See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/11464026

Computation of a Spectrum from a Single-Beam Fourier-Transform Infrared

Interferogram

Article  in  Applied Optics · March 2002

DOI: 10.1364/AO.41.001181 · Source: PubMed

CITATIONS READS

26 606

2 authors:

Avishai Ben-David Agustin I Ifarraguerri

Edgewood Chemical Biological Center, Aberdeen Proving Ground, Maryland, Unit… SAIC
95 PUBLICATIONS   725 CITATIONS    8 PUBLICATIONS   37 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Avishai Ben-David on 15 March 2014.

The user has requested enhancement of the downloaded file.

Computation of a spectrum from a single-beam
Fourier-transform infrared interferogram

Avishai Ben-David and Agustin Ifarraguerri

A new high-accuracy method has been developed to transform asymmetric single-sided interferograms
into spectra. We used a fraction 共short, double-sided兲 of the recorded interferogram and applied an
iterative correction to the complete recorded interferogram for the linear part of the phase induced by the
various optical elements. Iterative phase correction enhanced the symmetry in the recorded interfero-
gram. We constructed a symmetric double-sided interferogram and followed the Mertz procedure 关In-
frared Phys. 7, 17 共1967兲兴 but with symmetric apodization windows and with a nonlinear phase correction
deduced from this double-sided interferogram. In comparing the solution spectrum with the source
spectrum we applied the Rayleigh resolution criterion with a Gaussian instrument line shape. The
accuracy of the solution is excellent, ranging from better than 0.1% for a blackbody spectrum to a few
percent for a complicated atmospheric radiance spectrum. © 2002 Optical Society of America
OCIS codes: 300.6300, 300.6340, 300.6190, 120.0280, 120.3180, 070.6020.

1. Introduction ferogram I共x兲 is the measured interference pattern

Much research has been devoted to the subject of
1– 6
converting interferograms measured with single-
beam Fourier-transform infrared 共FTIR兲 instruments
into spectra. FTIR spectrometers can measure si-
multaneously many of the thousands of gaseous
chemicals that have absorption features in the infra-
red and are used routinely in air-pollution monitoring
of toxic gases released by industrial processes and in
atmospheric research in which the chemical compo-
sition of the atmosphere is of interest. Two issues
that are important but are not well understood in
converting an interferogram into a spectrum are
phase-correction techniques and apodization. As
the signal-to-noise ratio in the Fourier spectroscopy
measurements increases, an accurate phase correc-
tion and high-quality apodization techniques become
more important and may set the limit on the achiev-
able accuracy of the deduced spectrum. The inter-
A. Ben-David 共avishai.bendavid@aphea.army.mil兲 is with the
Science and Technology Corporation, Suite 206, 500 Edgewood
Road, Edgewood, Maryland 21040. A. Ifarraguerri is with the
Science Applications International Corporation, Suite 300, 4001
Fairfax Drive, Arlington, Virginia 22203; while this research was
performed he was with the U.S. Army Edgewood Chemical Biolog-
ical Center, Aberdeen Proving Ground, Maryland 21010-5424.
Received 12 June 2001; revised manuscript received 14 Septem-
ber 2001.
0003-6935兾02兾061181-09$15.00兾0
© 2002 Optical Society of America
for a displacement x 共in centimeters兲 between two
interfering beams in a classic Michelson interferom-
eter.1 The recorded interferogram 共Fig. 1兲 is a se-
quence of measurements that span the range ⫺xmin
to ⫹xmax, where xmax is the maximum optical-path
difference and is determined by the wave-number
resolution 共inverse centimeters兲 requirement and
xmin depends on the instrument design. In Fig. 1,
only the modulated part 共ac兲 of the interferogram,
which contains information on the spectrum, is
shown, and the constant 共dc兲 component is omitted.
The recorded interferogram is in general asymmetric.
If xmax ⫽ xmin the interferogram is termed double-
sided. In some instruments xmax ⬎⬎ xmin, and the
interferogram is termed a single-sided asymmetric
interferogram. The maximum of the interference
pattern is located at an optical-path difference x0 and
is called the center burst. At this location all the
wavelengths of the light source that produced the
共ideal兲 interferogram are in phase, making this point
in the recorded interferogram a unique reference
point for phase-correction procedures. There are
two components of phase in the recorded interfero-
gram. The first is due to a phase shift in the incident
radiance that results from the optical components
such as beam splitter, lenses, and mirrors in the in-
strument, all of which have different wavelength-
dependent refractive indices and optical thickness,
and from the electronics 共e.g., amplifiers and filters兲,
which exhibit a frequency-dependent time delay.

