Directed negative-weight percolation

C. Norrenbrock,∗ M. M. Mkrtchian,† and A. K. Hartmann‡

Institut für Physik, Universität Oldenburg, 26111 Oldenburg, Germany
(Dated: October 5, 2018)
We consider a directed variant of the negative-weight percolation model in a two-dimensional,
periodic, square lattice. The problem exhibits edge weights which are taken from a distribution
that allows for both positive and negative values. Additionally, in this model variant all edges are
directed. For a given realization of the disorder, a minimally weighted loop/path configuration
arXiv:1701.04726v1 [cond-mat.dis-nn] 17 Jan 2017

is determined by performing a non-trivial transformation of the original lattice into a minimum

weight perfect matching problem. For this problem, fast polynomial-time algorithms are available,
thus we could study large systems with high accuracy. Depending on the fraction of negatively
and positively weighted edges in the lattice, a continuous phase transition can be identified, whose
characterizing critical exponents we have estimated by a finite-size scaling analyses of the numerically
obtained data. We observe a strong change of the universality class with respect to standard directed
percolation, as well as with respect to undirected negative-weight percolation. Furthermore, the
relation to directed polymers in random media is illustrated.

PACS numbers: 64.60.Ak,75.40.Cx,68.35.Rh

Keywords:

I. INTRODUCTION in directed percolation (DP) [5], which is a variant of

standard percolation, where the links carry a direction,
In statistical physics, one of the central targets is to leading to an anisotropic behavior. Note that this di-
study systems that exhibit continuous phase transitions. rectionality can be interpreted as time direction, making
Due to the diverging correlation length in the critical DP relevant for the description of non-equilibrium pro-
region, long-range correlations are not affected by de- cesses. In particular, because of the anisotropic nature
tails of microscopic interactions, but depend on symme- of the cluster building process in DP, correlations are not
try properties of the underlying model only. For that governed by one but two correlation lengths: ξk and ξ⊥ .
reason models that exhibit continuous phase transition Also, in standard percolation, the links do not carry
can be grouped in universality classes, which are charac- any weights, thus, one can assume all weights being one,
terized by a set of critical exponents and their functional i.e., they are in particular positive. Recently, a percola-
relations, i.e. scaling laws [1]. One of the most basic and tion model called “negative-weight percolation” (NWP)
intensively studied universality classes is that of stan- was introduced [6], where random weights are attached
dard percolation [2, 3], which addresses the question of to the links, and, in particular, weights of either sign are
connectivity. Based on a tunable parameter p, sites or allowed. Algorithmically, this means special global op-
links in a given lattice get occupied or stay empty. Then, timization polynomial-time “matching” algorithms have
the central objects of interest are clusters consisting of to applied, see Sec. II. This leads, interestingly, to a new
adjacent and occupied sites. Above a certain value of type of behavior giving rise to a different universality
p = pc , i.e. the critical point, a lattice-spanning cluster class compared to standard percolation. In a series of
emerges in the thermodynamic limit. Even the model is papers [7–13], NWP has been studied in different dimen-
probably the simplest possible model exhibiting a phase sions and different variants.
transition, its importance comes from the following facts.
First, it allows to study basically all fundamental aspects It has been shown that two distinct phases can be
of phase transitions within a very basic framework. Sec- identified depending on a disorder parameter ρ, which
ond, many much more complex phase transitions can be controls the amount of negative weights. (i) for small ρ
traced back to an underlying percolation transition, e.g., the geometric objects are rather small and straight-lined,
the percolation of Fortuin-Kasteleyn clusters [4] in the which reflects a self-affine scaling, (ii) for large ρ the geo-
Ising model. metric objects scale self-similar and can wind around the
In standard percolation, there is no directional infor- lattice. In Ref. [6] the disorder-driven phase transition
mation in the connectivity pattern. Thus, not surpris- was investigated by means of finite-size scaling analyses
ingly, the critical exponents describing this phase transi- and it turned out that the critical exponents were univer-
tion differ from those that describe the phase transition sal in 2D (different lattice geometries and disorder distri-
bution were studied). Further studies regarding isotropic
NWP address the influence of dilution on the critical
properties [7], the upper critical dimension (du = 6) [8],
∗ Electronic address: [email protected] another upper critical dimension (dDPL u = 3) for densely
† Electronic address: [email protected] packed loops far above the critical point [9], the mean-
‡ Electronic address: [email protected] field behavior on a random graph with fixed connectivity
 2

