ARTICLES

PUBLISHED ONLINE: 13 JULY 2015 | DOI: 10.1038/NPHOTON.2015.112

Seeing through chaos in multimode fibres

Martin Plöschner1, Tomáš Tyc2 and Tomáš Čižmár1*

In a similar fashion to diffusers or other highly scattering media, multimode fibres deliver coherent light signals in the
form of apparently random speckled patterns. In contrast to other optically random environments, multimode fibres
feature remarkably faithful cylindrical symmetry. Our experimental studies challenge the commonly held notion that
classifies multimode fibres as unpredictable optical systems. Instead, we demonstrate that commercially available
multimode fibres are capable of performing as extremely precise optical components. We show that, with a sufficiently
accurate theoretical model, light propagation within straight or even significantly deformed segments of multimode fibres
may be predicted up to distances in excess of hundreds of millimetres. Harnessing this newly discovered predictability in
imaging, we demonstrate the unparalleled power of multimode fibre-based endoscopes, which offer exceptional
performance both in terms of resolution and instrument footprint. These results thus pave the way for numerous exciting
applications, including high-quality imaging deep inside motile organisms.

T
he theoretical description of light transport processes within memory of a computer, all light transport processes can be emulated
ideal multimode fibres (MMFs) has been developed for over with great accuracy, and the optical fibre can be subjected to
half a century1–4. This elaborate theoretical model is, detailed investigation24.
however, frequently considered inadequate to describe real-life While working with a relatively large numbers of modes (∼500),
MMFs, which are manufactured by drawing melted silica preforms. we initially used the shortest practical fibre segment (∼10 mm) in
Such fibres are commonly seen as unreliable, and the inherent ran- order to minimize the complexity of the problem. This intermediate
domization of light propagating through them is typically attributed step allowed us to eliminate influences caused by imperfections
to undetectable deviations from the ideal fibre structure. It is a com- (unavoidable misalignment) in our experimental settings and,
monly held belief that this additional chaos is unpredictable and importantly, establish the parameters of the fibre (core diameter,
that its influence grows with the length of the fibre. Despite this, numerical aperture) with sufficiently high precision.
light transport through MMFs remains deterministic. Only when corrections for these aspects were implemented could
The prospect of deterministic light propagation within MMFs we see a perfect agreement with the theoretical model used. This not
has only recently been used through methods of digital holography only confirmed the existence of propagation-invariant modes
and by adopting the concept of empirical measurement of the trans- within the fibre, but also matched their output phases very accu-
formation matrix (TM)5–11. This technique, developed in studies of rately, which is vitally important for imaging applications.
light propagation through highly turbid media12–17, has opened a Progressing to ∼100-mm-long fibre segments, we have faced the
new window of opportunity for MMFs to become extremely first challenges related to polarization coupling and significant devi-
narrow and minimally invasive endoscopes, allowing sub-micro- ations of the refractive index from the ideal step-index profile, which
metre resolution imaging in deep regions of sensitive tissues9,18. required major enhancement of our numerical modelling and
However attractive, this technology suffers from several major experimental methods.
limitations, the most critical being the lack of flexible operation. With such a newly developed ability to predict light propagation,
Any bending or looping of the fibre results in changes to its TM, even through such distances within the fibre, we could, for the first
rendering the imaging heavily impaired. All current methods time, rigorously investigate the influence of fibre deformation
exploiting MMFs for imaging require open optical access to the (bending) on the resulting TM. The study has shown that even sig-
distal end of the fibre during the time-consuming measurement of nificantly bent fibres are perfectly predictable, thus allowing compu-
the TM. Furthermore, this characterization must be repeated tation of their TM based purely on the observation of the fibre
upfront for every intended configuration (deformation) and any geometry. Finally, all of these new aspects were brought together
axial distance of the focal plane behind the fibre before the system to test the performance of imaging in such an enhanced system.
can be used for imaging7,19. The necessity to determine the TM We have shown that imaging can be achieved without experimental
empirically is therefore a major bottleneck of the technology, and TM acquisition, in both straight and deformed fibres at an arbitrary
it would be immensely advantageous to obtain the TM by another distance behind the distal fibre facet.
route, ideally on the basis of numerical modelling.
The aim of this Article is to determine whether such modelling is Identification of propagation-invariant modes
feasible in the supposedly chaotic environment of MMFs, given that, Our experimental geometry, introduced in the Methods, allows
unlike other random media, they feature a remarkable cylindrical measurement of the TM in the representation of diffraction-
symmetry. Indeed, numerous studies have already indicated that limited focal points (FPs), sometimes also called ‘delta peaks’,
at least some aspects of the ordered behaviour (propagation which, from the experimental perspective, are the most convenient
constants of modes) can ‘survive’ over very large distances20–23. choice. The input focal points (at the proximal end of the fibre) can
Our study represents a major advancement on these efforts. With be generated using a spatial light modulator (SLM) that simply
the availability of experimentally measured TMs stored in the steers a focused laser beam into a required location at the input

1
School of Engineering, Physics and Mathematics, College of Art, Science & Engineering, University of Dundee, Nethergate, Dundee DD1 4HN, UK.
2
Department of Theoretical Physics and Astrophysics, Masaryk University, Kotlarska 2, 61137 Brno, Czech Republic. * e-mail: [email protected]

NATURE PHOTONICS | VOL 9 | AUGUST 2015 | www.nature.com/naturephotonics 529

© 2015 Macmillan Publishers Limited. All rights reserved

ARTICLES NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112

a
1

b Acquired transformation matrix c Propagation-invariant modes

4

Amplitude
m-index
0

l-index −8 −6 −4 −2 0 2 4 6 8 10
d Conversion matrix e Iteration 1 f Final iteration

0
−π phase π

Figure 1 | Analysing a short segment of fibre. a, Organization of input and output modes. b, Experimentally measured TM. c, Theoretically predicted
LP PIMs. d, Conversion matrix between the representation of focal points and LP PIMs. e,f, Converted TM before (e) and after (f) the optimization
procedure, respectively.

