Develop Reference

Are these functions equivalent?

I am building a neural network that makes use of T-distribution noise. I am using functions defined in the numpy library np.random.standard_t and the one defined in tensorflow tf.distributions.StudentT. The link to the documentation of the first function is here and that to the second function is here. I am using the said functions like below:
a = np.random.standard_t(df=3, size=10000) # numpy's function
t_dist = tf.distributions.StudentT(df=3.0, loc=0.0, scale=1.0)
sess = tf.Session()
b = sess.run(t_dist.sample(10000))
In the documentation provided for the Tensorflow implementation, there's a parameter called scale whose description reads
The scaling factor(s) for the distribution(s). Note that scale is not technically the standard deviation of this distribution but has semantics more similar to standard deviation than variance.
I have set scale to be 1.0 but I have no way of knowing for sure if these refer to the same distribution.
Can someone help me verify this? Thanks

I would say they are, as their sampling is defined in almost the exact same way in both cases. This is how the sampling of tf.distributions.StudentT is defined:
def _sample_n(self, n, seed=None):
# The sampling method comes from the fact that if:
# X ~ Normal(0, 1)
# Z ~ Chi2(df)
# Y = X / sqrt(Z / df)
# then:
# Y ~ StudentT(df).
seed = seed_stream.SeedStream(seed, "student_t")
shape = tf.concat([[n], self.batch_shape_tensor()], 0)
normal_sample = tf.random.normal(shape, dtype=self.dtype, seed=seed())
df = self.df * tf.ones(self.batch_shape_tensor(), dtype=self.dtype)
gamma_sample = tf.random.gamma([n],
0.5 * df,
beta=0.5,
dtype=self.dtype,
seed=seed())
samples = normal_sample * tf.math.rsqrt(gamma_sample / df)
return samples * self.scale + self.loc # Abs(scale) not wanted.
So it is a standard normal sample divided by the square root of a chi-square sample with parameter df divided by df. The chi-square sample is taken as a gamma sample with parameter 0.5 * df and rate 0.5, which is equivalent (chi-square is a special case of gamma). The scale value, like the loc, only comes into play in the last line, as a way to "relocate" the distribution sample at some point and scale. When scale is one and loc is zero, they do nothing.
Here is the implementation for np.random.standard_t:
double legacy_standard_t(aug_bitgen_t *aug_state, double df) {
double num, denom;
num = legacy_gauss(aug_state);
denom = legacy_standard_gamma(aug_state, df / 2);
return sqrt(df / 2) * num / sqrt(denom);
})
So essentially the same thing, slightly rephrased. Here we have also have a gamma with shape df / 2 but it is standard (rate one). However, the missing 0.5 is now by the numerator as / 2 within the sqrt. So it's just moving the numbers around. Here there is no scale or loc, though.
In truth, the difference is that in the case of TensorFlow the distribution really is a noncentral t-distribution. A simple empirical proof that they are the same for loc=0.0 and scale=1.0 is to plot histograms for both distributions and see how close they look.
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
np.random.seed(0)
t_np = np.random.standard_t(df=3, size=10000)
with tf.Graph().as_default(), tf.Session() as sess:
tf.random.set_random_seed(0)
t_dist = tf.distributions.StudentT(df=3.0, loc=0.0, scale=1.0)
t_tf = sess.run(t_dist.sample(10000))
plt.hist((t_np, t_tf), np.linspace(-10, 10, 20), label=['NumPy', 'TensorFlow'])
plt.legend()
plt.tight_layout()
plt.show()
Output:
That looks pretty close. Obviously, from the point of view of statistical samples, this is not any kind of proof. If you were not still convinced, there are some statistical tools for testing whether a sample comes from a certain distribution or two samples come from the same distribution.

