scipy.signal.fftconvolve doesn't give the required results - python

I have a question regarding python's fftconvolve. In my current research I've been required to calculate some convolution between two functions. To do so I'm calculating it using fourier transform (which I used numpy.fft and normalize it) . The thing is that if I want to compare it using fftconvolve package, it fails to give the correct results. Here is my code:
#!/usr/bin/python
import numpy as np
from scipy.signal import fftconvolve , convolve
def FFT(array , sign):
if sign==1:
return np.fft.fftshift(np.fft.fft(np.fft.fftshift(array))) * dw / (2.0 * np.pi)
elif sign==-1:
return np.fft.fftshift(np.fft.ifft(np.fft.fftshift(array))) * dt * len(array)
def convolve_arrays(array1,array2,sign):
sign = int(sign)
temp1 = FFT(array1 , sign,)
temp2 = FFT(array2 , sign,)
temp3 = np.multiply(temp1 , temp2)
return FFT(temp3 , -1 * sign) / (2. * np.pi)
""" EXAMPLE """
dt = .1
N = 2**17
t_max = N * dt / 2
time = dt * np.arange(-N / 2 , N / 2 , 1)
dw = 2. * np.pi / (N * dt)
w_max = N * dw / 2.
w = dw * np.arange(-N / 2 , N / 2 , 1)
eta_fourier = 1e-10
Gamma = 1.
epsilon = .5
omega = .5
G = zeros(N , complex)
G[:] = 1. / (w[:] - epsilon + 1j * eta_fourier)
D = zeros(N , complex)
D[:] = 1. / (w[:] - omega + 1j * eta_fourier) - 1. / (w[:] + omega + 1j * eta_fourier)
H = convolve_arrays(D , G , 1)
J = fftconvolve(D , G , mode = 'same') * np.pi / (2. * N)
If you plot the real/imaginary part of H, J you'll see a shift in the w axes and also I had to multiply the J's results in order to get somehow close (but still not) to the correct results.
Any suggestions?
Thanks!


Boundary conditions are important when you compute convolutions.
When you convolve two signals, the edges of the result depend on what values you assume outside the edges of the inputs. fftconvolve computes convolution assuming zero-padded boundaries.
Take a look at the source code of fftconvolve. Notice the shenanigans they go through to achieve zero-padded boundary conditions, in particular, these lines:
size = s1 + s2 - 1
...
fsize = 2 ** np.ceil(np.log2(size)).astype(int) #For speed; often suboptimal!
fslice = tuple([slice(0, int(sz)) for sz in size])
...
ret = ifftn(fftn(in1, fsize) * fftn(in2, fsize))[fslice].copy()
...
return _centered(ret, s1) #strips off padding
This is good stuff! It's probably worth reading fftconvolve's code carefully, and a good education if you want to understand Fourier-based convolution.
Brief sketch
The forward FFT zero-pads each signal to prevent periodic boundary conditions:
a = np.array([3, 4, 5])
b = np.fft.ifftn(np.fft.fftn(a, (5,))).real
print b #[ 3. 4. 5. 0. 0.]
the inverse FFT of the product of the forward FFTs gives a padded result:
a = np.array([3, 4, 5])
b = np.array([0., 0.9, 0.1])
b = np.fft.ifftn(np.fft.fftn(a, (5,))*
np.fft.fftn(b, (5,))
).real
print b #[ 0. 2.7 3.9 4.9 0.5]
and the _centered function strips off the extra padding pixels at the end (assuming you use the mode='same' option).

