Theoretical Computer Science 412 (2011) 5205–5210

Contents lists available at ScienceDirect

Theoretical Computer Science

journal homepage: www.elsevier.com/locate/tcs

Shortest paths between shortest paths✩

Marcin Kamiński a,1 , Paul Medvedev b,∗ , Martin Milanič c
a
Département d’Informatique, Université Libre de Bruxelles, Brussels, Belgium
b
Department of Computer Science, University of Toronto, Toronto, Canada
c
FAMNIT and PINT, University of Primorska, Koper, Slovenia

article info abstract

Article history: We study the following problem on reconfiguring shortest paths in graphs: Given two
Received 26 August 2010 shortest s–t paths, what is the minimum number of steps required to transform one into the
Received in revised form 7 February 2011 other, where each intermediate path must also be a shortest s–t path and must differ from
Accepted 11 May 2011
the previous one by only one vertex. We prove that the shortest reconfiguration sequence
Communicated by G. Italiano
can be exponential in the size of the graph and that it is NP-hard to compute the shortest
Keywords: reconfiguration sequence even when we know that the sequence has polynomial length.
Reconfiguration © 2011 Elsevier B.V. All rights reserved.
Shortest path
Reconfigurability
NP-hard
Diameter

1. Introduction

One of the biggest impacts of algorithmic graph theory has been its usefulness in modeling real-world problems, where
the domain of the problem is modeled as a graph and the constraints on the solution define feasible solutions. For example,
consider the problem of routing a certain commodity between two nodes in a transportation network, using as few hops
as possible. The transportation network can be modeled as a graph, each route can be modeled as a path, and the feasible
solutions are all the shortest paths between the two nodes. Traditionally, the real-world user first defines a problem instance
and then uses an algorithm to find a feasible solution which she then ‘‘implements’’ in the real world. However, some real-
world situations do not follow this simple paradigm and are more dynamic because they allow the solution to ‘‘evolve’’ over
time. For example, consider the situation where the commodity is already being transferred along a shortest route, but the
operator has been instructed to use a different route, which is also a shortest path. She can physically switch the route only
one node at a time, but does not wish to interrupt the transfer. Thus, she would like to switch between the two routes in as
few steps as possible, while maintaining a shortest path route at every intermediate step.
In general, this type of situation gives rise to a reconfiguration framework, where we consider an algorithmic problem P
and a way of transforming one feasible solution of an instance I of P to another (reconfiguration rule). Given two feasible
solutions s1 , sk of I, we want to find a reconfiguration sequence s1 , . . . , sk such that each si (1 ≤ i ≤ k) is a feasible solution of I,
and the transition between si and si+1 is allowed by the reconfiguration rule. An alternate definition is via the reconfiguration
graph, where the vertices are the feasible solutions of I, and two solutions are adjacent if and only if one can be obtained from
the other by the reconfiguration rule. The reconfiguration sequence is then a path between s1 and sk in the reconfiguration

✩ A version of this work has appeared in the Proceedings of IWOCA 2010, 21st International Workshop on Combinatorial Algorithms, London, 26–28 July
2010 (Marcin Kamiński et al., 2011) [10].
∗ Corresponding author. Tel.: +1 416 946 3924.
E-mail addresses: [email protected] (M. Kamiński), [email protected] (P. Medvedev), [email protected] (M. Milanič).
1 Chargé de Recherches du FRS-FNRS.

0304-3975/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.tcs.2011.05.021
5206 M. Kamiński et al. / Theoretical Computer Science 412 (2011) 5205–5210

