Optical computing ⋆

Damien Woods a
a Department of Computer Science and Artificial Intelligence
University of Seville, Spain

Thomas J. Naughton b,c

b Department of Computer Science

National University of Ireland, Maynooth, County Kildare, Ireland

c University of Oulu, RFMedia Laboratory, Oulu Southern Institute, Vierimaantie 5, 84100 Ylivieska,

Finland

Abstract

We consider optical computers that encode data using images and compute by transforming such
images. We give an overview of a number of such optical computing architectures, including
descriptions of the type of hardware commonly used in optical computing, as well as some
of the computational efficiencies of optical devices. We go on to discuss optical computing
from the point of view of computational complexity theory, with the aim of putting some old,
and some very recent, results in context. Finally, we focus on a particular optical model of
computation called the continuous space machine. We describe some results for this model
including characterisations in terms of well-known complexity classes.

Key words:
PACS:

1. Introduction

In this survey we consider optical computers that encode data using images and com-
pute by transforming such images. We try to bring together, and thus give context to, a
large range of architectures and algorithms that come under the term optical computing.
⋆ DW acknowledges support from Junta de Andalucı́a grant TIC-581. TN acknowledges support from
the European Commission Framework Programme 6 through a Marie Curie Intra-European Fellowship.
Email addresses: [email protected] (Damien Woods), [email protected] (Thomas J. Naughton).
URLs: http://www.cs.ucc.ie/∼dw5 (Damien Woods), http://www.cs.nuim.ie/∼tnaughton/
(Thomas J. Naughton).

Preprint submitted to Elsevier 8 July 2008

 In Section 2 we begin by stating what we mean by the term optical computing, and we
discuss some common optical computing architectures. Unlike a number of other areas of
nature inspired computing, optical computers have existed (mostly in laboratories) for
many years, and so in Section 3 we describe hardware components that are found in many
optical computing systems. In Section 4 we describe some of the physical principles that
underlie the parallel abilities and efficiencies found in optical computing. These include
high fan-in, high interconnection densities, and low energy consumption.
In order to understand the abilities of optics as a computing medium we would argue
that computational complexity theory is an indispensable tool. So in Section 5 we collect
a number of existing algorithmic results for optical computers. In particular we focus
on algorithms and models for matrix-vector multipliers, and some other architectures.
We also offer suggestions for future work directions for optical algorithm designers. In
Section 6, we take a detailed look at a particular model of optical computing (the contin-
uous space machine, or CSM) that encompasses most of the functionality that coherent
optical information processing has to offer. We begin by defining the CSM and a total of
seven complexity measures that are inspired by real-world (optical) resources. We go on
to discuss how the CSM’s operations could be carried out physically. Section 7 contains
some example datastructures and algorithms for the CSM. In Section 8 we motivate and
introduce an important restriction of the model called the C2 -CSM, and in Section 9
we briefly describe a number of C2 -CSM computational complexity results, and their
implications.

2. Brief overview of optical computing architectures

Traditionally, in optical information processing a distinction was made between sig-

nal/image processing through optics and numerical processing through optics, with only
the latter (and often only the digital version of the latter) being called optical comput-
ing [57,48,29,101]. However, it was always difficult to clearly delineate between the two,
since it was largely a question of the interpretation the programmer attached to the
output optical signals. The most important argument for referring to the latter only as
optical computing had to do with the fact that the perceived limits (or at least, ambi-
tions) of the former was simply for special-purpose signal/image processing devices while
the ambitions for the latter was general-purpose computation. Given recent results on
the computational power of optical image processing architectures [63,98,93], it is not
the case that such architectures are limited to special-purpose tasks.
Furthermore, as the field become increasingly multidisciplinary, and in particular as
computer scientists play a more prominent role, it is necessary to bring the definition
of optical computing in line with the broad definition of computing. In particular, this
facilitates analysis from the theoretical computer science point of view. The distinction
between analog optical computing and digital optical computing is similarly blurred given
the prevalence of digital multiplication schemes effected through analog convolution [48].
Our broad interpretation of the term optical computing has been espoused before [19].

2
2.1. Optical pattern recognition

Arguably, optical computing began with the design and implementation of optical
systems to arbitrarily modify the complex valued spatial frequencies of an image. This
concept, spatial filtering [67,86,89], is at the root of optics’ ability to perform efficient
convolution and correlation operations. In a basic pattern recognition application, spatial
filtering is called matched filtering, where a filter is chosen that matches (i.e. conjugates)
the spectrum of the sought object. Employing this operation for advanced pattern recog-
nition [12,18,43,47], effort focused on achieving systems invariant to scaling, rotation,
out-of-plane rotation, deformation, and signal dependent noise, while retaining the exist-
ing invariance to translating, adding noise to, and obscuring parts of the input. Effort also
went into injecting nonlinearities into these inherently linear systems to achieve wider
functionality [46]. Improvements were made to the fundamental limitations of the basic
matched filter design, most notably the joint transform correlator architecture [92].
Optical correlators that use incoherent sources of illumination (both spatially and tem-
porally) rather than lasers are also possible [11,71]. The simplest incoherent correlator
would have the same basic spatial filtering architecture as that used for matched fil-
tering. While coherent systems in principle are more capable than incoherent systems
(principally because the former naturally represents complex functions while the latter
naturally represents real functions), incoherent systems require less precise positioning
when it comes to system construction and are less susceptible to noise.
Trade-offs between space and time were proposed and demonstrated. These included
time integrating correlators [90] (architecturally simpler linear time variants of the con-
stant time space integrating matched filter mentioned already) and systolic architec-
tures [21]. In addition to pattern recognition, a common application for these classes of
architectures was numerical calculation.

2.2. Analog optical numerical computation

An important strand of image-based optical computation involved numerical calcula-

tions: analog computation as well as multi-level discrete computation. Matrix-vector and
matrix-matrix multiplication systems were proposed and demonstrated [42,57,90,48,4,29,51].
The capability to expand a beam of light and to focus many beams of light to a common
point directly corresponded to high fan-out and fan-in capabilities, respectively. The limi-
tations of encoding a number simply as an intensity value (finite dynamic range and finite
intensity resolution in the modulators and detectors) could be overcome by representing
the numbers in some base. Significant effort went into dealing with carry operations so
that in additions, subtractions, and multiplications each digit could be processed in par-
allel. Algorithms based on convolution to multiply numbers in this representation were
demonstrated [48], with a single post-processing step to combine the sub-calculations
and deal with the carry operations.
An application that benefited greatly from the tightly-coupled parallelism afforded by
optics was the solving of sets of simultaneous equations and matrix inversion [17,1]. An
application that, further, was tolerant to the inherent inaccuracies and noise of analog
optics was optical neural networks [28,20] including online neural learning in the presence
of noise [62].

3
2.3. Digital optical computing

The next major advances came in the form of optical equivalents of digital comput-
ers [44]. The flexibility of digital systems over analog systems in general was a major
factor behind the interest in this form of optical computation [78]. Specific drawbacks of
the analog computing paradigm in optics that this new paradigm addressed included no
perceived ability to perform general purpose computation, accumulation of noise from
one computation step to another, and systematic errors introduced by imperfect analog
components. The aim was to design digital optical computers that followed the same
principles as conventional electronic processors but which could perform many binary
operations in parallel. These systems were designed from logic gates using nonlinear op-
tical elements: semitransparent materials whose transmitted intensity has a nonlinear
dependence on the input intensity.
Digital optical computing was also proposed as an application of architectures designed
originally for image-based processing, for example logic effected through symbolic substi-
tution [10]. At the confluence of computing and communication, optical techniques were
proposed for the routing of signals in long-haul networks [29,101].

3. Optical computing hardware

The three most basic hardware components of an optical information processing system
are a source, a modulator, and a detector. A source generates the light, a modulator
multiplies the light by a (usually, spatially varying) function, and a detector senses the
resulting light. These and others are introduced in this section.

3.1. Sources

Lasers are a common source of illumination because at some levels they are math-
ematically simpler to understand, but incoherent sources such as light-emitting diodes
are also used frequently for increased tolerance to noise and when nonnegative functions
are sufficient for the computation. Usually, the source is monochromatic to avoid the
problem of colour dispersion as the light passes through refracting optical components,
unless this dispersion is itself the basis for the computation.