Python: Kernel Density Estimation for positive values

I want to get kernel density estimation for positive data points. Using Python Scipy Stats package, I came up with the following code.
def get_pdf(data):
a = np.array(data)
ag = st.gaussian_kde(a)
x = np.linspace(0, max(data), max(data))
y = ag(x)
return x, y
This works perfectly for most data sets, but it gives an erroneous result for "all positive" data points. To make sure this works correctly, I use numerical integration to compute the area under this curve.
def trapezoidal_2(ag, a, b, n):
h = np.float(b - a) / n
s = 0.0
s += ag(a)[0]/2.0
for i in range(1, n):
s += ag(a + i*h)[0]
s += ag(b)[0]/2.0
return s * h
Since the data is spread in the region (0, int(max(data))), we should get a value close to 1, when executing the following line.
b = 1
data = st.pareto.rvs(b, size=10000)
data = list(data)
a = np.array(data)
ag = st.gaussian_kde(a)
trapezoidal_2(ag, 0, int(max(data)), int(max(data))*2)
But it gives a value close to 0.5 when I test.
But when I intergrate from -100 to max(data), it provides a value close to 1.
trapezoidal_2(ag, -100, int(max(data)), int(max(data))*2+200)
The reason is, ag (KDE) is defined for values less than 0, even though the original data set contains only positive values.
So how can I get a kernel density estimation that considers only positive values, such that area under the curve in the region (o, max(data)) is close to 1?

The choice of the bandwidth is quite important when performing kernel density estimation. I think the Scott's Rule and Silverman's Rule work well for distribution similar to a Gaussian. However, they do not work well for the Pareto distribution.
Quote from the doc:
Bandwidth selection strongly influences the estimate obtained from
the KDE (much more so than the actual shape of the kernel). Bandwidth selection
can be done by a "rule of thumb", by cross-validation, by "plug-in
methods" or by other means; see [3], [4] for reviews. gaussian_kde
uses a rule of thumb, the default is Scott's Rule.
Try with different bandwidth values, for example:
import numpy as np
import matplotlib.pyplot as plt
from scipy import stats
b = 1
sample = stats.pareto.rvs(b, size=3000)
kde_sample_scott = stats.gaussian_kde(sample, bw_method='scott')
kde_sample_scalar = stats.gaussian_kde(sample, bw_method=1e-3)
# Compute the integrale:
print('integrale scott:', kde_sample_scott.integrate_box_1d(0, np.inf))
print('integrale scalar:', kde_sample_scalar.integrate_box_1d(0, np.inf))
# Graph:
x_span = np.logspace(-2, 1, 550)
plt.plot(x_span, stats.pareto.pdf(x_span, b), label='theoretical pdf')
plt.plot(x_span, kde_sample_scott(x_span), label="estimated pdf 'scott'")
plt.plot(x_span, kde_sample_scalar(x_span), label="estimated pdf 'scalar'")
plt.xlabel('X'); plt.legend();
gives:
integrale scott: 0.5572130540733236
integrale scalar: 0.9999999999968957
and:
We see that the kde using the Scott method is wrong.

How to properly sample truncated distributions?

