h_of_dist: Finds the distance of each element in a matrix x to the its...

In genscore: Generalized Score Matching Estimators

Finds the distance of each element in a matrix x to the its boundary of the domain while fixing the others in the same row (dist(x, domain)), and calculates element-wise h(dist(x, domain)) and h\'(dist(x, domain)) (w.r.t. each element in x).

Description

Finds the distance of each element in a matrix x to its boundary of the domain while fixing the others in the same row (dist(x, domain)), and calculates element-wise h(dist(x, domain)) and h\'(dist(x, domain)) (w.r.t. each element in x).

Usage

h_of_dist(h_hp, x, domain, log = FALSE)

Arguments

h_hp	A function, the h and hp (the derivative of h) functions. h_hp(x) should return a list of elements hx (h(x)) and hpx (hp(x)), both of which have the same size as x.
x	An n by p matrix, the data matrix, where n is the sample size and p the dimension.
domain	A list returned from make_domain() that represents the domain.
log	A logical, defaults to FALSE. If TRUE, assumes that h_hp contains in fact the log of h and hp, and this function will return the log of h(dist(x, domain)) and abs(h\'(dist(x, domain))) along with the sign of h\'(dist(x, domain)).

Details

Define dist(x, domain) as the matrix whose i,j-th component is the distance of x_{i,j} to the boundary of domain, assuming x_{i,-j} are fixed. The matrix has the same size of x (n by p), or if domain$type == "simplex" and x has full dimension p, it has p-1 columns.
Define dist\'(x, domain) as the component-wise derivative of dist(x, domain) in its components. That is, its i,j-th component is 0 if x_{i,j} is unbounded or is bounded from both below and above or is at the boundary, or -1 if x_{i,j} is closer to its lower boundary (or if its bounded from below but unbounded from above), or 1 otherwise.
h_of_dist(h_hp, x, domain) simply returns h_hp(dist(x, domain))$hx and h_hp(dist(x, domain))$hpx * dist\'(x, domain) (element-wise derivative of h_hp(dist(x, domain))$hx w.r.t. x).

Value

If log == FALSE, a list that contains h(dist(x, domain)) and h\'(dist(x, domain)).

hdx	h(dist(x, domain)).
hpdx	hp(dist(x, domain)).

If log == TRUE, a list that contains the log of h(dist(x, domain)) and abs(h\'(dist(x, domain))) as well as the sign of h\'(dist(x, domain)).

log_hdx	log(h(dist(x, domain))).
log_hpdx	log(abs(hp(dist(x, domain)))).
sign_hpdx	sign(hp(dist(x, domain))).

Examples

n <- 20
p <- 10
eta <- rep(0, p)
K <- diag(p)
dm <- 1 + (1-1/(1+4*exp(1)*max(6*log(p)/n, sqrt(6*log(p)/n))))

# Gaussian on R^p:
domain <- make_domain("R", p=p)
x <- mvtnorm::rmvnorm(n, mean=solve(K, eta), sigma=solve(K))
# Equivalently:

x2 <- gen(n, setting="gaussian", abs=FALSE, eta=eta, K=K, domain=domain,
       finite_infinity=100, xinit=NULL, burn_in=1000, thinning=100, verbose=FALSE)

h_hp <- get_h_hp("pow", 2) # For demonstration only
hd <- h_of_dist(h_hp, x, domain)
# hdx is all Inf and hpdx is all 0 since each coordinate is unbounded with domain R
c(all(is.infinite(hd$hdx)), all(hd$hpdx==0))


# exp on R_+^p:
domain <- make_domain("R+", p=p)
x <- gen(n, setting="exp", abs=FALSE, eta=eta, K=K, domain=domain, finite_infinity=100, 
       xinit=NULL, seed=2, burn_in=1000, thinning=100, verbose=FALSE)
h_hp <- get_h_hp("pow", 2) # For demonstration only
hd <- h_of_dist(h_hp, x, domain)
# hdx is x^2 and hpdx is 2*x; with domain R+, the distance of x to the boundary is just x itself
c(max(abs(hd$hdx - x^2)), max(abs(hd$hpdx - 2*x)))


