def __init__(self, log_pdf, xlo, xhi, nx, ylo, yhi, ny, xstep='linear',
ystep='linear'):
self.log_pdf = log_pdf
self.xlo, self.xhi = xlo, xhi
self.nx = nx
if xstep == 'linear' or xstep=='lin':
self.xlinstep = True
self.xvals = array( [x for x in lin_stepper(xlo, xhi, self.nx)] )
self.dx = (xhi-xlo)/(self.nx-1)
elif xstep=='log' or xstep=='logarithmic':
self.xlinstep = False
self.xvals = array( [x for x in log_stepper(xlo, xhi, self.nx)] )
self.dx = log(xhi/xlo)/(self.nx-1)
else:
raise ValueError, 'Invalid xstep type!'
self.ylo, self.yhi = ylo, yhi
self.ny = ny
if ystep == 'linear' or ystep=='lin':
self.ylinstep = True
self.yvals = array( [y for y in lin_stepper(ylo, yhi, self.ny)] )
self.dy = (yhi-ylo)/(self.ny-1)
elif ystep=='log' or ystep=='logarithmic':
self.ylinstep = False
self.yvals = array( [y for y in log_stepper(ylo, yhi, self.ny)] )
self.dy = log(yhi/ylo)/(self.ny-1)
else:
raise ValueError, 'Invalid ystep type!'
logpdf = zeros((nx, ny), float)
for i,x in enumerate(self.xvals):
for j,y in enumerate(self.yvals):
logpdf[i,j] = log_pdf(x,y)
self.logpdf = logpdf
self.max = logpdf.max()
self.spdf = exp(logpdf-self.max) # Scaled PDF
# For log steps, change variables to log(x or y).
if not self.xlinstep:
for i, x in enumerate(self.xvals):
self.spdf[i,:] = self.spdf[i,:]*x
if not self.ylinstep:
for j, y in enumerate(self.yvals):
self.spdf[:,j] = self.spdf[:,j]*y
# Find the overall normalization for spdf.
self.norm = qgt2d(self.spdf, self.spdf.min())*self.dx*self.dy
# lml is the log marginal likelihood if logpdf = prior*like.
self.lml = log(self.norm) + self.max
self.probs = []
self.deltas = []
self.levels = []
def fracgt(self, logratio):
"""
The fraction of the posterior with log density > maximum + logratio.
"""
ratio = exp(logratio)
return qgt2d(self.spdf, ratio)*self.dx*self.dy/self.norm