[10], the Schramm-Loewner evolution properties of paths

in 2D lattices [12], and loop-length distributions in sev-
eral dimensions [11].
Nevertheless, all this work was for non-directed lattices
or graphs. Thus, as compared to the change which occurs
when moving from standard to directed percolation, it is
valid to ask whether the directed variant of NWP, which
is introduced and studied in this work, gives again rise
to a new type of behavior.
Note that, while DP is defined as a local growth pro- FIG. 1: Illustration of minimum-weight configurations con-
cess, the path-like clusters in NWP emerge due to global sisting of loops (gray) and one path (black) in a directed 2D
optimization. This is also true for directed polymers in square lattice of side length L = 32 with periodic boundary
random media (DPRM) [14], but unlike NWP, DPRM conditions. The path is forced to start at the right bottom
does not feature a phase transition. Nevertheless, it will corner. For small values of ρ, there does not appear a loop
be outlined in this article that NWP and DPRM are par- and also the path does not span the lattice. At ρ = ρc one
tially related to each other. percolating loop occurs. For large values of ρ, there are many
spanning loops and also the path is percolating.
Next, we outline the model and the numerical proce-
dures. NWP can be defined on any graph, in particular
finite-dimensional lattices, which we consider here. In
particular we study here a directed, weighted, periodic, ing path and no percolating loop might appear for some
simple square lattices with side length L. The direction realizations. Above the critical point ρ > ρc the num-
of the edges is arranged as follows: All horizontal edges ber of lattice-spanning loops increases and even the path
point to the left and all vertical ones point up. Edge winds around the lattice (cf. Fig. 1(c)). Thus, this article
weights are drawn from a distribution which provides we study the disorder-driven, geometric phase transition
both positive and negative values. The proportion of and determine its characterizing critical exponents, in
negative and positive weights can be tuned by a disorder particular with respect to the parallel and perpendicular
parameter ρ. correlation lengths.
We study paths and loops in the lattice. For each path As an interpretation of the NWP problem, one can
or loops a weight is defined which consist of the sum imagine an agent that takes a trip in a graph along the
of the weights of the edges contributing to the loop or path.Whenever he travels along a positively weighted
path. For a given realization of the disorder, an optimal edge, the agent has to pay some resource according to
configuration consisting of one path and zero or possibly the positive value. On the other hand, he will harvest
a finite number of system-spanning loops is determined. some resource, if he travels along a negatively weighted
The configuration must fulfill some constrains: The path edge. Therefore, the optimal path/loop configuration ob-
must be fixed at the bottom right, the loops and the tained in the context of the NWP problem provides the
path are not allowed to intersect one another, and the optimal route of the agent (path), possibly in competi-
total sum of all weights of the loops and the path must tion with other agents (loops), to gain as many resources
be an exact minimum. The algorithm is very similar to as possible. Only paths or loops which lead to a larger
the undirected case, the changes which are necessary are amount of harvested resources as compared to the paid
explained in Sec. II. resources will occur.
Fig. 1 shows such optimal configurations for different The remainder of this article is organized as follows.
values of ρ in a lattice of size L = 32. The main question In Sec. II, we introduce the model in more detail and
we are interested in is, whether there is a system span- explain the algorithm. In Sec. III, we describe the finite-
ning path or system spanning loops, i.e., whether at least size scaling technique that has been used to estimate the
one object percolates. Note that for analyzing geomet- critical exponents numerically and present our results.
rical properties we are using the path. The loops arise We close with a summary in Sec. IV.
(unavoidably) from the global nature of the underlying
optimization problem, see the technical details in Sec.
III and allow us to study the percolation transition at
ρ = ρc , i.e., whether the full lattice admits one or several
II. MODEL AND ALGORITHM
percolating objects. Note that here, due to the construc-
tion of the underlying lattice with directed edges, small
loops can not occur, in contrast to the undirected NWP. The underlying graph G = (V, E) at hand is a 2D
In particular, this is true for ρ < ρc (cf. Fig. 1(a)) where directed square lattice whose edges point either to the
just a finite path appears and loops are absent. At the left or up. Its boundaries are periodic meaning the lattice
critical point ρ = ρc there appears a lattice-spanning loop can be considered as placed on a torus in a topological
in the depicted example Fig. 1(b), since there are enough sense. Each edge eij ∈ E carries a weight ωij that is
negatively weighted edges in the lattice. Also a percolat- taken from a “Gauss-like” distribution characterized by
 3

