conditional simulation of max-stable processesribatet/docs/dombry2011.pdf · in classical...
TRANSCRIPT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
Biometrika (2012), xx, x, pp. 1–22
C© 2007 Biometrika Trust
Printed in Great Britain
Conditional simulation of max-stable processes
BY C. DOMBRY, F. EYI-MINKO,
Laboratoire de Mathematiques et Application, Universite de Poitiers, Teleport 2, BP 30179,
F-86962 Futuroscope-Chasseneuil cedex, France
[email protected], [email protected]
AND M. RIBATET
Department of Mathematics, Universite Montpellier 2, 4 place Eugene Bataillon, 34095 cedex
2 Montpellier, France
SUMMARY
Since many environmental processes are spatial in extent, a single extreme event may affect
several locations, and their spatial dependence has to be appropriately taken into account. This
paper proposes a framework for conditional simulation of max-stable processes and gives closed
forms for the regular conditional distributions of Brown–Resnick and Schlather processes. We
test the method on simulated data and give an application to extreme rainfall around Zurich and
extreme temperatures. The proposed framework provides accurate conditional simulations and
can handle real-sized problems.
Some key words: Conditional simulation; Markov chain Monte Carlo; Max-stable process; Precipitation; Regular
conditional distribution; Temperature.
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
2 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
1. INTRODUCTION
Max-stable processes arise naturally when studying extremes of stochastic processes and
therefore play a major role in the statistical modelling of spatial extremes (Buishand et al, 2008;
Padoan et al., 2010; Davison et al., 2012). Although a different spectral characterization of
max-stable processes exists (de Haan, 1984), for our purposes the most useful representation
is (Schlather, 2002)
Z(x) = maxi≥1
ζiYi(x), x ∈ Rd, (1)
where {ζi}i≥1 are the points of a Poisson process on (0,∞) with intensity dΛ(ζ) = ζ−2dζ and
the Yi are independent replicates of a non-negative stochastic process Y such that E{Y (x)} = 1
for all x ∈ Rd. It is well known that Z is a max-stable process on Rd with unit Frechet margins
(Schlather, 2002; de Haan & Fereira, 2006, page 307). Although (1) takes the pointwise maxi-
mum over an infinite number of points {ζi} and processes Yi, it is possible to get approximate
realizations from Z (Schlather, 2002; Oesting et al., 2012).
Based on (1) several parametric max-stable models have been proposed (Schlather, 2002;
Brown & Resnick, 1977; Kabluchko et al., 2009; Davison et al., 2012) that share the same finite-
dimensional distribution functions
pr{Z(x1) ≤ z1, . . . , Z(xk) ≤ zk} = exp
[−E
{maxj=1,...,k
Y (xj)
zj
}],
where k ∈ N, z1, . . . , zk > 0 and x1, . . . , xk ∈ Rd.
Apart from the Smith model (Genton et al., 2011), only the bivariate cumulative distribution
functions are explicitly known. To bypass this hurdle to inference based on max-stable processes,
de Haan & Pereira (2006) propose a semi-parametric estimator and Padoan et al. (2010) suggest
the use of the maximum pairwise likelihood estimator.
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
Conditional simulation of max-stable processes 3
Paralleling the use of the variogram in classical geostatistics, the extremal coefficient function
(Schlather & Tawn, 2003; Cooley et al., 2006)
θ(x1 − x2) = −z log pr{Z(x1) ≤ z, Z(x2) ≤ z}
is widely used to summarize the spatial dependence of extremes for stationary processes. It takes
values in the interval [1, 2]; the lower bound indicates complete dependence, and the upper bound
independence.
The last decade has seen many advances to develop geostatistics for extremes, and software is
available to practitioners (Wang, 2010; Schlather, 2012; Ribatet, 2011). However an important
tool currently missing is conditional simulation of max-stable processes. In classical geostatistic
based on Gaussian models, conditional simulation is well established (Chiles & Delfiner, 1999)
and provides a framework to assess the distribution of a Gaussian random field given values
observed at fixed locations. For example, conditional simulations of Gaussian processes have
been used to model land topography (Mandelbrot, 1982).
Conditional simulation of max-stable processes is a long-standing problem (Davis & Resnick,
1989, 1993). Wang & Stoev (2011) provide a first solution, but their framework is limited to
processes having a discrete spectral measure and thus may be too restrictive to appropriately
model spatial dependence in complex situations.
Based on the recent developments on the regular conditional distribution of max-infinitely di-
visible processes, the aim of this paper is to provide a methodology for conditional simulation
of max-stable processes with continuous spectral measures. More formally for a study region
X ⊂ Rd, our goal is to derive an algorithm to sample from the regular conditional distribu-
tion ofZ | {Z(x1) = z1, . . . , Z(xk) = zk} for some z1, . . . , zk > 0 and k conditioning locations
x1, . . . , xk ∈ X .
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
4 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
2. CONDITIONAL SIMULATION OF MAX-STABLE PROCESSES
2·1. General framework
This section reviews some key results of an unpublished paper available from the first author
with a particular emphasis on max-stable processes. Our goal is to give a more practical inter-
pretation of their results from a simulation perspective. To this aim, we recall two key results and
propose a procedure to generate conditional realizations of max-stable processes.
