LETTER

Reconstructing state mixtures from diffraction measurements

Pierre Thibault1 & Andreas Menzel2

doi:10.1038/nature11806

Progress in imaging and metrology depends on exquisite control over and comprehensive characterization of wave fields. As reflected in its name, coherent diffractive imaging relies on high coherence when reconstructing highly resolved images from diffraction intensities alone without the need for image-forming lenses13. Fully coherent light can be described adequately by a single pure state. Yet partial coherence and imperfect detection often need to be accounted for, requiring statistical optics or the superposition of states4,5. Furthermore, the dynamics of samples are increasingly the very objectives of experiments6. Here we provide a general analytic approach to the characterization of diffractive imaging systems that can be described as low-rank mixed states. We use experimental data and simulations to show how the reconstruction technique compensates for and characterizes various sources of decoherence quantitatively. Based on ptychography7,8, the procedure is closely related to quantum state tomography and is equally applicable to high-resolution microscopy, wave sensing and fluctuation measurements. As a result, some of the most stringent experimental conditions in ptychography can be relaxed, and susceptibility to imaging artefacts is reduced. Furthermore, the method yields high-resolution images of mixed states within the sample, which may include quantum mixtures or fast stationary stochastic processes such as vibrations, switching or steady flows. Recent experimental advances have provided cleaner probes to investigate ever more complex systems. However, state mixtures often have to be used to accurately describe the statistical nature of the system under investigation. In the context of imaging, we categorize the occurrence of mixed states into three distinct groups, illustrated in Fig. 1: fluctuations in the probing radiation, commonly grouped under the concept of partial coherence (Fig. 1a), mixed states within the object of interest (Fig. 1b) and detector point spread (Fig. 1c). Although any such source of decoherence affects the measured intensity distribution in a specific way, they all share the common signature of reducing the so-called fringe or speckle visibility in the diffracted intensities (Fig. 1d). Decoherence is often an unwanted complication, and much experimental effort is spent in reducing its impact. The need for coherence effectively imposes severe constraints on the setup, such as a highbrilliance source, negligible vibrations of all components, a static sample or ultra-short pulses and often a drastic reduction of usable flux. At hard X-ray synchrotron sources, for instance, the necessary coherence filtering reduces flux by several orders of magnitude. In the past few years, progress has been made to accommodate more permissive conditions. In some instances, imaging with a partially coherent wavefront can be cast as a blind deconvolution problem9. However, preconditions for this approach are frequently unsatisfied. Until now the general problem of how to mitigate decoherence in data analysis has largely remained unexplored. Here we show that stationary mixed states of any origin can be reconstructed by using ptychography. Used with X-rays10, electrons11 and visible light12, ptychography is a coherent diffractive imaging technique that allows the simultaneous reconstructions of the incident
1

wavefront and the sample transmission. Producing a ptychogram requires a sequence of diffraction measurements in which the object is scanned through a spatially confined illumination, or probe. The probe is assumed constant, and the specimens interaction with it is assumed to be adequately described by a multiplicative transmission function. Consequently, a fully coherent ptychography measurement can be written as a set of simple products of probe and object. Sampling takes place in reciprocal space, in which the phase information in the scattered field is discarded and only intensities I are recorded as a function of the momentum transfer q. We show in the Supplementary Information how such a measurement can be recast in probabilistic terms by following von Neumanns analysis of quantum mechanics13. Thus, the expectation value of any observable can conveniently be expressed in terms of an operator, here ^ Ijq , where j indexes the position in the ptychographic scan, and a density operator, usually represented by a density matrix r: I jq r Ijq ~Tr ^ 1

where Tr denotes the trace operator. Reconstructing the density matrix from such measurements of a set of observables is precisely the task of quantum state tomography. The key requirement for a reconstruction to succeed is that the measurements are tomographically complete; for example, that they span the entire parameter space of the system. Formally, this requires the density matrix to be of sufficiently low rank and that it can be sampled by sufficiently many independent measurements. Stating that r is of rank r means that there exists a spectral decomposition of r with r orthogonal states. Beyond the fully coherent
a d

Figure 1 | Decoherence in scattering experiments. a, Mixed states within the probing radiation may include all sources of mixing that manifest themselves as transverse partial coherence or finite bandwidth. b, Mixed states in the sample may occur from quantum mixtures and fast stationary stochastic processes such as vibrations, switching or steady flows. c, The spatial extent over which two detections cannot be discerned corresponds effectively to a mixed state in the detector plane. d, The signature of decoherence in far-field diffraction is a decrease in visibility of the scattered intensity distribution. Here the right half of the simulated diffraction pattern shows the effect of sample vibration.