fibre facet. The output (distal) facet is imaged on a charge-coupled transformation matrix M, its conversion into the representation of
device (CCD) chip, and focal points are acquired from the values of PIMs should result in a purely diagonal matrix, indicating that
individual pixels. each input mode is perfectly conserved.
The chosen set of focal points is arranged across an orthogonal The converted transformation matrix M
0 = T M T† is shown in
grid as shown in Fig. 1a, and ordered as indicated by the red line.
0 is not diagonal, which might lead to the con-
Fig. 1e. Apparently, M
The experimentally measured transformation matrix M for a clusion that the optical fibre does not follow the theoretical model.
10-mm-long fibre segment is shown in Fig. 1b. Each row of M rep- However, off-diagonal components can also appear as a result of
resents the amplitudes and phases of all output focal points for a even a very small misalignment of the fibre. The misalignment
single input focal point mode sent into the fibre. Due to space con- space of the degrees of freedom is very large, comprising three-
straints, here (and also in Fig. 1d–f ), the basis of modes for the TMs dimensional position, two tilts and one defocus, each on both
shown has been reduced to one-third of the full dimension. The sides of the fibre. Moreover, these 12 degrees of freedom are intrin-
complete transformation matrices are presented in Supplementary sically intertwined with uncertainty in the radius of the fibre core
Figs 5–7. Following TM acquisition we can numerically emulate (a) and the numerical aperture (NA). We have developed an
the optical fibre as an optical system and predict the outcome of optimization procedure (described in the Methods; source available
any optical field being sent into the fibre and thus validate the at http://complexphotonics.dundee.ac.uk/) that simultaneously cor-
correctness of any theoretical prediction. rects the TM for alignment imperfections and adjusts the values of a
The simplest form of theoretical description (see the scalar and and NA. The optimized result, M
final (Fig. 1f ), carries 93% of the
paraxial approach reviewed in Supplementary Methods 1) predicts optical power on the main diagonal, showing an excellent match
the existence of linearly polarized modes that do not change their between the scalar theoretical prediction and the corrected
field distribution during propagation through the fibre. A series of experimental data.
such propagation-invariant modes (PIMs), also known as eigen-
vectors of propagation operators or eigenmodes, are shown in Polarization coupling of modes
Fig. 1c. Again, due to space constraints we only show PIMs of a Each mode can be defined in two orthogonal polarization states, and
fibre with many fewer modes. PIMs are defined by a pair of only when both are taken into account can the TM be considered
indices m and l. Index l refers to the orbital angular momentum complete. An explanation of how the polarization in our geometry
of a given mode with magnitude equal to lħ/photon25. can be controlled to take such complete measurements of the TM is
Whether such modes remain unchanged after travelling through provided in Methods. After optimization and conversion into PIMs,
the real fibre can now be tested using the experimentally measured such a complete TM will now have four quadrants. Those contain-
TM. Each of these theoretically predicted PIMs can be constructed ing the main diagonal indicate that the polarization of a given mode
as a superposition of input focal points. Such a vector can then be has been conserved, and the remaining ones indicate mutual coup-
sent virtually into the fibre, which is implemented by its matrix mul- ling between polarization states. The optimized TM with input
tiplication with the TM. The resulting output vector of focal points PIMs defined by two orthogonal linear polarization states for the
should contain the identical PIM, differing only by a phase constant. 10-mm-long fibre is presented in Fig. 2a (see Supplementary
Carrying out such an operation for all modes simultaneously is Figs 7–10 for complete data sets). Coupling between polarization
mathematically equivalent to converting the experimentally states is clearly present, but relatively weak. The change in the polar-
measured M into the representation of PIMs. This is achieved by ization state of individual linearly polarized PIMs (LP PIMs) can be
constructing a conversion matrix T (shown in Fig. 1d), in which efficiently visualized on the Poincaré sphere shown in Fig. 2b. The
each line represents a single theoretically predicted PIM expressed polarizations of all modes remain linear, but their orientation is
as a superposition of input focal points. PIMs in T are ordered as rotated by up to 45° (corresponding to a shift of 90° along the
indicated by the white line in Fig. 1c. If the theoretically predicted equator on the Poincaré sphere). The same data are visualized by
PIMs are the true eigenvectors of the experimentally measured placing the polarization states of the output PIMs into the shape

530 NATURE PHOTONICS | VOL 9 | AUGUST 2015 | www.nature.com/naturephotonics

© 2015 Macmillan Publishers Limited. All rights reserved

NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112 ARTICLES
10 mm, linear polarization S3 100 mm, linear polarization S3
a d e
b

ψS

ψS
ψS

ψS
→

→
→

→
ψ

ψ
ψ

ψ
P

P
S

ψP
ψP
ψP

ψP
S1 S2 S1 S2

→
→
→

ψ
ψ
ψ

P
S
S

P
c f

100 mm, circular polarization S3 SOI – experiment Synthesis of PIM

g j 10 π m
h σ = −1
ψ−
ψ−

m
→
1
→
1

ψ
ψ

+1
−1

0 −π
SOI – simulation
k 10 π
Input → output

m
σ = +1
0 −π
ψ+
ψ+

S1 S2 Amplitudes of modes
→
→

1
1

l 10 1
ψ
ψ

+1
−1

m
0 0
l→ −10 0 10 20

Figure 2 | Polarization coupling effects in MMF. a–c, Data for the 10-mm-long fibre: TM of LP PIMs (a); Poincaré sphere depicting polarization change of
LP PIMs (b); projections of LP PIM polarization state on the mode pyramid (c). d–f, As in a–c, but for the 100-mm-long fibre. g–i, Data for the 100-mm-long
fibre using CP PIMs. j, Experimentally measured phase difference between CP PIMs with opposite spin (influence of the SOI). k, Numerically simulated
equivalent of j. l, CP PIMs output amplitude. m, Experimental synthesis of input PIMs and their corresponding output.

of the l–m pyramid defined in Fig. 1c. The colour defining the polar- tested our results in our experimental geometry, taking into
ization state corresponds to the colour of the Poincaré sphere in account all the detected alignment imperfections, and experimentally
Fig. 2b. synthesized the full orthogonal set of CP PIMs (see Methods for
Equivalent measurements were taken for the 100-mm-long fibre details). The PIMs were recorded in both circular polarization
(Fig. 2d–f). Once again, the output polarization is observed to be states before entering the fibre as well as after leaving the fibre. One
linear (for all PIMs except for the ‘misbehaving’ modes of l = ±1, as example showing a superposition of two such modes with the same
discussed later), but this time the rotation of polarization is much m but opposite l indices is presented in Fig. 2m. The full sets of
stronger. Although this effect remains highly ordered, the observed PIMs for the 10- and 100-mm-long fibre segments are presented in
polarization change proves that the LP PIMs can no longer be con- Supplementary Movies 2–5.
sidered truly propagation-invariant. To account for this effect, we
enhanced the theoretical description to the fully vectorial model Examination of optical phases
(Supplementary Methods 2). This advanced description shows that Evaluation of the ability to predict the optical phases of PIMs is of
only circularly polarized (CP) PIMs remain unaffected by propa- paramount importance to this study and conclusive in answering
gation through the fibre. Coincidentally, CP PIMs have an almost whether theoretically predicted TMs could be used for imaging.
identical distribution of the field amplitude and phase as LP PIMs, The phases of CP PIMS measured in the 10-mm-long fibre
which considerably simplifies their modelling. The experimental segment are shown in Fig. 3a. The phases are wrapped within an
study using circularly polarized modes propagating through a interval of 〈–π,π]. In consequence, no ordered behaviour is immedi-
100-mm-long segment of fibre is summarized in Fig. 2g–i. As ately apparent. The only sign of the ordered behaviour is the mirror
expected, with these CP PIMs we no longer see transitions far from symmetry. That can only be seen because the fibre is very short and
the original polarization state. The CP PIMs thus indeed represent therefore the contribution of the SOI is negligible. Simulations for
propagation-invariant modes, as predicted by our vectorial theoreti- the phases of PIMs using our numerical model are shown in
cal model. Interestingly, there is a broken symmetry between the Fig. 3b (unwrapped) and Fig. 3c (wrapped). The difference
modes with the same l and m indices but opposite circular polariz- between the simulation and experiment is shown in Fig. 3d. This
ation state (spin). Their optical fields are no longer symmetric, as very good agreement only occurs if the length of the fibre is
the wavefront helicity and spin are oriented the same way in one known with very high precision (on the order of units of micro-
case, but opposite in the other. The vectorial description predicts metres). We determined the fibre length by seeking the highest
that the degeneracy in such pairs of modes is removed and, as a con- value of the quantity referred to as total phase agreement (PA):
sequence, they do not propagate with the same phase velocity4,26. This  2
effect is well known in quantum optics as the spin–orbit interaction 1
N 
 i·Δfj 
(SOI). Such behaviour can be very precisely identified in our exper- PA =  e  (1)
N j=1 
imental data (Fig. 2j) and also predicted by our theoretical model
(Fig. 2k). Figure 2l shows the amplitudes of the output PIMs (diag-
onal components of the TM in Fig. 2g). The most dominant losses where N is the number of PIMs and Δϕj is the phase difference
are seen in the modes of l = ±1, which, as explained in between the experimentally measured and theoretically predicted
Supplementary Methods 2, are caused by the highly exceptional be- phases of jth PIM. This quantity equals 1 for perfect agreement.
haviour of these modes, for which the CP PIMs-based model is not PA as a function of assumed fibre length is plotted in Fig. 3e. Its
valid. To conclusively confirm the correctness of these findings we peak value of ∼83% is very satisfactory, given that the propagation

