu, Shirley, and Wang’s originalmulti-jittered sampling paper [CSW94] as it is the conceptualprecursor to orthogonal arrays.
2. Then read Kensler’s correlated multi-jittered tech report [Ken13].While the title suggests the main contribution is to incorporatefavorable correlations, more importantly, this paper shows howto construct multi-jittered samples in-place, one sample at a timewithout requiring any storage.
Orthogonal array basics
3. The preface and first chapter of Hedayat, Sloane, and Stufken[HSS99] provides a great non-technical introduction to orthogo-nal arrays from the experimental design perspective.
4. Sections 10.3 and 10.4 of Owen [Owe13] provide a great MonteCarlo integration-centric introduction to Latin hypercubes andorthogonal arrays. We suggest reading it before diving into moredetail in the original publications.
5. [Tan93] first introduced the idea of enforcing both OA stratifica-tion and Latin hypercube stratification with “orthogonal array-based Latin hypercube designs” also called “U-designs” (similarto Chiu, Shirley, and Wang’s multi-jittered stratification, but forpoints of arbitrary dimension).
6. Many more general OA construction routines require arithmeticon finite/Galois fields. This is more involved than our routinesusing modular arithmetic, but many open-source libraries exist(particularly for binary finite fields) since these are used in cryp-tography. Hedayat, Sloane, and Stufken [HSS99] provides a goodintroduction to Galois fields within the context of orthogonalarrays.
Additional avenues
After becoming familiar with the basics from the main paper and thereferences above, there are several promising off-shoots that may beuseful for graphics:
7. Strong orthogonal arrays: These enforce stratification in alldimensional projections, and are strongly related to (t,m,s)-netsfrom QMC. [HT12] originally proposed them, and there has beenactive work in this area since then [HCT18; HT14; LL15; Wen14].
8. Nested orthogonal arrays/Latin hypercube designs: Theseare a stepping stone to progressive sample sequences. They al-low creating two or more layers of nested point sets [HHQ16;AZ16; HQ10; HQ11; QA10; WL13; ZWD19; RHVD10; Qia09;QAW09; YLL14; SLQ14].
9. Correlation-controlled Latin hypercubes/orthogonal arrays:Instead of independently randomizing the factors and levels ofan orthogonal array, variance can be reduced by controlling thecorrelations [Owe94; CQ14; Tan98]. This is the same idea usedby correlated multi-jittered sampling [Ken13].
2. Additional and alternate code listings
Here we include alternate code listings and additional visualizationsand variance graphs.
Listing 1: A modified version of Listing 1 from the main documentwhich computes all dimensions (instead of just one) of an arbitrarysample from a Bose OA pattern.
1 vector<float> boseOA(unsigned i, // sample index2 unsigned d, // number of dimensions (<= s+1)3 unsigned s, // number of levels/strata4 unsigned p[], // array of d pseudo-random seeds5 OffsetType ot) { // J, MJ, or CMJ6 vector<float> Xi(d);7 i = permute(i % (s*s), s*s, p[0]);8 unsigned Ai0 = i % s;9 unsigned Ai1 = i / s;
10 unsigned stratumX = permute(Ai0, s, p[0] * 0x51633e2d);11 unsigned stratumY = permute(Ai1, s, p[1] * 0x51633e2d);12 float sstratX = offset(Ai0, Ai1, s, p[0] * 0x68bc21eb, ot);13 float sstratY = offset(Ai1, Ai0, s, p[1] * 0x68bc21eb, ot);14 float jitterX = randfloat(i, p[0] * 0x02e5be93);15 float jitterY = randfloat(i, p[1] * 0x02e5be93);16 Xi[0] = (stratumX + (sstratumX + jitterX) / s) / s;17 Xi[1] = (stratumY + (sstratumY + jitterY) / s) / s;18 for (unsigned j = 2; j < d; ++j) {19 unsigned Aij = (Ai0 + (j-1) * Ai1) % s;20 unsigned j2 = (j % 2) ? j-1 : j+1;21 unsigned Aij2 = (Ai0 + (j2-1) * Ai1) % s;22 unsigned stratumJ = permute(Aij, s, p[j] * 0x51633e2d);23 unsigned sstratJ = offset(Aij, Aij2, s, p[j] * 0x68bc21eb, ot);24 float jitterJ = randfloat(i, p[j] * 0x02e5be93);25 Xi[j] = (stratumJ + (sstratJ + jitterJ) / s) / s;26 }27 return Xi;28 }
submitted to Eurographics Symposium on Rendering (2019)
2 Jarosz, Enayet, Kensler, Kilpatrick, and Christensen / Orthogonal Array Sampling for Monte Carlo Rendering:supplemental document
Listing 2: A modified version of Listing 2 from the main documentwhich computes all dimensions (instead of just one) of an arbitrarysample from a Bush OA pattern.
