--- title: "Exact unconditional sampling from max-stable random vectors" author: "Léo Belzile, HEC Montréal" date: "`r Sys.Date()`" output: rmarkdown::html_vignette: toc: false vignette: > %\VignetteIndexEntry{Exact unconditional sampling from max-stable random vectors} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} bibliography: mevvignette.bib --- The `mev` package was originally introduced to implement the exact unconditional sampling algorithms in @Dombry:2016. The two algorithms therein allow one to simulate simple max-stable random vectors. The implementation will work efficiently for moderate dimensions. # Functions and use There are two main functions, `rmev` and `rmevspec`. `rmev` samples from simple max-stable processes, meaning it will return an $n \times d$ matrix of samples, where each of the column has a sample from a unit Frechet distribution. In constrast, `rmevspec` returns sample on the unit simplex from the spectral (or angular) measure. One could use this to test estimation based on spectral densities, or to construct samples from Pareto processes. The syntax is ```{r} #| eval: true #| echo: true library(mev) #Sample of size 1000 from a 5-dimensional logistic model x <- rmev(n=1000, d=5, param=0.5, model="log") #Marginal parameters are all standard Frechet, meaning GEV(1,1,1) apply(x, 2, function(col){ismev::gev.fit(col, show=FALSE)$mle}) #Sample from the corresponding spectral density w <- rmevspec(n=1000, d=5, param=0.5, model="log") #All rows sum to 1 by construction head(rowSums(w)) #The marginal mean is 1/d round(colMeans(w),2) ``` # Description of the models implemented The different models implemented are described in @Dombry:2016, but some other models can be found and are described here. Throughout, we consider $d$-variate models and let $\mathbb{B}_d$ be the collection of all nonempty subsets of $\{1, \ldots, d\}$. ## Logistic The logistic model (`log`) of @Gumbel:1960 has distribution function \begin{align*} \Pr(\boldsymbol{X} \leq \boldsymbol{x})= \exp \left[ - \left(\sum_{i=1}^{n} {x_i}^{-\alpha}\right)^{\frac{1}{\alpha}}\right] \end{align*} for $\alpha>1$. The spectral measure density is \begin{align*} h_{\boldsymbol{W}}(\boldsymbol{w})=\frac{1}{d}\frac{\Gamma(d-\alpha)}{\Gamma(1-\alpha)}\alpha^{d-1}\left( \prod_{j=1}^d w_j\right)^{-(\alpha+1)}\left(\sum_{j=1}^d w_j^{-\alpha}\right)^{1/\alpha-d}, \qquad \boldsymbol{w} \in \mathbb{S}_d \end{align*} ## Asymmetric logistic distribution The `alog` model was proposed by @Tawn:1990. It shares the same parametrization as the `evd` package, merely replacing the algorithm for the generation of logistic variates. The distribution function of the $d$-variate asymmetric logistic distribution is \begin{align*} \Pr(\boldsymbol{X} \le \boldsymbol{x}) = \exp \left[ -\sum_{b \in \mathbb{B}_d}\left(\sum_{i \in b} \left(\frac{\theta_{i, b}}{x_i}\right)^{\alpha_b}\right)^{\frac{1}{\alpha_b}}\right], \end{align*} The parameters $\theta_{i, b}$ must be provided in a list and represent the asymmetry parameter. The sampling algorithm, from @Stephenson:2003 