Error using complex variables

[pizqleh]

Hello,

I have been working with CVXPY over a convex problem with complex variables and no constraints of the form

min C_a(x),

where x = cp.Variable(n, complex=True). It worked perfectly, even using DCCP.

Now, it turns out that this is an approximation of a convex-concave problem with complex variables and no constraints of the form

min C_{a_1}(x) - C_{a_2}(x),

where x = cp.Variable(n, complex=True).

Then, I tried to solve it with DCCP using the standard reformulation

min C_{a_1}(x) - t
s.t. C_{a_2}(x) = t,

where x = cp.Variable(n, complex=True), t = cp.Variable(1). But I couldn't, the solver threw this error:

File "/Users/P/Desktop/thesis/wave-field-reconstruction/sinusoidal.py", line 139, in minimice_dccp
optimum = p_problem.solve(method='dccp')
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/problems/problem.py", line 396, in solve
return solve_func(self, *args, **kwargs)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/dccp/problem.py", line 54, in dccp
result_temp = iter_dccp(
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/dccp/problem.py", line 241, in iter_dccp
temp = convexify_constr(arg)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/dccp/constraint.py", line 76, in convexify_constr
right = linearize(constr.args[1])
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/dccp/linearize.py", line 58, in linearize
grad_map = expr.grad
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/atoms/atom.py", line 399, in grad
grad_arg = arg.grad
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/atoms/atom.py", line 399, in grad
grad_arg = arg.grad
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/atoms/atom.py", line 399, in grad
grad_arg = arg.grad
[Previous line repeated 2 more times]
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/atoms/atom.py", line 393, in grad
grad_self = self._grad(arg_values)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/atoms/affine/affine_atom.py", line 131, in _grad
canon_mat = canonInterface.get_problem_matrix(
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/cvxcore/python/canonInterface.py", line 327, in get_problem_matrix
build_lin_op_tree(lin, linPy_to_linC)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/cvxcore/python/canonInterface.py", line 518, in build_lin_op_tree
make_linC_from_linPy(linPy, linPy_to_linC)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/cvxcore/python/canonInterface.py", line 492, in make_linC_from_linPy
set_linC_data(linC, linPy)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/cvxcore/python/canonInterface.py", line 466, in set_linC_data
set_matrix_data(linC, linPy)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/cvxcore/python/canonInterface.py", line 438, in set_matrix_data
linC.set_dense_data(format_matrix(linPy.data, shape=linPy.shape))
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/cvxcore/python/cvxcore.py", line 200, in set_dense_data
return _cvxcore.LinOp_set_dense_data(self, matrix)
TypeError: Cannot cast array data from dtype('complex128') to dtype('float64') according to the rule 'safe'

Maybe the error is related to the utilization of complex and real variables. It's wierd, beacuse if I put t = cp.Variable() instead of t = cp.Variable(1) I get a different error:

File "/Users/P/Desktop/thesis/wave-field-reconstruction/sinusoidal.py", line 139, in minimizar_dccp
optimum = p_problem.solve(method='dccp')
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/problems/problem.py", line 396, in solve
return solve_func(self, *args, **kwargs)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/dccp/problem.py", line 50, in dccp
dccp_ini(
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/dccp/problem.py", line 142, in dccp_ini
value_para[count_para].value = np.random.randn(var.size) * 10
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/expressions/constants/parameter.py", line 82, in value
self._value = self._validate_value(val)
File "/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/cvxpy/expressions/leaf.py", line 379, in _validate_value
raise ValueError(
ValueError: Invalid dimensions (1,) for Parameter value.

Thanks in advance,
P.

[SteveDiamond]

Can you post a code snippet that reproduces the bug?

[algebravic]

I made up a simple dccp problem, but when I tried to solve it I got the error
"ValueError: Invalid dimensons (1,) for Parameter value."

The version of dccp is 1.0.3 and cvxpy is 1.1.15, and python is 3.8.5

Here's the program
x,y,z = [cp.Variable(name=_, nonneg=True) for _ in ['x', 'y', 'z']]
cons = [cp.square(x + y) == cp.square(x) + cp.square(y) - 2 * z, x + y == 1]
prob = cp.Problem(cp.Maximize(z), cons)

---

when I try

prob.solve(method='dccp')

I get the above error in line 382 of of leaf.py called fro line 142 in dccp_ini

[xinyueshen]

Hi @algebravic, I do not have this issue using CVXPY 1.1.15 and the most updated DCCP. Did you install DCCP from this repository? If not, you could download the master branch and install by running

python setup.py install

under the dccp directory.

To get a result for this particular example, the parameter tau needs to be set to 1 by

prob.solve(method="dccp", tau=1)

[algebravic]

It appears that the problem is that x = cvxpy.Variable(name='x', nonneg=True) gives x.shape = (). If I set it to (1,) instead it works. However, I also was unable to do something like

prob.solve(method='dccp', times=10)

because the times parameter is then passed to ecos which then complains.

[xinyueshen]

To run the algorithm many times with different random initial points, the parameter that needs to be set is ccp_times, not times.