1 vector<float> bushOA(unsigned i, // sample index2 unsigned d, // number of dimensions (<= s+1)3 unsigned s, // number of levels/strata4 unsigned t, // strength of OA (0 < t <= d)5 unsigned p[], // array of d pseudo-random seeds6 OffsetType ot) { // J or MJ7 vector<float> Xi(d);8 unsigned N = pow(s, t);9 i = permute(i, N, p[0]);
1011 // get digits of ‘i‘ in base ‘s‘12 auto iDigits = toBaseS(i, s, t);1314 unsigned stm = N / s; // pow(s, t-1)15 for (unsigned j = 0; j < s; ++j) {16 unsigned k = (j % 2) ? j - 1 : j + 1;17 unsigned phi = evalPoly(iDigits, j);18 unsigned stratum = permute(phi % s, s, p[j] * 0x51633e2d);19 unsigned sstratum = offset(i, s, stm, p[j] * 0x68bc21eb, ot);20 float jitter = randfloat(i, p[j] * 0x02e5be93);21 Xi[j] = (stratum + (sstratum + jitter) / stm) / s;22 }23 return Xi;24 }
Listing 3: A modified version of Listing 3 from the main documentwhich computes all dimensions (instead of just one) of a CMJ patterngeneralized to arbitrary dimensions d.
1 vector<float> cmjdD(unsigned i, // sample index2 unsigned s, // number of levels/strata3 unsigned t, // strength of OA (t=d)4 unsigned p[] { // array of d pseudo-random seeds5 vector<float> Xi(t);6 unsigned N = pow(s, t);7 i = permute(i, N, p[0]);89 // get digits of ‘i‘ in base ‘s‘
10 auto iDigits = toBaseS(i, s, t);1112 unsigned stm1 = N / s; // pow(s, t-1)13 for (unsigned j = 0; j < s; ++j) {14 unsigned stratum = permute(iDigits[j], s, p[j] * 0x51633e2d);15 auto pDigits = allButJ(iDigits, j);16 unsigned sstratum = evalPoly(pDigits, s);17 sstratum = permute(sstrata, stm1, p[j] * 0x68bc21eb);18 float jitter = randfloat(i, p[j] * 0x02e5be93);19 Xi[d] = (stratum + (sstratum + jitter) / stm1) / s;20 }21 return Xi;22 }
submitted to Eurographics Symposium on Rendering (2019)
Jarosz, Enayet, Kensler, Kilpatrick, and Christensen / Orthogonal Array Sampling for Monte Carlo Rendering:supplemental document 3
1
2
3
0 1 2 0 1 2 0 1 2
1
2
3
0 1 2 0 1 2 0 1 2
1
2
3
0 1 2 0 1 2 0 1 2(a) Jittered (b) Multi-Jittered (c) Correlated Multi-Jittered
Figure 1: All six bivariate projections of a 4D strength-2 Bush-OA with 121, 289 and 1369 samples using jittered, multi-jittered, and correlatedmulti-jittered offsets.
