Abstract
Compressive sensing (CS) of sparse gigahertz-band RF signals using microwave photonics may achieve better performances with smaller size, weight, and power than electronic CS or conventional Nyquist rate sampling. The critical element in a CS system is the device that produces the CS measurement matrix (MM). We show that passive speckle patterns in multimode waveguides potentially provide excellent MMs for CS. We measure and calculate the MM for a multimode fiber and perform simulations using this MM in a CS system. We show that the speckle MM exhibits the sharp phase transition and coherence properties needed for CS and that these properties are similar to those of a sub-Gaussian MM with the same mean and standard deviation. We calculate the MM for a multimode planar waveguide and find dimensions of the planar guide that give a speckle MM with a performance similar to that of the multimode fiber. The CS simulations show that all measured and calculated speckle MMs exhibit a robust performance with equal amplitude signals that are sparse in time, in frequency, and in wavelets (Haar wavelet transform). The planar waveguide results indicate a path to a microwave photonic integrated circuit for measuring sparse gigahertz-band RF signals using CS.
© 2016 Optical Society of America
High-resolution, Nyquist rate sampling of gigahertz-band RF signals rapidly generates huge amounts of data. Compressive sensing (CS) has been developed to address this general issue for sparse signals and images [1–3]. In CS systems, a sparse input signal (dimension ) is recovered from a measurement vector (dimension ) with . The vector is obtained from after multiplication by a measurement matrix , as in Eq. (1):
where is a sparse vector with a small number of nonzero elements and is the transform that shows the sparsity of . If the measurement matrix (MM) satisfies certain properties [1–3], sparse can be recovered by a range of algorithms provided that is somewhat greater than [1,3]. The MM and the matrix multiplication are performed in the analog domain, and an important issue for CS systems is finding a practical way to do this. Since CS recovery calculations require accurate knowledge of [4], it is also mandatory that be reproducible and amenable to calibration.Both electronic [5,6] and microwave photonic CS systems [4,7–16] have been demonstrated that recover sparse RF signals. In the gigahertz band, electronic CS systems suffer from the same sources of error as Nyquist rate analog-to-digital converters (ADCs), timing jitter and amplitude noise. Photonic CS systems have the equivalent of timing jitter and amplitude noise, but in many cases, the distortions are static or lower in frequency and amenable to calibration. Previous photonic systems have used pseudo-random bit sequences (PRBS) for the MM and modulated them on optical carriers with light valves [7–9] or optical modulators [10–16]. Here we demonstrate that the propagation of an optically chirped signal through a multimode optical fiber or planar waveguide performs the function of a CS MM.
The CS system proposed here in Fig. 1 uses a multimode waveguide to replace the two-dimensional spatial light modulator in earlier work [8, Fig. 1] and performs the function of the PRBS in the modulated wideband converter [6, Fig. 3]. Pulses from a femtosecond mode-locked laser (MLL) pass though a dispersion-compensating fiber (DCF) or other dispersive device with dispersion chosen to stretch the pulse to the interpulse time, pass through a Mach–Zehnder modulator (MZM) that impresses the RF signal on the optical intensity, enter a multimode waveguide, and finally are split spatially at the output of the guide and directed to an array of photodiodes. The integration times of the photodiodes are matched to the MLL period, and the electrical signals from the photodiode array are digitized by an array of ADCs clocked to the MLL pulse repetition frequency (PRF). Optical pulse compression, by placing after the MZM a dispersive element of opposite sign to the DCF, can be used to facilitate signal integration [14]. The components on the left and the right of the multimode waveguide are similar to earlier works [7–9], and again we exploit time-wavelength mapping as depicted by the rainbow-colored pulse icons. At the output of the multimode waveguide are formed speckle patterns that vary with the wavelength (and hence the time via the time-wavelength mapping of the MLL plus DCF combination), and small changes in the wavelength can give completely different patterns after modest propagation distances [17]. For example, Fig. 2 shows speckle patterns at the end of a 1 m-long, 105 μm-core diameter, 0.22 NA step-index fiber observed at 1539.44 and 1539.52 nm. An optional fiber mode scrambler (Newport Corporation model FM-1) is used near the input end of the fiber to fully excite the fiber modes. Without the mode scrambler, of multimode fiber is required to realize speckle patterns of similar spatial-frequency content. As seen in Fig. 3, wavelength scans at 4 different locations within the output image of the fiber appear random and uncorrelated. Each of these wavelength scans would correspond to a row of the MM , with the sampling in time set by the time-wavelength mapping property. (In generating Fig. 3, speckle pattern images were recorded while sweeping the wavelength of a single-frequency tunable laser. The combination of the camera frame rate and laser sweep rate provided a wavelength resolution of 0.02 nm.)
Using the grid of red dots in Fig. 2 as locations of the output photodiode array, we derive an MM for the multimode fiber from the speckle pattern images as a function of the wavelength. For apertures, this yields 112 measurements of the optical intensity as a function of the wavelength, of which 4 are shown in Fig. 3, and an MM with dimensions .