a tunable disorder parameter ρ: (a) (b)

√
P (ω) = (1 − ρ)δ(ω − 1) + ρ exp(−ω 2 )/ 2π,
0 ≤ ρ ≤ 1. (1)

The shape of the lattice is quadratic in all simulations,

so the number of nodes is N = |V | = L2 .
Given such a graph and a realization of the disorder,
an optimal configuration consisting of an arbitrary num-
ber of loops, i.e., closed paths, and one additional path
(possibly with zero length is computed. The configura-
tion must fulfill following requirements: i) One endpoint (c) (d)
of the path must be pinned at the bottom right corner.
However, it is also allowed that no path occurs. ii) The
loops and the paths are not allowed to cross or touch
each other. iii) The configurational energy
X
E= ωL (2)
L∈C

has to be minimized. Here ωL denotes the total of all

edge weights belonging to loop or path L. Note that
Eq. 1 provides real numbers, so the optimal configura- FIG. 2: Illustration of the algorithmic procedure for a peri-
tion is unique for each realization of the disorder. Since odic lattice of size L = 3. For the sake of clarity, the procedure
the number of loops is not specified and even the path is just depicted for a directed lattice that does not contain a
might not appear (zero length), also an empty configu- finite path. In Sec. II it is described how the construction
ration might be valid. This would be the case, e.g., if all of the auxiliary graph must be altered in order to force a
edges carried a positive weight. Loops and also the path path in the lattice that starts at the bottom right corner and
can solely appear, if their weight is negative, otherwise E terminates at any node. (a) Original lattice with weighted,
would not be minimal, since an empty configuration has directed edges. (b) Auxiliary graph with proper weight as-
E = 0. Furthermore, since all edges point either to the signment. The thick edges carry the weight as the respective
left or up, loops can appear only, if they span the lattice edges in the original graph. The weights of all other edges are
zero. (c) Illustration of the MWPM: black edges are matched
in either horizontal or vertical direction. Therefore, their
and gray ones are unmatched. For the sake of clarity, edge
smallest length is L. As a matter of fact, the typical weights are not depicted. (d) Reconstruction to the original
length of the loops is 2L in the vicinity of the critical lattice taking the MWPM result into account.
point.
In order to find the optimal configuration, we trans-
form the original graph to an appropriate auxiliary graph
first. Subsequently, a minimum-weight perfect matching 2, this node is white). Subsequently, each of the original
(MWPM) [15–17] provides all information to reconstruct edges is replaced by a path of three edges and two ad-
the original graph exhibiting the correct loop/path con- ditional nodes (in Fig. 2(b), these nodes are depicted as
figuration. Fig. 2 is an illustration of the algorithmic squares). One of the two edges that are connected with
procedure for a given realization of the disorder for a pe- the duplicates of the original nodes is assigned with the
riodic lattice of size L = 3. For reasons of clarity, only weight of the original edge. The other two edges carry
loops but no path can occur here. Guided by Fig. 2, zero weight. The resulting auxiliary graph is illustrated
we give a concise description how the algorithm works. in Fig. 2(b), where bold edges carry the original weights
Afterwards, we explain in which way the algorithm must and the thin ones carry a weight of zero. A more ex-
be altered, so that the appearance of a path becomes tensive and pedagogical description of the mapping (for
possible. the undirected variant of the model, where the auxiliary
(1) First all original nodes are duplicated. For each graph is slightly different) can be found in Ref. [18].
pair of duplicated nodes, one additional edge with zero (2) A MWPM is determined on the auxiliary graph
weight is added linking both nodes of a pair. The two du- via exact combinatorial optimization algorithms [19]. A
plicated nodes of one pair (represented black and white perfect matching is a subset of edges M which ensures
in Fig. 2(b)) are treated differently. Considering just one that each node in the graph has exactly one incident edge
pair, if an adjacent edge has pointed to the original node, ∈ M . There are several subsets that fulfill this condition.
this edge will be linked to one of the duplicated nodes (in The MWPM is that perfect matching which has the low-
Fig. 2(b) this node is the black one). On the other hand, est total weight. For the given example, the MWPM is
if the edge has pointed away from the original node, it illustrated in Fig. 2(c). Edges that belong to the MWPM
will be connected to the other duplicated node (in Fig. are represented bold and black.
 4

