Develop Reference

How to find period of signal (autocorrelation vs fast fourier transform vs power spectral density)?

Suppose one wanted to find the period of a given sinusoidal wave signal. From what I have read online, it appears that the two main approaches employ either fourier analysis or autocorrelation. I am trying to automate the process using python and my usage case is to apply this concept to similar signals that come from the time-series of positions (or speeds or accelerations) of simulated bodies orbiting a star.
For simple-examples-sake, consider x = sin(t) for 0 ≤ t ≤ 10 pi.
import numpy as np
from scipy import signal
import matplotlib.pyplot as plt
## sample data
t = np.linspace(0, 10 * np.pi, 100)
x = np.sin(t)
fig, ax = plt.subplots()
ax.plot(t, x, color='b', marker='o')
ax.grid(color='k', alpha=0.3, linestyle=':')
plt.show()
plt.close(fig)
Given a sine-wave of the form x = a sin(b(t+c)) + d, the period of the sine-wave is obtained as 2 * pi / b. Since b=1 (or by visual inspection), the period of our sine wave is 2 * pi. I can check the results obtained from other methods against this baseline.
Attempt 1: Autocorrelation
As I understand it (please correct me if I'm wrong), correlation can be used to see if one signal is a time-lagged copy of another signal (similar to how cosine and sine differ by a phase difference). So autocorrelation is testing a signal against itself to measure the times at which the time-lag repeats said signal. Using the example posted here:
result = np.correlate(x, x, mode='full')
Since x and t each consist of 100 elements and result consists of 199 elements, I am not sure why I should arbitrarily select the last 100 elements.
print("\n autocorrelation (shape={}):\n{}\n".format(result.shape, result))
autocorrelation (shape=(199,)):
[ 0.00000000e+00 -3.82130761e-16 -9.73648712e-02 -3.70014208e-01
-8.59889695e-01 -1.56185995e+00 -2.41986054e+00 -3.33109112e+00
-4.15799070e+00 -4.74662427e+00 -4.94918053e+00 -4.64762251e+00
-3.77524157e+00 -2.33298717e+00 -3.97976240e-01 1.87752669e+00
4.27722402e+00 6.54129270e+00 8.39434617e+00 9.57785701e+00
9.88331103e+00 9.18204933e+00 7.44791758e+00 4.76948221e+00
1.34963425e+00 -2.50822289e+00 -6.42666652e+00 -9.99116299e+00
-1.27937834e+01 -1.44791297e+01 -1.47873668e+01 -1.35893098e+01
-1.09091510e+01 -6.93157447e+00 -1.99159756e+00 3.45267493e+00
8.86228186e+00 1.36707567e+01 1.73433176e+01 1.94357232e+01
1.96463736e+01 1.78556800e+01 1.41478477e+01 8.81191526e+00
2.32100171e+00 -4.70897483e+00 -1.15775811e+01 -1.75696560e+01
-2.20296487e+01 -2.44327920e+01 -2.44454330e+01 -2.19677060e+01
-1.71533510e+01 -1.04037163e+01 -2.33560966e+00 6.27458308e+00
1.45655029e+01 2.16769872e+01 2.68391837e+01 2.94553896e+01
2.91697473e+01 2.59122266e+01 1.99154591e+01 1.17007613e+01
2.03381596e+00 -8.14633251e+00 -1.78184255e+01 -2.59814393e+01
-3.17580589e+01 -3.44884934e+01 -3.38046447e+01 -2.96763956e+01
-2.24244433e+01 -1.26974172e+01 -1.41464998e+00 1.03204331e+01
2.13281784e+01 3.04712823e+01 3.67721634e+01 3.95170295e+01
3.83356037e+01 3.32477037e+01 2.46710643e+01 1.33886439e+01
4.77778141e-01 -1.27924775e+01 -2.50860560e+01 -3.51343866e+01
-4.18671622e+01 -4.45258983e+01 -4.27482779e+01 -3.66140001e+01
-2.66465884e+01 -1.37700036e+01 7.76494745e-01 1.55574483e+01
2.90828312e+01 3.99582426e+01 4.70285203e+01 4.95000000e+01
4.70285203e+01 3.99582426e+01 2.90828312e+01 1.55574483e+01
7.76494745e-01 -1.37700036e+01 -2.66465884e+01 -3.66140001e+01
-4.27482779e+01 -4.45258983e+01 -4.18671622e+01 -3.51343866e+01
