使用Julia Convex包时diagind函数报错的正定矩阵优化问题问询

Let's break down why you're hitting that diagind error and fix your code step by step.

The Root Cause #

Convex.jl's variables (like Semidefinite) aren't regular Julia arrays—they're symbolic representations that track constraints and expressions under the hood. Directly modifying them with diagind and .= doesn't work because Convex can't track these in-place, low-level changes to its variable structures.

Corrected Code #

Here's the fixed version of your code, with explanations for each key change:

using Convex
using SCS  # Required: Convex needs a numerical solver; SCS works well for semidefinite problems

m = 5;
Σ = randn(m, m);  # Renamed to match your problem's known matrix Σ

# Define the symmetric positive definite variable (Semidefinite enforces symmetry + positive semidefiniteness)
x = Semidefinite(m)

# Construct XX: X with diagonal elements set to 0 (using Convex-compatible operations)
xx = x - Diagonal(x)

# Define the objective function as per your problem statement
obj = 0.5 * norm(Σ - x, :fro) + 0.01 * sum(abs(xx))

# Set up the minimization problem
pro = minimize(obj)

# Add the diagonal constraint: all diagonal elements of x must equal 1
pro.constraints += [diag(x) == ones(m)]

# Solve the problem with the SCS solver
solve!(pro, SCS.Optimizer)

# Inspect results
println("Optimal objective value: ", pro.optval)
println("Optimal X matrix:\n", evaluate(x))

Key Fixes Explained #

• Building xx correctly: Instead of trying to zero out diagonal elements with diagind, we use x - Diagonal(x)—this creates a valid Convex expression that represents x with its diagonal removed, which the library can properly track and use in the objective.
• Enforcing diagonal elements = 1: Instead of x[diagind(x)].=1, we add an equality constraint diag(x) == ones(m). This is the correct way to tell Convex to enforce that all diagonal entries of x equal 1, as it fits into the library's constraint-tracking system.
• Adding a solver: Convex.jl doesn't include a solver by default—we added SCS, a popular open-source solver for semidefinite and convex optimization problems, and specified it in the solve! call.
• Clarifying the objective: We used norm(Σ - x, :fro) for explicit Frobenius norm notation (it's equivalent to your original vecnorm call) and aligned variable names with your problem statement for readability.

Why This Works #

All operations now use Convex.jl's symbolic expression framework, so the library can correctly model your optimization problem, track all constraints, and pass a valid problem formulation to the solver without errors.

内容的提问来源于stack exchange，提问作者Xia.Song