Professional Documents
Culture Documents
An Optical Solution For The Traveling
An Optical Solution For The Traveling
Salesman Problem
Tobias Haist and Wolfgang Osten
Institut für Technische Optik, Universität Stuttgart
Pfaffenwalring 9, 70569 Stuttgart, Germany
haist@ito.uni-stuttgart.de
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10473
1. Introduction
The traveling salesman problem (TSP) is one of the most important unsolved problems in com-
puter science. Apart from practical applications like transportation, plotting, chip fabrication,
record balancing or routing (see e.g. Ref. [1] for an overview), the importance stems from the
fact that it belongs to the class of so called NP–complete problems. “NP–complete” more or
less means that for such problems the runtime for finding the optimum solution increases faster
than polynomial with the problem size N. It can be shown that once a solution to one NP–
complete problem is found it will be possible to map that result (in polynomial time) to any
other NP–complete problem[2]. Therefore fast optimum solutions to the TSP are of consider-
able importance for all NP–complete problems.
Figure 1 depicts the TSP: A salesman wants to visit a set of N cities. The task is to find the
shortest tour length through all the cities with the additional constraint to visit each city exactly
once. In the classical TSP the trip should end at the same city where it has started. That means
we have a round–trip. It is easy to see that for N cities there are (N − 1)!/2 possible tours.
So even for a moderate number of cities one would have to check an astronomical number of
possible tours (e.g. 30 cities lead to 4.4x1030 tours) in order to find the shortest one. For 60
cities we would have more tours than there are atoms in the universe.
It has been proven[2], that no polynomial time heuristic can guarantee the solution of the
TSP of a given size (assuming the widely believed conjecture P 6= NP). So the global optimum
for large problems in general cannot be found. But a lot of methods, e.g. based on local opti-
mization (e.g. 3-opt[3]) or global optimization and heuristics (e.g. simulated annealing[4], neu-
ral networks[5], tabu search[6], Lin–Kernighan algorithm[7], greedy-Algorithm[8]), for find-
ing good approximate solution have been used in the past. Even for large numbers of cities
(N > 100) very good solutions can be found today in reasonable time using conventional
computers[9]. Ref. [10] gives a good overview. Polynomial time algorithms are available if
one is satisfied with obtaining good solutions where one can even set upper bounds for the de-
viation from global optimality[9]. For large practical problem sizes typically solutions which
differ only by about 1-2 % from optimality can be found[10] because — compared to other
NP–complete problems - the TSP is somewhat of good nature for real life examples. Also
unconventional strategies like using quantum computers[11], DNA-computing[12], or optical
hardware[13] have been proposed in the last years. Shaked et al recently proposed an all optical
Fig. 1. The traveling salesman problem: Find the shortest round-trip through N cities.
method for solving the TSP[13]. Their approach doesn’t reduce the complexity of the problem
but rather introduces a fast method for solving the TSP by a parallel matrix–vector multiplica-
tion based on optical correlation.
Apart from all these advances finding the global solution for pathologically posed TSPs with
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10474
large N is still not possible by conventional computing[9]. In this paper we will propose the
first — to our knowledge — method that finds the global optimum of the TSP by reducing the
NP complexity of the problem to N 2 complexity. We achieve this by using white light interfer-
ometry (WLI)[14]. From the mathematical point of view we unfortunately haven’t disapproved
the commonly believed conjecture that NP 6= P (the Clay mathematics institute set out a price
of 1.000.000 US-$ for doing that) but we apparently reduced the complexity by a trick, namely
replacing operations by photons (and the physics of interference).
In section 2 we introduce the method as a gedankenexperiment. After that we explain in
section 3 in more detail how we avoid wrong interferences and in section 4 we show that the
practical implementation will be unrealistic for large problem sizes due to the low signal to
noise ratio. Finally we will state our opinion about the question whether the proposed method
should be regarded to be a quantum computer.
2. Basic idea
The proposed system (as shown in Fig. 2) is mainly a white light interferometer with a compli-
cated signal path. A wave-packet with limited temporal extension (due to the short coherence
length) entering the signal tour will travel through the network of N connected cities and finally
might exit at the final (=initial) city. If the optical path length through the network corresponds
to the reference path length, interference at the detector will be observed.
