Sunteți pe pagina 1din 11

C. Dorrer and I. A. Walmsley Vol. 19, No. 5 / May 2002 / J. Opt. Soc. Am.

B 1019

Accuracy criterion for ultrashort pulse


characterization techniques:
application to spectral
phase interferometry for
direct electric field reconstruction

Christophe Dorrer and Ian A. Walmsley


The Institute of Optics, University of Rochester, Rochester, New York 14627

Received March 13, 2001; revised manuscript received October 15, 2001
We propose a procedure for the comparison of two ultrashort optical pulses based on the root-mean-square field
error. This procedure supplies a practical criterion that allows one to test the accuracy of pulse characteriza-
tion techniques. We apply this procedure to spectral phase interferometry for direct electric field reconstruc-
tion (SPIDER), a method that uses shearing interferometry in the frequency domain to measure ultrashort
light pulses. Two novel reconstruction algorithms are described in detail, one based on the reconstruction of
the spectral phase by concatenation, the other by integration. The range of validity of these algorithms and
the influence of noise on the accuracy of the reconstruction are studied using the proposed criterion. 2002
Optical Society of America
OCIS codes: 120.3180, 120.5050, 320.7100.

1. INTRODUCTION sible experimental artifacts. The field is then retrieved


by use of the algorithms that would be used under experi-
The electric field is a fundamental entity in optics. A
mental conditions. The difference between the test field
number of methods are known for measuring it, or a
and the retrieved field is then evaluated according to a
closely related quantity, such as the coherence function.1
Most recently, attention has been paid to the character- chosen criterion. Accuracy criteria are useful for the
ization of the field of ultrashort pulses, in its most general comparison of different techniques aimed at reconstruct-
form a function of both temporal and spatial coordinates. ing the same quantity and also for determining the range
The knowledge of the field is important in a wide variety of working parameters for the successful operation of a
of scientific and technological applications, and this has given technique.
lent increasing importance to metrology. The definition of such a criterion is somewhat subjec-
The main difficulty in determining the optical field is tive, and many different accuracy measures have been
that most optical detectors are absorptive, and such used, most of them adopted from the field of image
square-law detectors are not sensitive to the phase of the reconstruction.3 It seems reasonable to base a test of ac-
optical wave. Therefore indirect techniques must be curacy on the retrieved quantity itself, rather than some
used for the characterization of the electric field of ul- functional of this quantity. Nonetheless, one might
trashort optical pulses. Several methods have been pro- choose to emphasize the importance of the reconstruction
posed and demonstrated. These methods may be classi- of some features of the field by using a weighting function
fied according to their experimental implementation and that gives more weight to these features.
the nature of the experimental trace, as well as the algo- In this paper, we use the root-mean-square (RMS) field
rithms used to extract the electric field from the experi- error to compare the fields of two ultrashort optical
mental data.2 In all methods the quantity of interest is pulses. This enables us to test the accuracy of any pulse
the electric field of the pulse as a function of the optical characterization technique by quantifying the difference
frequency or time. between a synthetic test pulse that is used to test the re-
The accuracy of these measurement methods is of ut- construction and the reconstructed pulse itself. Al-
most importance. It is critical in many applications that though this is a simple way of comparing two continuous
the reconstruction of the physical quantity from the mea- functions and is widely used in the domain of image re-
sured quantity matches well the actual physical quantity construction (and more generally in mathematics, where
from which the measured data were derived. This is it is known as the Euclidian distance), to our knowledge
clearly a theoretical task, relying on simulations or equa- little attention has been paid to it in the context of ul-
tions, because the actual field is unknown before a mea- trashort pulses. Other studies have relied on more com-
surement has been made. Therefore we adopt the usual plicated functions.4,5
procedure: One first simulates an experimental trace The RMS measure provides a natural test of accuracy
corresponding to a test field, including simulations of pos- that is commensurate with the aim of current methods of

0740-3224/2002/051019-11$15.00 2002 Optical Society of America


1020 J. Opt. Soc. Am. B / Vol. 19, No. 5 / May 2002 C. Dorrer and I. A. Walmsley

pulse characterization: reconstruction of the electric lytic signal is usually decomposed as E(t) E(t)
field. The RMS error is, of course, sensitive to both the expi (t) 0t 0. Conversely, in the spectral
intensity and the phase of the reconstructed field, and of- domain the analytic signal is decomposed as E( )
fers a naturally weighted balance between these entities. E( ) expi () t0 0, where t 0 is the delay re-
Moreover, it can be calculated in the time or frequency do- garding a chosen time origin and 0 is the absolute phase.
main and gives the same result in either.
I( ) E( ) 2 is the spectral density, () is the spectral
The paper is organized as follows. We first give gen-
phase, I(t) E(t) 2 is the instantaneous power, and (t)
eral definitions and properties of the RMS error as it re-
is the temporal phase.
lates to the characterization of ultrashort optical pulses.
The RMS field error between two fields E 1 and E 2 is
We also demonstrate its equivalence to the Wigner error,5
defined as the norm of the difference E 1 E 2 :
a more complicated and less computationally efficient cri-
terion. The RMS field error is then applied to studying
the accuracy of pulse reconstruction by spectral phase in-
terferometry for direct electric field reconstruction68
E 1 E 2


dt E 1 t E 2 t 2 1/2

(SPIDER). Two algorithms for the reconstruction of the


pulse shape from the SPIDER experimental data are pre-
sented and discussed, one based on integration tech-
1
2


d E 1 E 2 2 1/2
. (6)