Let RX be the space of real-valued functions on X ⊂ Rd and let Φ = {ϕi}i≥1 be a Poisson
point process on RX where ϕi(x) = ζiYi(x) (i = 1, 2, . . .) with ζi and Yi as in (1). We write
f(x) = {f(x1), . . . , f(xk)} for all random functions f : X → R and x = (x1, . . . , xk) ∈ X k. It
is not difficult to show that for all Borel sets A ⊂ Rk, the Poisson process {ϕi(x)}i≥1 defined
on Rk has intensity measure
Λx(A) =
∫ ∞0
pr {ζY (x) ∈ A} ζ−2dζ.
The point process Φ is called regular if the intensity measure Λx has an intensity function λx,
i.e., Λx(dz) = λx(z)dz, for all x ∈ X k.
The first key result is that provided the Poisson process Φ is regular, the intensity function λx
and the conditional intensity function
λs|x,z(u) =λ(s,x)(u, z)
λx(z), (s, x) ∈ Xm+k, u ∈ Rm, z ∈ (0,+∞)k, (2)
drive how the conditioning terms {Z(xj) = zj} (j = 1, . . . , k) are met, see Steps 1 and 2 in
Theorem 1.
The second key result is that, conditionally on Z(x) = z, the Poisson process Φ can be de-
composed into two independent point processes, say Φ = Φ− ∪ Φ+, where
Φ− = {ϕ ∈ Φ: ϕ(xi) < zi, i = 1, . . . , k} , Φ+ =k⋃i=1
{ϕ ∈ Φ: ϕ(xi) = zi} .
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
Conditional simulation of max-stable processes 5
Before introducing a procedure to get conditional realizations of max-stable processes, we
introduce notation and make connections with the pioneering work of Wang & Stoev (2011).
A function ϕ ∈ Φ+ such that ϕ(xi) = zi for some i ∈ {1, . . . , k} is called an extremal func-
tion associated to xi and denoted by ϕ+xi . It is easy to show that there exists almost surely a
unique extremal function associated to xi. Although Φ+ = {ϕ+x1 , . . . , ϕ
+xk} almost surely, it
might happen that a single extremal function contributes to the random vector Z(x) at sev-
eral locations xi, e.g., ϕ+x1 = ϕ+
x2 . To take such repetitions into account, we define a random
partition θ = (θ1, . . . , θ`) of the set {x1, . . . , xk} into ` = |θ| blocks, and extremal functions
(ϕ+1 , . . . , ϕ
+` ) such that Φ+ = {ϕ+
1 , . . . , ϕ+` } and ϕ+
j (xi) = zi if xi ∈ θj and ϕ+j (xi) < zi if
xi /∈ θj (i = 1, . . . , k; j = 1, . . . , `). Wang & Stoev (2011) call the partition θ the hitting sce-
nario. The set of all possible partitions of {x1, . . . , xk}, denoted Pk, identifies all possible hitting
scenarios.
From a simulation perspective, the fact that Φ− and Φ+ are independent given Z(x) = z
is especially convenient and suggests a three–step procedure to sample from the conditional
distribution of Z given Z(x) = z.
THEOREM 1. Suppose that the point process Φ is regular and let (x, s) ∈ X k+m. For τ =
(τ1, . . . , τ`) ∈Pk and j = 1, . . . , `, define Ij = {i : xi ∈ τj}, xτj = (xi)i∈Ij , zτj = (zi)i∈Ij ,
xτcj = (xi)i/∈Ij and zτcj = (zi)i/∈Ij . Consider the three–step procedure:
Step 1. Draw a random partition θ ∈Pk with distribution
πx(z, τ) = pr {θ = τ | Z(x) = z} =1
C(x, z)
|τ |∏j=1
λxτj (zτj )
∫{uj<zτc
j}λxτc
j|xτj ,zτj (uj)duj ,
where the normalization constant is
C(x, z) =∑τ∈Pk
|τ |∏j=1
λxτj (zθj )
∫{uj<zτc
j}λxτc
j|xτj ,zτj (uj)duj .
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
6 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
Step 2. Given τ = (τ1, . . . , τ`), draw ` independent random vectors ϕ+1 (s), . . . , ϕ+
` (s) from
the distribution
pr{ϕ+j (s) ∈ dv | Z(x) = z, θ = τ
}=
1
Cj
{∫1{u<zτc
j}λ(s,xτc
j)|xτj ,zτj (v, u)du
}dv,
where 1{·} is the indicator function and
Cj =
∫1{u<zτc
j}λ(s,xτc
j)|xτj ,zτj (v, u)dudv,
and define the random vector Z+(s) = maxj=1,...,` ϕ+j (s).
Step 3. Independently draw a Poisson point process {ζi}i≥1 on (0,∞) with intensity ζ−2dζ
and {Yi}i≥1 independent copies of Y , and define the random vector
Z−(s) = maxi≥1
ζiYi(s)1{ζiYi(x)<z}.
Then the random vector Z(s) = max {Z+(s), Z−(s)} follows the conditional distribution of
Z(s) given Z(x) = z.
The corresponding conditional cumulative distribution function is
pr {Z(s) ≤ a | Z(x) = z} =
∑τ∈Pk
πx(z, τ)
|τ |∏j=1
Fτ,j(a)
pr{Z(s) ≤ a, Z(x) ≤ z}pr{Z(x) ≤ z}
, (3)
where
Fτ,j(a) = pr{ϕ+j (s) ≤ a | Z(x) = z, θ = τ
}=
∫{u<zτc
j,v<a} λ(s,xτc
j)|xτj ,zτj (v, u)dudv∫
{u<zτcj} λxτc
j|xτj ,zτj (u)du
.