(3) After determining the MWPM, the original graph (parallel and perpendicular to the natural diagonal ori-
can be reconstructed. If and only if the edge that links entation), there are two different correlation lengths that
two additional nodes (in Fig. 2(c) these nodes are illus- have a different asymptotic behavior
trated as squares) does belong to the MWPM, the cor-
responding edge in the original graph is not part of the ξk ∼ |ρ − ρc |−νk , ξ⊥ ∼ |ρ − ρc |−ν⊥ (3)
optimal loop/path configuration. If, on the other hand,
the two additional nodes are not matched to each other, in the thermodynamic limit, with νk and ν⊥ being the
by the definition of the MWPM, they have to be matched critical exponents describing the power-law divergence of
to duplicated nodes, respectively. In this case the corre- the correlation lengths, respectively. At the critical point,
sponding edge of the original graph is part of a loop. In their finite-size scaling is assumed to be [20]
this way the complete optimal loop/path configuration
can be determined. In the presented example (cf. Fig. ξk ∼ Lθk
2(d)) the optimal configuration consists of one loop with ξ⊥ ∼ Lθ⊥ . (4)
total weight −2.
As the algorithm has been presented, it is not possible For anisotropic percolation models a phenomenological
to find a path that is pinned in the bottom right corner. finite-size scaling theory is introduced in Ref. [20]. It is
In order to enable such a path, the auxiliary graph must expected that cluster related quantities y(L, ρ) can be
be expanded. After constructing the auxiliary graph as rescaled according to
described above, the white duplicate of the original node
in the bottom right corner gets connected via a path y(L, ρ) = L−b θk /νk f [(ρ − ρc )Lθk /νk ]
consisting of three edges (all carry zero weight) and two = L−b θ⊥ /ν⊥ f [(ρ − ρc )Lθ⊥ /ν⊥ ], (5)
nodes to the black duplicates of all other original nodes.
This means technically the path is also a loop, but the where f [·] is an unknown scaling function and b rep-
“returning” part of the loop is “hidden” with respect resents a dimensionless critical exponent that describes
to the original lattice, such that it appears as a path the asymptotic behavior of y(L, ρ) in the thermody-
there. Such an auxiliary graph is not planar and contains namic limit. According to Eq. 5, if ρc , θk /νk and b
many additional edges, therefore, we do not depict this
are chosen properly, all data points of y(L, ρ)Lb θk /νk
additional specification in the illustration Fig. 2.
have to lie on one single curve. Therefore, y(L, ρ) can
be measured numerically for different values of L and ρ
and, subsequently, y(L, ρ)Lb θk /νk can be plotted against
III. RESULTS (ρ − ρc )Lθk /νk . Then, the unknown constants ρc , θk /νk
and b can be adjusted until the data “collapses” to one
The NWP model exhibits a geometrical continuous curve indicating that the correct values of the constants
phase transition. For a small amount of negative weights, are found. The same also applies for θ⊥ /ν⊥ instead of
the path would appear rather short and loops would not θk /νk . Note, that Eq. 5 shows the scaling behavior of
appear at all, if the system size were chosen sufficiently systems that are sufficiently large only [21]. All data col-
large. This can be seen in Fig. 1. For small values of ρ, lapses in this article are made with a computer-assisted
the formation of loops is suppressed, because each pos- scaling analysis [22].
sible loop has length O(L) and thus would collect too This data collapse approach allows only to determine
many positively weighted edges. This is clearly different the ratios θ⊥ /ν⊥ and θk /νk . In order to find an estimate
in the undirected variant of the model, where also small for νk and ν⊥ , we additionally determine θk and θ⊥ di-
loops will appear, even if ρ is small [6]. On the other rectly by applying Eqs. 4. For that reason, the path is
hand, if ρ is large, the path might grow very long and forced on the lattice, because, as evident from Fig. 3(a),
even multiple loops will occur. the correlation lengths can be estimated by taking mea-
The two regions, in which lattice-spanning, i.e., perco- surements of the path. The measurements are taken at
lating, loops or paths will or, respectively, will not occur the estimated critical point ρc = 0.3789 that has been
with high probability, are separated by a certain value found with the data collapse technique described above
of ρ = ρc (L), the critical point. In the thermodynamic and will be presented below. For the ease of presen-
limit, i.e., L → ∞, there are no lattice-spanning objects tation, we do not have to deal with the ratios θ/ν, we
in the lattice, if ρ < ρc = ρc (∞). On the other hand, have switched the order here and show the determina-
if ρ > ρc , there will appear some percolating objects al- tion of θk and θ⊥ first. Fig. 3(b) shows that a very clean
ways. power law behavior is visible, leading to θk = 0.83(2) and
In this section we determine the critical point and es- θ⊥ = 0.53(2).
timate the critical exponents that characterize the phase To actually determine the critical point and obtain the
transition via a finite-size scaling analysis. Note that a other critical exponents, we have monitored several ob-
common scaling assumption [3] that is typically used for servables in the vicinity of the expected value of the
undirected models cannot be applied here. Therefore, critical point (p ∈ [0.377, 0.382]) for different system
due to the anisotropic nature of the underlying lattice sizes. Since we could use fast optimization algorithms,
 5

