Cvxpy Python

CVXPY Python

Translate an optimization problem into CVXPY without changing its mathematics. Make dimensions, domains, curvature, solver capability, status handling, and numerical validation explicit.

Core objects

Model construction does not solve. solve() returns an objective value and updates problem.status, problem.value, variable values, constraint duals, and solver statistics when available. Problems are immutable; use Parameters for data changes or construct a new Problem for mathematical changes. Read the symbolic and curvature model.

Ordered workflow

1. Write the mathematical sets, indices, units, decision variables, objective, constraints, and allowed approximation before code.
2. Determine shapes and domains; distinguish scalar (), vector (n,), and column (n, 1) deliberately.
3. Build symbolic expressions using CVXPY atoms and @ for matrix products.
4. Check problem.is_dcp() or the intended ruleset before solver selection. Use is_dcp(dpp=True) for a repeated DCP parameterization.
5. Select a solver from the problem class and locally installed capabilities; pass only supported options.
6. Solve, branch on status, and reject missing/nonfinite values before use.
7. Independently validate primal feasibility, integrality, objective, and—when needed—duality/KKT residuals at tolerances justified by scale.

Decision rules

• Use a Parameter when coefficients or bounds change but the optimization structure remains. Require DPP for claimed recompilation speedups.
• Use variable attributes (nonneg, boolean, integer, PSD, sparsity) when analyzer information or reduced representation matters. Use explicit constraints when dual variables are required; attributes do not record their own duals.
• Use cp.sum, cp.sum_squares, cp.norm, cp.maximum, and other atoms on expressions. Python sum or NumPy functions can be slow, object-typed, or semantically wrong.
• Use separate constraints x >= 0 and x <= 1; chained comparisons and strict inequalities do not define valid CVXPY constraints.
• If a convex expression is marked unknown, rewrite it using a recognized atom or equivalent formulation; never suppress DCP errors without proving a different ruleset applies.
• Choose mixed-integer, conic, quadratic, or nonlinear solvers from installed solver support and licensing. Solver name is deployment configuration, not a universally portable constant.

Read modeling, solving, and validation for DCP/DPP, status gates, tolerances, solver choice, and repeated solves.

Canonical anchor

import cvxpy as cp
import numpy as np


def bounded_least_squares(design: np.ndarray, target: np.ndarray) -> tuple[np.ndarray, float]:
    coefficients = cp.Variable(design.shape[1])
    problem = cp.Problem(
        cp.Minimize(cp.sum_squares(design @ coefficients - target)),
        [coefficients >= 0, coefficients <= 1],
    )
    if not problem.is_dcp():
        raise ValueError("model must satisfy DCP")
    value = problem.solve()
    if problem.status not in {cp.OPTIMAL, cp.OPTIMAL_INACCURATE}:
        raise RuntimeError(f"solve failed: {problem.status}")
    if value is None or not np.isfinite(value):
        raise RuntimeError("solver returned no finite objective")
    if coefficients.value is None or not np.isfinite(coefficients.value).all():
        raise RuntimeError("solver returned no finite primal solution")
    return coefficients.value.copy(), float(value)

Production code must decide whether OPTIMAL_INACCURATE is acceptable and validate residuals accordingly. Do not return .value before the status gate.

High-risk rules

• Do not multiply decision variables together in a DCP model. Inspect curvature and use a recognized atom/formulation if the mathematics is convex.
• /, *, and @ have different scalar/elementwise/matrix meanings. Assert every expression shape; do not rely on accidental broadcasting.
• Do not read variable.value or constraint.dual_value until a successful solve and appropriate status check.
• Infeasible, unbounded, and infeasible_or_unbounded are distinct outcomes. Handle them separately; an infinite objective is not a usable solution.
• OPTIMAL_INACCURATE is evidence of a solver tolerance/status, not equivalent to exact optimality. Validate domain residuals.
• Warm start and Parameters do not guarantee DPP or speed. Check the compiled model and measure repeated solves.
• Scaling affects numerical reliability. Normalize units or use solver options based on diagnostics, not by accepting a loose tolerance until tests pass.
• Solver-specific options and availability drift. Query cp.installed_solvers() and inspect the environment.

Version grounding and completion

The official current tutorial documents DCP, DPP, solver status, variable attributes, and performance guidance; CVXPY is not installed locally at this authoring stage. Inspect version, installed solvers, atom/signature, and solver options before execution. Read verification.

Completion requires the code to match the written mathematics; ruleset checks to pass; solver capability to be explicit; all statuses to be handled; primal values and residuals to be validated; repeated solve behavior to be tested when claimed; and deterministic fixtures for feasible, infeasible, and boundary instances.

References

• Symbolic and curvature model
• Modeling, solving, and validation
• Verification and grounding