-2.50860560e+01 -1.27924775e+01 4.77778141e-01 1.33886439e+01
2.46710643e+01 3.32477037e+01 3.83356037e+01 3.95170295e+01
3.67721634e+01 3.04712823e+01 2.13281784e+01 1.03204331e+01
-1.41464998e+00 -1.26974172e+01 -2.24244433e+01 -2.96763956e+01
-3.38046447e+01 -3.44884934e+01 -3.17580589e+01 -2.59814393e+01
-1.78184255e+01 -8.14633251e+00 2.03381596e+00 1.17007613e+01
1.99154591e+01 2.59122266e+01 2.91697473e+01 2.94553896e+01
2.68391837e+01 2.16769872e+01 1.45655029e+01 6.27458308e+00
-2.33560966e+00 -1.04037163e+01 -1.71533510e+01 -2.19677060e+01
-2.44454330e+01 -2.44327920e+01 -2.20296487e+01 -1.75696560e+01
-1.15775811e+01 -4.70897483e+00 2.32100171e+00 8.81191526e+00
1.41478477e+01 1.78556800e+01 1.96463736e+01 1.94357232e+01
1.73433176e+01 1.36707567e+01 8.86228186e+00 3.45267493e+00
-1.99159756e+00 -6.93157447e+00 -1.09091510e+01 -1.35893098e+01
-1.47873668e+01 -1.44791297e+01 -1.27937834e+01 -9.99116299e+00
-6.42666652e+00 -2.50822289e+00 1.34963425e+00 4.76948221e+00
7.44791758e+00 9.18204933e+00 9.88331103e+00 9.57785701e+00
8.39434617e+00 6.54129270e+00 4.27722402e+00 1.87752669e+00
-3.97976240e-01 -2.33298717e+00 -3.77524157e+00 -4.64762251e+00
-4.94918053e+00 -4.74662427e+00 -4.15799070e+00 -3.33109112e+00
-2.41986054e+00 -1.56185995e+00 -8.59889695e-01 -3.70014208e-01
-9.73648712e-02 -3.82130761e-16 0.00000000e+00]
Attempt 2: Fourier
Since I am not sure where to go from the last attempt, I sought a new attempt. To my understanding, Fourier analysis basically shifts a signal from/to the time-domain (x(t) vs t) to/from the frequency domain (x(t) vs f=1/t); the signal in frequency-space should appear as a sinusoidal wave that dampens over time. The period is obtained from the most observed frequency since this is the location of the peak of the distribution of frequencies.
Since my values are all real-valued, applying the Fourier transform should mean my output values are all complex-valued. I wouldn't think this is a problem, except for the fact that scipy has methods for real-values. I do not fully understand the differences between all of the different scipy methods. That makes following the algorithm proposed in this posted solution hard for me to follow (ie, how/why is the threshold value picked?).
omega = np.fft.fft(x)
freq = np.fft.fftfreq(x.size, 1)
threshold = 0
idx = np.where(abs(omega)>threshold)[0][-1]
max_f = abs(freq[idx])
print(max_f)
This outputs 0.01, meaning the period is 1/0.01 = 100. This doesn't make sense either.
Attempt 3: Power Spectral Density
According to the scipy docs, I should be able to estimate the power spectral density (psd) of the signal using a periodogram (which, according to wikipedia, is the fourier transform of the autocorrelation function). By selecting the dominant frequency fmax at which the signal peaks, the period of the signal can be obtained as 1 / fmax.
freq, pdensity = signal.periodogram(x)
fig, ax = plt.subplots()
ax.plot(freq, pdensity, color='r')
ax.grid(color='k', alpha=0.3, linestyle=':')
plt.show()
plt.close(fig)
The periodogram shown below peaks at 49.076... at a frequency of fmax = 0.05. So, period = 1/fmax = 20. This doesn't make sense to me. I have a feeling it has something to do with the sampling rate, but don't know enough to confirm or progress further.
I realize I am missing some fundamental gaps in understanding how these things work. There are a lot of resources online, but it's hard to find this needle in the haystack. Can someone help me learn more about this?