submitted to Eurographics Symposium on Rendering (2019)
4 Jarosz, Enayet, Kensler, Kilpatrick, and Christensen / Orthogonal Array Sampling for Monte Carlo Rendering:supplemental document
y
u
v
x y u x y u x y u
(a) XORed-padded 2D (0,2) sequence [KK02] (b) XORed+shuffle-padded 2D (0,2) sequence (c) Shuffle-padded 2D CMJ points [Ken13]
Figure 2: Alternate version of Fig. 2 from the main document with 3D projections of the 4D pointsets instead of power spectra. A commonway to create samples for higher-dimensional integration is to pad together high-quality 2D point sets. Kollig and Keller [KK02] proposedscrambling each dimension of a (0,2) sequence by XORing it with a different random bit vector. While odd-even dimension pairings producehigh-quality point sets, even-even (xu) or odd-odd (yv) dimensions are poorly distributed. These issues can be eliminated by additionallyrandomly shuffling the points across dimension pairs, but this decorrelates all cross dimensions, destroying 2D stratification.
No shear Slight shear Extreme shear
101 102 103 104 105 106
Samples
10−10
10−9
10−8
10−7
10−6
10−5
10−4
10−3
10−2
Var
ianc
e −0.99, Random−1.01, Latin Hypercube−1.50, (0,2)-seq−1.49, BushMJ(t = 2)
−1.32, BushMJ(t = 3)
−1.30, CMJND(t = 3)
−1.59, Sobol (2,14,15)
−1.50, Sobol (0,1,2)
101 102 103 104 105 106
Samples
10−10
10−9
10−8
10−7
10−6
10−5
10−4
10−3
10−2
Var
ianc
e −0.99, Random−1.01, Latin Hypercube−1.03, (0,2)-seq−1.04, BushMJ(t = 2)
−1.34, BushMJ(t = 3)
−1.32, CMJND(t = 3)
−1.43, Sobol (2,14,15)
−1.39, Sobol (0,1,2)
101 102 103 104 105 106
Samples
10−9
10−8
10−7
10−6
10−5
10−4
10−3
10−2V
aria
nce −1.00, Random
−1.00, Latin Hypercube−1.01, (0,2)-seq−1.01, BushMJ(t = 2)
−1.32, BushMJ(t = 3)
−1.34, CMJND(t = 3)
−1.38, Sobol (2,14,15)
−1.32, Sobol (0,1,2)
Figure 3: Variance graphs for a cylinder sheared by different amounts. This integrand is defined by two circles that lie in the z = 0 andz = 1 planes with centers (x0,y0,0) and (x1,y1,1). The analytic integral is defined as
∫ 10 (x,y,z) = πr2√(x1− x0)2 +(y1− y0)2 +1 where the
radius r is an arbitrary constant. Since an extruded circle with no shear (left) is actually a two-dimensional integrand, samplers that have2D stratification perform best in this case. Once any shear is applied, however, the integrand is 3D and, asymptotically, 2D stratificationdegrades to O(N−1) convergence. With only slight shear (middle), 2D stratified samplers initially perform best, but their variance graphscurve and start to perform worse than 3D stratification at higher sampler counts. For larger shear angles the tipping point starts earlier, with3D stratification dominating even at lower sample counts.
submitted to Eurographics Symposium on Rendering (2019)
Jarosz, Enayet, Kensler, Kilpatrick, and Christensen / Orthogonal Array Sampling for Monte Carlo Rendering:supplemental document 51-
addi
tive( fD 4,
1+)
2-ad
ditiv
e( fD 4,
2+)
2-m
ultip
licat
ive( fD 4,
2×)
4D-r
adia
l( fD 4)
Gaussian(D=∞)
101
102
103
104
105
106
Sam
ples
10−
14
10−
12
10−
10
10−
8
10−
6
10−
4
10−
2
100
Variance
−1.
00,R
ando
m−
2.01
,Jitt
ered
2Dpa
d−
2.97
,Lat
inhy
perc
ube
−2.
99,C
MJ2
Dpa
d−
1.94
,(0,
2)-s
eq.p
ad−
3.00
,Bos
eCM
J(t=
2)−
2.99
,Bus
hMJ(
t=2)
−3.
01,B
ushM
J(t=
3)−
2.93
,Bus
hMJ(
t=4)
−2.
