Measuring Ultrashort Laser Pulses III: FROG tricks Dealing with error in FROG measurements Random error (noise) and how to suppress it; error bars N pixels FROG trac e--ex pa nd ed 60 50 40 30 20 10 10 20 30 40 50 60 Delay remove Nonrandom error (systematic error), how toRandomly know when it’s about half of the datapoints from the trace. there, and how to correct for it The FROG marginals Extremely simple FROG beam geometry Rick Trebino, Georgia Tech, rick.trebino@physics.gatech.edu Random and Systematic Error in Pulse Measurement Consider an autocorrelation measurement. Frequency (or phase) Intensity The FROG trace overdetermines the pulse. This has advantages. Advantages: 1. Natural √N averaging occurs, reducing noise. 2. Can perform filtering operations to reduce noise further. 3. Can run algorithm with some points removed to determine error bars in the intensity and phase—independent of the source of noise. 4. Can identify the presence of systematic error—independent of the source. 5. Can remove systematic error—independent of the source. 6. Can understand distortions in the autocorrelation due to systematic error. Noise and its Suppression in FROG 3.00 Frequency (1/pulse width) Fortunately, there are many techniques for suppressing the noise with minimal distortion to the retrieved pulse. Frequency (1/pulse width) Noise can corrupt a FROG trace and yield an incorrect pulse measurement. Without noise 0.00 -3.00 -3.00 With noise 3.00 0.00 -3.00 0.00 3.00 Time (pulse widths) -3.00 0.00 3.00 Time (pulse widths) 1. Background subtraction The FROG trace should be an island in a sea of zeroes. Otherwise, data are missing. So we can subtract off any background. 2. Corner suppression No data should be in the corners of the trace; what’s there can only be noise, so set it to ~zero by multiplying by exp(-r4/d4). 3. Low-pass filtering Noise varies from pixel to pixel, that is, with a high frequency. The FROG trace has only slower variations. Fittinghoff, et al., JOSA B, 12, 1955 (1995). Noise and its Suppression in FROG: Example 0.00 Normalized Intensity 3.00 Intensity 1.2 0.8 0.4 0 -40 -5 -20 -2.5 00 20 2.5 40 5 Time (pulse widths) -3.00 -3.00 0.00 3.00 Time (pulse widths) This pulse has a narrow glitch in its intensity vs. time, and it has a phase jump of ~2 radians, a difficult feature to reproduce. We’ll add noise to this trace. Phase 16 Phase (radians) Frequency (1/pulse width) FROG trace for a complex pulse: 12 8 4 0 -40 -5 -20 -2.5 00 20 2.5 Time (pulse widths) 40 5 Corrupting a FROG Trace with Noise 3.00 Frequency (1/pulse width) Frequency (1/pulse width) Adding 10% additive noise turns this clear trace into this mess: 0.00 -3.00 -3.00 0.00 3.00 Time (pulse widths) 3.00 0.00 -3.00 -3.00 0.00 3.00 Time (pulse widths) (Noise is Gaussian distributed with a mean of 10%.) Note the resulting large background in the noisy trace. Noise in the FROG trace can yield a noisy retrieved intensity and phase. 3.00 -5 -2.5 0 2.5 5 Time (pulse widths) 0.00 1.2 -3.00 -3.00 Normalized Intensity Frequency (1/pulse width) Background at large delays yields wings in the intensity. Background at large frequency offsets yields noise in those wings. Actual Retrieved 0.8 0.00 3.00 0.4 Time (pulse widths) The retrieved pulse is very noisy! 0 It looks nothing like the actual pulse. -5 -2.5 0 2.5 Time (pulse widths) 5 Frequency (1/pulse width) Subtracting off the background improves the retrieved intensity and phase. FROG trace with 10% additive noise—after subtracting the mean of the noise 3.0 0.0 -5 -2.5 0 2.5 5 Time (pulse widths) -3.0 -3.0 0.0 3.0 Time (pulse widths) Note the suppression of the wings and of the noise in the wings of the pulse. -5 -2.5 0 2.5 Time (pulse widths) 5 Suppressing the corners of the trace also improves the retrieved intensity and phase. Frequency (1/pulse width) FROG trace with 10% additive noise—after subtracting off the background mean and suppressing the corners 3.0 -5 0.0 -2.5 0 2.5 5 Time (pulse widths) -3.0 -3.0 0.0 3.0 Time (pulse widths) Trace was multiplied by a “superGaussian”: exp(-r4/d4), where r = distance from trace center. -5 Note further improvement in the wings. -2.5 0 2.5 Time (pulse widths) 5 Wavelength (1/pulse width) Low-pass filtering further improves the retrieved intensity and phase. FROG trace with 10% additive noise after mean subtraction, super-Gaussian filtering and lowpass filtering 3.0 -5 0.0 -2.5 0 2.5 5 Time (pulse widths) -3.0 -3.0 0.0 3.0 Time (pulse widths) Fourier-transforming the trace, retaining only the center region, and transforming back. The resulting intensity and phase now look very much like the actual curves! -5 -2.5 0 2.5 Time (pulse widths) 5 Filtering summary: Always do it! Without