Source code for multipletau.core

#!/usr/bin/python
# -*- coding: utf-8 -*-
"""
A multiple-τ algorithm for Python 2.7 and 3.x.

Copyright (c) 2014 Paul Müller

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:

  1. Redistributions of source code must retain the above copyright
     notice, this list of conditions and the following disclaimer.

  2. Redistributions in binary form must reproduce the above copyright
     notice, this list of conditions and the following disclaimer in
     the documentation and/or other materials provided with the
     distribution.

  3. Neither the name of multipletau nor the names of its contributors
     may be used to endorse or promote products derived from this
     software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INFRAE OR
CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"""
from __future__ import division

import numpy as np
import warnings

__all__ = ["autocorrelate", "correlate", "correlate_numpy"]


#: Defines the cutoff when the absolute mean trace divided by the median
#: of the absolute value of the trace is treated as zero.
ZERO_CUTOFF = 1e-15


class DtypeWarning(UserWarning):
    pass


class InvalidMWarning(UserWarning):
    pass


[docs]def autocorrelate(a, m=16, deltat=1, normalize=False, copy=True, dtype=None, compress="average", ret_sum=False): """ Autocorrelation of a 1-dimensional sequence on a log2-scale. This computes the correlation similar to :py:func:`numpy.correlate` for positive :math:`k` on a base 2 logarithmic scale. :func:`numpy.correlate(a, a, mode="full")[len(a)-1:]` :math:`z_k = \\Sigma_n a_n a_{n+k}` Parameters ---------- a: array-like input sequence m: even integer defines the number of points on one level, must be an even integer deltat: float distance between bins normalize: bool normalize the result to the square of the average input signal and the factor :math:`M-k`. copy: bool copy input array, set to ``False`` to save memory dtype: object to be converted to a data type object The data type of the returned array and of the accumulator for the multiple-tau computation. compress: str strategy for propagating values to the next register - `"average"` (default): average two measurements when pushing to the next level of the correlator. - `"first"`: use only the first value when pushing to the next level of the correlator. - `"second"`: use only the second value when pushing to the next level of the correlator. Using only the first or the second values during propagation completely removes the systematic error at the cost of increasing the statistical error. See https://doi.org/10.1063/1.3491098 for a discussion on the effect of averaging. ret_sum: bool return the exact sum :math:`z_k = \\Sigma_n a_n a_{n+k}`. In addition :math:`M-k` is returned as an array of length N. Returns ------- autocorrelation: ndarray of shape (N,2) the lag time (1st column) and the autocorrelation (2nd column). count: ndarray of length N only returned if `ret_sum` is True; the value of :math:`M-k` for each row in `autocorrelation`. Notes ----- .. versionchanged:: 0.1.6 Compute the correlation for zero lag time. The algorithm computes the correlation with the convention of the curve decaying to zero. For experiments like e.g. fluorescence correlation spectroscopy, the signal can be normalized to :math:`M-k` by invoking ``normalize=True``. For normalizing according to the behavior of :py:func:`numpy.correlate`, use ``normalize=False``. For complex arrays, this method falls back to the method :func:`correlate`. Examples -------- >>> from multipletau import autocorrelate >>> autocorrelate(range(42), m=2, dtype=np.float_) array([[ 0.00000000e+00, 2.38210000e+04], [ 1.00000000e+00, 2.29600000e+04], [ 2.00000000e+00, 2.21000000e+04], [ 4.00000000e+00, 2.03775000e+04], [ 8.00000000e+00, 1.50612000e+04]]) """ if not isinstance(normalize, bool): raise ValueError("`normalize` must be boolean!") if not isinstance(copy, bool): raise ValueError("`copy` must be boolean!") if not isinstance(ret_sum, bool): raise ValueError("`ret_sum` must be boolean!") if normalize and ret_sum: raise ValueError("'normalize' and 'ret_sum' must not both be True!") compress_values = ["average", "first", "second"] if compress not in compress_values: raise ValueError("Invalid value for `compress`! Possible values " "are '{}'.".format(','.join(compress_values))) if dtype is None: dtype = np.dtype(a[0].