1

Reconstruction of a yeast cell from x-ray diffraction data

Pierre Thibault,a Veit Elser,a * Chris Jacobsen,b,c David Shapirob and

David Sayreb

a Department of Physics, Cornell University, Ithaca, NY 14853-2501 USA,

b Department of Physics and Astronomy, Stony Brook University, Stony Brook, NY

11794-3800 USA, and c Center for Functional Nanomaterials, Brookhaven National

Laboratory, Upton, NY 11973, USA. E-mail: [email protected]

(Received 0 XXXXXXX 0000; accepted 0 XXXXXXX 0000)

Abstract

We provide details of the algorithm used for the reconstruction of yeast cell images

in the recent demonstration of diffraction microscopy by Shapiro et al. (2005). Two

refinements of the iterative, constraint-based scheme are developed to address the

current experimental realities of this imaging technique, which include missing central

data and noise. We define a constrained power operator whose eigenmodes help us

identify a small number of degrees of freedom in the reconstruction that are negli-

gibly constrained as a result of the missing data. To achieve reproducibility in the

algorithm’s output, a special intervention is required for these modes. Weak incom-

patibility of the constraints caused by noise in both direct and Fourier space leads to

residual phase fluctuations. We address this problem by supplementing the algorithm

with an averaging method. The effect of averaging may be interpreted in terms of an

effective modulation transfer function, as used in optics, to quantify the resolution.

We preface the reconstruction details with simulations of wave propagation through a

model yeast cell. These show that the yeast cell is a strong phase contrast object for

the conditions in the experiment.

PREPRINT: Acta Crystallographica Section A A Journal of the International Union of Crystallography
 2

1. Introduction

The proposal for using oversampled x-ray diffraction patterns as the basis of a new

type of microscopy (Sayre, 1980) was advanced by many recent experiments. After

its first demonstration with a manufactured two-dimensional specimen by Miao et al.

(1999), the method was successfully used on other 2D objects (Marchesini et al.,

2003), on 3D engineered structures (Miao et al., 2002; Chapman et al., 2005), on

stained bacteria (Miao et al., 2003) and on micro-crystals (Williams et al., 2003). A

recent advance was made by Shapiro et al. (2005) by the measurement of a highly

detailed diffraction pattern produced by a single, unstained, freeze-dried yeast cell

with soft x-rays (750 eV). An integral part of the proposed “diffraction microscope”

is the reconstruction algorithm that interprets the diffraction pattern; this action is

performed in a conventional microscope by physical apparatus (lenses, etc.). We report

here on the development of this algorithmic component of the instrument that resulted

from efforts directed at the yeast cell data.

Reconstruction of the image of an isolated object is made possible if its diffraction

pattern is measured on a fine enough grid. In direct space, this oversampling results in

a field of view empty but for a small region occupied by the object. If this region, called

the “support”, is small enough, the phase problem has an overwhelmingly high prob-

ability of having a unique solution. For a more rigorous treatment of these questions

see Miao & Sayre (2000) and Elser & Millane (2005). All recent work in diffraction

microscopy make use of iterative reconstruction algorithms. The most popular to date

is Fienup’s (1982) hybrid input-output (HIO) algorithm, which was initially inspired

by Gerchberg & Saxton’s (1972) algorithm. Since then, many variations have been

proposed (Wu et al., 2004; Marchesini et al., 2003; Elser, 2003a; Russel Luke, 2005).

The algorithm used in this work is the difference map (Elser, 2003a), a generalization

of the hybrid input-output which makes it applicable to a wide variety of problems

IUCr macros version 2.0β15: 2004/05/19

 3

(Elser, 2003b; Elser & Rankenburg, 2005). An overview of this algorithm will be given

in Section 4.

Before reconstruction is attempted, some important attributes of the object to

be reconstructed can be determined directly from the data or with only a mini-

mum amount of processing. One of the data sets is reproduced in Figure 1 (Shapiro

et al., 2005). The streaks of intensity at variance with 2D Friedel symmetry call for

an explanation even before the data is phased and the image reconstructed. Is the

asymmetry with respect to 180◦ rotation due to the curvature of the Ewald sphere, or

evidence of strong contrast? By simulating the wave propagation through a model cell,

we show in Section 2 that, although these mechanisms do in principle contribute, by

far the strongest source of asymmetry is the large optical thickness of the specimen.

From this exercise we learn that the contrast mechanism is far from satisfying the

Born approximation, and that the data corresponds to the far field diffraction pattern

of a strongly phase shifted exit wave. This information is important when formulating

the constraints used by the phasing algorithm (Section 4).

The detail of the diffraction pattern (Fig. 1, inset) shows a continuous speckle

pattern — qualitative evidence that the oversampling of the diffraction pattern may

be sufficient for phasing. A more quantitative test is provided by the autocorrelation

of the object/exit-wave, as obtained by Fourier transforming the intensity distribution

in the diffraction pattern. Because of the non-Friedel asymmetry this will be complex-

valued; the real part is shown in Figure 2. We see evidence of a rather sharply defined

object support. The outline of the oval shape has perfect symmetry with respect

to 180◦ rotation and corresponds to the autocorrelation of the object support. The

diameter is twice that of the actual yeast cell. We have direct evidence of the cell

from the faint “ghosts” surrounding the symmetric central oval. These are likely the

result of phase interference between the yeast cell and isolated point-like scatterers

IUCr macros version 2.0β15: 2004/05/19

 4

located a few cell diameters distant, thus forming something like a collection of Fourier-

transform holograms (Stroke & Falconer, 1964; McNulty et al., 1992; Eisebitt et al.,

2004). The autocorrelation image thus provides very detailed information about the

support constraint used by the phasing algorithm (Fienup, 1978). Estimating the area

in pixels of the support we arrive at an oversampling ratio of about 25 (Miao &

Sayre, 2000), or about 5 in each dimensions. A quantity more relevant to the difficulty

of a reconstruction is the overdetermination ratio (Elser & Millane, 2005), given by

1 Asupport
Ω= , (1)
2 Aautocorrelation

where Asupport and Aautocorrelation are the number of pixels in the support and in the

autocorrelation. This number can be seen as the ratio of the number of constraints

to the number of degrees of freedom: the solution is more overdetermined when Ω

is larger. For the yeast cell problem, the overdetermination ratio is almost exactly

2 because the support is convex and nearly centrosymmetric. This represents the

most challenging case in terms of computation since Ω cannot be less than 2 in two

dimensions. Non-convex supports (as is seen in Chapman et al., 2005, for instance)

have a higher overdetermination ratio and are known to be easier to solve (Fienup,

1987).

Data at the very center of the diffraction data is missing because of the beamstop;

this beamstop is required in present experiments to avoid saturation and damage

of the x-ray detector by the very strong undiffracted beam, and saturation by the

stronger diffraction intensity at low spatial frequency. In crystallographic phasing,

where the primary constraint, atomicity, has mostly high spatial frequency content,

missing reflections at small angles do not pose a serious challenge. In speckle diffraction

reconstructions the situation is different, and we address the problem of missing data

in Section 3.