Related

How to correctly extract the phase of the spectrum in python

I describe a pulse in the time domain and do a Fourier Transform to convert it to the frequency domain.
I add an e-index polynomial phase e^{i*phase}to it in the frequency domain,phase is a polynomial.
At this time, I use the angle function under numpy to extract the phase, and what I get is such dense peaks as shown in the figure. I don't know if this is correct and I don't know how should I extract the polynomial again.
import numpy as np
import matplotlib.pyplot as plt
fs = 1e-15
THz = 1e12
nm = 1e-9
c = 3e8
N = 2 ** 13
time_window = 3000 * fs
wavelength = 800 * nm
t = np.linspace(-time_window / 2,time_window / 2, N)
df = np.append(np.linspace(0, N / 2, int(N / 2)),(np.linspace(-N / 2, -1, int(N / 2))))/ time_window
f = c/wavelength + df
dw = 2 * np.pi * df
FWHM = 50 * fs
m = 4 * np.log(2)
A_t = np.exp(-m * t ** 2 * (1 / 2) / FWHM ** 2)
A_w = np.fft.fft(A_t)
GDD = 500 * fs*fs
TOD = 0 * fs*fs*fs
FOD = 0
A_w = np.exp(1j * (GDD / 2.0) * dw**2 +
1j * (TOD / 6.0) * dw ** 3+
1j * (FOD / 24.0) * dw ** 4) * A_w
fig_1 = plt.figure(1, facecolor='w', edgecolor='k')
ax_1 = fig_1.add_subplot(1, 1, 1)
ax_2 = ax_1.twinx()
ax_1.plot(np.fft.fftshift(f/THz),np.fft.fftshift(np.abs(A_w) ** 2 / max(np.abs(A_w) ** 2)),'b')
ax_2.plot(np.fft.fftshift(f/THz),np.fft.fftshift(np.angle(A_w)),'r')
ax_1.set_ylabel('Intensity / a.u.')
ax_2.set_ylabel('Phase / rad')
ax_1.tick_params(axis='y', colors='b')
ax_2.tick_params(axis='y', colors='r')
plt.xlim(300,450)
plt.show()

Two things. To center your data relative to phase, you need to either fftshift your data before the FFT, or flip the sign of the imaginary component in every other result element.
Then look at the magnitude result. When the magnitudes go sufficiently near zero, the phase becomes that of random numerical noise, rather than informative. So the phase of near zero magnitudes can be zeroed to make the plot look cleaner.

How to use Gradient Descent to solve this multiple terms trigonometry function?

Question is like this:
f(x) = A sin(2π * L * x) + B cos(2π * M * x) + C sin(2π * N * x)
and L,M,N are constants integer, 0 <= L,M,N <= 100
and A,B,C can be any possible integers.
Here is the given data:
x = [0,0.01,0.02,0.03,0.04,0.05,0.06,0.07,0.08,0.09,0.1,0.11,0.12,0.13,0.14,0.15,0.16,0.17,0.18,0.19,0.2,0.21,0.22,0.23,0.24,0.25,0.26,0.27,0.28,0.29,0.3,0.31,0.32,0.33,0.34,0.35,0.36,0.37,0.38,0.39,0.4,0.41,0.42,0.43,0.44,0.45,0.46,0.47,0.48,0.49,0.5,0.51,0.52,0.53,0.54,0.55,0.56,0.57,0.58,0.59,0.6,0.61,0.62,0.63,0.64,0.65,0.66,0.67,0.68,0.69,0.7,0.71,0.72,0.73,0.74,0.75,0.76,0.77,0.78,0.79,0.8,0.81,0.82,0.83,0.84,0.85,0.86,0.87,0.88,0.89,0.9,0.91,0.92,0.93,0.94,0.95,0.96,0.97,0.98,0.99]
y = [4,1.240062433,-0.7829654986,-1.332487982,-0.3337640721,1.618033989,3.512512389,4.341307895,3.515268061,1.118929599,-2.097886967,-4.990538967,-6.450324073,-5.831575611,-3.211486891,0.6180339887,4.425660706,6.980842552,7.493970785,5.891593744,2.824429495,-0.5926374511,-3.207870455,-4.263694544,-3.667432785,-2,-0.2617162175,0.5445886005,-0.169441247,-2.323237059,-5.175570505,-7.59471091,-8.488730333,-7.23200463,-3.924327772,0.6180339887,5.138501587,8.38127157,9.532377045,8.495765687,5.902113033,2.849529206,0.4768388529,-0.46697525,0.106795821,1.618033989,3.071952496,3.475795162,2.255463709,-0.4905371745,-4,-7.117914956,-8.727599664,-8.178077181,-5.544088451,-1.618033989,2.365340134,5.169257268,5.995297102,4.758922924,2.097886967,-0.8873135564,-3.06024109,-3.678989552,-2.666365632,-0.6180339887,1.452191817,2.529722611,2.016594378,-0.01374122059,-2.824429495,-5.285215072,-6.302694708,-5.246870619,-2.210419738,2,6.13956874,8.965976562,9.68000641,8.201089581,5.175570505,1.716858387,-1.02183483,-2.278560533,-1.953524751,-0.6180339887,0.7393509358,1.129293593,-0.02181188158,-2.617913164,-5.902113033,-8.727381729,-9.987404016,-9.043589913,-5.984648344,-1.618033989,2.805900027,6.034770001,7.255101454,6.368389697]
enter image description here
How to use Gradient Descent to solve this multiple terms trigonometry function?