20 February 2002 兾 Vol. 41, No. 6 兾 APPLIED OPTICS 1181


Fig. 1. Single-sided asymmetric interferogram I共⫺xmin ⱕ x ⱕ
xmax兲 computed for blackbody source B共500 ⬍ ␯ ⬍ 1500 cm⫺1兲 at a
temperature of 300 K. Short, double-sided interferogram Is共x兲 ⫽
I共⫺xmin ⱕ x ⱕ xmin兲 and the location of center burst x0 are shown.
Only the modulated part 共ac兲 of the interferogram that contains
information on the spectrum is shown, and the constant 共dc兲 com-
ponent is omitted.
The second part of the phase is due to the location of
the true center burst in the recorded measurement
sequence, which in general will fall between sample
points.
The two commonly used methods of converting an
interferogram into a spectrum are the Mertz7 and the
Forman8 methods 共see Ref. 1, Chap. 3, for a compre-
hensive discussion兲. In the Mertz method a phase
estimate of the optical components 共i.e., the first part
of the phase兲 is computed from a short region about
the center burst of the interferogram 共usually de-
noted a short double-sided interferogram; Fig. 1兲.
This phase estimate is deduced for only a limited
number of frequencies and is then interpolated to all
the frequencies that are contained within the mea-
sured interferogram. Then the recorded asymmet-
ric interferogram is apodized by an asymmetric
triangular window and rotated to zero out the phase
shift introduced by the sampled location of the inter-
ferogram. One converts the interferogram into a
spectrum by taking the Fourier transform; it is phase
corrected by use of the phase estimate deduced from
the short double-sided interferogram. This phase
correction in a sense transfers residual energy from
the imaginary frequency components of the Fourier
transform operation back to the real frequency com-
ponents of the spectrum. In the Forman method the
phase correction is performed in the time domain
共i.e., the interferogram domain兲, whereas in the
Mertz method the phase correction is done in the
frequency domain 共i.e., the spectrum domain兲. In
the Forman method no apodization is applied to the
short double-sided interferogram, and no interpola-
tion of the phase is performed. Chase2 stated that
the Forman method is better than the Mertz method
1182 APPLIED OPTICS 兾 Vol. 41, No. 6 兾 20 February 2002
because the triangular apodization 共used in the
Mertz method兲 has a discontinuity in its first deriv-
ative, and thus the phase estimate is less accurate,
and suggested that no fewer than 100 data points
共preferably several hundreds兲 should be used in the
phase estimate.
In this paper we present a method that combines
the Forman and Mertz methods of converting an in-
terferogram into a spectrum. In our method we use
the short double-sided interferogram to apply an it-
erative phase-correction procedure 共for the linear
part of the phase induced by the various optical ele-
ments and for the linear phase delay that is due to the
electronics兲 to an asymmetric single-sided recorded
interferogram. From the linearly phase-corrected
interferogram we construct a complete symmetric
double-sided interferogram for the range ⫺xmax to
⫹xmax. We then compute the Fourier transform, us-
ing any user-selected apodizing window 共e.g., Ham-
ming, Blackman, Chebyshev, and Kaiser兲. These
windows have characteristics 共e.g., lower sidelobes兲
that are superior to those of the triangular apodizing
window. A final phase correction for the nonlinear
part of the phase induced by the various optical ele-
ments is performed on the derived Fourier transform
to transfer energy from the imaginary frequency com-
ponents to the real frequency components, as is done
in the Mertz method. In our method the nonlinear
phase is deduced directly from the constructed sym-
metric double-sided interferogram, and thus no inter-
polation on the nonlinear phase is needed. We
discuss practical issues of numerical implementa-
tions, such as zero-fill interpolation and a deresolu-
tion algorithm for comparing the computed spectrum
with the source spectrum that produced the interfero-
gram, and show some examples of the method’s re-
sults.
2. Phase Correction of the Interferogram I(x)
The modulated 共ac兲 component of the recorded inter-
ferogram I共 x兲 共in volts兲 at n locations xi , ⫺xmin ⱕ x ⱕ
xmax uniformly spaced every ␦ centimeters that re-
sults from a source B共␯兲 关W兾共cm2 sr cm⫺1兲兴, where ␯ is
the wave number 共in inverse centimeters兲 is given1 by
兰
␯2
I共 x兲 ⫽ K共␯兲0.5B共␯兲cos关2␲␯x ⫹ ⌽共␯兲兴d␯, (1a)
␯1
⌽共␯兲 ⫽ ⌽ 1共␯兲 ⫹ ⌽ 2共␯兲, (1b)
⌽ 2共␯兲 ⫽ ⫺2␲␯x 0, (1c)
where K共␯兲 关共V cm2 sr兲兾W兴 is the instrument constant
that combines the responsivity 共volts per watt兲, the
aperture area 共square centimeters兲 and the field of
view 共steradians兲 of the instrument. For conve-
nience we assume that K共␯兲 ⫽ 1 for ␯2 ⬍ ␯ ⬍ ␯1. The
total phase ⌽共␯兲 is combined phase shift ⌽1共␯兲 from all
optical elements and electronics components and
phase shift ⌽2共␯兲 that is due to the shift x0 of the
center burst in the recorded interferogram. The in-
tegration limits ␯1 and ␯2 are for a band-limited
source B共␯兲 or are due to a detector band-limited