graph. We can then ask for the shortest reconfiguration sequence, or, in the reconfigurability problem, to simply check if the
two solutions are reconfigurable (i.e., if such a sequence exists).
The reconfiguration framework has recently been applied in a number of settings, including vertex coloring [4,5,3,2],
list-edge coloring [9], clique, set cover, integer programming, matching, spanning tree, matroid bases [8], block puzzles [7],
independent set [7,8,10], and satisfiability [6]. In the well-studied vertex coloring problem, for example, we are given two
k-colorings of a graph, and the reconfiguration rule allows us to change the color of a single vertex. In a different example, we
are given two independent sets, which we imagine to be two sets of tokens placed on the vertices, and the reconfiguration
rule is to slide a single token along an edge.
A topic certainly related to graph reconfiguration problems is the reconfiguration of Boolean formulas studied in [6].
The authors define a class of Boolean formulas that can be built from tight relations and prove three interesting dichotomy
theorems. First, they consider the problem of checking if two assignments of a Boolean formula are connected. They show
that for the formulas built from tight relations the question can be answered in linear time and is PSPACE-complete for the
formulas not in that class. Then, they study the problem of determining if the reconfiguration graph of a Boolean formula
is connected. The set of formulas that can be built from tight relations gives instances that are in coNP; the problem is
PSPACE-complete for the formulas not in that class. Finally, they focus on the diameter of connected components of the
reconfiguration graph. They show that the diameter is linear for the formulas that can be built from tight relations; and can
be exponential otherwise.
Though the complexities of each of the many reconfiguration problems may each be studied independently, a
fundamental question is whether there exists any systematic relationship between the complexity of the original problem
and that of its reconfigurability problem. To this end, current studies have revealed a pattern where most ‘‘natural’’ problems
in P have their reconfigurability problems in P as well, while problems whose reconfigurability versions are at least NP-hard
are NP-complete. For example, spanning tree, matching, and matroid problems in general (all in P) lead to polynomially
solvable reconfigurability problems when using the most natural reconfiguration rule, while the reconfigurability of
independent set, set cover, and integer programming (all NP-complete) are PSPACE-complete [8].
Ito et al. [8] have conjectured that this relationship is not true in general, and that there exist problems in P which give
rise, in a natural way, to NP-hard reconfigurability problems. Indeed, the problem of deciding whether two k-colorings are
reconfigurable is PSPACE-complete for (i) bipartite graphs and k ≥ 4, and (ii) planar graphs, for 4 ≤ k ≤ 6 [2]. Clearly,
4-coloring of bipartite or planar graphs is in P. However, these are not ‘‘natural’’ problems in the sense that the colorings
used in the PSPACE-hardness proof constructions are not optimal. It is interesting to ask if there exists a ‘‘natural’’ problem
in P whose reconfiguration version is NP-hard.
Another systematic relationship that has been pursued is between the complexity of a reconfigurability problem and the
diameter of the reconfiguration graph. When the diameter is polynomial, a reconfiguration sequence is a trivial certificate for
the reconfigurability of two instances, guaranteeing that the problem is in NP. However, current evidence further suggests
that for reconfigurability problems that are solvable in polynomial time, the diameter is also polynomial. In the study of
k-coloring, it was found that for k ≤ 3, the reconfigurability problem is solvable in polynomial time and the diameter of
the reconfiguration graph is at most quadratic in the number of vertices of the colored graph. For satisfiability, the formulas
built from tight relations (whose reconfigurability is polynomial) lead to reconfiguration graphs with linear diameter [6].
We are not aware of any natural problems with the property that the diameter can be exponential while reconfigurability
can be decided in polynomial time2 ; however, such an example, if found, would indicate that the diameter cannot serve as
a reliable indicator of the reconfigurability complexity.
In this paper, we introduce the reconfiguration version of the shortest path problem, which arises naturally, such as in the
routing example above. We show (in Section 2) that the reconfiguration graph can have exponential diameter, implying that
the shortest path reconfiguration problem probably breaks one of the two established patterns described above. On the one
hand, if reconfigurability of shortest paths can be decided in polynomial time, then it is the first example of a reconfigurability
problem in P with exponential diameter. On the other hand, if it is NP-hard, it is the first example of a ‘‘natural’’ problem in P
whose reconfigurability version is NP-hard. For these reasons, we believe that shortest path reconfiguration is an important
problem to study, not only for its practical application but also for our understanding of the systematic relationship between
the hardness of a problem, the diameter of its reconfiguration graph, and the hardness of its reconfigurability problem.
Towards this end, we give (in Section 3) a reduction from SAT to show that it is NP-hard to find the shortest reconfiguration
sequence between two shortest paths.

