Is get_delay correct? about pypte

[patrk]

Hillebrand 2016 in fact states:

prediction delay δ was set as (Ns x Nch)/N±, with Ns and Nch the number of samples and channels (ROIs), respectively, and N± the number of times the phase changes sign across time and channels.

It seems to be a classical case of a magic number, which should have been documented. I will update the docs.

So, I shifted the time series by two, and dropped the first and last data point for the multiplication. For that reason, the number of samples is also reduced by 2. At this moment, it might be questionable what happens if a zero crossing appears right at the boundaries.

[patrk]

Perhaps it is advisable to look into other ways to compute the delay.

A convenience class will be provided by #3

[jshanna100]

But why shift it by 2 instead of 1?

[patrk]

You are right. The choice of the shift interval seems to be random. Bolzano's theorem says there is at least one zero-crossing if the signs of the interval boundaries differs.

The code was inspired by Hillebrand's Matlab implementation :

%% Compute delay (if required)  
if isempty(delay)  
% delay is based on the number of times the phase flips across time and channels, as in Brainwave (C.J. Stam)  
    counter1 = 0; counter2 = 0;  
    for j=1:N  
        for i=2:L-1  
            counter1 = counter1 + 1;  
            if (phase_data(i-1,j)-pi)*(phase_data(i+1,j)-pi)<0, % make sure phase is in range [-pi pi]  
                counter2 = counter2 + 1;  
            end; %if  
        end; %for  
    end; %for  
    delay = round(counter1/counter2);  
end; %if;  

I will leave this implementation, which conforms to Hillebrand and implement another zero-crossing method, additionally.

Thank you for pointing this out.

[patrk]

This will be picked up and unit-tests should be provided.

[patrk]

Shifting by 2 doesn't actually lead to double-counting zero-crossings, but I get why it's a bit off. I’m planning to switch it up to shifting by 1 to stick closer to what's generally expected for zero-crossing detection. Also, I'll stop dropping the first and last data points. It looks like the original method by Hillebrand dropped those points for simplicity's sake