transfer function, and the modulation efficiency is
assumed to be 1. For an ideal spectrometer for
which ⌽共␯兲 ⫽ 0 and interferogram I共x兲 is sampled at
all x values, I共x兲 and B共␯兲 are a Fourier pair given by
兰
I共 x兲 ⫽ K共␯兲 B共␯兲cos共2␲x␯兲d␯ ⫽ F ⫺1关K共␯兲 B共␯兲兴,
B共␯兲 ⫽ K共␯兲 ⫺1
兰 I共 x兲cos共2␲x␯兲dx ⫽ K共␯兲 ⫺1F关I共 x兲兴,
(2)
where F共 兲 and F⫺1共 兲 denote the cosine Fourier-
transform operation and the inverse cosine Fourier
transform, respectively. In this research the diffi-
culty in implementing the straightforward Fourier
transform 关Eqs. 共2兲兴 because of phase ⌽共␯兲 and the
finite sampling range ⫺xmin ⱕ x ⱕ xmax are ad-
dressed.
We apply a phase correction to I共x兲 for the linear
part of phase shift ⌽1共␯兲 and to the portion of phase
2␲␯共x0 ⫺ 关 x0兴兲 in ⌽2共␯兲, where 关 x0兴 is the rounded
value of center-burst location x0 共关 x0兴 ⫽ k␦, where 1 ⬍
k ⬍ n is an integer sampling point number兲 in an
iterative procedure that usually takes only a few it-
erations. Because of the finite number of sampling
intervals, the center burst is not sampled exactly 共it
usually falls between sampling points兲. Phase cor-
rection 2␲␯共x0 ⫺ 关 x0兴兲 shifts the center burst to the
location 关 x0兴 that is one of the sampled xi locations.
Phase shift 2␲␯关 x0兴 is a phantom phase 共owing to the
arbitrary placement of the center-burst location in
the interferogram vector兲 that is easily handled with
a rotation process by rearrangement of the elements
of I共x兲 to place the center burst at the first element of
the vector. As our iterative procedure is a linear
procedure, only the linear part of phase ⌽1共␯兲 is cor-
rected. For a general function f 共x兲, phase ⌽f 共␯兲 is
given by the arctangent of the ratio between the
imaginary part and the real part of the Fourier trans-
form of f 共x兲. If f 共x兲 is real and symmetric about the
origin, phase ⌽f 共␯兲 is zero because the imaginary part
of a complex Fourier transform of a real and symmet-
ric function is zero. Because unknown source spec-
trum B共␯兲 is purely a real function, we would like to
transform the real interferogram I共x兲 to be symmetric
共i.e., an even function兲 about x0; then, shifting the x
axis 关i.e., forming the function I共x ⫺ x0兲兴 will yield a
phase that has become zero.
The general objective of our procedure is to esti-
mate with high accuracy the location of center burst
x0, using a short double-sided fraction of the recorded
interferogram, and impose a time shift 共i.e., a shift of
the x axis, by a resampling process兲 on the full re-
corded interferogram. When this is done, the sym-
metry of the interferogram tends to improve,
resulting in a smaller phase component. The con-
cept is applied iteratively; the correction improves
following each iteration.
A first estimate of center-burst location x̂0 is easily
found from the maximum of the absolute value of
I共 x兲, I共 x̂0兲 ⫽ max关兩I共 x兲兩兴. A short double-sided inter-
ferogram Is共x兲 of length 共2n0 ⫺ 1兲 is formed, with I共x̂0兲
at the center and 共n0 ⫺ 1兲 samples on each side 共n0 is
the location, point number i for x ⫽ x̂0兲. Short
double-sided interferogram Is共x兲 is in general not
symmetric about I共 x̂0兲, and this departure from sym-
metry is due to the phase shift. Average departure
⌬ 共in centimeters兲 from symmetry is computed from
⌬
2n0⫺1
兺 i兩 I s共 x i 兲兩
⫽ i⫽1
2n0⫺1 ⫺ n 0, (3)
␦
兺i⫽1
兩 I s共 x i 兲兩
where n0 is the point number of center burst location
I共 x̂0兲, and ⌬兾␦ can be viewed as an estimate of the
center of gravity of Is共x兲 or as the first moment of i
2n0⫺1
with 兩Is共xi 兲兩兾¥i⫽1 兩Is共xi 兲兩 taken as a weighing func-
tion. If Is共x兲 were exactly symmetric 关where Eq. 共3兲
is used as a measure of symmetry兴 about n0, then ⌬
would be zero. It is important to note that this pro-
cedure tends to improve the symmetry of the inter-
ferogram but does not guarantee complete symmetry,
for which all odd moments of Is共x兲 are zero and
Is共x0 ⫺ x兲 ⫽ Is共x0 ⫹ x兲. A common method of esti-
mating ⌬ is to perform a polynomial fit about I共x̂0兲
共where the proper polynomial order and the number
of points for the fit affect ⌬ and should be predeter-
mined兲 and computing the location of the maximum.
In practice, because the signal-to-noise ratio in the
measured interferogram decreases rapidly as x
moves farther from the center-burst location, we use
fewer points 共⬃200兲 and not all the 共2n0 ⫺ 1兲 points
about the center burst to compute ⌬ in Eq. 共3兲.
A linear phase-shift correction can be thought of as
a convolution of I共x兲 by a shifted Dirac delta function
␦共x ⫺ ⌬兲, given by
I共 x ⫺ ⌬兲 ⫽
兰 I共␣兲␦关␣ ⫺ 共 x ⫺ ⌬兲兴d␣
兰
⫽ I共␣兲␦共 x ⫺ ⌬ ⫺ ␣兲d␣ (4)
关using the symmetry ␦共t兲 ⫽ ␦共⫺t兲兴 or with the Fourier-
transform operation given as
I共 x ⫺ ⌬兲 ⫽ F ⫺1兵F关I共 x兲兴F关␦共 x ⫺ ⌬兲兴其
⫽ F ⫺1关F关I共 x兲兴exp共⫺j2␲␯⌬兲兴, (5)
where j is the imaginary number 公⫺1. For a vector
I共x兲 of length n uniformly spaced every ␦ cm, the n
frequencies ␯ are 0, 共n␦兲⫺1, 2共n␦兲⫺1, 3共n␦兲⫺1, . . . ,
共n ⫺ 1兲共n␦兲⫺1, and frequencies ␯ for Eq. 共5兲 are
再
␯ ␯ ⱕ ␯ Nyquist
␯⫽ , (6)
⫺共2␯ Nyquist ⫺ ␯兲 ␯ ⬎ ␯ Nyquist
where ␯Nyquist ⫽ 共2␦兲⫺1.
20 February 2002 兾 Vol. 41, No. 6 兾 APPLIED OPTICS 1183