(a) Step A: Finding the minimum tour length (b) Step B: Finding the tour
The cities are connected by input and output fibers with a length corresponding to their
distance in reality. We denote the fiber length between city i and city j by di j . It should be
noted that by using fibers we can implement an arbitrary distance metric as well as asymmetric
connections between the cities (so called asymmetric TSP). The input signals from other cities
to the exit(=initial) city are connected directly to the white light interferometer.
Figure 3 shows the structure of one city. It has N input and N output channels. A photon
entering one of the input channels will travel through a fiber optical delay line before enter-
ing a fan–out element that outputs the photon to one of the output channels. In the quantum
mechanical particle model the choice of the output port is given purely randomly. In the wave-
optical model of course we have a splitting of the wave to the N output channels. Different
methods (e.g. Y-couplers[15]) can be used for a practical implementation and by using direc-
tional couplers we can circumvent that light leaks back to the input channels. To make the cities
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10475
Fig. 3. Assembly of one city: N input and N output ports (for the asymmetrical TSP) are
connected by a delay line of variable length. The overall delay for photons in city i is
given by Di (see Eq. 1). Fan-in and Fan–out elements (e.g. star-couplers) can be used for a
practical implementation.
distinguishable we use additional delay lines of different length within the cities.
The overall idea for globally solving the TSP is to perform the following four steps:
2. Computation of a lower bound for the overall optical path length for a correct solution.
3. Increasing the reference path length of the WLI, starting from the lower bound while
checking for correct interference to obtain the global minimum of the tour length.
The most critical step is to check for correct interferences (step 3). To this end a careful
choice of the delays within the cities is mandatory. For the TSP we chose (other choices are
possible) for city number i the length Di of the delay line
Di = α 2i + d (1)
with i = 0, 1, ..., N − 1 and with a small d in the range of tens of microns. The reason for this
choice will become clear in section 3. The constant α is chosen such that it is considerably
larger than N times the largest distance max(di j ) between two cities.
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10476
2) Computation of a lower bound for the overall optical path length for a correct solu-
tion
We first connect the white light source and the interferometer to one of the cities (Fig. 2 (a)).
For searching the global solution it doesn’t matter which city we chose as the starting city of
the trip because it is anyway a round–trip.
According to Eq. 1 visiting all the cities means that the overall path length L is at least
N−1
L = Nd+ ∑ α 2i + P (4)
i=0
> N d + α (2N − 1) (5)
with P being the overall path length traveled between the cities:
N
P = ∑ dPath i (6)
i=1
3) Increasing the reference path length of the WLI, starting from the lower bound while
checking for correct interference to obtain the global minimum of the tour length.
Now we continue to increase the reference path length until we detect interference. The first
interference signal will occur if the shortest tour is approached. This detection of interference is
a little bit more complicated as in a conventional white light interferometer because we have to
check for additional interferences due to wrong tours within our network of cities, e.g. if not all
cities were visited exactly once. We will describe the interference detection in detail in section
3.
If we have found the first position of interference we have completed step A and the path
length of the reference arm minus α (2N − 1) + N d exactly equals the minimum path length
that the salesman has to travel during his trip.
It should be pointed out that this was the hard part of the problem and itself corresponds to
a NP–complete problem. We continue by finding also the correct tour through the network of
cities.
∆Li = Li + di j (7)
with j being the index of the exit city. Of course we will also disconnect the input connections
to the final city.
If we still have a correct interference pattern (see section 3), city i is the correct last but
one city. We now regard this city as the exit city and continue the same way, now with N − 2
remaining cities.
Therefore we have to do less than
N−1
N2 − N
Nmax = (N − 1) + (N − 2) + ...2 = ∑ i= 2
−1 (8)
i=2
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10477
tests to completely find the optimum tour for the salesman. The complexity is the same as for
setting up the experiment and proportional to N 2 . Therefore we have reduced the NP–complete
problem to a problem with cost proportional N 2 . For the example at the beginning with N=30 we
have to perform 900 comparisons which is negligible compared to the original cost proportional
to 1030 .
#H
I0 + ∑#K
k=0 | ∑h=0 Wkh | is the (very strong) background intensity due to the light not interfering
k 2
with the reference. One might use heterodyne detection in order to get rid of this background
signal but we have to look at the signal–to–noise problem which leads to the practical limit for
the method more closely in section 4.