# Gaussian on sum(x^2) > p with x allowed to be negative
domain <- make_domain("polynomial", p=p,
       ineqs=list(list("expression"=paste("sum(x^2)>", p), abs=FALSE, nonnegative=FALSE)))
x <- gen(n, setting="gaussian", abs=FALSE, eta=eta, K=K, domain=domain, finite_infinity=100, 
       xinit=NULL, seed=2, burn_in=1000, thinning=100, verbose=FALSE)
dist <- get_dist(x, domain)
quota <- p - (rowSums(x^2) - x^2) # How much should xij^2 at least be so that sum(xi^2) > p?
# How far is xij from +/-sqrt(quota), if quota >= 0?
dist_to_bound <- abs(x[quota >= 0]) - abs(sqrt(quota[quota >= 0]))
# Should be equal to our own calculations
max(abs(dist$dx[is.finite(dist$dx)] - dist_to_bound))
# dist'(x) should be the same as the sign of x
all(dist$dpx[is.finite(dist$dx)] == sign(x[quota >= 0]))
# quota is negative <-> sum of x_{i,-j}^2 already > p <-> xij unbounded given others
#      <-> distance to boundary is Inf
all(quota[is.infinite(dist$dx)] < 0)

h_hp <- get_h_hp("pow", 2) # For demonstration only
# Now confirm that h_of_dist indeed applies h and hp to dists
hd <- h_of_dist(h_hp, x, domain)
# hdx = dist ^ 2
print(max(abs(hd$hdx[is.finite(dist$dx)] - dist$dx[is.finite(dist$dx)]^2)))
# hdx = Inf if dist = Inf
print(all(is.infinite(hd$hdx[is.infinite(dist$dx)])))
 # hpdx = 2 * dist' * dist
print(max(abs(hd$hpdx[is.finite(dist$dx)] - 2*(dist$dpx*dist$dx)[is.finite(dist$dx)])))
print(all(hd$hpdx[is.infinite(dist$dx)] == 0)) # hpdx = 0 if dist = Inf


# gamma on ([0, 1] v [2,3])^p
domain <- make_domain("uniform", p=p, lefts=c(0,2), rights=c(1,3))
x <- gen(n, setting="gamma", abs=FALSE, eta=eta, K=K, domain=domain,
       xinit=NULL, seed=2, burn_in=1000, thinning=100, verbose=FALSE)
dist <- get_dist(x, domain)
# If 0 <= xij <= 1, distance to boundary is min(x-0, 1-x)
max(abs(dist$dx - pmin(x, 1-x))[x >= 0 & x <= 1])
# If 0 <= xij <= 1, dist'(xij) is 1 if it is closer to 0, or -1 if it is closer 1,
#   assuming xij %in% c(0, 0.5, 1) with probability 0
all((dist$dpx == 2 * (1-x > x) - 1)[x >= 0 & x <= 1])
# If 2 <= xij <= 3, distance to boundary is min(x-2, 3-x)
max(abs(dist$dx - pmin(x-2, 3-x))[x >= 2 & x <= 3])
# If 2 <= xij <= 3, dist'(xij) is 1 if it is closer to 2, or -1 if it is closer 3,
#   assuming xij %in% c(2, 2.5, 3) with probability 0
all((dist$dpx == 2 * (3-x > x-2) - 1)[x >= 2 & x <= 3])
h_hp <- get_h_hp("pow", 2) # For demonstration only
# Now confirm that h_of_dist indeed applies h and hp to dists
hd <- h_of_dist(h_hp, x, domain)
# hdx = dist ^ 2
print(max(abs(hd$hdx - dist$dx^2)))
# hpdx = 2 * dist' * dist
print(max(abs(hd$hpdx - 2*dist$dpx*dist$dx)))


