I originally posted this on stackoverflow.com and then deleted it and moved it here

My question is similar to Similarity of two discrete fourier tranforms (specifically the selected answer). I've also gleaned some helpful information from this R help thread.

I'm stuck on the actual implementation of this process, however.

I have two Fourier series; my ultimate goal is to calculate a sum of weighted coherence. (In actuality I have millions of series corresponding to pixels, but, let's just call it two for now).

    TS1 <- c(5051.29, 5355.31, 5602.18, 5784.4, 5896.45, 5934.9, 5898.6, 
5788.64, 5608.37, 5363.27, 5060.78, 4710.09, 4321.86, 3907.89, 
3480.75, 3053.42, 2638.89, 2249.75, 1897.82, 1593.81, 1346.93, 
1164.71, 1052.67, 1014.21, 1050.52, 1160.47, 1340.74, 1585.84, 
1888.33, 2239.02, 2627.25, 3041.22, 3468.36, 3895.69, 4310.22, 
4699.36, 5051.29)
    TS2 <- c(4192.83, 4532.62, 4836.41, 5094.96, 5300.41, 5446.53, 5528.88, 
5544.95, 5494.25, 5378.33, 5200.71, 4966.78, 4683.65, 4359.93, 
4005.45, 3630.98, 3247.9, 2867.85, 2502.38, 2162.59, 1858.8, 
1600.25, 1394.8, 1248.67, 1166.33, 1150.26, 1200.95, 1316.87, 
1494.5, 1728.43, 2011.55, 2335.28, 2689.76, 3064.23, 3447.31, 
3827.36, 4192.83)

From the above links I see that I need to calculate coherence of the cross-spectrum at each frequency, and weight the frequencies that have higher power spectral density.

I know I can extract spectral density from

spectrum(data.frame(TS1, TS2))$spec

I know I can extract coherence using:

spectrum(data.frame(TS1, TS2))$coh

I can see that the value of coherence extracted is always equal to 1. It seems the "solution" to this is to lag the time series somehow.

I'm not sure how to 1) calculate optimal lag, especially because I'd like it to be comparable across all the coherence calculations, and 2) how to specify lag in the spectrum(). spec.pgram takes "spans", but those seem to have more to do with smoothing than offsetting.

I tried a maximum ccf function defined here, but in many cases (not in the example set I've provided, you'll have to trust me) the "optimal lag" returned is 0, which makes sense because the ccf is 1, but isn't actually what I'm looking for.

Should I be using acf() or ccf() to manually calculate coherence? What steps (R functions would be really helpful) do I need to perform to get from where I am to having a single weighted coherence value?

  • $\begingroup$ The number of points you have in the two time-series seems very small to practically perform fourier or cross-spectral analysis. $\endgroup$
    – hearse
    Apr 7, 2014 at 19:36
  • 2
    $\begingroup$ I shortened the data for the post- what I have are Fourier coefficients, for a fundamental frequency and two harmonics. What I've done is used those to calculate the frequency using radian values from 0 to 2*pi with steps of 0.0001 and I've been working with that data, so I actually have 36000 points for each time series. $\endgroup$
    – Nan
    Apr 7, 2014 at 20:14
  • $\begingroup$ Ok . that's positive. $\endgroup$
    – hearse
    Apr 7, 2014 at 20:29

3 Answers 3


The following are the issues that I summarize succinctly in sequence, as each of them are an interesting problem in itself and then follow it up with my solutions:

i) How to unlag two time-series so as to lose the least data due to the lagging, and at the same time to maximize the cross-correlation or spectral coherence.

ii) how would this be done when you are dealing with multivariate time-series. As it's greater than the two, its like solving a rubiks cube optimally while lagging various time-series in the dataset so as to preserve a summed up measure of (i).

iii) How to compute a weighted-coherence?

I will skip ii) and answer i) and iii) due to reasons of conflict of interest and research collaboration/confidential agreements on my work on ii). I understand that iii) happens to be the central question to start with followed by i).

My solution for iii)

Pre-Processing Steps: Say you had two time -series t1 and t2. Determine a CCF maximizing lag, say optLag and unlag t1 and t2 as unlaggedTS = lag(t2,optLag) and perform a time-series union of the dataset as unionTS = ts.union(t1,unlaggedTS). Following that, as the time-series unlagging would produce a few NA's as you lose entries by lagging, you need to set these values to zero in unionTS as required-meaning that nothing was observed at the NA points.

Step 2: Now, obtain an abinded time-series, with the span parameters as required.

bindTSPair= abind(((spectrum(unionTS,span=c(16,16),plot=F)$spec)[,1]),((spectrum(unionTS,span=c(16,16),plot=F)$spec)[,2])) 

Following it, just normalize the data in bindTSPair as say:


Then obtained the following two halves:



and then finally obtain 'a weighted' version of spectral coherence as follows:

 weightedCoherence=sum((10*(TSHalf1 + TSHalf2)) * (spectrum(unionTS,span=c(10,10),plot=F)$coh))

You may change the weighting as required or using 'prior' info.

My solution for i) to obtain a CCF maximizing lag:

 for(i in 2:ncol(myTS))
    for(j in 2:ncol(myTS))
    if(abs(cp) > abs(cn))
    		if(abs(cn) > abs(cp))
    if(abs(cn) > abs(cp))
if(length(tp) != 0)



  • $\begingroup$ I think that in the normalization TSHalf1 lines, object cbind1 is actually what you named bindTSpair? $\endgroup$
    – Nan
    Apr 7, 2014 at 21:24
  • $\begingroup$ You are welcome. That's right. $\endgroup$
    – hearse
    Apr 7, 2014 at 21:30
  • 1
    $\begingroup$ @PraneethVepakomma Thanks for a very useful answer. Just two clarifications. What is the normalization that is being used here in data.Normalization? Is this optimal lag also the estimate of "time delay"? Maybe we can use ts.intersect to bind the two series as that will save the step of putting NA to 0. $\endgroup$
    – Anusha
    Sep 17, 2014 at 2:40

unless your task is to actually implement coherence calculation, I wouldn't do it, because there's got to be a package for that. e.g. I see coh package, haven't used it myself. this function seems to do all that you need for a specific frequency f, so you can loop a call to it over the frequency range.

In matlab there's a function called mscohere, which produces an entire graph like this one, it's squared magnitude of coherence enter image description here

I apolologise for using Matlab example, but I didn't do any spectral analysis in R myself yet. The concepts are the same though.

  • $\begingroup$ @Aksakal- I agree that mscohere() in matlab does exactly this, but I don't have access to Matlab and was unable to find an R package that performed the same functionality as plainly. $\endgroup$
    – Nan
    Apr 7, 2014 at 21:25

Use seewave::coh() to extract the coherence and lag values for the maximum coherence point:

coherence <- coh(ts1, ts2, f=100, plot=TRUE)
lags      <- coherence[,1]
vals      <- coherence[,2]
val.max   <- vals[which.max(vals)]
lag.max   <- lags[which.max(vals)]

The coh() function is designed to work with ts objects. This is a simple way of doing the calculation Praneeth describes here.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.