Fig. 2. Sinc共100x兲 with ⫺xmin ⱕ x ⱕ xmax, exact shifted function
sinc关100共x ⫺ ⌬兲兴, and approximations I共x ⫺ ⌬兲 from the shift
theorem, where I共x ⫺ ⌬兲 computed with Eq. 共5兲 using single-sided
and reconstructed double-sided I共⫺xmin ⱕ x ⱕ xmax兲 关Eq. 共7兲兴 is
used. The residual ⑀共x兲 ⫽ I共x ⫺ ⌬兲 ⫺ sinc共x ⫺ ⌬兲 shows the
improved accuracy when Eq. 共7兲 is used. The centerburst at x ⫽
0 is not sampled on this x-axis grid and is located between two
adjacent values of x. ⌬ ⫽ 0.075 ⫹ ␦兾4, where ␦ ⫽ 10⫺4 is the
interval between adjacent points.
A. Implementation of the Shift Theorem for I共x ⫺ ⌬兲
The shift theorem implemented with a fast-Fourier-
transform 共FFT兲 operation implies cyclical 共periodic兲
functions 共which are due to the FFT operation兲. But
sampled interferogram I共x兲 is asymmetric and is
sampled at ⫺xmin ⬍ x ⬍ xmax. Thus shift theorem
I共 x ⫺ ⌬兲 implemented with a FFT moves 共clockwise
rotation for ⌬ ⬎ 0兲 in regions near xmax to regions
near xmin owing to the cyclical rotation. We solve
this problem by creating a double-sided interfero-
gram Id共x兲 for the complete range ⫺xmax ⱕ x ⱕ xmax
by using the portion I共xmin ⱕ x ⱕ xmax兲 of the recorded
interferogram for ⫺xmax ⱕ x ⱕ ⫺xmin with the rest of
the recorded interferogram I共⫺xmin ⱕ x ⱕ xmax兲 be-
fore the shift operation with the FFT shift theorem
关Eq. 共5兲兴, where Id共x兲 is given by
I d共 x兲 ⫽ 再 I共⫺x兲
I共 x兲
⫺x max ⱕ x ⱕ ⫺x min
⫺x min ⬍ x ⱕ x max
. (7)
A small discontinuity at ⫺xmin is introduced in Id共x兲.
Then we take the new linearly phase-corrected region
⫺xmin ⬍ x ⬍ xmax to be I共x ⫺ ⌬兲. This improvement
is modest when small shifts ⌬ are involved.
In Fig. 2 we demonstrate the difference between
the shift theorem 关Eq. 共5兲兴 for I共x兲 and 关Eq. 共7兲兴 for
Id共x兲. For illustration of the significant difference
we chose a sinc共100x兲 function for I共x兲 with ⫺xmin ⱕ
x ⱕ xmax uniformly spaced from xmin ⫽ 0.1 ⫹ ␦兾2 to
xmax ⫽ 0.25 ⫹ ␦兾2, where ␦ ⫽ 10⫺4 is the interval
between adjacent points. The center burst at x ⫽ 0
is not sampled on this x-axis grid and is located be-
tween two adjacent values of x. A sinc function is
the interferogram of rectangular B共␯兲 in Eqs. 共2兲.
We chose a value ⌬ ⫽ 0.075 ⫹ ␦兾4 for the desired
1184 APPLIED OPTICS 兾 Vol. 41, No. 6 兾 20 February 2002
shift, which falls at point 750.25 on the x-axis grid
共i.e., between grid points兲. In Fig. 2 we show the
function sinc共100x兲, the exact shifted function
sinc关100共x ⫺ ⌬兲兴, and the approximation I共x ⫺ ⌬兲
computed with Eqs. 共5兲 and 共7兲. The residual ⑀共x兲 ⫽
I共x ⫺ ⌬兲 ⫺ sinc共x ⫺ ⌬兲 shows the improved accuracy
with Eq. 共7兲.
B. Iteration Process for Linear Phase Correction
We implemented the linear phase correction for the
sampled interferogram in an iterative process with
which, from the first iteration phase-corrected inter-
ferogram, I共x ⫺ ⌬兲, we extracted a new short double-
sided Is共x兲 and ⌬ 关Eq. 共3兲兴 and computed a second
iteration phase-corrected interferogram with Eq. 共5兲.
Usually no more than a few iterations are needed for
a phase correction to be achieved with an accuracy of
⌬兾␦ ⬍ 0.001. At the end of the process we obtained
interferogram I共x兲, where the center burst location
was resampled with high accuracy and the linear
part of the phase, ⌽共␯兲 in Eqs. 共1兲, was corrected to a
large degree. In addition, the symmetry of short-
double-sided region Is共x兲 was improved. For the rest
of this paper we refer to I共 x ⫺ ⌬兲 in Eq. 共5兲 as phase-
corrected asymmetric 共single-sided兲, ⫺xmin ⱕ x ⱕ
xmax, interferogram I共x兲.
3. Apodizing Functions
Spectrum B共␯兲 was obtained from I共 x兲 关Eqs. 共1兲兴 by a
Fourier-transform operation 关Eqs. 共2兲兴, B̂共␯兲 ⫽
F关I共x兲w共x兲兴, where w共x兲 is an apodizing function 共i.e.,
a window function兲 that is designed9 to minimize
sidelobes and thereby to reduce the distortion of the
derived estimated spectrum B̂共␯兲 that results from
the transfer of energy 共leakage兲 across frequencies ␯,
a leakage that can even produce negative values and
therefore an unphysical spectrum. The design of
the apodization function involved a trade-off between
minimizing the sidelobes 共reducing energy leakage兲
of the power spectrum of the window at the expense
of increasing the width of the main lobe, an increase
that reduces the resolution 共i.e., a smearing effect兲.
The resolution width is given in units of spectral bins,
in which a bin size is the fundamental frequency
resolution N⫺1, where N is the number of points in
the window w共x兲. Usually the spectral bin widths of
the apodization windows range from 1.2 to 3 spectral
bins. Boxcar window w共x兲 has the narrowest main
lobe 共1.21 spectral bins兲 and the highest sidelobes
共⫺13 dB, falling at a rate of ⫺6 dB兾octave兲. The
traditionally used triangular window in a FTIR spec-
trometer has a resolution width of 1.78 spectral bins,
a highest sidelobe level of ⫺27 dB, and a sidelobe
falloff rate of ⫺12 dB兾octave, a decay rate that is
usually too slow for accurate spectral analysis. A
Chebychev window designed to have a constant level
of sidelobes of ⫺80 dB has a resolution width of 2.31
spectral bins and for a constant level of sidelobes of
⫺50 dB has a resolution width of 1.85 spectral bins.
With the phase correction, interferogram I共x兲 is
single sided where the region about center-burst lo-
cation x0 appears twice. Without proper weighing of
Fig. 3. Asymmetrical hybrid apodization windows. xmin ⫽ 5% of