(a) (b)
1000
θ|| ζ = 0.53(2) 1
θ|| ω = 0.26(2)
θ|| = 0.83(2) 0.9 0.8

〈N〉
100 θ⊥ = 0.53(2) ξ||
0.6
0.7 0.4
ξ⊥

〈N〉
10
xend 0.378 0.38
0.5
ρ
ρc = 0.3789(2) L = 256
σE L = 362
1 ν|| = 1.18(10) L = 512
0.3
100 1000 ν⊥ = 0.75(6) L = 724
L
-0.1 0 0.1 0.2 0.3
FIG. 3: (a) Sketch of ξk , ξ⊥ and xend . (b) Plot shows ξk (ρ-ρc)Lθ|| / ν||
(red. χ2 = 1.1), ξ⊥ (red. χ2 = 1.5), xend (red. χ2 = 2.1) and
σE (red. χ2 = 3.8) as a function of L. Merely system sizes
from L = 181 to 724 have been considered for the power- FIG. 5: Average number of spanning loops hN i as a function
law regression curves. The measurements are taken at the of ρ in the vicinity of the critical point (inset). The data is
estimated value of the critical point ρc = 0.3789. collapsed to one curve by using the scaling assumption Eq. 5
(main plot).

L = 256
0.6 L = 362 to standard percolation, more than one object can be
L = 512 spanning. By using again the data-collapse approach,
L = 724 we have found ρc = 0.3789(2), νk = 1.18(10) and ν⊥ =
0.5 0.75(6) with quality S = 2.3.
ρc = 0.3791(2)
Pperc(ρ)