We performed several tests to test the measured speckle MM for application to CS. First, we tried to recover several different types of sparse signals: sparse in time (identity transform, ), sparse in frequency (discrete cosine transform), and sparse after the Harr wavelet transform. Figure 4 shows the probabilities that the RF signals are recovered as a function of the small dimension of the MM. A signal is classified as recovered if all of its unknown frequencies, pulse locations, or Haar coefficients are recovered and the amplitudes are recovered to better than 1 part in 10,000. For each basis, the sparse vector consists of equal-amplitude numbers randomly placed on a 2048-point grid, and each curve is for 100 realizations of the sparse vector . We used a standard LASSO code [18] to recover the vector and varied the small dimension of the MM by stripping rows off the . The results are consistent with the well-known formula for the minimum dimension of the MM, [2,3], and the measured speckle MM works well for sparse signals in all three bases. Note that the pulses with the identity matrix for are recovered somewhat better than the sinusoids or Haars. Apparently, the speckle MM is more correlated with the Haar and cosine basis vectors than with the identity basis.
A second test involves the coherence between the rows of the MM , which ideally should be uncorrelated with each other so that each component of is an independent measurement of the input [19]. This can be quantified with the normalized mutual coherence :
for all . Figure 5 (left) overlays the 12,432 coherences calculated from the measured multimode-fiber speckle MM, and for comparison, Fig. 5 (right) shows the coherences for an MM composed of Gaussian random numbers with the same mean and standard deviation as elements of the speckle MM. The mutual coherence for the measured speckle MM is a bit broader than the random matrix, but the measured MM has very good coherence. (Note that many CS MMs use positive and negative numbers as opposed to the positive numbers used here. An MM with positive and negative numbers can be achieved in our speckle system by subtracting photodiode signals from one another as suggested in another context by Ref. [12]. Should more photodiode signals be needed to reach rows, the output of the MZM in Fig. 1 can be split and input into an additional multimode waveguide and photodiode array.)A third well-known test for a CS MM is the restricted isometry property (RIP) [1–3]. Unfortunately, proving RIP appears to be computationally intractable. A surrogate to proving RIP suggested by Ref. [20] is performing numerical experiments to determine if the MM produces a recovery “phase transition.” The “phase transition” is seen as a sharp boundary between regions of high probability of recovery and regions of low probability of recovery in a three-dimensional plot of probability as a function of and . Figure 6 shows the phase transitions for the multimode-fiber speckle MM and the random MM for sparse signals in time, and the difference is minimal. CS MMs must be equally effective against all possible bases in which an input signal may be sparse, and this occurs when the rows of the MM are uncorrelated with the bases of interest. It is not possible to test all possible bases, but we have also performed CS recovery calculations with sparse signals in cosines and Haars with the results shown in Fig. 7. It can be seen that the measured fiber MM yields very similar sharp phase transitions when used with all three bases.
Analytical solutions for the modes of multimode waveguides [21–23] can be used to calculate speckle patterns and MMs. For a cylindrical guide like our fiber (105-μm core, 1-m long, 0.22 NA) with equal power in each mode to approximate the mode-scrambler, rows of the calculated MM look qualitatively the same as for the measured MM in Fig. 3. For planar silicon-on-insulator (SOI) guides with equal power per mode at the entrances of the guides, we similarly calculate output speckle patterns as functions of wavelength. Figure 8(a) shows output 1D speckle patterns for two closely spaced wavelengths for several 10-cm long guides, and Fig. 8(b) shows speckle patterns for a band of wavelengths for several 25.4-μm wide guides. Based on Fig. 8, we select a guide and calculate its MM. MMs for both the planar and cylindrical waveguides are used in CS simulations to recover signals sparse in time. Figure 9 (left) shows phase transition plots for the calculated fiber MM that are nearly the same as those for the measured MM in Fig. 6, and Fig. 9 (right) shows similar results for the calculated planar waveguide MM.
The practical use of speckle in a multimode waveguide for CS requires that the speckle MM, the MLL pulse, and any other dispersion in the system be stable and amenable to calibration. Previous work [8,9,14,15] has found that commercial mode-locked lasers were sufficiently stable for CS. The stability of speckle from a multimode waveguide depends on control of the temperature and mechanical stresses (e.g., bending of the fiber). In our laboratory measurements with the multimode fiber, basic precautions were taken to ensure stable speckle patterns, namely securing the fiber from perturbations. Figure 10 shows the match between two wavelength calibrations taken more than 1 h apart (using a tunable single-frequency laser) for a single row of an MM obtained with a single photodiode placed in the image plane of the multimode fiber output. Other work using speckle in a similar 1 m, 105 μm, 0.22 NA fiber for spectroscopic applications discusses the stability of the speckle pattern in detail [17], Sections 8 and 9, and states that “for a 1 m long fiber, the temperature would need to change by to decorrelate the speckle pattern.”