3.2. Spatial light modulators (SLMs)

It is possible to encode a spatial function (a 2D image) in an optical wavefront. A page

of text when illuminated with sunlight, for example, does this job perfectly. This would
be called an amplitude-modulating reflective SLM. Modulation is a multiplicative effect,
so an image encoded in the incoming wavefront will be pointwise multiplied by the image
on the SLM. Modulators can also act on phase and polarisation, and can be transmissive
rather than reflective. They include photographic film, and electro-optic, magneto-optic,
and acousto-optic devices [57,4,48,29,35]. One class of note are the optically-addressed
SLMs, in which, typically, a 2D light pattern falling on a photosensitive layer on one side
of the SLM spatially varies (with an identical pattern) the reflective properties of the

4
other side of the SLM. A beam splitter then allows one to read out this spatially-varying
reflectance pattern. The liquid-crystal light valve [45] is one instance of this class. Other
classes of SLMs such as liquid-crystal display panels and acousto-optic modulators allow
one to dynamically alter the pattern using electronics. It is possible for a single device
(such as an electronically programmed array of individual sources) to act as both source
and modulator.

3.3. Detectors and Nature’s square law

Optical signals can be regarded as having both an amplitude and phase. However,
detectors will measure only the square of the amplitude of the signal (referred to as its
intensity). This phenomenon is known as Nature’s detector square law and applies to
detectors from photographic film to digital cameras to the human eye. Detectors that
obey this law are referred to as square-law detectors. This law is evident in many physical
theories of light. In quantum theory, the measurement of a complex probability function
is formalised as a projection onto the set of real numbers through a squaring operation.
Square-law detectors need to be augmented with a interferometric or holographic ar-
rangement to measure both amplitude and phase rather than intensity [13], or need to
be used for multiple captures in different domains to heuristically infer the phase.
Since it squares the absolute value of a complex function, this square law can be
used for some useful computation (for example, in the joint transform correlator [92]).
Detectors most commonly used include high range point (single pixel) detectors such as
photodiodes, highly sensitive photon detectors such as photomultiplier tubes, and 1D
and 2D array detectors such as CCD- or CMOS-digital cameras. Intensity values outside
the range of a detector (outside the lowest and highest intensities that the detector
can record) are thresholded accordingly. The integration time of some detectors can be
adjusted to sum all of the light intensity falling on them over a period of time. Other
detectors can have quite large light sensitive areas and can sum all of the light intensity
falling in a region of space.

3.4. Other optical computing hardware

Lenses can be used to effect high fan-in and fan-out interconnections, to rescale images
linearly in either one or two dimensions, and for taking Fourier transforms. In fact, a co-
herent optical wavefront naturally evolves into its Fresnel transform, and subsequently
into its Fourier transform at infinity, and the lens simply images those frequency compo-
nents at a finite fixed distance.
A mirror changes the direction of the wavefront and simultaneously reflects it along
some axis. A phase conjugate mirror [25] returns an image along the same path at which
it approached the mirror.
Prisms can be used to for in-plane flipping (mirror image), in-plane rotations, and out-
of-plane tilting. A prism or diffraction grating can be used to separate by wavelength the
components of a multi-wavelength optical channel. For optical fiber communications ap-
plications, more practical (robust, economical, and scalable) alternatives exist to achieve
the same effect [101].

5
 Polarisation is an important property of wavefronts, in particular in coherent optical
computing, and is the basis for how liquid crystal displays work. At each point, an
optical wavefront has an independent polarisation value dependent on the angle, in the
range [0, 2π), of its electrical field. This can be independent of its successor (in the case
of randomly polarised wavefronts), or dependent (as in the case of linear polarisation),
or dependent and time varying (as in the case of circular or elliptical polarisation).
Mathematically, a polarisation state, and the transition from one polarisation state to
another, can be described using the Mueller calculus or the Jones calculus.
Photons have properties such as phase, polarisation, and quantum state that can be
used for computation. For example, quantum computers using linear optical elements
(such as mirrors, polarisers, beam splitters, and phase shifters) have been proposed and
demonstrated [50].

4. Efficiencies in optical computing

Optical computing is an inherently multidisciplinary subject whose study routinely

involves a spectrum of expertise that threads optical physics, materials science, optical
engineering, electrical engineering, computer architecture, computer programming, and
computer theory. Applying ideas from theoretical computer science, such as analysis of
algorithms and computational complexity, enables us to place optical computing in a
framework where we can try to answer a number of important questions. For example,
which problems are optical computers suitable for solving? Also, how does the resource
usage on optical computers compare with more standard (e.g. digital electronic) architec-
tures? The physical principles behind some efficiencies in optical computing are outlined
here.

4.1. Fan-in efficiency

Kirchoff’s Law is well understood in analog electronics as a natural and constant-time

means of summing the current at the intersection of an arbitrary number of wires [58]. In
optics, the same thing is possible by directing several light beams towards a point detector
with a linear response to incident light. Such an optical arrangement sums n nonnegative
integers in O(1) addition steps. On a model of a sequential digital electronic computer
this would require n − 1 addition operations and even many typical (bounded fan-in)
parallel models, with n or more processors, take O(log n) timesteps. Tasks that rely on
scalar summation operations (such as matrix multiplication) would benefit greatly from
an optical implementation of the scalar sum operation. Similarly, O(1) multiplication
and O(1) convolution operations can be realised optically. Very recently, an optics-based
digital signal processing platform has been marketed that claims digital processing speeds
of tera (1012 ) operations per second [53].

4.2. Efficiency in interconnection complexity

As optical pathways can cross in free space without measurable effect on the infor-
mation in either channel, high interconnection densities are possible with optics [20].

6
Architectures with highly parallel many-to-many interconnections between parallel sur-
faces have already been proposed for common tasks such as sorting [7]. Currently, intra-
chip, inter-chip, and inter-board connections are being investigated for manufacturing
feasibility [59].

4.3. Energy efficiency

Electrical wires suffer from induced noise and heat, which increases dramatically when-
ever wires are made thinner or placed closer together, or whenever the data throughput
is increased [59]. As a direct consequence of their resistance-free pathways and noise-
reduced environments, optical systems have the potential to generate less waste heat and
so consume less energy per computation step than electronic systems [14]. This has been
demonstrated experimentally with general-purpose digital optical processors [38].

5. Optical models of computation and computational complexity

There has been a lot of effort put into designing optical computers to emulate conven-
tional microprocessors (digital optical computing), and to image processing over contin-
uous wavefronts (analog optical computing and pattern recognition). Numerous physical
implementations of the latter class exist, and example applications include fast pattern
recognition and matrix-vector algebra [35,90]. There have been much resources devoted
to designs, implementations and algorithms for such optical information processing ar-
chitectures (for example see [4,15,29,35,52,55,57,62,77,90,101,27] and their references).
However the computational complexity theory of optical computers (that is, finding
lower and upper bounds on computational power in terms of known complexity classes)
has received relatively little attention when compared with other nature-insired com-
puting paradigms. Some authors have even complained about the lack of suitable mod-
els [29,55]. Many other areas of natural computing (e.g. [41,2,54,100,82,37,70,60,61]) have
not suffered from this problem. Despite this, we review a number of algorithmically orien-
tated results related to optical computing. We then go on to suggest classes of problems
where optical computers might be usefully applied.
Reif and Tyagi [77] study two optically inspired models. One model is a 3D VLSI model
augmented with a 2D discrete Fourier transform (DFT) primitive and parallel optical
interconnections. The other is a DFT circuit with operations (multiplication, addition,
comparison of two inputs, DFT) that compute over an ordered ring. Time complexity
is defined for both models as number of (parallel) steps. For the first model, volume
complexity is defined as the volume of the smallest convex box enclosing an instance of
the model. For the DFT circuit, size is defined as the number of edges plus gates. Constant
time, polynomial size/volume, algorithms for a number of problems are reported including
matrix multiplication, sorting and string matching [77]. These interesting results are
built upon the ability of their models to compute the 2D DFT in one step. The authors
suggest that the algorithm designer and optical computing architecture communities
should identify other primitive optical operations, besides the DFT, that might result
in efficient parallel algorithms. Barakat and Reif [6], and Tyagi and Reif [76] have also
shown lower bounds on the optical VLSI model.