NATURE PHOTONICS | VOL 9 | AUGUST 2015 | www.nature.com/naturephotonics 531

© 2015 Macmillan Publishers Limited. All rights reserved

ARTICLES NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112

10 mm
a Experiment c Simulation wrapped e Phase agreement
m
10 π 10 π 1

m
0 −π 0 −π

PA
b Simulation unwrapped d Difference – Δϕ
10 160 10 π
×2π
m

m
0 0 0 −π 0
l→ −10 0 10 20 l→ −10 0 10 20 0 10 15 20
Length (mm)
100 mm
f Profile of refractive index g Phase agreement h Difference – Δϕ
0.1 10 π
1.45
nc

m
Ideal step-index n0
n

1.43 0 0 −π
i j k
1 10 π
1.45

m
n

Dopant diffusion
1.43 0 0 −π
l 1 × 10−5 m n
1 10 π
n–nc

m
Fine modulation
−2 × 10−5 0 0 −π
0 20 30 0 100 150 200 l→ −10 0 10 20
Radius (μm) Length (mm)
300 mm
o Profile of refractive index p Phase agreement q Difference – Δϕ
1 × 10−5 1 10 π
n–nc

m
Fine modulation
−2 × 10−5 0 0 −π
0 20 30 0 300 450 600 l→ −10 0 10 20
Radius (μm) Length (mm)

Figure 3 | Optical phases of PIMs. a–e, Data for the 10-mm-long fibre: experimentally measured phases (a); phases predicted by the numerical model,
unwrapped (b) and 2π-wrapped (b); difference between experimentally measured and theoretically predicted phases (d); phase agreement (e). f–n, Data for
the 100-mm-long fibre: assumed profile of the refractive index, phase agreement and difference between experimentally obtained and theoretically predicted
phases of PIMs for an ideal step-index fibre (f–h), a model including dopant diffusion (i–k), and with correction for diffusion as well as fine index modulation
across the fibre core (l–n). o–q, As in l–n, but for the 300-mm-long fibre.

constants of PIMs have been matched with a relative accuracy better elements of this series are required to entirely explain the observed
than 10–5. Moreover, it is clearly seen that the deviations are not phase deviations. Even with such minuscule modifications to the
random (noisy), but instead form a smooth surface (Fig. 3d). refractive index profile (shown in Fig. 3l), the PA was enhanced
Repeating the same experiment with the 100-mm-long fibre beyond the value of 0.95 (Fig. 3m). The difference between the
results in the PA peaking around the expected value of fibre experimental and theoretical phases of the PIMs is shown in
length (Fig. 3g), but the PA peak value has decreased dramatically. Fig. 3n. On further extending the length of the fibre to 300 mm,
The difference between the experimental and theoretical phases is we found that for the satisfactory agreement shown in Fig. 3o–q
shown in Fig. 3h, which clearly demonstrates that the aberration we had to add one more free parameter of fine modulation.
seen in Fig. 3d has grown proportionally with fibre length. This
effect could not be explained by any aberration of our optical Transformation matrix for deformed fibres
system. We therefore conjectured that the observed phase differ- Our experimental geometry was designed to allow bending of the
ences could be caused by deviations of the refractive index from optical fibre with curvatures reaching the long-term damage
the ideal step-index profile shown in Fig. 3f, originating, for threshold given by the manufacturer. We experimentally measured
example, from the diffusion of dopant material between core and TMs for several different fibre geometries by translating and rotating
cladding27,28 (Fig. 3i). To quantify this deviation we used pertur- the output end of the fibre, which was monitored with high pre-
bation calculus, as explained in the Methods. The diffusion is com- cision. Using this information we reconstructed the shape of the
monly quantified by the diffusion length d, which became an fibre numerically by minimizing its elastic energy. These shapes
additional free parameter in our optimization procedure are represented in Fig. 4a by solid lines, complemented by the
(see Methods). With this new model taking diffusion into actual layout of the fibre (marked by black dots) obtained from
account, the PA was enhanced substantially (Fig. 3j), but it still camera images. The changes introduced to the TM by such defor-
could not fully explain the observed deviation seen in Fig. 3k. We mations can best be characterized by a deformation operator (DO, D̂),
therefore searched for further deviations of the refractive index which describes the transition of the straight fibre TM (M
s ) to the
within the fibre core, and using perturbation calculus we arrived TM of the bent fibre (M
b ):
at their description using zero-order Zernike polynomials
(Z40 (r/a), Z60 (r/a), Z80 (r/a), … ), finding that only the first two M
s

b = D̂M (2)

532 NATURE PHOTONICS | VOL 9 | AUGUST 2015 | www.nature.com/naturephotonics

© 2015 Macmillan Publishers Limited. All rights reserved

NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112 ARTICLES
a Deformation – bending b DO (V) – experiment c DO (V) – simulation
1
100 I
II
90
III
80
IV
70

60
z (mm)

V
50

40

30

Amplitude
20 e f
PIMs – experiment PIMs – simulation
10 I I
0
0 20 40 60 II II
x (mm)
d Curvature agreement
35 III III
Optimized curvature (m−1)

30 I
II
25 III
IV IV IV
20 V
15
10
V V
5 ξ = 0.77 ± 0.02
0 0
−π π
0 5 10 15 20 25 30 35 −20 −10 0 10 20 −20 −10 0 10 20 Phase
Observed curvature (m−1) l-index l-index

Figure 4 | Influence of fibre deformation. a, Arrangements of deformed fibre used in the experiment. b,c, Experimentally measured and theoretically
predicted DO corresponding to deformation V. d, Empiric estimation of scaling factor ξ. e,f, Experimentally measured (e) and theoretically predicted (f)
influence of deformation on PIMs.