t Mu nchen, 85748 Garching bei Mu nchen, Germany. 2Paul Scherrer Institut, 5232 Villigen PSI, Switzerland. Physics Department, Technische Universita

6 8 | N AT U R E | VO L 4 9 4 | 7 F E B R U A RY 2 0 1 3

2013 Macmillan Publishers Limited. All rights reserved

LETTER RESEARCH
case (r 5 1), ptychography can be generalized to mixed states (r . 1) of various origins, and the reconstruction problem amounts to finding tomographically complete sets for probe and object states. When the object is in a pure state, this spectral decomposition is equivalent to the mode expansion of the mutual coherence function14, as recently used for coherent diffractive imaging to allow reconstructions with partial transverse15 and longitudinal coherence16 from a single diffraction pattern. Determining the conditions under which a solution is trustworthy is a central concern for any inverse problem. Increasing the dimension of the density matrix adds a large number of degrees of freedom, which intuitively calls for an equal increase in measured data. This claim is supported by recent work in compressed sensing17,18, which confirms that, if the ensemble of measurement operators satisfies certain conditions, the increase in measured data should scale with the rank r of the density matrix to be reconstructed. The formal result does not apply completely to ptychography because of the need to concurrently reconstruct the object transmission function. However, simulations indicate that the rank scaling rule seems to apply in a wide range of situations. In practice this condition amounts to ensuring that the scanned area is covered by a larger number of measurement points. The fact that the use of finer sampling, or a larger overlap of adjacent probe positions, helps in solving the problem is consistent with the Wigner distribution point of view, taken in early formulations of both ptychography19,20 and quantum state tomography21. In these implementations, dense sampling effectively maps the complete phase space. Conversely, this general result also explains why ptychographic reconstructions can succeed with a relatively small number of diffraction patterns in comparison with the extents of the reconstructed images. Adapting existing algorithms to the new problem formulation is rather straightforward (see also Supplementary Information). Most reconstruction algorithms use projections22,23, one of which is to ensure that the model for the wave leaving the object complies with measured data. This Fourier magnitude projection can be seen as projecting any point in the complex plane onto a circle whose square radius is equal to the measured intensity. Reconstruction of the lowrank density matrix uses an equivalent projection in which the circle is replaced by a higher-dimensional quadric form. Maximum-likelihood refinement24,25 is also readily adapted through a modification of the intensity model. Any of these modified algorithms yields, if it converges, a set of r modes that span the rank space of r. All the reconstructions below use the difference map implementation7 followed by maximum-likelihood refinement25.
a

We demonstrate our findings with X-ray data from a simple ptychography experiment. The setup, shown in Fig. 2a, is described elsewhere7. The monochromatic beam was apertured by slits and focused by a Fresnel zone plate onto a test sample made from nanofabricated gold pillars. First, the slit opening was smaller than the area that can be considered coherent. Then the horizontal slits were opened fourfold, significantly exceeding the coherence length of the incident X-ray beam. Typical far-field diffraction patterns obtained in either case are shown in Fig. 2. As expected, reconstruction from the high-coherence data set poses no problem. However, even in this case, only about 80% of the illumination is carried by the dominant coherent mode, and imaging artefacts can be reduced by allowing for multiple modes (see Supplementary Fig. 3). Reconstructions from the low-coherence data set are shown in Fig. 3. An attempt that assumed full coherence failed to produce a faithful image of the sample. When reconstructing without a priori wavefront characterization12 modes in the illumination, image retrieval converged to the result shown in Fig. 3b. The five most significant modes obtained from this reconstruction are shown in Fig. 3c. The relative power of the dominant mode, as measure of coherence, is seen to be slightly less than 40%. The immediate consequence of our result is a reduced susceptibility to systematic errors while simultaneously relaxing experimental conditions. Acquisition times can be reduced by admitting higheralbeit less coherentflux, thereby alleviating long-term stability requirements for the instrumentation. Reconstruction of the illumination modes also has intrinsic interest, for instance for metrological applications to complement interferometric techniques26,27. Derived quantities such as the complex degree of coherence are readily computed from the reconstructed modes and are in good agreement with other experimental measurements (see Supplementary Fig. 2). Through detailed simulations we have verified that multiple additional sources of decoherence are similarly accounted for, including sample vibrations28 and partial coherence in an unfocused probe. Similarly, the loss of visibility caused by the detector point spread can be included in the density matrix (see Supplementary Fig. 3). Decoherence effects caused by the detector point spread are thus transferable to illumination properties, a fact that can be interpreted as a consequence of Cowleys reciprocity principle29. The formalism we present also opens new imaging applications, because the reconstruction of sample mixtures (Fig. 1b) can be implemented in a manner completely analogous to the cases discussed so far. The wide-sense stationary processes to be thus characterized include continuous sample movement, stochastic equilibrium fluctuations,

