Inconsistent results when defining the hessianstructure() method in the Problem interface

[lfaucheux]

Here and there, it reads that

there seems to be options to supply either the upper or lower triangular part (maybe dependent on which IPOPT interface is used).

Are you sure that the ipopt_interface that cyipopt mobilizes does not require the supplied hessian to be upper triangular ? Because that's what cyipopt does in the self.__hessianstructure is None case. And it turns out to work better that way...

[lfaucheux]

Why prevent us from transposing the indices of the lower triangle (and get the upper triangle's) ? You do this in the case where self.__hessianstructure is None :

[...]
        row, col = np.nonzero(np.tril(np.ones((self.__n, self.__n))))
        np_iRow = np.array(col, dtype=DTYPEi)   # <- `np_iRow` defined from `col` 🕊️🕊️🕊️
        np_jCol = np.array(row, dtype=DTYPEi)   # <- and vice versa               🕊️🕊️🕊️

... but you forbid us from doing so when self.__hessianstructure is not None.

[...]
        if not(np_iRow >= np_jCol).all():
            msg = b"Indices are not lower triangular in hessianstructure"
            log(msg, logging.ERROR)  # 💥💥💥
            return False             # 💥💥💥

[lfaucheux]

ok. I can reproduce the not self.__hessianstructure's results by 1) commenting out the if not(np_iRow >= np_jCol).all(): part and 2) defining the following problem's methods

def hessianstructure(...):
    [...]
    return np.nonzero(np.tril(np.ones((n, n))))[::-1]

def hessian(...):
    [...]
    return H[self.hessianstructure[::-1]]