How to add (source to target) causality between two 1D time series? (without using numpy tril)

[bartmch]

Is there a better (computationally cheaper) way to quantify the relationship between a source>target time series instead of setting the lower triangle matrix to inf after computing it?

# These 2 lines are not really important but adding for any recommendations
s1 = zscore(dummy['ets_flow']); s2 = zscore(dummy['return_water_temp'])
s1_diff = preprocessing.differencing(s1.values[None,:], smooth=0.15).squeeze()
s2_diff = preprocessing.differencing(s2.values[None,:], smooth=0.15).squeeze()
# Calculate full matrix
d, paths = dtw.warping_paths(s1_diff, s2_diff, window=30, use_pruning=True)
# Set lower triangle part of the matrix to inf so it will not be considered by best_path
lower = np.tril_indices(paths.shape[0], 0)
paths[lower] = paths[0,-1] #inf
best_path = dtw.best_path(paths)
# Distance is inf?
dtwvis.plot_warpingpaths(s1, s2, paths, best_path)
plt.plot([0, len(paths)], [0, len(paths)],color='white')

Besides the computational overhead, I cannot seem to compute the best path if values above the matrix diagonal are set inf:

[wannesm]

There is no argument that does just that. But this seems easy to achieve by using a window and pre-/post-pending values to make it a onesided window (and optionally psi-relaxation to avoid the extra penalty from pre-/post-pending values).

# s2 can only map to later timestamps in s1
onesidedwindow = 20
window=int(onesidedwindow/2)
s1b = np.concatenate((np.full((window,), s1[0]), s1))
s2b = np.concatenate((s2, np.full((window,), s2[0])))
d, paths = dtw.warping_paths(s1b, s2b, window=window,
                             psi=(window,0,0,window))
best_path = dtw.best_path(paths)
dtwvis.plot_warpingpaths(s1b, s2b, paths, best_path)