7
 Reif, Tygar and Yoshida [75] examined the computational complexity of ray tracing
problems. In such problems we are concerned about the geometry of an optical system
where diffraction is ignored and we wish to predict the position of light rays after passing
through some system of mirrors and lenses. They gave undecidability and PSPACE hard-
ness results, which gives an indication of the power of these systems as computational
models.
Feitelson [29] gives a call to theoretical computer scientists to apply their knowledge
and techniques to optical computing. He then goes on to generalise the concurrent read,
concurrent write parallel random access machine, by augmenting it with two optically
inspired operations. The first is the ability to write the same piece of data to many global
memory locations at once. Secondly, if many values are concurrently written to a single
memory location then a summation of those values is computed in a single timestep.
Essentially Feitelson is using ‘unbounded fan-in with summation’ and ‘unbounded fan-
out’. His architecture mixes a well known discrete model with some optical capabilities.
A symbolic substitution model of computation has been proposed by Huang and Bren-
ner, and a proof sketched of its universality [10]. This model of digital computation oper-
ates over discrete binary images and derives its efficiency by performing logical operations
on each pixel in the image in parallel. It has the functionality to copy, invert, and shift
laterally individual images, and OR and AND pairs of images. Suggested techniques for
its optical implementation are outlined.
In computer science there are two famous classes of problems called P and NP [68]. P
contains those problems that are solvable in polynomial time on a standard sequential
computer, while NP is the class of problems that are solvable in polynomial time on a
nondeterministic computer. NP contains P, and it is widely conjectured that they are not
equal. A direct consequence of this conjecture is that there are (NP-hard) problems for
which we strongly believe there is no polynomial time algorithm on a standard sequential
computer.
It is known that it is possible to solve any NP (and even any PSPACE) problem in
polynomial time on optical computers, albeit with exponential use of some other, space-
like, resources [97,93,95]. These results were shown on the CSM, a general model of a
wide range of optical computers. The lower bound results were shown by generating
appropriate Boolean masks, of exponential size, and manipulating the masks via parallel
multiplications and additions to simulate space bounded Turing machines in a time-
efficient way. The model was designed on the one hand to be close to the realities of
optical computing, and on the other hand to be relatively straightforward to analyse from
the point of view of computational complexity theory (e.g. see Section 6). In Section 9.1
we discuss the computational abilities of this computational model.
Since these general results, there have been a number of specific examples of optical
systems (with exponential resource usage) for NP-hard problems.
Shaked et al. [79–81] design an optical system for solving the NP-hard travelling sales-
man problem in polynomial time. Basically they use an optical matrix-vector multiplier
to generate the (exponentially large) matrix of all possible tours, then they multiply this
tour matrix by the vector of intercity weights, and finally the lowest value in the resulting
vector corresponds to the shortest tour. Interestingly, they give both optical experiments
and simulations. They note that solving travelling salesman problems (or Hamiltonian
path problems) with more than 15 nodes is problematic. However they argue that for less
nodes (e.g. 5) their system works in real-time, which is faster than digital-electronic ar-

8
chitectures. Problems with such bounds on input size (i.e. constant) lie in the class NC1 ,
and moreover in AC0 . As argued below, perhaps this suggests that the optical comput-
ing community should be focusing on problems where optics excels over digital-electronic
architectures, such as problems in P or NC, rather than NP-hard problems.
Dolev and Fitoussi [26] give optical algorithms that make use of (exponentially large)
masks to solve a number of NP-hard problems. Oltean [66], and Haist and Osten [39],
give architectures for Hamiltonian path, and travelling salesman problem, respectively,
via light travelling through optical cables. As is to be expected, both suffer from expo-
nential resource use. The paper by MacKenzie and Ramachandran [56] is an example of
algorithmic work, and lower bounds, on dynamically reconfigurable optical networks.

5.1. A possible way forward

Nature-inspired systems that apparently solve NP-hard problems in polynomial time,

while using an exponential amount of some other resource(s), have been around for many
years. So the existence of massively parallel optical systems for NP-hard problems should
not really suprise the reader.
One could argue that it is interesting to know the computational abilities, limitations,
and resource trade-offs of such optical architectures, as well as to find particular (tractable
or intractable) problems which are particularly suited to optical algorithms. However,
“algorithms” that use exponential space-like resources (such as number of pixels, number
of images, number of amplitude levels, etc.) are not realistic to implement for large input
instances. What happens to highly parallel optical architectures if add the requirement
that the amount of space-like resources are bounded in some reasonable way? We could,
for example, stipulate that the optical machine use no more than a polynomially bounded
amount of space-like resources. If the machine runs in polynomial time, then it is not
difficult to see that it characterises P [99] (by characterise we mean that the model solves
exactly those problems in P), for a wide range of reasonable parallel and sequential
optical models. Many argue that the reason for using parallel architectures is to speed-
up computations. Asking for an exponential speed-up motivates the complexity class NC.
The class NC can be thought of as those problems in P that can be solved exponentially
faster on parallel computers than on sequential computers. NC is contained in P and it is
an major open question whether this containment is strict: it is widely conjectured that
this is indeed the case [36].
How does this relate to optics? It turns out that a wide range of optical computers that
run for at most polylogarithmic time, and use at most polynomial space-like resources,
solve exactly NC [97,93,95] (this can be shown to be a corollary of the PSPACE char-
acterisation cited earlier in Section 5). In effect this means that we have an algorithmic
way (in other words, a compiler) to convert existing NC algorithms into optical algo-
rithms that use similar amounts of resources. There is scope for further work here, on
the CSM in particular, in order to find exact characterisations, or as close as possible for
NCk for given k. On a technical note, NC can be defined as ∪∞ k k
k=0 NC , where NC is the
k
class of problems solvable on a PRAM that runs for O(log n) time and uses polynomial
processors/space, in input length n. Equivalently NCk can be defined as those problems
solvable by circuits of O(log n)k depth (parallel time), and polynomial size. From the
practical side of things, perhaps we can use these kinds of results to find problems within

9
NC, where optical architectures can be shown to excel. Obvious examples for which this
is already known are matrix-vector multiplication (which lies in NC2 ), or Boolean matrix
multiplication (which is in NC1 ). Another example is the NC1 unordered search prob-
lem [99,98]. Another closely related idea is to exploit the potential unbounded fan-in
of optics to compute problems in the AC, and TC, (parallel) circuit classes. These are
defined similarly to NC circuits except we allow unbounded fan-in gates, and threshold
gates, respectively. The results in the above mentioned paper of Reif and Tyagi [77], and
Caulfield’s observation on the benefits of unbounded fan-in [16], can be interpreted as
exploiting this important and efficient aspect of optics.

6. Continuous space machine (CSM)

For the remainder of this paper we focus on an optical model of computation called the
CSM. The model was originally proposed by Naughton [63,64]. The CSM is inspired by
analog Fourier optical computing architectures, specifically pattern recognition and ma-
trix algebra processors [35,62]. For example, these architectures have the ability to do unit
time Fourier transformation using coherent (laser) light and lenses. The CSM computes
in discrete timesteps over a number of two-dimensional images of fixed size and arbitrary
spatial resolution. The data and program are stored as images. The (constant time) oper-
ations on images include Fourier transformation, multiplication, addition, thresholding,
copying and scaling. The model is designed to capture much of the important features
of optical computers, while at the same time be amenable to analysis from a computer
theory point of view. Towards these goals we give an overview of how the model relates
to optics as well as giving a number of computational complexity results for the model.
Section 6.1 begins by defining the model. We analyse the model in terms of seven
complexity measures inspired by real-world resources, these are described in Section 6.2.
In Section 6.3 we discuss possible optical implementations for the model. We then go
on to give example algorithms and datastructures in Section 7. The CSM definition is
rather general, and so in Section 8 we define a more restricted model called the C2 -
CSM. Compared to the CSM, the C2 -CSM is somewhat closer to optical computing as
it happens in the laboratory. Finally, in Section 9 we show the power and limitations
of optical computing, as embodied by the C2 -CSM, in terms computational complexity
theory. Optical information processing is a highly parallel form of computing and we
make this intuition more concrete by relating the C2 -CSM to parallel complexity theory
by characterising the parallel complexity class NC. For example, this shows the kind of
worst case resource usage one would expect when applying CSM algorithms to problems
that are known to be suited to parallel solutions.

6.1. CSM definition

We begin this section by describing the CSM model in its most general setting, this
brief overview is not intended to be complete and more details are to be found in [93].
A complex-valued image (or simply, image) is a function f : [0, 1) × [0, 1) → C, where
[0, 1) is the half-open real unit interval. We let I denote the set of complex-valued images.
Let N+ = {1, 2, 3, . . .}, N = N+ ∪ {0}, and for a given CSM M let N be a countable set
of images that encode M ’s addresses. An address is an element of N × N. Additionally,

10
for a given M there is an address encoding function E : N → N such that E is Turing
machine decidable, under some reasonable representation of images as words.

