harminv - extract mode frequencies from time-series data
harminv [OPTION]... [freq-min-freq-max]...
harminv is a program designed to solve the problem of "harmonic
inversion": given a time series consisting of a sum of sinusoids
("modes"), extract their frequencies and amplitudes. It can also
handle the case of exponentially-decaying sinusoids, in which case it
extracts their decay rates as well.
harminv is often able to achieve much greater accuracy and robustness
than Fourier-transform methods, essentially because it assumes a
specific form for the input.
It uses a low-storage "filter-diagonalization method" (FDM), as
described in V. A. Mandelshtam and H. S. Taylor, "Harmonic inversion of
time signals," J. Chem. Phys. 107, 6756 (1997). See also erratum, ibid
109, 4128 (1998).
harminv reads in a sequence of whitespace-separated real or complex
numbers from standard input, as well as command-line arguments
indicating one or more frequency ranges to search, and outputs the
modes that it extracts from the data. (It preferentially finds modes
in the frequency range you specify, but may sometimes find additional
modes outside of that range.) The data should correspond to equally-
spaced time intervals, but there is no constraint on the number of
Complex numbers in the input should be expressed in the format RE+IMi
(no whitespace). Otherwise, whitespace is ignored. Also, comments
beginning with "#" and extending to the end of the line are ignored.
A typical invocation is something like
harminv -t 0.02 1-5 < input.dat
which reads a sequence of samples, spaced at 0.02 time intervals (in
ms, say, corresponding to 50 kHz), and searches for modes in the
frequency range 1-5 kHz. (See below on units.)
harminv writes six comma-delimited columns to standard output, one line
for each mode: frequency, decay constant, Q, amplitude, phase, and
error. Each mode corresponds to a function of the form:
amplitude * exp[-i (2 pi frequency t - phase) - decay t]
Here, i is sqrt(-1), t is the time (see below for units), and the other
parameters in the output columns are:
The frequency of the mode. If you don’t recognize that from the
expression above, you should recall Euler’s formula: exp(i x) =
cos(x) + i sin(x). Note that for complex data, there is a
distinction between positive and negative frequencies.
The exponential decay constant, indicated by decay in the above
formula. The inverse of this is often called the "lifetime" of
the mode. The "half-life" is ln(2)/decay.
Q A conventional, dimensionless expression of the decay lifetime:
Q = pi |frequency| / decay. Q, which stands for "quality
factor", is the number of periods for the "energy" in the mode
(the squared amplitude) to decay by exp(-2 pi). Equivalently,
if you look at the power spectrum (|Fourier transform|^2), 1/Q
is the fractional width of the peak at half maximum.
The (real, positive) amplitude of the sinusoids. The amplitude
(and phase) information generally seems to be less accurate than
the frequency and decay constant.
phase The phase shift (in radians) of the sinusoids, as given by the
error A crude estimate of the relative error in the (complex)
frequency. This is not really an error bar, however, so you
should treat it more as a figure of merit (smaller is better)
for each mode.
Typically, harminv will find a number of spurious solutions in addition
to the desired solutions, especially if your data are noisy. Such
solutions are characterized by large errors, small amplitudes, and/or
small Q (large decay rates / broad linewidths). You can omit these
from the output by the error/Q/amplitude screening options defined
By default, modes with error > 0.1 and Q < 10 are automatically
omitted, but it is likely that you will need to set stricter limits.
The frequency (and decay) values, both input and output, are specified
in units of 1/time, where the units of time are determined by the
sampling interval dt (the time between consecutive inputs). dt is by
default 1, unless you specify it with the -t dt option.
In other words, pick some units (e.g. ms in the example above) and use
them to express the time step. Then, be consistent and use the inverse
of those units (e.g. kHz = 1/ms) for frequency.
Note that the frequency is the usual 1/period definition; it is not the
-h Display help on the command-line options and usage.
-V Print the version number and copyright info for harminv.
-v Enable verbose output, printed to standard output as comment
lines (starting with a "#" character). Also, any "#" comments
in the input are echoed to the output.
-T Specify period-ranges instead of frequency-ranges on the command
line (in units of time corresponding to those specified by -t).
The output is still frequency and not period, however.
-w Specify angular frequencies instead of frequencies, and output
angular frequency instead of frequency. (Angular frequency is
frequency multiplied by 2 pi).
-n Flip the sign of the frequency (and phase) convention used in
harminv. (The sign of the frequency is only important if you
have complex-valued input data, in which case the positive and
negative frequency amplitudes can differ.)
-t dt Specify the sampling interval dt; this determines the units of
time used throughout the input and output. Defaults to 1.0.
-d d Specify the spectral "density" d to search for modes, where a
density of 1 indicates the usual Fourier resolution. That is,
the number of basis functions (which sets an upper bound on the
number of modes) is given by d times (freq-max - freq-min) times
dt times the number of samples in your dataset. A maximum of
300 is used, however, to prevent the matrices from getting too
big (you can force a larger number with -f, below).
Note that the frequency resolution of the outputs is not limited
by the spectral density, and can generally be much greater than
the Fourier resolution. The density determines how many modes,
at most, to search for, and in some sense is the density with
which the bandwidth is initially "searched" for modes.
The default density is 0.0, which means that the number of basis
functions is determined by -f (which defaults to 100). This
often corresponds to a much larger density than the usual
Fourier resolution, but the resulting singularities in the
system matrices are automatically removed by harminv.
-f nf Specify a lower bound nf on the number of spectral basis
functions (defaults to 100), setting a lower bound on the number
of modes to search for. This option is often a more convenient
way to specify the number of basis functions than the -d option,
above, which is why it is the default.
-f also allows you to employ more than 300 basis functions, but
careful: the computation time scales as O(N nf) + O(nf^3), where
N is the number of samples, and very large matrices can also
have degraded accuracy.
Specify how the outputs are sorted, where sort is one of
frequency/error/Q/decay/amplitude. (Only the first character of
sort matters.) All sorts are in ascending order. The default
is to sort by frequency.
-e err Omit any modes with error (see above) greater than err times the
largest error among the computed modes. Defaults to no limit.
-E err Omit any modes with error (see above) greater than err.
Defaults to 0.1.
-F Omit any modes with frequencies outside the specified range.
(Such modes are not necessarily spurious, however.)
-a amp Omit any modes with amplitude (see above) less than amp times
the largest amplitude among the computed modes. Defaults to no
-A amp Omit any modes with amplitude (see above) less than amp.
Defaults to no limit.
-Q q Omit any modes with |Q| (see above) less than q. Defaults to
Send bug reports to S. G. Johnson, firstname.lastname@example.org.
Written by Steven G. Johnson. Copyright (c) 2005 by the Massachusetts
Institute of Technology.