An example of an experimentally measured DO (in the represen- 1 – (1 – 2σ)(n – 1)/n, where σ is Poisson’s ratio (Methods). With a
tation of PIMs) is shown in Fig. 4b (with only one polarization com- value of σ = 0.17, which is a typical material constant of fused silica
ponent; for a full DO see Supplementary Fig. 12), which corresponds given by many manufacturers, we reached a value of ξ = 0.79,
to the largest curvature (V) introduced to the fibre. It is clear that, which is in perfect agreement with our experiment. The theoretically
even for a curvature reaching the damage threshold, the DO predicted DO corresponding to fibre deformation V (with the
remains highly diagonal, indicating that a majority of PIMs were empirical scaling factor applied) is shown in Fig. 4c (see
very well conserved. The main effect of deformation is therefore Supplementary Fig. 13 for all polarization components), and the
only a small relative phase shift of the PIMs. Diagonal components diagonal components for the theoretically predicted DOs are pre-
of the DO (influencing the PIMs) for all fibre deformations are pre- sented in Fig. 4f.
sented in Fig. 4e. Apparently, PIMs with low l and m indices are the Importantly, the theoretical model predicts that the DO does not
most vulnerable to such deformations. Our theoretical description of change if the plane in which the deformed fibre lies is oriented dif-
the bent fibres is detailed in Methods. This analytical model can ferently. This was confirmed by extending this study into the third
directly handle only those fibre segments with uniform curvature. dimension, as presented in Supplementary Results 1.
Therefore, to model our experimental conditions as authentically
as possible, we constructed a corresponding theoretical TM from Application to imaging
300 segments, each with constant curvature corresponding to the Being able to predict the correct spatial distribution of the optical
minimal elastic energy fibre layout (Fig. 4a). The results of the simu- fields of PIMs together with their correct phases after propagating
lation indicated effects of the same kind as observed experimentally, through the fibre allows us to construct a TM of a MMF in any rep-
but noticeably stronger. We therefore searched numerically for an resentation of modes, purely based on the grounds of theoretical
empirical scaling factor with which to reduce the curvature of the modelling. For imaging purposes the representation of focal
fibre in our model to obtain optimal agreement with the experiment. points is the natural choice. In this representation, each column of
For all our fibre deformations (II–V) the scaling factors were very the TM tells us what superposition of input focal points we need
similar (Fig. 4d), with a value estimated to be 0.77 ± 0.02. Further to provide in order to get a single diffraction-limited focal point
theoretical investigation revealed that this effect could be caused by at the distal end of the fibre. Consequently, we can design a corre-
linear variation of the refractive index with respect to the material sponding holographic modulation for our SLM (taking into
density, which changes under the influence of deformation-induced account all detected misalignment imperfections) to synthesize
stress. This leads to exactly the same effect as fibre bending, but with such an output focal point experimentally. Using various forms of
opposite action. Assuming that (n – 1) is directly proportional to theoretically predicted TMs, we synthesized a set of focal points
the material density29, the scaling factor can be expressed as ξ = across an orthogonal grid of 75 × 75 different positions, which

NATURE PHOTONICS | VOL 9 | AUGUST 2015 | www.nature.com/naturephotonics 533

© 2015 Macmillan Publishers Limited. All rights reserved

ARTICLES NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112

Empiric TM Theoretically predicted TM

Ideal step index Diffusion corrected Full correction
a b c d

10 mm

50 μm

e f g h
100 mm

i j k l
300 mm

Figure 5 | Application to imaging. a–l, Imaging of USAF 1951 resolution target is performed for three lengths of fibre: 10 mm (a–d), 100 mm (e–h) and
300 mm (i–l). In each case we show imaging with an experimentally acquired TM (a,e,i) and three theoretically predicted TMs corresponding to an ideal
step-index fibre (b,f,j), a profile with correction for dopant diffusion (c,g,k) and a profile with full correction (d,h,l).

were sequentially lit by the SLM to efficiently raster-scan the object profile (Fig. 5d). We see that the imaging was successful in all cases
during image acquisition. Each pixel value of the resultant image and that a small aberration due to refractive index profile imperfections
was obtained by measuring the total optical power transmitted (Fig. 3d) has a minimal influence on imaging quality. Equivalent tests
through the object while exposed to a single focal point. More were performed for 100-mm-long fibre (Fig. 5e–h) and 300-mm-long
details on this approach, as well as a demonstration of proximal fibre (Fig. 5i–l). Using the theoretically predicted TM of ideal step-
end imaging, are provided in Supplementary Results 2 and index fibre at these length scales entirely fails to produce a recognizable
Supplementary Movie 7. For our demonstration of imaging we image, which clearly demonstrates the importance of the phase correc-
chose a negative USAF 1951 resolution target as the imaging tions studied earlier. Much better but still heavily impaired imaging
object, which was placed in close proximity to the distal end of can be obtained after taking into account the diffusion effect in the
the fibre. In Supplementary Results 3 we show that imaging is not refractive index profile (Fig. 5g,k). However, only after implementing
restricted to the proximity of the fibre facet, but with the use of a the additional correction for fine modulation of the refractive index
free-space propagation operator the image plane can be easily (Fig. 5h,l) does the imaging quality approach that of an experimentally
shifted to an arbitrary distance behind the fibre facet. measured TM (Fig. 5e,i).
Our evaluation began with the 10-mm-long fibre. Figure 5a The influence of bending is demonstrated in Fig. 6a, where the
shows traditional imaging with an experimentally measured TM measured on a straight segment of fibre was used for imaging
TM7–9,18. The imaging performance of this was compared with after the fibre was bent, as illustrated by deformation V in Fig. 4a.
images obtained using three different theoretically predicted TMs The results obtained after applying an appropriate theoretically pre-
that correspond to the ideal step-index profile (Fig. 5b), the profile dicted DO (Fig. 4c) are shown in Fig. 6b. We further tested the DO
with diffusion of dopant material (Fig. 5c), and the fully corrected with all off-diagonal components set to zero (Fig. 6c), and also after

a Deformed fibre (V) b Full correction c Diagonal correction d Phase-only correction

Figure 6 | Imaging with deformed fibre. a, Imaging with an empirical TM for a straight fibre. b, Imaging with TM after full DO correction. c, TM with only
diagonal components of the DO applied. d, TM with only phases of the diagonal components of DO applied.