niques, the other based on the reconstruction of multiple As a consequence of Parsevals theorem, the fidelities of
sets of the electric field sampled according to the Shannon reconstruction of a pulse in the temporal or the spectral
theorem. The error measure is used to quantify the ac- domain are equivalent, according to this measure.
curacy of SPIDER under simulated experimental condi-
tions and for various parameters with these algorithms. B. Properties
In a companion paper tests of consistency and precision For any pair of normalized fields, ranges between 0 (sig-
will be presented for SPIDER.9 We emphasize that the nifying that the two fields are equal) and 2 (signifying
procedures outlined in these two papers have wider appli- that they are inversions of each other). Let us consider a
cation than the specific example we have chosen. They physical field E and a measured field E . Ideally, E is
can be applied equally well to other methods of pulse equal to E, but because of the limitations of the measure-
measurement. ment technique itself or experimental imperfections,
these fields might differ. The influence of differences be-
tween E and E can be studied by writing the fields in
2. RECONSTRUCTION OF THE ELECTRIC terms of their amplitudes and phases:
FIELD AND THE RMS FIELD ERROR
A. General Definitions E E exp i
The real electric field E(t) of a short optical pulse can be
E E exp i ,
decomposed into its complex analytic signal E(t) by use of
the following relations: (7)

E

dtE t exp i t , (1)
where E( ) and () are small variations of the spec-
tral amplitude and phase of the field. Expanding the
field to first order in these small quantities gives

E t
2
1
0

d E exp i t . E E E

(2) i E exp i . (8)


This temporal representation is related to a frequency The error function can then be calculated directly as


representation by a Fourier transform:
1/2
1
E

dtE t exp i t , (3)

2
d E 2 2 E 2

E 2 E 2 1/2. (9)
and vice versa;


thus depends on the variation of the amplitude differ-
1
E t d E exp i t . (4) ence E and the variation of the phase difference
2 weighted by the spectral density E . The weighting
of the phase difference at frequency by the spectral den-
Either of the representations E or E completely describe
sity at that frequency has a simple physical interpreta-
the pulse. The norm E can be calculated in both do-
mains by use of Parsevals theorem: tion: E( ) 2 is the energy at frequency . The phase
takes an arbitrary value when the spectral density is

E2

dt E t 2
1
2


d E 2 . (5)
zero, so this weighting ensures that large variations of
the phase difference near do not significantly affect the
reconstructed pulse shape. (Experimentally, a low
With the reference (carrier) frequency 0 and the signal-to-noise ratio is expected where the spectral den-
carrier-envelope phase offset (CEO) 0 the temporal ana- sity is low.) This is why all error measures that give use-
C. Dorrer and I. A. Walmsley Vol. 19, No. 5 / May 2002 / J. Opt. Soc. Am. B 1021

ful results use a function of the phase weighted by the representation of ultrashort pulses (Ref. 15 and refer-
intensity.4 The spectral density appears naturally here ences therein). It is defined in terms of the analytic sig-
as the weighting function.10 An identical argument can nal as


be made for the time domain.
In our simulations we consider that values of below
0.02 correspond to very good reconstructions, accuracies
W t,


dt E t
t
2
E* t
t
2
exp i t dt


between 0.02 and 0.1 correspond to average reconstruc-
tions, and accuracies above 0.1 correspond to poor recon-
structions. Such values are calculated for E E
1
2


d E

2
1. We recognize that this choice of delineation is sub-
jective, however.
Most pulse characterization techniques do not measure
the time delay t 0 and the CEO phase 0 that appear in
E *
2
exp i t d . (12)

the spectral domain representation of the field. Mea- This two-dimensional real distribution is attractive for
surement of these quantities is meaningful only if a ref- representation purposes, since it provides some intuition
erence time is available in the experiment. In the case of about ill-defined but appealing notions such as instanta-
the time delay, this might be provided, for example, by an- neous frequency or group delay. The Wigner distribution
other short pulse. Measurement of the second quantity has recently been used as a test of the resemblance of two
requires different approaches from the well-established electric fields and as a test of accuracy of pulse character-
pulse characterization techniques, and new techniques ization techniques.3 We show in Appendix B that this ap-
have only recently been proposed.11,12 The RMS error proach is equivalent to the approach presented here, since
depends, however, on these two quantities, and can be the Wigner error can be derived from the RMS field error,
used to compare directly only the full analytic signals of so that the intensive work associated with direct calcula-
two pulses. In most experimental cases, however, these tion of the Wigner error is not required. Similar conclu-
two quantities are indeterminate, and one wishes to mini- sions can be drawn for error measures involving the two-
mize without regard to their value. An identical con- time correlation function and the ambiguity function.
sideration exists in image reconstruction, as the piston
(constant phase term) and tilt (linear phase term) be-
tween two wavefronts are usually arbitrary or irrelevant 3. TEST OF SPECTRAL PHASE
to the information content of an image.13 The RMS error INTERFEROMETRY FOR DIRECT
function makes this minimization quick and simple, as is ELECTRIC-FIELD RECONSTRUCTION
demonstrated in Appendix A. A. SPIDER
SPIDER is a technique for the characterization of the
C. Equivalence With Other Error Functions electric field of ultrashort optical pulses based on shear-
1. Temporal Phase and Intensity Error Functions ing interferometry in the frequency domain. It relies on
One criterion used to test the accuracy of the frequency- interference between two frequency-sheared replicas of
resolved optical gating technique,4 and later applied to the field. The shear is obtained by nonlinear frequency
SPIDER,14 relies on a temporal intensity error function I conversion of two replicas of the pulse, delayed in time by
and a temporal phase error function , defined as , with a chirped pulse. We consider here an ideal situa-
tion in which the nonlinear conversion is performed for


1/2 each replica with a perfectly monochromatic frequency, 0
I dt I t I t 2 , (10) and 0 , in a nonlinear crystal with infinite band-
width. Practical limitations on the validity of these as-
sumptions were studied previously.3,8