Definition 1 (CSM) A CSM is a quintuple M = (E, L, I, P, O), where

E : N → N is the address encoding function,
L = ((sξ , sη ) , (aξ , aη ) , (bξ , bη )) are the addresses: sta, a and b, where a 6= b,
I and O are finite sets of input and output addresses, respectively,
P = {(ζ1 , p1ξ , p1η ), . . . , (ζr , prξ , prη )} are the r programming symbols ζj and
their addresses (pjξ , pjη ) where ζj ∈ ({h, v, ∗, ·, +, ρ, st, ld, br, hlt} ∪ N ) ⊂ I.
Each address is an element from {0, . . . , Ξ − 1} × {0, . . . , Y − 1} where Ξ, Y ∈ N+ .
Addresses whose contents are not specified by P in a CSM definition are assumed
to contain the constant image f (x, y) = 0. We interpret this definition to mean that
M is (initially) defined on a grid of images bounded by the constants Ξ and Y, in the
horizontal and vertical directions respectively. The grid of images may grow in size as the
computation progresses. In our grid notation the first and second elements of an address
tuple refer to the horizontal and vertical axes of the grid respectively, and image (0, 0) is
located at the lower left-hand corner of the grid. The images have the same orientation
as the grid. For example the value f (0, 0) is located at the lower left-hand corner of the
image f .
In Definition 1 the tuple P specifies the CSM program using programming symbol
images ζj that are from the (low-level) CSM programing language [93,98]. We refrain from
giving a description of this programming language and instead describe a less cumbersome
high-level language [93]. Figure 1 gives the basic instructions of this high-level language.
The copy instruction is illustrated in Figure 2. There are also if/else and while control
flow instructions with conditional equality tests of the form (fψ == fφ ) where fψ and fφ
are binary symbol images (see Figures 3(a) and 3(b)).
Address sta is the start location for the program so the programmer should write
the first program instruction at sta. Addresses a and b define special images that are
frequently used by some program instructions. The function E is specified by the pro-
grammer and is used to map addresses to image pairs. This enables the programmer to
choose her own address encoding scheme. We typically don’t want E to hide complicated
behaviour thus the computational power of this function should be somewhat restricted.
For example we put such a restriction on E in Definition 7. At any given timestep, a
configuration is defined in a straightforward way as a tuple hc, ei where c is an address
called the control and e represents the grid contents.

6.2. Complexity measures

In this section we define a number of CSM complexity measures. As is standard, all

resource bounding functions map from N into N and are assumed to have the usual
properties [5]. We begin by defining CSM time complexity in a manner that is standard
among parallel models of computation.

Definition 2 The time complexity of a CSM M is the number of configurations in the

computation sequence of M , beginning with the initial configuration and ending with the
first final configuration.

11
h(i1 ;i2 ) : replace image at i2 with horizontal 1D Fourier transform of i1 .

v(i1 ;i2 ) : replace image at i2 with vertical 1D Fourier transform of image at i1 .

∗(i1 ;i2 ) : replace image at i2 with the complex conjugate of image at i1 .

· (i1 ,i2 ;i3 ) : pointwise multiply the two images at i1 and i2 . Store result at i3 .

+(i1 ,i2 ;i3 ) : pointwise addition of the two images at i1 and i2 . Store result at i3 .

ρ(i1 ,zl ,zu ;i2 ) : filter the image at i1 by amplitude using zl and zu as lower and upper
amplitude threshold images, respectively. Place result at i2 .
[ξ1 , ξ2 , η1 , η2 ] ← [ξ1 , ξ2 , η1 , η2 ] : copy the rectangle of images whose bottom left-hand
′ ′ ′ ′

address is (ξ1 , η1 ) and whose top right-hand address is (ξ2 , η2 ) to the

rectangle of images whose bottom left-hand address is (ξ1′ , η1′ ) and whose
top right-hand address is (ξ2′ , η2′ ). See illustration in Figure 2.
Fig. 1. CSM high-level programming language instructions. In these instructions i, zl , zu ∈ N × N are
image addresses and ξ, η ∈ N. The control flow instructions are described in the main text.

η
i
ξ ξ+3

Fig. 2. Illustration of the instruction i ← [ξ, ξ + 3, η, η] that copies four images to a single image that is
denoted i.

(a) (b) (c) (d) (e) (f)

Fig. 3. Representing binary data. The shaded areas denote value 1 and the white areas denote value
0. (a) Binary symbol image representation of 1 and (b) of 0, (c) list (or row) image representation of
the word 1011, (d) column image representation of 1011, (e) 3 × 4 matrix image, (f) binary stack image
representation of 1101. Dashed lines are for illustration purposes only.

The first of our six space-like resources is called grid.

Definition 3 The grid complexity of a CSM M is the minimum number of images,

arranged in a rectangular grid, for M to compute correctly on all inputs.

Let S : I × (N × N) → I, where S(f (x, y), (Φ, Ψ)) is a raster image, with ΦΨ constant-
valued pixels arranged in Φ columns and Ψ rows, that approximates f (x, y). If we choose
a reasonable and realistic S then the details of S are not important.

Definition 4 The spatialRes complexity of a CSM M is the minimum ΦΨ such that if

each image f (x, y) in the computation of M is replaced with
S(f (x, y), (Φ, Ψ)) then M computes correctly on all inputs.

12
One can think of spatialRes as a measure of the number of pixels needed during a
computation. In optical image processing terms, and given the fixed size of our images,
spatialRes corresponds to the space-bandwidth product of a detector or SLM.

Definition 5 The dyRange complexity of a CSM M is the ceiling of the maximum of

all the amplitude values stored in all of M ’s images during M ’s computation.
In optical processing terms dyRange corresponds to the dynamic range of a signal.
We also use complexity measures called amplRes, phaseRes and freq [93,98]. Roughly
speaking, the amplRes of a CSM M is the number of discrete, evenly spaced, amplitude
values per unit amplitude of the complex numbers in M ’s images. For example, we would
need amplRes of 3 to directly store values from the set {0, ±1/3, ±2/3, ±1, ±4/3, . . .} as
complex values in an image. Thus amplRes corresponds to the amplitude quantisation
of a signal. The phaseRes of M is the total number (per 2π) of discrete evenly spaced
phase values in M ’s images, and so phaseRes corresponds to the phase quantisation of
a signal. For example, we would need phaseRes of 3 to directly store values from the set
{eix | x ∈ {0, 2/3π, 4/3π}} as complex values in an image. Finally, the freq complexity
of a CSM M is the minimum optical frequency necessary for M to compute correctly,
this concept is explained further in [98].
Often we wish to make analogies between space on some well-known model and CSM
‘space-like’ resources. Thus we define the following convenient term.

Definition 6 The space complexity of a CSM M is the product of the grid, spatialRes,
dyRange, amplRes, phaseRes and freq complexities of M .

6.3. Optical realisation

In this section, we outline how some of the elementary operations of the CSM could
be carried out physically. We do not intend to specify the definitive realisation of any of
the operations, but simply convince the reader that the model’s operations have physical
interpretations. Furthermore, although we concentrate on implementations employing
visible light (optical frequencies detectable to the human eye) the CSM definition does
not preclude employing other portion(s) of the electromagnetic spectrum.
A complex-valued image could be represented physically by a spatially coherent optical
wavefront. Spatially coherent illumination (light of a single wavelength and emitted with
the same phase angle) can be produced by a laser. A SLM could be used to encode the
image onto the expanded and collimated laser beam. One could write to a SLM offline
(expose photographic film, or laser print or relief etch a transparency) or online (in the
case of a liquid-crystal display [102,62,91] or holographic material [24,74]). The functions
h and v could be effected using two convex cylindrical lenses, oriented horizontally and
vertically, respectively [90,35,62,34].
A coherent optical wavefront will naturally evolve into its own Fourier spectrum as
it propagates to infinity. What we do with a convex lens is simply image, at a finite
distance, this spectrum. This finite distance is called the focal length of the lens. The
constant θ used in the definitions of h and v could be effected using Fourier spectrum
size reduction techniques [90,35] such as varying the focal length of the lens, varying the
separation of the lens and SLM, employing cascaded Fourier transformation, increasing