The first term in the right hand side of (3) corresponds to Steps 1 and 2 while the ratio is a
consequence of Step 3. It is clear from (3) that the conditional random field Z | {Z(x) = z} is
not max-stable.
2·2. Distribution of the extremal functions
In this section we derive closed forms for the intensity function λx(z) and the conditional in-
tensity function λs|x,z(u) for two widely used max-stable processes; the Brown–Resnick (Brown
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
Conditional simulation of max-stable processes 7
& Resnick, 1977; Kabluchko et al., 2009) and the Schlather (2002) processes. Details of the
derivations are given in the Appendix.
The Brown–Resnick process corresponds to the case where Y (x) = exp{W (x)− γ(x)} (x ∈
Rd) in (1) with W a centered Gaussian process with stationary increments, semivariogram γ and
such that W (o) = 0 almost surely where o denotes the origin of Rd. For x ∈ X k and provided
the covariance matrix Σx of the random vector W (x) is positive definite, the intensity function
is
λx(z) = Cx exp
(−1
2log zTQx log z + Lx log z
) k∏i=1
z−1i , z ∈ (0,∞)k,
with 1k = (1)i=1,...,k, γx = {γ(xi)}i=1,...,k,
Qx = Σ−1x −Σ−1x 1k1
Tk Σ−1x
1Tk Σ−1x 1k, Lx =
(1Tk Σ−1x γx − 1
1Tk Σ−1x 1k1k − γx
)TΣ−1x ,
Cx = (2π)(1−k)/2|Σx|−1/2(1Tk Σ−1x 1k)−1/2 exp
{1
2
(1Tk Σ−1x γx − 1)2
1Tk Σ−1x 1k− 1
2γTx Σ−1x γx
},
and for all (s, x) ∈ Xm+k, (u, z) ∈ (0,∞)m+k and provided the covariance matrix Σ(s,x) is pos-
itive definite, the conditional intensity function corresponds to a multivariate log-normal proba-
bility density function
λs|x,z(u) = (2π)−m/2|Σs|x|−1/2 exp
{−1
2(log u− µs|x,z)TΣ−1s|x(log u− µs|x,z)
} m∏i=1
u−1i ,
where µs|x,z ∈ Rm and Σs|x are the mean and covariance matrix of the underlying normal dis-
tribution and are given by
Σ−1s|x = JTm,kQ(s,x)Jm,k, µs|x,z ={L(s,x)Jm,k − log zT J Tm,kQ(s,x)Jm,k
}Σs|x,
with
Jm,k =
Im
0k,m
, Jm,k =
0m,k
Ik
,
where Ik denotes the k × k identity matrix and 0m,k the m× n null matrix.
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
8 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
The Schlather process considers the case Y (x) = (2π)1/2 max{0, ε(x)} (x ∈ Rd) in (1) with
ε a standard Gaussian process with correlation function ρ. The associated point process Φ is
not regular and it is more convenient to consider the equivalent representation where Y (x) =
(2π)1/2ε(x) (x ∈ Rd). For x ∈ X k and provided the covariance matrix Σx of the random vector
ε(x) is positive definite, the intensity function is
λx(z) = π−(k−1)/2|Σx|−1/2ax(z)−(k+1)/2Γ
(k + 1
2
), z ∈ Rk,
where ax(z) = zTΣ−1x z.
For (s, x) ∈ Xm+k, (u, z) ∈ Rm+k and provided that the covariance matrix Σ(s,x) is positive
definite, the conditional intensity function λs|x,z(u) corresponds to the density of a multivariate
Student distribution with k + 1 degrees of freedom, location parameter µ = Σs:xΣ−1x z, and scale
matrix
Σ =ax(z)
k + 1
(Σs − Σs:xΣ−1x Σx:s
), Σ(s,x) =
Σs Σs:x
Σx:s Σx
.
3. MARKOV CHAIN MONTE CARLO SAMPLER
Section 2 introduced a procedure to get realizations from the regular conditional distribution
of max-stable processes. This sampling scheme amounts to sampling from a discrete distribution
whose state space corresponds to all possible partitions of the set of conditioning points, see
Step 1 of Theorem 1. Hence, even for a moderate number k of conditioning locations, the state
space Pk becomes very large and the distribution πx(z, ·) cannot be computed exactly. A Gibbs
sampler is especially convenient for Monte Carlo sampling from πx(z, ·).
For τ ∈Pk, let τ−j be the restriction of τ to the set {x1, . . . , xk} \ {xj}. Our goal is to
simulate from the conditional distribution
pr(θ ∈ · | θ−j = τ−j), (4)
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
Conditional simulation of max-stable processes 9
where θ ∈Pk is a random partition which follows the target distribution πx(z, ·).
Since the number of possible updates is always less than k, a combinatorial explosion is
avoided. Indeed for τ ∈Pk of size `, the number of partitions τ∗ ∈Pk such that τ∗−j = τ−j
for some j ∈ {1, . . . , k} is
b+ =
` if {xj} is a partitioning set of τ ,
`+ 1 if {xj} is not a partitioning set of τ ,
since the point xj may be reallocated to any partitioning set of τ−j or to a new one.