1/2 1/2 We consider in our simulations that the spectrometer
dt t t 2 I 2 t dtI 2 t . used to measure the interferogram has a flat temporal re-
sponse. The primary effect of finite spectral resolution
(11) and pixellation of the detector is to decrease the contrast
of the fringes, thus reducing the signal-to-noise ratio.16
These criteria weight the phase difference between the The phase difference between the two electric fields, (
two fields by the square of the temporal intensity. Al- 0 ) ( 0 ) , is extracted from the in-
though this emphasizes the phase differences at times terferogram by Fourier transform spectral inter-
when the corresponding intensity is large, there is no ferometry.17 The interferogram is directly transformed
simple connection to similar errors defined in the fre- without any interpolation using a fast Fourier transform.
quency domain. (We note, however, that because of the It is then filtered and Fourier transformed back.16 The
Parseval theorem two fields with nearly identical tempo- removal of the linear phase term is performed by use of
ral phase and intensity will have nearly identical spectral an error-free reference phase at the fundamental
phase and intensity.) wavelength.18 The reference phase is considered to be
accurate because it is obtained from an a priori calibra-
2. Wigner Error tion of the device. (The calibration interferogram can
The Wigner distribution, widely used in quantum me- therefore be taken with enough signal averaging to yield
chanics and signal processing, has been applied to the a noiseless trace.) Although the spectrum of the pulse
1022 J. Opt. Soc. Am. B / Vol. 19, No. 5 / May 2002 C. Dorrer and I. A. Walmsley

I( ) can be measured on the same experimental trace,19 function whose average value on every interval of
we consider here a noiseless independent measurement of length is equal to the function that is obtained from
the spectral intensity to analyze the differences in the the experimental data. can be obtained directly from
phase reconstruction. After subtraction of the reference by integration with the use of a trapezoidal approxima-
phase, and a shift along the frequency axis, one obtains tion, noting that is sampled on a frequency interval .
the spectral phase ( ) ( ) ( ). There are a number of different options available for the
Two numerical recipes for the reconstruction of the calculation of the integral on the right-hand side. Find-
electric field from the experimental data I( ) and () ing the function is the inverse problem to the integra-
are explored. Although some reconstruction techniques tion of a function on an interval 0 , 0 by using
in spatial shearing interferometry are based on a modal only values of the function at discrete points. The gen-
approach (i.e., the decomposition of the phase in known eral procedure for the latter problem is to consider linear
modes, such as two-dimensional polynomials defined over combinations of the values of the function in the interval
the pupil of an optical system), it is here more appropriate that ensure the accuracy of the integration to be high for
to use a zonal approach20 (i.e., local reconstruction). In- a particular class of functions. When the class of poly-
deed, the spectral phases of ultrashort optical pulses that nomic functions of order lower than a given integer is cho-
one encounters in a practical situation cannot be easily sen, these combinations are known as Newton-Cotes for-
classified in modes (in contrast to the spatial aberrations mulas, of which the best-known examples are Simpsons
of an optical system that are constrained by the systems or Bodes rules of integration.18 These can be adapted to
symmetries). The techniques presented here rely on the the phase reconstruction problem in the following way.
fact that the pixellated detector at the Fourier plane of First, we write the integral as a linear combination of the
the spectrometer samples the SPIDER interferogram at a unknown values of the phase:


rate that is smaller than the spectral shear . This 0
naturally arises in the experimental implementation of
SPIDER. The spectral sampling of the detector 0
d a
j
j 0 j , (14)

means that the input pulse sequence must lie in the time where a j and j are sets of real numbers between 0
window [/, /]. Therefore the correct arrange- and 1. The right-hand side of this equation can be writ-
ment of the apparatus requires that the delay between ten as a convolution of the (continuous) spectral phase
the two interfering pulses be set so that / . The function with a sum of Dirac functions , so that Eq. (13)
sampling of the spectral phase at frequency intervals is equivalent to
means that any field whose temporal support is included
in [/, /] can be reconstructed faithfully. To
achieve this in SPIDER, the frequency sweep in the
m, where m a .
j
j j (15)
chirped pulse during the interaction with each test pulse
replica must be kept small compared with the spectral From the convolution theorem, this is equivalent to (t)
shear. The frequency sweep is proportional to the tem- (t)m(t), where the tilde represents a Fourier trans-
poral support of the pulse that one is characterizing, form. If m does not have any zeros, then /m, from
while the spectral shear is proportional to the delay be- which can be determined by an inverse Fourier trans-
tween the two interfering pulses. These considerations form.
then give / , i.e., . In practice, the ratio of For given sets a j and j , the reconstruction of the
these two quantities is of the order of ten. phase requires only the following steps: Integration of
the phase difference , which is itself extracted from the
B. Reconstruction of the Spectral Phase in SPIDER interferogram by Fourier processing, to yield (); a fast
There are several different algorithms that can be used to Fourier transform to yield (t); multiplication by the
reconstruct the electric field from the experimental data function 1/m to yield (t) and finally an inverse fast Fou-
I( ) and (). We present two in detail here. rier transform to yield ().
This method provides exact reconstruction of spectral
1. Reconstruction by Integration phases that are polynomic of order N on each interval of
The spectral phase ( ) ( ) ( ) extracted length , provided that the appropriate integration for-
from the interferogram can be integrated between mula is chosen. Specific choices of m are available to
and 0 to give the phase : give exact integration for such classes of functions, as we
show in succeeding paragraphs. Furthermore, such an
0
0
d
integration formula leads to exact reconstruction for spec-
tral phases that are polynomials of order N 2 on the
full spectral support of the pulse.