High coherence Low coherence

Figure 2 | Ptychography with partly coherent X-rays. a, A portion of the incident beam, selected by slits, is focused with a Fresnel zone plate onto the specimen, which is scanned in both transverse directions. The order-sorting aperture is not shown, for clarity. The scattered intensities are collected on a

detector in the far field. The inset shows two different openings of the horizontal slits as well as elongated ellipses representing the approximate area over which the beam is coherent. b, c, One of the measured diffraction patterns in high-coherence mode (b) and in low-coherence mode (c).
7 F E B R U A RY 2 0 1 3 | VO L 4 9 4 | N AT U R E | 6 9

2013 Macmillan Publishers Limited. All rights reserved

RESEARCH LETTER
a b a b

38.9%

27.6%

16.7%

52.3%

44.7%

54.0%

42.5% 0

7.8%

2.7%

Figure 4 | Imaging a simulated Ising model. a, In the sample layout, grey squares mark individual spins, which incur a 6p/2 phase shift on the illumination. Blue lines represent ferromagnetic couplings, J 5 21, red lines antiferromagnetic ones, J 5 11. Periodic boundary conditions apply. The configurations are Boltzmann distributed with a temperature five times the Curie temperature for the infinite two-dimensional Ising model. b, The four dominant modes after subtraction of the mean. Consistent with the probe size, long-range symmetries are not reconstructed. c, d, The results of orthogonalization of neighbouring spins, outlined by blue and red frames in b, for a ferromagnetic bond (c) and an antiferromagnetic bond (d). The relative mode amplitudes are in good agreement with expected values.

Figure 3 | Reconstruction from the low-coherence X-ray experiment. a, Reconstructed transmission function when the incoming wavefield is assumed to be fully coherent, that is, r is of rank one. b, Reconstructed transmission function with an incoming wavefield composed of r 5 12 independent modes. c, The five dominant modes reconstructed simultaneously with the sample, represented in Fourier space. Complex amplitudes in the sample plane are shown in the insets. In accordance with the shorter horizontal coherence length, variation in all modes is essentially only along the horizontal direction.

steady flows and periodic variations. Typically, to characterize such dynamics, intensity fluctuation spectroscopy6 or related techniques are employed and, when spatially inhomogeneous, statistics can be characterized by scanning-probe techniques. Applying ptychography to such fluctuation microscopy30 enhances its resolving power and sensitivity. To demonstrate the potential of mixed-sample-state reconstruction, we simulated an Ising toy model comprising 4 3 4 quantum dots, illustrated in Fig. 4a. Each quantum dot can be in two possible spin states, which we assume to produce contrast in a scattering experiment. The equilibrium configuration follows the usual rules of Ising models given by identical ferromagnetic couplings between all first neighbours and a temperature well above the Curie point. The four central couplings, shown in red, were chosen to be antiferromagnetic. This model is simple enough to have a configuration space of manageable size and is sufficiently well understood for validation of a simulated imaging experiment. In our simulation we assumed that the dynamics are fast; that is, that all possible configurations are explored during each single exposure. The extent of the simulated probe function is such that two neighbouring quantum dots can be illuminated simultaneously. Reconstruction from the simulated data set follows the same principles as described above. Here, the illumination was assumed to be a pure state, and the density matrix representing the mixed state of the sample was sufficient to describe the statistics of the entire experiment. Sixteen modes were reconstructed in total, of which the four dominant ones are shown in Fig. 4b. As a result of symmetry, the expectation
7 0 | N AT U R E | VO L 4 9 4 | 7 F E B R U A RY 2 0 1 3