A potential factor limiting the RF bandwidth of a speckle-based CS system is that the frequency content of the RF signal itself may modify the speckle pattern. At an optical wavelength of 1550 nm, should the speckle pattern vary on a 0.05 nm wavelength scale, the MM will be frequency dependent for RF signals with frequency content above 6.24 GHz. However, the speckle MM can still be used for CS if the system is calibrated by measuring the response for all basis vectors in which the RF signal is sparse. Referring to Eq. (1), the calibration consists of measuring for each basis vector . This matrix, in which is now the number of possible basis vectors in the RF signal, can be used as the dictionary to recover the signal using an orthogonal matching pursuit recovery algorithm [9]. The variation of the speckle MM with RF frequency suggests that it may be possible to measure RF signals modulated on a stable single-frequency laser directly from the change in speckle pattern, a subject for future investigation.
To conclude, we show that optical speckle in multimode fibers and planar waveguides satisfies 3 tests for a compressive sensing measurement matrix: (1) CS simulations show the expected recovery as a function of the number of measurements, (2) rows of the speckle MMs show coherence properties similar to an MM formed from Gaussian random numbers, and (3) recovery plotted in the sparsity/measurement plane () shows sharp phase transitions for all measured and calculated speckle MMs and for 3 classes of sparse signals. The next step is to couple an array of photodiodes and ADCs to the output of the multimode guide and demonstrate a full CS system.
REFERENCES
1. D. L. Donoho, IEEE Trans. Inf. Theory 52, 1289 (2006). [CrossRef]
2. E. J. Candès and M. B. Wakin, IEEE Signal Process. Mag. 25(2), 21 (2008). [CrossRef]
3. R. G. Baraniuk, IEEE Signal Process. Mag. 24(4), 118 (2007). [CrossRef]
4. J. T. Parker, V. Cevher, and P. Schniter, “Compressive sensing under matrix uncertainties: an approximate message passing approach,” in 45th Asilomar Conference on Signals, Systems and Computers, Pacific Grove, California (IEEE, 2011), pp. 804–808.
5. J. A. Tropp, J. N. Laska, M. F. Duarte, J. K. Romberg, and R. G. Baraniuk, IEEE Trans. Inf. Theory 56, 520 (2010). [CrossRef]
6. M. Mishali and Y. Eldar, IEEE J. Sel. Top. Signal Process. 4, 375 (2010). [CrossRef]
7. G. C. Valley and G. A. Sefler, Proc. SPIE 7797, 14 (2010).
8. G. C. Valley, G. A. Sefler, and T. J. Shaw, Opt. Lett. 37, 4675 (2012). [CrossRef]
9. G. C. Valley, G. A. Sefler, and T. J. Shaw, Proc. SPIE 8645, 86450P (2013). [CrossRef]
10. H. Nan, Y. Gu, and H. Zhang, IEEE Photon. Technol. Lett. 23, 67 (2011). [CrossRef]
11. J. M. Nichols and F. Bucholtz, Opt. Express 19, 7339 (2011). [CrossRef]
12. H. Chi, Y. Mei, Y. Chen, D. Wang, S. Zheng, X. Jin, and X. Zhang, Opt. Lett. 37, 4636 (2012). [CrossRef]
13. H. Chi, Y. Chen, Y. Jei, X. Jin, S. Zheng, and X. Zhang, Opt. Lett. 38, 136 (2013). [CrossRef]
14. B. T. Bosworth and M. A. Foster, Opt. Lett. 38, 4892 (2013). [CrossRef]
15. B. T. Bosworth, J. R. Stroud, D. N. Tran, T. D. Tran, S. Chin, and M. A. Foster, Opt. Lett. 40, 3045 (2015). [CrossRef]
16. Y. Chen, X. Yu, H. Chi, S. Zheng, X. Zhang, X. Jin, and M. Galili, Opt. Commun. 338, 428 (2015). [CrossRef]
17. B. Redding, S. M. Popoff, and H. Cao, Opt. Express 21, 6584 (2013). [CrossRef]
18. I. Loris, Comput. Phys. Commun. 179, 895 (2008). (We also used Section 1.1.3 “Lasso” at http://scikit-learn.org/stable/modules/linear_model.html.) [CrossRef]
19. A. Juditsky and A. Nemirovski, Math. Program. 127, 57 (2011). [CrossRef]
20. D. Donoho and J. Tanner, Philos. Trans. R. Soc. London Ser. A 367, 4273 (2009). [CrossRef]
21. D. Gloge, Appl. Opt. 10, 2252 (1971). [CrossRef]
22. D. Marcuse, Theory of Dielectric Waveguides (Academic, 1974).
23. A. Yariv, IEEE J. Quantum Electron. 9, 919 (1973). [CrossRef]