I am trying to learn how to sample truncated distributions. To begin with I decided to try a simple example I found here example
I didn't really understand the division by the CDF, therefore I decided to tweak the algorithm a bit. Being sampled is an exponential distribution for values x>0 Here is an example python code:
# Sample exponential distribution for the case x>0
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import norm
def pdf(x):
return x*np.exp(-x)
xvec=np.zeros(1000000)
x=1.
for i in range(1000000):
a=x+np.random.normal()
xs=x
if a > 0. :
xs=a
A=pdf(xs)/pdf(x)
if np.random.uniform()<A :
x=xs
xvec[i]=x
x=np.linspace(0,15,1000)
plt.plot(x,pdf(x))
plt.hist([x for x in xvec if x != 0],bins=150,normed=True)
plt.show()
Ant the output is:
The code above seems to work fine only for when using the condition if a > 0. :, i.e. positive x, choosing another condition (e.g. if a > 0.5 :) produces wrong results.
Since my final goal was to sample a 2D-Gaussian - pdf on a truncated interval I tried extending the simple example using the exponential distribution (see the code below). Unfortunately, since the simple case didn't work, I assume that the code given below would yield wrong results.
I assume that all this can be done using the advanced tools of python. However, since my primary idea was to understand the principle behind, I would greatly appreciate your help to understand my mistake.
Thank you for your help.
EDIT:
# code updated according to the answer of CrazyIvan
from scipy.stats import multivariate_normal
RANGE=100000
a=2.06072E-02
b=1.10011E+00
a_range=[0.001,0.5]
b_range=[0.01, 2.5]
cov=[[3.1313994E-05, 1.8013737E-03],[ 1.8013737E-03, 1.0421529E-01]]
x=a
y=b
j=0
for i in range(RANGE):
a_t,b_t=np.random.multivariate_normal([a,b],cov)
# accept if within bounds - all that is neded to truncate
if a_range[0]<a_t and a_t<a_range[1] and b_range[0]<b_t and b_t<b_range[1]:
print(dx,dy)
EDIT:
I changed the code by norming the analytic pdf according to this scheme, and according to the answers given by, #Crazy Ivan and #Leandro Caniglia , for the case where the bottom of the pdf is removed. That is dividing by (1-CDF(0.5)) since my accept condition is x>0.5. This seems again to show some discrepancies. Again the mystery prevails ..
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import norm
def pdf(x):
return x*np.exp(-x)
# included the corresponding cdf
def cdf(x):
return 1. -np.exp(-x)-x*np.exp(-x)
xvec=np.zeros(1000000)
x=1.
for i in range(1000000):
a=x+np.random.normal()
xs=x
if a > 0.5 :
xs=a
A=pdf(xs)/pdf(x)
if np.random.uniform()<A :
x=xs
xvec[i]=x
x=np.linspace(0,15,1000)
# new part norm the analytic pdf to fix the area
plt.plot(x,pdf(x)/(1.-cdf(0.5)))
plt.hist([x for x in xvec if x != 0],bins=200,normed=True)
plt.savefig("test_exp.png")
plt.show()
It seems that this can be cured by choosing larger shift size
shift=15.
a=x+np.random.normal()*shift.
which is in general an issue of the Metropolis - Hastings. See the graph below:
I also checked shift=150
Bottom line is that changing the shift size definitely improves the convergence. The misery is why, since the Gaussian is unbounded.

You say you want to learn the basic idea of sampling a truncated distribution, but your source is a blog post about
Metropolis–Hastings algorithm? Do you actually need this "method for obtaining a sequence of random samples from a probability distribution for which direct sampling is difficult"? Taking this as your starting point is like learning English by reading Shakespeare.
Truncated normal
For truncated normal, basic rejection sampling is all you need: generate samples for original distribution, reject those outside of bounds. As Leandro Caniglia noted, you should not expect truncated distribution to have the same PDF except on a shorter interval — this is plain impossible because the area under the graph of a PDF is always 1. If you cut off stuff from sides, there has to be more in the middle; the PDF gets rescaled.
It's quite inefficient to gather samples one by one, when you need 100000. I would grab 100000 normal samples at once, accept only those that fit; then repeat until I have enough. Example of sampling truncated normal between amin and amax:
import numpy as np
n_samples = 100000
amin, amax = -1, 2
samples = np.zeros((0,)) # empty for now
while samples.shape[0] < n_samples:
s = np.random.normal(0, 1, size=(n_samples,))
accepted = s[(s >= amin) & (s <= amax)]
samples = np.concatenate((samples, accepted), axis=0)
samples = samples[:n_samples] # we probably got more than needed, so discard extra ones
And here is the comparison with the PDF curve, rescaled by division by cdf(amax) - cdf(amin) as explained above.
from scipy.stats import norm
_ = plt.hist(samples, bins=50, density=True)
t = np.linspace(-2, 3, 500)
plt.plot(t, norm.pdf(t)/(norm.cdf(amax) - norm.cdf(amin)), 'r')
plt.show()
Truncated multivariate normal
Now we want to keep the first coordinate between amin and amax, and the second between bmin and bmax. Same story, except there will be a 2-column array and the comparison with bounds is done in a relatively sneaky way:
(np.min(s - [amin, bmin], axis=1) >= 0) & (np.max(s - [amax, bmax], axis=1) <= 0)
This means: subtract amin, bmin from each row and keep only the rows where both results are nonnegative (meaning we had a >= amin and b >= bmin). Also do a similar thing with amax, bmax. Accept only the rows that meet both criteria.
n_samples = 10
amin, amax = -1, 2
bmin, bmax = 0.2, 2.4
mean = [0.3, 0.5]
cov = [[2, 1.1], [1.1, 2]]
samples = np.zeros((0, 2)) # 2 columns now
while samples.shape[0] < n_samples:
s = np.random.multivariate_normal(mean, cov, size=(n_samples,))
accepted = s[(np.min(s - [amin, bmin], axis=1) >= 0) & (np.max(s - [amax, bmax], axis=1) <= 0)]
samples = np.concatenate((samples, accepted), axis=0)
samples = samples[:n_samples, :]
Not going to plot, but here are some values: naturally, within bounds.
array([[ 0.43150033, 1.55775629],
[ 0.62339265, 1.63506963],
[-0.6723598 , 1.58053835],
[-0.53347361, 0.53513105],
[ 1.70524439, 2.08226558],
[ 0.37474842, 0.2512812 ],
[-0.40986396, 0.58783193],
[ 0.65967087, 0.59755193],
[ 0.33383214, 2.37651975],
[ 1.7513789 , 1.24469918]])