To illustrate consider the set {x1, x2, x3} and let τ = ({x1, x2}, {x3}). Then the possible
partitions τ∗ such that τ∗−2 = τ−2 are ({x1, x2}, {x3}), ({x1}, {x2}, {x3}), ({x1}, {x2, x3}),
while there exist only two partitions such that τ∗−3 = τ−3, i.e., ({x1, x2}, {x3}), ({x1, x2, x3}).
The distribution (4) has nice properties. Since for all τ∗ ∈Pk such that τ∗−j = τ−j we have
pr (θ = τ∗ | θ−j = τ−j) =πx(z, τ∗)∑
τ∈Pk
πx(z, τ)1{τ−j=τ−j}∝∏|τ∗|j=1wτ∗,j∏|τ |j=1wτ,j
, (5)
where
wτ,j = λxτj (zτj )
∫{u<zτc
j}λxτc
j|xτj ,zτj (u)du.
Since many factors cancel out on the right hand side of (5), the Gibbs sampler is especially
convenient.
The most computationally demanding part of (5) is the evaluation of the integral∫{u<zτc
j}λxτc
j|xτj ,zτj (u)du.
For the Brown–Resnick and Schlather processes, we follow the lines of Genz (1992) and compute
these probabilities using a separation of variables method which provides a transformation of the
original integration problem to the unit hyper-cube. A quasi Monte Carlo scheme and antithetic
variables are used to improve efficiency.
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
10 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
Since it is not obvious how to implement a Gibbs sampler whose target distribution has support
Pk, the remainder of this section gives practical details. For any fixed locations x1, . . . , xk ∈ X ,
we first describe how each partition of {x1, . . . , xk} is stored. To illustrate this consider the set
{x1, x2, x3} and the partition ({x1, x2}; {x3}). This partition is defined as (1, 1, 2), indicating
that x1 and x2 belong to the partitioning set labeled 1 and x3 belongs to the partitioning set 2.
There exist several equivalent notations for this partition: for example one can use (2, 2, 1) or
(1, 1, 3). Since there is a one-one mapping between Pk and the set
P∗k =
{(a1, . . . , ak) : i ∈ {2, . . . , k}, 1 = a1 ≤ ai ≤ max
1≤j<iaj + 1, ai ∈ Z
},
we shall restrict our attention to the partitions that live in P∗k , and going back to our example we
see that (1, 1, 2) is valid but (2, 2, 1) and (1, 1, 3) are not.
For τ ∈P∗k of size `, let r1 =
∑ki=1 1{τi=aj} and r2 =
∑ki=1 1{τi=b}, i.e., the number of
conditioning locations that belong to the partitioning sets aj and b where b ∈ {1, . . . , b+} with
b+ =
` (r1 = 1),
`+ 1 (r1 6= 1).
Then the conditional probability distribution (5) satisfies
pr(τj = b | τi = ai, i 6= j) ∝
1 (b = aj), (6a)
wτ∗,b/(wτ,bwτ,aj ) (r1 = 1, r2 6= 0, b 6= aj), (6b)
wτ∗,bwτ∗,aj/(wτ,bwτ,aj ) (r1 6= 1, r2 6= 0, b 6= aj), (6c)
wτ∗,bwτ∗,aj/wτ,aj (r1 6= 1, r2 = 0, b 6= aj), (6d)
where τ∗ = (a1, . . . , aj−1, b, aj+1, . . . , ak). Although τ∗ may not belong to P∗k , it corresponds
to a unique partition of Pk and we can use the bijection Pk →P∗k to recode τ∗ into an element
of P∗k . In (6a)–(6d) the event {r1 = 1, r2 = 0, b 6= aj} is missing since {r1 = 1, r2 = 0} implies
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
Conditional simulation of max-stable processes 11
Table 1. Sample path properties of the max-stable models. For the Brown–Resnick
model, the variogram parameters are set to ensure that the extremal coefficient function
satisfies θ(115) = 1·7 while for the Schlather model the correlation function parame-
ters are set to ensure that θ(100) = 1·5.
Brown–Resnick: γ(h) = (h/λ)κ Schlather: ρ(h) = exp{−(h/λ)κ}
γ1: Very wiggly γ2: Wiggly γ3: Smooth ρ1: Very wiggly ρ2: Wiggly ρ3: Smooth
λ 25 54 69 208 144 128
κ 0·5 1·0 1·5 0·5 1·0 1·5
that τ∗ = τ , where the equality has to be understood in terms of elements of Pk, and this case
has been already taken into account with (6a).
Once these conditional weights have been computed, the Gibbs sampler proceeds by up-
dating each element of τ successively. We use a random scan implementation (Liu et al.,
1995). One iteration selects an element of τ at random according to a given distribution, say
p = (p1, . . . , pk), and then updates this element. Throughout this paper we will use the uniform
random scan Gibbs sampler for which the selection distribution is a discrete uniform distribution,
i.e., p = (k−1, . . . , k−1).
4. SIMULATION STUDY
In this section we check if our algorithm is able to produce realistic conditional simulations
of Brown–Resnick and Schlather processes. For each model, we consider three different sample
path properties, as summarized in Table 1. These configurations were chosen such that the spatial
dependence structures are similar to our applications in Section 5.