Gradient descent is not well suited for optimisation over integers. You can try a navie relaxation where you solve in floats, and hope the rounded solution is still ok.
from autograd import grad, numpy as jnp
import numpy as np
def cast(params):
[A, B, C, L, M, N] = params
L = jnp.minimum(jnp.abs(L), 100)
M = jnp.minimum(jnp.abs(M), 100)
N = jnp.minimum(jnp.abs(N), 100)
return A, B, C, L, M, N
def pred(params, x):
[A, B, C, L, M, N] = cast(params)
return A *jnp.sin(2 * jnp.pi * L * x) + B*jnp.cos(2*jnp.pi * M * x) + C * jnp.sin(2 * jnp.pi * N * x)
x = [0,0.01,0.02,0.03,0.04,0.05,0.06,0.07,0.08,0.09,0.1,0.11,0.12,0.13,0.14,0.15,0.16,0.17,0.18,0.19,0.2,0.21,0.22,0.23,0.24,0.25,0.26,0.27,0.28,0.29,0.3,0.31,0.32,0.33,0.34,0.35,0.36,0.37,0.38,0.39,0.4,0.41,0.42,0.43,0.44,0.45,0.46,0.47,0.48,0.49,0.5,0.51,0.52,0.53,0.54,0.55,0.56,0.57,0.58,0.59,0.6,0.61,0.62,0.63,0.64,0.65,0.66,0.67,0.68,0.69,0.7,0.71,0.72,0.73,0.74,0.75,0.76,0.77,0.78,0.79,0.8,0.81,0.82,0.83,0.84,0.85,0.86,0.87,0.88,0.89,0.9,0.91,0.92,0.93,0.94,0.95,0.96,0.97,0.98,0.99]
y = [4,1.240062433,-0.7829654986,-1.332487982,-0.3337640721,1.618033989,3.512512389,4.341307895,3.515268061,1.118929599,-2.097886967,-4.990538967,-6.450324073,-5.831575611,-3.211486891,0.6180339887,4.425660706,6.980842552,7.493970785,5.891593744,2.824429495,-0.5926374511,-3.207870455,-4.263694544,-3.667432785,-2,-0.2617162175,0.5445886005,-0.169441247,-2.323237059,-5.175570505,-7.59471091,-8.488730333,-7.23200463,-3.924327772,0.6180339887,5.138501587,8.38127157,9.532377045,8.495765687,5.902113033,2.849529206,0.4768388529,-0.46697525,0.106795821,1.618033989,3.071952496,3.475795162,2.255463709,-0.4905371745,-4,-7.117914956,-8.727599664,-8.178077181,-5.544088451,-1.618033989,2.365340134,5.169257268,5.995297102,4.758922924,2.097886967,-0.8873135564,-3.06024109,-3.678989552,-2.666365632,-0.6180339887,1.452191817,2.529722611,2.016594378,-0.01374122059,-2.824429495,-5.285215072,-6.302694708,-5.246870619,-2.210419738,2,6.13956874,8.965976562,9.68000641,8.201089581,5.175570505,1.716858387,-1.02183483,-2.278560533,-1.953524751,-0.6180339887,0.7393509358,1.129293593,-0.02181188158,-2.617913164,-5.902113033,-8.727381729,-9.987404016,-9.043589913,-5.984648344,-1.618033989,2.805900027,6.034770001,7.255101454,6.368389697]
def loss(params):
p = pred(params, np.array(x))
return jnp.mean((np.array(y)-p)**2)
params = np.array([np.random.random()*100 for _ in range(6)])
for _ in range(10000):
g = grad(loss)
params = params - 0.001*g(params)
print("Relaxed solution", cast(params), "loss=", loss(params))
constrained_params = np.round(cast(params))
print("Integer solution", constrained_params, "loss=", loss(constrained_params))
print()
Since the problem will have a lot of local minima, you might need to run it multiple times.

