I am using NumPy's linalg.eig on square matrices. My square matrices are a function of a 2D domain, and I am looking at its eigenvectors' complex angles along a parameterized circle on this domain. As long as the path I am considering is smooth, I expect the complex angles of each eigenvector's components to be smooth. However, for some cases, this is not the case with Python (although it is with other programming languages). For the parameter M=0 (some argument in my matrix that appears on its diagonal), I have components that look like:
when they should ideally look like (M=0.1):
What I have tried:
I verified that the matrices are Hermitian in both cases.
When I use linalg.eigh, M=0.1 becomes discontinuous while M=0 sometimes becomes continuous.
Using np.unwrap did nothing.
The difference between component phases (i.e. np.angle(v1-v2) for eigenvector v=[[v1],[v2]]) is smooth/continuous, but this is not what I want.
Fixing the NumPy seed before solving did nothing for different values of the seed. For example: np.random.seed(1).
What else can I do? I am trying to use Sympy's eigenvects just because I am running out of options, and I asked another question asking about another potential approach here: How do I force first component of NumPy eigenvectors to be real? . But, I do not know what else I can try.
Here is a minimal working example that works nicely in a Jupyter notebook:
import numpy as np
from numpy import linalg as LA
import matplotlib.pyplot as plt
M = 0.01; # nonzero M is okay
M = 0.0; # M=0 causes problems
def matrix_generator(kx,ky,M):
a = 2.46; t = 1; k = np.array((kx,ky));
d1 = (a/2)*np.array((1,np.sqrt(3)));d2 = (a/2)*np.array((1,-np.sqrt(3)));d3 = -a*np.array((1,0));
sx = np.matrix([[0,1],[1,0]]);sy = np.matrix([[0,-1j],[1j,0]]);sz = np.matrix([[1,0],[0,-1]]);
hx = np.cos(k#d1)+np.cos(k#d2)+np.cos(k#d3);hy = np.sin(k#d1)+np.sin(k#d2)+np.sin(k#d3);
return -t*(hx*sx - hy*sy + M*sz)
n_segs = 200; #number of segments in (kx,ky) loop
evecs_along_loop = np.zeros((n_segs,2,2),dtype=float)
# parameterize circular loop
kx0 = 0.5; ky0 = 1; r1=0.2; r2=0.2;
a = np.linspace(0.0, 2*np.pi, num=n_segs+2)
kloop=np.zeros((n_segs+2,2))
for i in range(n_segs+2):
kloop[i,:]=np.array([kx0 + r1*np.cos(a[i]), ky0 + r2*np.sin(a[i])])
# assign eigenvector complex angles
for j in np.arange(n_segs):
np.random.seed(2)
H = matrix_generator(kloop[j][0],kloop[j][1],M)
eval0, psi0 = LA.eig(H)
evecs_along_loop[j,:,:] = np.angle(psi0)
# plot eigenvector complex angles
for p in np.arange(2):
for q in np.arange(2):
print(f"Phase for eigenvector element {p},{q}:")
fig = plt.figure()
ax = plt.axes()
ax.plot((evecs_along_loop[:,p,q]))
plt.show()
Clarification for anon01's comment:
For M=0, a sample matrix at some value of (kx,ky) would look like:
a = np.matrix([[0.+0.j, 0.99286437+1.03026667j],
[0.99286437-1.03026667j, 0.+0.j]])
For M =/= 0, the diagonal will be non-zero (but real).
I think that in general this is a tough problem. The fundamental issue is that eigenvectors (unlike eigenvalues) are not unambiguously defined. An eigenvector v of M with eigenvalue c is any non-zero vector for which
M*v = c*v
In particular for any non zero scalar s, multiplying an eigenvector by s yields an eigenvector, and even if you demand (as usual) that eigenvectors have length 1, we are still free to multiply by any scalar of absolute value 1. Even worse, if v1,..vd are orthogonal eigenvectors for c, then any non-zero linear combination of the v's is also an eigenvector for c.
Different eigendecomposition routines might well, therefore, come up with very different eigenvectors and still be doing their job. Moreover some routines might produce eigenvectors that are far apart for matrices that are close together.