Let's first look at your signal (I've added endpoint=False to make the division even):
t = np.linspace(0, 10*np.pi, 100, endpoint=False)
x = np.sin(t)
Let's divide out the radians (essentially by taking t /= 2*np.pi) and create the same signal by relating to frequencies:
fs = 20 # Sampling rate of 100/5 = 20 (e.g. Hz)
f = 1 # Signal frequency of 1 (e.g. Hz)
t = np.linspace(0, 5, 5*fs, endpoint=False)
x = np.sin(2*np.pi*f*t)
This makes it more salient that f/fs == 1/20 == 0.05 (i.e. the periodicity of the signal is exactly 20 samples). Frequencies in a digital signal always relate to its sampling rate, as you have already guessed. Note that the actual signal is exactly the same no matter what the values of f and fs are, as long as their ratio is the same:
fs = 1 # Natural units
f = 0.05
t = np.linspace(0, 100, 100*fs, endpoint=False)
x = np.sin(2*np.pi*f*t)
In the following I'll use these natural units (fs = 1). The only difference will be in t and hence the generated frequency axes.
Autocorrelation
Your understanding of what the autocorrelation function does is correct. It detects the correlation of a signal with a time-lagged version of itself. It does this by sliding the signal over itself as seen in the right column here (from Wikipedia):
Note that as both inputs to the correlation function are the same, the resulting signal is necessarily symmetric. That is why the output of np.correlate is usually sliced from the middle:
acf = np.correlate(x, x, 'full')[-len(x):]
Now index 0 corresponds to 0 delay between the two copies of the signal.
Next you'll want to find the index or delay that presents the largest correlation. Due to the shrinking overlap this will by default also be index 0, so the following won't work:
acf.argmax() # Always returns 0
Instead I recommend to find the largest peak instead, where a peak is defined to be any index with a larger value than both its direct neighbours:
inflection = np.diff(np.sign(np.diff(acf))) # Find the second-order differences
peaks = (inflection < 0).nonzero()[0] + 1 # Find where they are negative
delay = peaks[acf[peaks].argmax()] # Of those, find the index with the maximum value
Now delay == 20, which tells you that the signal has a frequency of 1/20 of its sampling rate:
signal_freq = fs/delay # Gives 0.05
Fourier transform
You used the following to calculate the FFT:
omega = np.fft.fft(x)
freq = np.fft.fftfreq(x.size, 1)
Thhese functions re designed for complex-valued signals. They will work for real-valued signals, but you'll get a symmetric output as the negative frequency components will be identical to the positive frequency components. NumPy provides separate functions for real-valued signals:
ft = np.fft.rfft(x)
freqs = np.fft.rfftfreq(len(x), t[1]-t[0]) # Get frequency axis from the time axis
mags = abs(ft) # We don't care about the phase information here
Let's have a look:
plt.plot(freqs, mags)
plt.show()
Note two things: the peak is at frequency 0.05, and the maximum frequency on the axis is 0.5 (the Nyquist frequency, which is exactly half the sampling rate). If we had picked fs = 20, this would be 10.
Now let's find the maximum. The thresholding method you have tried can work, but the target frequency bin is selected blindly and so this method would suffer in the presence of other signals. We could just select the maximum value:
signal_freq = freqs[mags.argmax()] # Gives 0.05
However, this would fail if, e.g., we have a large DC offset (and hence a large component in index 0). In that case we could just select the highest peak again, to make it more robust:
inflection = np.diff(np.sign(np.diff(mags)))
peaks = (inflection < 0).nonzero()[0] + 1
peak = peaks[mags[peaks].argmax()]
signal_freq = freqs[peak] # Gives 0.05
If we had picked fs = 20, this would have given signal_freq == 1.0 due to the different time axis from which the frequency axis was generated.
Periodogram
The method here is essentially the same. The autocorrelation function of x has the same time axis and period as x, so we can use the FFT as above to find the signal frequency:
pdg = np.fft.rfft(acf)
freqs = np.fft.rfftfreq(len(x), t[1]-t[0])
plt.plot(freqs, abs(pdg))
plt.show()
This curve obviously has slightly different characteristics from the direct FFT on x, but the main takeaways are the same: the frequency axis ranges from 0 to 0.5*fs, and we find a peak at the same signal frequency as before: freqs[abs(pdg).argmax()] == 0.05.
Edit:
To measure the actual periodicity of np.sin, we can just use the "angle axis" that we passed to np.sin instead of the time axis when generating the frequency axis:
freqs = np.fft.rfftfreq(len(x), 2*np.pi*f*(t[1]-t[0]))
rad_period = 1/freqs[mags.argmax()] # 6.283185307179586
Though that seems pointless, right? We pass in 2*np.pi and we get 2*np.pi. However, we can do the same with any regular time axis, without presupposing pi at any point:
fs = 10
t = np.arange(1000)/fs
x = np.sin(t)
rad_period = 1/np.fft.rfftfreq(len(x), 1/fs)[abs(np.fft.rfft(x)).argmax()] # 6.25
Naturally, the true value now lies in between two bins. That's where interpolation comes in and the associated need to choose a suitable window function.