In order to check if our sampling procedure is accurate and given a single conditional event
{Z(x) = z} for each configuration, we generated 1000 conditional realizations with standard
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
12 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
−100 −50 0 50 100
−1
0
1
2
3
4
x
Z(x
)
−100 −50 0 50 100
−1
0
1
2
3
4
xZ
(x)
−100 −50 0 50 100
−1
0
1
2
3
4
x
Z(x
)
−100 −50 0 50 100
−1
0
1
2
3
4
x
Z(x
)
−100 −50 0 50 100
−1
0
1
2
3
4
x
Z(x
)
−100 −50 0 50 100
−1
0
1
2
3
4
xZ
(x)
Fig. 1. Pointwise sample quantiles estimated from 1000
conditional simulations of max-stable processes with stan-
dard Gumbel margins with k = 5, 10, 15 conditioning lo-
cations. The top row shows results for the Brown–Resnick
models with semivariograms γ3, γ2, γ1, from left to right.
The bottom row shows results for the Schlather models
with correlation functions ρ3, ρ2, ρ1, from left to right. The
solid black lines shows the pointwise 0·025, 0·5, 0·975
sample quantiles and the dashed grey lines that of a stan-
dard Gumbel distribution. The squares show the condi-
tional points {(xi, zi)}i=1,...,k. The solid grey lines show
the simulated paths used to get the conditioning events.
Gumbel margins. Figure 1 shows the pointwise sample quantiles obtained from these 1000 sim-
ulated paths and compares them to unit Gumbel quantiles. As expected the conditional sample
paths inherit the regularity driven by the shape parameter κ and there is less variability in re-
gions close to conditioning locations. Since the Brown–Resnick processes considered are ergodic
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
Conditional simulation of max-stable processes 13
Table 2. Timings for conditional simulations of max-stable processes on a 50× 50 grid de-
fined on the square [0, 100× 21/2]2 for a varying number k of conditioning locations uniformly
distributed over the region. The times, in seconds, are mean values over 100 conditional simu-
lations; standard deviations are reported in brackets.
Brown–Resnick: γ(h) = (h/25)0·5 Schlather: ρ(h) = exp{−(h/208)0·50
}Step 1 Step 2 Step 3 Overall Step 1 Step 2 Step 3 Overall
k = 5 0·21 (0·01) 49 (11) 1·4 (0·1) 50 (11) 1·4 (0·02) 1·9 (0·7) 0·9 (0·3) 4·2 (0·8)
k = 10 8 (2) 76 (18) 1·4 (0·1) 85 (19) 12 (4) 2·4 (0·8) 1·0 (0·3) 15 (4)
k = 25 95 (38) 117 (30) 1·4 (0·1) 214 (61) 86 (42) 4 (1) 1·0 (0·3) 90 (43)
k = 50 583 (313) 348 (391) 1·5 (0·1) 931 (535) 367 (222) 62 (113) 1·0 (0·3) 430 (262)
Conditional simulations with k = 5 do not use a Gibbs sampler.
(Kabluchko and Schlather, 2010), for regions far away from any conditioning location the sam-
ple quantiles converges to that of a standard Gumbel distribution, indicating that the conditional
event has no influence. This is not the case for the non-ergodic Schlather process. Most of the
time the sample paths used to get the conditional events belong to the 95% pointwise confidence
intervals, corroborating that our sampling procedure seems to be accurate.
Table 2 gives timings for conditional simulations of max-stable processes on a 50× 50 grid
with a varying number of conditioning locations. Due to the combinatorial complexity of the
partition set Pk, the timings increase rapidly with respect to the number of conditioning points
k. It is however reassuring that the algorithm is tractable when k ≤ 50; this covers many practical
situations.
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
14 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
1000 2000 3000 4000
0 25 50 75 100(km)
(m)
1970 1980 1990 2000
0
20
40
60
80
Years
Maxim
um
rain
fall
in Z
uri
ch (
mm
)
0 50 100 150 200
1.0
1.2
1.4
1.6
1.8
2.0
h (km)
θ(h
)
Fig. 2. Left: Map of Switzerland showing the stations of
the 24 rainfall gauges used for the analysis, with an insert
showing the altitude. The station marked with a triangle
corresponds to Zurich. Middle: Summer maximum rain-
fall values for 1962–2008 at Zurich. Right: Comparison
between the pairwise extremal coefficient estimates for the
51 original weather stations and the extremal coefficient
function derived from a fitted Brown–Resnick processes
having semi variogram γ(h) = (h/λ)κ. The grey points
are pairwise estimates; the black ones are binned estimates
and the curve is the fitted extremal coefficient function.
5. APPLICATION
5·1. Extreme precipitation around Zurich
The data considered here were previously analyzed by Davison et al. (2012), who showed that
Brown–Resnick processes were among the best models for these data. The data are summer max-
imum rainfall for the years 1962–2008 at 51 weather stations in the Plateau region of Switzerland,
provided by the national meteorological service, MeteoSuisse. To ensure strong dependence be-
tween the conditioning locations, we consider as conditional locations the 24 weather stations
that are at most 30km distant from Zurich and set as the conditional values the rainfall amounts
recorded in the year 2000, the year of the largest precipitation event recorded there between
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
Conditional simulation of max-stable processes 15
Table 3. Distribution of the partition size for the rainfall and temperature data estimated from
simulated Markov chains of length 15000 and 10000 respectively
Rainfall Temperature
Partition size 1 2 3 4 5 6 7–24 1 2 3 4 5–16
Frequencies (%) 66·2 28·0 4·8 0·5 0·2 0·2 <0·05 2·47 21·55 64·63 10·74 0·61
1962–2008, see Figure 2. The largest and smallest distances between the conditioning locations
are around 55km and just over 4km respectively.