value of any spin is zero. Still, the reconstruction reliably images each quantum dot because of the decoherence that its fluctuation causes. What is more, correlations between dots on length scales shorter than the illumination are encoded in the diffraction data. These short-range correlations are well reconstructed when orthogonalization is applied to two neighbouring spins (Fig. 4c, d). The obtained relative powers are quantitatively consistent with the ferromagnetic and antiferromagnetic couplings. Beyond magnetic systems, we see potential applications of our method in the direct imaging of spin-density wave domains, of quantum oscillator modes and of steady flows in nanofluidics. Our simulations indicate that reconstruction can succeed even in cases where decoherence occurs concurrently in the illumination and the object. However, in such cases, coupling of object and probe modes may be unavoidable because, for instance, decoherence caused by global translations of the sample has the same effect as translations of the illumination. The mixed-state formulation of ptychography presented in this letter helps to address some of the most limiting experimental challenges. Acceptance of some degree of partial coherence and robustness to detector point spread improves reconstruction quality and reliability and makes further instruments useable or a higher flux accessible, thus widening the applicability of ptychographic imaging. Bridging intensity fluctuation techniques with coherent diffractive imaging, the principles introduced in this paper allow dynamical systems to be imaged with only the most basic of assumptions; that is, wide-sense stationary statistics that can be described by low-rank density matrices. Steady-state fluctuations in both quantum and classical systems can be reconstructed unambiguously with high resolving power, offering new insight into fluctuations at phase boundaries, defects and points of excitation.

METHODS SUMMARY
Experiments were conducted at the cSAXS beamline (X12SA) at the Swiss Light . LongiSource, Paul Scherrer Institut, Villigen, at an X-ray wavelength of 2 A tudinal coherence was determined by the bandwidth of DE/E < 0.02% of the double-crystal Si(111) monochromator. As described in the text, transversecoherence filtering was achieved with movable slits about 33 m from the source. The coherence length at the slits was estimated to be about 200 mm 3 20 mm (horizontal 3 vertical). The same slits determined the part of the Fresnel zone

