Univariate optimization code for Julia
Author matthewclegg
2 Stars
Updated Last
8 Years Ago
Started In
September 2012


Univariate optimization and root-finding code for Julia


min(F::Function, x0::Float64, x1::Float64)
max(F::Function, x0::Float64, x1::Float64)
argmin(F::Function, x0::Float64, x1::Float64)
argmax(F::Function, x0::Float64, x1::Float64)
zero(F::Function, x0::Float64, x1::Float64)
inv(F::Function, x0::Float64, x1::Float64)


The function min(F, x0, x1) (respectively max(F, x0, x1)) searches for a local minimum (maximum) of F in the closed interval [x0,x1]. It returns the value F(x) that is found to minimize (maximize) F.

The algorithm first searches for a bracket that is known to contain a local minimum (maximum). From there, a modified version of Brent's method is used to precisely locate the minimum (maximum). Convergence is guaranteed to be linear and is often much more rapid. For further details, see section 10.3 of "Numerical Recipes".

The function argmin(F, x0, x1) is identical to min(F, x0, x1) except that the value x that minimizes F is returned, rather than returning F(x). The function argmax(F, x0, x1) is analogous.

The function zero(F, x0, x1) searches for a value x within the closed interval [x0,x1] such that F(x) = 0. Unlike the R function uniroot(), it is not a requirement that F(x0) and F(x1) have opposite signs. Rather, the algorithm initially searches for an interval [a,b] such F(a) and F(b) have opposite signs. Once this is found, a modified version of Ridder's method with bisection steps is used to precisely locate the zero. Convergence is guaranteed to be at least linear, and for well-behaved functions, super-linear convergence is obtained. For further details, see section 9.2 of "Numerical Recipes".

The function inv(F, x0, x1) constructs the inverse of F in the interval [x0,x1]. In other words, a function G:Float64 -> Float64 is returned such that G(F(x)) = x for all x in [x0,x1]. If F is strictly monotone and continuous, then G also will be strictly monotone and continuous. If these conditions are not satisfied, then all bets are off.

All of the preceding functions accept an optional fourth parameter, tolerance::Float64, which specifies the tolerance to be used in assessing the objective function.

The function polynomial_roots(a::Vector{Float64}) accepts as input a vector of length n representing the coefficients of the polynomial

a[1] + a[2] x + a[3] x^2 + ... + a[n] x^(n-1)

As output, it produces a pair (roots::Vector{Float64}, mult::Vector{Float64}) representing the real roots and their multiplicities, e.g., roots[i] is a real root of multiplicity multi[i]. This routine was provided as an illustration of the zero-finding procedure, and no claims are made about its optimality. When performance is an issue or when complex roots are needed, the user might wish to consult the literature, starting with section 9.5 of "Numerical Recipes".

A few simple performance tests were conducted to assess the performance of this code vis a vis the corresponding R routines. In the following, v is a vector of length 3,001, and w is a vector of length 30,001. These tests were conducted on a Mac Pro with 2.66 GHz Xeon processors.

                                                              Julia        R
 10000 reps of A:  min(x->x^2-1, -5., 0.)                     0.479 sec    0.681 sec
 10000 reps of B:  min(x->x^2-1, 1., 10.)                     0.477 sec    0.722 sec
 10000 reps of C:  min(x->^2-1, 1., 10.)                      0.159 sec    0.277 sec
 10000 reps of D:  zero(x->x^2-1, -5., 2.)                    0.107 sec    0.464 sec
  1000 reps of E:  min(x->sum((v - x).^2), -10., 20.)         0.391 sec    0.513 sec
   100 reps of F:  zero(x->sum(x - w.^2), -100., 200          0.581 sec    0.528 sec

See the benchmark() routine for further details.


julia> min(x->(x-2)*(x+3), -10., 10.)

julia> argmin(x->(x-2)*(x+3), -10., 10.)

julia> zero(x->(x-2)*(x+3), -10., 10.)

julia> max(x->sin(x), 0., pi)

julia> argmax(x->sin(x), 0., pi)

julia> zero(x->cos(x)-x, 0., 2*pi)

julia> atan2 = inv(tan, -1.57, 1.57)

julia> 4.0*atan2(1.)

julia> atan2(1.)-atan(1.)  # Compare our arctan with the supplied version

julia> polynomial_roots([-6., 1., 1.])  # x^2+x-6 = (x-2)*(x+3)
([-3.0, 2.0],[1, 1])

julia> H10=[-30240., 0., 302400., 0., -403200., 0., 161280., 0., -23040., 0., 1024.]
11-element Float64 Array:

julia> @time polynomial_roots(H10)
elapsed time: 0.055845022201538086 seconds
([-3.43616, -2.53273, -1.75668, -1.03661, -0.342901, 0.342901, 1.03661, 1.75668, 2.53273, 3.43616],[1, 1, 1, 1, 1, 1, 1, 1, 1, 1])

See also

optim.jl by John Myles White. This is a package for multivariate optimization. In the tests that I have conducted, I have found this implementation of the Nelder-Mead method to be quite competitive with the R implementation. https://github.com/johnmyleswhite/optim.jl

glm.jl by Douglas Bates. This is a package for fitting generalized linear models. https://github.com/dmbates/glm.jl


William H. Press, Saul A. Teukolsky, William T. Vetterling and Brian P. Flannery. 2007. "Numerical Recipes: The Art of Scientific Computing, 3rd Edition" Cambridge: Cambridge University Press.


Matthew Clegg

Comments and suggestions are of course welcome.