93,C
MJ4
D(t
=4)
−1.
86,H
alto
n−
1.95
,Sob
ol(0
,2,1
4,15
)−
1.94
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
13
10−
11
10−
9
10−
7
10−
5
10−
3
10−
1
Variance
−1.
00,R
ando
m−
1.01
,Jitt
ered
2Dpa
d−
1.00
,Lat
inhy
perc
ube
−1.
00,C
MJ2
Dpa
d−
1.01
,(0,
2)-s
eq.p
ad−
1.52
,Bos
eCM
J(t=
2)−
1.78
,Bus
hMJ(
t=2)
−1.
69,B
ushM
J(t=
3)−
1.53
,Bus
hMJ(
t=4)
−1.
48,C
MJ4
D(t
=4)
−1.
87,H
alto
n−
2.01
,Sob
ol(0
,2,1
4,15
)−
1.95
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
14
10−
12
10−
10
10−
8
10−
6
10−
4
Variance
−1.
00,R
ando
m−
1.00
,Jitt
ered
2Dpa
d−
1.00
,Lat
inhy
perc
ube
−0.
99,C
MJ2
Dpa
d−
1.00
,(0,
2)-s
eq.p
ad−
1.00
,Bos
eCM
J(t=
2)−
1.00
,Bus
hMJ(
t=2)
−1.
00,B
ushM
J(t=
3)−
1.49
,Bus
hMJ(
t=4)
−1.
32,C
MJ4
D(t
=4)
−1.
97,H
alto
n−
1.99
,Sob
ol(0
,2,1
4,15
)−
2.02
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
14
10−
12
10−
10
10−
8
10−
6
10−
4
10−
2
Variance
−1.
00,R
ando
m−
1.00
,Jitt
ered
2Dpa
d−
1.00
,Lat
inhy
perc
ube
−1.
00,C
MJ2
Dpa
d−
1.00
,(0,
2)-s
eq.p
ad−
1.07
,Bos
eCM
J(t=
2)−
1.02
,Bus
hMJ(
t=2)
−1.
05,B
ushM
J(t=
3)−
1.51
,Bus
hMJ(
t=4)
−1.
35,C
MJ4
D(t
=4)
−1.
95,H
alto
n−
2.09
,Sob
ol(0
,2,1
4,15
)−
2.03
,Sob
ol(0
,1,2
,3)
Linearstep(D=1)
101
102
103
104
105
106
Sam
ples
10−
14
10−
12
10−
10
10−
8
10−
6
10−
4
10−
2
Variance
−1.
00,R
ando
m−
2.00
,Jitt
ered
2Dpa
d−
2.97
,Lat
inhy
perc
ube
−3.
00,C
MJ2
Dpa
d−
1.92
,(0,
2)-s
eq.p
ad−
2.99
,Bos
eCM
J(t=
2)−
2.98
,Bus
hMJ(
t=2)
−3.
00,B
ushM
J(t=
3)−
3.02
,Bus
hMJ(
t=4)
−3.
02,C
MJ4
D(t
=4)
−1.
84,H
alto
n−
1.92
,Sob
ol(0
,2,1
4,15
)−
1.92
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
11
10−
9
10−
7
10−
5
10−
3
10−
1
101
Variance
−1.
00,R
ando
m−
1.02
,Jitt
ered
2Dpa
d−
1.00
,Lat
inhy
perc
ube
−1.
00,C
MJ2
Dpa
d−
1.03
,(0,
2)-s
eq.p
ad−
1.72
,Bos
eCM
J(t=
2)−
1.99
,Bus
hMJ(
t=2)
−1.
65,B
ushM
J(t=
3)−
1.50
,Bus
hMJ(
t=4)
−1.
73,C
MJ4
D(t
=4)
−1.
86,H
alto
n−
2.01
,Sob
ol(0
,2,1
4,15
)−
1.95
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
11
10−
9
10−
7
10−
5
10−
3
10−
1
Variance
−1.
00,R
ando
m−
1.00
,Jitt
ered
2Dpa
d−
0.99
,Lat
inhy
perc
ube
−0.