Inverse FFT returns negative values when it should not

I have several points (x,y,z coordinates) in a 3D box with associated masses. I want to draw an histogram of the mass-density that is found in spheres of a given radius R.
I have written a code that, providing I did not make any errors which I think I may have, works in the following way:
My "real" data is something huge thus I wrote a little code to generate non overlapping points randomly with arbitrary mass in a box.
I compute a 3D histogram (weighted by mass) with a binning about 10 times smaller than the radius of my spheres.
I take the FFT of my histogram, compute the wave-modes (kx, ky and kz) and use them to multiply my histogram in Fourier space by the analytic expression of the 3D top-hat window (sphere filtering) function in Fourier space.
I inverse FFT my newly computed grid.
Thus drawing a 1D-histogram of the values on each bin would give me what I want.
My issue is the following: given what I do there should not be any negative values in my inverted FFT grid (step 4), but I get some, and with values much higher that the numerical error.
If I run my code on a small box (300x300x300 cm3 and the points of separated by at least 1 cm) I do not get the issue. I do get it for 600x600x600 cm3 though.
If I set all the masses to 0, thus working on an empty grid, I do get back my 0 without any noted issues.
I here give my code in a full block so that it is easily copied.
import numpy as np
import matplotlib.pyplot as plt
import random
from numba import njit
# 1. Generate a bunch of points with masses from 1 to 3 separated by a radius of 1 cm
radius = 1
rangeX = (0, 100)
rangeY = (0, 100)
rangeZ = (0, 100)
rangem = (1,3)
qty = 20000 # or however many points you want
# Generate a set of all points within 1 of the origin, to be used as offsets later
deltas = set()
for x in range(-radius, radius+1):
for y in range(-radius, radius+1):
for z in range(-radius, radius+1):
if x*x + y*y + z*z<= radius*radius:
deltas.add((x,y,z))
X = []
Y = []
Z = []
M = []
excluded = set()
for i in range(qty):
x = random.randrange(*rangeX)
y = random.randrange(*rangeY)
z = random.randrange(*rangeZ)
m = random.uniform(*rangem)
if (x,y,z) in excluded: continue
X.append(x)
Y.append(y)
Z.append(z)
M.append(m)
excluded.update((x+dx, y+dy, z+dz) for (dx,dy,dz) in deltas)
print("There is ",len(X)," points in the box")
# Compute the 3D histogram
a = np.vstack((X, Y, Z)).T
b = 200
H, edges = np.histogramdd(a, weights=M, bins = b)
# Compute the FFT of the grid
Fh = np.fft.fftn(H, axes=(-3,-2, -1))
# Compute the different wave-modes
kx = 2*np.pi*np.fft.fftfreq(len(edges[0][:-1]))*len(edges[0][:-1])/(np.amax(X)-np.amin(X))
ky = 2*np.pi*np.fft.fftfreq(len(edges[1][:-1]))*len(edges[1][:-1])/(np.amax(Y)-np.amin(Y))
kz = 2*np.pi*np.fft.fftfreq(len(edges[2][:-1]))*len(edges[2][:-1])/(np.amax(Z)-np.amin(Z))
# I create a matrix containing the values of the filter in each point of the grid in Fourier space
R = 5
Kh = np.empty((len(kx),len(ky),len(kz)))
#njit(parallel=True)
def func_njit(kx, ky, kz, Kh):
for i in range(len(kx)):
for j in range(len(ky)):
for k in range(len(kz)):
if np.sqrt(kx[i]**2+ky[j]**2+kz[k]**2) != 0:
Kh[i][j][k] = (np.sin((np.sqrt(kx[i]**2+ky[j]**2+kz[k]**2))*R)-(np.sqrt(kx[i]**2+ky[j]**2+kz[k]**2))*R*np.cos((np.sqrt(kx[i]**2+ky[j]**2+kz[k]**2))*R))*3/((np.sqrt(kx[i]**2+ky[j]**2+kz[k]**2))*R)**3
else:
Kh[i][j][k] = 1
return Kh
Kh = func_njit(kx, ky, kz, Kh)
# I multiply each point of my grid by the associated value of the filter (multiplication in Fourier space = convolution in real space)
Gh = np.multiply(Fh, Kh)
# I take the inverse FFT of my filtered grid. I take the real part to get back floats but there should only be zeros for the imaginary part.
Density = np.real(np.fft.ifftn(Gh,axes=(-3,-2, -1)))
# Here it shows if there are negative values the magnitude of the error
print(np.min(Density))
D = Density.flatten()
N = np.mean(D)
# I then compute the histogram I want
hist, bins = np.histogram(D/N, bins='auto', density=True)
bin_centers = (bins[1:]+bins[:-1])*0.5
plt.plot(bin_centers, hist)
plt.xlabel('rho/rhom')
plt.ylabel('P(rho)')
plt.show()
Do you know why I'm getting these negative values? Do you think there is a simpler way to proceed?
Sorry if this is a very long post, I tried to make it very clear and will edit it with your comments, thanks a lot!
-EDIT-
A follow-up question on the issue can be found [here].1