2. Instances with exponential diameter

We define the reconfiguration rule for shortest paths in the natural way: two shortest (s, t )-paths are adjacent in the
reconfiguration graph of shortest (s, t )-paths if and only if they differ, as sequences, in exactly one vertex.
We now present a family of graphs Gk whose size is linear in k but the diameter of the reconfiguration graph is Ω (2k ). The
graph G1 contains vertices {x1i | 1 ≤ i ≤ 7}∪{y1i | 1 ≤ i ≤ 6}∪{s, t } and edges {(x1i , y1i ), (x1i+1 , y1i ), (y1i , t ) | i ≤ 6}∪{(s, x1i ) |

2 For a very artificial one, consider the problem in which instances are n-bit words and two instances are adjacent if they differ by 1 modulo 2n . The
diameter of the reconfiguration graph is 2n−1 but all pairs of instances are reconfigurable.
 M. Kamiński et al. / Theoretical Computer Science 412 (2011) 5205–5210 5207

Fig. 1. The graph Gk for k = 4, where the reconfiguration distance between pkb = s, xk1 , yk1 , . . . , x11 , y11 , t and pke = s, xk7 , yk6 , xk1−1 , xk1−1 , . . . , x11 , y11 , t is Θ (2k ).
An edge with a circle end means that the vertex is connected to all the vertices in the next layer.

1 ≤ i ≤ 7}. The graph Gk is defined recursively with vertices {xki | 1 ≤ i ≤ 7} ∪ {yki | 1 ≤ i ≤ 6} ∪ V (Gk−1 ) and the edges
{(xki , yki ), (xki+1 , yki ) | i ≤ 6} ∪ {(yki , xkj −1 ) | i ∈ {1, 3, 5}, j ≤ 7} ∪ {(yk2 , x1k−1 ), (yk4 , xk7−1 ), (yk6 , xk1−1 )} ∪ E (Gk−1 \ {s}) ∪ {(s, xki ) |
1 ≤ i ≤ 7} (see Fig. 1). Let pkb = s, xk1 , yk1 , . . . , x11 , y11 , t, and let pke = s, xk7 , yk6 , xk1−1 , xk1−1 , . . . , x11 , y11 , t. We will consider the
problem of reconfiguring pkb to pke in Gk .

Lemma 1. Let p be a shortest path in Gk that goes through yk1 , and let q be a path that goes through yk6 . Then the reconfiguration
distance between p and q is at least 9(2k − 1).
Proof. We prove by induction on k, where the base case is clear. Let ρ = p1 , . . . , pn be the shortest reconfiguration sequence
between p and q. First, let i′ be the smallest integer such that pi′ +1 contains yk4 , and let i ≤ i′ be the smallest integer such
that every path pi , . . . , pi′ contains yk3 . By construction, we know that pi−1 , and hence pi , contains yk1−1 and pi′ +1 , and hence
pi′ , contains y6k−1 . Hence, by the induction hypothesis, the length of this first phase, i′ − i + 1, is at least 9(2k−1 − 1).
Next, let j′ be the smallest integer such that pj′ +1 contains yk6 , and let j ≤ j′ be the smallest integer such that every path
pj , . . . , pj′ contains yk5 . By construction, we know that pj−1 , and hence pj , contains y6k−1 and pj′ +1 , and hence pj′ , contains yk1−1 .
Hence, by the induction hypothesis, the length of this second phase, j′ − j + 1, is at least 9(2k−1 − 1).
Observe from the graph construction that ρ must always visit ykx−1 before visiting ykx , hence i′ < j, and so the length of
ρ is at least the sum of the two phases plus the moves of the first and second vertex of the path necessary to percolate yk1
down to yk6 , proving the lemma. 
On the other hand, there exists an asymptotically matching lower bound:
Lemma 2. The reconfiguration distance between pkb and pke is at most 11(2k − 1).
Proof. We prove by induction on k, where the base case is clear. It will be helpful to formally treat a reconfiguration sequence
not as a sequence of paths but as a sequence of vertices, each one representing the switched vertex at that step. Applying
the induction hypothesis, let ρ be the shortest reconfiguration sequence in Gk−1 , and let rev(ρ) be that sequence in the
reverse direction (from pke−1 to pkb−1 ). We construct the sequence as ρ ′ = xk2 , yk2 , xk3 , yk3 , ρ, xk4 , yk4 , xk5 , yk5 , rev(ρ), xk6 , yk6 , xk7 .
This sequence of moves reconfigures pkb into pke with the number of steps satisfying the lemma. 
We therefore have the following theorem:
Theorem 1. The reconfiguration distance in Gk between pkb and pke is in Θ (2k ).