0
d
However, higher accuracy for a particular class of func-
tions (in this case polynomials of order N 2) does not
mean higher accuracy in the general case of arbitrary
0
0
d . (13)
spectral phase. Also, because the SPIDER phase at
any frequency is the difference between spectral phase at
two frequencies separated by , any phase with periodic-
Reconstruction of the spectral phase of the pulse from the ity is discarded.
SPIDER interferogram is thus equivalent to identifying a We now consider specific integration formulas:
C. Dorrer and I. A. Walmsley Vol. 19, No. 5 / May 2002 / J. Opt. Soc. Am. B 1023

Midpoint rule of integration: The integral is writ-


ten as m t

3

exp i
t
2
* 2 cos
t
2
. (21)

0
d 0
2
. (16) As m does not have any zeroes in the time domain for this
integration formula, one can take its inverse, and this
leads to the unique determination of . In this case, the
The reconstructed phase is then directly obtained as
error is dominated by the fourth order of the phase. Per-
0 1/ 0 /2 , (17) fect reconstruction is obtained for functions that are up to
third-order polynomials on each interval of length or
without the need for any calculation in the Fourier do- that are global polynomials of up to fifth order.
main, since the convolution is done with a single Dirac Higher-order approximations: It turns out that
delta function. higher-order NewtonCotes formulas do not uniquely
Taking an expansion up to second order in frequency of specify the phase , since the function m(t) has some ze-
the function , one can write roes in the time domain. However, it is possible that, in


principle, other linear combinations that provide accurate
0 0
d d 0 integration for functions other than polynomials (such as
0 0 2 Gaussian quadrature formulas) could be used.


0

2
0

2
2. Reconstruction by Concatenation
This algorithm makes use of the Shannon theorem and is


exact for arbitrary phase functions corresponding to fields
1 2
2 satisfying the criterion of this theorem. That is, any elec-
0 f ,
2 2 2 tric field pulse whose temporal support is included in the
interval [/, /] will be perfectly reconstructed. The
(18)
phase is set to zero at a starting point and is concatenated
where the second-order derivative of the phase is taken at every following the formula
an unknown point f( ) in the interval 0 /2 , for
each . This integration yields 0 0,

0
0

d * 0 /2
2
24
O
2
2
,
n 1 n n .

An interpolation of the spectrum completes the measure-


(22)

where O( / ) is an upper bound of the second-order


2 2 ment, leading to the knowledge of the electric field on the
derivative of the phase on the interval 0 , 0 . sampling grid 0, , 2..., (N 1) .
The error in this integration technique is bounded by Because of the oversampling performed by the detector
1/24 2 O( 2 / 2 ), which is small compared with for of the spectrometer, multiple sets of the concatenated
most realistic phase functions. This technique will then phase can be obtained with different starting pixels. For
perform well in most cases. For a parabolic phase, example, one can use a sampling grid shifted by m and
2 / 2 f( ) is independent of and adds only a con- integrating following
stant on the retrieved phase, which does not compromise
the accuracy. For a cubic phase, a linear phase is added m 0,
on the reconstructed phase, so that the accuracy is not
m n 1 m n
compromised. The choice of the midpoint rule of integra-
tion thus gives error-free reconstruction of phase func- m n . (23)
tions that are linear on each interval of length or that
are global polynomials of up to third order on the full The number M of different determinations of the elec-
spectral support of the pulse, whatever the coefficients of tric field is of the order of /. Now, since every field
these polynomials are. with temporal support in the interval t,t is de-
Simpsons rule: In this case, the integral is written scribed by a spectral representation with a sampling in-
as terval smaller than /t, each and every reconstruction


from the SPIDER interferogram exactly represents the

0
0
d

6
0 4 0

2
field. Therefore it is possible to use them as a set of rep-
resentations of a single pulse and to determine the statis-


tical properties of the ensemble, that is, the mean field
and the variance at each spectral or temporal point.
0 . (19) These equivalent determinations cannot be averaged
directly in the spectral domain since they are defined on
One then gets different sampling grids. However, they can be averaged


in the temporal domain by choosing a common sampling
grid 0, T ,..., (N 1)T and calculating the Fourier
m 4 ,
6 2 transform of each spectral representation on that com-
(20) mon sampling grid. This is done by writing
1024 J. Opt. Soc. Am. B / Vol. 19, No. 5 / May 2002 C. Dorrer and I. A. Walmsley

1
E m nT
NJ
E m j
j

exp i m j nT
1
exp imn nT
N
E m j
j

exp i 2 jn
N
, (24)

where the sum can be calculated with a fast Fourier


transform algorithm. (Note that the temporal linear
phase mn T must be added to E m .)
The average field is calculated by use of

1
E nT
M
E
m
m nT . (25)

The temporal fields E m are not shifted in time to optimize


their overlap since they are retrieved from the same ex-
tracted phase () and should have the same time refer-
ence. Their respective absolute CEO phase is, however,
arbitrary since it depends only on the choice of the initial
phase for each set, and this cannot be determined from
the SPIDER phase (). The CEO phase is set to an ar-
bitrary constant at a point where the temporal intensity
is high, although a more time-consuming procedure could
be used to optimize the relative phases (as well as the
temporal overlap). A spectral representation can finally
be obtained by Fourier transforming the temporal repre-
sentation

E n E jT exp
j
i
2 nj
N
. (26)

The reconstructed mean field can be compared with a


field whose spectrum is sampled every by padding the
time-domain representation with zeros before taking the
appropriate Fourier transform. This reconstruction algo-
rithm then reveals whether the apparatus settings gave
an insufficient sampling rate. The reconstructed tempo-
ral electric field should be equal to zero at the edges of the
time window [/, /], since this is the sampling that
is used. If this is not the case, then there are systematic
errors in the SPIDER interferogram.