IUCr macros version 2.0β15: 2004/05/19

 5

2. Multi-slice simulations

To arrive at a better understanding of the contrast mechanism in this experiment,

we simulated the propagation of x-rays through a model, freeze dried yeast cell. Our

model cell is a 3 µm sphere made of lipids and proteins. All the lipids are concentrated

in the 50 nm thick cell membrane. The protein material inside the cell is modeled as

a binary-valued distribution occupying 25% of the cell volume. The exact spatial

distribution of the protein material is random, but the envelope of its power spectrum

was forced to follow a power law of the form |F (q)| ∝ q −α . The exponent of this power

law controls the relative contributions of high and low spatial frequencies. Figure 3(a)

shows the protein material in one slice of the cell when generated with our choice of

exponent, α = 1.6. For the simulations, the refractive index of the material inside the

cell was sampled on a grid with (25 nm)3 voxels.

The refractive indices n = 1−δ −iβ, of a model lipid (62.3% H, 31.4% C and 6.3% O

as number fraction) and protein (48.6% H, 32.9% C, 8.9% N, 8.9% O and 0.6% S)

were calculated using the data of Henke et al. (1993). At the 750 eV x-ray energy of

the experiment these are

δ β

lipid 4.23 × 10−4 6.88 × 10−5

protein 5.44 × 10−4 1.05 × 10−4

From this we see that the wave acquires a phase shift of ∆ϕ = 2πtδ/λ or about π/2

as it passes through the center of a 3 µm diameter cell with a quarter of its volume

filled with protein so that t = 0.75 µm. The magnitude of this phase shift already

implies that the Born approximation will not be valid in the interpretation of the

diffraction pattern.

We used the multi-slice propagation formalism derived in appendix A to calculate

exit waves for our model yeast cell. The multi-slice method is based on a discretization
IUCr macros version 2.0β15: 2004/05/19
 6

of the wave propagation along the direction of the incident wave (set along the z-axis

in this work), from one plane to anoother. The main result derived in the appendix A

is the following formula:

" # √

 
e
Ψq (z + ∆z) = Ψe q (z) + p ik ∆z f
δn(z) e
∗ Ψ(z) ei∆z k2 −q 2 −k
(2)
1 − q 2 /k 2 q

e q (z) is the two dimensional Fourier transform of the wavefield in the transverse
where Ψ
f q is the Fourier
plane at position z, k = 2π/λ, ∗ is the convolution operation and δn
1 2
transform (in the transverse plane) of δn(r) = 2 (n − 1) ≈ −δ(r) − iβ(r). This

multi-slice formula was preferred to Cowley & Moodie’s (1957) standard multi-slice

formula [Eq. (27) in appendix A] because it does not rely on the small scattering

angle approximation. Equation (2) becomes exact for arbitrary large ∆z when δn = 0,

resulting in the simple free space propagation equation:

√
e q (z + ∆z) = Ψ
e q (z)ei∆z( k2 −q 2 −k)
Ψ . (3)

Figure 3(b) shows a typical exit wave obtained with (2) and rendered such that hue

and brightness correspond to phase and magnitude respectively. The distribution of

the exit wave values in the complex plane, shown in Figure 3(d ), is itself useful in

that it might be used to constrain the reconstruction of the object/exit-wave. That

these values lie on an approximate spiral is explained by the Eikonal approximation

(24) and is obtained from (2) in the limit q ≪ k. The distribution of exit wave values

Ψ is simply related to the distribution, transverse to the beam, of projected protein

thicknesses t by

Ψ ∝ exp(i k t δn). (4)

Broadening of the spiral is due to Ewald sphere lift-off effects, that is, terms in the

propagation equations [Eq. (21) in appendix A] of higher order in the transverse spatial

frequency, q/k.

IUCr macros version 2.0β15: 2004/05/19

 7

The non-negligible curvature of the Ewald sphere over the measured diffraction

pattern makes the question of focus important for reconstructions (Spence et al., 2002).

By definition, features in the exit waves are more in focus if they are close to the exit

plane, while they are slightly dispersed by the wave propagation if they are farther

from that plane. For a spherical cell, the plane for which the wavefield has the smallest

support is the plane going through its center, corresponding to the exit wave back-

propagated [in vacuum, using Eq. (3)] by half the diameter of the cell. We define this

plane as the “focal plane”.

Figure 4(a) illustrates the simulated perturbation of the incident wave as it propa-

gates through the cell. At the scale of this image, the lateral dispersion of the wavefield

is unnoticeable. Figure 4(b) shows the wavefield propagated (in free space) both for-

ward and backward from the exit plane, with the axis in the direction of propagation

compressed by a factor 5. As emphasized in figure 4(c), there is a noticeable spread

of the field perturbation. This observation is very important for the understanding of

the reconstruction method described in Section 4.

It is the exit wave back propagated to the focal plane that the constraints of the

reconstruction algorithm are applied to. This two-dimensional object is subject both

to the support constraint and, because its forward propagation leads to the measured

diffraction pattern, to the Fourier constraint (Fienup, 1982; Elser, 2003a).

The missing data at the center of the diffraction pattern, which accounts for most

of the power in the exit wave, precludes the direct use of value constraints in the

reconstruction, and the spiral distribution in particular. With reduced missing data,

and use of the unconstrained mode analysis described in Section 3, it may be possible

to exploit value constraints in the future.

IUCr macros version 2.0β15: 2004/05/19

 8

3. Unconstrained modes

Any reconstruction effort that uses a support constraint and is faced with a significant

region of missing data at the center of the diffraction data must contend with the

following form of ambiguity. Consider adding a broad Gaussian feature at the center

of the support. If the tails of the Gaussian are small at the support boundary, then

the added feature is still consistent with the support constraint. On the other hand,

if the Gaussian is sufficiently broad, its counterpart in Fourier space will be narrow

enough to fit inside the region of missing data and not upset the constrained diffraction

intensities along that region’s boundary. Clearly there is an optimal Gaussian mode,

with respect to constraints in both direct and Fourier space. The amplitude of this

mode is negligibly constrained by the data and support, and as such represents a

source of ambiguity in the reconstruction.

There may be other modes, in addition to the Gaussian just described, that con-

tribute to the ambiguity. In appendix C we develop a formalism for identifying all the

“unconstrained modes” for any particular support/missing-data combination. This

analysis provides a set of eigenmodes of a “constrained power operator”, ranked by

their corresponding eigenvalues. The constrained power of a mode ranges, by defini-

tion, between 0 and 2, where 2 would apply to a mode having most of its power in

the measured diffraction pattern and, in direct space, outside the object support. We

are concerned here with modes having negligible constrained power. A useful estimate

of the number of such modes is derived in appendix C and is given by the formula

M = NC /σ, where NC is the number of missing Fourier samples at the center of

the diffraction pattern and σ is the oversampling ratio (Miao & Sayre, 2000). For the

data shown in Figure 1, NC = 362; this together with our earlier estimate σ = 25

gives M = 14.5. A euristic approach consists in describing the unconstrained degrees

of freedom as “missing speckles”; in the present work, M has essentially the same

IUCr macros version 2.0β15: 2004/05/19

 9

signification than the “number of speckles” introduced recently by Miao et al. (2005).

Figure 5 shows the results of a detailed unconstrained mode analysis. The number of

modes (four) with constrained power less than 0.02 gives some sense of the number of

unconstrained degrees of freedom in the reconstruction. Since all of these modes lie, in

Fourier space, at the center of the diffraction pattern, their actual power is large. This

makes value constraints such as positivity problematic. In the case of complex-valued

reconstructions with pixel values having special distributions [for example Fig. 3(c)],

the use of value constraints can in principle be beneficial. We chose not to use such

constraints for the yeast cell reconstruction.