534 NATURE PHOTONICS | VOL 9 | AUGUST 2015 | www.nature.com/naturephotonics

© 2015 Macmillan Publishers Limited. All rights reserved

NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112 ARTICLES
setting the absolute values of the diagonal components to unity 6. Čižmár, T. & Dholakia, K. Shaping the light transmission through a multimode
(Fig. 6d). Both simplifications had negligible effects on the optical fibre: complex transformation analysis and applications in biophotonics.
Opt. Express 19, 18871–18884 (2011).
imaging performance, thus confirming that DOs are highly diagonal 7. Čižmár, T. & Dholakia, K. Exploiting multimode waveguides for pure fibre-
and almost perfectly commute with one another. This will immen- based imaging. Nature Commun. 3, 1027 (2012).
sely simplify predicting the DOs of more complex deformations, 8. Bianchi, S. & Di Leonardo, R. A multi-mode fiber probe for holographic
because the order in which individual bends appear along the micromanipulation and microscopy. Lab Chip 12, 635–639 (2012).
fibre will not play an important role. 9. Choi, Y. et al. Scanner-free and wide-field endoscopic imaging by using a single
multimode optical fiber. Phys. Rev. Lett. 109, 203901 (2012).
10. Papadopoulos, I. N., Farahi, S., Moser, C. & Psaltis, D. Focusing and scanning
Discussion light through a multimode optical fiber using digital phase conjugation. Opt.
In this Article we have demonstrated that MMFs are highly predict- Express 20, 10583–10590 (2012).
able optical systems. We have developed new methods to measure 11. Mahalati, R. N., Gu, R. Y. & Kahn, J. M. Resolution limits for imaging through
multi-mode fiber. Opt. Express 21, 1656–1668 (2013).
and analyse the TMs of real-life MMFs, as well as a complex theor- 12. Vellekoop, I. M. & Mosk, A. P. Focusing coherent light through opaque strongly
etical framework to describe the light-transport processes within, scattering media. Opt. Lett. 32, 2309–2311 (2007).
taking into account several important deviations of the refractive 13. Popoff, S. M. et al. Measuring the transmission matrix in optics: an approach to
index profile and fibre curvature. We have not identified any signifi- the study and control of light propagation in disordered media. Phys. Rev. Lett.
cant randomizing processes in 300-mm-long fibre segments and, 104, 100601 (2010).
14. Čižmár, T., Mazilu, M. & Dholakia, K. In situ wavefront correction and its
based on the behaviour observed, it is very unlikely that highly application to micromanipulation. Nature Photon. 4, 388–394 (2010).
ordered light propagation would be obstructed on a length scale 15. Popoff, S., Lerosey, G., Fink, M., Boccara, A. C. & Gigan, S. Image transmission
of even several metres. through an opaque material. Nature Commun. 1, 81 (2010).
From our results it is clear which modes are the most vulnerable 16. Vellekoop, I. M., Lagendijk, A. & Mosk, A. P. Exploiting disorder for perfect
to fibre deformations and, in turn, which are the most immune. focusing. Nature Photon. 4, 320–322 (2010).
17. Ji, N., Milkie, D. E. & Betzig, E. Adaptive optics via pupil segmentation for high-
This is of crucial importance not only for endoscopy, but also for resolution imaging in biological tissues. Nature Methods 7, 141–147 (2009).
many other disciplines, including telecommunications (mode- 18. Papadopoulos, I. N., Farahi, S., Moser, C. & Psaltis, D. High-resolution, lensless
division multiplexing). endoscope based on digital scanning through a multimode optical fiber. Biomed.
We have found that the strength of deformation-induced TM Opt. Express 4, 260 (2013).
19. Farahi, S., Ziegler, D., Papadopoulos, I. N., Psaltis, D. & Moser, C. Dynamic
changes can be tuned by selection of the refractive index and bending compensation while focusing through a multimode fiber. Opt. Express
Poisson’s ratio. This may bring about an exciting new possibility 21, 22504–22514 (2013).
to design fibres featuring suppressed bending effects (minimizing 20. Gambling, W. A., Payne, D. N. & Matsumura, H. Mode conversion coefficients
the value of ξ). However, the ultimate goal of designing fibres in optical fibers. Appl. Opt. 14, 1538–1542 (1975).
with ξ = 0, where light transport processes are entirely invariant to 21. Friesem, A. A., Levy, U. & Silberberg, Y. Parallel transmission of images through
single optical fibers. Proc. IEEE 71, 208–221 (1983).
fibre deformation, would require either a negative refractive index 22. Kreysing, M. et al. Dynamic operation of optical fibres beyond the single-mode
or a negative Poisson’s ratio, both of which are very difficult to regime facilitates the orientation of biological cells. Nature Commun. 5,
achieve experimentally. We speculate that the latter could possibly 5481 (2014).
be approached by providing a highly anisotropic sleeve around 23. von Hoyningen-Huene, J., Ryf, R. & Winzer, P. LCoS-based mode shaper for
the fibre. few-mode fiber. Opt. Express 21, 18097–18110 (2013).
24. Carpenter, J., Eggleton, B. J. & Schröder, J. 110×110 optical mode transfer matrix
The startling agreement of the theoretical predictions with exper- inversion. Opt. Express 22, 96–101 (2014).
imental reality allowed us to perform imaging using theoretically 25. Leach, J., Padgett, M., Barnett, S., Franke-Arnold, S. & Courtial, J. Measuring
predicted TMs in straight and curved fibre segments that were suffi- the orbital angular momentum of a single photon. Phys. Rev. Lett. 88,
ciently long for many practical applications in life sciences, medi- 257901 (2002).
cine and other disciplines. These new possibilities bring a step 26. Bliokh, K. Y., Niv, A., Kleiner, V. & Hasman, E. Geometrodynamics of spinning
light. Nature Photon. 2, 748–753 (2008).
change for this technology by eliminating the most critical draw- 27. Lyytikäinen, K. et al. Dopant diffusion during optical fibre drawing. Opt. Express
backs related to the requirement for empirically measured TMs as 12, 972–977 (2004).
well as a lack of flexibility. We believe that there are currently no 28. Gibson, B. C. et al. Controlled modification and direct characterization of
fundamental problems standing in the way of further fast develop- multimode-fiber refractive-index profiles. Appl. Opt. 42, 627–633 (2003).
ment of this exciting technology, and with sufficiently sophisticated 29. Skinner, B. J. & Appleman, D. E. Melanophlogite, a cubic polymorph of silica.
Am. Mineral. 48, 854–867 (1963).
engineering we might soon see a multitude of exciting applications
such as observations of single or multiple neurons in deep brain Acknowledgements
structures of unrestrained and awake animal models. T.Č. and M.P. acknowledge the Scottish Universities Physics Alliance (SUPA)
(PaLS initiative) and the University of Dundee for funding. T.T. acknowledges support
(grant no. P201/12/G028) by the Czech Science Foundation and SUPA Graduate School
Methods Distinguished Visitor Scheme. The authors thank H.I. Campbell Dalgarno, A. Danner and
Methods and any associated references are available in the online W.A. Gillespie for useful comments and proofreading of the manuscript.
version of the paper.
Author contributions
Received 15 January 2015; accepted 2 June 2015; T.Č. conceived and led the project, designed the instrumentation, performed all
experiments and wrote the manuscript with contributions from all authors. M.P. and T.T.
published online 13 July 2015 developed the theoretical model. All authors contributed equally to analysis of the results.