A simple tractable case is where you know that all your eigenvalues are non-degenerate (i.e. each eigenspace is of dimension 1) and you happen to know that for a particular i, the i'th component of each eigenvector will be non zero. Then you could multiply the eigenvector v by a scalar, of absolute value 1, chosen so that after the multiplication v[i] is a positive real number. In C
s = conj(v[i])/cabs(v[i])
where
conj(z) is the complex conjugate of the complex number z,
and cabs(z) is the absolute value of the complex number z
Note that the above supposes that we are using the same index for every eigenvector, though the factor s varies from eigenvector to eigenvector.
This would impose a uniqueness on the eigenvectors, and, one would hope, mean that they varied continuously with the parameters of your matrix.
Related
The scipy.linalg.eigh function can take two matrices as arguments: first the matrix a, of which we will find eigenvalues and eigenvectors, but also the matrix b, which is optional and chosen as the identity matrix in case it is left blank.
In what scenario would someone like to use this b matrix?
Some more context: I am trying to use xdawn covariances from the pyRiemann package. This uses the scipy.linalg.eigh function with a covariance matrix a and a baseline covariance matrix b. You can find the implementation here. This yields an error, as the b matrix in my case is not positive definitive and thus not useable in the scipy.linalg.eigh function. Removing this matrix and just using the identity matrix however solves this problem and yields relatively nice results... The problem is that I do not really understand what I changed, and maybe I am doing something I should not be doing.
This is the code from the pyRiemann package I am using (modified to avoid using functions defined in other parts of the package):
# X are samples (EEG data), y are labels
# shape of X is (1000, 64, 2459)
# shape of y is (1000,)
from scipy.linalg import eigh
Ne, Ns, Nt = X.shape
tmp = X.transpose((1, 2, 0))
b = np.matrix(sklearn.covariance.empirical_covariance(tmp.reshape(Ne, Ns * Nt).T))
for c in self.classes_:
# Prototyped response for each class
P = np.mean(X[y == c, :, :], axis=0)
# Covariance matrix of the prototyper response & signal
a = np.matrix(sklearn.covariance.empirical_covariance(P.T))
# Spatial filters
evals, evecs = eigh(a, b)
# and I am now using the following, disregarding the b matrix:
# evals, evecs = eigh(a)
If A and B were both symmetric matrices that doesn't necessarily have to imply that inv(A)*B must be a symmetric matrix. And so, if i had to solve a generalised eigenvalue problem of Ax=lambda Bx then i would use eig(A,B) rather than eig(inv(A)*B), so that the symmetry isn't lost.
One practical application is in finding the natural frequencies of a dynamic mechanical system from differential equations of the form M (d²x/dt²) = Kx where M is a positive definite matrix known as the mass matrix and K is the stiffness matrix, and x is displacement vector and d²x/dt² is acceleration vector which is the second derivative of the displacement vector. To find the natural frequencies, x can be substituted with x0 sin(ωt) where ω is the natural frequency. The equation reduces to Kx = ω²Mx. Now, one can use eig(inv(K)*M) but that might break the symmetry of the resultant matrix, and so I would use eig(K,M) instead.
A - lambda B x it means that x is not in the same basis as the covariance matrix.
If the matrix is not definite positive it means that there are vectors that can be flipped by your B.
I hope it was helpful.
I am trying to find the top k leading eigenvalues of a numpy matrix (using python dot product notation)
L#L + a*Y#Y.T, where L and Y are a symmetric nxn and an nxd matrix, respectively.
According to the below text from this paper, I should be able to calculate these leading eigenvalues with L#(L#v) + a*X#(X.T#v), where I guess v is an arbitrary vector. The Lanczos paper they cite is here.
I'm not quite sure where to start. I know that scipy has scipy.sparse.linalg.eigsh here, and from the notes it looks like it uses the Lanczos algorithm - but I am at a loss as to whether it's possible to use sparse.linalg.eigsh for my specific use case. I googled around and didn't find a Python implementation for this very quickly -- does anybody know if I can use sparse.linalg.eigsh to calculate this somehow? I definitely don't want to write this algorithm myself.
I also wasn't sure whether to post this in math.stackexchange or here, since it's a question about the Python implementation of a very mathy thing.
You could check scipy.sparse.linalg.eigsh.
import numpy as np;
from scipy.sparse.linalg import eigsh;
from numpy.linalg import eigh
a = 1.4
n = 20;
d = 7;
# random symmetric n x n matrix
L = np.random.randn(n, n)
L = L + L.T
# random n x d matrix
Y = np.random.randn(n, d)
A = L # L.T + a * Y # Y.T # your equation
A must be positive-definite to use eigsh, this is guaranteed to be true if a>0.