How does one resolve this M -parameter ambiguity in the reconstruction? With-

out the benefit of supplementary data or additional a priori knowledge, these mode

amplitudes are completely unconstrained. Since the reconstruction algorithm we use

is based on the available constraints, the amplitudes of the unconstrained modes drift

only very slowly with iteration number and are therefore strongly subject to the ran-

dom start that initiates the algorithm. To eliminate these very slow degrees of freedom

and to improve the reproducibility of results, our algorithm orthogonalizes the recon-

struction with respect to a set of unconstrained modes that are computed in advance.

For the yeast reconstruction we used 4 modes, the highest having constrained power

0.0162. In the resulting reconstructions the artificial zeroing of these mode amplitudes

is very noticeable, owing to the fact that their relative power should be quite large.

To mitigate this largely aesthetic problem we use an ad hoc prescription for restoring

the mode amplitudes that still leads to reproducible results. This is simply to deter-

mine the amplitudes by the condition that the variance of the distribution of complex

pixel values within the support is minimized. If Ψ0 is the reconstruction that has been

orthogonalized with respect to the M modes, the complex amplitude ai of the ith

IUCr macros version 2.0β15: 2004/05/19

 10

mode χi is given by the minimization of

 D E 
PM PM 2
(∆Ψ)2 = Ψ0 + i=1 ai χi − Ψ0 + i=1 ai χi , (5)
S S

where h iS is the complex-valued pixel average inside the support.

We find that reconstructions appear more “plausible” when unconstrained modes

are restored by this procedure than if their amplitudes are set to zero. Figure 6 illus-

trates the effect of the mode replacement method. Fig. 6(a) is an example of a recon-

struction obtained when the mode amplitudes are allowed to vary freely. Fig. 6(b)

shows the result of zeroing the mode amplitudes. The final reconstruction, with the

modes restored with the ad hoc rule is shown in Fig. 6(c).

4. Reconstruction algorithm

There are a number of different algorithms that can be used for speckle diffraction

reconstructions (Fienup, 1982; Bauschke et al., 2003). Our reconstruction used the

difference map algorithm (Elser, 2003a; Elser, 2003b) with support and Fourier mag-

nitude as the two constraints. The support projector PS zeroes the reconstruction

outside the support,

(
Ψr if r ∈ S,
PS (Ψr ) = (6)
0 otherwise,
while PF rescales the corresponding Fourier transform to have the measured magni-

tudes Fq wherever the latter are known (and leaves others unchanged):

e
F Ψq
 e q| =
if Fq is known and |Ψ 6 0,
e q
PF (Ψq ) = e q|
|Ψ (7)

e
Ψq otherwise.
For the standard choice of γ parameters, and β = 1, this algorithm reduces to Fienup’s

hybrid input-output algorithm. For the yeast cell reconstruction we used the simple

alternative β = −1, for which the iteration is given by

Ψn+1 = Ψn + (ΨF n − ΨSn ), (8)

IUCr macros version 2.0β15: 2004/05/19
 11

where
 
ΨF n = PF 2PS (Ψn ) − Ψn and ΨSn = PS (Ψn ) (9)

are the nth Fourier and support estimates, respectively1 . It is clear in equation (8)

that the iteration reaches a fixed point if the two estimates are equal, in which case the

reconstruction is equal to either of these estimates. The Fourier and support estimates

differ in general and the difference map error, used to monitor convergence, is defined

as

ǫn = kΨn+1 − Ψn k = kΨF n − ΨSn k. (10)

The reconstruction iterate Ψn is complex-valued since the Fourier magnitudes in the

data lack Friedel symmetry.

Details about the implementation of the algorithm that are special to the yeast

reconstruction and deserve elaboration are the treatment of the support constraint

and the averaging procedure used to achieve reproducible results in the presence of

noisy constraints.

The support was initially defined as a rectangle of dimensions half the size of the

oval shape observed in the autocorrelation (Figure 2). It was then refined with var-

ious methods, going from simple thresholding to pixel-by-pixel editing in a drawing

program. A more systematic approach has been demonstrated by Marchesini et al.

(2003) who use a combination of low-pass filtering and thresholding to automatically

generate a tightening support at each iteration in a procedure called “shrink wrap”.

We did not use this approach in our present work. Instead, each time a new support

was obtained, a new reconstruction was attempted (without most of the refinements

described below) for a few hundred iterations only. The new reconstruction attempt

was then used to define a yet tighter support, until no further improvement was possi-
1
Note that 2PS (Ψ) − Ψ, appearing here with our particular choice of parameters, is called in other
work the “reflector” (Bauschke et al., 2003) or the “charge-flipping” operation (Wu et al., 2004).

IUCr macros version 2.0β15: 2004/05/19

 12

ble. Excessively tight supports were avoided by noticing sharp upturns in the difference

map error.

A new procedure was found to be necessary for obtaining reproducible results. The

noise in the Fourier data creates incompatibility in the constraints with the result

that the reconstruction iterate never reaches a fixed point. Instead, the algorithm

enters a steady state, characterized by an error that fluctuates around a constant,

non-zero value. The relatively small error in this regime suggests that the region

explored by the algorithm contains the best solution, that is, the image that would be

reconstructed in absence of noise in the data. In the absence of convergence, however,

the definition of “solution” becomes ambiguous. We decided to adopt an averaging

method to obtain a unique solution, both independent of the initial conditions and

of the number of iterations. The final reconstruction is defined as the average of

many Fourier estimates in the steady state regime. We take advantage of the chaotic

dynamics of the algorithm to form averages out of estimates taken in a single run; we

did not notice any benefit from the averaging of estimates from different runs, although

it has been observed recently by Chapman et al. (2005) that using different runs could

help identifying reconstruction pathologies (phase vortices). A tight support is needed

to avoid compromising the average by a drift of the reconstruction within the support.

Cross-correlation between estimates could also be used to improve alignment before

computing the average, though this procedure was not needed in the present case.

Besides reproducibility, another benefit of the averaging process is a systematic

cancellation of the highly fluctuating (thus not well determined) degrees of freedom.

Since the phase fluctuation is larger in the high spatial frequency region of the Fourier

data, the global effect of averaging is a decrease in the detail seen in the reconstruction.

Apart from achieving reproducibility, we believe that reconstructions not averaged over

residual fluctuations convey a misleading degree of detail (resolution). As described

IUCr macros version 2.0β15: 2004/05/19

 13

in Section 5, the power reduction at high spatial frequencies, as a result of averaging,

provides a way to measure the spatial resolution of the final reconstruction.

Averaging a complex-valued reconstruction introduces additional complications.

Both projectors are blind to a constant phase factor in the reconstruction iterate

(that is, PS (eiθ Ψ) = eiθ PS (Ψ) and PF (eiθ Ψ) = eiθ PF (Ψ)). Prior to averaging, we

therefore reset the global phase factor such that the sum of the pixel values in ΨF

lies on the real axis. This rotation in the complex plane has to be done only after the

unconstrained modes are projected out of the image, since these free amplitudes bias

the evaluation of the global phase factor.

