def get_Erad(self,beta,zs=None,Nterms=None,interpolation='linear'):
if hasattr(beta,'__len__'):
if isinstance(beta,numpy.ndarray):
if not beta.ndim: beta=beta.tolist()
else: beta=numpy.array(beta)
if isinstance(beta,numpy.ndarray): beta=beta.reshape((len(beta),1,1))
if zs is None: zs=self.zs
Rs=self.evaluate_residues(zs,Nterms)#,interpolation)
Ps=self.evaluate_poles(zs,Nterms)#,interpolation)
#Should broadcast over freqs if beta has an additional first axis
#The offset term is absolutely critical, offsets false z-dependence arising from first terms
approach=numpy.sum(Rs*(1/(beta-Ps)+1/Ps),axis=-1)
axes=[zs]; axis_names=['z/a']
if hasattr(beta,'__len__'):
approach=approach.transpose()
if isinstance(beta,AWA):
axes=axes+[beta.axes[0]]
axis_names=axis_names+[beta.axis_names[0]]
else:
axes=axes+[None]
axis_names=axis_names+[None]
signals=AWA(approach,axes=axes,axis_names=axis_names).squeeze()
if not signals.ndim: signals=signals.tolist()
return signals
def get_signal_from_nodes(self,beta,nodes=[(0,1)],Nterms=None,interpolation='linear'):
if not hasattr(beta,'__len__'): beta=[beta]
if not isinstance(beta,AWA): beta=AWA(beta)
#`Frequency` axis will be first
if isinstance(beta,numpy.ndarray): beta=beta.reshape((len(beta),1,1))
#Weights apply across z-values
zs,ws=list(zip(*nodes))
ws_grid=numpy.array(ws).reshape((1,len(ws),1))
zs=numpy.array(zs)
#Evaluate at all nodal points
Rs=self.evaluate_residues(zs,Nterms,interpolation)
Ps=self.evaluate_poles(zs,Nterms,interpolation)
#Should broadcast over freqs if beta has an additional first axis
#The offset term is absolutely critical, offsets false z-dependence arising from first terms
approach=numpy.sum(numpy.sum(Rs*(1/(beta-Ps)+1/Ps)*ws_grid,axis=-1),axis=-1)#+Rs/Ps,axis=-1)
axes=[zs]; axis_names=['z/a']
if hasattr(beta,'__len__'):
approach=approach.transpose()
if isinstance(beta,AWA):
axes=axes+[beta.axes[0]]
axis_names=axis_names+[beta.axis_names[0]]
else:
axes=axes+[None]
axis_names=axis_names+[None]
signals=AWA(approach); signals.adopt_axes(beta)
signals=signals.squeeze()
if not signals.ndim: signals=signals.tolist()
return signals
def invert_signal(self,signals,nodes=[(1,0)],Nterms=10,\
interpolation='linear',\
select_by='continuity',\
closest_pole=0,\
scaling=10):
"""The inversion is not unique, consequently the selected solution
will probably be wrong if signal values correspond with
"beta" values that are too large (`|beta|~>min{|Poles|}`).
This can be expected to break at around `|beta|>2`."""
#Default is to invert signal in contact
#~10 terms seem required to converge on e.g. SiO2 spectrum,
#especially on the Re(beta)<0 (low signal) side of phonons
global roots,poly,root_scaling
#global betas,all_roots,pmin,rs,ps,As,Bs,roots,to_minimize
if self.verbose:
Logger.write('Inverting `signals` based on the provided `nodes` to obtain consistent beta values...')
if not hasattr(signals,'__len__'): signals=[signals]
if not isinstance(signals,AWA): signals=AWA(signals)
zs,ws=list(zip(*nodes))
ws_grid=numpy.array(ws).reshape((len(ws),1)) #last dimension is to broadcast over all `Nterms` equally
zs=numpy.array(zs)
Rs=self.Rs[:,:Nterms].interpolate_axis(zs,axis=0,kind=interpolation,
bounds_error=False,extrapolate=True)
Ps=self.Ps[:,:Nterms].interpolate_axis(zs,axis=0,kind=interpolation,
bounds_error=False,extrapolate=True)
#`rs` and `ps` can safely remain as arrays for `invres`
rs=(Rs*ws_grid).flatten()
ps=Ps.flatten()
k0=numpy.sum(rs/ps).tolist()
#Rescale units so their order of magnitude centers around 1
rscaling=numpy.exp(-(numpy.log(numpy.abs(rs).max())+\
numpy.log(numpy.abs(rs).min()))/2.)