gives some insight on the construction mechanism as a max-mixture of logistic distributions. Consider sampling $\boldsymbol{Z}_b$ from a logistic distribution of dimension $|b|$ (or Fréchet variates if $|b|=1)$ with parameter $\alpha_b$ (possibly recycled). Each marginal value corresponds to the maximum of the weighted corresponding entry. That is, $X_{i}=\max_{b \in \mathbb{B}_d}\theta_{i, b}Z_{i,b}$ for all $i=1, \ldots, d$. The max-mixture is valid provided that $\sum_{b \in \mathbb{B}_d} \theta_{i,b}=1$ for $i=1, \ldots, d.$ As such, empirical estimates of the spectral measure will almost surely place mass on the inside of the simplex rather than on subfaces. ## Negative logistic distribution The `neglog` distribution function due to @Galambos:1975 is \begin{align*} \Pr(\boldsymbol{X} \le \boldsymbol{x}) = \exp \left[ -\sum_{b \in \mathbb{B}_d} (-1)^{|b|}\left(\sum_{i \in b} {x_i}^{\alpha}\right)^{-\frac{1}{\alpha}}\right] \end{align*} for $\alpha \geq 0$ [@Dombry:2016]. The associated spectral density is \begin{align*} h_{\boldsymbol{W}}(\boldsymbol{w}) = \frac{1}{d} \frac{\Gamma(1/\alpha+1)}{\Gamma(1/\alpha + d-1)} \alpha^d\left(\prod_{i=1}^d w_j\right)^{\alpha-1}\left(\sum_{i=1}^d w_i^{\alpha}\right)^{-1/\alpha-d} \end{align*} ## Asymmetric negative logistic distribution The asymmetric negative logistic (`aneglog`) model is alluded to in @Joe:1990 as a generalization of the Galambos model. It is constructed in the same way as the asymmetric logistic distribution; see Theorem~1 in @Stephenson:2003. Let $\alpha_b \leq 0$ for all $b \in \mathbb{B}_d$ and $\theta_{i, b} \geq 0$ with $\sum_{b \in \mathbb{B}_d} \theta_{i, b} =1$ for $i=1, \ldots, d$; the distribution function is \begin{align*} \Pr(\boldsymbol{X} \le \boldsymbol{x}) = \exp \left[ -\sum_{b \in \mathbb{B}_d}(-1)^{|b|} \left\{\sum_{i \in b} \left(\frac{\theta_{i, b}}{x_i}\right)^{\alpha_b}\right\}^{\frac{1}{\alpha_b}}\right]. \end{align*} In particular, it does not correspond to the ``negative logistic distribution'' given in e.g., Section 4.2 of @Coles:1991 or Section3.5.3 of @Kotz:2000. The latter is not a valid distribution function in dimension $d \geq 3$ as the constraints therein on the parameters $\theta_{i, b}$ are necessary, but not sufficient. @Joe:1990 mentions generalizations of the distribution as given above but the constraints were not enforced elsewhere in the literature. The proof that the distribution is valid follows from Theorem~1 of @Stephenson:2003 as it is a max-mixture. Note that the parametrization of the asymmetric negative logistic distribution does not match the bivariate implementation of `rbvevd`. ## Multilogistic distribution This multivariate extension of the logistic, termed multilogistic (`bilog`) proposed by @Boldi:2009, places mass on the interior of the simplex. Let $\boldsymbol{W} \in \mathbb{S}_d$ be the solution of \begin{align*} \frac{W_j}{W_d}=\frac{C_jU_j^{-\alpha_j}}{C_dU_d^{-\alpha_d}}, \quad j=1, \ldots, d \end{align*} where $C_j=\Gamma(d-\alpha_j)/\Gamma(1-\alpha_j)$ for $j=1, \ldots, d$ and $\boldsymbol{U} \in \mathbb{S}_d$ follows a $d$-mixture of Dirichlet with the $j$th component being $\mathcal{D}(\boldsymbol{1}-\delta_{j}\alpha_j)$, so that the mixture has density function \begin{align*} h_{\boldsymbol{U}}(\boldsymbol{u})=\frac{1}{d} \sum_{j=1}^d \frac{\Gamma(d-\alpha_j)}{\Gamma(1-\alpha_j)} u_j^{-\alpha_j} \end{align*} for $0<\alpha_j <1, j=1, \ldots, d$. The spectral density of the multilogistic distribution is thus \begin{align*} h_{\boldsymbol{W}}(\boldsymbol{w}) = \frac{1}{d} \left(\sum_{j=1}^d \alpha_ju_j\right)^{-1} \left(\prod_{j=1}^d \alpha_ju_d \right)\left(\sum_{j=1}^d \frac{\Gamma(d-\alpha_j)}{\Gamma(1-\alpha_j)}u_j^{-\alpha_j}\right)\prod_{j=1}^d w_j^{-1} \end{align*} for $\alpha_j \in (0,1)$ $(j=1, \ldots, d)$. ## Coles and Tawn Dirichlet distribution The Dirichlet (`ct`) model of @Coles:1991 \begin{align*} h_{\boldsymbol{W}}(\boldsymbol{w}) = \frac{1}{d} \frac{\Gamma \left(1+\sum_{j=1}^d \alpha_j\right)}{\prod_{j=1}^d \alpha_jw_j} \left(\sum_{j=1}^d \alpha_jw_j\right)^{-(d+1)}\prod_{j=1}^d \alpha_j \prod_{j=1}^d \left(\frac{\alpha_jw_j}{\sum_{k=1}^d \alpha_kw_k}\right)^{\alpha_j-1} \end{align*} for $\alpha_j>0.$ ## Scaled extremal Dirichlet The angular density of the scaled extremal Dirichlet (`sdir`) model with parameters $\rho > -\min(\boldsymbol{\alpha})$ and $\boldsymbol{\alpha} \in \mathbb{R}^{d}_{+}$ is given, for all $\boldsymbol{w} \in \mathbb{S}_d$, by \begin{align*} h_{\boldsymbol{W}}(\boldsymbol{w})=\frac{\Gamma(\bar{\alpha}+\rho)}{d\rho^{d-1}\prod_{i=1}^d\Gamma(\alpha_i)} \bigl\langle\{\boldsymbol{c}(\boldsymbol{\alpha},\rho)\}^{1/\rho},\boldsymbol{w}^{1/\rho}\bigr\rangle^{-\rho-\bar{\alpha}}\prod_{i=1}^{d} \{c(\alpha_i,\rho)\}^{\alpha_i/\rho}w_i^{\alpha_i/\rho-1}. \end{align*} where $\boldsymbol{c}(\boldsymbol{\alpha},\rho)$ is the $d$-vector with entries $\Gamma(\alpha_i+\rho)/\Gamma(\alpha_i)$ for $i=1, \ldots, d$ and $\langle \cdot, \cdot \rangle$ denotes the inner product between two vectors. ## Huesler--Reiss The Huesler--Reiss model (`hr`), due to @Husler:1989, is a special case of the Brown--Resnick process. While @Engelke:2015 state that H\"usler--Reiss variates can be sampled following the same scheme, the spatial analog is conditioned on a particular site ($\boldsymbol{s}_0$), which complicates the comparisons with the other methods. Let $I_{-j}=\{1, \ldots, d\} \setminus \{j\}$ and $\lambda_{ij}^2 \geq 0$ be entries of a strictly conditionally negative definite matrix $\boldsymbol{\Lambda}$, for which $\lambda_{ij}^2=\lambda_{ji}^2$. Then, following @Nikoloulopoulos:2009 (Remark~2.5) and @Huser:2013, we can write the distribution function as \begin{align*} \Pr(\boldsymbol{X} \le \boldsymbol{x}) = \exp \left[ -\sum_{j=1}^d \frac{1}{x_j} \Phi_{d-1, \boldsymbol{\Sigma}_{-j}} \left( \lambda_{ij}- \frac{1}{2\lambda_{ij}} \log\left(\frac{x_j}{x_i}\right), i \in I_{-j}\right)\right]. \end{align*} where the partial correlation matrix $\boldsymbol{\Sigma}_{-j}$ has elements \begin{align*} \varrho_{i,k; j}= \frac{\lambda_{ij}^2+\lambda_{kj}^2-\lambda_{ik}^2}{2\lambda_{ij}\lambda_{kj}} \end{align*} and $\lambda_{ii}=0$ for all $i \in I_{-j}$ so that the diagonal entries $\varrho_{i,i; j}=1$. @Engelke:2015 uses the covariance matrix with entries are $\varsigma=2(\lambda_{ij}^2+\lambda_{kj}^2-\lambda_{ik}^2)$, so the resulting expression is evaluated at $2\boldsymbol{\lambda}_{.j}^2-\log({x_j}/{\boldsymbol{x}_{-j}})$ instead. We recover the same expression by standardizing, since this amounts to division by the standard deviations $2\boldsymbol{\lambda}_{.j}$ The \texttt{evd} package implementation has a bivariate implementation of the H\"usler--Reiss distribution with dependence parameter $r$, with $r_{ik}=1/\lambda_{ik}$ or $2/r=\sqrt{2\gamma(\boldsymbol{h})}$ for $\boldsymbol{h}=\|\boldsymbol{s}_i-\boldsymbol{s}_i\|$ for the Brown--Resnick model. In this setting, it is particularly easy since the only requirement is non-negativity of the parameter. For inference in dimension $d>2$, one needs to impose the constraint $\boldsymbol{\Lambda}=\{\lambda_{ij}^2\}_{i, j=1}^d \in \mathcal{D}$ (cf. @Engelke:2015, p.3), where \begin{multline*} \mathcal{D}=\Biggl\{\mathbf{A}\in [0, \infty)^{d\times d}: \boldsymbol{x}^\top\!\!\mathbf{A}\boldsymbol{x} <0, \ \forall \ \boldsymbol{x} \in \mathbb{R}^{d} \setminus\{\boldsymbol{0}\} \\ \qquad \text{ with } \sum_{i=1}^d x_i=0, a_{ij}=a_{ji}, a_{ii}=0 \ \forall \ i, j \in \{1,\ldots, d\}\Biggr\} \end{multline*} denotes the set of symmetric conditionally negative definite matrices with zero diagonal entries. An avenue to automatically satisfy these requirements is to optimize over a symmetric positive definite matrix parameter $\boldsymbol{\varSigma}=\mathbf{L}^\top\mathbf{L}$, where $\mathbf{L}$ is an upper triangular matrix whose diagonal element are on the log-scale to ensure uniqueness of the Cholesky factorization; see @Pinheiro:1996. By taking \begin{align*} \boldsymbol{\Lambda}(\boldsymbol{\varSigma})= \begin{pmatrix} 0 & \mathrm{diag} (\boldsymbol{\varSigma})^\top \\ \mathrm{diag}(\boldsymbol{\varSigma}) & \boldsymbol{1}\mathrm{diag}(\boldsymbol{\varSigma})^\top + \mathrm{diag}(\boldsymbol{\varSigma})\boldsymbol{1}^\top - 2 \boldsymbol{\varSigma} \end{pmatrix} \end{align*} one can perform unconstrained optimization for the non-zero elements of $\mathbf{L}$ which are in one-to-one correspondence with those of $\boldsymbol{\Lambda}$. It easily follows that generating $\boldsymbol{Z}$ from a $d-1$ dimensional log-Gaussian distribution with covariance $\mathsf{Co}(Z_i, Z_k)=2(\lambda_{ij}^2+\lambda_{kj}^2-\lambda_{ik}^2)$ for $i, k \in I_{-j}$ with mean vector $-2\lambda_{\bullet j}^2$ gives the finite dimensional analog of the Brown--Resnick process in the mixture representation of @Dombry:2016. The \texttt{rmev} function checks conditional negative definiteness of the matrix. The easiest way to do so negative definiteness of $\boldsymbol{\Lambda}$ with real entries is to form $\tilde{\boldsymbol{\Lambda}}=\mathbf{P}\boldsymbol{\Lambda}\mathbf{P}^\top$, where $\mathbf{P}$ is an $d \times d$ matrix with ones on the diagonal, $-1$ on the $(i, i+1)$ entries for $i=1, \ldots