Noise and support size play an even more subtle role in the particular case of

the reconstruction of a complex-valued image. In fact, it has been often observed

(Fienup, 1987; Spence et al., 2002; Faulkner & Rodenburg, 2004) that very tight sup-

ports are needed for the reconstruction of complex-valued objects, especially when

noise is present in the data. This can be explained by the defocus ambiguity inherent

to any reconstruction lacking value constraints (such as reality or positivity). The mea-

surement of the cell’s diffraction pattern gives the magnitude of the in-plane Fourier

transform of the exit wave, |Ψ e q (z + ∆z)

e q (z) = Ψ
e q |. By virtue of equation (3), Ψ

for any ∆z. The Fourier constraint is therefore insensitive to the focus plane of the

reconstruction. As illustrated in Figure 4, if ∆z is small enough (see Appendix B for

a more precise definition of ∆z), an exit wave propagated backward or forward using

Eq. (3) still almost satisfies the support constraint as well. Hence, the reconstruc-

tion will never be uniquely determined since a whole ensemble of defocus planes will

nearly satisfy both constraints. If there was no noise in the diffraction pattern and no

unknown scatterers outside the support, the reconstruction algorithm should in prin-

ciple converge to what we defined earlier as the focal plane. Therefore, in any realistic

situation and in absence of value constraints in the support, the reproducibility of

IUCr macros version 2.0β15: 2004/05/19

 14

the reconstruction is compromised, not only by the non-convergence of the algorithm

but also by this defocus ambiguity. This problem is not present in the case of real-

valued exit waves since the exponential factor in (3) compromises Friedel symmetry if

∆z 6= 0. Very similar conclusions about this defocus abiguity were recently published

by Chapman et al. (2005).

For the defocus issues as well, the averaging method provides a unique reconstruc-

tion, at the cost of a resolution decrease. It is shown in appendix B that the averaging

procedure on the defocus ensemble results in an effective low-pass filter on the recon-

struction. This source of resolution loss should become dominant for loose supports

and/or noisy data.

Future applications of the averaging method could depend less on a tight support if

one instead translates the reconstruction to a unique position using the transformation

rules in Fourier space. In addition to translations r transverse to the beam, we must

also consider “defocus” translations z along the beam. The corresponding phase shifts

in Fourier space are given by

q
φ′q = φq + φ0 + r · q + z( k 2 − q 2 − k) (11)

where we have also included the global phase ambiguity, φ0 . Arbitrary values of the

parameters φ0 , r and z result in a new Fourier estimate also consistent with the

diffraction data. The values of r and z are normally discovered by the algorithm in the

process of satisfying the support constraint. Alternatively, once the reconstruction has

entered the averaging stage, one could relax the support constraint and fix r and z by,

say, zeroing the mean phase ramp and the isotropic part of the phase curvature. This

may turn out to be a more robust centering procedure, not plagued by uncertainties

in estimating the precise shape of the support.

IUCr macros version 2.0β15: 2004/05/19

 15

5. Results

The reconstructed yeast cell image is shown in Figure 7(a). It is the average of 1000

Fourier estimates, taken every 50 iterations in a single run. We let the averaging begin

only after 50, 000 iterations to avoid transient effects. Figure 7(b) shows the difference

map error of the first 10, 000 iterations. As suggested by this plot, 5000 iterations

would have been sufficient to avoid transients.

One important point has to be made regarding normalization: prior to the recon-

struction, the diffraction data was normalized in the standard way, that is by setting

the average intensity per pixel equal to 1. However, the missing central data seriously

underestimates the total intensity. The absolute scale of the difference map error thus

does not have a staightforward interpretation. One can compare the mean value of the

error in the steady-state regime ǫ with the Fourier magnitudes. The resolution shell

where the Fourier magnitudes are of the order of ǫ is independent of the normalization,

and evaluates to 42 nm in the current reconstruction. We expect this number to be

proportional to the resolution of the reconstruction, although, as explained below, we

judge that it is an overestimate in the present case.

Defining the resolution of the reconstruction is at first not obvious. While resolution

is limited by the apparatus in conventional microscopy and by the quality of the speci-

men in x-ray crystallography, both effects seem to be playing a role in x-ray diffraction

microscopy. For a biological specimen, one has additionally a source of measurement

noise associated with dose limits. The resolution of a particular reconstruction is there-

fore a function of the specimen observed. As explained above, the averaging method

seems very well suited to reveal the effect of noise on the resolution.

The dimension of the smallest features in Fig. 7(a) is about 3 pixels, leading to a

rough evaluation of the resolution of 30 nm. Figure 8 shows a more rigorous way of

determining the resolution. For any q in Fourier space, the ratio of the reconstructed

IUCr macros version 2.0β15: 2004/05/19

 16

e rec (q)|2 /|Fobs (q)|2 , is the result of averaging over

intensity to the observed intensity, |Ψ

residual phase fluctuations in the steady state of the algorithm. The solid curve on

the graph is the angular average of this ratio. This curve can be regarded as being

analogous to a modulation transfer function (MTF), since the MTF of an optical

system indicates the fraction of power in the object captured in the image as a function

of spatial frequency. For comparison, two classical optics MTF curves (dashed lines)

have been added to the graph, showing that the 30 nm resolution estimate is sensible.

We verified that the finest details in the final reconstruction were reproduced when

the algorithm was given different starting phases. When the algorithm is given the

opposite sign of the difference map parameter β, the agreement is only partial, though

unnoticeable to the eye, as illustrated in figures 9(a) and 9(b). Figure 9(c) shows a

cross-section of the absolute value of the reconstruction. This figure shows that the

high spatial frequency features are very well reproduced and that the disagreement

lies mostly in low spatial frequencies. That the reconstructions are not identical for

different sets of parameters is to be expected since the averaging process depends on

the ensemble of estimates sampled, which depends on the dynamics of the algorithm.

Hence, the high similarity between the two reconstructions in figure 9 increases our

confidence in our results. Contamination by the reconstruction twin (enantiomorph)

could be a source of concern, especially, as in our case, when the support only weakly

favors one twin over the other. Certainly to get optimal results, the averaging of

reconstructions should commence only after obvious signs of transients, such as seen

in the difference map error, have passed.

As suggested by the autocorrelation (Fig. 2), weak scatterers were present around

the yeast cell during the measurement of the diffraction pattern. Figure 10 shows

the result of our attempt to reconstruct the “dust” surrounding the cell. This recon-

struction was realized by a simple relaxation of the support constraint: the modified

IUCr macros version 2.0β15: 2004/05/19

 17

support projector sets to zero only those pixels outside the support with amplitude

below a predetermined threshold c:

(
Ψr if r ∈ S or |Ψr | > c
PSc (Ψr ) = (12)
0 otherwise.

In the present work, we found that setting this threshold to about 6 times the ampli-

tude of the error was appropriate. Of course, too low a value for the threshold weakens

the constraint and slows down the dynamics of the algorithm. This modified projector

can be unstable if most of the power is not already inside the support. The recon-

struction shown on Figure 10 was made by starting with the final iterate of a previous

run with the regular support projection. A total of 20, 000 iterations was needed to

generate this average of 400 estimates. In future experiments, a small quantity of

strong scatterers could be placed around the main specimen (see for instance Eisebitt

et al., 2004). The resulting increase in the overdetermination ratio should help the

reconstruction algorithm and also resolve the defocus ambiguity if the specimen is

expected to be complex-valued. In the present work, these scatterers were too weak

to be useful (less than 0.02% of the total power comes from the reconstructed dust).

6. Conclusion

The reconstruction of a real space image from the x-ray diffraction pattern of a sin-

gle, unstained cell has stimulated several developments in reconstruction algorithms.