3. NP-hardness of Min-SPR

Given (G, s, t , pb , pe , k), where pb and pe are shortest (s, t )-paths and k is an integer, the Min-SPR problem is to determine
whether there is a reconfiguration sequence between pb and pe of length at most k. Let φ be a Boolean formula in conjunctive
normal form with variables x1 , . . . , xn and clauses C1 , . . . , Cm . We will create an instance (Gφ , s, t , pb , pe , 2m(n + 2)) and
show that φ is satisfiable if and only if this instance is in Min-SPR. For ease of presentation, the graph Gφ will be directed.
However, our result holds for undirected graphs because the directed shortest (s, t )-paths in Gφ are exactly the shortest
paths in the underlying undirected graph of Gφ .
For every variable xℓ and its possible value v s ∈ {0, 1}, we build a gadget G(ℓ, v s). The vertex set is {v(ℓ, v s, cs, d) | cs ∈
{0, 1}, 1 ≤ d ≤ 2m}. The values ℓ, v s, cs, and d for a vertex are referred to as its level, v-state (short for variable state), c-state
(short for clause state), and depth, and denoted by ℓ(v), v s(v), cs(v), and d(v), respectively. For every 1 ≤ d ≤ 2m − 1, and
every cs, there is an edge from v(ℓ, v s, cs, d) to v(ℓ, v s, cs, d + 1). For all 1 ≤ d ≤ m − 1, there is an edge from v(ℓ, v s, 0, 2d)
to v(ℓ, v s, 1, 2d + 1), and from v(ℓ, v s, 1, 2d) to v(ℓ, v s, 0, 2d + 1). We also add edges, called formula edges, that are formula
5208 M. Kamiński et al. / Theoretical Computer Science 412 (2011) 5205–5210

A C
c-state = 0

c-state = 1

Fig. 2. The reduction from a formula φ to a graph Gφ for the case of three clauses and three variables. Panel (A) shows the internal connections of a gadget,
with the potential formula edges that depend on φ given in red (dashed). Panel (B) shows the way we connect two given gadgets, while (C) shows the
structure of the whole graph. Each of the rectangles represents a gadget, with the lines showing which gadgets are connected together.