A Brown–Resnick process having semivariogram γ(h) = (h/λ)κ was fitted using the max-
imum pairwise likelihood estimator introduced by Padoan et al. (2010) to simultaneously es-
timate the marginal parameters and the spatial dependence parameters λ and κ. As in Davison
et al. (2012), the marginal parameters were described by η(x) = β0,η + β1,ηlon(x) + β2,ηlat(x),
σ(x) = β0,σ + β1,σlon(x) + β2,σlat(x), ξ(x) = β0,ξ, where η(x), σ(x), ξ(x) are the location,
scale and shape parameters of the generalized extreme value distribution and lon(x), lat(x) de-
note the longitude and latitude of the stations at which the data are observed. The maximum
pairwise likelihood estimates and standard errors for λ and κ are 38 (14) and 0·69 (0·07) and
give a practical extremal range, i.e., the distance h+ such that θ(h+) = 1·7, of around 115km;
see the right panel of Figure 2.
Table 3 shows the distribution of the partition size estimated from a Markov chain of length
15000. Around 65% of the time the summer maxima observed at the 24 conditioning locations
belonged to a single extremal function, i.e., only one storm event, and around 30% of the time
to two different storms. Since the simulated Markov chain keeps a trace of all the simulated
partitions, we looked at the partitions of size two and saw that around 65% of the time, at least
one of the four up–north conditioning locations was impacted by one extremal function while
the remaining 20 locations were always influenced by another one.
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
16 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
20
40
60
80
30
40
50
60
70
80
90
40
60
80
100
1.00
1.02
1.04
1.06
1.08
1.10
Fig. 3. From left to right, maps on a 50× 50 grid of the
pointwise 0·025, 0·5 and 0·975 sample quantiles for rain-
fall (mm) obtained from 10000 conditional simulations of
Brown–Resnick processes having semi variogram γ(h) =
(h/38)0·69. The rightmost panel plots the ratio of the point-
wise confidence intervals with and without taking estima-
tion uncertainties into account. The squares show the con-
ditional locations.
Figure 3 plots the pointwise 0·025, 0·5 and 0·975 sample quantiles obtained from 10000 con-
ditional simulations of our fitted Brown–Resnick process. The conditional median provides a
point estimate for the rainfall at an ungauged location and the 0·025 and 0·975 conditional quan-
tiles a 95% pointwise confidence interval. As indicated by Figure 1, the shape parameter κ has
a major impact on the regularity of paths and on the width of the confidence interval. The value
κ ≈ 0·69 corresponds to very wiggly sample paths and wider confidence intervals. To assess the
impact of parameter uncertainties on conditional simulations, the ratio of the width of the confi-
dence intervals with and without parameter uncertainty is shown in the right panel of Figure 3.
The uncertainties were taken into account by sampling from the asymptotic distribution of the
maximum composite likelihood estimator, and drawing one conditional simulation for each re-
alization. These ratios show no clear spatial pattern and the width of the confidence interval is
increased by at most 10%.
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
Conditional simulation of max-stable processes 17
1000 2000 3000 4000
Arosa (1840)
Bad Ragaz (496)
Basel (316)
Bern (565)
Chateau d’Oex (985)
Davos (1590)Engelberg (1035)
Gd−St−Bernard (2472)
Locarno−Monti (366)
Lugano (273)
Montana (1508)
Montreux (405)
Neuchatel (485)
Oeschberg (483)
Santis (2490)Zurich (556)
0 25 50 75 100(km)
(m)
2.5 3 3.5 4 4.5
(°C)
Fig. 4. Left: Topographical map of Switzerland showing
the sites and altitudes in metres above sea level of 16
weather stations for which annual maxima temperature
data are available. Right: Map of temperature anomalies
(◦C), i.e., the difference between the pointwise medians
obtained from 10000 conditional simulations and uncondi-
tional medians estimated from the fitted Schlather process.
5·2. Extreme temperatures in Switzerland
The data considered here are annual maximum temperatures recorded at 16 sites in Switzer-
land during the period 1961–2005, see Figure 4. Following Davison and Gholamrezaee (2012),
we fit a Schlather process with an isotropic powered exponential correlation function and trend
surfaces η(x) = β0,η + β1,ηalt(x), σ(x) = β0,σ, ξ(x) = β0,ξ + β1,ξalt(x), where alt(x) de-
notes the altitude above mean sea level in kilometres and {η(x), σ(x), ξ(x)} are the location,
scale and shape parameters of the generalized extreme value distribution at location x. The spatial
dependence parameter estimates and their standard errors are λ = 260 (149) and κ = 0.52 (0.12)
and yield a fitted extremal coefficient function similar to our test case ρ3 in Section 4.
In year 2003, western Europe endured a heatwave believed to be most severe recorded since
1540 (Beniston, 2004). Switzerland was severely impacted by this event: the nationwide record
temperature of 41·5◦C was recorded that year in Grono, Graubunden, near Lugano. For our anal-
ysis we condition on the maximum temperatures observed in 2003. Based on the fitted Schlather
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
18 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
model, we simulate a Markov chain of effective length 10000 with a burn-in period of length
500 and a thinning lag of 100 iterations. The distribution of the partition size estimated from
these Markov chains is shown in Table 3. Around 90% of the time the conditional realizations
were a consequence of at most three extremal functions. Since our original observations were not
summer maxima but maximum daily values, a close inspection of the times series in year 2003
reveals that the hottest temperatures occurred between the 11th and 13th of August and, to some
extent, corroborates the distribution of Table 3.