You could check the four eigenvalues as follows
eigsh(La, 4)[0]
For reference you can compare based on numpy.linalg.eigh that compute all the eigenvalues. Sort them, and take the last four elements of the sorted array, the results should be close.
np.sort(eigh(La)[0])[-4:]
I want to implement ifft2 using DFT matrix. The following code works for fft2.
import numpy as np
def DFT_matrix(N):
i, j = np.meshgrid(np.arange(N), np.arange(N))
omega = np.exp( - 2 * np.pi * 1j / N )
W = np.power( omega, i * j ) # Normalization by sqrt(N) Not included
return W
sizeM=40
sizeN=20
np.random.seed(0)
rA=np.random.rand(sizeM,sizeN)
rAfft=np.fft.fft2(rA)
dftMtxM=DFT_matrix(sizeM)
dftMtxN=DFT_matrix(sizeN)
# Matrix multiply the 3 matrices together
mA = dftMtxM # rA # dftMtxN
print(np.allclose(np.abs(mA), np.abs(rAfft)))
print(np.allclose(np.angle(mA), np.angle(rAfft)))
To get to ifft2 I assumd I need to change only the dft matrix to it's transpose, so expected the following to work, but I got false for the last two print any suggesetion please?
import numpy as np
def DFT_matrix(N):
i, j = np.meshgrid(np.arange(N), np.arange(N))
omega = np.exp( - 2 * np.pi * 1j / N )
W = np.power( omega, i * j ) # Normalization by sqrt(N) Not included
return W
sizeM=40
sizeN=20
np.random.seed(0)
rA=np.random.rand(sizeM,sizeN)
rAfft=np.fft.ifft2(rA)
dftMtxM=np.conj(DFT_matrix(sizeM))
dftMtxN=np.conj(DFT_matrix(sizeN))
# Matrix multiply the 3 matrices together
mA = dftMtxM # rA # dftMtxN
print(np.allclose(np.abs(mA), np.abs(rAfft)))
print(np.allclose(np.angle(mA), np.angle(rAfft)))
I am going to be building on some things from my answer to your previous question. Please note that I will try to distinguish between the terms Discrete Fourier Transform (DFT) and Fast Fourier Transform (FFT). Remember that DFT is the transform while FFT is only an efficient algorithm for performing it. People, including myself, however very commonly refer to the DFT as FFT since it is practically the only algorithm used for computing the DFT
The problem here is again the normalization of the data. It's interesting that this is such a fundamental and confusing part of any DFT operations yet I couldn't find a good explanation on the internet. I will try to provide a summary at the end about DFT normalization however I think the best way to understand this is by working through some examples yourself.
Why the comparisons fail?
It's important to note, that even though both of the allclose tests seemingly fail, they are actually not a very good method of comparing two complex number arrays.
Difference between two angles
In particular, the problem is when it comes to comparing angles. If you just take the difference of two close angles that are on the border between -pi and pi, you can get a value that is around 2*pi. The allclose just takes differences between values and checks that they are bellow some threshold. Thus in our cases, it can report a false negative.
A better way to compare angles is something along the lines of this function:
def angle_difference(a, b):
diff = a - b
diff[diff < -np.pi] += 2*np.pi
diff[diff > np.pi] -= 2*np.pi
return diff
You can then take the maximum absolute value and check that it's bellow some threshold:
np.max(np.abs(angle_difference(np.angle(mA), np.angle(rAfft)))) < threshold
In the case of your example, the maximum difference was 3.072209153742733e-12.
So the angles are actually correct!
Magnitude scaling
We can get an idea of the issue is when we look at the magnitude ratio between the matrix iDFT and the library iFFT.
print(np.abs(mA)/np.abs(rAfft))
We find that all the values in mA are 800, which means that our absolute values are 800 times larger than those computed by the library. Suspiciously, 800 = 40 * 20, the dimensions of our data! I think you can see where I am going with this.