It's quite hard to use gradient descent to find a solution to this problem, because it tends to get stuck when changing the L, M, or N parameters. The gradients for those can push it away from the right solution, unless it is very close to an optimal solution already.
There are ways to get around this, such as basinhopping or random search, but because of the function you're trying to learn, you have a better alternative.
Since you're trying to learn a sinusoid function, you can use an FFT to find the frequencies of the sine waves. Once you have those frequencies, you can find the amplitudes and phases used to generate the same sine wave.
Pardon the messiness of this code, this is my first time using an FFT.
import scipy.fft
import numpy as np
import math
import matplotlib.pyplot as plt
def get_top_frequencies(x, y, num_freqs):
x = np.array(x)
y = np.array(y)
# Find timestep (assume constant timestep)
dt = abs(x[0] - x[-1]) / (len(x) - 1)
# Take discrete FFT of y
spectral = scipy.fft.fft(y)
freq = scipy.fft.fftfreq(y.shape[0], d=dt)
# Cut off top half of frequencies. Assumes input signal is real, and not complex.
spectral = spectral[:int(spectral.shape[0] / 2)]
# Double amplitudes to correct for cutting off top half.
spectral *= 2
# Adjust amplitude by sampling timestep
spectral *= dt
# Get ampitudes for all frequencies. This is taking the magnitude of the complex number
spectral_amplitude = np.abs(spectral)
# Pick frequencies with highest amplitudes
highest_idx = np.argsort(spectral_amplitude)[::-1][:num_freqs]
# Find amplitude, frequency, and phase components of each term
highest_amplitude = spectral_amplitude[highest_idx]
highest_freq = freq[highest_idx]
highest_phase = np.angle(spectral[highest_idx]) / math.pi
# Convert it into a Python function
function = ["def func(x):", "return ("]
for i, components in enumerate(zip(highest_amplitude, highest_freq, highest_phase)):
amplitude, freq, phase = components
plus_sign = " +" if i != (num_freqs - 1) else ""
term = f"{amplitude:.2f} * math.cos(2 * math.pi * {freq:.2f} * x + math.pi * {phase:.2f}){plus_sign}"
function.append(" " + term)
function.append(")")
return "\n ".join(function)
x = [0,0.01,0.02,0.03,0.04,0.05,0.06,0.07,0.08,0.09,0.1,0.11,0.12,0.13,0.14,0.15,0.16,0.17,0.18,0.19,0.2,0.21,0.22,0.23,0.24,0.25,0.26,0.27,0.28,0.29,0.3,0.31,0.32,0.33,0.34,0.35,0.36,0.37,0.38,0.39,0.4,0.41,0.42,0.43,0.44,0.45,0.46,0.47,0.48,0.49,0.5,0.51,0.52,0.53,0.54,0.55,0.56,0.57,0.58,0.59,0.6,0.61,0.62,0.63,0.64,0.65,0.66,0.67,0.68,0.69,0.7,0.71,0.72,0.73,0.74,0.75,0.76,0.77,0.78,0.79,0.8,0.81,0.82,0.83,0.84,0.85,0.86,0.87,0.88,0.89,0.9,0.91,0.92,0.93,0.94,0.95,0.96,0.97,0.98,0.99]
y = [4,1.240062433,-0.7829654986,-1.332487982,-0.3337640721,1.618033989,3.512512389,4.341307895,3.515268061,1.118929599,-2.097886967,-4.990538967,-6.450324073,-5.831575611,-3.211486891,0.6180339887,4.425660706,6.980842552,7.493970785,5.891593744,2.824429495,-0.5926374511,-3.207870455,-4.263694544,-3.667432785,-2,-0.2617162175,0.5445886005,-0.169441247,-2.323237059,-5.175570505,-7.59471091,-8.488730333,-7.23200463,-3.924327772,0.6180339887,5.138501587,8.38127157,9.532377045,8.495765687,5.902113033,2.849529206,0.4768388529,-0.46697525,0.106795821,1.618033989,3.071952496,3.475795162,2.255463709,-0.4905371745,-4,-7.117914956,-8.727599664,-8.178077181,-5.544088451,-1.618033989,2.365340134,5.169257268,5.995297102,4.758922924,2.097886967,-0.8873135564,-3.06024109,-3.678989552,-2.666365632,-0.6180339887,1.452191817,2.529722611,2.016594378,-0.01374122059,-2.824429495,-5.285215072,-6.302694708,-5.246870619,-2.210419738,2,6.13956874,8.965976562,9.68000641,8.201089581,5.175570505,1.716858387,-1.02183483,-2.278560533,-1.953524751,-0.6180339887,0.7393509358,1.129293593,-0.02181188158,-2.617913164,-5.902113033,-8.727381729,-9.987404016,-9.043589913,-5.984648344,-1.618033989,2.805900027,6.034770001,7.255101454,6.368389697]
print(get_top_frequencies(x, y, 3))
That produces this function:
def func(x):
return (
5.00 * math.cos(2 * math.pi * 10.00 * x + math.pi * 0.50) +
4.00 * math.cos(2 * math.pi * 5.00 * x + math.pi * -0.00) +
2.00 * math.cos(2 * math.pi * 3.00 * x + math.pi * -0.50)
)
Which is not quite the format you specified - you asked for two sins and one cos, and for no phase parameter. However, using the trigonometric identity cos(x) = sin(pi/2 - x), you can convert this into an equivalent expression that matches what you want:
def func(x):
return (
5.00 * math.sin(2 * math.pi * -10.00 * x) +
4.00 * math.cos(2 * math.pi * 5.00 * x) +
2.00 * math.sin(2 * math.pi * 3.00 * x)
)
And there's the original function!