The right panel of Figure 4 shows the spatial distribution of temperature anomalies, i.e., the
difference between the pointwise conditional medians obtained from 10000 conditional simu-
lations and the pointwise unconditional medians estimated from the fitted Schlather model. As
expected, the largest deviations occur in the plateau region of Switzerland while appreciably
smaller values are found in the Alps. The differences range between 2·5◦C and 4·75◦C and are
consistent with the values reported by climatologists for mean values (Beniston, 2004).
ACKNOWLEDGEMENTS
M. Ribatet was partly funded by the MIRACCLE-GICC and McSim ANR projects. The
authors thank MeteoSuisse and Dr. S. A. Padoan for providing the precipitation data, Prof.
A. C. Davison and Dr. M. M. Gholamrezaee for providing the temperature data and Dr. M. Oest-
ing for pointing out an error in the computations for the Brown–Resnick model.
APPENDIX
Brown–Resnick model
For all x ∈ X k and Borel set A ⊂ Rk
Λx(A) =
∫ ∞0
pr [ζ exp{W (x)− γ(x)} ∈ A] ζ−2dζ =
∫ ∞0
∫Rk
1{ζ exp{y−γ(x)}∈A}fx(y)dyζ−2dζ,
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
Conditional simulation of max-stable processes 19
where fx denotes the density of the random vector W (x), i.e., a centered Gaussian random vector with
covariance matrix Σx and variance 2γ(x). The change of variables z = ζ exp{y − γ(x)} and r = log ζ
yields
Λx(A) =
∫ ∞−∞
∫A
fx{log z − r + γ(x)}k∏i=1
z−1i dze−rdr =
∫A
λx(z)dz
with
λx(z) =
k∏i=1
z−1i
∫ ∞−∞
fx{log z − r + γ(x)}e−rdr.
Since
fx{log z − r + γ(x)}e−r = (2π)−k/2|Σx|−1/2 exp
{−1
2P (r)
},
with
P (r) = r21Tk Σ−1x 1k − 2r[1Tk Σ−1x {log z + γ(x)} − 1
]+ {log z + γ(x)}TΣ−1x {log z + γ(x)},
standard computations for Gaussian integrals give
λx(z) = Cx exp
(−1
2log zTQx log z + Lx log z
) k∏i=1
z−1i .
The conditional intensity function is
λs|x,z(u) =C(s,x)
Cxexp
{−1
2log (u, z)
TQ(s,x) log(u, z) + L(s,x) log(u, z) +
1
2log zTQx log z − Lx log z
} m∏i=1
u−1i ,
and since log(u, z) = Jm,k log u+ Jm,k log z, it is not difficult to show that
λs|x,z(u) =C(s,x)
Cxexp
{−1
2(log u− µs|x,z)TΣ−1s|x(log u− µs|x,z)
} m∏i=1
u−1i .
Finally, the relation C(s,x)/Cx = (2π)−m/2|Σs|x|−1/2 is a simple consequence of the normalization∫λs|x,z(u)du = 1.
Schlather model
For all x ∈ X k and Borel sets A ⊂ Rk
Λx(A) =
∫ ∞0
pr{
(2π)1/2ζε(x) ∈ A}ζ−2dζ =
∫ ∞0
∫Rk
1{(2π)1/2ζy∈A}fx(y)dyζ−2dζ,
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
20 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
where fx denotes the density of the random vector ε(x), i.e., a centered Gaussian random vector with
covariance matrix Σx. The change of variable z = (2π)1/2ζy gives
Λx(A) = (2π)−k/2∫ ∞0
∫A
fx
{z
(2π)1/2ζ
}ζ−(k+2)dzdζ
= (2π)−k|Σx|−1/2∫ ∞0
∫A
exp
(− 1
4πζ2zTΣ−1x z
)ζ−(k+2)dzdζ
= (2π)−k|Σx|−1/2∫A
2π
zTΣ−1x zE[Xk−1]dz, X ∼Weibull
{(4π
zTΣ−1x z
)1/2
, 2
}
= (2π)−k|Σx|−1/2∫A
2π
zTΣ−1x z
(4π
zTΣ−1x z
)(k−1)/2
Γ
(k + 1
2
)dz
=
∫A
λx(z)dz,
where λx(z) = π−(k−1)/2|Σx|−1/2ax(z)−(k+1)/2Γ {(k + 1)/2} and ax(z) = zTΣ−1x z.
For all u ∈ Rm the conditional intensity function is
λs|x,z(u) = π−m/2|Σ(s,x)|−1/2
|Σx|−1/2
{a(s,x)(u, z)
ax(z)
}−(m+k+1)/2
ax(z)−m/2Γ(m+k+1
2
)Γ(k+12
) .
We start by focusing on the ratio a(s,x)(u, z)/ax(z). Since Σs Σs:x
Σx:s Σx
−1
=
(Σs − Σs:xΣ−1x Σx:s)−1 −(Σs − Σs:xΣ−1x Σx:s)
−1Σs:xΣ−1x
−Σ−1x Σx:s(Σs − Σs:xΣ−1x Σx:s)−1 Σ−1x + Σ−1x Σx:s(Σs − Σs:xΣ−1x Σx:s)
−1Σs:xΣ−1x
,
straightforward algebra shows that
a(s,x)(u, z)
ax(z)= 1 +
(u− µ)T Σ−1(u− µ)
k + 1, µ = Σs:xΣ−1x z, Σ =
ax(z)
k + 1
(Σs − Σs:xΣ−1x Σx:s
).