filtering With filtering Intensity: -5 -2.5 0 2.5 5 Phase: -5 -2.5 0 2.5 Time (pulse widths) 5 -5 -2.5 0 2.5 Time (pulse widths) 5 Dramatic improvements in the retrieval occur with little distortion. After filtering,10% additive noise yields ~1% error; even less with multiplicative noise. We can place error bars on the retrieved intensity and phase using the “Bootstrap” method. N pixels Frequency (or phase) Intensity Frequency Frequency FROG trac e--ex pa nd ed 60 50 40 30 20 10 10 20 30 40 50 60 Delay Randomly remove about half of the datapoints from the trace. Munroe, et al., CLEO Proceedings, 1998. Press, et al., Numerical Recipes Intensity Calculate the mean and standard deviation of the intensity and phase (or frequency) for each time. Frequency (or phase) Repeat the above procedure several times, removing different points each time. Time Error Bars in the Intensity and Phase Using the Bootstrap Method—Theory Introducing 1% additive noise to the FROG trace: Phase Analytic intensity Retrieved intensity with noise Time (arb. units) Errors in the intensity are similar everywhere (slightly larger at the peak). Because the noise was additive, noise exists in the wings also. 4 Phase (radians) Intensity (arb. units) Intensity Analytic phase Retrieved phase with noise 2 0 -2 -4 Time (arb. units) Errors in the phase are much larger in the wings, where the intensity is near-zero and the phase is necessarily undefined. Error Bars in the Intensity and Phase Using the Bootstrap Method—Exp’t In practice, SHG FROG traces have mostly multiplicative noise: Phase Phase (radians) Intensity (arb. units) Intensity 2 1 0 -1 -2 -3 -400 -200 0 200 400 Time (fs) Errors in the intensity are much larger at the peak. Because the noise was multiplicative, there is almost no noise in the wings. -400 -200 0 200 400 Time (fs) The phase error is low, except in the wings, where, as before, the intensity is near-zero and the phase is necessarily undefined. Sources of Systematic Error in FROG Source: Check? Correct? Variation in spectral response of optics √ √ Variation in spectral response of camera √ √ Dispersion of nonlinearity √ √ Group-velocity mismatch/phase-matching bandwidth √ √ Variable alignment of beam overlap √ Unknown √ Possibly! It is possible, not only to check for systematic error, but also to correct it in most pulse measurements using FROG, even when its origin in unknown. Geometrical time-smearing could yield systematic error. Avoiding Geometrical Time-Smearing Wavelength-dependent SHG phase-matching efficiency yields systematic error. Group-velocity mismatch yields a wavelength-dependent SHG efficiency. Usually, it’s a sinc2 curve, but even when two such curves fortuitously overlap, there’s wavelength-dependent SHG efficiency: Phase-matching efficiency vs. wavelength Phasematched wavelength 60-µm thick KDP crystal It’s impossible to achieve the desired flat curve. Taft, et al., J. Selected Topics in Quant. Electron., 3, 575 (1996) Even very thin SHG crystals may lack sufficient bandwidth for a 10-fs pulse. The FROG Marginals M ( ) 50 40 30 20 10 10 I FR OG ( , ) d The frequency marginal is the integral of the FROG trace over all delays: It is a function of frequency only. M ( ) Frequency The delay marginal is the integral of the FROG trace over all frequencies. It is a function of delay only. SHG FROG trace--expanded 60 I F ROG ( , ) d The marginals are essential in checking for systematic error. 20 30 40 50 60 Delay The FROG marginals can be related to easily measured quantities: M ( ) The Autocorrelation M ( ) The Autoconvolution of the Spectrum DeLong, et al., JQE, 32, 1253 (1996). Applications of the SHG FROG Marginals Taft, et al., J. Selected Topics in Quant. Electron., 3, 575 (1996) Correcting for Systematic Error: Example Wavelength (nm) Attempts to measure a ~10-fs pulse produced this trace and pulse: 350 370 390 410 430 450 470 490 -40 FROG trace measured 350 370 390 410 430 450 470 490 -20 0 20 Delay (fs) 40 -40 FROG trace Retrieved pulse retrieved -20 0 20 Delay (fs) 40 Independently measured spectrum Usually, systematic error yields poor convergence. Here, however, despite good convergence, the retrieved spectrum disagrees with the independently measured spectrum. (This is due to insufficient phasematching bandwidth in a 60-µm KDP crystal.) Comparing the FROG frequency marginal with the spectrum autoconvolution 1.2 FROG Marginal Autoconvolution Intensity 1 0.8 0.6 0.4 0.2 0 360 420 450 480 390 Wavelength (nm) 510 Although they should agree, they don’t! This is because the SHG crystal did not phase-match the longer wavelengths of the pulse. Forcing the frequency marginal to agree with the spectrum autoconvolution yields an improved trace. Wavelength (nm) Multiplying the measured FROG trace by the ratio of the spectrum autoconvolution and frequency marginal: 350 370 390 410 430 450 470 490 -40 FROG trace 350 corrected370 390 410 430 450 470 490 measured -20 0 20 Delay (fs) 40 -40 1.2 FROG trace retrieved Retrieved pulse 3 2.5 1 0.8 2 0.6 1.5 0.4 1 0.2 0.5 0 680 765 850 935 0 1020 Wavelength (nm) -20 0 20 Delay (fs) The retrieved spectrum now agrees with the measured spectrum. The spectral phase has also changed. 