How to use numy linalg lstsq to fit two datasets with same slope but different intercept?

I am trying to do weighted least-square fitting, and came across numpy.linalg.lstsq. I need to fit the weighted least squares. So, the following works:
# Generate some synthetic data from the model.
N = 50
x = np.sort(10 * np.random.rand(N))
yerr = 0.1 + 0.5 * np.random.rand(N)
y = 10.0 * x + 15
y += yerr * np.random.randn(N)
#do the fitting
err = 1/yerr**2
W = np.sqrt(np.diag(err))
x = x.flatten()
y = y.flatten()
A = np.vstack([x, np.ones(len(x))]).T
xw = np.dot(W,A)
yw = np.dot(W,y)
m, b = np.linalg.lstsq(xw, yw)[0]
which gives me the best-fit slope and intercept. Now, suppose I have two datasets with same slope but different intercepts? How would I do a joint fit such that I get best-fit slope plus two intercepts. I still need to have the weighted least square version. For an unweighted case, I found that the following works:
(m,b1,b2),_,_,_ = np.linalg.lstsq(np.stack([np.concatenate((x1,x2)),
np.concatenate([np.ones(len(x1)),np.zeros(len(x2))]),
np.concatenate([np.zeros(len(x1)),np.ones(len(x2))])]).T,
np.concatenate((y1,y2)))