13
the dimensions/reducing the spatial resolution of the SLM, or using light with a shorter
wavelength.
The function ∗ could be implemented using a phase conjugate mirror [25]. The func-
tion · could be realised by placing a SLM encoding an image f in the path of a wavefront
encoding another image g [90,35,89]. The wavefront immediately behind the SLM would
then be ·(f, g). The function + describes the superposition of two optical wavefronts.
This could be achieved using a 50:50 beam splitter [90,25,92]. The function ρ could be
implemented using an electronic camera or a liquid-crystal light valve [91]. The parame-
ters zl and zu would then be physical characteristics of the particular camera/light valve
used. zl corresponds to the minimum intensity value that the device responds to, known
as the dark current signal, and zu corresponds to the maximum intensity (the saturation
level).
A note will be made about the possibility of automating these operations. If suitable
SLMs can be prepared with the appropriate 2D pattern(s), each of the operations h, v, ∗, ·,
and + could be effected autonomously and without user intervention using appropriately
positioned lenses and free space propagation. The time to effect these operations would
be the sum of the flight time of the image (distance divided by velocity of light) and
the response time of the analog 2D detector; both of which are constants independent
of the size or resolution of the images if an appropriate 2D detector is chosen. Examples
of appropriate detectors would be holographic material [24,74] and a liquid-crystal light
valve with a continuous (not pixellated) area [91]. Since these analog detectors are also
optically-addressed SLMs, we can very easily arrange for the output of one function to act
as the input to another, again in constant time independent of the size or resolution of the
image. A set of angled mirrors will allow the optical image to be fed back to the first SLM
in the sequence, also in constant time. It is not known, however, if ρ can be carried out
completely autonomously for arbitrary parameters. Setting arbitrary parameters might
fundamentally require offline user intervention (adjusting the gain of the camera, and so
on), but at least for a small range of values this can be simulated online using a pair of
liquid-crystal intensity filters.
We have outlined some optics principles that could be employed to implement the op-
erations of the model. The simplicity of the implementations hides some imperfections in
our suggested realisations. For example, the implementation of the + operation outlined
above results in an output image that has been unnecessarily multiplied by the constant
factor 0.5 due to the operation of the beam splitter. Also, in our suggested technique,
the output of the ρ function is squared unnecessarily. However, each of these effects can
be compensated for with a more elaborate optical setup and/or at the algorithm design
stage.
A more important issue concerns the quantum nature of light. According to our cur-
rent understanding, light exists as individual packets called photons. As such, in order
to physically realise the CSM one would have to modify it such that images would have
discrete, instead of continuous, amplitudes. The atomic operations outlined above, in
particular the Fourier transform, are not affected by the restriction to quantised ampli-
tudes, as the many experiments with electron interference patterns indicate. We would
still assume, however, that in the physical world space is continuous.
A final issue concerns how a theoretically infinite Fourier spectrum could be represented
by an image (or encoded by a SLM) of finite extent. This difficulty is addressed with the
freq complexity measure [98].

14
7. Example CSM datastructures and algorithms

In this section we give some example data representations. We then to go on to give

an example CSM algorithm that efficiently squares a binary matrix.

7.1. Representing data as images

There are many ways to represent data as images and interesting new algorithms
sometimes depend on a new data representation. Data representations should be in some
sense reasonable, for example it is unreasonable that the input to an algorithm could (non-
uniformly) encode solutions to NP-hard or even undecidable problems. From Section 8.1,
the CSM address encoding function gives the programmer room to be creative, so long
as the representation is logspace computable (assuming a reasonable representation of
images as words).
Here we mention some data representations that are commonly used. Figures 3(a)
and 3(b) are the binary symbol image representations of 1 and 0 respectively. These im-
ages have an everywhere constant value of 1 and 0 respectively, and both have spatialRes
of 1. The row and column image representations of the word 1011 are respectively given
in Figures 3(c) and 3(d). These row and column images both have spatialRes of 4. In
the matrix image representation in Figure 3(e), the first matrix element is represented at
the top left corner and elements are ordered in the usual matrix way. This 3 × 4 matrix
image has spatialRes of 12. Finally the binary stack image representation, which has
exponential spatialRes of 16, is given in Figure 3(f).
Figure 2 shows how we might form a list image by copying four images to one in a single
timestep. All of the above mentioned images have dyRange, amplRes and phaseRes
of 1.
Another useful representation is where the value of a pixel directly encodes a number,
in this case dyRange becomes crucial. We can also encode values as phase values, and
naturally phaseRes becomes a useful measure of the resources needed to store such
values.

7.2. A matrix squaring algorithm

Here we give an example CSM algorithm (taken from [95]) that makes use of the data
representations described above. The algorithm squares a n × n matrix in O(log n) time
and O(n3 ) spatialRes (number of pixels), while all other CSM resources are constant.

Lemma 1 Let n be a power of 2 and let A be a n × n binary matrix. The matrix A2

is computed by a C2 -CSM, using the matrix image representation, in time O(log n),
spatialRes O(n3 ), grid O(1), dyRange O(1), amplRes 1 and phaseRes 1.

Proof. (Sketch) In this proof the matrix and its matrix image representation (see Fig-
ure 3(e)) are both denoted A. We begin with some precomputation, then one parallel
pointwise multiplication step, followed by log n additions to complete the algorithm.
We generate the matrix image A1 that consists of n vertically juxtaposed copies of
A. This is computed by placing one copy of A above the other, scaling to one image,

15
and repeating to give a total of log n iterations. The image A1 is constructed in time
O(log n), grid O(1) and spatialRes O(n3 ).
Next we transpose A to the column image A2 . The first n elements of A2 are row 1 of
A, the second n elements of A2 are row 2 of A, etc. This is computed in time O(log n),
grid O(1) and spatialRes O(n2 ) as follows.
Let A′ = A and i = n. We horizontally split A′ into a left image A′L and a right
image A′R . Then A′L is pointwise multiplied (or masked) by the column image that
represents (10)i , in time O(1). Similarly A′R is pointwise multiplied (or masked) by the
column image that represents (01)i . The masked images are added. The resulting image
has half the number of columns as A′ and double the number of rows, and for example:
row 1 consists of the first half of the elements of row 1 of A′ and row 2 consists of the
latter half of the elements of row 1 of A′ . We call the result A′ and we double the value
of i. We repeat the process to give a total of log n iterations. After these iterations the
resulting column image is denoted A2 .
We pointwise multiply A1 and A2 to give A3 in time O(1), grid O(1) and spatialRes O(n3 ).
To facilitate a straightforward addition we first transpose A3 in the following way:
A3 is vertically split into a bottom and a top image, the top image is placed to the
left of the bottom and the two are scaled to a single image, this splitting and scaling is
repeated to give a total of log n iterations and we call the result A4 . Then to perform
the addition, we vertically split A4 into a bottom and a top image. The top image is
pointwise added to the bottom image and the result is thresholded between 0 and 1.
This splitting, adding and thresholding is repeated a total of log n iterations to create
A5 . We ‘reverse’ the transposition that created A4 : image A5 is horizontally split into a
left and a right image, the left image is placed above the right and the two are scaled
to a single image, this splitting and scaling is repeated a total of log n iterations to give
A2 . 2
The algorithm highlights a few points of interest about the CSM. The CSM has quite
a number of space-like resources, and it is possible to have trade-offs between them. For
example in the algorithm above, if we allow grid to increase from O(1) to O(n) then
the spatialRes can be reduced from O(n3 ) to O(n2 ). In terms of optical architectures
modelled by the CSM this phenomenon could be potentially very useful as certain re-
sources may well be more economically viable than others. The algorithm is used in the
proof that that polynomial time CSMs (and C2 -CSMs, see below) compute problems
that are in the PSPACE class of languages. PSPACE includes the famous NP class. Such
computational complexity results are discussed further in Section 9 below.
There are a number of existing CSM algorithms, for these we point the reader to the
literature [63–65,93,95,97,98].

8. C2 -CSM

In this section we define the C2 -CSM. One of the motivations for this model is the need
to put reasonable upper bounds on the power of reasonable optical computers. As we’ve
shown elsewhere [96], it turns out that CSMs can very quickly use massive amounts of
resources, and the C2 -CSM definition is an attempt to define a more reasonable model,
especially towards the goal of providing useful upper bounds on its power.

16
8.1. C2 -CSM

Motivated by a desire to apply standard complexity theory tools to the model, we

defined [93,96] the C2 -CSM, a restricted and more realistic class of CSM.

Definition 7 (C2 -CSM) A C2 -CSM is a CSM whose computation time is defined for
t ∈ {1, 2, . . . , T (n)} and has the following restrictions:
– For all time t both amplRes and phaseRes have constant value of 2.
– For all time t each of grid, spatialRes and dyRange is 2O(t) and space is rede-
fined to be the product of all complexity measures except time and freq.
– Operations h and v compute the discrete Fourier transform in the horizontal and ver-
tical directions respectively.
– Given some reasonable binary word representation of the set of addresses N , the ad-
dress encoding function E : N → N is decidable by a logspace Turing machine.
Let us discuss these restrictions. The restrictions on amplRes and phaseRes imply
that C2 -CSM images are of the form f : [0, 1) × [0, 1) → {0, ± 21 , ±1, ± 23 , . . .}. We have
replaced the Fourier transform with the discrete Fourier transform [9], this essentially
means that freq is now solely dependent on spatialRes; hence freq is not an interest-
ing complexity measure for C2 -CSMs and we do not analyse C2 -CSMs in terms of freq
complexity [93,96]. Restricting the growth of space is not unique to our model, such
restrictions are to be found elsewhere [33,69,73].
In Section 6.1 we stated that the address encoding function E should be Turing ma-
chine decidable, here we strengthen this condition. At first glance sequential logspace
computability may perhaps seem like a strong restriction, but in fact it is quite weak.
From an optical implementation point of view it should be the case that E is not compli-
cated, otherwise we cannot assume fast addressing. Other (sequential/parallel) models
usually have a very restricted ‘addressing function’: in most cases it is simply the identity
function on N. Without an explicit or implicit restriction on the computational complex-
ity of E, finding non-trivial upper bounds on the power of C2 -CSMs is impossible as E
could encode an arbitrarily complex Turing machine. As a weaker restriction we could
give a specific E. However, this restricts the generality of the model and prohibits the
programmer from developing novel, reasonable, addressing schemes.