Special attention was devoted to the reproducibility of the reconstruction. The miss-

ing data in the center of the diffraction pattern gives rise to very weakly constrained

degrees of freedom (modes) that we identified as the eigenvectors of a constrained

power operator. Very weakly constrained modes were projected out of the recon-

struction and replaced using a well-defined ad hoc rule. The lack of convergence was

addressed by adopting an averaging method. The final reconstruction therefore does

IUCr macros version 2.0β15: 2004/05/19

 18

not depend on the starting phases or the number of iterations. The averaging results in

an improvement of image reliability along with an effective decrease of the resolution.

This can be traced to noise in the data and, in the special case of a complex-valued

object without value constraint, the ambiguity in the defocus plane position.

A number of issues still stand in the way of the application of diffraction microscopy

to the three dimensional imaging of optically thick biological specimens. While miss-

ing central data poses a problem even for reconstructions of optically thin specimens

(valid Born approximation), the problem is more acute when value constraints (posi-

tivity, etc.) cannot be applied directly to the reconstructed object (exit wave). Present

experiments are working at minimizing the region of missing data, and it may soon

be possible to obtain 2D data that can be reconstructed without the benefit of value

constraints. Merging a series of 2D reconstructions tomographically, to yield a 3D map

of the specimen’s refractive index, poses another challenge. For thick specimens this is

probably feasible only when the relationship between the 3D refractive index contrast

and the exit wave is simple (Rytov, 1937). An attractive possibility is to carry out the

full set of 2D reconstructions collectively, rather than individually. We are optimistic

that some version of these strategies will succeed in the 3D problem, since animated

sequences of 2D reconstructions of the yeast cell over a few degrees of tilt (Shapiro

et al., 2005) already convey a fair degree of 3D structure.

Acknowledgement

This work was supported by Department of Energy grants DoE-FG02-05ER46198

and DoE-FG02-04ER46128 and National Institute of Health grant 1 R01 GM64846-01.

Appendix A
Multi-slice propagation formula

IUCr macros version 2.0β15: 2004/05/19

 19

We derive the multi-slice formula used in Section 2 to propagate an incident plane

wave through an object with refractive index distribution n. The radiation field Φ is

treated in the scalar wave approximation:

∇2 Φ + k 2 n 2 Φ = 0 . (13)

Here k is the wave number of the incident and elastically scattered waves. We let
e q (z) denote the Fourier transform of Φ in the dimensions transverse to the incident
Φ

radiation direction, given by z. Introducing the refractive contrast 2δn = n2 − 1,

equation (13) becomes

 
e q (z) + (k 2 − q 2 )Φ
∂z2 Φ f ∗ Φ(z)
e q (z) + 2k 2 δn e =0, (14)
q

where ∗ denotes convolution in the transverse Fourier space. The modes q are sampled

on a periodic grid for efficient computation of the convolution by 2D FFTs. In free

space (δn = 0) the most general solution of (14) has the form
√ √
e q (z) = Aq eiz k2 −q 2 k2 −q 2
Φ + Bq e−iz , (15)

where Aq = δ(q) (the Dirac delta function), Bq = 0 corresponds to the incident wave.

The scattering is calculated in the modulation approximation

√
e q (z) = Aq (z)eiz k2 −q 2
Φ , (16)

where Aq (z) = δ(q) for z < z0 , and is slowly varying for z > z0 . This precludes

calculation of the backscattered waves (second term in Eq. 15), or waves with large

q. Mathematically the approximation corresponds to the neglect of the third term in

 q  √
e q (z) = −(k 2 − q 2 )Aq (z) + 2i k 2 − q 2 ∂z Aq (z) + ∂ 2 Aq (z) eiz k2 −q 2
∂z2 Φ z . (17)

Substituting just the first two terms into Eq. 14, we obtain
√ k2  √ 
k2 −q 2 f ∗ A(z)eiz k2 −q2
∂z Aq (z)eiz ≈ ip 2 δn , (18)
k − q2 q
IUCr macros version 2.0β15: 2004/05/19
 20

a set of first order differential equations, one for each q, coupled by the convolution

term. The first order structure of the equations enables the multi-slice approach, where

the incident wave is propagated unidirectionally through the refractive medium. The

exit wave will have z-independent modulation amplitudes Aq (z), and (16) ensures

that the scattered waves have the correct wave number k. To cast (18) into a form

more amenable to computations we define

e q (z) = Φ
Ψ e q (z)e−ikz , (19)

and substitute
√

e q (z)e−iz k2 −q 2 −k
Ψ = Aq (z) (20)

into (18). The result is

q  2  
e q (z) = i
∂z Ψ e q (z) + i p k
k2 − q2 − k Ψ f ∗ Ψ(z)
δn e . (21)
k2 − q2 q

In the limit q ≪ k we recover from (21) the Eikonal approximation since

 
∂z Ψ f ∗ Ψ(z)
e q (z) = ik δn e (22)
q

implies

∂z Ψ(z) = ik δn Ψ(z) (23)

with solution
 Z z 
Ψ(z) = Ψ(z0 ) exp ik δn dz . (24)
z0

Our numerical multi-slice formula is given by the finite difference integration of

(21):
" q #
  1  
e q (z + ∆z) = Ψ
Ψ e q (z) + ik ∆z 1− q 2 /k 2 e q (z) + p
−1 Ψ f ∗ Ψ(z)
δn e .
1 − q 2 /k 2 q

(25)

IUCr macros version 2.0β15: 2004/05/19

 21

This is equivalent, to first order in ∆z, to the form that was implemented in our

computations:
" # √

 
e
Ψq (z + ∆z) = Ψe q (z) + p ik ∆z f ∗ Ψ(z)
δn e ei∆z k2 −q 2 −k
. [Equation (2)]
1 − q 2 /k 2 q

When δn = 0, this equation becomes of the same form as Eq. (16), making it valid for

arbitrary large ∆z. The exponential factor to the right is the free-space propagator.

The simulations in this work involve a specimen for which the condition k ≫ |∇δn| is

always satisfied, so that the Rytov approximation (Rytov, 1937; Davis, 1994) of Eq.

(13) should be valid. We have verified that this is indeed the case.

A.1. Implementation details

Practical implementations make use of the Fast Fourier Transform (denoted as F

here) to compute the convolution appearing in (2), so that this equation can be written

as the iteration of
e j },
∆Ψj = ik ∆z δnj × F −1 {Ψ
" # q
1 (26)
e j+1
Ψ ej + p
= Ψ × F {∆Ψ j } × exp[i∆z k 2 − q 2 − k)],
1 − q 2 /k 2

with δnj = δn(r, zj ). This particular formulation emphasizes the similarity with the

standard multi-slice formula (Cowley & Moodie, 1957; Ishizuka & Uyeda, 1977), which

does not conserve the wave number of the scattered radiation:

e j },
Ψ′j = exp [ik∆z δnj ] × F −1 {Ψ
n o (27)
e j+1 = F Ψ′ × exp[− 1 i∆z q 2 /k].
Ψ j 2

In this approximation, the square root in the second term’s denominator of equation

(2) is set to 1 and the propagator term is approximated by its first non-zero order in

q, becoming the Fresnel propagator.