the window width 共100 points兲. 共a兲 Time domain where a ramp
function is substituted for the range ⫺xmin ⱕ x ⱕ xmin. Triangle,
Blackman, Kaiser, and Chebyshev apodization windows for the
range xmin ⱕ x ⱕ xmax. 共b兲 Power spectral density magnitude 10
log10兵兩F关w共x兲兴兩2其 in the frequency domain where the Nyquist fre-
quency is 1. Attenuation R ⫽ 100 dB for Chebyshev and Kaiser
windows.
the interferogram, intensities of sharp absorption
bands 共frequencies兲 in the deduced spectrum will be
distorted. A proper weighing function w共x兲 such
that w共x0 ⫺ x兲 ⫹ w共x0 ⫹ x兲 ⫽ 1 for the range ⫺xmin
ⱕ x ⱕ xmin will correct this distortion. Mertz chose
the weighing function w共x兲 to be a ramp, and Rah-
melow and Hubner5 chose a fifth-order polynomial for
⫺xmin ⱕ x ⱕ xmin, w共x兲 ⫽ 0.5 ⫹ 1.25共x兾xmin兲3 ⫺
0.75共x兾xmin兲5, which is smoother than a ramp func-
tion and thus is somewhat better for a Fourier-
transform operation in which one wants to avoid
sharp discontinuities.
We tried to combine ramp function w共x兲 共or a fifth-
order polynomial兲 for the range ⫺xmin ⱕ x ⱕ xmin with
different windows w共x兲 for the range xmin ⱕ x ⱕ xmax
to achieve the proper weighing of the interferogram
in the short double-sided region and an improved
apodizing window for the rest of the range. Exam-
ples of hybrid asymmetrical windows 共a ramp com-
bined with a triangle, Blackman, Kaiser, and
Chebyshev windows兲 are presented in Fig. 3共a兲 for
xmin ⫽ 5% of the window width; in Fig. 3共b兲 the cor-
responding power spectral density magnitude, 10
log10关兩F共w共x兲兲兩2兴 共in decibels兲, of the windows w共x兲 is
shown in the frequency domain. These asymmetric
windows are not significantly better than the tradi-
tional Mertz asymmetric triangular window. The
good properties 共i.e., narrow main lobe and low side-
lobes兲 of the more-sophisticated windows 共e.g., Black-
man, Kaiser, and Chebyshev兲 were lost owing to the
asymmetry. The symmetrical apodization windows
for triangle, Blackman, Kaiser, and Chebyshev func-
tions are shown in Fig. 4 共R ⫽ 100 dB for Kaiser and
Chebyshev windows兲, where parameter ␣ for a Kaiser
Fig. 4. Symmetrical apodization windows 共xmin ⫽ xmax兲 for which
the peak location of the window 共in the time domain兲 is at the
middle 共a兲 in the time domain and 共b兲 in the frequency domain.
Attenuation R ⫽ 100 dB for Chebyshev and Kaiser windows.
window is given for attenuation R 共decibels兲 as ␣ ⫽
0.5842共R ⫺ 21兲0.4 ⫹ 0.07886共R ⫺ 21兲 for 50 ⱖ R ⱖ 21
and ␣ ⫽ 0.1102共R ⫺ 8.7兲 for R ⬎ 50. The superiority
of the symmetrical windows to the asymmetrical
windows is evident. To take full advantage of the
symmetrical apodization we constructed from the
single-sided phase-corrected interferogram I共⫺xmin ⱕ
x ⱕ xmax兲 a double-sided phase-corrected interfero-
gram Id共x兲 关Eq. 共7兲兴. In this process we introduced a
discontinuity at x ⫽ ⫺xmin, but this discontinuity is
small because of the phase correction and will be
reduced with the apodization window. In double-
sided phase-corrected interferogram Id共x兲 the center
burst location is in the middle, and all points in in-
terferogram Id共x兲 appear twice 共therefore no special
weighing is necessary兲, and thus a symmetrical apo-
dization window 共Fig. 4兲 that is superior to the asym-
metrical triangular window can be used. A
demonstration of the superiority of a spectrum com-
puted with the double-sided phase-corrected inter-
ferogram to a spectrum computed with a zero-fill
vector for the missing data ⫺xmax ⱕ x ⱕ ⫺xmin is
presented in Section 6 below.
4. Transforming the Recorded Interferogram I共x兲 to a
Solution Spectrum S共␯兲
Given recorded interferogram I共x兲 sampled at n
points, unknown spectrum B共␯兲 can be deduced at a
set of discrete frequencies 共wave numbers兲 ␯ that are
uniformly spaced between a zero frequency and a
maximum frequency, the Nyquist frequency, ␯max ⫽
␯Nyquist ⫽ 共2␦兲⫺1, where ␦, the distance between con-
secutive optical-path difference x values, is usually
determined by an external clock 共e.g., a He–Ne laser兲.
Padding the interferogram with m0 zeros helps to
reduce the discontinuities that are inherent in a dis-
crete Fourier-transform operation in which the
n-point vector is treated as a periodic sequence of
20 February 2002 兾 Vol. 41, No. 6 兾 APPLIED OPTICS 1185
many n-point vectors. It is also sometimes of inter-
est to obtain a representation of an estimated spec-
trum B̂共␯兲 at a different 共smaller兲 frequency interval
⌬␯, an objective that can be achieved with appropri-
ate zero padding. The number of available discrete
frequencies n␯ from ␯ ⫽ 0 to ␯max for an n-point vector
padded with m0 zeros is
n␯ ⫽ 再 0.5共n ⫹ m 0兲 ⫹ 1
0.5共n ⫹ m 0 ⫹ 1兲
共n ⫹ m 0兲 even
共n ⫹ m 0兲 odd
, S共␯兲 ⫽ Re关B共␯兲兴
and the spacing 共intervals兲 between consecutive fre-
quencies is ⌬␯ ⫽ 1兾关共n ⫹ m0兲␦兴 ⫽ 2共n ⫹ m0兲⫺1␯Nyquist.
However, the minimum frequency 共i.e., ␯ ⫽ 0兲 and the
maximum frequency remain unchanged with any
zero padding.
We apodized the phase-corrected double-sided in-
terferogram 关Eq. 共7兲兴, Idw共x兲 ⫽ Id共x兲w共x兲, where the
superscript w denotes the windowing 共apodizing兲 pro-
cess. To obtain a specific interval ⌬␯ ⬍ 共2xmax兲⫺1 we
padded Id共x兲 with m0 zeros, m0 ⫽ 共␦⌬␯兲⫺1 ⫺ nd, where
nd ⫽ 2共n ⫺ n0兲 ⫹ 1 ⫽ 2xmax兾␦ ⫹ 1 is the number of
points in Id共x兲. The phase shift ⌽2共␯兲 that is due to
the location of the centerburst, at point n0d ⫽ 共n ⫺
n0 ⫹ 1兲 in double-sided phase-corrected interfero-