2013 Macmillan Publishers Limited. All rights reserved

LETTER RESEARCH
plate that was illuminated to create the probe. The zone plate had been manufactured as described in ref. 31 with 1,000-nm-thick gold structures, a diameter of 200 mm and a finest zone width of 100 nm. The same manufacturing process was used for the sample, which consisted of a pseudorandom array of gold structures with diameter 500 nm on a rectangular grid with pitch 750 nm. A PILATUS-II single-photon counting hybrid pixel detector32 was positioned 7.18 m from the sample. Over roughly 7 m the volume between sample and detector was flushed with helium to reduce absorption and scattering. Scans, shown in Fig. 3, comprised sample movements along concentric circles. Six such circles, with an increase in diameter of Dr 5 0.25 mm and 5n points on the nth shell10, resulted in data acquisition at 90 positions for a field of view of about 5 mm diameter. The acquisition time was typically 2 s per position.
Received 26 June; accepted 16 November 2012. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. Miao, J., Charalambous, P., Kirz, J. & Sayre, D. Extending the methodology of X-ray crystallography to allow imaging of micrometre-sized non-crystalline specimens. Nature 400, 342344 (1999). Eisebitt, S. et al. Lensless imaging of magnetic nanostructures by X-ray spectroholography. Nature 432, 885888 (2004). Abbey, B. et al. Keyhole coherent diffractive imaging. Nature Phys. 4, 394398 (2008). Mandel, L. & Wolf, E. Coherence properties of optical fields. Rev. Mod. Phys. 37, 231287 (1965). Goodman, J. W. Statistical Optics (Wiley, 2000). Sutton, M. A review of X-ray intensity fluctuation spectroscopy. C. R. Phys. 9, 657667 (2008). Thibault, P. et al. High-resolution scanning X-ray diffraction microscopy. Science 321, 379382 (2008). Maiden, A. M. & Rodenburg, J. M. An improved ptychographical phase retrieval algorithm for diffractive imaging. Ultramicroscopy 109, 12561262 (2009). Clark, J. N. & Peele, A. G. Simultaneous sample and spatial coherence characterisation using diffractive imaging. Appl. Phys. Lett. 99, 154103 (2011). Dierolf, M. et al. Ptychographic X-ray computed tomography at the nanoscale. Nature 467, 436439 (2010). Putkunz, C. et al. Atom-scale ptychographic electron diffractive imaging of boron nitride cones. Phys. Rev. Lett. 108, 14 (2012). Maiden, A. M., Rodenburg, J. M. & Humphry, M. J. Optical ptychography: a practical implementation with useful resolution. Opt. Lett. 35, 25852587 (2010). von Neumann, J. Wahrscheinlichkeitstheoretischer Aufbau der ttingen Math.-Phys. Kl. 245272 (1927). Quantenmechanik. Nachr. Ges. Wiss. Go Wolf, E. New theory of partial coherence in the space-frequency domain Part I: spectra and cross spectra of steady-state sources. J. Opt. Soc. Am. 72, 343351 (1982). Whitehead, L. W. et al. Diffractive imaging using partially coherent X rays. Phys. Rev. Lett. 103, 243902 (2009). Abbey, B. et al. Lensless imaging using broadband X-ray sources. Nature Photon. 5, 420424 (2011). ` s, E. J. & Recht, B. Exact matrix completion via convex optimization. Found. Cande Comput. Math. 9, 717772 (2009). 18. Gross, D. Recovering low-rank matrices from few coefficients in any basis. IEEE Trans. Inf. Theory 57, 15481566 (2011). 19. Rodenburg, J. M. & Bates, R. H. T. The theory of super-resolution electron microscopy via Wigner-distribution deconvolution. Phil. Trans. R. Soc. Lond. A 339, 521553 (1992). 20. Chapman, H. N. Phase-retrieval X-ray microscopy by Wigner-distribution deconvolution: signal processing. Scanning Microsc. 11, 6780 (1997). 21. Raymer, M. G. Measuring the quantum mechanical wave function. Contemp. Phys. 38, 343355 (1997). 22. Fienup, J. R. Phase retrieval algorithms: a comparison. Appl. Opt. 21, 27582769 (1982). 23. Elser, V. Phase retrieval by iterated projections. J. Opt. Soc. Am. A Opt. Image Sci. Vis. 20, 4055 (2003). 24. Guizar-Sicairos, M. & Fienup, J. R. Phase retrieval with transverse translation diversity: a nonlinear optimization approach. Opt. Express 16, 72647278 (2008). 25. Thibault, P. & Guizar-Sicairos, M. Maximum-likelihood refinement for coherent diffractive imaging. N. J. Phys. 14, 063004 (2012). 26. Pfeiffer, F. et al. Shearing interferometer for quantifying the coherence of hard X-ray beams. Phys. Rev. Lett. 94, 164801 (2005). 27. Cerbino, R. et al. X-ray-scattering information obtained from near-field speckle. Nature Phys. 4, 238243 (2008). 28. Clark, J. N. et al. Dynamic sample imaging in coherent diffractive imaging. Opt. Lett. 36, 19541956 (2011). 29. Cowley, J. M. Image contrast in a transmission scanning electron microscope. Appl. Phys. Lett. 58, 5859 (1969). 30. Treacy, M. M. J., Gibson, J. M., Fan, L., Paterson, D. J. & McNulty, I. Fluctuation microscopy: a probe of medium range order. Rep. Prog. Phys. 68, 28992944 (2005). 31. Gorelick, S., Guzenko, V. A., Vila-Comamala, J. & David, C. Direct e-beam writing of dense and high aspect ratio nanostructures in thick layers of PMMA for electroplating. Nanotechnology 21, 295303 (2010). 32. Henrich, B. et al. PILATUS: a single photon counting pixel detector for x-ray applications. Nucl. Instrum. Methods Phys. Res. A 607, 247249 (2009). Supplementary Information is available in the online version of the paper. Acknowledgements We thank M. Dierolf for discussions and help in the algorithm implementation; A. Diaz for help during the measurements; C. Kewish for providing the sample, which had been produced by J. Vila Comamala; V. Elser for pointing us to relevant literature; and F. Pfeiffer, M. Bech and I. Zanette for helping to improve the manuscript. This work is supported in part by a European Research Council Starting Grant, under project OptImaX (no. 279753). Author Contributions A.M. and P.T. designed and conducted the experiment. P.T. analysed the data and prepared the simulations. Both authors worked together to refine the methods, interpret results, write the manuscript and create the figures. Author Information Reprints and permissions information is available at www.nature.com/reprints. The authors declare no competing financial interests. Readers are welcome to comment on the online version of the paper. Correspondence and requests for materials should be addressed to P.T. (pierre.thibault@tum.de).

7 F E B R U A RY 2 0 1 3 | VO L 4 9 4 | N AT U R E | 7 1

2013 Macmillan Publishers Limited. All rights reserved