Appendix B
Averaging over defocus planes
IUCr macros version 2.0β15: 2004/05/19
 22

We model one slice of the specimen being reconstructed as a uniform disk of radius

R. The goal of the first part of this calculation is to find the approximate amplitude of

the scattered field at distance z downstream from the disk and a distance r from the

axis passing through its center. The Fourier transform of the wavefield at the plane z

is given by
√

e q (0)eiz
e q (z) = Ψ k2 −q 2 −k
Ψ , (28)

e q (0) is the Fourier transform of the exit wave at z = 0. Since the specimen is
where Ψ

assumed to be a uniform disk,

(
Ψ0 if |r| < R
Ψ(r, z = 0) = (29)
0 otherwise.
e q (0) is just an Airy disk:
Then Ψ

e q (0) = Ψ0 R2 J1 (qR)
Ψ , (30)
qR

The inverse Fourier transform of (28) can be written as:

Z
1 e q (z) eir·q d2 q
Ψ(r, z) = Ψ
2π  
Z
1 2 J1 (Rq) − 21 izq 2 /k
≈ Ψ0 R e eir·q d2 q
2π qR
Z ∞ 1 2
= Ψ0 kR J0 (krx) J1 (kRx) e− 2 ikzx dx. (31)
0

Using the method of stationary phase, we find that, when R ≫ (r−R) and z ≫ (r−R),

z
Ψ(r, z) ∼ Ψ0 , (32)
k(r − R)2

which is also, as it turns out, the asymptotic behavior of the Fresnel integral for a

semi-infinite plane (that is, Fresnel diffraction of a plane wave by a straight edge).

As explained in Section 4, in the presence of noise in the Fourier data, the recon-

struction algorithm enters in a steady-state regime. In the absence of convergence, the

reconstruction is defined as an average over the Fourier estimates [Eq. (9)]. Because
IUCr macros version 2.0β15: 2004/05/19
 23

of noise, the region outside the support in these estimates is never zero but fluctuates

with an rms value ε which scales like the diffence map error. ε can than be seen as

the tolerance of the algorithm to fluctuations outside the support. The goal of this

Appendix is to point out that this finite tolerance allows in turn a range of defocus

values, as long as ε > |Ψ/Ψ0 |, that is:

z
ε > |Ψ/Ψ0 | ∼ , (33)
k(RS − R)2

where we have set r = RS , the radius of the circular support used in a reconstruction.

It is then reasonable to assume that the extent of the defocus region |z| < z0 tolerated

by the algorithm is given by z0 ∼ εk(RS − R)2 .

The last step in this calculation consists in computing the effect of averaging on the

ensemble of defocus planes. Taking z0 as the standard deviation of the distribution of

planes, we have
D 1 2 /k
E D 1 2 /k
E
e q e− 2 izq
Ψ e q e− 2 izq
= Ψ
1 2 /k 2
e q e− 4 (z0 q
= Ψ )

2 (R 4 q4
e q e−αε
= Ψ S −R)
. (34)

The averaging over defocus planes results in a low-pass filter both dependent on the

noise (through ε) and the tightness of the support, |RS − R|. It is argued at the end

of Section 5 that an effective alternative consists in translating the defocus planes,

thereby avoiding the construction of a highly detailed support.

Appendix C
Unconstrained mode analysis

In direct space there is no constraint on the value of pixels (voxels) r ∈ S, where S

is the support. Similarly, there is no Fourier space constraint at frequencies q that lie
IUCr macros version 2.0β15: 2004/05/19
 24

within the region of missing data C, usually at the center of the diffraction pattern.

This motivates the following definition for the constrained power in the reconstructed

image Ψ:
Z Z
hΨ|W |Ψi = dr |Ψ(r)|2 + e
dq |Ψ(q)| 2
, (35)
r∈S
/ q∈C
/

where
Z
e
Ψ(q) = (2π)−D/2 dr eiq·r Ψ(r) (36)

is the Fourier transformed image. W is a bounded, self-adjoint operator on the space of

square-integrable functions. Its spectrum, bounded by 0 and 2, enables us to identify

modes that are negligibly constrained in both direct and Fourier space. When the

regions S and C are sufficiently large, the relevant features of the spectrum yield to

a semi-classical, wavepacket analysis. Consider wavepacket modes inside the region S

in direct space. The density of these modes in Fourier (momentum) space is given by

the well known expression (Rayleigh, 1900; Ashcroft & Mermin, 1976)

V (S)
dM = dq , (37)
(2π)D

where V (S) is the area/volume of the region S. To count the modes that also have

negligible power in the region C of Fourier space, we integrate (37) over C:

Z
V (S)Ve (C)
M= dM = . (38)
q∈C (2π)D

When M ≫ 1, the condition for semi-classical analysis to be valid, there will be

approximately M negligibly constrained modes with W ≈ 0. We adopt the simpler

terminology “unconstrained” for these modes from now on.

The structure of the spectrum of W , for modes with higher constrained power, is

also simple. There is an infinite near-degeneracy of modes with W ≈ 1, and relatively

few modes with 0 < W < 1. This too can be understood in semi-classical terms.

Consider wavepackets of very small width in the region S. For suitably small widths,
IUCr macros version 2.0β15: 2004/05/19
 25

the number of independent wavepacket modes within S can be made arbitrarily large.

On the other hand, all of these modes will have Fourier transforms with widths so

broad that close to the maximum penalty, W ≈ 1, is incurred by the second term in

(35).

The number of unconstrained modes can be estimated from the number of missing

Fourier data samples NC in the region C, and the oversampling ratio, defined as

LD
σ= , (39)
V (S)

where L is the linear size of the direct space field of view. Since the density of Fourier

samples for this field of view is (L/2π)D , we have

 D
L
NC = Ve (C) . (40)
2π

Combining (38), (39), and (40), we obtain

NC
M= . (41)
σ

As in Weyl’s formula, for the density of modes of the Laplacian on a bounded region,

there are corrections associated with the boundary that diminish the density of modes.

We therefore expect (41) to be an overestimate when the regions S and C have large

surface to volume ratios.

To complete the correspondence between constraints in direct and Fourier space,

we can define a missing data ratio

ΛD
σ̃ = , (42)
Ve (C)
where Λ is the range of spatial frequencies in the diffraction pattern. The number of

pixels (voxels) in the support S is given by

 D
Λ
NS = V (S) , (43)
2π

and
NS
M= . (44)
σ̃
IUCr macros version 2.0β15: 2004/05/19
 26

C.1. Numerical methods for unconstrained mode computations

The two ratios associated with constraints in direct and Fourier space typically

satisfy the relationship 1 ≪ σ ≪ σ̃. From (41) and (44) we then have M ≪ NC ≪ NS ,

and neither the pixels in the support nor the missing samples in the diffraction pattern

are well matched in number to serve as efficient bases for the unconstrained modes.

Moreover, since the matrix elements of the operator W are not sparse in either of

these bases, their computation would be costly.