To compute the truncated density function pdf_t from the entire density function pdf, do the following:
Let [a, b] be the truncation interval; (x axis)
Let A := cdf(a) and B := cdf(b); (cdf = non-truncated cumulative distribution function)
Then pdf_t(x) := pdf(x) / (B - A) if x in [a, b] and 0 elsewhere.
In cases where a = -infinity (resp. b = +infinity), take A := 0 (resp. B := 1).
As for the "mystery" you see
please note that your blue curve is wrong. It is not the pdf of your truncated distribution, it is just the pdf of the non-truncated one, scaled by the correct amount (division by 1-cdf(0.5)). The actual truncated pdf curve starts with a vertical line on x = 0.5 which goes up until it reaches your current blue curve. In other words, you only scaled the curve but forgot to truncate it, in this case to the left. Such a truncation corresponds to the "0 elsewhere" part of step 3 in the algorithm above.

Least Squares method in practice

Very simple regression task. I have three variables x1, x2, x3 with some random noise. And I know target equation: y = q1*x1 + q2*x2 + q3*x3. Now I want to find target coefs: q1, q2, q3 evaluate the
performance using the mean Relative Squared Error (RSE) (Prediction/Real - 1)^2 to evaluate the performance of our prediction methods.
In the research, I see that this is ordinary Least Squares Problem. But I can't get from examples on the internet how to solve this particular problem in Python. Let say I have data:
import numpy as np
sourceData = np.random.rand(1000, 3)
koefs = np.array([1, 2, 3])
target = np.dot(sourceData, koefs)
(In real life that data are noisy, with not normal distribution.) How to find this koefs using Least Squares approach in python? Any lib usage.

#ayhan made a valuable comment.
And there is a problem with your code: Actually there is no noise in the data you collect. The input data is noisy, but after the multiplication, you don't add any additional noise.
I've added some noise to your measurements and used the least squares formula to fit the parameters, here's my code:
data = np.random.rand(1000,3)
true_theta = np.array([1,2,3])
true_measurements = np.dot(data, true_theta)
noise = np.random.rand(1000) * 1
noisy_measurements = true_measurements + noise
estimated_theta = np.linalg.inv(data.T # data) # data.T # noisy_measurements
The estimated_theta will be close to true_theta. If you don't add noise to the measurements, they will be equal.
I've used the python3 matrix multiplication syntax.
You could use np.dot instead of #
That makes the code longer, so I've split the formula:
MTM_inv = np.linalg.inv(np.dot(data.T, data))
MTy = np.dot(data.T, noisy_measurements)
estimated_theta = np.dot(MTM_inv, MTy)
You can read up on least squares here: https://en.wikipedia.org/wiki/Linear_least_squares_(mathematics)#The_general_problem
UPDATE:
Or you could just use the builtin least squares function:
np.linalg.lstsq(data, noisy_measurements)

In addition to the #lhk answer I have found great scipy Least Squares function. It is easy to get the requested behavior with it.
This way we can provide a custom function that returns residuals and form Relative Squared Error instead of absolute squared difference:
import numpy as np
from scipy.optimize import least_squares
data = np.random.rand(1000,3)
true_theta = np.array([1,2,3])
true_measurements = np.dot(data, true_theta)
noise = np.random.rand(1000) * 1
noisy_measurements = true_measurements + noise
#noisy_measurements[-1] = data[-1] # (1000 * true_theta) - uncoment this outliner to see how much Relative Squared Error esimator works better then default abs diff for this case.
def my_func(params, x, y):
res = (x # params) / y - 1 # If we change this line to: (x # params) - y - we will got the same result as np.linalg.lstsq
return res
res = least_squares(my_func, x0, args=(data, noisy_measurements) )
estimated_theta = res.x
Also, we can provide custom loss with loss argument function that will process the residuals and form final loss.