ν|| = 1.17(14) 0.5 Another quantity that has been under scrutiny is the
0.4 ν⊥ = 0.75(9) order parameter
0.4
Pperc(ρ)

hli
0.3 Pnode ≡ , (6)
0.3 Ld
0.377 ρ 0.379 0.381 which is the probability that an edge belongs to either
0.2
-0.1 0 0.1 0.2 0.3 a percolating loop or percolating path. The total num-
(ρ-ρc)Lθ|| / ν|| ber of all edges that belong to the percolating objects
is given by l. d = 2 signifies the dimension of the lat-
tice. The asymptotic behavior of the order parameter is
FIG. 4: Percolation probability Pperc (ρ) as a function of ρ in governed by an additional critical exponent β, the per-
the vicinity of the critical point (inset). The data is collapsed colation strength [23]. As evident from Fig. 6, we have
to one curve by using the scaling assumption Eq. 5 (main found ρc = 0.3788(2), νk = 1.18(18), ν⊥ = 0.75(11) and
plot). β = 1.42(21) with quality S = 1.3. It should be noted
that several combinations of ρc and the exponents pro-
vide valid data collapses. Therefore, we considered Pnode
we could study rather large system sizes in the range versus L at the critical point (plot not shown here), which
L = 256 to L = 724 with good statistics: The data have exhibits only one fitting parameter. We found for large
been obtained by averaging over 20000 (L = 256), 16000 system sizes a power law behavior, which is compatible
(L = 362), 10000 (L = 512) and 8000 (L = 724) realiza- with β = 1.42(21), which we therefore take as final esti-
tions of the disorder, respectively. mate.
Fig. 4 shows the percolation probability Pperc (ρ) as Next, we consider the associated finite-size susceptibil-
a function of the disorder parameter ρ as well as the ity
rescaled data collapse. Since the percolation probabil-
ity is a dimensionless quantity, b = 0 is set in Eq. χL = L−d (hl2 i − hli2 ), (7)
5. The estimates ρc = 0.3791(2), νk = 1.17(14) and
ν⊥ = 0.75(9) provide the best data collapse with quality whose asymptotic behavior is guided by the critical ex-
S = 1.2, which denotes the mean-square distance of the ponent γ. As can be seen from Fig. 7, the best data
data points to the unknown scaling function in units of collapse is provided by ρc = 0.3789(3), νk = 1.18(26),
the standard error [22]. ν⊥ = 0.76(17) and γ = 0.00(5) with quality S = 0.8.
We have also measured the average number of lattice- Right at the critical point, we studied the distribution
spanning objects hN i, see Fig. 5. Note that, in contrast of path-lengths excluding the lattice-spanning ones. As
 6

0.1
1.55 L = 256
L = 362
1.4 L = 512
L = 724
Lβ θ|| / ν|| Pnode(ρ)

1.2
ρc = 0.3788(2) nl ~ l -0.780(2)
1 β = 1.42(21) 0.01
ν|| = 1.18(18)

nl
3*10-3
0.8 ν⊥ = 0.75(11)
2*10-3

Pnode(ρ)
0.6 1*10-3
0.001
0.4 0.377 ρ 0.379 0.381 L = 256
-0.1 0 0.1 0.2 0.3
1 10 100
(ρ-ρc)Lθ|| / ν||
l

FIG. 6: Order parameter Pnode (ρ) as a function of ρ in the FIG. 8: Distribution of the path lengths l at the critical point
vicinity of the critical point (inset). The data is collapsed to excluding those which percolate. 1200000 realizations of the
one curve by using the scaling assumption Eq. 5 (main plot). disorder have been considered. For the fit (red. χ2 = 1.0)
path lengths from l = 2 to 100 have been taken into account
only.

3 3 ρc = 0.3789(3) ζ ◦ is the roughness exponent defined by D ∼ tζ . D de-

◦

γ = 0.00(5)
χL(ρ)

2
scribes the mean distance between the base center and
2.5
L-γ θ|| / ν|| χL(ρ)

the endpoint of the path and t is the size of the triangle.