C. Accuracy of the SPIDER Reconstruction


Fig. 1. (a) Spectrum (solid curve) and spectral phases (dashed
1. Test Pulses curves) used in the simulations. (b) Temporal intensity for the
Phase retrieval with the various algorithms presented in flat phase, the parabolic phase, the cubic phase, and the phase
the previous section has been tested on a benchmark set jump, from top to bottom; in the time domain, the plotted window
of test pulses (see Fig. 1). For each pulse in the set, the is the Nyquist window corresponding to the SPIDER shear.
spectral amplitude is taken to be E( ) exp(
0)/2 with 0 corresponding to 800 nm and 1.1 ps1 . The shear (obtained with a delay line
50 ps1 . This represents a pulse with a Gaussian with quadratic spectral phase equal to 30,000 fs2 and a
temporal intensity of 47 fs at FWHM given a flat spectral delay equal to 1500 fs) is equal to 5 ps1 , i.e., one
phase. The detector at the output of the spectrometer tenth of . The spectral phase structures we have ex-
has 512 pixels and is linearly calibrated in wavelength be- plored are:
tween 750 and 850 nm. The SPIDER interferogram at
400 nm is measured at the second order of diffraction of Flat phase, ( ) 0. This corresponds to the
the grating using the same array calibrated between 375 minimum pulse duration for a given spectrum.
and 425 nm. This provides a sampling interval of Parabolic phase, ( ) (1/2) ( 2 ) ( 0 )/ 2 .
C. Dorrer and I. A. Walmsley Vol. 19, No. 5 / May 2002 / J. Opt. Soc. Am. B 1025

Cubic phase, ( ) (1/6) ( 3 ) ( 0 )/ 3 .


Quartic phase, ( ) 1/24 ( 4 ) ( 0 )/ 4 .
Phase jumps. An infinitely steep phase jump does
not satisfy the Nyquist criterion for finite temporal sup-
port. We choose a spectral phase function of the form
( /2)tanh(20 / ), which does satisfy a Nyquist crite-
rion.

Two classes of phase function were used in these tests:


(1) Polynomial phase functions, where the coefficients
( 2 ) , ( 3 ) , and ( 4 ) were chosen so that the group delay at
the frequency 0 is equal to one third of the
Nyquist limit /, and (2) spectral phases with more com-
plicated forms. In the latter case, these spectral phase
Fig. 2. Difference between the test and the reconstructed
coefficients were chosen so that the group delay at that
phases in the case of the flat phase (three lower curves), the
same frequency is equal to /. This provides a situa- parabolic phase (three middle curves), and the cubic phase (three
tion where significant aliasing is present for a sampling upper curves). In each case, the dashed curve corresponds to I1,
interval . the dotted curve to I2 and the solid curve to the concatenation
We considered the accuracy of the midpoint rule of in- technique.
tegration (I1), the Simpson integration technique (I2) and
the concatenation technique (C).

2. Reconstruction from Noise-Free Experimental Traces


Determining the accuracy of reconstruction for a noiseless
experimental trace addresses the question of the accuracy
of the reconstruction algorithm itself and allows one to
track systematic errors in the apparatus parameters or
the failure of the method. One looks for ambiguous or
badly reconstructed pulse shapes.
An important parameter in the reconstruction algo-
rithm is the filtering of the field autocorrelation compo-
nent in the pseudo-time-domain conjugate to the wave-
length domain in which the interferogram is measured.
This is a key step for Fourier transform spectral interfer- Fig. 3. Difference between the test and the reconstructed
ometry. In the rest of this paper, the filtering in this phases by the concatenation technique (solid curves) and the in-
tegration technique I1 (dashed curves) in the case of the second-
pseudo-temporal domain is performed by use of a rectan- and third-order phases corresponding to fields not satisfying the
gular window with width of 220 pixels, unless otherwise Nyquist limit.
specified. This width does not modify the results in the
absence of noise as long as all the information contained
in the interferometric component is kept. In the pres- forms poorly, as can be seen in Fig. 3. This situation can
ence of noise, this window can be used to remove some be realized if ( 2 ) or ( 3 ) are set so that the group delay at
noise components at pseudo-time points outside the re- 0 is equal to /. The accuracy is of the or-
gion of interest and thus to reduce the overall noise power der of 0.2 in the two cases. It must be pointed out, how-
in the reconstructed trace. An appropriate choice of the ever, that the sinusoidal behavior of the reconstructed
shape of the filter might optimize the accuracy of Fourier field is due to the padding operation performed before
transform spectral interferometry (FTSI) in the presence Fourier transforming the reconstructed temporal field to
of noise, although to our knowledge this has never been get a spectral field sampled every . Visual inspection
thoroughly studied. Reconstruction from noiseless inter- of the reconstructed field on the spectral sampling is
ferograms is excellent, because the retrieval of the field much more satisfying. One must keep in mind, however,
from the experimental trace is algebraic. In contrast to that the latter field has the value zero outside the interval
several other pulse characterization techniques, which re- [/, /], which is not the case for the test field, and
quire iterative algorithms applied to large two- which explains the low accuracy. For the same phase
dimensional sets of data, we require here only two Fou- structure, no significant error is present with the two in-
rier transforms performed on a single one-dimensional tegration techniques, and is smaller than 103 . This
vector. As a consequence, for noise-free interferograms, confirms that integration techniques can reconstruct any
the RMS error is of the order of 104 for the flat phase sufficiently regular function even if the resulting field
and the polynomial modulations for the three algorithms. does not satisfy the Nyquist limit. In the case of a
As can be seen in Fig. 2, this corresponds to very small fourth-order polynomial for which the group delay at
errors in the spectral phase. 0 is equal to /, a systematic parabolic er-
When aliasing is present with the sampling (i.e., if ror is present in the case of integration method I1 (Fig. 4).
the electric field pulse has significant energy outside the However, this error remains very small compared with
window [/, /]) the concatenation technique per- the value of the phase (of the order of tens of radians).
1026 J. Opt. Soc. Am. B / Vol. 19, No. 5 / May 2002 C. Dorrer and I. A. Walmsley