A better basis, and one that exploits the symmetry between direct and Fourier

space, is provided by the quantum harmonic oscillator modes. These modes have the

property of having some number of oscillations within a classically limited area, yet

they rapidly taper off outside the classical limit and thus can make a nice transition

between missing and measured data regions. We illustrate this in one dimension, for

support and missing data regions given by

S = {x : |x| < α b} C = {k : |k| < α/b} . (45)

In 2D and 3D we would construct a product basis from 1D bases specified by widths

along the Cartesian axes. The properly scaled harmonic oscillator modes in direct

space are
r
1 2
ψn (x) = Hn (x/b) e−(x/b) /2 , (46)
b
where Hn is the nth Hermite polynomial. For computing the constraint penalty in

Fourier space we use the Fourier transform

√ 2 /2
ψen (k) = in b Hn (k b) e−(k b) . (47)

Since at the boundaries of S and C we have

2 /2
ψn (α b) ∼ ψen (α/b) ∼ Hn (α) e−α , (48)

the constraint penalty is equally divided between direct and Fourier space.
IUCr macros version 2.0β15: 2004/05/19
 27

To determine the unconstrained modes the constrained power operator W is eval-

uated in the basis above and diagonalized. Since one is only interested in the M least

constrained modes, the basis can be truncated at a small multiple of M . In typical

applications NC ≪ NS , and it makes sense to evaluate the unconstrained modes on

the smaller number of Fourier samples within and surrounding the region C. The

corresponding modes in direct space may then be computed efficiently using the FFT.

C.2. Example of a mode calculation

For the 1D regions (45) we use the mode expansion

PN −1
Ψ(x) = n=0 cn ψn (x) (49)

where Ψn (x) is given by (46), and the mode number cut-off N is chosen to be a few

times the number of unconstrained modes M . Using (38) we obtain M = (2/π)α2 in

our 1D example. The truncated mode expansion of the constrained power operator

takes the form

PN −1 PN
hΨ|W |Ψi ≈ m=0 n=0 wmn cm ∗ cn , (50)

where
( R
−ξ 2
wmn = |ξ|>α Hm (ξ)Hn (ξ)e dξ if m ≡ n (mod 4)
(51)
0 otherwise.

By diagonalizing the matrix wmn we obtain constrained power eigenvalues wp and

corresponding mode amplitudes cnp for the W -eigenmodes χp (p = 1, . . . , N ):

PN −1
wp cmp = n=0 wmn cnp (52)

PN −1
χp (x) = n=0 cnp ψn (x) (53)

Figure 11 shows the results of a computation for α = 5, for which M ≈ 15.9.

IUCr macros version 2.0β15: 2004/05/19

 28

C.3. Numerical implementation

Application of the mode analysis on experimental data requires that both S and

C are well known. These sets are represented as masks in logical arrays (that is,

characteristic functions sampled on a grid):

( (
1 if r ∈ S 1 if q ∈ C
XS (r) = , XC (q) = .
0 otherwise 0 otherwise

For simplicity, we assume that the arrays are two-dimensional N × N arrays. The

main steps are:

1. Evaluation of the number of missing modes The upper bound to the

number of unconstrained modes can be written

NS NC
M= , (54)
N2
P P
where NS = r XS (r), NC = q XC (q).

2. Recentering If needed, recentering of the sets simply requires a translation of

the indices to bring to 0 the position of the “center of mass”, given by

1 P
r= rXS (r), (55)
NS r

with the obvious counterpart in Fourier space.

3. Quantum harmonic oscillator modes. As explained above, the eigenstates

of a quantum harmonic oscillator form a well-matched basis for the expansion

of the modes of the power operator. The properly normalized two dimensional

version of (46) above is

1 1
Hn (x/σx ) Hm (y/σy ) e− 2 (x /σx +y /σy ) ,
2 2 2 2
ψn,m = √ (56)
πσx σy

and in Fourier space,

r
σx σy 1
Hn (qx σx ) Hm (qy σy ) e− 2 (qx σx +qy σy ) .
2 2 2 2
ψen,m = i(n+m) (57)
π
IUCr macros version 2.0β15: 2004/05/19
 29

σx and σy can be defined as

s s
∆x ∆y
σx = , σy = , (58)
∆qx ∆qy

where

1 P 2
(∆x)2 = x XS (r)
NS r
1 P 2
(∆y)2 = y XS (r),
NS r

and the analogous definition for (∆qx )2 and (∆qy )2 .

4. Computation of the unconstrained modes The expansion of the uncon-

strained modes in the basis (56) should in general include all ψk,l−k , for all

possible k and all l up to a cutoff l0 . The matrix elements of the constrained

power operator are given by

P ∗ P e∗ e
wi,j = r XS ψ i ψ j + q XC ψ i ψ j , (59)

where ψi = ψni ,mi is a single-index relabelling of the expanding functions. Diag-

onalization of w yields the sought modes [Eq. (53) above].

References

Ashcroft, N. W. & Mermin, N. D. (1976). Solid State Physics, p. 35. Harcourt College.
Bauschke, H. H., Combettes, P. L. & Russel Luke, D. (2003). J. Opt. Soc. Am. A, pp. 1025–
1034.
Chapman, H. N., Barty, A., Marchesini, S., Noy, A., Cui, C., Howells, M. R., Rosen, R., He,
H., Spence, J. C. H., Weierstall, U., Beetz, T., Jacobsen, C. & Shapiro, D. (2005). J. Opt.
Soc. Am. A. (to be published).
Cowley, J. M. & Moodie, A. F. (1957). Acta Cryst. A, 10(10), 609–619.
Davis, T. J. (1994). Acta Cryst. A, 50, 686–690.
Eisebitt, S., Luning, J., Schlotter, W. F., Lorgen, M., Hellwig, O., Eberhardt, W. & Stohr, J.
(2004). Nature, 432(7019), 885 – 888.
Elser, V. (2003a). J. Opt. Soc. Am. A, 20(1), 40–55.
Elser, V. (2003b). Acta Cryst. A, 59(3), 201–209.
Elser, V. & Millane, R. D. (2005). In preparation.
Elser, V. & Rankenburg, I. (2005). Phys. Rev. E. Submitted to Physical Review E.
Faulkner, H. M. L. & Rodenburg, J. M. (2004). Phys. Rev. Lett. 93, 023903.
Fienup, J. R. (1978). Opt. Lett. 3(1), 27–29.
IUCr macros version 2.0β15: 2004/05/19
 30

Fienup, J. R. (1982). AO, 21(15), 2758–2769.