The filter you create in the frequency domain is only an approximation to the filter you want to create. The problem is that we are dealing with the DFT here, not the continuous-domain FT (with its infinite frequencies). The Fourier transform of a ball is indeed the function you describe, however this function is infinitely large -- it is not band-limited!
By sampling this function only within a window, you are effectively multiplying it with an ideal low-pass filter (the rectangle of the domain). This low-pass filter, in the spatial domain, has negative values. Therefore, the filter you create also has negative values in the spatial domain.
This is a slice through the origin of the inverse transform of Kh (after I applied fftshift to move the origin to the middle of the image, for better display):
As you can tell here, there is some ringing that leads to negative values.
One way to overcome this ringing is to apply a windowing function in the frequency domain. Another option is to generate a ball in the spatial domain, and compute its Fourier transform. This second option would be the simplest to achieve. Do remember that the kernel in the spatial domain must also have the origin at the top-left pixel to obtain a correct FFT.
A windowing function is typically applied in the spatial domain to avoid issues with the image border when computing the FFT. Here, I propose to apply such a window in the frequency domain to avoid similar issues when computing the IFFT. Note, however, that this will always further reduce the bandwidth of the kernel (the windowing function would work as a low-pass filter after all), and therefore yield a smoother transition of foreground to background in the spatial domain (i.e. the spatial domain kernel will not have as sharp a transition as you might like). The best known windowing functions are Hamming and Hann windows, but there are many others worth trying out.
Unsolicited advice:
I simplified your code to compute Kh to the following:
kr = np.sqrt(kx[:,None,None]**2 + ky[None,:,None]**2 + kz[None,None,:]**2)
kr *= R
Kh = (np.sin(kr)-kr*np.cos(kr))*3/(kr)**3
Kh[0,0,0] = 1
I find this easier to read than the nested loops. It should also be significantly faster, and avoid the need for njit. Note that you were computing the same distance (what I call kr here) 5 times. Factoring out such computation is not only faster, but yields more readable code.