9. Optical computing and computational complexity

As we saw in Section 5, there are number of optical algorithms that use the inherent
parallelism of optics to provide fast solutions to certain problems. An alternative approach
is to ask the following question: How does a given optical model relate to standard se-
quential and parallel models? Establishing a relationship with computational complexity
theory, by describing both upper and lower bounds on the model, gives immediate access
to a large collection of useful algorithms and proof techniques.
The parallel computation thesis [31,23,49,87,69] states that parallel time (polynomi-
ally) corresponds to sequential space, for reasonable parallel and sequential models. An
example would be the fact that the class of problems solvable in polynomial space on a
number of parallel models is equivalent to PSPACE, the class of problems solvable on
Turing machines that use at most polynomial space [40,8,30,32,22,33,88,85,83,3,84].

17
 Of course the thesis can never be proved, it relates the intuitive notion of reasonable
parallelism to the precise notion of a Turing machine. When results of this type were
first shown researchers were suitably impressed; their parallel models truly had great
power. For example if model M verifies the thesis then M decides PSPACE (includ-
ing NP) languages in polynomial time. However there is another side to this coin. It is
straightforward to verify that given our current best algorithms, M will use at least a
superpolynomial amount of some other resource (like space or number of processors) to
decide a PSPACE-complete or NP-complete language. Since the composition of polyno-
mials is itself a polynomial, it follows that if we restrict the parallel computer to use at
most polynomial time and polynomial other resources, then it can at most solve problems
in P.
Nevertheless, asking if M verifies the thesis is an important question. Certain problems,
such as those in the class NC, are efficiently parallelisable. NC can be defined as the class
of problems that are solvable in polylogarithmic time on a parallel computer that uses a
polynomial amount of hardware. So one can think of NC as those problems in P which
are solved exponentially faster on parallel computation thesis models than on sequential
models. If M verifies the thesis then it may be useful to apply M to these problems. We
also know that if M verifies the thesis then there are (P-complete) problems for which
it is widely believed that we will not find exponential speed-up using M .

9.1. C2 -CSM and parallel complexity theory

Here we summarise some characterisations of the computing power of optical comput-

ers. Such characterisations enable the algorithm designer to know what kinds of problems
are solvable with resource bounded optical algorithms.
Theorem 2 below gives lower bounds on the computational power of the C2 -CSM by
showing that it is at least as powerful as models that verify the parallel computation
thesis.

Theorem 2 ([95,97]) NSPACE(S(n)) ⊆ C2 -CSM–TIME(O(S 2 (n)))

In particular, polynomial time C2 -CSMs accept the PSPACE languages. PSPACE is
the class of problems solvable by Turing machines that use polynomial space, which
includes the famous class NP, and so NP-complete problems can be solved by C2 -CSMs
in polynomial time. However, any C2 -CSM algorithm that we could presently write to
solve PSPACE or NP problems would require exponential space.
Theorem 2 is established by giving a C2 -CSM algorithm that efficiently generates, and
squares, the transition matrix of a S(n) = Ω(log n) space bounded Turing machine.
This transition matrix represents all possible computations of the Turing machine and
is of size O(2S ) × O(2S ). The matrix squaring part was already given as an example
(Lemma 1), and the remainder of the algorithm is given in [95]. The algorithm uses
space that is cubic in one of the matrix dimensions. In particular the algorithm uses
cubic spatialRes, O(23S ), and all other space-like resources are constant. This theorem
improves upon the time overhead of a previous similar result [93,97] that was established
via C2 -CSM simulation of the vector machines [72,73] of Pratt, Rabin, and Stockmeyer.
From the resource usage point of view, it is interesting to see that the older of these two
algorithms uses grid, dyRange, and spatialRes that are each O(2S ), while the newer

18
algorithm shows that if we allow more spatialRes we can in fact use only constant grid
and dyRange. It would be interesting to find other such resource trade-offs within the
model.
Since NP is contained in PSPACE, Theorem 2, and the corresponding earlier results
in [93,97], show that this optical model solves NP-complete problems in polynomial time.
As described in Section 5, this has also been shown experimentally, for example Shaked
et al. [80] have recently given a polynomial time, exponential space, optical algorithm
to solve the NP-complete travelling salesperson problem. Their optical setup can be
implemented on the CSM.
The other of the two inclusions that are necessary in order to verify the parallel com-
putation thesis have also been shown: C2 -CSMs computing in time T (n) are no more
powerful than T O(1) (n) space bounded deterministic Turing machines. More precisely,
we have:

Theorem 3 ([93,94]) C2 -CSM-TIME(T (n)) ⊆ DSPACE(O(T 2 (n)))

This result gives an upper bound on the power of C2 -CSMs and was established via C2 -
CSM simulation by logspace uniform circuits of size and depth polynomial in space and
time respectively [94].
These computational complexity results for the C2 -CSM have shown that the model is
capable of parallel processing in much the same way as models that verify the parallel
computation thesis (and models that are known to characterise the parallel class NC).
These results strongly depend on their use of non-constant spatialRes. The algorithms
exploit the ability of optical computers, and the CSM in particular, to operate on large
numbers of pixels in parallel. But what happens when we do not have arbitrary numbers
of pixels? If allow images to have only a constant number of pixels then we need to
find new CSM algorithms. It turns out that such machines also characterise PSPACE in
polynomial time.

Theorem 4 ([99]) PSPACE is characterised by C2 -CSMs that are restricted to use poly-
nomial time T = O(nk ), spatialRes O(1), grid O(1), and generalised to use amplRes
T T
O(22 ), dyRange O(22 ).
So by treating images as registers and generating exponentially large, and exponentially
small, values we can solve seemingly intractable problems. Of course this kind of CSM is
quite unrealistic from the point of view of optical implementations. In particular, accurate
multiplication of such values is difficult to implement in optics [99].
To restrict the model we could replace arbitrary multiplication, by multiplication by
constants, which can be easily simulated by a constant number of additions. If we disallow
multiplication in this way, we characterise P.

Theorem 5 ([99]) C2 -CSMs without multiplication, that compute in polynomial time,

polynomial grid O(nk ), and spatialRes O(1), characterise P.
We can swap the roles of grid and spatialRes, and still obtain a P characterisation:

Theorem 6 ([99]) CSMs without multiplication, that compute in polynomial time,

polynomial spatialRes O(nk ), and grid O(1), characterise P.

19
 Theorems 5 and 6 give conditions under which our optical model essentially looses its
parallel abilities and acts like a standard sequential Turing machine.
Via the proofs of Theorems 2 and 3 we can show that C2 -CSMs that simultaneously
use polynomial space and polylogarithmic time solve exactly those problems in the
class NC.

Corollary 7 C2 -CSM-SPACE, TIME(nO(1) , logO(1) n) = NC

Problems in NC highlight the power of parallelism, as these problems can be solved
exponentially faster on a polynomial amount of parallel resources than on polynomial
time sequential machines. As further work in this area one could try to find alternate
characterisations of NC in terms of the C2 -CSM. In particular, one could try to find
further interesting trade-offs between the various space-like resources of the model. In
the real world this would correspond to computing over various different optical resources.
As discussed in Section 5 it would be interesting for optical algorithm designers to try to
design (implementable) optical algorithms for NC problems in an effort to find problems
that are well-suited to optical solutions.

