CODEWITHSUNDEEP
pythonpandasnumpylinear-algebraeigenvalue
najeem
najeem

Reputation: 1921

Eigenvalues from pandas rows

I have a pandas DataFrame with rows representing a symmetric matrix component. 

                            sxx         syy             szz         sxy         syz         sxz                                                                                           
    NodeID      time
    1500000     20921.0     2504729.0   -16524560.0     -3966213.0  5058878.0   8026349.0   390275.7
                20923.0     2541577.0   -16459500.0     -3930280.0  5047995.0   8019404.0   393201.3
                20925.0     2582004.0   -16384690.0     -3891037.0  5035703.0   8011226.0   396850.2
                20927.0     2618859.0   -16313310.0     -3855520.0  5024095.0   8003384.0   400578.7
                20933.0     2703961.0   -16133460.0     -3773937.0  4995101.0   7985394.0   411183.2

The matrix will look like the following.

[[sxx, sxy, sxz],
 [sxy, syy, syz],
 [sxz, syz, szz]]

What is the fastest way to calculate the eigenvalue from each row?

I tried 'applying' np.linalg.eigvalsh on every row. However, it takes quite long when I have close to a million lines.

Edit

To give the complete context, I should also mention that this DataFrame is part of an object definition. object.df is the DataFrame. Below is the related code.

    def s1(self):
        """Returns the first principal stress for every node every timepoint"""
        return self.df.apply(principal, axis=1, label="s1")

def principal(s, label):
    principals = np.linalg.eigvalsh(
        np.array(
            [s.sxx, s.sxy, s.sxz, s.sxy, s.syy, s.syz, s.sxz, s.syz, s.szz]
        ).reshape(3, 3)
    )
    if label.lower() == "s3":
        return principals[0]
    elif label.lower() == "s2":
        return principals[1]
    elif label.lower() == "s1":
        return principals[2]
    else:
        raise ValueError("Invalid Input, choose from s1, s2, or s3.")

Upvotes: 2

Views: 1681

Answers (1)

Tarifazo
Tarifazo

Reputation: 4343

You can set the order of the columns to generate a view and then pass it to an array using .values (faster than np.array(..)), then apply eigvalsh to an (n, 3, 3) array:

values = df[['sxx', 'sxy', 'sxz', 'sxy', 'syy', 'syz', 'sxz', 'syz', 'szz']].values.reshape(-1,3))
eigh = eigvalsh(values.reshape((-1, 3, 3)))
eigh

>>array([[-21253030.07083309,  -1397298.11167328,   4664284.18250638],
       [-21184732.23304478,  -1361435.36228467,   4697964.59532944],
       [-21106512.77176102,  -1322433.70013306,   4735223.47189408],
       [-21032246.72681734,  -1287171.41922922,   4769447.14604654],
       [-20847979.70886149,  -1205613.19093403,   4850156.89979552]])

Upvotes: 2

Related Questions