AlgorithmAnalysis.jl is a Julia package for the automated analysis of algorithms.

This package provides a generic way to analyze algorithms in a systematic manner in the Julia programming language. Algorithm analysis seeks to find a mathematically proven guarantee of an algorithm's performance over a class of problems. AlgorithmAnalysis.jl includes both the performance estimation (PEP) and control theoretic methodologies to analysis.

Installation

The package can be installed using the Julia package manager:

import Pkg; Pkg.add("AlgorithmAnalysis")

Example

This example code finds the worst-case convergence rate of the (squared) distance to optimality of the gradient descent algorithm at minimizing $L$-smooth and $\mu$-strongly convex functions.

using AlgorithmAnalysis

@alg begin

    # variables
    α, μ, L, ρ ∈ R
    x, xs ∈ Rⁿ
    f ∈ differentiable_functional(Rⁿ)

    # algorithm update
    x₊ = x - α * f'(x)

    # state transitions
    trans = (x → x₊) ∧ (xs → xs)

    # constraints
    c1 = smooth_strongly_convex(f, μ, L)
    c2 = f'(xs)^2 == zero(R)
    con = c1 ∧ c2

    # performance measure
    perf = (x - xs)^2
    perf₊ = (x₊ - xs)^2

    # performance estimation problem
    pep = maximize(perf₊, con ∧ (perf ≤ 1))

    # Lyapunov-based stability certification
    cert = certify(ρ, perf, con ∧ trans)

    # optimal rate that is certifiable (using bisection)
    ρopt = rate(perf, con ∧ trans)
end

with_numerics(parameters = Dict(α => 0.1, μ => 1.0, L => 10.0)) do

    isapprox(evaluate(simplify(pep)), 0.81, atol=1e-6)  # true
end

with_numerics(parameters = Dict(α => 0.1, μ => 1.0, L => 10.0, ρ => 0.81000001)) do

    evaluate(simplify(cert))  # true
end

with_parameters(Dict(α => 0.1, μ => 1.0, L => 10.0)) do
    
    ρopt_simplified = simplify(ρopt)

    evaluate(ρopt_simplified) ≈ 0.81  # true
end

Documentation structure

• Manual: describes the data structures used by AlgorithmAnalysis.jl

• API: a comprehensive list of all public objects exported by AlgorithmAnalysis.jl

• Developer Guide: helps get researchers started in how to contribute novel algorithms or analysis techniques

License

The software is licensed under the MIT License.

Acknowledgement

This material is based upon work supported by the National Science Foundation under Award No. 2347121. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.