First of all I rewrite your first approach as it can be written clearer in my opinion like this
weights = 1 / yerr
m, b = np.linalg.lstsq(np.c_[weights * x, weights], weights * y, rcond=None)[0]
To fit 2 datasets you can stack 2 arrays but make 0 some elements of matrix.
np.random.seed(12)
N = 3
x = np.sort(10 * np.random.rand(N))
yerr = 0.1 + 0.5 * np.random.rand(N)
y = 10.0 * x + 15
y += yerr * np.random.randn(N)
M = 2
x1 = np.sort(10 * np.random.rand(M))
yerr1 = 0.1 * 0.5 * np.random.rand(M)
y1 = 10.0 * x1 + 25
y1 += yerr1 * np.random.randn(M)
#do the fitting
weights = 1 / yerr
weights1 = 1 / yerr1
first_column = np.r_[weights * x, weights1 * x1]
second_column = np.r_[weights, [0] * x1.size]
third_column = np.r_[[0] * x.size, weights1]
a = np.c_[first_column, second_column, third_column]
print(a)
# [[ 4.20211437 2.72576342 0. ]
# [ 24.54293941 9.32075195 0. ]
# [ 13.22997409 1.78771428 0. ]
# [126.37829241 0. 26.03711851]
# [686.96961895 0. 124.44253391]]
c = np.r_[weights * y, weights1 * y1]
print(c)
# [ 83.66073785 383.70595203 159.12058215 1914.59065915 9981.85549321]
m, b1, b2 = np.linalg.lstsq(a, c, rcond=None)[0]
print(m, b1, b2)
# 10.012202998026055 14.841412336510793 24.941219918240172
EDIT
If you want different slopes and one intercept you can do it this way. Probably it is better to grasp the general idea on the one slope 2 intercepts case. Take a look to array a: you construct it from weights as well as c so now it is unweighted problem. You try to find such vector = [slope, intercept1, intercept2] that a # vector = c (as much as possible by minimizing sum of squares of differences). By putting zeros in a we make it separable: upper part of matrix a vary slope and intercept1 and down part of a vary slope and intercept2. Similar to 2 slopes case with vector = [slope1, slope2, intercept].
first_column = np.r_[weights * x, [0] * x1.size]
second_column = np.r_[[0] * x.size, weights1 * x1]
third_column = np.r_[weights, weights1]

Is there a faster way of repeating a chunk of code x times and taking an average?

Starting with:
a,b=np.ogrid[0:n+1:1,0:n+1:1]
B=np.exp(1j*(np.pi/3)*np.abs(a-b))
B[z,b] = np.exp(1j * (np.pi/3) * np.abs(z - b +x))
B[a,z] = np.exp(1j * (np.pi/3) * np.abs(a - z +x))
B[diag,diag]=1-1j/np.sqrt(3)
this produces an n*n grid that acts as a matrix.
n is just a number chosen to represent the indices, i.e. an a*b matrix where a and b both go up to n.
Where z is a constant I choose to replace a row and column with the B[z,b] and B[a,z] formulas. (Essentially the same formula but with a small number added to the np.abs(a-b))
The diagonal of the matrix is given by the bottom line:
B[diag,diag]=1-1j/np.sqrt(3)
where,
diag=np.arange(n+1)
I would like to repeat this code 50 times where the only thing that changes is x so I will end up with 50 versions of the B np.ogrid. x is a randomly generated number between -0.8 and 0.8 each time.
x=np.random.uniform(-0.8,0.8)
I want to generate 50 versions of B with random values of x each time and take a geometric average of the 50 versions of B using the definition:
def geo_mean(y):
y = np.asarray(y)
return np.prod(y ** (1.0 / y.shape[0]), axis=-1)
I have tried to set B as a function of some index and then use a for _ in range(): loop, this doesn't work. Aside from copy and pasting the block 50 times and denoting each one as B1, B2, B3 etc; I can't think of another way of working this out.
EDIT:
I'm now using part of a given solution in order to show clearly what I am looking for:
#A matrix with 50 random values between -0.8 and 0.8 to be used in the loop
X=np.random.uniform(-0.8,0.8, (50,1))
#constructing the base array before modification by random x values in position z
a,b = np.ogrid[0:n+1:1,0:n+1:1]
B = np.exp(1j * ( np.pi / 3) * np.abs( a - b ))
B[diag,diag] = 1 - 1j / np.sqrt(3)
#list to store all modified arrays
randomarrays = []
for i in range( 0,50 ):
#copy array and modify it
Bnew = np.copy( B )
Bnew[z, b] = np.exp( 1j * ( np.pi / 3 ) * np.abs(z - b + X[i]))
Bnew[a, z] = np.exp( 1j * ( np.pi / 3 ) * np.abs(a - z + X[i]))
randomarrays.append(Bnew)
Bstack = np.dstack(randomarrays)
#calculate the geometric mean value along the axis that was the row in 2D arrays
B0 = geo_mean(Bstack)
From this example, every iteration of i uses the same value of X, I can't seem to get a way to get each new loop of i to use the next value in the matrix X. I am unsure of the ++ action in python, I know it does not work in python, I just don't know how to use the python equivalent. I want a loop to use a value of X, then the next loop to use the next value and so on and so forth so I can dstack all the matrices at the end and find a geo_mean for each element in the stacked matrices.