Just a guess:
Where do you get the idea that the imaginary part MUST be zero? Have you ever tried to take the absolute values (sqrt(re^2 + im^2)) and forget about the phase instead of just taking the real part? Just something that came to my mind.

Moving average produces array of different length?

This question has a lot of useful answers on how to get a moving average.
I have tried the two methods of numpy convolution and numpy cumsum and both worked fine on an example dataset, but produced a shorter array on my real data.
The data are spaced by 0.01. The example dataset has a length of 50, the real data tens of thousands. So it must be something about the window size that is causing the problem and I don't quite understand what is going on in the functions.
This is how I define the functions:
def smoothMAcum(depth,temp, scale): # Moving average by cumsum, scale = window size in m
dz = np.diff(depth)
N = int(scale/dz[0])
cumsum = np.cumsum(np.insert(temp, 0, 0))
smoothed=(cumsum[N:] - cumsum[:-N]) / N
return smoothed
def smoothMAconv(depth,temp, scale): # Moving average by numpy convolution
dz = np.diff(depth)
N = int(scale/dz[0])
smoothed=np.convolve(temp, np.ones((N,))/N, mode='valid')
return smoothed
Then I implement it:
scale = 5.
smooth = smoothMAconv(dep,data, scale)
but print len(dep), len(smooth)
returns 81071 80572
and the same happens if I use the other function.
How can I get the smooth array of the same length as the data?
And why did it work on the small dataset? Even if I try different scales (and use the same for the example and for the data), the result in the example has the same length as the original data, but not in the real application.
I considered an effect of nan values, but if I have a nan in the example, it doesn't make a difference.
So where is the problem, if possible to tell without the full dataset?

The second of your approaches is easy to modify to preserve the length, because numpy.convolve supports the parameter mode='same'.
np.convolve(temp, np.ones((N,))/N, mode='same')
This is made possible by zero-padding the data set temp on both sides, -
which will inevitably have some effect at the boundaries unless your data happens to be 0 near the boundaries. Example:
N = 10
x = np.linspace(0, 2, 100)
y = x**2 + np.random.uniform(size=x.shape)
y_smooth = np.convolve(y, np.ones((N,))/N, mode='same')
plt.plot(x, y, 'r.')
plt.plot(x, y_smooth)
plt.show()
The boundary effect of zero-padding is very visible at the right end, where the data points are about 4-5 but are padded by 0.
To reduce this undesired effect, use numpy.pad for more intelligent padding; reverting to mode='valid' for convolution. The pad width must be such that in total N-1 elements are added, where N is the size of moving window.
y_padded = np.pad(y, (N//2, N-1-N//2), mode='edge')
y_smooth = np.convolve(y_padded, np.ones((N,))/N, mode='valid')
Padding by edge values of an array looks much better.

Construct an array spacing proportional to a function or other array

I have a function (f : black line) which varies sharply in a specific, small region (derivative f' : blue line, and second derivative f'' : red line). I would like to integrate this function numerically, and if I distribution points evenly (in log-space) I end up with fairly large errors in the sharply varying region (near 2E15 in the plot).
How can I construct an array spacing such that it is very well sampled in the area where the second derivative is large (i.e. a sampling frequency proportional to the second derivative)?
I happen to be using python, but I'm interested in a general algorithm.
Edit:
1) It would be nice to be able to still control the number of sampling points (at least roughly).
2) I've considered constructing a probability distribution function shaped like the second derivative and drawing randomly from that --- but I think this will offer poor convergence, and in general, it seems like a more deterministic approach should be feasible.

Assuming f'' is a NumPy array, you could do the following
# Scale these deltas as you see fit
deltas = 1/f''
domain = deltas.cumsum()
To account only for order of magnitude swings, this could be adjusted as follows...
deltas = 1/(-np.log10(1/f''))

