Try something like this —
format compact
a = randn(2, 1)
x = linspace(-1, 1, 20) % equispaced x
f = a(1) + a(2)*x % function values
dfdx = ddx(f, x(2)-x(1)) % derivatives should be exact for linear
computeError = a(2) - dfdx % should be zeros to 1e-15
function dfdx = ddx(f,h)
dfdx(1) = (f(2) - f(1))/h;
dfdx(2:numel(f)) = (f(2:end) - f(1:end-1))/h;
end
EDIT — The gradient function already exists to do this, however I’m assuming here that you want to write your own function to do the numerical derivative.
.