dependent. For all d, if xℓ = v s satisfies Cd , we add an edge from v(ℓ, v s, 1, 2d − 1) to v(ℓ, v s, 0, 2d). This gadget is shown
in Fig. 2A.
We now connect some of these gadgets together. The gadgets we connect are G(ℓ, v s) to G(ℓ+ 1, 0) and to G(ℓ+ 1, 1), for
all ℓ ≤ n − 1 and all v s. Given two gadgets, G(ℓ, v s) and G(ℓ + 1, v s′ ), the meaning of connecting G(ℓ, v s) to G(ℓ + 1, v s′ ) is
given as follows (shown in Fig. 2BC). For all d ≤ 2m − 1 and cs, there is an edge from v(ℓ + 1, v s′ , cs, d − 1) to v(ℓ, v s, cs, d).
Also, for all d ≤ m − 1 and cs, there is an edge from v(ℓ + 1, v s′ , cs, 2d) to v(ℓ, v s, 1 − cs, 2d + 1).
We next add a begin and end gadget to the graph, consisting of vertices begd and endd , respectively, for 1 ≤ d ≤ 2m.
These are connected in a path, with edges (begd , begd+1 ) and (endd , endd+1 ) for d ≤ 2m − 1. The level of the vertices in the
begin (end) gadget is 0 (n + 1), the c-state is 0 (1), and the depth of begd or endd is d. For all v s, d ≤ 2m − 1, there is an edge
from v(1, v s, 0, d) to begd+1 , and from endd to v(n, v s, 1, d + 1).
Finally, we add vertices s and t to the graph, and make an edge from s to every depth 1 vertex, and from every depth 2m
vertex to t. The depth of s (t) is defined to be 0 (2m + 1). We call the resulting directed graph Gφ . Let pb = s, beg1 , . . . , beg2m , t
and pe = s, end1 , . . . , end2m , t be two paths in this graph. Then, (Gφ , s, t , pb , pe , 2m(n + 2)) is the instance of the Min-SPR
problem that we will consider here.
The intuition behind the reduction is that in order for the path to percolate down from pb to pe in a minimal number
of steps, it must pass consecutively through exactly one of G(ℓ, 0) or G(ℓ, 1) for every variable xℓ . The choice of which one
corresponds to assigning xℓ the corresponding value. Furthermore, each shortest path that goes through a gadget can visit
the vertex at depth 2d with a c-state of 0 or 1. This corresponds to having the dth clause satisfied or not. Initially, the path
goes only through vertices with c-state 0, and the only way to switch the c-state at a given depth is via a formula edge. By
going through a gadget G(ℓ, v s), there is an opportunity to use the formula edges to switch the c-state of all clauses that
xℓ = v s would satisfy. In order to reach the final path pe , the c-state of all the vertices in the path must become 1, hence all
the clauses must be satisfied.
Each edge (a, b) is considered to be either inter-clause or intra-clause, depending on the parity of d(a). The edge is inter-
clause if d(a) is even. We call edges that connect vertices on the same level (exactly those that belong to the same gadget) as
intra-level, while the edges that connect vertices on different levels are called inter-level. The following facts about Gφ follow
from definitions and capture most of the properties of the reduction that are needed to prove completeness and soundness.
Fact 1. Let e = (a, b) be an edge in Gφ . The statements below follow directly from the construction:
1. ℓ(b) ≤ ℓ(a) ≤ ℓ(b) + 1.
2. If e is an intra-level intra-clause edge, cs(a) = 0 implies that cs(b) = 0.
3. If e is a non-formula intra-clause edge, then cs(a) = cs(b).
4. If e is intra-level, then v s(a) = v s(b).
First, we will show that the reduction is sound. Let p = s, v1 , . . . , v2m , t be a shortest path and consider an arbitrary
move that switches vd with vd′ . The move graph is the subgraph induced by vd−1 , vd , vd′ , vd+1 , referred to by the tuple
(vd−1 , vd , vd′ , vd+1 ).
Lemma 3. The length of a reconfiguration sequence is at least 2m(n + 2). Moreover, each move in a sequence that has this length
must either increase the c-state and leave the level of the switched vertex unchanged, or increase the level by one but leave the
c-state unchanged.
 M. Kamiński et al. / Theoretical Computer Science 412 (2011) 5205–5210 5209

