# This program is public domain
"""
Error propogation algorithms for simple arithmetic
Warning: like the underlying numpy library, the inplace operations
may return values of the wrong type if some of the arguments are
integers, so be sure to create them with floating point inputs.
"""
from __future__ import division # Get true division
import numpy
[docs]def div(X, varX, Y, varY):
"""Division with error propagation"""
# Direct algorithm:
# Z = X/Y
# varZ = (varX/X**2 + varY/Y**2) * Z**2
# = (varX + varY * Z**2) / Y**2
# Indirect algorithm to minimize intermediates
Z = X/Y # truediv => Z is a float
varZ = Z**2 # Z is a float => varZ is a float
varZ *= varY
varZ += varX
T = Y**2 # Doesn't matter if T is float or int
varZ /= T
return Z, varZ
[docs]def mul(X, varX, Y, varY):
"""Multiplication with error propagation"""
# Direct algorithm:
Z = X * Y
varZ = Y**2 * varX + X**2 * varY
# Indirect algorithm won't ensure floating point results
# varZ = Y**2
# varZ *= varX
# Z = X**2 # Using Z to hold the temporary
# Z *= varY
# varZ += Z
# Z[:] = X
# Z *= Y
return Z, varZ
[docs]def sub(X, varX, Y, varY):
"""Subtraction with error propagation"""
Z = X - Y
varZ = varX + varY
return Z, varZ
[docs]def add(X, varX, Y, varY):
"""Addition with error propagation"""
Z = X + Y
varZ = varX + varY
return Z, varZ
[docs]def exp(X, varX):
"""Exponentiation with error propagation"""
Z = numpy.exp(X)
varZ = varX * Z**2
return Z, varZ
[docs]def log(X, varX):
"""Logarithm with error propagation"""
Z = numpy.log(X)
varZ = varX / X**2
return Z, varZ
# Confirm this formula before using it
# def pow(X,varX, Y,varY):
# Z = X**Y
# varZ = (Y**2 * varX/X**2 + varY * numpy.log(X)**2) * Z**2
# return Z,varZ
#
[docs]def pow(X, varX, n):
"""X**n with error propagation"""
# Direct algorithm
# Z = X**n
# varZ = n*n * varX/X**2 * Z**2
# Indirect algorithm to minimize intermediates
Z = X**n
varZ = varX / X
varZ /= X
varZ *= Z
varZ *= Z
varZ *= n**2
return Z, varZ
[docs]def div_inplace(X, varX, Y, varY):
"""In-place division with error propagation"""
# Z = X/Y
# varZ = (varX + varY * (X/Y)**2) / Y**2 = (varX + varY * Z**2) / Y**2
X /= Y # X now has Z = X/Y
T = X**2 # create T with Z**2
T *= varY # T now has varY * Z**2
varX += T # varX now has varX + varY*Z**2
del T # may want to use T[:] = Y for vectors
T = Y # reuse T for Y
T **= 2 # T now has Y**2
varX /= T # varX now has varZ
return X, varX
[docs]def mul_inplace(X, varX, Y, varY):
"""In-place multiplication with error propagation"""
# Z = X * Y
# varZ = Y**2 * varX + X**2 * varY
T = Y**2 # create T with Y**2
varX *= T # varX now has Y**2 * varX
del T # may want to use T[:] = X for vectors
T = X # reuse T for X**2 * varY
T **=2 # T now has X**2
T *= varY # T now has X**2 * varY
varX += T # varX now has varZ
X *= Y # X now has Z
return X, varX
[docs]def sub_inplace(X, varX, Y, varY):
"""In-place subtraction with error propagation"""
# Z = X - Y
# varZ = varX + varY
X -= Y
varX += varY
return X, varX
[docs]def add_inplace(X, varX, Y, varY):
"""In-place addition with error propagation"""
# Z = X + Y
# varZ = varX + varY
X += Y
varX += varY
return X, varX
[docs]def pow_inplace(X, varX, n):
"""In-place X**n with error propagation"""
# Direct algorithm
# Z = X**n
# varZ = abs(n) * varX/X**2 * Z**2
# Indirect algorithm to minimize intermediates
varX /= X
varX /= X # varX now has varX/X**2
X **= n # X now has Z = X**n
varX *= X
varX *= X # varX now has varX/X**2 * Z**2
varX *= n**2 # varX now has varZ
return X, varX