d-1$ and zeros elsewhere. If the matrix $\boldsymbol{\Lambda} \in \mathcal{D}$, then the eigenvalues of the leading $(d-1) \times (d-1)$ submatrix of $\tilde{\boldsymbol{\Lambda}}$ will all be negative. For a set of $d$ locations, one can supply the variogram matrix as valid input to the method. ## Brown--Resnick process The Brown--Resnick process (`br`) is the functional extension of the H\"usler--Reiss distribution, and is a max-stable process associated with the log-Gaussian distribution. It is often in the spatial setting conditioned on a location (typically the origin). Users can provide a variogram function that takes distance as argument and is vectorized. If `vario` is provided, the model will simulate from an intrinsically stationary Gaussian process. The user can alternatively provide a covariance matrix `sigma` obtained by conditioning on a site, in which case simulations are from a stationary Gaussian process. See @Engelke:2015 or @Dombry:2016 for more information. ## Extremal Student The extremal Student (`extstud`) model of @Nikoloulopoulos:2009, eq. 2.8, with unit Fréchet margins is \begin{align*} \Pr(\boldsymbol{X} \le \boldsymbol{x}) = \exp \left[-\sum_{j=1}^d \frac{1}{x_j} T_{d-1, \nu+1, \mathbf{R}_{-j}}\left( \sqrt{\frac{\nu+1}{1-\rho_{ij}^2}} \left[\left(\frac{x_i}{x_j}\right)^{1/\nu}\!\!\!-\rho_{ij}\right], i \in I_{-j} \right)\right], \end{align*} where $T_{d-1}$ is the distribution function of the $d-1 $ dimensional Student-$t$ distribution and the partial correlation matrix $\mathbf{R}_{-j}$ has diagonal entry \[r_{i,i;j}=1, \qquad r_{i,k;j}=\frac{\rho_{ik}-\rho_{ij}\rho_{kj}}{\sqrt{1-\rho_{ij}^2}\sqrt{1-\rho_{kj}^2}}\] for $i\neq k, i, k \in I_{-j}$. The user must provide a valid correlation matrix (the function checks for diagonal elements), which can be obtained from a variogram. ## Dirichlet mixture The Dirichlet mixture (`dirmix`) proposed by @Boldi:2007, see @Dombry:2016 for details on the mixture. The spectral density of the model is \begin{align*} h_{\boldsymbol{W}}(\boldsymbol{w}) = \sum_{k=1}^m \pi_k \frac{\Gamma(\alpha_{1k}+ \cdots + \alpha_{dk})}{\prod_{i=1}^d \Gamma(\alpha_{ik})} \left(1-\sum_{i=1}^{d-1} w_i\right)^{\alpha_{dk}-1}\prod_{i=1}^{d-1} w_{i}^{\alpha_{ik}-1} \end{align*} The argument `param` is thus a $d \times m$ matrix of coefficients, while the argument for the $m$-vector `weights` gives the relative contribution of each Dirichlet mixture component. ## Smith model The Smith model (`smith`) is from the unpublished report of @Smith:1990. It corresponds to a moving maximum process on a domain $\mathbb{X}$. The de Haan representation of the process is \begin{align*} Z(x)=\max_{i \in \mathbb{N}} \zeta_i h(x-\eta_i), \qquad \eta_i \in \mathbb{X} \end{align*} where $\{\zeta_i, \eta_i\}_{i \in \mathbb{N}}$ is a Poisson point process on $\mathbb{R}_{+} \times \mathbb{X}$ with intensity measure $\zeta^{-2}\mathrm{d} \zeta \mathrm{d} \eta$ and $h$ is the density of the multivariate Gaussian distribution. Other $h$ could be used in principle, but are not implemented. # References {-}