We eliminate the last double sum for example by making a second measurement with the
reference blocked:
2 2
#C #W #K #Hk
I2 = I0 + ∑ Ci + ∑ WC j + ∑ ∑ Wkh (10)
1 1 k=1 h=1
In the following it will become clear why we have chosen the delays in the cities according
to Eq. 1. We will show that by this choice we can effectively remove the WC j .
When connecting all the the cities by fibers we chose inter-city fiber lengths di j as well as
α (see Eq. 1) in multiples of 1 mm. Consequently, it follows that P then is also a multiple of 1
mm:
P = G x 1 mm (12)
with a positive integer G.
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10478
For a practical problem we might chose max(di j ) = 1 m. That means the resolution of our
problem is 1:1000 (e.g. 1 km for a trip with maximum inter city distances of 1000 km) which
is enough for non-artificial problems for finding the best trip but of course other choices —
with a better resolution — are possible. The computational cost does not depend on the chosen
resolution.
d is chosen larger than twice the coherence length of the light source. E.g. we might chose
d = 10 µ m.
Even if one doesn’t like such an argument based on the probability of detection one could
detect such a wrong tour by continuing with step B (finding the corresponding tour). In this
case the worst case computational complexity would be increased and more complicated to
calculate.
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10479
One might also check for the very unlikely case that ∑ Ci = 0 by randomly increasing some
(randomly chosen) distances between cities by a very small amount (e.g. λ /5) that is perform-
ing phase shifting.
Additional note: Since both directions of a tour are equally valid we always will have at least
two solutions. These two solutions lead to positive interference due to the equal path length of
both ways.
In conclusion we have shown that by detecting a modulation in the difference equation 11
while scanning (using the conventional axial shift of the white light interferometer) we detect
a possible solution of the TSP. If we start scanning from the lower bound α (2N − 1) + N d the
first solution that we will find is the global solution to the TSP.
4. Signal-to-Noise Ratio
For a TSP with N cities we have to use fan–out elements with N connections. That means that
the intensity of a wave packet entering the network will be attenuated by a factor of (1/N)N
before being detected by the interferometer.
Since we always have two possible solutions this results in an amplitude of 2 (1/N)N/2 and
the interference with the reference (which should have a large amplitude, we take here 2) leads
to an interference of
2
I = 2 + 2 N −N/2 eiφ ≈ 4 + 8 N −N/2 cos(φ ) (15)
1e+25
1e+20
necessary energy for 1 measurement [J]
1e+15
1e+10
100000
1e-05
1e-10
1e-15
1e-20
1 10 100
number of cities
Fig. 4. Necessary energy per measurement (at λ = 1 µ m)) for solving the TSP.
In principle photon noise (due to the Poisson distribution for coherent light) will set the
lower bound for the number of necessary photons in order to achieve detection. For the Poisson
distribution[16] the standard deviation sdev of the photonp
number is directly proportional to the
square root of the total photon number Ng , that is sdev = Ng .
If we claim that the standard deviation of the overall photon number (on the detector) is at
least larger than the expected signal we have to claim that
Ng < M16 N −N/2
p
(16)
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10480
where we denoted the total number of photons coming out of the network with M and the total
number of photons (dominated by the reference light) by Ng .
Therefore we can conclude that
6. Conclusions
We have shown that the complexity of the traveling salesman problem can be dramatically
reduced from N! to N 2 by optical means. To this end we employed white light interferometry
and a fiber optic model of the network of cities that the salesman should travel through.
The maximum number of cities N (problem size) is fundamentally limited by the number
of photons. Problems with N = 20 cities can be solved if 1 kW of power is available (1 s
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10481
integration time per individual measurement). Since for practical (non-pathological) problems
by purely electronic means very good solutions to even large size problems can be found, our
proposed method is not meant to solve real–world traveling salesman problems but rather as a
gedankenexperiment to show how photons and the laws of physics can considerably reduce the
computational complexity of difficult mathematical problems.
We want to thank Thomas Schuster, Christof Pruss, and Martin Schönleber for fruitful dis-
cussions on this topic.
#83676 - $15.00 USD Received 5 Jun 2007; revised 3 Jul 2007; accepted 3 Jul 2007; published 3 Aug 2007
(C) 2007 OSA 6 August 2007 / Vol. 15, No. 16 / OPTICS EXPRESS 10482