With the higher-order integration technique I2, no such


systematic error is present, resulting in a better accuracy
(3 104 for I2 versus 1.6 103 for I1). Since the re-
sulting field does not satisfy the Nyquist limit, the value
of for the concatenation technique is slightly higher
(4 103 ), but this is small enough to yield good recon-
struction. For I1, perfect reconstruction is obtained for
any polynomial up to third order. For I2, perfect recon-
struction is extended to any polynomial up to fifth order.
For rapidly varying phases, such as a phase jump at
the center of the spectrum, differences between the vari-
ous reconstruction techniques are confirmed (see Fig. 5).
The concatenation technique performs poorly, since alias-
ing is significant. A systematic error is present on the
Fig. 5. Difference between the test and the reconstructed
spectral phase reconstructed by I1. This error is reduced
phases in the case of a phase jump for the concatenation tech-
when I2 is used. The accuracy is equal to 4.2 102 for nique (solid curve), the integration technique I1 (dashed curve)
I2 and 6.5 102 for I1. and the integration technique I2 (dotted curve).
When no aliasing is present, all these techniques give
excellent results. In the presence of rapidly varying
phases that have significant aliasing with sampling ,
the results of the concatenation technique are poor. Not
surprisingly, better results are obtained when the spec-
tral shear is corrected. Integration techniques perform
better even with some undersampling. The low-order
technique I1 tends to smooth out rapidly varying spectral
phase structures. The higher-order technique I2, which
increases the contribution of higher-order polynomials in
the overall phase, reduces this smoothing, leading to a
higher-fidelity reconstruction.

3. Reconstruction in the Presence of Noise


A reconstruction algorithm that performs well for noise- Fig. 6. RMS field error of the reconstruction of a flat phase by
free data is of little use if it falls apart when experimental the integration technique I1 as a function of the noise fraction.
imperfections are present. Thus the influence of noise on The error for I2 and C exactly overlap with the plotted curve.
The higher left-hand inset represents a simulated interferogram
these different algorithms has also been examined.
with a noise fraction of 104 ; the lower right-hand inset repre-
The most common class of noise present in experimen- sents a simulated interferogram with a noise fraction of 101 .
tal signals is pixel-uncorrelated additive Gaussian noise.
This kind of noise describes, for example, the shot noise
and electronic noise due to the pixellated detector at the The maximum amplitude of the uncorrupted, calculated
Fourier plane of the spectrometer. We modeled pulse re- interferogram was set to one, and noise was added. The
construction from interferograms with this type of noise. variance of the noise ranged from 104 (low noise; see up-
per left-hand inset in Fig. 6) to 101 (strong noise; see
lower right-hand inset in Fig. 6). The experimental noise
usually present on a SPIDER interferogram is of the or-
der of 103 , but this value obviously depends on the dy-
namics of the detection and analog-to-digital conversion.
The curves plotted in Fig. 7 were obtained by simulating
interferograms with 100 different noise fractions. For
each noise fraction, ten interferograms were calculated by
using different random noise in each case, and recon-
structions were performed on each. The statistics of
these reconstructions were used in determining the accu-
racy of the reconstruction according to the RMS-error
measure.
The influence of the noise on the accuracy for the inte-
gration technique I1 is summarized in Fig. 6. The three
Fig. 4. Difference between the test and the reconstructed techniques were found to be equally sensitive to noise
phases in the case of a fourth-order polynomial phase corre- when there is no systematic error for the reconstruction of
sponding to a field not satisfying the Nyquist limit for the con- the phase. The phases reconstructed by the integration
catenation technique (solid curve), the integration technique I1
(dashed curve) and the integration technique I2 (dotted curve).
technique I1 are plotted in Fig. 7 for different levels of
The small systematic error present when I1 is used disappears noise. The corresponding values of are 8 103 ,
when I2 is used. 3 102 , 102 , and 4 101 . Note that the phase is
C. Dorrer and I. A. Walmsley Vol. 19, No. 5 / May 2002 / J. Opt. Soc. Am. B 1027