gram Idw, is a phantom phase that is easily corrected
by a simple rotation, as was demonstrated by Mertz,7
to place the center burst at the first element.
The first estimate spectrum B̂共␯兲 共a complex quan-
tity兲 from Idw is given by
F共Idw兲兾K共␯兲 nd
B̂共␯兲 ⫽ 2
nd 2␯ Nyquist
⫽ 2␦F共Idw兲兾K共␯兲
F共Idw兲兾K共␯兲
⬵ 关W兾共cm2 sr cm⫺1兴,
n d共2x max兲 ⫺1兾2
where the factor 2 compensates for the factor 0.5 in
Eqs. 共1兲, nd in the denominator normalizes the length
of 共unpadded兲 vector Id共x兲, factor 2nd⫺1␯Nyquist ⬵
共2xmax兲⫺1 is the spectral bin size 共inverse centime-
ters兲 in the FFT operation to produce a spectrum per
wave number. For nd ⫹ m0 ⫽ odd, spectrum B̂共␯兲
has to be multiplied by 2 for all frequencies other
than the zero frequency to account for the energy in
the imaginary frequencies in the FFT operation.
For nd ⫹ m0 ⫽ even, B̂共␯兲 has to be multiplied by 2 for
all frequencies other than the zero frequency and the
Nyquist frequency, which appears once in the FFT
vector.
In the Mertz7 procedure, the residual nonlinear
phase in ⌽̂1共␯兲 is estimated from a Fourier transform
of the short double-sided interferogram for a limited
number of frequencies and then interpolated 关assum-
ing that ⌽̂1共␯兲 is a smooth function兴 to all frequencies
for B̂共␯兲. In our procedure, phase ⌽̂1共␯兲 that was not
corrected in the iterative phase correction can be es-
timated directly for all frequencies from B̂共␯兲 and is
given by ⌽̂1共␯兲 ⫽ tan⫺1兵Im关B共␯兲兴兾关Re共B̂共␯兲兴其, where
Im共 兲 denotes the imaginary part and special care is
given to the signs of Im关B̂共␯兲兴 and Re关B̂共␯兲兴 to preserve
1186 APPLIED OPTICS 兾 Vol. 41, No. 6 兾 20 February 2002
the sign of ⌽̂1共␯兲 兵e.g., for Im关B̂共␯兲兴 ⫽ ⫺1 and Re关B̂共␯兲兴
⫽ ⫺1, ⌽̂1共␯兲 ⫽ 45° in the first quadrant when the
correct angle is ⫺135° in the third quadrant其. The
estimated spectrum B̂共␯兲 is corrected for the residual
nonlinear part of phase ⌽̂1共␯兲 to give the sought-after
spectrum B共␯兲 and is denoted solution spectrum S共␯兲:
⫽ Re关B̂共␯兲兴cos关⌽̂ 1共␯兲兴 ⫹ Im关B̂共␯兲兴sin关⌽̂ 1共␯兲兴.
(9)
With this phase correction we neglected the small
imaginary part in the derived solution spectrum:
Im关S共␯兲兴 ⫽ ⫺Re关B̂共␯兲兴sin关⌽̂1共␯兲兴 ⫹ Im关B̂共␯兲兴cos关⌽̂1共␯兲兴.
5. Comparison of Solution S共␯兲 and Source Spectrum
B共␯兲
Solution S共␯兲 is a lower-resolution estimate 共some-
times also called a deresolved spectrum兲 of true input
source spectrum B共␯兲. The source spectrum is de-
graded 共smeared兲 as a result of the instrument trans-
fer function 关also called the instrument line shape
共ILS兲兴 and the maximum optical path difference xmax
that was used in the interferogram measurements.
Thus, to compare solution spectrum S共␯兲 derived with
our method with B共␯兲 we must first degrade the res-
olution of B共␯兲 to the spectrum that would have been
measured by the instrument 共i.e., without regard to
the algorithm used to derive a spectrum from the
recorded interferogram兲.
The wavelength resolution of S共␯兲 is given by the
(8)
Rayleigh resolution criterion1 as ⌬ ⫽ 2⌬␯ ⬵ 1兾xmax.
With the Rayleigh resolution criterion, two mono-
chromatic sources of a unit intensity at two frequen-
cies, ␯1 and ␯2 ⫽ ␯1 ⫹ ⌬, are considered to be just
resolved when they produce a combined spectrum
关for an instrument line shape in the form of a
sinc2共 x兲 function兴 with a dip of 81% in the middle of
the combined peak. For convenience 共it is easier to
convolve Gaussian functions than sinc functions兲
we implement the Rayleigh resolution criterion
for the ILS in the form of a Gaussian function given
by
exp关⫺0.5共␯ ⫺ ␯ 1兲 2兾共⌬兾2.638兲 2兲兴
g共␯, ␯ 1兲 ⫽ , (10)
共2␲兲 0.5共⌬兾2.638兲
for which the combined spectrum g共␯兲 ⫽ g共␯, ␯1兲 ⫹ g共␯,
␯1 ⫹ ⌬兲 is a good approximation of the Rayleigh cri-
terion, as is shown in Fig. 5 for ⌬ ⫽ 2⌬␯ ⬵ 1兾xmax ⫽
4 cm⫺1. The two Gaussians10 intersect at x ⫽ ⌬兾2,
and each has a FWHM value of 0.8927⌬. When
spectrum B共␯兲 is measured with a Gaussian ILS the
resultant reduced resolution spectrum is denoted
Fig. 5. Rayleigh resolution criterion implemented with Gaussian
functions. g共␯兲 ⫽ g共␯, ␯1兲 ⫹ g共␯, ␯1 ⫹ ⌬兲 is the combined spectrum
of two monochromatic sources, g1 and g2, at frequencies ␯1 ⫽ 1000
cm⫺1 and ␯2 ⫽ ␯1 ⫹ ⌬ separated by ⌬ ⫽ 4 cm⫺1.
L共␯兲 and is given by convolution of the ILS 关Eq. 共10兲兴
with B共␯兲, given by
兰
⬁ FWHM and transmissions of 0.1 to 0.5, and atmo-
L共␯兲 ⫽ B共t兲 g共t, ␯兲dt
0 the MODTRAN program11 for a 1976 U.S. Standard At-
兰
⬁ exp关⫺0.5共t ⫺ ␯兲 2兾共⌬兾2.638兲 2兲兴
⫽ B共t兲 dt.
0
共2␲兲 0.5共⌬兾2.638兲
(11)
Fig. 6. Spectrum B共␯兲 W兾共cm2 sr cm⫺1兲. 共a兲 Planck function at a
temperature of 300 K; 共b兲 Planck function at a temperature of 300
K with five Lorentzian absorption lines of 10-cm⫺1 FWHM and
transmissions 0.1, 0.2, 0.3, 0.4, and 0.5 at frequencies 600, 800,
1000, 1200 and 1400 cm⫺1, respectively; 共c兲 atmospheric radiance
共1-cm⫺1 resolution兲 computed with the MODTRAN program for a
1976 U.S. Standard Atmosphere for an observer on the ground
looking up. Wavelength range, 510 –1500 cm⫺1.
Fig. 7. Nonlinear phase error ⌽1共␯兲 rad 共obtained from a FTIR
instrument兲 representing the dispersive optical elements in a spec-
trometer.
6. Results
We simulated interferograms 关Eqs. 共1兲兴 for three
types of source spectrum B共␯兲 shown in Fig. 6: a
Planck function 共i.e., a blackbody兲, a Planck function
with five Lorentzian1 absorption lines of 10-cm⫺1
spheric radiance 共1-cm⫺1 resolution兲 computed with