References Additional information

1. Snyder, A. W. & Love, J. Optical Waveguide Theory (Springer, 1983). Supplementary information is available in the online version of the paper. Reprints and
2. Gloge, D. Weakly guiding fibers. Appl. Opt. 10, 2252–2258 (1971). permissions information is available online at www.nature.com/reprints. Correspondence and
3. Snitzer, E. Cylindrical dielectric waveguide modes. J. Opt. Soc. Am. 51, requests for materials should be addressed to T.Č.
491–498 (1961).
4. Liberman, V. S. & Zel’dovich, B. Y. Spin–orbit polarization effects in isotropic Competing financial interests
multimode fibres. Pure Appl. Opt. 2, 367–382 (1993). The office of Intellectual Property Rights of the University of Dundee has submitted a
5. Di Leonardo, R. & Bianchi, S. Hologram transmission through multi-mode provisional patent application (GB 1509418.8, filing date 1 June 2015) related to some
optical fibers. Opt. Express 19, 247–254 (2011). aspects of the work.

NATURE PHOTONICS | VOL 9 | AUGUST 2015 | www.nature.com/naturephotonics 535

© 2015 Macmillan Publishers Limited. All rights reserved

ARTICLES NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112

∞ 2π
Methods where Ni = [ ∫0 rdr ∫0 dφ|ψ i (r, φ)|2 ]1/2 is the normalization constant of the ith
Calculating the changes of propagation constants due to deviations in the mode. Equation (7) is in fact an eigenvalue problem for 1/β′2. The eigenvectors
refractive index profile. Light transport processes in MMFs are well understood (sequences of coefficients ci) then determine the corresponding modes of the bent
theoretically, and the necessary theoretical considerations are briefly reviewed in fibre. This equation can be further simplified by taking advantage of the fact that the
Supplementary Methods 1. To find corrections to the propagation constants arising propagation constants βi are limited by the condition ncladdingk0 < βi < ncorek0
from slight deviations of the refractive index profile, we used perturbation calculus and thus are all very close to ncorek0. Making the substitutions βi = ncorek0 + Δβi and
with respect to basic scalar theory. These corrections were then added to β′ = ncorek0 + Δβ′, inserting into equation (7), neglecting the term containing the
propagation constants calculated by the weak guidance approximation of the full product AijΔβi and returning back from Δβi, Δβ′ to βi, β′, we finally get the equation
vector model to obtain corrected propagation constants.
The refractive index of the ideal step-index fibre is denoted n(r) and the ncore k0

βi ci − Aij cj = β′ ci (9)
perturbed refractive index is denoted n′(r). Starting from the scalar wave ρ j
equation ΔΨ − (n/c)2 ∂2t Ψ = 0 and separating the time and z-coordinate along the
fibre as Ψ(r, φ, z, t) = ψ(r, φ)exp[i(βz – ωt)]), we arrive at the Helmholtz equation for This equation shows that the propagation constants are eigenvalues of the matrix B
the transversal part of the wave ψ(r, φ) in the step refractive index n(r): with entries Bij = βiδij – (ncorek0/ρ)〈ψi|x|ψj|〉. The advantage of equation (9)
compared to equation (7) is that its eigenvalues are directly the new propagation
[Δ⊥ + k20 n2 − β2 ]ψ(r, φ) = 0 (3) constants. The evolution operator of the state along the fibre segment of length L is
thus simply eiBL. Evolution along a fibre that is bent non-uniformly can be expressed
where Δ⊥ is the transversal part of the Laplacian and k0 is the wavenumber. The
as a product of such operators corresponding to sufficiently short fibre segments in
solutions can be indexed by the angular index l (angular momentum, or topological
which the curvature can be considered constant.
charge) and radial index m, as mentioned in the main text. A similar equation holds
Previously, we have assumed that the refractive index profile of the deformed
for the perturbed mode functions ψ′(r, φ), but with n replaced by n′ and β replaced
fibre remains identical to the original straight fibre. Fibre deformation, however,
by the perturbed propagation constants β′. Next we write the square of the modified
causes local density changes and consequently also refractive index changes, which
refractive index as n′2(r) = n 2(r) + g(r), where g(r) is a small perturbation, express the
influences light propagation. When the fibre is bent, its outer side becomes longer
mode functions ψ′lm as superpositions of the unperturbed mode functions ψlm, and
and the inner side becomes shorter. Such longitudinal changes of the length of
perform the standard perturbation calculation. In this way we arrive at the following
infinitesimal fibre elements cause corresponding lateral changes of their width that
first-order correction to the propagation constant of the mode ψ′lm:
are of opposite sign and smaller in magnitude by a factor of the Poisson ratio σ. In
k20 other words, the diagonal elements of the deformation tensor are related as
Δβlm = 〈ψ |g|ψ lm 〉 ϵxx = ϵyy = −σϵzz . The relative changes of the element volume and density are then
2βlm lm
(4) Trϵ̂ = ϵzz (1 − 2σ) and −Trϵ̂ = −ϵzz (1 − 2σ), respectively, provided that the
∞
k20 ∫0 [n′2 (aR) − n2 (aR)]Flm
2
(R)RdR deformation is small. If we assume that the refractive index has the property
= ∞ 2
2βlm ∫0 Flm (R)RdR that n – 1 is proportional to the density, which seems reasonable for the fused
silica from which the fibre is made, we obtain a modified refractive index
Here βlm is the unperturbed propagation constant of the mode ψ lm , and Flm is the n′ = n – (n – 1)ϵzz (1 – 2σ). Inserting this modified index n′ into equation (5) instead
radial part of this mode expressed as a function of r scaled by the radius of the core a: of n, together with ϵzz = –x/ρ, we find, after neglecting one insignificant term,
ψ lm (r, φ) exp (ilφ)Flm (r/a)  
2x n−1
Δ⊥ + k20 n2 − β′ ψ′ = 0
2
1+ 1 − (1 − 2σ)
For a given modified index n′(r), these corrections can be readily calculated by ρ n
numerical integration. The polarization index σ of the mode does not influence
the correction. We see that the resulting equation differs from equation (5) by a factor
ξ = 1 –(1 – 2σ)(n – 1)/n standing in front of the curvature 1/ρ. This means that the
Calculating the changes to the TM due to fibre bending. Eigenmodes of a bent fibre behaves the same as if the index had not changed, but the bending were weaker
fibre differ from those of a straight fibre. Taking into account the fact that in practice by a factor of ξ. This also means that the effect of index change alone (due to fibre
the radius of curvature is larger than the core radius by several orders of magnitude, deformation) is of a similar character as the effect of the bending itself, but of
one could be tempted to think that the two sets of modes will differ only slightly. opposite sign.
This is not true, however, due to the very small index difference between the core and In general, the transverse deformation leads to a change in the shape of the
the cladding. In fact, bending introduces effective index changes that are comparable cross-section of the fibre core, which could also affect the modes. However, it is
to this index difference, so perturbation theory would yield inaccurate results. not hard to show that for the particular deformation in question, that is,
Fortunately, one can still describe modes of the bent fibre approximately using a ϵxx = ϵyy = −σϵzz = σx/ρ, the cross-section remains circular (still with radius a)
calculation that is reminiscent of perturbation theory, as we show in the following. up to the first order in a/ρ; the deviation from the circular shape is of the
Consider a short element of the fibre centred at the origin of Cartesian coordinates second order in a/ρ and therefore completely negligible.
with the fibre axis oriented along the z axis. Suppose now that the fibre is bent in the
x–z plane with radius of curvature ρ ≫ a with the centre of curvature at the point The experimental system. All the present studies were accomplished using a
(ρ, 0, 0). The local longitudinal wavenumber (along the z axis) will no longer be step-index MMF with radius a = 25.0 ± 0.5 μm and numerical aperture (NA) of
constant across the fibre cross-section, but will depend on x as kz(x) = β′(1 + x/ρ), 0.22 ± 0.02 (these values were provided by the fibre supplier Thorlabs). Our system
where β′ is the value of the longitudinal wavenumber on the axis—the propagation operating at the wavelength of 1,064 nm is based on a standard digital holographic
constant. This yields the following equation for the mode in the x–y plane: geometry in a Fourier regime where an SLM is imaged onto the back aperture of a
microscope objective (MO1, ×20, NA = 0.25) by a 4–f system formed by lenses L3
   and L4 (Supplementary Fig. 14 shows a simplified scheme with labelled
x 2 ′
Δ⊥ + k20 n2 − β′2 1 + ψ =0 (5) components). The signal from the SLM is sent to three different zones within the
ρ L3–L4 telescope simultaneously. One part is separated by mirror M1 and used as a
 reference signal during acquisition of the TM. The remaining two are sent into a
Next we express ψ′ as a superposition of the modes of the straight fibre, ψ ′ = ci ψ i system of half wave plates (HWP1–3) and polarizing beam displacers (PBD1,2) to
i
(for simplicity we have replaced the pair of indices l, m by a single index i) and create and overlap two fields with orthogonal linear polarization states. Even though
substitute into equation (5). Using also equation (3), we get this is sufficient for the generation of optical fields with an arbitrary spatial