I'm just spitballing here ... (as I don't have time to try this out for real)...
Your data looks (roughly) linear on a log-log plot (at least, each segment seems to be... So, I might consider doing a sort-of integration in log-space.
log_x = log(x)
log_y = log(y)
Now, for each of your points, you can get the slope (and intercept) in log-log space:
rise = np.diff(log_y)
run = np.diff(log_x)
slopes = rise / run
And, similarly, the the intercept can be calculated:
# y = mx + b
# :. b = y - mx
intercepts = y_log[:-1] - slopes * x_log[:-1]
Alright, now we have a bunch of (straight) lines in log-log space. But, a straight line in log-log space, corresponds to y = log(intercept)*x^slope in real space. We can integrate that easily enough: y = a/(k+1) x ^ (k+1), so...
def _eval_log_log_integrate(a, k, x):
return np.log(a)/(k+1) * x ** (k+1)
def log_log_integrate(a, k, x1, x2):
return _eval_log_log_integrate(a, k, x2) - _eval_log_log_integrate(a, k, x1)
partial_integrals = []
for a, k, x_lower, x_upper in zip(intercepts, slopes, x[:-1], x[1:]):
partial_integrals.append(log_log_integrate(a, k, x_lower, x_upper))
total_integral = sum(partial_integrals)
You'll want to check my math -- It's been a while since I've done this sort of thing :-)

1) The Cool Approach
At the moment I implemented an 'adaptive refinement' approach inspired by hydrodynamics techniques. I have a function which I want to sample, f, and I choose some initial array of sample points x_i. I construct a "sampling" function g, which determines where to insert new sample points.
In this case I chose g as the slope of log(f) --- since I want to resolve rapid changes in log space. I then divide the span of g into L=3 refinement levels. If g(x_i) exceeds a refinement level, that span is subdivided into N=2 pieces, those subdivisions are added into the samples and are checked against the next level. This yields something like this:
The solid grey line is the function I want to sample, and the black crosses are my initial sampling points.
The dashed grey line is the derivative of the log of my function.
The colored dashed lines are my 'refinement levels'
The colored crosses are my refined sampling points.
This is all shown in log-space.
2) The Simple Approach
After I finished (1), I realized that I probably could have just chosen a maximum spacing in in y, and choose x-spacings to achieve that. Similarly, just divide the function evenly in y, and find the corresponding x points.... The results of this are shown below:

A simple approach would be to split the x-axis-array into three parts and use different spacing for each of them. It would allow you to maintain the total number of points and also the required spacing in different regions of the plot. For example:
x = np.linspace(10**13, 10**15, 100)
x = np.append(x, np.linspace(10**15, 10**16, 100))
x = np.append(x, np.linspace(10**16, 10**18, 100))
You may want to choose a better spacing based on your data, but you get the idea.

Fast, elegant way to calculate empirical/sample covariogram

Does anyone know a good method to calculate the empirical/sample covariogram, if possible in Python?
This is a screenshot of a book which contains a good definition of covariagram:
If I understood it correctly, for a given lag/width h, I'm supposed to get all the pair of points that are separated by h (or less than h), multiply its values and for each of these points, calculate its mean, which in this case, are defined as m(x_i). However, according to the definition of m(x_{i}), if I want to compute m(x1), I need to obtain the average of the values located within distance h from x1. This looks like a very intensive computation.
First of all, am I understanding this correctly? If so, what is a good way to compute this assuming a two dimensional space? I tried to code this in Python (using numpy and pandas), but it takes a couple of seconds and I'm not even sure it is correct, that is why I will refrain from posting the code here. Here is another attempt of a very naive implementation:
from scipy.spatial.distance import pdist, squareform
distances = squareform(pdist(np.array(coordinates))) # coordinates is a nx2 array
z = np.array(z) # z are the values
cutoff = np.max(distances)/3.0 # somewhat arbitrary cutoff
width = cutoff/15.0
widths = np.arange(0, cutoff + width, width)
Z = []
Cov = []
for w in np.arange(len(widths)-1): # for each width
# for each pairwise distance
for i in np.arange(distances.shape[0]):
for j in np.arange(distances.shape[1]):
if distances[i, j] <= widths[w+1] and distances[i, j] > widths[w]:
m1 = []
m2 = []
# when a distance is within a given width, calculate the means of
# the points involved
for x in np.arange(distances.shape[1]):
if distances[i,x] <= widths[w+1] and distances[i, x] > widths[w]:
m1.append(z[x])
for y in np.arange(distances.shape[1]):
if distances[j,y] <= widths[w+1] and distances[j, y] > widths[w]:
m2.append(z[y])
mean_m1 = np.array(m1).mean()
mean_m2 = np.array(m2).mean()
Z.append(z[i]*z[j] - mean_m1*mean_m2)
Z_mean = np.array(Z).mean() # calculate covariogram for width w
Cov.append(Z_mean) # collect covariances for all widths
However, now I have confirmed that there is an error in my code. I know that because I used the variogram to calculate the covariogram (covariogram(h) = covariogram(0) - variogram(h)) and I get a different plot:
And it is supposed to look like this:
Finally, if you know a Python/R/MATLAB library to calculate empirical covariograms, let me know. At least, that way I can verify what I did.