__class__) else: dtype = np.dtype(dtype) # Complex data if dtype.kind == "c": # run cross-correlation return correlate(a=a, v=a, m=m, deltat=deltat, normalize=normalize, copy=copy, dtype=dtype) elif dtype.kind != "f": warnings.warn("Input dtype is not float; casting to np.float_!", DtypeWarning) dtype = np.dtype(np.float_) # If copy is false and dtype is the same as the input array, # then this line does not have an effect: trace = np.array(a, dtype=dtype, copy=copy) # Check parameters if m // 2 != m / 2: mold = m m = np.int_((m // 2 + 1) * 2) warnings.warn("Invalid value of m={}. Using m={} instead" .format(mold, m), InvalidMWarning) else: m = np.int_(m) N = N0 = trace.shape[0] # Find out the length of the correlation function. # The integer k defines how many times we can average over # two neighboring array elements in order to obtain an array of # length just larger than m. k = np.int_(np.floor(np.log2(N / m))) # In the base2 multiple-tau scheme, the length of the correlation # array is (only taking into account values that are computed from # traces that are just larger than m): lenG = m + k * (m // 2) + 1 G = np.zeros((lenG, 2), dtype=dtype) normstat = np.zeros(lenG, dtype=dtype) normnump = np.zeros(lenG, dtype=dtype) traceavg = np.average(trace) # We use the fluctuation of the signal around the mean if normalize: trace -= traceavg if np.abs(traceavg) / np.median(np.abs(trace)) < ZERO_CUTOFF: raise ValueError("Cannot normalize: Average of `a` is zero!") # Otherwise the following for-loop will fail: if N < 2 * m: raise ValueError("`len(a)` must be >= `2m`!") # Calculate autocorrelation function for first m+1 bins # Discrete convolution of m elements for n in range(0, m + 1): G[n, 0] = deltat * n # This is the computationally intensive step G[n, 1] = np.sum(trace[:N - n] * trace[n:]) normstat[n] = N - n normnump[n] = N # Now that we calculated the first m elements of G, let us # go on with the next m/2 elements. # Check if len(trace) is even: if N % 2 == 1: N -= 1 # compress every second element if compress == compress_values[0]: trace = (trace[:N:2] + trace[1:N:2]) / 2 elif compress == compress_values[1]: trace = trace[:N:2] elif compress == compress_values[2]: trace = trace[1:N:2] N //= 2 # Start iteration for each m/2 values for step in range(1, k + 1): # Get the next m/2 values via correlation of the trace for n in range(1, m // 2 + 1): npmd2 = n + m // 2 idx = m + n + (step - 1) * m // 2 if len(trace[:N - npmd2]) == 0: # This is a shortcut that stops the iteration once the # length of the trace is too small to compute a corre- # lation. The actual length of the correlation function # does not only depend on k - We also must be able to # perform the sum with respect to k for all elements. # For small N, the sum over zero elements would be # computed here. # # One could make this for-loop go up to maxval, where # maxval1 = int(m/2) # maxval2 = int(N-m/2-1) # maxval = min(maxval1, maxval2) # However, we then would also need to find out which # element in G is the last element... G = G[:idx - 1] normstat = normstat[:idx - 1] normnump = normnump[:idx - 1] # Note that this break only breaks out of the current # for loop. However, we are already in the last loop # of the step-for-loop. That is because we calculated # k in advance. break else: G[idx, 0] = deltat * npmd2 * 2**step # This is the computationally intensive step G[idx, 1] = np.sum(trace[:N - npmd2] * trace[npmd2:]) normstat[idx] = N - npmd2 normnump[idx] = N # Check if len(trace) is even: if N % 2 == 1: N -= 1 # compress every second element if compress == compress_values[0]: trace = (trace[:N:2] + trace[1:N:2]) / 2 elif compress == compress_values[1]: trace = trace[:N:2] elif compress == compress_values[2]: trace = trace[1:N:2] N //= 2 if normalize: G[:, 1] /= traceavg**2 * normstat elif not ret_sum: G[:, 1] *= N0 / normnump if ret_sum: return G, normstat else: return G