99,C
MJ2
Dpa
d−
1.01
,(0,
2)-s
eq.p
ad−
1.03
,Bos
eCM
J(t=
2)−
1.02
,Bus
hMJ(
t=2)
−1.
09,B
ushM
J(t=
3)−
1.48
,Bus
hMJ(
t=4)
−1.
42,C
MJ4
D(t
=4)
−1.
92,H
alto
n−
1.89
,Sob
ol(0
,2,1
4,15
)−
1.96
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
10
10−
8
10−
6
10−
4
10−
2
Variance
−1.
00,R
ando
m−
1.00
,Jitt
ered
2Dpa
d−
0.99
,Lat
inhy
perc
ube
−0.
99,C
MJ2
Dpa
d−
1.01
,(0,
2)-s
eq.p
ad−
1.01
,Bos
eCM
J(t=
2)−
1.01
,Bus
hMJ(
t=2)
−1.
05,B
ushM
J(t=
3)−
1.46
,Bus
hMJ(
t=4)
−1.
38,C
MJ4
D(t
=4)
−1.
79,H
alto
n−
1.77
,Sob
ol(0
,2,1
4,15
)−
1.73
,Sob
ol(0
,1,2
,3)
Binarystep(D=0)
101
102
103
104
105
106
Sam
ples
10−
12
10−
10
10−
8
10−
6
10−
4
10−
2
Variance
−1.
00,R
ando
m−
1.57
,Jitt
ered
2Dpa
d−
2.09
,Lat
inhy
perc
ube
−2.
09,C
MJ2
Dpa
d−
1.92
,(0,
2)-s
eq.p
ad−
2.05
,Bos
eCM
J(t=
2)−
2.05
,Bus
hMJ(
t=2)
−2.
04,B
ushM
J(t=
3)−
1.89
,Bus
hMJ(
t=4)
−1.
89,C
MJ4
D(t
=4)
−1.
84,H
alto
n−
1.92
,Sob
ol(0
,2,1
4,15
)−
1.92
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
8
10−
6
10−
4
10−
2
100
Variance
−1.
01,R
ando
m−
1.03
,Jitt
ered
2Dpa
d−
1.00
,Lat
inhy
perc
ube
−1.
00,C
MJ2
Dpa
d−
1.03
,(0,
2)-s
eq.p
ad−
1.49
,Bos
eCM
J(t=
2)−
1.50
,Bus
hMJ(
t=2)
−1.
32,B
ushM
J(t=
3)−
1.23
,Bus
hMJ(
t=4)
−1.
32,C
MJ4
D(t
=4)
−1.
56,H
alto
n−
1.54
,Sob
ol(0
,2,1
4,15
)−
1.53
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
8
10−
6
10−
4
10−
2
Variance
−1.
00,R
ando
m−
1.01
,Jitt
ered
2Dpa
d−
0.99
,Lat
inhy
perc
ube
−1.
01,C
MJ2
Dpa
d−
1.01
,(0,
2)-s
eq.p
ad−
1.01
,Bos
eCM
J(t=
2)−
1.01
,Bus
hMJ(
t=2)
−1.
10,B
ushM
J(t=
3)−
1.24
,Bus
hMJ(
t=4)
−1.
28,C
MJ4
D(t
=4)
−1.
45,H
alto
n−
1.42
,Sob
ol(0
,2,1
4,15
)−
1.44
,Sob
ol(0
,1,2
,3)
101
102
103
104
105
106
Sam
ples
10−
8
10−
6
10−
4
10−
2
Variance
−1.
00,R
ando
m−
1.00
,Jitt
ered
2Dpa
d−
0.99
,Lat
inhy
perc
ube
−0.
99,C
MJ2
Dpa
d−
1.00
,(0,
2)-s
eq.p
ad−
0.99
,Bos
eCM
J(t=
2)−
0.99
,Bus
hMJ(
t=2)
−1.
04,B
ushM
J(t=
3)−
1.25
,Bus
hMJ(
t=4)
−1.