1 Furthermore, it is also shown in Ref. [24] that ω ◦ = 1/3,
◦
2 which is defined by σE ∼ tω . σE denotes the standard
0.378 0.38 0.382 deviation of the weight of the optimal path. A relation
1.5 ρ L = 256 between these two exponents is given by the scaling re-
L = 362
L = 512 lation ω ◦ = 2ζ ◦ − 1 [24]. There are some differences
1 ν|| = 1.18(26) L = 724
ν⊥ = 0.76(17) between the optimal path in DPRM and the path that
appears in directed NWP. First of all, the directed NWP
-0.1 0 0.1 0.2 includes the disorder parameter ρ which allows us to in-
(ρ-ρc)Lθ|| / ν|| vestigate a percolation transition, which is completely
absent for DPRM, since all paths are system spanning
by construction. There are also smaller technical differ-
FIG. 7: Fluctuations of the order parameter χL (ρ) over ρ in ences: For the directed NWP, one looks for an optimal
the vicinity of the critical point (inset). The data is collapsed configuration of loops plus one path in NWP. The loops,
to one curve by using the scaling assumption Eq. 5 (main
which cannot be crossed by the path, have to be nega-
plot).
tively weighted as well and, therefore, block several neg-
atively weighted edges that cannot be picked up by the
path. Thus, this path can not be considered as optimal
evident from Fig. 8, the distribution is in good agreement
on its own. Furthermore, while the lengths of the paths
with a power law decay nl ∼ l−τ with τ = 0.780(2). in DPRM are always equal, the lengths of the paths in
NWP is defined as a global optimization problem (cf. NWP differ considerably.
Sec. II) to find the minimally weighted configuration con-
sisting of loops plus one path. Since in the directed poly- Nevertheless, in spite of one big and the two smaller
mer problem [14] minimally weighted paths are also se- differences, directed NWP and DPRM exhibit some scal-
lected by global optimization in a random media, these ing which is comparable. This is not unnatural, since
two models might be related. The DPRM can be de- both models describe some optimal paths in disordered
scribed as follows: A weighted square lattice, in which lattices: In order to compare both models, we iden-
all edges carry a positive weight, gets cut along its di- tify D ↔ xend , where xend is the distance between the
agonal and then it is oriented as a triangle, whose right endpoint of the path and the line of predominant di-
angle is up. Then all edges become directed and point rection (cf. Fig. 3(a)) and t ↔ ξk . Then we consider
either to bottom right or bottom left. On such a lattice, xend ∼ ξkζ ∼ Lθk ζ and σE ∼ ξkω ∼ Lθk ω for the NWP
for a given realization of the disorder, one looks for the model. As evident from Fig. 3(b), xend scales with
minimally weighted path that goes from the apex to the θk ζ = 0.53(2) and σE with θk ω = 0.26(2). Consequently,
base. It has been shown in Ref. [24] that ζ ◦ = 2/3, where ω = 0.31(3) and ζ = 0.64(4), which are in good agree-
 7

ment with the exponents of the DPRM model. the average number hN i of percolating loops). These val-
Note that in our model the path is included to deter- ues are compatible with the estimates from the scaling
mine its geometrical properties, in particular its exten- of other quantities. Finally, we tested the scaling rela-
sion parallel and perpendicular to the preferred (diag- tion 2β = νk + ν⊥ − γ [25], which is a standard relation
onal) lattice direction. Nevertheless, we have also per- for directed percolation. For the left side we get 2.84(42)
formed simulations for the model without a path, just to while for the right side we get 1.93(21). Thus within one-
study the percolation properties of loops alone. All re- sigma, the scaling relation is not fulfilled, while within
sults (for somehow smaller system sizes, not shown here) two-sigma, the left and right side are compatible. Thus,
for the percolation properties remain the same within er- it is presently not fully clear whether the scaling relation
ror bars. is fulfilled. If not, it could be due to the fact that the per-
colating objects are line-like rather than bulk-like. Note
that for standard directed percolation near a wall, for the
IV. SUMMARY results obtained using a series expansion the scaling rela-
tion is clearly violated [26]. Nevertheless, in the case of a
In this work with have studied the directed variant of violation it would be different from the undirected NWP
the negative-weight percolation model. This model de- case, where the standard scaling relations for percolation
fined as a global optimization problem. The model can hold [6, 8].
be studied numerically efficiently, since a mapping to the Additionally, we have shown that the directed
minimum-weight perfect matching problem exist, such negative-weight percolation model is related to directed
that fast polynomial-time optimization algorithms can polymers in random media (DPRM), although the
be applied. Thus, large systems can be studied numer- DPRM does not exhibit a percolation transition (except
ically with good statistics giving rise to high-quality re- when diluting the system where just the standard perco-
sults. The model exhibits a continuous phase transition, lation transition appears.)
that is characterized by the appearance of loops and a
path where at least one of them is large, i.e., system-
spanning. We have studied this percolation transition by Acknowledgments
extended numerical simulations and their analysis based
on a finite-size scaling method. By investigating sev- Financial support was obtained via the Lower Saxony
eral cluster-related observables we found estimates for research network “Smart Nord” which acknowledges the
the percolation threshold, which we summarized here as support of the Lower Saxony Ministry of Science and Cul-
ρc = 0.3789(3), several critical exponents νk = 1.18(10), ture through the “Niedersächsisches Vorab” grant pro-
ν⊥ = 0.75(6), β = 1.42(21), γ = 0.00(5) and an exponent gram (grant ZN 2764/ ZN 2896). The simulations were
that describes the power-law decay of the path-length dis- performed at the HERO cluster for scientific computing
tribution τ = 0.780(2). For the values of the correlation of the University of Oldenburg jointly funded by the DFG
lengths, we have taken the estimates which yielded the (INST 184/108-1 FUGG) and the ministry of Science and
smallest statistical error bars (from the data collapse of Culture (MWK) of the Lower Saxony State.