mosphere and an observer on the ground looking up.
The wavelength range was 510 –1500 cm⫺1, and the
temperature for the Planck blackbody function was
300 K. The maximum optical path difference was
xmax ⫽ 0.25 cm, and there were n ⫽ 2048 sampled
points from xmin ⫽ ⫺0.05 cm to xmax. The Nyquist
frequency was 3411.7 cm⫺1. We tried to make the
simulations as realistic as possible. Thus the loca-
tion of the center burst 共x ⫽ 0兲 was not sampled in
interferogram I共 x兲, and a nonlinear phase error,
⌽1共␯兲, obtained from a FTIR instrument and shown in
Fig. 7, was used in simulating the interferograms.
Derived spectrum S共␯兲 was constructed at equal in-
tervals of ⌬␯ ⫽ 1.9995 cm⫺1 for zero padding, and a
corresponding spectrum L共␯兲 was computed from B共␯兲
with ⌬ ⫽ xmax⫺1 ⫽ 4 cm⫺1. As noted above 共Section
3兲, the apodization function is a trade-off between
minimizing the sidelobes 关and thereby minimizing
the leakage from a nearby spectral features of B共␯ ⫽
␯0兲 to derived spectrum S共␯0兲兴 at the expense of in-
creasing the width of the main lobe and thus smear-
ing and spreading S共␯0兲 共and thereby distorting its
magnitude兲. Assuming that we would like to obtain
a large dynamic range, of the order of 105, in derived
spectra S共␯兲 to exploit the high signal-to-noise ratio
共SNR兲 in the interferogram measurements, an atten-
uation R 共in the Kaiser and Chebyshev apodization
windows兲 of 50 –100 dB will be sufficient. The SNR
for a shot-noise process 共owing to the statistical fluc-
tuations in the incident photon flux兲 is the maximum
attainable SNR. For a blackbody source B共500 ⬍
␯ ⬍ 1500 cm⫺1兲 at a temperature of 300 K the SNR for
20 February 2002 兾 Vol. 41, No. 6 兾 APPLIED OPTICS 1187
Fig. 8. Radiance spectrum L共␯兲 W兾共cm2 sr cm⫺1兲 关Eq. 共11兲兴, with
⌬ ⫽ 4 cm⫺1 for B共␯兲, and solution spectrum S共␯兲 W兾共cm2 sr cm⫺1兲
共top兲. Residual percent difference 100兩L共␯兲 ⫺ S共␯兲兩兾L共␯兲 共middle兲.
Absolute radiance difference 兩L共␯兲 ⫺ S共␯兲兩 W兾共cm2 sr cm⫺1兲 共bot-
tom兲. Solution spectrum S共␯兲 is derived for Planck function B共␯兲
关Fig. 6共a兲兴 for a Chebyshev apodizing window with R ⫽ 100 dB.
a shot-noise process is of the order of 107, and the
SNR that is due to detector noise and background
photon flux in FTIR spectrometers is of the order of
103 to 104. It was also noted previously6 that Kaiser
and Chebyshev apodization windows are good choices
for use in retrievals.
In Figs. 8 –10 we show spectra L共␯兲关W兾共cm2 sr
cm⫺1兲兴 and solution spectrum S共␯兲关W兾共cm2 sr cm⫺1兴,
the residual percentage difference 100兩L共␯兲 ⫺ S共␯兲兩兾
Fig. 9. Same as Fig. 8 but here source spectrum B共␯兲 关Fig. 6共b兲兴 is
a Planck function with five Lorentzian absorption peaks 共10-cm⫺1
FWHM兲 with transmissions of 0.1, 0.2, 0.3, and 0.5 at frequencies
600, 800, 1000, 1200, and 1400 cm⫺1. Two spectra, S1共␯兲 and
S2共␯兲 共computed with a Chebyshev window with R ⫽ 100 dB兲, are
shown. For S2共␯兲 the double-sided phase-corrected interferogram
was used, and for S1共␯兲 zeros were substituted for the missing data
⫺xmax ⱕ x ⱕ ⫺xmin in phase-corrected interferogram Id共x兲 关Eq. 共7兲兴.
1188 APPLIED OPTICS 兾 Vol. 41, No. 6 兾 20 February 2002
Fig. 10. Same as Fig. 8 but here source spectrum B共␯兲 关Fig. 6共c兲兴
is atmospheric radiance 共1-cm⫺1 resolution兲 computed with MODT-
RAN and S共␯兲 is computed with a Kaiser window with R ⫽ 50 dB.
L共␯兲, and the absolute radiance difference 兩L共␯兲 ⫺
S共␯兲兩关W兾共cm2 sr cm⫺1兴. In Fig. 8, spectra S共␯兲 are
shown for Planck function B共␯兲 关Fig. 6共a兲兴 for a Cheby-
shev apodizing window with R ⫽ 100 dB. The figure
shows an error that is in general less than 0.1% and a
radiance difference of less than 10⫺8 W兾共cm2 sr cm⫺1兲.
The importance of constructing a double-sided
phase-corrected interferogram 关expression 共8兲兴 is
shown in Fig. 9, where two spectra, S1共␯兲 and S2共␯兲
共both computed with a Chebyshev window with R ⫽
100 dB兲, are shown for radiance spectrum B共␯兲 of Fig.
6共b兲. For S1共␯兲, zeros were substituted for the miss-
ing data, ⫺xmax ⱕ x ⱕ ⫺xmin, in phase-corrected
interferogram Id共x兲 关Eq. 共7兲兴, and for S2共␯兲 the double-
sided phase-corrected interferogram was used. The
figure shows the degradation in accuracy 共of the order
of 5–10兲 for S1共␯兲, a degradation that increases as the
range of missing data increases. Substituting zeros
for the missing data and using symmetrical apodiza-
tion functions yields an interferogram that is not
weighted properly 共i.e., not all the fringes in the in-
terferogram are counted equally兲. As a result, de-
duced spectrum S1共␯兲 is distorted. In general,
fringes in the interferogram located farther from the
center burst 共large values of x兲 contain spectral infor-
mation on sharp spectral features of B共␯兲 共i.e., sharp
absorption bands兲 and fringes closer to the center
burst contain spectral information on smooth spec-
tral features of source B共␯兲.
In Fig. 10, the atmospheric radiance 共1-cm⫺1 reso-
lution兲 of source spectrum B共␯兲 关Fig. 6共c兲兴 computed
with the MODTRAN program in which there are many
narrow absorption lines, and S共␯兲 computed with a
Kaiser window with R ⫽ 50 dB, are shown. The
accuracy of S共␯兲 is within a few percent, and radiance
the difference is smaller than 10⫺7 W兾共cm2 sr cm⫺1兲.
7. Summary
A method for transforming symmetric single-sided
interferograms into spectra that combines the For-
man and the Mertz methods has been developed. In
this new method we use any symmetric apodizing