28,C
MJ4
D(t
=4)
−1.
28,H
alto
n−
1.27
,Sob
ol(0
,2,1
4,15
)−
1.28
,Sob
ol(0
,1,2
,3)
Figu
re4:
Fig.
4fr
omth
em
ain
pape
rex
tend
edw
ithth
eG
auss
ian
inte
gran
d.Va
rian
cebe
havi
orof
13sa
mpl
ers
on4D
anal
ytic
inte
gran
dsof
diffe
rent
com
plex
ity(c
olum
ns)a
ndco
ntin
uity
(row
s).W
elis
tthe
best
-fits
lope
ofea
chte
chni
que,
whi
chge
nera
llym
atch
esth
eth
eore
tical
lypr
edic
ted
conv
erge
nces
rate
s(T
able
3).O
ursa
mpl
ers
alw
ays
perf
orm
bette
rth
antr
aditi
onal
padd
ing
appr
oach
es,b
utar
eas
ympt
otic
ally
infe
rior
tohi
gh-d
imen
sion
alQ
MC
sequ
ence
sfo
rge
nera
lhig
h-di
men
sion
alin
tegr
ands
.Whe
nst
reng
tht<
d(r
ight
two
colu
mns
),co
nver
genc
ede
grad
esto−
1,bu
thig
her
stre
ngth
sat
tain
low
erco
nsta
ntfa
ctor
s.
submitted to Eurographics Symposium on Rendering (2019)
6 Jarosz, Enayet, Kensler, Kilpatrick, and Christensen / Orthogonal Array Sampling for Monte Carlo Rendering:supplemental document
References[AZ16] ATKINS, J. and ZAX, D. B. “Nested orthogonal arrays”.
arXiv:1601.06459 [stat] (Jan. 2016). arXiv: 1601.06459 [stat] 1.
[CQ14] CHEN, J. and QIAN, P. Z. “Latin hypercube designs with controlledcorrelations and multi-dimensional stratification”. Biometrika 101.2 (Feb.2014), 319–332. ISSN: 1464-3510. DOI: 10/f5549t 1.
[CSW94] CHIU, K., SHIRLEY, P., and WANG, C. “Multi-jittered sampling”.Graphics Gems IV. Ed. by HECKBERT, P. S. San Diego, CA, USA:Academic Press, 1994, 370–374. ISBN: 0-12-336155-9 1.
[HCT18] HE, Y., CHENG, C.-S., and TANG, B. “Strong orthogonal arraysof strength two plus”. The Annals of Statistics 46.2 (Apr. 2018), 457–468.ISSN: 0090-5364. DOI: 10/gfznb6 1.
[HHQ16] HWANG, Y., HE, X., and QIAN, P. Z. “Sliced orthogonal array-based Latin hypercube designs”. Technometrics 58.1 (Jan. 2016), 50–61.ISSN: 1537-2723. DOI: 10/f8cjrk 1.
[HQ10] HAALAND, B. and QIAN, P. Z. G. “An approach to constructingnested space-filling designs for multi-fidelity computer experiments”.Statistica Sinica 20.3 (2010), 1063–1075. ISSN: 1017-0405 1.
[HQ11] HE, X. and QIAN, P. Z. G. “Nested orthogonal array-based Latinhypercube designs”. Biometrika 98.3 (Sept. 2011), 721–731. ISSN: 0006-3444. DOI: 10/cz9pj6 1.
[HSS99] HEDAYAT, A. S., SLOANE, N. J. A., and STUFKEN, J. Orthogo-nal Arrays: Theory and Applications. Springer-Verlag, 1999. ISBN: 978-1-4612-7158-1. DOI: bqh3jj 1.
[HT12] HE, Y. and TANG, B. “Strong orthogonal arrays and associatedLatin hypercubes for computer experiments”. Biometrika 100.1 (Dec.2012), 254–260. ISSN: 1464-3510. DOI: 10/gfznb5 1.
[HT14] HE, Y. and TANG, B. “A characterization of strong orthogonalarrays of strength three”. The Annals of Statistics 42.4 (2014), 1347–1360.ISSN: 00905364. DOI: 10/gfznb4 1.
