Generic function for the computation of asymmetric total variation distance of two distributions
Generic function for the computation of asymmetric total variation distance of two distributions
Generic function for the computation of asymmetric total variation distance dv(rho)
of two distributions P and Q where the distributions may be defined for an arbitrary sample space (Omega,A). For given ratio of inlier and outlier probability rho, this distance is defined as [REMOVE_ME]dv(ρ)(P,Q)=∫(dQ−cdP)+dv(rho)(P,Q)=∫max(dQ−cdP,0)[REMOVEME2]
for c defined by [REMOVE_ME]ρ∫(dQ−cdP)+=∫(dQ−cdP)−rho∫max(dQ−cdP,0)=∫max(cdP−dQ,0)[REMOVEME2]
It coincides with total variation distance for rho=1.
Description
Generic function for the computation of asymmetric total variation distance dv(rho)
of two distributions P and Q where the distributions may be defined for an arbitrary sample space (Omega,A). For given ratio of inlier and outlier probability rho, this distance is defined as
asis.smooth.discretize: possible methods are "asis", "smooth" and "discretize". Default is "discretize".
n.discr: if asis.smooth.discretize is equal to "discretize" one has to specify the number of lattice points used to discretize the abs. cont. distribution.
low.discr: if asis.smooth.discretize is equal to "discretize" one has to specify the lower end point of the lattice used to discretize the abs. cont. distribution.
up.discr: if asis.smooth.discretize is equal to "discretize" one has to specify the upper end point of the lattice used to discretize the abs. cont. distribution.
h.smooth: if asis.smooth.discretize is equal to "smooth" -- i.e., the empirical distribution of the provided data should be smoothed -- one has to specify this parameter.
rho: ratio of inlier/outlier radius
rel.tol: relative tolerance for distrExIntegrate and uniroot
maxiter: parameter for uniroot
Ngrid: How many grid points are to be evaluated to determine the range of the likelihood ratio?
,
TruncQuantile: Quantile the quantile based integration bounds (see details)
IQR.fac: Factor for the scale based integration bounds (see details)
...: further arguments to be used in particular methods -- (in package distrEx: just used for distributions with a.c. parts, where it is used to pass on arguments to distrExIntegrate).
diagnostic: logical; if TRUE, the return value obtains an attribute "diagnostic" with diagnostic information on the integration, i.e., a list with entries method ("integrate"
or "GLIntegrate"), call, result (the complete return value of the method), args (the args with which the method was called), and time (the time to compute the integral).
Details
For distances between absolutely continuous distributions, we use numerical integration; to determine sensible bounds we proceed as follows: by means of min(getLow(e1,eps=TruncQuantile),getLow(e2,eps=TruncQuantile)), max(getUp(e1,eps=TruncQuantile),getUp(e2,eps=TruncQuantile)) we determine quantile based bounds c(low.0,up.0), and by means of s1 <- max(IQR(e1),IQR(e2));m1<- median(e1);
m2 <- median(e2)
and low.1 <- min(m1,m2)-s1*IQR.fac, up.1 <- max(m1,m2)+s1*IQR.fac
we determine scale based bounds; these are combined by low <- max(low.0,low.1), up <- max(up.0,up1).
Again in the absolutely continuous case, to determine the range of the likelihood ratio, we evaluate this ratio on a grid constructed as follows: x.range <- c(seq(low, up, length=Ngrid/3),q.l(e1)(seq(0,1,length=Ngrid/3)*.999),q.l(e2)(seq(0,1,length=Ngrid/3)*.999))
Finally, for both discrete and absolutely continuous case, we clip this ratio downwards by 1e-10 and upwards by 1e10
In case we want to compute the total variation distance between (empirical) data and an abs. cont. distribution, we can specify the parameter asis.smooth.discretize
to avoid trivial distances (distance = 1).
Using asis.smooth.discretize = "discretize", which is the default, leads to a discretization of the provided abs. cont. distribution and the distance is computed between the provided data and the discretized distribution.
Using asis.smooth.discretize = "smooth" causes smoothing of the empirical distribution of the provided data. This is, the empirical data is convoluted with the normal distribution Norm(mean = 0, sd = h.smooth)
which leads to an abs. cont. distribution. Afterwards the distance between the smoothed empirical distribution and the provided abs. cont. distribution is computed.
Diagnostics on the involved integrations are available if argument diagnostic is TRUE. Then there is attribute diagnostic
attached to the return value, which may be inspected and accessed through showDiagnostic and getDiagnostic.
Returns
Asymmetric Total variation distance of e1 and e2
Methods
e1 = "AbscontDistribution", e2 = "AbscontDistribution":: total variation distance of two absolutely continuous univariate distributions which is computed using distrExIntegrate.
e1 = "AbscontDistribution", e2 = "DiscreteDistribution":: total variation distance of absolutely continuous and discrete univariate distributions (are mutually singular; i.e., have distance =1).
e1 = "DiscreteDistribution", e2 = "DiscreteDistribution":: total variation distance of two discrete univariate distributions which is computed using support and sum.
e1 = "DiscreteDistribution", e2 = "AbscontDistribution":: total variation distance of discrete and absolutely continuous univariate distributions (are mutually singular; i.e., have distance =1).
e1 = "numeric", e2 = "DiscreteDistribution":: Total variation distance between (empirical) data and a discrete distribution.
e1 = "DiscreteDistribution", e2 = "numeric":: Total variation distance between (empirical) data and a discrete distribution.
e1 = "numeric", e2 = "AbscontDistribution":: Total variation distance between (empirical) data and an abs. cont. distribution.
e1 = "AbscontDistribution", e1 = "numeric":: Total variation distance between (empirical) data and an abs. cont. distribution.
e1 = "AcDcLcDistribution", e2 = "AcDcLcDistribution":: Total variation distance of mixed discrete and absolutely continuous univariate distributions.
References
to be filled; Agostinelli, C and Ruckdeschel, P. (2009): A simultaneous inlier and outlier model by asymmetric total variation distance.