Proof. A single move cannot increase the level or the c-state by more than one (Fact 1.1). Moreover, we claim that it cannot
increase both of these at the same time. Let m be an arbitrary move with move graph (a, b, c , d). If m increases the level
by one, then the properties of the construction (Fact 1) applied to the move graph imply that the c-state does not increase.
Specifically, for the case that the depth of b is odd, Fact 1.1 applied to the edges (b, d) and (c , d) implies that they are intra-
and inter-level, respectively. If the c-state of b is 1 then it trivially cannot increase, but if it is zero then Fact 1.2 implies that
cs(d) = 0, and Fact 1.3 implies that cs(c ) = 0. A similar argument can be applied to the edges (a, b) and (a, c ) for the case
the depth of b is even. Thus, the level and c-state cannot both increase.
The sum of the levels and c-states in the starting path pb is 0 and in the final path pe is 2m(n + 2). Since we showed that
the sum cannot increase by more than one in a single move, a reconfiguration sequence of length at 2m(n + 2) must increase
this sum by exactly one each move, and a shorter reconfiguration sequence is not possible. 
Lemma 4. No path can contain two vertices with the same level but different v -state.
Proof. In any path, the level of the vertices is non-increasing (by Fact 1.1). Therefore, all the vertices that have the same level
must appear consecutively in the path. Since the edges connecting them are intra-level, Fact 1.4 implies that their v -states
are identical. 
We say that a reconfiguration sequence ρ visits a vertex if there exists p ∈ ρ that contains that vertex.
Lemma 5. Suppose there exists a reconfiguration sequence ρ of length 2m(n + 2). Then ρ visits at least one vertex at every level,
and all the vertices that it visits at a given level have the same v -state.
Proof. First, since the level of a switched vertex can never increase by more than one, pb has the vertices of the smallest
level and pe of the biggest level, ρ must visit at least one vertex at every level.
Next, consider all the paths in ρ that contain a level ℓ vertex, for some ℓ. We claim that these paths form a contiguous
subsequence of ρ . After ρ visits its first level ℓ vertex, it can never have a path with just lower level vertices (by Lemma 3),
so the next time it reaches a path with no level ℓ vertex, all the path’s vertices will have a higher level. After that point, ρ
can never visit a level ℓ vertex again (by Lemma 3).
Now, for the sake of contradiction, suppose that ρ visits two vertices of the same level but different v -states. Consider
the first time this happens, going from a path p to p′ via move (a, b, c , d). We know that b and c are the only level ℓ vertices
in p and p′ , respectively (by Lemma 4), and that they have different c-states (by Lemma 3). Since the levels of a and d are not
ℓ, the c-states of b and d must be the same (Fact 1.3), a contradiction. 
Suppose there exists a reconfiguration sequence ρ of length 2m(n + 2). Lemma 5 allows us to build a truth assignment
θ ∈ {0, 1}n by assigning θℓ the v -state of the vertices of level ℓ in ρ .
Lemma 6. The assignment θ is satisfying for φ .
Proof. Consider an arbitrary clause Cd , and the vertices at depth 2d − 1. Each p ∈ ρ contains exactly one vertex at this
depth. In pb , the c-state of this vertex is 0, while in pe it is 1, so there exists some first move (a, b, c , d) at depth 2d − 1 that
increases the c-state. By Lemma 3, this move cannot also change the level. Therefore, either (b, d) or (c , d) is a formula edge,
otherwise Fact 1.3 would give us the contradiction 0 = cs(b) = cs(d) = cs(c ) = 1. Since the c-state of b is 0, (b, d) cannot
be a formula edge, meaning (c , d) is a formula edge. Then we know from the construction that xl(c ) = v s(c ) satisfies Cd ,
meaning Cd is satisfied by θℓ(c ) = v s(c ). 
We now prove that the reduction is complete.
Lemma 7. If φ is satisfiable, then there exists a reconfiguration sequence of length at most 2m(n + 2).
Proof. Let θ ∈ {0, 1}n be a satisfying truth assignment. Let sat (ℓ, d) = 1 if Cd is satisfied by θ1 , . . . , θℓ , and 0 otherwise. We
define a path p(ℓ, ℓ′ ) which goes through the vertices of the gadget G(ℓ, θℓ ) with c-states at depth 2d and 2d − 1 derived
from sat (ℓ′ , d). Specifically, let p(ℓ, ℓ′ ) = s, v(ℓ, θℓ , sat (ℓ′ , 1), 1), v(ℓ, θℓ , sat (ℓ′ , 1), 2), . . ., v(ℓ, θℓ , sat (ℓ′ , d), 2d − 1),
v(ℓ, θℓ , sat (ℓ′ , d), 2d), . . . , t. We will build a reconfiguration sequence ρ that starts from pb , and then goes to p(ℓ, ℓ − 1)
and p(ℓ, ℓ) for every ℓ, finally finishing with pe .
Let us fill in the intermediate moves of ρ . The vertices of p(ℓ, ℓ) can be switched in order of increasing depth using inter-
level edges to get p(ℓ + 1, ℓ) in 2m steps. The paths p(ℓ, ℓ − 1) and p(ℓ, ℓ) are different only when Cd is satisfied by θℓ , in
which case there is a formula edge from v(ℓ, θℓ , 1, 2d − 1) to v(ℓ, θℓ , 0, 2d). Using these edges, the vertices of p(ℓ, ℓ − 1)
can be switched in order of increasing depth to get p(ℓ, ℓ).
The number of moves required to switch between p(ℓ, ℓ) and p(ℓ + 1, ℓ) is 2m. The total number of moves to switch
between p(ℓ, ℓ − 1) and p(ℓ, ℓ) is 2k, where k is the number of clauses satisfied by θℓ but not satisfied by θ1 , . . . , θℓ−1 . When
summed over ρ , these add up to at most 2m, since each clause can become satisfied for the first time only once. Finally, we
can switch between pb and p(1, 0) and between p(n, n) and pe using 2m steps each. The length of ρ is therefore 2m(n + 2). 
Combining Lemmas 6 and 7 together with the fact that the reduction can be clearly done in polynomial time, we have
the following theorem.
Theorem 2. The Min-SPR problem is NP-hard, even if k is polynomial in |V (G)|.
5210 M. Kamiński et al. / Theoretical Computer Science 412 (2011) 5205–5210