windows 共e.g., triangular, Hamming, Hanning,
Blackman, Chebyshev, and Kaiser兲. The measured
interferogram is asymmetric and contains phase er-
rors that are due to missampling of the interferogram
center burst’s location and a phase error from disper-
sive optical elements and electronics in the spectrom-
eter. We use a short double-sided fraction of the
asymmetric interferogram and apply an iterative
phase correction for the linear portion of the phase to
the recorded single-sided asymmetric interferogram.
The linear phase correction produces a phase-
corrected single-sided interferogram and is imple-
mented with an iterative method in which the
location of the center burst is computed by estimation
of the center of gravity of the short, double-sided
portion of the measured interferogram, and the shift
theorem is applied to a reconstructed double-sided
interferogram. The iterative phase correction
enhances the symmetry in the double-sided interfero-
gram computed from the phase-corrected single-
sided recorded interferogram.
Because of the poor properties of the asymmetric
apodization functions, we constructed a symmetric
double-sided interferogram from the single-sided
phase-corrected interferogram and followed the
Mertz procedure to produce the solution spectrum
but with symmetrical apodization windows and with
nonlinear correction derived from the double-sided
interferogram without the need to interpolate over
frequencies as is done in the Mertz procedure.
Examples of the accuracy of the derived spectrum
have been shown for a Planck function of tempera-
ture of 300 K, for a Planck function of temperature of
300 K with five Lorentzian absorption lines of 10-
cm⫺1 FWHM and transmissions of 0.1 to 0.5, and for
a moderate resolution radiance 共1 cm⫺1兲 computed
with the MODTRAN program for a standard atmosphere
for an observer on the ground looking up. A nonlin-
ear phase error 共obtained from a FTIR spectrometer兲
was included in the simulated interferograms, and
the location of the center burst 共x ⫽ 0兲 was not sam-
pled so realistic simulated interferograms could be
produced. In comparing the solution spectrum with
the source spectrum we applied the Rayleigh resolu-
tion criterion with a Gaussian instrument line shape.
The accuracy of the solution has been shown to be
excellent, ranging from better than 0.1% for a Planck
blackbody 关Fig. 6共a兲兴 and a radiance difference of less
than 10⫺8 W兾共cm2 sr cm⫺1兲 to a few percent for a
View publication stats
complicated atmospheric radiance spectrum 关Fig.
6共c兲兴 and radiance difference smaller than 10⫺7
W兾共cm2 sr cm⫺1兲.
This study was supported by the U.S. Army Soldier
and Biological Chemical Command, Edgewood
Chemical Biological Center 共ECBC; formerly Edge-
wood Research Development and Engineering Cen-
ter兲 under contract DAAM01-94-C-0079. We thank
Thomas Gruber and Rich Vanderbeek for many use-
ful discussions and Bill Loerop of the ECBC for ad-
ministrative support and encouragement. We are
very grateful for the constructive and insightful sug-
gestions of the anonymous reviewer, especially re-
garding the FFT operation and the shift theorem.
References and Notes
1. P. R. Griffiths and J. A. de Haseth, Fourier Transform Infrared
Spectrometry 共Wiley, New York, 1986兲.
2. D. B. Chase, “Phase correction in FT-IR,” Appl. Spectrosc. 36,
240 –244 共1982兲.
3. H. E. Revercomb, H. Buijs, H. B. Howell, D. D. LaPorte, W. L.
Smith, and L. A. Sromovsky, “Radiometric calibration of IR
Fourier transform spectrometers: solution to a problem with
the High-Resolution Interferometer Sounder,” Appl. Opt. 27,
3210 –3218 共1988兲.
4. R. C. M. Learner, A. P. Thorne, I. Wynne-Jones, J. W. Brault,
and M. C. Abrams, “Phase correction of emission line Fourier
transform spectra,” J. Opt. Soc. Am. A 12, 2165–2171 共1995兲.
5. K. Rahmelow and W. Hubner, “Phase correction in Fourier
transform spectroscopy: subsequent displacement correction
and error limit,” Appl. Opt. 36, 6678 – 6686 共1997兲.
6. C. D. Barnet, J. M. Blaisdell, and J. Susskind, “Practical
method for rapid and accurate computation of interferometric
spectra for remote sensing application,” IEEE Trans. Geosci.
Remote Sens. 38, 169 –183 共2000兲.
7. L. Mertz, “Auxiliary computation for Fourier spectrometry,”
Infrared Phys. 7, 17–23 共1967兲.
8. M. L. Forman, W. H. Steel, and G. A. Vanasse, “Correction of
asymmetric interferograms obtained in Fourier spectroscopy,”
J. Opt. Soc. Am. 56, 59 – 63 共1966兲.
9. F. J. Harris, “On the use of windows for harmonic analysis with
the discrete Fourier transform,” Proc. IEEE 66, 51– 83 共1978兲.
10. Based on probability theory, the probability of measuring a
signal with a Gaussian probability function g共␯, ␯1 ⫽ ⌬兲 in the
presence of zero-mean noise with probability-density function
⬁
g共␯, ␯1 ⫽ 0兲 is 兰⌬兾2 g共␯, ⌬兲d␯ ⫽ 90% when a threshold of
detection ␯ ⱖ ⌬兾2 is chosen.
11. A. Berk, G. P. Anderson, L. S. Bernstein, P. K. Acharya, H.
Dothe, M. W. Matthew, S. M. Adler-Golden, J. H. Chetwynd,
Jr., S. C. Richtsmeier, B. Pukall, C. L. Allred, L. S. Jeong, and
M. L. Hoke, “MODTRAN4 radiative transfer modeling for atmo-
spheric correction,” in Optical Spectroscopic Techniques and
Instrumentation for Atmospheric and Space Research III, A. M.
Larar, ed., Proc. SPIE 3756, 348 –353 共1999兲.
20 February 2002 兾 Vol. 41, No. 6 兾 APPLIED OPTICS 1189