pscaling=numpy.exp(-(numpy.log(numpy.abs(ps).max())+\
numpy.log(numpy.abs(ps).min()))/2.)
root_scaling=1/pscaling
#rscaling=1
#pscaling=1
if self.verbose:
Logger.write('\tScaling residues by a factor %1.2e to reduce floating point overflow...'%rscaling)
Logger.write('\tScaling poles by a factor %1.2e to reduce floating point overflow...'%pscaling)
rs*=rscaling; ps*=pscaling
k0*=rscaling/pscaling
signals=signals*rscaling/pscaling
#highest order first in `Ps` and `Qs`
#VERY SLOW - about 100ms on practical inversions (~60 terms)
As,Bs=invres(rs, ps, k=[k0], tol=1e-16, rtype='avg') #tol=1e-16 is the smallest allowable to `unique_roots`..
dtype=numpy.complex128 #Double precision offers noticeable protection against overflow
As=numpy.array(As,dtype=dtype)
Bs=numpy.array(Bs,dtype=dtype)
signals=signals.astype(dtype)
#import time
betas=[]
for i,signal in enumerate(signals):
#t1=time.time()
#Root finding `roots` seems to give noisy results when `Bs` has degree >84, with dynamic range ~1e+/-30 in coefficients...
#Pretty fast - 5-9 ms on practical inversions with rank ~60 companion matrices, <1 ms with ~36 terms
#@TODO: Root finding chokes on `Nterms=9` (number of eigenfields) and `Nts=12` (number of nodes),
# necessary for truly converged S3 on resonant phonons, probably due to
# floating point overflow - leading term increases exponentially with
# number of terms, leading to huge dynamic range.
# Perhaps limited by the double precision of DGEEV.
# So, replace with faster / more reliable root finder?
# We need 1) speed, 2) ALL roots (or at least the first ~10 smallest)
poly=As-signal*Bs
roots=find_roots(poly,scaling=scaling)
roots=roots[roots.imag>0]
roots*=root_scaling #since all beta units scaled by `pscaling`, undo that here
#print time.time()-t1
#How should we select the most likely beta among the multiple solutions?
#1. Avoids large changes in value of beta
if select_by=='difference' and i>=1:
if i==1 and self.verbose:
Logger.write('\tSelecting remaining roots by minimizing differences with prior...')
to_minimize=numpy.abs(roots-betas[i-1])
#2. Avoids large changes in slope of beta (best for spectroscopy)
#Nearly guarantees good beta spectrum, with exception of very loosely sampled SiC spectrum
#Loosely samples SiO2-magnitude phonons still perfectly fine
elif select_by=='continuity' and i>=2:
if i==2 and self.verbose:
Logger.write('\tSelecting remaining roots by ensuring continuity with prior...')
earlier_diff=betas[i-1]-betas[i-2]
current_diffs=roots-betas[i-1]
to_minimize=numpy.abs(current_diffs-earlier_diff)
#3. Select specifically which pole we want |beta| to be closest to
else:
if i==0 and self.verbose:
Logger.write('\tSeeding inversion closest to pole %i...'%closest_pole)
reordering=numpy.argsort(numpy.abs(roots)) #Order the roots towards increasing beta
roots=roots[reordering]
to_minimize=numpy.abs(closest_pole-numpy.arange(len(roots)))
beta=roots[to_minimize==to_minimize.min()].squeeze()
betas.append(beta)
if not i%5 and self.verbose:
Logger.write('\tProgress: %1.2f%% - Inverted %i signals of %i.'%\
(((i+1)/numpy.float(len(signals))*100),\
(i+1),len(signals)))
betas=AWA(betas); betas.adopt_axes(signals)
betas=betas.squeeze()
if not betas.ndim: betas=betas.tolist()
return betas