We now simplify the ratio |Σ(s,x)|/|Σx|. Using the fact that
Σ(s,x) =
Σs Σs:x
Σx:s Σx
=
Im Σs:x
0k,m Σx
Σs − Σs:xΣ−1x Σx:s 0m,k
Σ−1x Σx:s Ik
,
combined with some more algebra yields
|Σ(s,x)||Σx|
= |Σs − Σs:xΣ−1x Σx:s| ={k + 1
ax(z)
}m|Σ|.
Using the two previous results it is easily found that
λs|x,z(u) = π−m/2(k + 1)−m/2|Σ|−1/2{
1 +(u− µ)T Σ−1(u− µ)
k + 1
}−(m+k+1)/2Γ(m+k+1
2
)Γ(k+12
) ,
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
Conditional simulation of max-stable processes 21
which corresponds to the density of a multivariate Student distribution with the stated parameters.
REFERENCES
BENISTON, M. (2004). The 2003 heat wave in Europe: A shape of things to come? An analysis based on Swiss
climatological data and model simulations. Geophys. Res. Lett. 31, L02202.
BROWN, B. M. & RESNICK, S. (1977). Extreme values of independent stochastic processes. J. Appl. Prob. 14,
732–739.
BUISHAND, T. A., DE HAAN, L. & ZHOU, C. (2008). On spatial extremes: With application to a rainfall problem.
Annals of Applied Statistics 2, 624–642.
CHILES, J.-P. & DELFINER, P. (1999). Geostatistics: Modelling Spatial Uncertainty. New York: Wiley.
COOLEY, D., NAVEAU, P. & PONCET, P. (2006). Variograms for spatial max-stable random fields. In Dependence in
Probability and Statistics, P. Bertail, P. Soulier, P. Doukhan, P. Bickel, P. Diggle, S. Fienberg, U. Gather, I. Olkin
& S. Zeger, eds., vol. 187 of Lecture Notes in Statistics. Springer New York, pp. 373–390.
DAVIS, R. & RESNICK, S. (1989). Basic properties and prediction of max-ARMA processes. Advances in Applied
Probability 21, 781–803.
DAVIS, R. & RESNICK, S. (1993). Prediction of stationary max-stable processes. Annals Of Applied Probability 3,
497–525.
DAVISON, A. C. & GHOLAMREZAEE, M. M. (2012). Geostatistics of extremes. Proceedings of the Royal Society
A: Mathematical, Physical and Engineering Science 468, 581–608.
DAVISON, A. C., PADOAN, S. A. & RIBATET, M. (2012). Statistical modelling of spatial extremes. Statistical
Science 7, 161–186.
DE HAAN, L. (1984). A spectral representation for max-stable processes. The Annals of Probability 12, 1194–1204.
DE HAAN, L. & PEREIRA, T. T. (2006). Spatial extremes: Models for the stationary case. The Annals of Statistics
34, 146–168.
DE HAAN, L. & FEREIRA, A. (2006). Extreme Value Theory: An Introduction. Springer, New-York.
GENTON, M. G., MA, Y. & SANG, H. (2011). On the likelihood function of Gaussian max-stable processes.
Biometrika 98, 481–488.
GENZ, A. (1992). Numerical computation of multivariate normal probabilities. J. Comp. Graph Stat 1, 141–149.
KABLUCHKO, Z., SCHLATHER, M. & DE HAAN, L. (2009). Stationary max-stable fields associated to negative
definite functions. Ann. Prob. 37, 2042–2065.
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
22 C. DOMBRY, F. EYI-MINKO AND M. RIBATET
KABLUCHKO, Z. & SCHLATHER, M. (2010). Ergodic properties of max-infinitely divisible processes. Stochastic
Processes and their Applications, 120, 281–295.
LIU, J., WONG, W. & KONG, A. (1995). Correlation structure and convergence rate of the Gibbs sampler with
various scans. Journal of the Royal Statistical Society Series B 57, 157–169.
MANDELBROT, B. (1982). The Fractal Geometry of Nature. New York: W. H. Freeman.
OESTING, M., KABLUCHKO, Z. & SCHLATHER, M. (2012). Simulation of Brown–Resnick processes. Extremes
15, 89–107.
PADOAN, S. A., RIBATET, M. & SISSON, S. (2010). Likelihood-based inference for max-stable processes. Journal
of the American Statistical Association 105, 263–277.
R DEVELOPMENT CORE TEAM (2012). R: A Language and Environment for Statistical Computing. R Foundation
for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0.
RIBATET, M. (2011). SpatialExtremes: Modelling Spatial Extremes. R package version 1.8-5.
SCHLATHER, M. (2002). Models for stationary max-stable random fields. Extremes 5, 33–44.
SCHLATHER, M. (2012). RandomFields: Simulation and Analysis of Random Fields. R package version 2.0.54.
SCHLATHER, M. & TAWN, J. (2003). A dependence measure for multivariate and spatial extremes: Properties and
inference. Biometrika 90, 139–156.
WANG, Y. (2010). maxLinear: Conditional sampling for max-linear models. R package version 1.0.
WANG, Y. & STOEV, S. A. (2011). Conditional sampling for spectrally discrete max-stable random fields. Advances
in Applied Probability 43, 461–483.