[docs]def correlate(a, v, m=16, deltat=1, normalize=False, copy=True, dtype=None, compress="average", ret_sum=False): """ Cross-correlation of two 1-dimensional sequences on a log2-scale. This computes the cross-correlation similar to :py:func:`numpy.correlate` for positive :math:`k` on a base 2 logarithmic scale. :func:`numpy.correlate(a, v, mode="full")[len(a)-1:]` :math:`z_k = \\Sigma_n a_n v_{n+k}` Note that only the correlation in the positive direction is computed. To obtain the correlation for negative lag times swap the input variables ``a`` and ``v``. Parameters ---------- a, v: array-like input sequences with equal length m: even integer defines the number of points on one level, must be an even integer deltat: float distance between bins normalize: bool normalize the result to the square of the average input signal and the factor :math:`M-k`. copy: bool copy input array, set to ``False`` to save memory dtype: object to be converted to a data type object The data type of the returned array and of the accumulator for the multiple-tau computation. compress: str strategy for propagating values to the next register - `"average"` (default): average two measurements when pushing to the next level of the correlator. - `"first"`: use only the first value when pushing to the next level of the correlator. - `"second"`: use only the second value when pushing to the next level of the correlator. Using only the first or the second values during propagation completely removes the systematic error at the cost of increasing the statistical error. See https://doi.org/10.1063/1.3491098 for a discussion on the effect of averaging. ret_sum: bool return the exact sum :math:`z_k = \\Sigma_n a_n v_{n+k}`. In addition :math:`M-k` is returned as an array of length N. Returns ------- cross_correlation: ndarray of shape (N,2) the lag time (1st column), the cross-correlation (2nd column). count: ndarray of length N only returned if `ret_sum` is True; the value of :math:`M-k` for each row in `autocorrelation`. Notes ----- .. versionchanged:: 0.1.6 Compute the correlation for zero lag time and correctly normalize the correlation for a complex input sequence `v`. The algorithm computes the correlation with the convention of the curve decaying to zero. For experiments like e.g. fluorescence correlation spectroscopy, the signal can be normalized to :math:`M-k` by invoking ``normalize=True``. For normalizing according to the behavior of :py:func:`numpy.correlate`, use ``normalize=False``. Examples -------- >>> from multipletau import correlate >>> correlate(range(42), range(1,43), m=2, dtype=np.float_) array([[ 0.00000000e+00, 2.46820000e+04], [ 1.00000000e+00, 2.38210000e+04], [ 2.00000000e+00, 2.29600000e+04], [ 4.00000000e+00, 2.12325000e+04], [ 8.00000000e+00, 1.58508000e+04]]) """ if not isinstance(normalize, bool): raise ValueError("`normalize` must be boolean!") if not isinstance(copy, bool): raise ValueError("`copy` must be boolean!") if not isinstance(ret_sum, bool): raise ValueError("`ret_sum` must be boolean!") if normalize and ret_sum: raise ValueError("'normalize' and 'ret_sum' must not both be True!") compress_values = ["average", "first", "second"] if compress not in compress_values: raise ValueError("Invalid value for `compress`! Possible values " "are '{}'.".format(','.join(compress_values))) # See `autocorrelation` for better documented code. traceavg1 = np.average(v) traceavg2 = np.average(a) if normalize: if np.abs(traceavg1) / np.median(np.abs(v)) < ZERO_CUTOFF: raise ValueError("Cannot normalize: Average of `v` is zero!") if np.abs(traceavg2) / np.median(np.abs(a)) < ZERO_CUTOFF: raise ValueError("Cannot normalize: Average of `a` is zero!") if dtype is None: dtype = np.dtype(v[0].__class__) dtype2 = np.dtype(a[0].__class__) if dtype != dtype2: if dtype.kind == "c" or dtype2.kind == "c": # The user might try to combine complex64 and float128. warnings.warn( "Input dtypes not equal; casting to np.complex_!", InvalidMWarning) dtype = np.dtype(np.complex_) else: warnings.warn("Input dtypes not equal; casting to np.float_!", InvalidMWarning) dtype = np.dtype(np.float_) else: dtype = np.dtype(dtype) if dtype.kind not in ["c", "f"]: warnings.warn("Input dtype is not float; casting to np.float_!", InvalidMWarning) dtype = np.dtype(np.float_) trace1 = np.array(v, dtype=dtype, copy=copy) # Prevent traces from overwriting each other if a is v: # Force copying trace 2 copy = True trace2 = np.array(a, dtype=dtype, copy=copy) if trace1.size != trace2.size: raise ValueError("`a` and `v` must have same length!") # Complex data if dtype.kind == "c": np.conjugate(trace1, out=trace1) # Check parameters if m // 2 != m / 2: mold = m m = np.int_(m // 2 + 1) * 2 warnings.warn("Invalid value of m={}. Using m={} instead" .format(mold, m), InvalidMWarning) else: m = np.int_(m) N = N0 = trace1.shape[0] # Find out the length of the correlation function. # The integer k defines how many times we can average over # two neighboring array elements in order to obtain an array of # length just larger than m. k = np.int_(np.floor(np.log2(N / m))) # In the base2 multiple-tau scheme, the length of the correlation # array is (only taking into account values that are computed from # traces that are just larger than m): lenG = m + k * m // 2 + 1 G = np.zeros((lenG, 2), dtype=dtype) normstat = np.zeros(lenG, dtype=dtype) normnump = np.zeros(lenG, dtype=dtype) # We use the fluctuation of the signal around the mean if normalize: trace1 -= np.conj(traceavg1) trace2 -= traceavg2 # Otherwise the following for-loop will fail: if N < 2 * m: raise ValueError("`len(a)` must be >= `2m`!") # Calculate autocorrelation function for first m+1 bins for n in range(0, m + 1): G[n, 0] = deltat * n G[n, 1] = np.sum(trace1[:N - n] * trace2[n:]) normstat[n] = N - n normnump[n] = N # Check if len(trace) is even: if N % 2 == 1: N -= 1 # compress every second element if compress == compress_values[0]: trace1 = (trace1[:N:2] + trace1[1:N:2]) / 2 trace2 = (trace2[:N:2] + trace2[1:N:2]) / 2 elif compress == compress_values[1]: trace1 = trace1[:N:2] trace2 = trace2[:N:2] elif compress == compress_values[2]: trace1 = trace1[1:N:2] trace2 = trace2[1:N:2] N //= 2 for step in range(1, k + 1): # Get the next m/2 values of the trace for n in range(1, m // 2 + 1): npmd2 = (n + m // 2) idx = m + n + (step - 1) * m // 2 if len(trace1[:N - npmd2]) == 0: # Abort G = G[:idx - 1] normstat = normstat[:idx - 1] normnump = normnump[:idx - 1] break else: G[idx, 0] = deltat * npmd2 * 2**step G[idx, 1] = np.sum( trace1[:N - npmd2] * trace2[npmd2:]) normstat[idx] = N - npmd2 normnump[idx] = N # Check if len(trace) is even: if N % 2 == 1: N -= 1 # compress every second element if compress == compress_values[0]: trace1 = (trace1[:N:2] + trace1[1:N:2]) / 2 trace2 = (trace2[:N:2] + trace2[1:N:2]) / 2 elif compress == compress_values[1]: trace1 = trace1[:N:2] trace2 = trace2[:N:2] elif compress == compress_values[2]: trace1 = trace1[1:N:2] trace2 = trace2[1:N:2] N //= 2 if normalize: G[:, 1] /= traceavg1 * traceavg2 * normstat elif not ret_sum: G[:, 1] *= N0 / normnump if ret_sum: return G, normstat else: return G
[docs]def correlate_numpy(a, v, deltat=1, normalize=False, dtype=None, copy=True): """ Convenience function that wraps around :py:func:`numpy.correlate` and returns the correlation in the same format as :func:`correlate` does. Parameters ---------- a, v: array-like input sequences deltat: float distance between bins normalize: bool normalize the result to the square of the average input signal and the factor :math:`M-k`. The resulting curve follows the convention of decaying to zero for large lag times. copy: bool copy input array, set to ``False`` to save memory dtype: object to be converted to a data type object The data type of the returned array. Returns ------- cross_correlation: ndarray of shape (N,2) the lag time (column 1) and the cross-correlation (column 2). Notes ----- .. versionchanged:: 0.1.6 Removed false normalization when `normalize==False`. """ ab = np.array(a, dtype=dtype, copy=copy) vb = np.array(v, dtype=dtype, copy=copy) if ab.size != vb.size: raise ValueError("`a` and `v` must have same length!") avg = np.average(ab) vvg = np.average(vb) if normalize: ab -= avg vb -= vvg if np.abs(avg) / np.median(np.abs(ab)) < ZERO_CUTOFF: raise ValueError("Cannot normalize: Average of `a` is zero!") if np.abs(vvg) / np.median(np.abs(vb)) < ZERO_CUTOFF: raise ValueError("Cannot normalize: Average of `v` is zero!") Gd = np.correlate(ab, vb, mode="full")[len(ab) - 1:] if normalize: N = len(Gd) m = N - np.arange(N) Gd /= m * avg * vvg G = np.zeros((len(Gd), 2), dtype=dtype) G[:, 1] = Gd G[:, 0] = np.arange(len(Gd)) * deltat return G