One could use scipy.cov, but if one does the calculation directly (which is very easy), there are more ways to speed this up.
First, make some fake data that has some spacial correlations. I'll do this by first making the spatial correlations, and then using random data points that are generated using this, where the data is positioned according to the underlying map, and also takes on the values of the underlying map.
Edit 1:
I changed the data point generator so positions are purely random, but z-values are proportional to the spatial map. And, I changed the map so that left and right side were shifted relative to eachother to create negative correlation at large h.
from numpy import *
import random
import matplotlib.pyplot as plt
S = 1000
N = 900
# first, make some fake data, with correlations on two spatial scales
# density map
x = linspace(0, 2*pi, S)
sx = sin(3*x)*sin(10*x)
density = .8* abs(outer(sx, sx))
density[:,:S//2] += .2
# make a point cloud motivated by this density
random.seed(10) # so this can be repeated
points = []
while len(points)<N:
v, ix, iy = random.random(), random.randint(0,S-1), random.randint(0,S-1)
if True: #v<density[ix,iy]:
points.append([ix, iy, density[ix,iy]])
locations = array(points).transpose()
print locations.shape
plt.imshow(density, alpha=.3, origin='lower')
plt.plot(locations[1,:], locations[0,:], '.k')
plt.xlim((0,S))
plt.ylim((0,S))
plt.show()
# build these into the main data: all pairs into distances and z0 z1 values
L = locations
m = array([[math.sqrt((L[0,i]-L[0,j])**2+(L[1,i]-L[1,j])**2), L[2,i], L[2,j]]
for i in range(N) for j in range(N) if i>j])
Which gives:
The above is just the simulated data, and I made no attempt to optimize it's production, etc. I assume this is where the OP starts, with the task below, since the data already exists in a real situation.
Now calculate the "covariogram" (which is much easier than generating the fake data, btw). The idea here is to sort all the pairs and associated values by h, and then index into these using ihvals. That is, summing up to index ihval is the sum over N(h) in the equation, since this includes all pairs with hs below the desired values.
Edit 2:
As suggested in the comments below, N(h) is now only the pairs that are between h-dh and h, rather than all pairs between 0 and h (where dh is the spacing of h-values in ihvals -- ie, S/1000 was used below).
# now do the real calculations for the covariogram
# sort by h and give clear names
i = argsort(m[:,0]) # h sorting
h = m[i,0]
zh = m[i,1]
zsh = m[i,2]
zz = zh*zsh
hvals = linspace(0,S,1000) # the values of h to use (S should be in the units of distance, here I just used ints)
ihvals = searchsorted(h, hvals)
result = []
for i, ihval in enumerate(ihvals[1:]):
start, stop = ihvals[i-1], ihval
N = stop-start
if N>0:
mnh = sum(zh[start:stop])/N
mph = sum(zsh[start:stop])/N
szz = sum(zz[start:stop])/N
C = szz-mnh*mph
result.append([h[ihval], C])
result = array(result)
plt.plot(result[:,0], result[:,1])
plt.grid()
plt.show()
which looks reasonable to me as one can see bumps or troughs at the expected for the h values, but I haven't done a careful check.
The main speedup here over scipy.cov, is that one can precalculate all of the products, zz. Otherwise, one would feed zh and zsh into cov for every new h, and all the products would be recalculated. This calculate could be sped up even more by doing partial sums, ie, from ihvals[n-1] to ihvals[n] at each timestep n, but I doubt that will be necessary.