Noisefun {intrinsicDimension} | R Documentation |
Transition functions f(s|r)
describing the shift in lengths of vectors
when Gaussian noise is added. Given a length r
, f(s|r)
is the
probability density for the length after noise is added to one endpoint.
dnoiseNcChi(r, s, sigma, k)
dnoiseGaussH(r, s, sigma, k)
r |
length or vector of lengths of original vector. |
s |
length or vector of lengths of perturbed vector. |
sigma |
noise standard deviation. |
k |
dimension of noise. |
dnoiseNcChi
is the true transition function density when the noise
is Gaussian, the other transition functions are approximations of this.
dnoiseGaussH
is the Gaussian approximation used in Haro et al.
If Gaussian noise is added to both endpoints of the vector, sigma
should be replaced by
sqrt(2)*sigma
.
Vector of probability densities.
Only r
or s
can be a vector.
Kerstin Johnsson, Lund University
Haro, G., Randall, G. and Sapiro, G. (2008) Translated Poisson Mixture Model for Stratification Learning. Int. J. Comput. Vis., 80, 358-374.
maxLikPointwiseDimEst
, maxLikGlobalDimEst
, maxLikLocalDimEst
# High SNR, high-dim noise
sigma <- 0.05
x <- seq(0, 1.5, length.out = 200)
y <- dnoiseNcChi(x, s = .5, sigma, k = 20)
plot(x, y, type = 'l', main = 'Noise dim = 20')
y2 <- dnoiseGaussH(x, s = .5, sigma, k = 20)
lines(x, y2, lty = 2)
# Low SNR
par(mfrow = c(2, 3))
sigma <- 0.2
x <- seq(0, 1.5, length.out = 200)
y <- dnoiseNcChi(x, s = .5, sigma, k = 4)
plot(x, y, type = 'l', main = 'Noise approximations')
y2 <- dnoiseGaussH(x, s = .5, sigma, k)
lines(x, y2, lty = 2)
# High SNR, low-dim noise
sigma <- 0.05
x <- seq(0, 1.5, length.out = 200)
y <- dnoiseNcChi(x, s = .5, sigma, k = 4)
plot(x, y, type = 'l', main = 'Noise dim = 4')
y2 <- dnoiseGaussH(x, s = .5, sigma, k)
lines(x, y2, lty = 2)