One pedestrian way would be to use a list comprehension or generator expression:
>>> def f(n, z, x):
... diag = np.arange(n+1)
... a,b=np.ogrid[0:n+1:1,0:n+1:1]
... B=np.exp(1j*(np.pi/3)*np.abs(a-b))
... B[z,b] = np.exp(1j * (np.pi/3) * np.abs(z - b +x))
... B[a,z] = np.exp(1j * (np.pi/3) * np.abs(a - z +x))
... B[diag,diag]=1-1j/np.sqrt(3)
... return B
...
>>> X = np.random.uniform(-0.8, 0.8, (10,))
>>> np.prod((*map(np.power, map(f, 10*(4,), 10*(2,), X), 10 * (1/10,)),), axis=0)
But in your concrete example we can do much better than that;
using the identity exp(a) x exp(b) = exp(a + b) we can convert the geometric mean after exponentiation to an arithmetic mean before exponentition. A bit of care is required because of the multivaluedness of the complex n-th root which occurs in the geometric mean. In the code below we normalize the angles occurring to range -pi, pi so as to always hit the same branch as the n-th root.
Please also note that the geo_mean function you provide is definitely wrong. It fails the basic sanity check that taking the average of copies of the same thing should return the same thing. I've provided a better version. It is still not perfect, but I think there actually is no perfect solution, because of the nonuniqueness of the complex root.
Because of this I recommend taking the average before exponentiating. As long as your random spread is less than pi this allows a well-defined averaging procedure with an average that is actually close to the samples
import numpy as np
def f(n, z, X, do_it_pps_way=True):
X = np.asanyarray(X)
diag = np.arange(n+1)
a,b=np.ogrid[0:n+1:1,0:n+1:1]
B=np.exp(1j*(np.pi/3)*np.abs(a-b))
X = X.reshape(-1,1,1)
if do_it_pps_way:
zbx = np.mean(np.abs(z-b+X), axis=0)
azx = np.mean(np.abs(a-z+X), axis=0)
else:
zbx = np.mean((np.abs(z-b+X)+3) % 6 - 3, axis=0)
azx = np.mean((np.abs(a-z+X)+3) % 6 - 3, axis=0)
B[z,b] = np.exp(1j * (np.pi/3) * zbx)
B[a,z] = np.exp(1j * (np.pi/3) * azx)
B[diag,diag]=1-1j/np.sqrt(3)
return B
def geo_mean(y):
y = np.asarray(y)
dim = len(y.shape)
y = np.atleast_2d(y)
v = np.prod(y, axis=0) ** (1.0 / y.shape[0])
return v[0] if dim == 1 else v
def geo_mean_correct(y):
y = np.asarray(y)
return np.prod(y ** (1.0 / y.shape[0]), axis=0)
# demo that orig geo_mean is wrong
B = np.exp(1j * np.random.random((5, 5)))
# the mean of four times the same thing should be the same thing:
if not np.allclose(B, geo_mean([B, B, B, B])):
print('geo_mean failed')
if np.allclose(B, geo_mean_correct([B, B, B, B])):
print('but geo_mean_correct works')
n, z, m = 10, 3, 50
X = np.random.uniform(-0.8, 0.8, (m,))
B0 = f(n, z, X, do_it_pps_way=False)
B1 = np.prod((*map(np.power, map(f, m*(n,), m*(z,), X), m * (1/m,)),), axis=0)
B2 = geo_mean_correct([f(n, z, x) for x in X])
# This is the recommended way:
B_recommended = f(n, z, X, do_it_pps_way=True)
print()
print(np.allclose(B1, B0))
print(np.allclose(B2, B1))