# a0.6_b0.7 on {x1 > 1 && log(1.3) < x2 < 1 && x3 > log(1.3) && ... && xp > log(1.3)}
domain <- make_domain("polynomial", p=p, rule="1 && 2 && 3",
       ineqs=list(list("expression"="x1>1", abs=FALSE, nonnegative=TRUE),
                      list("expression"="x2<1", abs=FALSE, nonnegative=TRUE),
                      list("expression"="exp(x)>1.3", abs=FALSE, nonnegative=FALSE)))
set.seed(1)
xinit <- c(1.5, 0.5, abs(stats::rnorm(p-2)) + log(1.3))
x <- gen(n, setting="ab_3/5_7/10", abs=FALSE, eta=eta, K=K, domain=domain, finite_infinity=100, 
       xinit=xinit, seed=2, burn_in=1000, thinning=100, verbose=FALSE)
dist <- get_dist(x, domain)
# x_{i1} has uniform bound [1, +Inf), so its distance to its boundary is x_{i1} - 1
max(abs(dist$dx[,1] - (x[,1] - 1)))
# x_{i2} has uniform bound [log(1.3), 1], so its distance to its boundary
#    is min(x_{i2} - log(1.3), 1 - x_{i2})
max(abs(dist$dx[,2] - pmin(x[,2] - log(1.3), 1 - x[,2])))
# x_{ij} for j >= 3 has uniform bound [log(1.3), +Inf), so its distance to its boundary is
#    simply x_{ij} - log(1.3)
max(abs(dist$dx[,3:p] - (x[,3:p] - log(1.3))))
# dist'(xi2) is 1 if it is closer to log(1.3), or -1 if it is closer 1,
#    assuming x_{i2} %in% c(log(1.3), (1+log(1.3))/2, 1) with probability 0
all((dist$dpx[,2] == 2 * (1 - x[,2] > x[,2] - log(1.3)) - 1))
all(dist$dpx[,-2] == 1) # x_{ij} for j != 2 is bounded from below but unbounded from above,
#    so dist'(xij) is always 1
h_hp <- get_h_hp("pow", 2) # For demonstration only
# Now confirm that h_of_dist indeed applies h and hp to dists
hd <- h_of_dist(h_hp, x, domain)
# hdx = dist ^ 2
print(max(abs(hd$hdx - dist$dx^2)))
# hpdx = 2 * dist' * dist
print(max(abs(hd$hpdx - 2*dist$dpx*dist$dx)))


# log_log model on {x in R_+^p: sum_j j * xj <= 1}
domain <- make_domain("polynomial", p=p,
       ineqs=list(list("expression"=paste(paste(sapply(1:p,
                           function(j){paste(j, "x", j, sep="")}), collapse="+"), "<1"),
                     abs=FALSE, nonnegative=TRUE)))
x <- gen(n, setting="log_log", abs=FALSE, eta=eta, K=K, domain=domain, finite_infinity=100, 
       xinit=NULL, seed=2, burn_in=1000, thinning=100, verbose=FALSE)
dist <- get_dist(x, domain)
# Upper bound for j * xij so that sum_j j * xij <= 1
quota <- 1 - (rowSums(t(t(x) * 1:p)) - t(t(x) * 1:p))
# Distance of xij to its boundary is min(xij - 0, quota_{i,j} / j - xij)
max(abs(dist$dx - pmin((t(t(quota) / 1:p) - x), x)))
h_hp <- get_h_hp("pow", 2) # For demonstration only
# Now confirm that h_of_dist indeed applies h and hp to dists
hd <- h_of_dist(h_hp, x, domain)
# hdx = dist ^ 2
print(max(abs(hd$hdx - dist$dx^2)))
# hpdx = 2 * dist' * dist
print(max(abs(hd$hpdx - 2*dist$dpx*dist$dx)))


# log_log_sum0 model on the simplex with K having row and column sums 0 (Aitchison model)
domain <- make_domain("simplex", p=p)
K <- -cov_cons("band", p=p, spars=3, eig=1)
diag(K) <- diag(K) - rowSums(K) # So that rowSums(K) == colSums(K) == 0
eigen(K)$val[(p-1):p] # Make sure K has one 0 and p-1 positive eigenvalues
x <- gen(n, setting="log_log_sum0", abs=FALSE, eta=eta, K=K, domain=domain,
       xinit=NULL, seed=2, burn_in=1000, thinning=100, verbose=FALSE)
# Note that dist$dx and dist$dpx only has p-1 columns -- excluding the last coordinate in x
dist <- get_dist(x, domain)
# Upper bound for x_{i,j} so that x_{i,1} + ... + x_{i,p-1} <= 1
quota <- 1 - (rowSums(x[,-p]) - x[,-p])
# Distance of x_{i,j} to its boundary is min(xij - 0, quota_{i,j} - xij)
max(abs(dist$dx - pmin(quota - x[,-p], x[,-p])))
h_hp <- get_h_hp("pow", 2) # For demonstration only
# Now confirm that h_of_dist indeed applies h and hp to dists
hd <- h_of_dist(h_hp, x, domain)
# hdx = dist ^ 2
print(max(abs(hd$hdx - dist$dx^2)))
# hpdx = 2 * dist' * dist
print(max(abs(hd$hpdx - 2*dist$dpx*dist$dx)))

genscore documentation built on May 31, 2023, 6:28 p.m.