recovered accurately where the spectral density is high, techniques perform some averaging, thus naturally can-
as expected, since the Fourier-transform spectral- celing the effect of high frequency noise.
interferometry-signal amplitude used to extract the spec- It is also useful to calculate the influence of the shear
tral phase difference at a frequency is proportional to on the accuracy in the presence of noise. The noise n( )
the spectral density at that frequency. The spectral on the extracted phase () is integrated to a random
0
phase degrades toward the leading and trailing edges of phase component 1/ d n( ) on the phase () (as-
the pulse, where the signal-to-noise ratio decreases. suming the low-order integration method). If the shear
When the reconstruction accuracy is low even for noise- is halved, the random phase is multiplied by 2, and so is
free interferograms, is first influenced by the systematic . This effect is confirmed by our simulations. The
errors at low levels of noise and then by the noise itself as shear should therefore be set as large as possible to re-
this reaches a sufficient level. Reconstruction of the duce the influence of noise, as long as it remains small
phase jump (see Fig. 8) clearly demonstrates this. In this enough that the Nyquist criterion is satisfied.
case, the high-order integration I2 is still the best choice, We have shown that the several reconstruction algo-
since it provides the best noise-free reconstruction and is rithms for SPIDER are equally sensitive to the influence
less sensitive to noise than the other techniques. of random noise on the interferogram. Systematic errors
Optimal filtering increases the reconstruction accuracy, compromise the accuracy even at low noise levels.
because filtering the interferometric component in the Therefore the technique that performs best for noise-free
Fourier domain removes the noise components outside reconstruction is the most useful tool in the general case.
the temporal range of interest. For example, in the case In all the cases we have tested, the higher-order integra-
of the flat phase, taking a 40-pixel window for the rectan- tion technique, based on the Simpsons rule of integration,
gular filter (instead of the 220 pixels used in the simula- gives the best results, although the lower-order technique
tions presented thus far) reduces by a factor of 1.25. also performs well. We have also shown that these tech-
One reason for this relatively small gain is that the noise niques naturally reduce the effect of noise on the pulse re-
after filtering scales as the square root of the number of construction and that a larger shear diminishes the effect
pixels of the amplitude filter (for uncorrelated, white of noise.
noise). Another reason is that all the reconstruction
4. Comments
It is striking that integration provides error-free recon-
struction of the spectral phase even when the temporal
electric field does not satisfy the Nyquist criterion. In
some sense, assuming that a phase is polynomial of a
given order modifies the number of parameters needed for
a complete specification of the field (for example, a poly-
nomial of order N is completely determined by its value at
N 1 points). However, one should note that, because
the information is recorded as a phase difference between
frequencies separated by , any periodic function with
this period is discarded. (This class of function is inaccu-
rately reconstructed by the concatenation algorithm.)
Thus higher accuracy is obtained by integration, except
for the class of functions with such periodicity. It is
Fig. 7. Reconstructed phases for a flat test phase for levels of therefore important to set the shear to sample the spec-
noise of 104 , 103 , 102 , and 101 , from bottom to top. The
corresponding values of the RMS field error are respectively trum correctly. Finally, we note that most linear optical
8 103 , 3 102 , 102 and 4 101 . systems used with ultrashort optical pulses have time-
shift-invariant transfer functions with relatively smooth
phases. (Examples of such include a compressor, a
stretcher, or propagation in a dispersive material.)
Therefore, integration using the polynomial approxima-
tion on the shear interval is the optimal reconstruction al-
gorithm.

4. CONCLUSION
We have presented the RMS-field-error technique as an
efficient test of accuracy of ultrashort pulse reconstruc-
tion techniques. This criterion is simpler than others
that have been proposed, provides for easy interpretation,
and can be calculated quickly. The application of this cri-
terion to SPIDER has enabled us to test several algo-
Fig. 8. Accuracy of the reconstruction of the phase jump by the
integration technique I1 (solid curve), the integration technique
rithms for the reconstruction of the spectral phase from
I2 (dotted curve), and the concatenation technique (dashed the interferogram. The concatenation technique is based
curve) as a function of the noise fraction. on the use of the sampling theorem and provides exact re-
1028 J. Opt. Soc. Am. B / Vol. 19, No. 5 / May 2002 C. Dorrer and I. A. Walmsley

construction as long as the shear satisfies the Nyquist cri-


terion. An additional small assumption regarding the lo-
cal regularity of the phase allows the use of more
dW E, E
2
1


dtd W t,

powerful integration techniques with numerical accuracy


comparable to concatenation. These methods are based
on the midpoint rule of integration and Simpsons for-
W t, 2 1/2
. (B1)

mula. Since the sensitivity to the noise on the experi- One can develop the Wigner error as:
mental data is the same for all three techniques, our re-
sults suggest that this latter integration technique is the
best choice for reconstruction in spectral shearing inter- dW E, E 2
2
1


dtd W t, 2
ferometry. All these techniques also allow a test of the
consistency of the reconstructioni.e., how well the re- 2W t, W t, W t, 2 .
constructed field matches the experimental dataas well
(B2)
as the precisioni.e., how similar are the various fields
19
extracted from the same redundant experimental data. Using the Moyals formula


The precision and consistency of pulse reconstruction us- 1
ing SPIDER is explored in the companion paper.9 dtd W t, W t,
2

APPENDIX A: MINIMIZATION OF THE RMS


ERROR FUNCTION REGARDING THE


dtE t E * t , 2

ABSOLUTE PHASE AND TIME DELAY this can be simplified as


For normalized fields, the square of the RMS field error is 2 2
dW E, E 2 dt E t 2 2 dtE t E * t

E E
2 2 2
1
2


d E E * c.c.


dt E t 2 . 2

(B3)

2 2 Re
1
2



d E E * . (A1) Thus for normalized fields

Minimization of this quantity is equivalent to the maxi-


mization of Re
d E( )expi( 0 t0)E*(), and
dW E, E 2 2 2


dtE t E * t . 2
(B4)

one has Following Appendix A, if the absolute phases of the


fields are free parameters, the RMS field error is


1
max 0 ,t 0 Re d E exp i 0 t 0 E * 1 1/2
2 22 d E E *
2

maxt 0
1
2



d E exp i t 0 E * (A2) 22

dtE t E * t 1/2
. (B5)


since 0 arg d E( )E * ( )exp(it0) maximizes Combining Eqs. (B4) and (B5) gives
the integral.
dW E, E 2 2 2 1 1/4 2 . (B6)
As the integral on the right-hand side is the Fourier
transform of the quantity u( ) E( )E * ( ), one gets The Wigner error can be directly calculated from the
2 2 * 1 maxt0u(t0), where u is the Fourier trans- RMS field error, with is a large gain in simplicity and
form of u. Minimization of the error can then be per- speed. The functional shape of Eq. (B6) proves that
formed by calculating u( ) and finding the maximum of these norms are equivalent (in other words, resemblance
its Fourier transform u. Because one uses a discrete rep- of two fields according to one of these criteria is equiva-
resentation of fields in the spectral and temporal domain, lent to resemblance according to the other criterion).
one can either pad the data in the spectral domain or fit There is thus no particular advantage of the Wigner cri-
the maximum of u(t) to get a better estimate of the terion for the description of the accuracy of the recon-
maximum.13 These two procedures were found neces- struction, as long as an analytic signal representation is
sary only for accuracy better than 103 . available. This is also true for the ambiguity function
and the two-time or two-frequency correlation function.

