uncontrolled parameter domain

[AxelBreuer]

Hi,

When specifying a CVXPY problem with parameters, we can define a domain of validity for these parameters (e.g., requiring positivity). It seems to me that CVXPY layers can override this domain silently (e.g., by setting a negative value to a parameter that is supposed to be positive).

For example, if we modify the toy example on the main webpage of cvxpylayers such that

import cvxpy as cp
import torch
from cvxpylayers.torch import CvxpyLayer

n, m = 2, 3
x = cp.Variable(n)
A = cp.Parameter((m, n))
# b = cp.Parameter(m)
b = cp.Parameter(m, pos=True)  # b is now a vector of positive numbers
constraints = [x >= 0]
objective = cp.Minimize(0.5 * cp.pnorm(A @ x - b, p=1))
problem = cp.Problem(objective, constraints)
assert problem.is_dpp()

cvxpylayer = CvxpyLayer(problem, parameters=[A, b], variables=[x])
A_tch = torch.randn(m, n, requires_grad=True)
# b_th = torch.randn(m, requires_grad=True)
b_tch = -abs(torch.randn(m, requires_grad=True))  # b_th is now a tensor of negative numbers

solution, = cvxpylayer(A_tch, b_tch)

solution.sum().backward()

The above code will work, but shouldn't it trigger an error saying that b_tch and/or b should be positive ?