References

[1] M. A. G. Abushagur and H. J. Caulfield. Speed and convergence of bimodal optical computers.
Optical Engineering, 26(1):22–27, Jan. 1987.
[2] L. M. Adleman. Molecular computation of solutions to combinatorial problems. Science, 266:1021–
1024, Nov. 1994.
[3] A. Alhazov and M. de Jesús Pérez-Jiménez. Uniform solution to QSAT using polarizationless
active membranes. In J. Durand-Lose and M. Margenstern, editors, Machines, Computations and
Universality (MCU), volume 4664 of LNCS, pages 122–133, Orléans, France, Sept. 2007. Springer.
[4] H. H. Arsenault and Y. Sheng. An Introduction to Optics in Computers, volume TT8 of Tutorial
Texts in Optical Engineering. SPIE Press, Bellingham, Washington, 1992.
[5] J. L. Balcázar, J. Dı́az, and J. Gabarró. Structural complexity, vols I and II. EATCS Monographs
on Theoretical Computer Science. Springer, Berlin, 1988.
[6] R. Barakat and J. H. Reif. Lower bounds on the computational efficiency of optical computing
systems. Applied Optics, 26(6):1015–1018, Mar. 1987.
[7] F. R. Beyette Jr., P. A. Mitkas, S. A. Feld, and C. W. Wilmsen. Bitonic sorting using an
optoelectronic recirculating architecture. Applied Optics, 33(35):8164–8172, Dec. 1994.
[8] A. Borodin. On relating time and space to size and depth. SIAM Journal on Computing, 6(4):733–
744, Dec. 1977.
[9] R. N. Bracewell. The Fourier transform and its applications. Electrical and electronic engineering
series. McGraw-Hill, second edition, 1978.
[10] K.-H. Brenner, A. Huang, and N. Streibl. Digital optical computing with symbolic substitution.
Applied Optics, 25(18):3054–3060, Sept. 1986.
[11] D. P. Casasent and G. P. House. Comparison of coherent and noncoherent optical correlators. In
Optical Pattern Recognition V, Proceedings of SPIE vol. 2237, pages 170–178, Apr. 1994.
[12] D. P. Casasent and D. Psaltis. Position, rotation, and scale invariant optical correlation. Applied
Optics, 15(7):1795–1799, 1976.
[13] H. J. Caulfield, editor. Handbook of Optical Holography. Academic Press, New York, 1979.
[14] H. J. Caulfield. The energetic advantage of analog over digital computing. In OSA Optical
Computing Technical Digest Series, volume 9, pages 180–183, 1989.
[15] H. J. Caulfield. Space-time complexity in optical computing. In B. Javidi, editor, Optical
information-processing systems and architectures II, volume 1347, pages 566–572. SPIE, July 1990.
[16] H. J. Caulfield. Space-time complexity in optical computing. Multidimensional Systems and Signal
Processing, 2(4):373–378, Nov. 1991. Special issue on optical signal processing.

20
[17] H. J. Caulfield and M. A. G. Abushagur. Hybrid analog-digital algebra processors. In Optical and
Hybrid Computing II, Proceedings of SPIE vol. 634, pages 86–95, Orlando, Florida, Apr. 1986.
[18] H. J. Caulfield and R. Haimes. Generalized matched filtering. Applied Optics, 19(2):181–183, Jan.
1980.
[19] H. J. Caulfield, S. Horvitz, and W. A. V. Winkle. Introduction to the special issue on optical
computing. Proceedings of the IEEE, 65(1):4–5, Jan. 1977.
[20] H. J. Caulfield, J. M. Kinser, and S. K. Rogers. Optical neural networks. Proceedings of the IEEE,
77:1573–1582, 1989.
[21] H. J. Caulfield, W. T. Rhodes, M. J. Foster, and S. Horvitz. Optical implementation of systolic
array processing. Optics Communications, 40:86–90, 1981.
[22] A. K. Chandra, D. C. Kozen, and L. J. Stockmeyer. Alternation. Journal of the ACM, 28(1):114–
133, Jan. 1981.
[23] A. K. Chandra and L. J. Stockmeyer. Alternation. In 17th annual symposium on Foundations of
Computer Science, pages 98–108, Houston, Texas, Oct. 1976. IEEE. (Preliminary Version).
[24] F. S. Chen, J. T. LaMacchia, and D. B. Fraser. Holographic storage in lithium niobate. Applied
Physics Letters, 13(7):223–225, oct 1968.
[25] A. E. Chiou. Photorefractive phase-conjugate optics for image processing, trapping, and
manipulation of microscopic objects. Proceedings of the IEEE, 87(12):2074–2085, Dec. 1999.
[26] S. Dolev and H. Fitoussi. The traveling beam: optical solution for bounded NP-complete problems.
In P. Crescenzi, G. Prencipe, and G. Pucci, editors, The fourth international conference on fun
with algorithms (FUN), pages 120–134, 2007.
[27] J. Durand-Lose. Reversible conservative rational abstract geometrical computation is Turing-
universal. In Logical Approaches to Computational Barriers, Second Conference on Computability
in Europe, (CiE), volume 3988 of Lecture Notes in Computer Science, pages 163–172, Swansea,
UK, 2006. Springer.
[28] N. H. Farhat and D. Psaltis. New approach to optical information processing based on the Hopfield
model. Journal of the Optical Society of America A, 1:1296, 1984.
[29] D. G. Feitelson. Optical Computing: A survey for computer scientists. MIT Press, Cambridge,
Massachusetts, 1988.
[30] S. Fortune and J. Wyllie. Parallelism in random access machines. In Proc. 10th Annual ACM
Symposium on Theory of Computing, pages 114–118, 1978.
[31] L. M. Goldschlager. Synchronous parallel computation. PhD thesis, University of Toronto,
Computer Science Department, Dec. 1977.
[32] L. M. Goldschlager. A unified approach to models of synchronous parallel machines. In Proc. 10th
Annual ACM Symposium on Theory of Computing, pages 89–94, 1978.
[33] L. M. Goldschlager. A universal interconnection pattern for parallel computers. Journal of the
ACM, 29(4):1073–1086, Oct. 1982.
[34] J. W. Goodman. Operations achievable with coherent optical information processing systems.
Proceedings of the IEEE, 65(1):29–38, Jan. 1977.
[35] J. W. Goodman. Introduction to Fourier Optics. Roberts & Company, Englewood, Colorado, third
edition, 2005.
[36] R. Greenlaw, H. J. Hoover, and W. L. Ruzzo. Limits to parallel computation: P-completeness
theory. Oxford university Press, Oxford, 1995.
[37] L. K. Grover. A fast quantum mechanical algorithm for database search. In Proc. 28th Annual
ACM Symposium on Theory of Computing, pages 212–219, May 1996.
[38] P. S. Guilfoyle, J. M. Hessenbruch, and R. V. Stone. Free-space interconnects for high-performance
optoelectronic switching. IEEE Computer, 31(2):69–75, Feb. 1998.
[39] T. Haist and W. Osten. An optical solution for the travelling salesman problem. Optics Express,
15(16):10473–10482, Aug. 2007. Erratum: Vol. 15(10), pp 12627. Aug. 2007.
[40] J. Hartmanis and J. Simon. On the power of multiplication in random access machines. In
Proceedings of the 15th annual symposium on switching and automata theory, pages 13–23, The
University of New Orleans, Oct. 1974. IEEE.
[41] T. Head. Formal language theory and DNA: an analysis of the generative capacity of specific
recombinant behaviors. Bulletin of Mathematical Biology, 49(6):737–759, 1987.
[42] J. L. Horner, editor. Optical Signal Processing. Academic Press, San Diego, 1987.
[43] Y.-N. Hsu and H. H. Arsenault. Optical pattern recognition using circular harmonic expansion.
Applied Optics, 21(22):4016–4019, Nov. 1982.