C. Dorrer is now with Bell Laboratories, Lucent Tech-


APPENDIX B: EQUIVALENCE OF THE nologies, 101 Crawfords Corner Road, Holmdel, NJ
WIGNER ERROR AND THE RMS FIELD 07733. He can be reached by e-mail at dorrer@lucent.com.
ERROR I. Walmsley is now with the Clarendon Laboratory, Uni-
For two fields E and E , with respective Wigner functions versity of Oxford, Parks Road, Oxford, OX1 3PU, UK.
W and W , the Wigner error is defined as: His e-mail address is walmsley@physics.ox.ac.uk.
C. Dorrer and I. A. Walmsley Vol. 19, No. 5 / May 2002 / J. Opt. Soc. Am. B 1029

REFERENCES and U. Keller, Techniques for the characterization of sub-


10-fs optical pulses: a comparison, Appl. Phys. B 70, S67
1. M. Born and E. Wolf, Principles of Optics 6th ed. (Cam- S75 (2000).
bridge U. Press, Cambridge, UK, 1980). 11. P. Dietrich, F. Krausz, and P. B. Corkum, Determining the
2. V. Wong and I. A. Walmsley, Linear filter analysis of meth- absolute carrier phase of a few-cycle laser pulse, Opt. Lett.
ods for ultrashort pulse shape measurements, J. Opt. Soc. 25, 1618 (2000).
Am. B 12, 14911499 (1995). 12. M. Kakehata, H. Takada, Y. Kobayashi, K. Torizuka, Y.
3. A. Eskicioglu and P. Fisher, Image quality measures and Fujihira, T. Homma, and H. Takahashi, Single-shot mea-
their performance, IEEE Trans. Commun. 43, 29592965 surement of carrier-envelope phase changes by spectral in-
(1995). terferometry, Opt. Lett. 26, 14361438 (2001).
4. D. N. Fittinghoff, K. W. Delong, R. Trebino, and C. L. Lad- 13. J. R. Fienup, Invariant error metrics for image reconstruc-
era, Noise sensitivity in frequency-resolved-optical-gating tion, Appl. Opt. 36, 83528357 (1997).
measurements of ultrashort optical pulses, J. Opt. Soc. 14. M. E. Anderson, L. E. E. de Araujo, E. M. Kosik, and I. A.
Am. B 12, 19551967 (1995). Walmsley, The effects of noise on ultrashort-optical-pulse
5. S. Yeremenko, A. Baltuska, M. S. Pshenichnikov, and D. A. measurement using spectral phase interferometry for di-
Wiersma, The criterion of pulse reconstruction quality rect electric field reconstruction, Appl. Phys. B 70, S85
based on Wigner representation, Appl. Phys. B 70, S109 S93 (2000).
S117 (2000). 15. J. Paye, The chronocyclic representation of ultrashort light
6. C. Iaconis and I. A. Walmsley, Spectral phase interferom- pulses, J. Opt. Soc. Am. B 28, 22622273 (1992).
etry for direct electric field reconstruction of ultrashort op- 16. C. Dorrer, N. Belabas, J. P. Likforman, and M. Joffre, Spec-
tical pulses, Opt. Lett. 23, 792794 (1998). tral resolution and sampling issues in Fourier-transform
7. L. Gallmann, D. H. Sutter, N. Matuschek, G. Steinmeyer, spectral interferometry, J. Opt. Soc. Am. B 17, 17951802
U. Keller, C. Iaconis, and I. A. Walmsley, Characterization (2000).
of sub-6-fs optical pulses with spectral phase interferom- 17. L. Lepetit, G. Cheriaux, and M. Joffre, Linear techniques
etry for direct electric field reconstruction, Opt. Lett. 24, of phase measurement by femtosecond spectral interferom-
13141316 (1999). etry for applications in spectroscopy, J. Opt. Soc. Am. B 12,
8. C. Dorrer, B. de Beauvoir, C. Le Blanc, S. Ranc, J. P. Rous- 24672474 (1995).
seau, P. Rousseau, and J. P. Chambaret, Single-shot real- 18. C. Dorrer, Influence of the calibration of the detector on
time characterization of chirped-pulse amplification sys- spectral interferometry, J. Opt. Soc. Am. B 16, 11601168
tems by spectral phase interferometry for direct electric- (1999).
field reconstruction, Opt. Lett. 24, 16441646 (1999). 19. C. Dorrer, Implementation of spectral phase interferom-
9. C. Dorrer and I. A. Walmsley, Precision and consistency etry for direct electric-field reconstruction with a simulta-
criteria for spectral phase interferometry for direct electric- neously recorded reference interferogram, Opt. Lett. 24,
field reconstruction, J. Opt. Soc. Am. B 19, 10301038 15321534 (1999).
(2002). 20. M. Murty, Lateral shearing interferometers in Optical
10. L. Gallmann, D. H. Sutter, N. Matuschek, G. Steinmeyer, Shop Testing, D. Malacara, ed. (Wiley, New York, 1978).

S-ar putea să vă placă și