[Ken13] KENSLER, A. Correlated Multi-Jittered Sampling. Tech. rep. 13-01. Pixar Animation Studios, Mar. 2013 1, 4.
[KK02] KOLLIG, T. and KELLER, A. “Efficient multidimensional sam-pling”. Computer Graphics Forum (Proceedings of Eurographics) 21.3(Sept. 2002), 557–563. ISSN: 0167-7055. DOI: 10/d2stpx 4.
[LL15] LIU, H. and LIU, M.-Q. “Column-orthogonal strong orthogonalarrays and sliced strong orthogonal arrays”. Statistica Sinica (2015). ISSN:1017-0405. DOI: 10/f7z56h 1.
[Owe13] OWEN, A. B. Monte Carlo Theory, Methods and Examples. Tobe published, 2013. URL: https://statweb.stanford.edu/~owen/mc/(visited on 06/07/2019) 1.
[Owe94] OWEN, A. B. “Controlling correlations in Latin hypercube sam-ples”. Journal of the American Statistical Association 89.428 (Dec.1994), 1517–1522. ISSN: 1537-274X. DOI: 10/gfzncf 1.
[QA10] QIAN, P. Z. G. and AI, M. “Nested lattice sampling: a new sam-pling scheme derived by randomizing nested orthogonal arrays”. Journalof the American Statistical Association 105.491 (2010), 1147–1155. ISSN:01621459. DOI: 10/fwjp8h 1.
[QAW09] QIAN, P. Z. G., AI, M., and WU, C. F. J. “Construction of nestedspace-filling designs”. The Annals of Statistics 37.6A (Dec. 2009), 3616–3643. ISSN: 0090-5364, 2168-8966. DOI: 10/fr79bb 1.
[Qia09] QIAN, P. Z. G. “Nested Latin hypercube designs”. Biometrika 96.4(2009), 957–970. ISSN: 0006-3444. DOI: 10/cpmd77 1.
[RHVD10] RENNEN, G., HUSSLAGE, B., VAN DAM, E. R., and DENHERTOG, D. “Nested maximin Latin hypercube designs”. Structuraland Multidisciplinary Optimization 41.3 (Apr. 2010), 371–395. ISSN:1615-1488. DOI: 10/cpqtq3 1.
[SLQ14] SUN, F., LIU, M.-Q., and QIAN, P. Z. G. “On the construction ofnested space-filling designs”. The Annals of Statistics 42.4 (2014), 1394–1425. ISSN: 0090-5364. DOI: 10/gf3nmd 1.
[Tan93] TANG, B. “Orthogonal array-based Latin hypercubes”. Journalof the American Statistical Association 88.424 (Dec. 1993), 1392–1397.ISSN: 1537-274X. DOI: 10/gfzncp 1.
[Tan98] TANG, B. “Selecting Latin hypercubes using correlation criteria”.Statistica Sinica 8.3 (1998), 965–977. ISSN: 10170405, 19968507 1.
[Wen14] WENG, J. “Maximin Strong Orthogonal Arrays”. M.Sc. Thesis.Simon Fraser University, 2014 1.
[WL13] WANG, K. and LI, Y. “Constructions of nested orthogonal arrays”.Journal of Combinatorial Designs 21.10 (2013), 464–477. ISSN: 1520-6610. DOI: 10/gf2zf9 1.
[YLL14] YIN, Y., LIN, D. K., and LIU, M.-Q. “Sliced Latin hypercube de-signs via orthogonal arrays”. Journal of Statistical Planning and Inference149 (June 2014), 162–171. ISSN: 0378-3758. DOI: 10/gfzncs 1.
[ZWD19] ZHANG, T.-F., WU, G., and DEY, A. “Construction of somenew families of nested orthogonal arrays”. Communications in Statistics -Theory and Methods 48.3 (Feb. 2019), 774–779. ISSN: 0361-0926. DOI:10/gf2zf4 1.
submitted to Eurographics Symposium on Rendering (2019)