Scipy约束最小化未遵守等式约束问题咨询

Let's cut to the chase: the L-BFGS-B optimizer in SciPy ignores your equality constraint entirely. Here's why, and how to fix it:

The Root Cause #

L-BFGS-B is built exclusively for optimization problems with only bound constraints (like your 0 ≤ q_i ≤ 1 limits). It has no support for equality or inequality constraints—any constraints parameter you pass to it will be silently ignored. That’s exactly why your solution’s elements don’t sum to 1: the constraint was never applied in the first place.

Fix 1: Switch to an Optimizer That Supports Constraints #

The simplest, most reliable fix is to use SciPy's SLSQP method, which natively handles both equality and bound constraints. Here’s how to adjust your code:

import numpy as np
from scipy.optimize import minimize

def EV(q, P):
    return (-1)*np.sum(100 * q * (2*P - 1))

# Replace with your actual P and initial guess X0
P = np.random.rand(12)  # Example given 12D array
X0 = np.ones(12) / 12   # Initial uniform distribution (sum=1)

# Define constraints and bounds
cons = {'type': 'eq', 'fun': lambda q: np.sum(q) - 1}
bds = [(0, 1)] * 12

# Use SLSQP instead of L-BFGS-B
result = minimize(
    EV,
    X0,
    args=(P,),
    method='SLSQP',
    bounds=bds,
    constraints=cons,
    tol=1e-8  # Optional: tighten tolerance for stricter constraint satisfaction
)

# Verify the solution
print(f"Sum of optimized q: {np.sum(result.x):.8f}")
print(f"Optimization success: {result.success}")

SLSQP will enforce your sum-to-1 rule while respecting the 0-1 bounds on each element. For most use cases, this is the best approach.

Fix 2: Reparameterize Variables to Eliminate the Equality Constraint #

If you specifically need to use L-BFGS-B (e.g., for performance with large-scale problems), you can reparameterize your variables to embed the sum-to-1 constraint directly into the objective function. Since sum(q) = 1, you only need 11 independent variables—derive the 12th as q_12 = 1 - sum(q_1 ... q_11).

For a cleaner fit with L-BFGS-B, use a softmax transformation to automatically enforce the probability simplex (sum=1, all elements ≥0):

import numpy as np
from scipy.optimize import minimize

def EV_softmax(z, P):
    # Convert unconstrained z values to valid q via softmax
    q = np.exp(z) / np.sum(np.exp(z))
    return (-1)*np.sum(100 * q * (2*P - 1))

# Example inputs
P = np.random.rand(12)
X0_softmax = np.zeros(12)  # Simple initial guess

# Use L-BFGS-B with loose bounds for z (softmax handles the simplex)
result = minimize(
    EV_softmax,
    X0_softmax,
    args=(P,),
    method='L-BFGS-B',
    bounds=[(-10, 10)]*12
)

# Convert back to the original q variable
q_opt = np.exp(result.x) / np.sum(np.exp(result.x))
print(f"Sum of optimized q: {np.sum(q_opt):.8f}")

The softmax transformation ensures q automatically meets the sum-to-1 and non-negativity rules, so you don’t need any constraints at all—perfect for L-BFGS-B’s bound-only design.

Key Takeaways #

• L-BFGS-B does not support equality/inequality constraints—only use it for problems with variable bounds.
• For constrained problems, SLSQP is the standard go-to method in SciPy.
• If you must use L-BFGS-B, reparameterize variables to embed constraints into the objective function (like softmax for simplex rules).

内容的提问来源于stack exchange，提问作者Davide Floriello