Confusing DFT normalization
We spot some indications why this is the case when we have a look at the FFT formulas as taken from the Numpy FFT documentation:
You will notice that while the forward transform doesn't normalize by anything. The reverse transform divides the output by 1/N. These are the 1D FFTs but the exact same thing applies in the 2D case, the inverse transform multiplies everything by 1/(N*M)
So in our example, if we update this line, we will get the magnitudes to agree:
mA = dftMtxM # rA/(sizeM * sizeN) # dftMtxN
A side note on comparing the outputs, an alternative way to compare complex numbers is to compare the real and imaginary components:
print(np.allclose(mA.real, rAfft.real))
print(np.allclose(mA.imag, rAfft.imag))
And we find that now indeed both methods agree.
Why all this normalization mess and which should I use?
The fundamental property of the DFT transform must satisfy is that iDFT(DFT(x)) = x. When you work through the math, you find that the product of the two coefficients before the sum has to be 1/N.
There is also something called the Parseval's theorem. In simple terms, it states that the energy in the signals is just the sum of square absolutes in both the time domain and frequency domain. For the FFT this boils down to this relationship:
Here is the function for computing the energy of a signal:
def energy(x):
return np.sum(np.abs(x)**2)
You are basically faced with a choice about the 1/N factor:
You can put the 1/N before the DFT sum. This makes senses as then the k=0 DC component will be equal to the average of the time domain values. However you will have to multiply the energy in frequency domain by N in order to match it with time domain frequency.
N = len(x)
X = np.fft.fft(x)/N # Compute the FFT scaled by `1/N`
# Energy related by `N`
np.allclose(energy(x), energy(X) * N) == True
# Perform some processing...
Y = X * H
y = np.fft.ifft(Y*N) # Compute the iFFT, remember to cancel out the built in `1/N` of ifft
You put the 1/N before the iDFT. This is, slightly counterintuitively, what most implementations, including Numpy do. I could not find a definitive consensus on the reasoning behind this, but I think it has something to do with the implementation efficiency. (If anyone has a better explanation for this, please leave it in the comments) As shown in the equations earlier, the energy in the frequency domain has to be divided by N to match the time domain energy.
N = len(x)
X = np.fft.fft(x) # Compute the FFT without scaling
# Energy, related by 1/N
np.allclose(energy(x), energy(X) / N) == True
# Perform some processing...
Y = X * H
y = np.fft.ifft(Y) # Compute the iFFT with the build in `1/N`
You can split the 1/N by placing 1/sqrt(N) before each of the transforms making them perfectly symmetric. In Numpy, you can provide the parameter norm="ortho" to the fft functions which will make them use the 1/sqrt(N) normalization instead: np.fft.fft(x, norm="ortho") The nice property here is that the energy now matches in both domains.
X = np.fft.fft(x, norm='orth') # Compute the FFT scaled by `1/sqrt(N)`
# Perform some processing...
# Energy are equal:
np.allclose(energy(x), energy(X)) == True
Y = X * H
y = np.fft.ifft(Y, norm='orth') # Compute the iFFT, with scaling by `1/sqrt(N)`
In the end it boils down to what you need. Most of the time the absolute magnitude of your DFT is actually not that important. You are mostly interested in the ratio of various components or you want to perform some operation in the frequency domain but then transform back to the time domain or you are interested in the phase (angles). In all of these case, the normalization does not really play an important role, as long as you stay consistent.
For the forward (multidimensional) FFTW algorithm you can specify that the input numpy.ndarray is real, and the output should be complex. This is done when creating the byte-aligned arrays that go in the arguments of the fft_object:
import numpy as np
import pyfftw
N = 256 # Input array size (preferrably 2^{a}*3^{b}*5^{c}*7^{d}*11^{e}*13^{f}, (e+f = 0,1))
dx = 0.1 # Spacing between mesh points
a = pyfftw.empty_aligned((N, N), dtype='float64')
b = pyfftw.empty_aligned((N, N//2+1), dtype='complex128')
fft_object = pyfftw.FFTW(a, b, axes=(0, 1), direction='FFTW_FORWARD')
The output array is not symmetric and the second axis is truncated up to the positive frequencies. For the complex FFT you can compute the laplacian with the following np.ndarray
kx, ky = np.meshgrid(np.fft.fftfreq(N, dx), np.fft.fftfreq(N, dx)) # Wave vector components
k2 = -4*np.pi**2*(kx*kx+ky*ky) # np.ndarray for the Laplacian operator in "frequency space"
How would it be done in the truncated case? I thought about using:
kx, ky = np.meshgrid(np.fft.fftfreq(N//2+1, dx), np.fft.fftfreq(N, dx)) # The axes conven-
# tions are different
But, would this really work? It seems like it is neglecting the negative frequencies in the "y" direction.