40 Independently measured spectrum The corrected pulse can now be used in comparisons with theory. The measured and predicted pulse vs. time: Measured pulse Intensity 0.6 0.4 0.8 6 0.6 4 0.4 2 0.2 0.2 0 0 -40 -20 0 Time (fs) 20 8 6 0.6 Intensity Phase 0.4 0 -20 8 0.8 0 Time (fs) 4 Wavelength -40 20 -20 40 0 6 4 2 0.2 0 -40 40 10 2 0 20 40 Time (fs) This pulse measurement verifies that material dispersion is the pulselength-limiting effect in this laser. 0 Phase (rad) Wavelength 10 Phase (rad) 0.8 Intensity Phase 1 8 Phase (rad) Intensity 1 Predicted pulse 1.2 10 Intensity 1.2 Intensity Phase 1 1.2 The FROG marginals can be used to understand the effects of systematic error on autocorrelation measurements. Why pulse autocorrelations can appear narrower when using a thick crystal = phase-matching bandwidth / pulse bandwidth Thick crystal: Thicker crystal: SHG FROG trace Cropped SHG FROG trace Very cropped SHG FROG trace Frequency Thin crystal: Gaussian spectrum w/ cubic spectral phase Delay Delay Delay Autocorrelation Incorrect (thick-crystal) Autocorrelations Correct (thin-crystal) autocorrelation = Delay Thick crystal suppresses wings! It’s difficult to know if the crystal is thin enough! Can we simplify FROG? Pulse to be measured FROG has 3 sensitive alignment degrees of freedom (q,f of a mirror and also delay). The thin crystal is also a pain. Pulse to be measured SHG crystal Variable delay 2 alignment q parameters q (q,f) q Camera Crystal must be very thin, which hurts sensitivity. 1 alignment q parameter q (delay) q Remarkably, we can design a FROG without these components! SpecCamera trometer The angular width of second harmonic varies inversely with the crystal thickness. Suppose white light with a large divergence angle impinges on an SHG crystal. The SH generated depends on the angle. And the angular width of the SH beam created varies inversely with the crystal thickness. Very thin crystal creates broad SH spectrum in all directions. Standard autocorrelators and FROGs use such crystals. Thin crystal creates narrower SH spectrum in a given direction and so can’t be used for autocorrelators or FROGs. Very Thin SHG crystal Thick crystal begins to separate colors. Thin SHG crystal Thick Very thick crystal acts like SHG crystal a spectrometer! Why not replace the spectrometer in FROG with a very thick crystal? Very thick crystal GRating-Eliminated No-nonsense Observation of Ultrafast Incident Laser Light E-fields (GRENOUILLE) Patrick O’Shea, Mark Kimmel, Xun Gu and Rick Trebino, Optics Letters, 2001; Trebino, et al., OPN, June 2001. GRENOUILLE Beam Geometry In GRENOUILLE, the GVM must be large! This is the opposite of the usual condition! In GRENOUILLE, the GVD must still be small. Putting it all together GVM is usually much greater than GVD. Testing GRENOUILLE Compare a GRENOUILLE measurement of a pulse with a tried-and-true FROG measurement of the same pulse: GRENOUILLE Measured: Retrieved: Retrieved pulse in the time and frequency domains FROG Really Testing GRENOUILLE GRENOUILLE Even for highly structured pulses, GRENOUILLE allows for accurate reconstruction of the intensity and phase. Measured: Retrieved: Retrieved pulse in the time and frequency domains FROG Advantages of GRENOUILLE Disadvantages of GRENOUILLE It currently only works for pulses between ~ 40 fs and ~ 300 fs. Like other single-shot techniques, it requires good spatial beam quality. Improvements on the horizon: Inclusion of GVD and GVM in FROG code to extend the range of operation to shorter and longer pulses. Folded beam geometry for even more compact arrangement. Disadvantages of FROG and its relatives FROG requires taking a lot of data. While this can be done easily with a readily available camera, and it allows error checking and correcting, multi-shot FROG measurements can take minutes. The algorithm can be slow, also taking minutes for complex pulses. (There is, however, a new algorithm, based on singularvalue decomposition, which is much faster: < 1 sec.) SHG FROG has an ambiguity in the direction of time. FROG has a few advantages! To learn more, see the FROG web site! www.physics.gatech.edu/frog Or read the cover story in the June 2001 issue of OPN Or read the book!