  distribution of amplitude, phase and polarization, a quarter wave plate (QWP1) is
1 + 2x
ci β2i − β′2 ψi = 0 (6) inserted into the geometry to give higher efficiency and reduce noise, as most of the
i
ρ experimental work requires circular polarization. Part of the signal sent to MO1 is
separated by a non-polarizing beamsplitter (NPBS1) and imaged onto CCD1 to
Multiplying this equation by ψ *j , integrating over the x–y plane, using the
eliminate optical aberrations and measure SLM irradiance (as detailed in the next
orthogonality of the functions ψi, rearranging the terms slightly and interchanging
section) and to monitor the field being sent into the MMF (the field recorded at
the indices i and j, we obtain
CCD1 is a scaled copy of the field at the input facet of the MMF). The optical signal
1 2
1 leaving the MMF is collected by microscope objective (MO2, with the same
ci + A c = ′2 ci (7) parameters as MO1) and focused by tube lens L6. Circularly polarized modes are
β2i ρ j ij j β converted to linearly polarized modes by HWP2, and individual polarization
components are separated on polarizing beamsplitter PBS and imaged on CCD2 and
Here we have defined the matrix element of the x coordinate CCD3. The reference signal is delivered to both CCD2 and CCD3 by a single-mode
∞ 2π
∫0 rdr ∫0 dφψ *i (r, φ)r cos φψ j (r, φ) fibre (SMF) with collimator lenses (L7 and L8) at each end. The reference optical
Aij ≡ 〈ψ i |x|ψ j 〉 = (8) pathway is merged with the imaging pathway before the polarization components
Ni Nj are separated by NPBS2. The geometry enables us to generate any optical field