Fienup, J. R. (1987). J. Opt. Soc. Am. A, 4(1), 118–123.
Gerchberg, R. W. & Saxton, W. O. (1972). Optik, 35, 237.
Henke, B. L., Gullikson, E. M. & Davis, J. C. (1993). Atomic Data and Nuclear Data Tables,
54, 181–342.
Ishizuka, K. & Uyeda, N. (1977). Acta Cryst. A, 33, 740–749.
Marchesini, S., He, H., Chapman, H. N., Hau-Riege, S. P., Noy, A., Howells, M. R., Weierstall,
U. & Spence, J. C. H. (2003). Phys. Rev. B, 68(140101).
McNulty, I., Kirz, J., Jacobsen, C., Anderson, E. H., Howells, M. R. & Kern, D. P. (1992).
Science, 256, 1009–1012.
Miao, J., Charalambous, P., Kirz, J. & Sayre, D. (1999). Nature, 400, 342–344.
Miao, J., Hodgson, K. O., Ishikawa, T., Larabell, C. A., LeGros, M. L. & Nishino, Y. (2003).
Proc. Nat. Ac. Sci. 100(1), 110–112.
Miao, J., Tetsuya Ishikawa, Bart Johnson, Erik H. Anderson, Barry Lai & Keith O. Hodgson
(2002). Phys. Rev. Lett. 89(088303), 1–4.
Miao, J., Yoshinori Nishino, Yoshiki Kohmura, Bart Johnson, Changyong Song, Subhash
H. Risbud & Tetsuya Ishikawa (2005). Phys. Rev. Lett. 95(085503), 1–4.
Miao, J. & Sayre, D. (2000). Acta Cryst. A, 56, 596–506.
Rayleigh, J. W. S. (1900). Phil. Mag. 49, 539 – 540. Republished by Dover (1964).
Russel Luke, D. (2005). Inverse Problems, 21(1), 37–50.
Rytov, S. M. (1937). Izv. Akad. Nauk. SSR Ser. Fiz. 2, 223.
Sayre, D. (1980). In Imaging Processes and Coherence in Physics, edited by J. Schlenker,
M. Fink, J. P. Goedgebuer, V. Malgrange, J. C. Viénot & R. H. Wade, vol. 112, pp. 229
– 235. Berlin: Springer-Verlag.
Shapiro, D., Thibault, P., Beetz, T., Elser, V., Howells, M. R., Jacobsen, C., Kirz, J., Lima, E.,
Miao, H., Nieman, A. M. & Sayre, D. (2005). Proc. Nat. Ac. Sci. 102(43), 15343–15346.
Spence, J. C. H., Weierstall, U. & Howells, M. R. (2002). Phil. Trans. R. Soc. Lond. A, 360,
875–895.
Stroke, G. W. & Falconer, D. G. (1964). Phys. Lett. 13(4), 306–309.
Williams, G. J., Pfeifer, M. A., Vartanyants, I. A. & Robinson, I. K. (2003). Phys. Rev. Lett.
90(175501), 1–4.
Wu, J. S., Weierstall, U., Spence, J. C. H. & Koch, C. T. (2004). Opt. Lett. 29(23), 2737–2739.

IUCr macros version 2.0β15: 2004/05/19

 31

Fig. 1. Soft x-ray (λ = 1.65 nm) diffraction pattern of a freeze dried yeast cell, on a log-
arithmic scale (Shapiro et al., 2005). This 1200 × 1200 array extends to (20.7 nm)−1
on the sides, giving the corresponding real-space array 10.3 nm wide pixels. Inset,
left: Magnified portion of the diffraction pattern showing the speckles. Inset, right:
Magnified central region showing the diamond-shaped missing data region.

IUCr macros version 2.0β15: 2004/05/19

 32

Fig. 2. Real part of the autocorrelation. This image is the high-pass filtered inverse
Fourier transform of the diffraction pattern shown in figure 1 (the high-pass filter
reduces the effect of the sharp discontinuity due to missing central data). The
central oval-shaped structure is the autocorrelation of the cell. Its contour is used
to determine the size of the support and to calculate the overdetermination ratio.
The surrounding shapes are faint images of the cell caused by the interference with
point-like scatterers around the cell.

IUCr macros version 2.0β15: 2004/05/19

 33

(a) (c)
ImΨ

ReΨ

(b) (d ) ImΨ

ReΨ

Fig. 3. Yeast cell model. (a) Slice of the model, through the origin. (b) Simulated wave-
field at the focus plane (i.e. exit wave back-propagated to the center of the cell). The
image is indistinguishable from the exit wave (not shown). (c) Spiral distribution
of the values in the complex plane at the focus plane. (d ) Value distribution at the
exit plane. The structure close to the origin is caused by free propagation just out-
side the support region (see figure 4). Figures (a) and (b) represent complex-valued
objects. The colors are defined such that phase is mapped to hue and magnitude is
mapped to brightness.

IUCr macros version 2.0β15: 2004/05/19

 34

(a) (b)

(c)

Fig. 4. Wavefield propagation simulations. (a) Longitudinal section through the center
of the cell (outlined in white) of the simulated wavefield perturbation. (b) Longitu-
dinal section of the simulated free-space forward and backward propagation of the
wavefield. (c) Same as (b) but with the higher amplitude values truncated to see
the perturbation propagating out of the cell. The coloring scheme is the same as in
Fig. 3. In (b) and (c), the propagation axis has been compressed by a factor 5.

Mode 1 (0.00024) Mode 2 (0.0033)

Mode 3 (0.0044) Mode 4 (0.0162)

Fig. 5. The 4 most weakly constrained modes (constrained power in parentheses). Each
mode is shown in direct and Fourier space. The phase of the mode is illustrated by
the hue and the amplitude by the saturation. Both the support (in real space), and
region of measured data (in Fourier space), are superimposed in gray.

IUCr macros version 2.0β15: 2004/05/19

 35

(a) (b) (c)

Fig. 6. Illustration of the effect of the unconstrained modes. (a) Reconstruction with
the freely varying mode amplitudes, (b) after the 4 least constrained modes have
been projected out, (c) after restoration of the modes using the ad hoc variance
minimization rule. Both (b) and (c) are reproducible but (b) has a distracting and
unnatural appearance.

IUCr macros version 2.0β15: 2004/05/19

 36

1µm

(a)
0.55 1

0.5 ǫn

ǫn 0.45 0
0 50000 100000
Iteration number

0.4

0.35
0 5000 10000
Iteration number
(b)

Fig. 7. (a) Yeast cell reconstruction, based on an average of 1000 Fourier estimates.
The weakly constrained modes have been restored using the variance minimization
rule. As in previous figures, the phase is represented by the hue and the magnitude
by the brightness. (b) Difference map error (ǫn ) for the first 10000 iterates (thin
line). The bold line is a running average (over a 1000 iterate window) to emphasize
the decay of the transient. The dotted line is the overall average, ǫ = 0.384. Inset:
the difference map error for the totality of the run. As explained in the text, the
overall scale of the error can be considered as arbitrary because of the missing
central data.

IUCr macros version 2.0β15: 2004/05/19

 37
Full period (nm)
1000 100 50 20
0.8
MTF

0.6
hIrec /Iobs i

0.4

15 nm

0.2
30 nm

1000 0
1 2 5 10 20 50
Spatial frequency (µm−1 )

Fig. 8. Resolution decrease caused by averaging . The black line shows the relative
decrease of the reconstructed intensity, as a result of averaging over residual phase
fluctuations. The two dotted lines show the classical MTF for an incoherent imaging
system with 75% efficiency and a Rayleigh resolution of 15 and 30 nm respectively.
[from Shapiro et al. (2005)]

IUCr macros version 2.0β15: 2004/05/19

 38

(a) (b)
position (µm)
−2 −1 0 1 2

15
Amplitude

10

0
400 450 500 550 600 650 700 750 800
pixel index
(c)

Fig. 9. Comparison of the reconstructions obtained with (a) β = −1 and (b) β = 1.

The dotted lines show the position of the cross section plotted below. (c) Cross
section of the absolute value of the two reconstructions for β = −1 (blue) and
β = 1 (red).

IUCr macros version 2.0β15: 2004/05/19

 39

Fig. 10. Reconstruction magnitudes of the small scatterers surrounding the cell, sat-
urated to allow weaker points to be seen. This image is the average of 400 Fourier
estimates.

hχi |W |χi i x
1 0 2 4 6 8 10

0.8 χ17
0.6
0.4
0.2
i χ14
5 10 15 20 25 30

Fig. 11. Left: Spectrum of the constrained power operator W in one dimension for
support |x| < 5 and missing frequencies |k| < 5. Right: Comparison of modes χ14
and χ17 shows the onset of constraint as power invades the region x > 5.
IUCr macros version 2.0β15: 2004/05/19
 40

Synopsis

IUCr macros version 2.0β15: 2004/05/19