statistics for histogram of periodic data

For a series of angle values in (-pi, pi) range, I make a histogram. Is there an effective way to calculate a mean and modal (post probable) value? Consider following examples:
import numpy as N, cmath
deg = N.pi/180.
d = N.array([-175., 170, 175, 179, -179])*deg
i = N.sum(N.exp(1j*d))
ave = cmath.phase(i)
i /= float(d.size)
stdev = -2. * N.log(N.sqrt(i.real**2 + i.imag**2))
print ave/deg, stdev/deg
Now, let's have a histogram:
counts, bins = N.histogram(data, N.linspace(-N.pi, N.pi, 360))
Is it possible to calculate mean, mode having counts and bins? For non-periodic data, calculation of a mean is straightforward:
ave = sum(counts*bins[:-1])
Calculations of a modal value requires more effort. Actually, I'm not sure my code below is correct: firstly, I identify bins which occur most frequently and then I calculate an arithmetic mean:
cmax = bins[N.argmax(counts)]
mode = N.mean(N.take(bins, N.nonzero(counts == cmax)[0]))
I have no idea, how to calculate standard deviation from such data, though. One obvious solution to all my problems (at least those described above) is to convert histogram data to a data series and then use it in calculations. This is not elegant, however, and inefficient.
Any hints will be very appreciated.
This is the partial solution I wrote.
import numpy as N, cmath
import scipy.stats as ST
d = [-175, 170.2, 175.57, 179, -179, 170.2, 175.57, 170.2]
deg = N.pi/180.
data = N.array(d)*deg
i = N.sum(N.exp(1j*data))
ave = cmath.phase(i) # correct and exact mean for periodic data
wrong_ave = N.mean(d)
i /= float(data.size)
stdev = -2. * N.log(N.sqrt(i.real**2 + i.imag**2))
wrong_stdev = N.std(d)
bins = N.linspace(-N.pi, N.pi, 360)
counts, bins = N.histogram(data, bins, normed=False)
# consider it weighted vector addition
nz = N.nonzero(counts)[0]
weight = counts[nz]
i = N.sum(weight * N.exp(1j*bins[nz])/len(nz))
pave = cmath.phase(i) # correct and approximated mean for periodic data
i /= sum(weight)/float(len(nz))
pstdev = -2. * N.log(N.sqrt(i.real**2 + i.imag**2))
print
print 'scipy: %12.3f (mean) %12.3f (stdev)' % (ST.circmean(data)/deg, \
ST.circstd(data)/deg)
When run, it gives following results:
mean: 175.840 85.843 175.360
stdev: 0.472 151.785 0.430
scipy: 175.840 (mean) 3.673 (stdev)
A few comments now: the first column gives mean/stdev calculated. As can be seen, the mean agrees well with scipy.stats.circmean (thanks JoeKington for pointing it out). Unfortunately stdev differs. I will look at it later. The second column gives completely wrong results (non-periodic mean/std from numpy obviously does not work here). The 3rd column gives sth I wanted to obtain from the histogram data (#JoeKington: my raw data won't fit memory of my computer.., #dmytro: thanks for your input: of course, bin size will influence result but in my application I don't have much choice, i.e. I have to reduce data somehow). As can be seen, the mean (3rd column) is properly calculated, stdev needs further attention :)

Have a look at scipy.stats.circmean and scipy.stats.circstd.
Or do you only have the histogram counts, and not the "raw" data? If so, you could fit a Von Mises distribution to your histogram counts and approximate the mean and stddev in that way.

Here's how to get an approximation.
Since Var(x) = <x^2> - <x>^2, we have:
meanX = N.sum(counts * bins[:-1]) / N.sum(counts)
meanX2 = N.sum(counts * bins[:-1]**2) / N.sum(counts)
std = N.sqrt(meanX2 - meanX**2)