I think you should rely more on numpy functionality, when approaching your problem. Not a numpy expert myself, so there is surely room for improvement:
from scipy.stats import gmean
n = 2
z = 1
a = np.arange(n + 1).reshape(1, n + 1)
#constructing the base array before modification by random x values in position z
B = np.exp(1j * (np.pi / 3) * np.abs(a - a.T))
B[a, a] = 1 - 1j / np.sqrt(3)
#list to store all modified arrays
random_arrays = []
for _ in range(50):
#generate random x value
x=np.random.uniform(-0.8, 0.8)
#copy array and modify it
B_new = np.copy(B)
B_new[z, a] = np.exp(1j * (np.pi / 3) * np.abs(z - a + x))
B_new[a, z] = np.exp(1j * (np.pi / 3) * np.abs(a - z + x))
random_arrays.append(B_new)
#store all B arrays as a 3D array
B_stack = np.stack(random_arrays)
#calculate the geometric mean value along the axis that was the row in 2D arrays
geom_mean_for_rows = gmean(B_stack, axis = 2)
It uses the geometric mean function from scipy.stats module to have a vectorised approach for this calculation.

Chaotic billiards simulation

I came to ask for some help with maths and programming.
What am I trying to do? I'm trying to implement a simulation of a chaotic billiard system, following the algorithm in this excerpt.
How am I trying it? Using numpy and matplotlib, I implemented the following code
def boundaryFunction(parameter):
return 1 + 0.1 * np.cos(parameter)
def boundaryDerivative(parameter):
return -0.1 * np.sin(parameter)
def trajectoryFunction(parameter):
aux = np.sin(beta - phi) / np.sin(beta - parameter)
return boundaryFunction(phi) * aux
def difference(parameter):
return trajectoryFunction(parameter) - boundaryFunction(parameter)
def integrand(parameter):
rr = boundaryFunction(parameter)
dd = boundaryDerivative (parameter)
return np.sqrt(rr ** 2 + dd ** 2)
##### Main #####
length_vals = np.array([], dtype=np.float64)
alpha_vals = np.array([], dtype=np.float64)
# nof initial phi angles, alpha angles, and nof collisions for each.
n_phi, n_alpha, n_cols, count = 10, 10, 10, 0
# Length of the boundary
total_length, err = integrate.quad(integrand, 0, 2 * np.pi)
for phi in np.linspace(0, 2 * np.pi, n_phi):
for alpha in np.linspace(0, 2 * np.pi, n_alpha):
for n in np.arange(1, n_cols):
nu = np.arctan(boundaryFunction(phi) / boundaryDerivative(phi))
beta = np.pi + phi + alpha - nu
# Determines next impact coordinate.
bnds = (0, 2 * np.pi)
phi_new = optimize.minimize_scalar(difference, bounds=bnds, method='bounded').x
nu_new = np.arctan(boundaryFunction(phi_new) / boundaryDerivative(phi_new))
# Reflection angle with relation to tangent.
alpha_new = phi_new - phi + nu - nu_new - alpha
# Arc length for current phi value.
arc_length, err = integrate.quad(integrand, 0, phi_new)
# Append values to list
length_vals = np.append(length_vals, arc_length / total_length)
alpha_vals = np.append(alpha_vals, alpha)
count += 1
print "{}%" .format(100 * count / (n_phi * n_alpha))
What is the problem? When calculating phi_new, the equation has two solutions (assuming the boundary is convex, which is.) I must enforce that phi_new is the solution which is different from phi, but I don't know how to do that. Are there more issues with the code?
What should the output be? A phase space diagram of S x Alpha, looking like this.
Any help is very appreciated! Thanks in advance.

One way you could try would be (given there really are only two solutions) would be
epsilon = 1e-7 # tune this
delta = 1e-4 # tune this
# ...
bnds = (0, 2 * np.pi)
phi_new = optimize.minimize_scalar(difference, bounds=bnds, method='bounded').x
if abs(phi_new - phi) < epsilon:
bnds_1 = (0, phi - delta)
phi_new_1 = optimize.minimize_scalar(difference, bounds=bnds_1, method='bounded').x
bnds_2 = (phi + delta, 2 * np.pi)
phi_new_2 = optimize.minimize_scalar(difference, bounds=bnds_2, method='bounded').x
if difference(phi_new_1) < difference(phi_new_2):
phi_new = phi_new_1
else:
phi_new = phi_new_2
Alternatively, you could introduce a penalty-term, e.g. delta*exp(eps/(x-phi)^2) with appropriate choices of epsilon and delta.

Categories

Resources