calculus (version 1.0.1)

hessian: Numerical and Symbolic Hessian

Description

Computes the numerical Hessian of functions or the symbolic Hessian of characters.

Usage

hessian(f, var, params = list(), accuracy = 4, stepsize = NULL, drop = TRUE)
f %hessian% var

Value

Hessian matrix for scalar-valued functions when drop=TRUE, array otherwise.

Arguments

f

array of characters or a function returning a numeric array.

var

vector giving the variable names with respect to which the derivatives are to be computed and/or the point where the derivatives are to be evaluated. See derivative.

params

list of additional parameters passed to f.

accuracy

degree of accuracy for numerical derivatives.

stepsize

finite differences stepsize for numerical derivatives. It is based on the precision of the machine by default.

drop

if TRUE, return the Hessian as a matrix and not as an array for scalar-valued functions.

Functions

• f %hessian% var: binary operator with default parameters.

Details

In Cartesian coordinates, the Hessian of a scalar-valued function \(F\) is the square matrix of second-order partial derivatives:

$$(H(F))_{ij} = \partial_{ij}F$$

When the function \(F\) is a tensor-valued function \(F_{d_1,\dots,d_n}\), the hessian is computed for each scalar component.

$$(H(F))_{d_1\dots d_n,ij} = \partial_{ij}F_{d_1\dots d_n}$$

It might be tempting to apply the definition of the Hessian as the Jacobian of the gradient to write it in arbitrary orthogonal coordinate systems. However, this results in a Hessian matrix that is not symmetric and ignores the distinction between vector and covectors in tensor analysis. The generalization to arbitrary coordinate system is not currently supported.

References

Guidotti E (2022). "calculus: High-Dimensional Numerical and Symbolic Calculus in R." Journal of Statistical Software, 104(5), 1-37. tools:::Rd_expr_doi("10.18637/jss.v104.i05")

See Also

Other differential operators: curl(), derivative(), divergence(), gradient(), jacobian(), laplacian()

Examples

### symbolic Hessian 
hessian("x*y*z", var = c("x", "y", "z"))

### numerical Hessian in (x=1, y=2, z=3)
f <- function(x, y, z) x*y*z
hessian(f = f, var = c(x=1, y=2, z=3))

### vectorized interface
f <- function(x) x[1]*x[2]*x[3]
hessian(f = f, var = c(1, 2, 3))

### symbolic vector-valued functions
f <- c("y*sin(x)", "x*cos(y)")
hessian(f = f, var = c("x","y"))

### numerical vector-valued functions
f <- function(x) c(sum(x), prod(x))
hessian(f = f, var = c(0,0,0))

### binary operator
"x*y^2" %hessian% c(x=1, y=3)