[1] H. E. Stanley, Rev. Mod. Phys. 71, 358 (1999). Rev. E 87, 032142 (2013).
[2] D. Stauffer, Physics Reports 54, 1 (1979). [13] O. Melchert, C. Norrenbrock, and A. K. Hartmann,
[3] D. Stauffer and A. Aharony, Introduction to Percolation Physics Procedia 57, 58 (2014).
Theory (Taylor & Francis, 1992). [14] M. Kardar and Y.-C. Zhang, Phys. Rev. Lett. 58, 2087
[4] C. M. Fortuin and P. W. Kasteleyn, Physica 57, 536 (1987).
(1972). [15] O. Melchert, PhD thesis (not published, 2009).
[5] H. Hinrichsen, Adv. Phys. 49, 815 (2000). [16] W. Cook and A. Rohe, INFORMS J. Computing 11, 138
[6] O. Melchert and A. K. Hartmann, New. J. Phys. 10, (1999).
043039 (2008). [17] A. K. Hartmann and H. Rieger, Optimization Algorithms
[7] L. Apolo, O. Melchert, and A. K. Hartmann, Phys. Rev. in Physics (Wiley-VCH, Weinheim, 2001).
E 79, 031103 (2009). [18] O. Melchert and A. K. Hartmann, Phys. Rev. B 76,
[8] O. Melchert, L. Apolo, and A. K. Hartmann, Phys. Rev. 174411 (2007).
E 81, 051108 (2010). [19] For the calculation of minimum-weighted per-
[9] O. Melchert and A. K. Hartmann, Eur. Phys. J. B 80, fect matchings we use Cook and Rohes blos-
155 (2011). som4 extension to the Concorde library., URL
[10] O. Melchert, A. K. Hartmann, and M. Mézard, Phys. http://www2.isye.gatech.edu/~ wcook/blossom4/.
Rev. E 84, 041106 (2011). [20] S. Sinha and S. B. Santra, Preprint: arXiv:0807.2300v1
[11] G. Claussen, O. Melchert, and A. K. Hartmann, Phys. (2008).
Rev. E 86, 056708 (2012). [21] K. Binder and D. W. Herrmann, Monte Carlo Simulation
[12] C. Norrenbrock, O. Melchert, and A. K. Hartmann, Phys. in Statistical Physics: An Introduction (Springer, Berlin,
 8

2002). [25] J. W. Essam, A. J. Guttmann, and D. K., J. Phys. A 21,

[22] O. Melchert, Preprint: arXiv:0910.5403v1 (2009). 3815 (1988).
[23] D. Stauffer and A. Aharony, Introduction to Percolation [26] I. J. J. W. Essam, A. J. Guttmann and D. TanlaKishani,
Theory (Taylor and Francis, London, 1994). J. Physics A 29, 1619 (1996).
[24] D. A. Huse and C. Henley, Phys. Rev. Lett. 54, 2708
(1985).