NEWS

pnd 0.1.13

Fix: relaxed tolerances in some tests
Fix: made a better default zero tolerance for generic derivative and accuracy orders

Feature: original kink-based algorithm for step size selection step.K()
Feature: added safety shrinking if FUN(x) is finite but FUN(x+h) is not in all SSS routines
Feature: added S3 printing methods for derivatives and step sizes
Feature: removed the diagnostics and report arguments; the iteration information is always saved, but not printed
Feature: added support for max.rel.error for all step-selection methods
Feature: all step-search methods now return both the truncation and the rounding-error estimate
Fix: corrected the wrong formula for the plug-in step size
Fix: now 1x1 Hessians can be computed (why, though, if second derivatives exist?)
Fix: added the v argument for numDeriv compatibility

Fix: GitHub issue #2 -- checkDimensions could not handle character h passed for auto-selection
Fix: GitHub issue #1 -- function arguments in ... did non propagate properly to step... functions

Fix: fixed a regression with the default step size
Fix: parallelised Hessians in the same manner as gradients
Feature: compatibility of Hessian() with the arguments for methods "Richardson" and "simple" from numDeriv

Fix: Derivatives of vectorised functions are working. Example: Grad(sin, 1:4)
Feature: Auto-detecting the number of cores available on multi-core machines to speed up computations
Feature: Added plug-in step size selection with an estimated f''' with a rule of thumb
Feature: Auto-detection of parallel type
Feature: Added zero tolerance to the default step for a fixed step

Feature: Extended the step-selection routines to gradients (vector input x)
Feature: Parallelisation of step selection in all algorithms
Feature: Mathur's AutoDX algorithm for step size selection step.M()
Feature: Added Hessian() that supports central differences (for the moment) and arbitrary accuracy
Feature: Separate Grad() and Jacobian() that call the workhorse, GenD(), for compatibility with numDeriv

Feature: Stepleman--Winarsky algorithm for step size selection step.SW()
Feature: Automated wrapper for step size selection gradstep()
Improvement: Safe handling of function errors and non-finite returns in step-size selection procedures
Improvement: Finite-difference coefficients gained attributes: Taylor expansion, coefficient on the largest truncated term, and effective accuracy (useful for custom stencils)
Improvement: added unit tests for core features

Feature: solveVandermonde() to solve ill-conditioned problems that arise in weight calculation
Feature: Dumontet--Vignes algorithm for step size selection step.DV()
Feature: Curtis--Reid algorithm for step size selection step.CR() and its modification
Feature: Different step sizes for the gradient
Fix: If the user supplies a short custom stencil and requests a high accuracy order, it will provide the best available order and produce a warning
Fix: The output of Grad() preserves the names of x and FUN(x), which prevents errors in cases where names are required