4. Concluding remarks

In this paper, we studied the reconfiguration variant of the shortest path problem. We believe that the major open
problem is to determine the complexity of deciding whether two shortest paths are reconfigurable. If the problem is
NP-hard, then it will be the first example of a ‘‘natural’’ problem in P whose reconfigurability version is NP-hard. If the
problem is polynomially solvable, then it will be the first example of an efficiently solvable reconfigurability problem with
reconfiguration graphs of large diameter.
Our results are somewhat orthogonal to previous research on reconfiguration since we consider the length of a shortest
path between two instances in the reconfiguration graph. If we assume that the bound k on the reconfiguration sequence
length is given in unary, Min-SPR is in NP and Theorem 2 says it is NP-complete. It would be interesting to analyze whether
similar results hold for other problems that have been studied in the context of reconfiguration.

Acknowledgements

We are grateful to Paul Bonsma, Takehiro Ito and Daniel Pellicer for interesting and fruitful discussions. We also thank
the referees for comments that helped to improve the paper.
The third author was supported in part by ‘‘Agencija za raziskovalno dejavnost Republike Slovenije’’, research program
P1-0285.

Addendum

While this paper was under review, Paul Bonsma proved that shortest path reconfigurability is PSPACE-complete [1].

References

[1] Paul S. Bonsma, Shortest path reconfiguration is PSPACE-hard, 2010. arXiv:1009.3217v1 [cs.CC].
[2] Paul S. Bonsma, Luis Cereceda, Finding paths between graph colourings: PSPACE-completeness and superpolynomial distances, Theoret. Comput. Sci.
410 (50) (2009) 5215–5226.
[3] Paul S. Bonsma, Luis Cereceda, Jan van den Heuvel, Matthew Johnson, Finding paths between graph colourings: computational complexity and possible
distances, Electron. Notes Discrete Math. 29 (2007) 463–469.
[4] Luis Cereceda, Jan van den Heuvel, Matthew Johnson, Connectedness of the graph of vertex-colourings, Discrete Math. 308 (5–6) (2008) 913–919.
[5] Luis Cereceda, Jan van den Heuvel, Matthew Johnson, Mixing 3-colourings in bipartite graphs, European J. Combin. 30 (2009) 1593–1606.
[6] Parikshit Gopalan, Phokion G. Kolaitis, Elitza N. Maneva, Christos H. Papadimitriou, The connectivity of Boolean satisfiability: computational and
structural dichotomies, SIAM J. Comput. 38 (6) (2009) 2330–2355.
[7] Robert A. Hearn, Erik D. Demaine, PSPACE-completeness of sliding-block puzzles and other problems through the nondeterministic constraint logic
model of computation, Theoret. Comput. Sci. 343 (1–2) (2005) 72–96.
[8] Takehiro Ito, Erik D. Demaine, Nicholas J.A. Harvey, Christos H. Papadimitriou, Martha Sideri, Ryuhei Uehara, Yushi Uno, On the complexity of
reconfiguration problems, in: ISAAC, in: Lecture Notes in Computer Science, vol. 5369, Springer, 2008, pp. 28–39.
[9] Takehiro Ito, Marcin Kamiński, Erik D. Demaine, Reconfiguration of list edge-colorings in a graph, in: WADS, in: Lecture Notes in Computer Science,
vol. 5664, Springer, 2009, pp. 375–386.
[10] Marcin Kamiński, Paul Medvedev, Martin Milanič, Shortest paths between shortest paths and independent sets, in: IWOCA, in: Lecture Notes in
Computer Science, vol. 6460, Springer, 2011, pp. 56–67.