I'm not familiar with pyfftw, but with the numpy.fft module it would work just fine (assuming you use rfftfreq as mentioned in the comments).
To recap: for a real array, a, the fourier transform, b, has a Hermtian-like property: b(-kx,-ky) is the complex conjugate of b(kx,ky).
The real version of the forward fft discards (most of) the redundant information by omitting the negative kys. The real version of the backward fft assumes that the values at the missing frequencies can be found by complex conjugating the appropriate elements.
If you had used the complex fft and kept all frequencies, -k2 * b would still have the Hermitian-like property. So the assumption made by the real backward fft still holds and would give the correct answer.
I guess with pyfftw it will work just fine provided that you specify a float64 array of the correct size for the output for the direction=FFT_BACKWARD case.
I tried to calculate the Pearson's correlation coefficients between every pairs of rows from two 2D arrays. Then, sort the rows/columns of the correlation matrix based on its diagonal elements. First, the correlation coefficient matrix (i.e., 'ccmtx') was calculated from one random matrix (i.e., 'randmtx') in the following code:
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import pearsonr
def correlation_map(x, y):
n_row_x = x.shape[0]
n_row_y = x.shape[0]
ccmtx_xy = np.empty((n_row_x, n_row_y))
for n in range(n_row_x):
for m in range(n_row_y):
ccmtx_xy[n, m] = pearsonr(x[n, :], y[m, :])[0]
return ccmtx_xy
randmtx = np.random.randn(100, 1000) # generating random matrix
#ccmtx = np.corrcoef(randmtx, randmtx) # cc matrix based on numpy.corrcoef
ccmtx = correlation_map(randmtx, randmtx) # cc matrix based on scipy pearsonr
#
ccmtx_diag = np.diagonal(ccmtx)
#
ids, vals = np.argsort(ccmtx_diag, kind = 'mergesort'), np.sort(ccmtx_diag, kind = 'mergesort')
#ids, vals = np.argsort(ccmtx_diag, kind = 'quicksort'), np.sort(ccmtx_diag, kind = 'quicksort')
plt.plot(ids)
plt.show()
plt.plot(ccmtx_diag[ids])
plt.show()
vals[0]
The issue here is when the 'pearsonr' was used, the diagonal elements of 'ccmtx' are exactly 1.0 which makes sense. However, the 'corrcoef' was used, the diagonal elements of 'ccmtrix' are not exactly one (and slightly less than 1 for some diagonals) seemingly due to a precision error of floating point numbers.
I found to be annoying that the auto-correlation matrix of a single matrix have diagnoal elements not being 1.0 since this resulted in the shuffling of rows/columes of the correlation matrix when the matrix is sorted based on the diagonal elements.
My questions are:
[1] is there any good way to accelerate the computation time when I stick to use the 'pearsonr' function? (e.g., vectorized pearsonr?)
[2] Is there any good way/practice to prevent this precision error when using the 'corrcoef' in numpy? (e.g. 'decimals' option in np.around?)
I have searched the correlation coefficient calculations between all pairs of rows or columns from two matrices. However, as the algorithms containe some sort of "cov / variance" operation, this kind of precision issue seems always existing.
Minor point: the 'mergesort' option seems to provide reliable results than the 'quicksort' as the quicksort shuffled 1d array with exactly 1 to random order.
Any thoughts/comments would be greatly appreciated!
For question 1 vectorized pearsonr see the comments to the question.
I will answer only question 2: how to improve the precision of np.corrcoef.
The correlation matrix R is computed from the covariance matrix C according to
.
The implementation is optimized for performance and memory usage. It computes the covariance matrix, and then performs two divisions by sqrt(C_ii) and by sqrt(Cjj). This separate square-rooting is where the imprecision comes from. For example:
np.sqrt(3 * 3) - 3 == 0.0
np.sqrt(3) * np.sqrt(3) - 3 == -4.4408920985006262e-16
We can fix this by implementing our own simple corrcoef routine:
def corrcoef(a, b):
c = np.cov(a, b)
d = np.diag(c)
return c / np.sqrt(d[:, None] * d[None, :])
Note that this implementation requires more memory than the numpy implementation because it needs to store a temporary matrix with size n * n and it is slightly slower because it needs to do n^2 square roots instead of only 2 n.