21
[44] A. Huang. Architectural considerations involved in the design of an optical digital computer.
Proceedings of the IEEE, 72(7):780–786, July 1984.
[45] A. D. Jacobson, T. D. Beard, W. P. Bleha, J. D. Morgerum, and S. Y. Wong. The liquid crystal light
value, an optical-to-optical interface device. In Proceedings of the Conference on Parallel Image
Processing, Goddard Space Flight Center, pages 288–299, Mar. 1972. Document X-711-72-308.
[46] B. Javidi. Nonlinear joint power spectrum based optical correlation. Applied Optics, 28(12):2358–
2367, June 1989.
[47] B. Javidi and J. Wang. Optimum distortion-invariant filter for detecting a noisy distorted target in
nonoverlapping background noise. Journal of the Optical Society of America A, 12(12):2604–2614,
Dec. 1995.
[48] M. A. Karim and A. A. S. Awwal. Optical Computing: An Introduction. Wiley, New York, 1992.
[49] R. M. Karp and V. Ramachandran. Parallel algorithms for shared memory machines. In J. van
Leeuwen, editor, Handbook of Theoretical Computer Science, volume A, chapter 17. Elsevier,
Amsterdam, 1990.
[50] E. Knill, R. LaFlamme, and G. J. Milburn. A scheme for efficient quantum computation with
linear optics. Nature, 409:46–52, 2001.
[51] J. N. Lee, editor. Design Issues in Optical Processing. Cambridge Studies in Modern Optics.
Cambridge University Press, Cambridge, Great Britain, 1995.
[52] J. N. Lee, editor. Design issues in optical processing. Cambridge studies in modern optics.
Cambridge University Press, 1995.
[53] Lenslet Labs. Enlight256. white paper report, Lenslet Ltd., 6 Galgalei Haplada St, Herzelia Pituach,
46733 Israel, Nov. 2003.
[54] R. J. Lipton. Using DNA to solve NP-complete problems. Science, 268:542–545, Apr. 1995.
[55] A. Louri and A. Post. Complexity analysis of optical-computing paradigms. Applied Optics,
31(26):5568–5583, Sept. 1992.
[56] P. D. MacKenzie and V. Ramachandran. Ercw prams and optical communication. Theoretical
Computer Science, 196:153–180, 1998.
[57] A. D. McAulay. Optical Computer Architectures: The Application of Optical Concepts to Next
Generation Computers. Wiley, New York, 1991.
[58] C. Mead. Analog VLSI and Neural Systems. Addison-Wesley, Reading, Massachusetts, 1989.
[59] D. A. Miller. Rationale and challenges for optical interconnects to electronic chips. Proceedings of
the IEEE, 88(6):728–749, June 2000.
[60] C. Moore. Generalized shifts: undecidability and unpredictability in dynamical systems.
Nonlinearity, 4:199–230, 1991.
[61] C. Moore. Majority-vote cellular automata, Ising dynamics and P-completeness. Journal of
Statistical Physics, 88(3/4):795–805, 1997.
[62] T. Naughton, Z. Javadpour, J. Keating, M. Klı́ma, and J. Rott. General-purpose acousto-optic
connectionist processor. Optical Engineering, 38(7):1170–1177, July 1999.
[63] T. J. Naughton. Continuous-space model of computation is Turing universal. In S. Bains and
L. J. Irakliotis, editors, Critical Technologies for the Future of Computing, Proceedings of SPIE
vol. 4109, pages 121–128, San Diego, California, Aug. 2000.
[64] T. J. Naughton. A model of computation for Fourier optical processors. In R. A. Lessard and
T. Galstian, editors, Optics in Computing 2000, Proc. SPIE vol. 4089, pages 24–34, Quebec,
Canada, June 2000.
[65] T. J. Naughton and D. Woods. On the computational power of a continuous-space optical model
of computation. In M. Margenstern and Y. Rogozhin, editors, Machines, Computations and
Universality: Third International Conference (MCU’01), volume 2055 of LNCS, pages 288–299,
Chişinău, Moldova, May 2001. Springer.
[66] M. Oltean. A light-based device for solving the Hamiltonian path problem. In Fifth International
Conference on Unconventional Computation (UC’06), volume 4135 of LNCS, pages 217–227, York,
UK, 2006. Springer.
[67] E. L. O’Neill. Spatial filtering in optics. IRE Transactions on Information Theory, 2:56–65, June
1956.
[68] C. H. Papadimitriou. Computational complexity. Addison-Wesley, 1995.
[69] I. Parberry. Parallel complexity theory. Wiley, 1987.
[70] G. Păun. Membrane computing: an introduction. Springer, 2002.

22
[71] A. Pe’er, D. Wang, A. W. Lohmann, and A. A. Friesem. Optical correlation with totally incoherent
light. Optics Letters, 24(21):1469–1471, Nov. 1999.
[72] V. R. Pratt, M. O. Rabin, and L. J. Stockmeyer. A characterisation of the power of vector machines.
In Proc. 6th annual ACM symposium on theory of computing, pages 122–134. ACM press, 1974.
[73] V. R. Pratt and L. J. Stockmeyer. A characterisation of the power of vector machines. Journal of
Computer and Systems Sciences, 12:198–221, 1976.
[74] A. Pu, R. F. Denkewalter, and D. Psaltis. Real-time vehicle navigation using a holographic memory.
Optical Engineering, 36(10):2737–2746, Oct. 1997.
[75] J. Reif, D. Tygar, and A. Yoshida. The computability and complexity of optical beam tracing. In
31st Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 106–114, St.
Louis, MO, Oct. 1990. IEEE.
[76] J. H. Reif and A. Tyagi. Energy complexity of optical computations. In 2nd IEEE Symposium on
Parallel and Distributed Processing, pages 14–21, Dallas, TX, Dec. 1990.
[77] J. H. Reif and A. Tyagi. Efficient parallel algorithms for optical computing with the discrete Fourier
transform (DFT) primitive. Applied Optics, 36(29):7327–7340, Oct. 1997.
[78] A. A. Sawchuk and T. C. Strand. Digital optical computing. Proceedings of the IEEE, 72(7):758–
779, July 1984.
[79] N. T. Shaked, S. Messika, S. Dolev, and J. Rosen. Optical solution for bounded NP-complete
problems. Applied Optics, 46(5):711–724, Feb. 2007.
[80] N. T. Shaked, G. Simon, T. Tabib, S. Mesika, S. Dolev, and J. Rosen. Optical processor for solving
the traveling salesman problem (TSP). In B. Javidi, D. Psaltis, and H. J. Caulfield, editors, Proc.
of SPIE, Optical Information Systems IV, volume 63110G, Aug. 2006.
[81] N. T. Shaked, T. Tabib, G. Simon, S. Messika, S. Dolev, and J. Rosen. Optical binary-
matrix synthesis for solving bounded NP-complete combinatorical problems. Optical Engineering,
46(10):108201–1–108201–11, Oct. 2007.
[82] P. Shor. Algorithms for quantum computation: Discrete logarithms and factoring. In Proceedings
35th Annual Symposium on Foundations Computer Science, pages 124–134, 1994.
[83] P. Sosı́k. The computational power of cell division in P systems: Beating down parallel computers?
Natural Computing, 2(3):287–298, 2003.
[84] P. Sosı́k and A. Rodrı́guez-Patón. Membrane computing and complexity theory: A characterization
of PSPACE. Journal of Computer and System Sciences, 73(1):137–152, 2007.
[85] J. Tromp and P. van Emde Boas. Associative storage modification machines. In K. Ambos-Spies,
S. Homer, and U. Schöning, editors, Complexity theory: current research, pages 291–313. Cambridge
University Press, 1993.
[86] G. L. Turin. An introduction to matched filters. IRE Transactions on Information Theory,
6(3):311–329, June 1960.
[87] P. van Emde Boas. Machine models and simulations. In J. van Leeuwen, editor, Handbook of
Theoretical Computer Science, volume A, chapter 1. Elsevier, Amsterdam, 1990.
[88] J. van Leeuwen and J. Wiedermann. Array processing machines. BIT, 27:25–43, 1987.
[89] A. VanderLugt. Signal detection by complex spatial filtering. IEEE Transactions on Information
Theory, 10(2):139–145, Apr. 1964.
[90] A. VanderLugt. Optical Signal Processing. Wiley, New York, 1992.
[91] P.-Y. Wang and M. Saffman. Selecting optical patterns with spatial phase modulation. Optics
Letters, 24(16):1118–1120, Aug. 1999.
[92] C. S. Weaver and J. W. Goodman. A technique for optically convolving two functions. Applied
Optics, 5(7):1248–1249, July 1966.
[93] D. Woods. Computational complexity of an optical model of computation. PhD thesis, National
University of Ireland, Maynooth, 2005.
[94] D. Woods. Upper bounds on the computational power of an optical model of computation. In
X. Deng and D. Du, editors, 16th International Symposium on Algorithms and Computation
(ISAAC 2005), volume 3827 of LNCS, pages 777–788, Sanya, China, Dec. 2005. Springer.
[95] D. Woods. Optical computing and computational complexity. In Fifth International Conference
on Unconventional Computation (UC’06), volume 4135 of LNCS, pages 27–40, York, UK, 2006.
Springer. Invited.
[96] D. Woods and J. P. Gibson. Complexity of continuous space machine operations. In S. B.
Cooper, B. Löewe, and L. Torenvliet, editors, New Computational Paradigms, First Conference

23
 on Computability in Europe (CiE 2005), volume 3526 of LNCS, pages 540–551, Amsterdam, June
2005. Springer.
[97] D. Woods and J. P. Gibson. Lower bounds on the computational power of an optical model of
computation. In C. S. Calude, M. J. Dinneen, G. Păun, M. J. Pérez-Jiménez, and G. Rozenberg,
editors, Fourth International Conference on Unconventional Computation (UC’05), volume 3699
of LNCS, pages 237–250, Sevilla, Oct. 2005. Springer.
[98] D. Woods and T. J. Naughton. An optical model of computation. Theoretical Computer Science,
334(1-3):227–258, Apr. 2005.
[99] D. Woods and T. J. Naughton. Sequential and parallel optical computing. In International
Workshop on Optical SuperComputing, 2008. To appear.
[100] T. Yokomori. Molecular computing paradigm – toward freedom from Turing’s charm. Natural
computing, 1(4):333–390, 2002.
[101] F. T. S. Yu, S. Jutamulia, and S. Yin, editors. Introduction to information optics. Academic Press,
San Diego, 2001.
[102] F. T. S. Yu, T. Lu, X. Yang, and D. A. Gregory. Optical neural network with pocket-sized liquid-
crystal televisions. Optics Letters, 15(15):863–865, Aug. 1990.

24