NATURE PHOTONICS | www.nature.com/naturephotonics

© 2015 Macmillan Publishers Limited. All rights reserved

NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112 ARTICLES
allowed to propagate in the optical fibre and to observe how it is transformed by its M̂ = Aout M̂ Ain . The resulting pre-optimization giving the best image symmetry and
transport through the fibre. Crucially, the geometry allows us to synthesize and sharpness is presented in Supplementary Fig. 16e–h.
observe all aspects of the input and output optical fields, respectively, that is, their Following this procedure the output modes are resampled using the scaling
intensity, phase and polarization. CCD4 was only used in tests of proximal end property of the Fourier transform in order to match the sampling of the input
imaging (as explained in Supplementary Results 2). modes. Finally, redundant data (manifesting itself as dark boundaries in
Supplementary Fig. 16) are removed, leading to a matrix with 729 input and 729
Wavefront correction. All the experimental studies in this Article require beam output modes, as presented in Supplementary Fig. 5. This operation is also achieved
control with extremely high fidelity, virtually free of optical aberrations. To eliminate by matrix multiplication: M = Bout M̂ Bin .
aberrations in our very complex optical system we used the method developed in
Optimization procedure. The optimization procedure gradually evolved over the
ref. 14. This algorithm efficiently manipulates the phases of small segments of the
course of the project to take into account a series of newly recognized factors. In the
SLM to gain the highest signal in the selected pixel of CCD1, signifying optimal
following we describe only the final one, optimized to study CP PIMs propagating
focusing. The quality of the resulting wavefront correction can be influenced by
through up to 300-mm-long fibre segments.
CCD noise and SLM flicker noise, both of which can be statistically averaged out by
With the well-established theoretical description of light transport in MMFs1
longer data collection. However, with increased duration of the procedure there is
(reviewed in Supplementary Methods 1) and its modifications introduced above, we
the risk of a strong influence of systematic drift. To overcome this tradeoff we
can compute the conversion matrix T that allows conversion of the measured TM
measured optical aberrations with a large number of different pixels (10,000)
= T M T† . Due to remaining misalignment
into the representation of PIMs: M
simultaneously, and combined the results. As pixels are organized across an
imperfections and insufficient precision in the estimation of the fibre parameters,
orthogonal grid, each of these measurements carries a specific linear phase modulation
this does not lead to satisfactory results and further optimization is necessary. While
(equivalent of a prism), which can be predicted and subtracted from the data before
compensation for misalignment can be introduced by matrix multiplications of M,
averaging. The enhancement is demonstrated in Supplementary Fig. 15.
correction of fibre parameters requires calculation of a new conversion matrix T at
every iteration, which is time-consuming, and M
does not have fixed size, which
Measurement of the TM. In our previous studies6,7 we defined the input modes for requires more elaborate optimization criteria to converge. These two parts of the
the TM as beamlets originating from a series of SLM segments. One considerable optimization are thus decoupled. In each iteration (indexed by j ) of the
problem of this approach is a relatively weak signal that is often significantly affected misalignment optimization procedure we implement misalignment corrections to
by various sources of scattered light. For the purposes of this study we used a the experimentally measured TM before its conversion into PIMs:
representation of optimally focused spots at the input facets of the MMF (as M
j = T C out M Cin T† . The quality metric that the algorithm optimizes across the
explained in the main text). This is beneficial not only for gaining much stronger 12-dimensional space of free parameters is defined on the resulting matrix M
j as the
signals (all the power reflected from the SLM is used for the analysis of every mode), fraction of the total optical power (absolute squares) carried by the TM’s diagonal
but also for simpler comparison with the theoretical prediction. During the TM components, multiplied by PA (the phase agreement introduced in the main text).
measurement in this regime the SLM splits the optical power between two pathways, During optimization of the fibre parameters (that directly influence the conversion
one of which is sent to the SMF, which acts as a reference signal, and the other results matrix T) the quality metric defined on the resulting M
j is a combination of four
in a focal point input mode at a selected location of the input facet of the MMF. Due factors. This time we could not use the total power ratio carried by diagonal
to the series of polarization optical elements shown in Supplementary Fig. 14, each components as this strongly favoured small-sized matrices corresponding to low NA
focal point can be generated by two different SLM modulations of the signal, leading and smaller core diameters, which led to divergence of the optimization algorithm.
to generation of a pair of mutually orthogonal polarization states. Output focal point Nor did the total power of the diagonal components prove itself to be a suitable
modes are monitored on selected (grouped) pixels of CCD2 and CCD3, where the criterion as it favoured large-sized matrices corresponding to high NA and larger core
output fibre facet is imaged. Their phases are established from interference with the diameters. The optimal power conversion was therefore achieved by reverse
reference signal delivered by a SMF. Using the SLM to uniformly alter the phase conversion of M
j back into the space of diffraction-limited points, Mjr = T† M
j T,
difference Δθ between the tested input mode and the reference field leads to a and comparison with the original experimentally measured one using a least-squares
 j 2
harmonic signal for each CCD pixel: approach, (|Mr | − |M|) . This value falls steeply while increasing both the NA and
 2 the diameter of the fibre core, but after passing the expected value it remains
  stationary. This was complemented by the introduction of a second criterion ensuring
ICCD = Fm · ei·Δθ + Fref 
(10) the maximal uniformity of the diagonal components. The standard deviation of the
 2  2   
j remains stationary until reaching the expected values and
= Fm  + Fref  + 2Fm Fref  · cos (fm + Δθ) diagonal components of M
then rises steeply. Similarly, as in the previous case, we use PA as the third
–1
optimization factor (PA , as we search for the minimum in this case). Finally, we have
where Fm and Fref are the fields of the tested mode and the reference field, found that much faster convergence is ensured by minimizing the value of a fourth
respectively. Typically, we use more that one period of 2π in such measurement, for criterion that represents the total power of elements directly neighbouring the
example, Δθ = (2π/4) [0,1,…,7]. main diagonal. The combination of these four criteria features a sharp minimum in
On analysing the detected signal we can immediately isolate the distribution of the multidimensional space of all fibre parameters: NA, core diameter, diffusion
phase of the tested input mode ϕm across the output modes. From the measured lengths and further refractive index corrections as described in the main text.
harmonic signal we can also establish the magnitude of the a.c. and d.c. components Both parts of the optimization procedure are cycled several times until a
that, based on equation (10), allow for recovery of the output mode amplitudes. In stationary point is reached. The result of the algorithm is the set of optimal fibre
this way, the measurement is not affected by the Gaussian envelope and small spatial parameters, optimized M
and T and fine misalignment matrices Cin and Cout. The
intensity non-uniformity of the reference signal (resulting from interference effects Matlab code for this part, together with detailed instructions and samples of
at various optical components of the reference pathway), which can significantly pre-optimized experimental data samples (M), is available for download from
affect measurement of the TM. Similarly to ref. 6, we use a feedback loop to eliminate http://complexphotonics.dundee.ac.uk/.
drift between the MMF optical pathway and the reference optical pathway. The The whole optimization procedure is executed separately for polarization
initial measurement of the TM contains 1,089 input modes and 5,625 output modes. components of M̃:
The input modes are measured sequentially, but the output modes are measured  −1 
+1
simultaneously and therefore they are favourably oversampled to reduce noise. M̃−1 M̃−1
−1 +1
M̃+1 M̃+1
Processing raw data. The data from direct measurement of the TM (M̃) are heavily
−1
affected by various sources of misalignment. Moreover, input and output modes are We only seek optimal fibre parameters in the first step of analysing M̃−1 , together
obtained with different sampling. Before such a TM can be used in a comparison with misalignment factors. A record of this procedure is presented in Supplementary
+1
with the theoretical prediction, the data must be processed to address these issues. Movie 1. This is followed by analysis of M̃+1 , where we only seek misalignment
The amount of misalignment can be estimated roughly from the total transmission parameters (light follows a different optical pathway, as shown in Supplementary
profiles of the TM, which, in the ideal case (no misalignment) would be perfectly Fig. 14), and for conversion into PIMs we use the fibre parameters obtained in the
symmetric and focused on both sides of the optical system. These profiles are +1 −1
previous step. For conversion of components M̃−1 and M̃+1 we know all the
obtained by averaging the absolute squares of either rows or columns of the TM, and −1 +1
misalignment parameters of the systems from analysis of M̃−1 and M̃+1 , so no further
they signify position misalignment of the fibre end with respect to the optical optimization is needed.
pathway. Using a Fourier transform we can convert the measured TM into the
representation of plane waves, which analogously provide transmission profiles, this Computing complex holographic modulations. Our LabView-based controlling
time signifying tilts and defocus. The transmission profiles are shown in system was designed to synthesize complex holographic modulations to be applied
Supplementary Fig. 16a–d. across the SLM. It uses the experimentally measured TM (M̃) to combine a series
All misalignment operators can be represented by phase-only linear or quadratic of input modes (in the representation of focal points) to achieve any output optical
modulations of the modes or their Fourier spectra and combined into one matrix for field leaving the optical fibre, with arbitrary distribution of amplitude, phase
the input (Ain) and one matrix for the output (Aout) of the fibre. The correction is and polarization. This is achieved using a modified version30 of the
then implemented by matrix multiplication of the TM from either side: Gerchberg–Saxton algorithm31.

NATURE PHOTONICS | www.nature.com/naturephotonics

© 2015 Macmillan Publishers Limited. All rights reserved

ARTICLES NATURE PHOTONICS DOI: 10.1038/NPHOTON.2015.112

Synthesis of holograms for imaging. Because imaging is achieved by scanning a controlling interface. Instead of the fully reversed conversion, we only convert
focal point behind the output fibre facet, performing imaging with an experimentally PIM
the input modes as M̃th = M
th T C†in B†in A†in . Matrix M̃PIM
th is a hybrid
measured TM is accomplished by the sequential indexing of rows of M̃, leading to transformation between the input, represented using focal points, and the output,
synthesis of a single focal point output mode at a time. For imaging with represented by PIMs. The complete set of PIMs recorded is presented in
theoretically predicted TMs M
th , we can utilize pairs of matrices A, B and C obtained Supplementary Movies 2–5.
during optimization of the data: M̃th = A†out B†out C†out T† M
th T C†in B†in A†in .
Here M
th represents any of the modifications introduced in the main text, including References
those for deformed fibres and axially shifted imaging planes. 30. Čižmár, T. & Dholakia, K. Tunable Bessel light modes: engineering the axial
propagation. Opt. Express 17, 15558–15570 (2009).
Generation of PIMs. Generation of individual PIMs uses the same procedure as 31. Gerchberg, R. W. & Saxton, W. O. A practical algorithm for the determination of
described in the previous case. The only difference is in the TM that we load into the the phase from image and diffraction plane pictures. Optik 35, 237–246 (1972).

NATURE PHOTONICS | www.nature.com/naturephotonics

© 2015 Macmillan Publishers Limited. All rights reserved