simple.nlsfit: NLS fit with without bootstrap
In hadron: Analysis Framework for Monte Carlo Simulation Data in Physics

simple.nlsfit

R Documentation

NLS fit with without bootstrap

Description

NLS fit with without bootstrap

Usage

simple.nlsfit(fn, par.guess, y, x, errormodel, priors = list(param = c(), p =
  c(), psamples = c()), ..., lower = rep(x = -Inf, times =
  length(par.guess)), upper = rep(x = +Inf, times = length(par.guess)), dy,
  dx, CovMatrix, boot.R = 0, gr, dfn, mask, use.minpack.lm = TRUE,
  error = sd, maxiter = 500, success.infos = 1:3,
  relative.weights = FALSE, na.rm = FALSE)

Arguments

`fn`	`fn(par, x, ...)`. The (non-linear) function to be fitted to the data. Its first argument must be the fit parameters named `par`. The second must be `x`, the explaining variable. Additional parameters might be passed to the function. Currently we pass `boot.r` which is `0` for the original data and the ID (1, ...) of the bootstrap sample otherwise. As more parameters might be added in the future it is recommended that the fit function accepts `...` as the last parameter to be forward compatible.
`par.guess`	initial guess values for the fit parameters.
`y`	the data as a one-dimensional numerical vector to be described by the fit function.
`x`	values of the explaining variable in form of a one-dimensional numerical vector.
`errormodel`	Either "yerrors" or "xyerrors", depending on the x-values having errors or not.
`priors`	List possessing the elements `param`, `p` and `psamples`. The vector `param` includes the indices of all fit parameters that are to be constrained and the vector `p` the corresponding paramater values (e.g. known from a previous fit). The list element `psamples` is a matrix of dimensions `(boot.R, length(param))` and contains the corresponding bootstrap samples. If this list is not specified priors are omitted within the fit.
`...`	Additional parameters passed to `fn`, `gr` and `dfn`.
`lower`	Numeric vector of length `length(par.guess)` of lower bounds on the fit parameters. If missing, `-Inf` will be set for all.
`upper`	Numeric vector of length `length(par.guess)` of upper bounds on the fit parameters. If missing, `+Inf` will be set for all.
`dy, dx`	Numeric vector. Errors of the dependent and independent variable, respectively. These do not need to be specified as they can be computed from the bootstrap samples. In the case of parametric bootstrap it might would lead to a loss of information if they were computed from the pseudo-bootstrap samples. They must not be specified if a covariance matrix is given.
`CovMatrix`	complete variance-covariance matrix of dimensions `c(length(y), length(y))` or `c(length(y)+length(x), length(y)+length(x))` depending on the errormodel. Pass `NULL` if the matrix has to be calculated from the `bsamples`. In that case, if the number of boostrap samples is small compared to the number of variables, singular value decomposition with small eigenvalue replacement will be used (see invertCovMatrix) to attempt a clean inversion. In case a variance-covariance matrix is passed, the inversion will simply be attempted using `solve` on the Cholesky decomposition. Finally, if `CovMatrix` is missing, an uncorrelated fit will be performed.
`boot.R`	If larger than 0, `boot.R` paramtetric bootstrap samples are generated on the fit results after fit and error calculation are finished. The original data is never boostraped in this function.
`gr`	`gr(par, x, ...)`. `gr=d(fn) / d(par)` is a function to return the gradient of `fn`. It must return an array with `length(x)` rows and `length(par)` columns.
`dfn`	`dfn(par, x, ...)`. `dfn=d(fn) / dx` is the canonical derivative of `fn` by `x` and only relevant if x-errors are provided.
`mask`	logical or integer index vector. The mask is applied to select the observations from the data that are to be used in the fit. It is applied to `x`, `y`, `dx`, `dy`, `bsamples` and `CovMatrix` as applicable.
`use.minpack.lm`	use the `minpack.lm` library if available. This is usually faster than the default `optim` but somtimes also less stable.
`error`	Function that takes a sample vector and returns the error estimate. This is a parameter in order to support different resampling methods like jackknife.
`maxiter`	integer. Maximum number of iterations that can be used in the optimization process.
`success.infos`	integer vector. When using `minpack.lm` there is the `info` in the return value. Values of 1, 2 or 3 are certain success. A value of 4 could either be a success or a saddle point. If you want to interpret this as a success as well just pass `1:4` instead of the default `1:3`.
`relative.weights`	are the errors on y (and x) to be interpreted as relative weights instead of absolute ones? If TRUE, the covariance martix of the fit parameter results is multiplied by chi^2/dof. This is the default in many fit programs, e.g. gnuplot.
`na.rm`	logical. If set to `true`, NAs in `y` and `dy` will be ignored. If x-errors are taken into account, NAs in `x` and `dx` will be ignored, too.

Value

Returns an object of class bootstrapfit, see bootstrap.nlsfit.

Examples

## Declare some data.
value <- c(0.1, 0.2, 0.3)
dvalue <- c(0.01, 0.01, 0.015)
x <- c(1, 2, 3)
dx <- c(0.1, 0.1, 0.1)

fn <- function (par, x, ...) par[1] + par[2] * x

fit.result <- simple.nlsfit(fn, c(1, 1), value, x, "xyerrors", dy=dvalue, dx=dx)
summary(fit.result)

hadron documentation built on Sept. 9, 2022, 5:06 p.m.