A SPECTRAL APPROACH TO THE

SHORTEST PATH PROBLEM

STEFAN STEINERBERGER
arXiv:2004.01163v2 [math.CO] 16 Apr 2020

Abstract. Let G = (V, E) be a simple, connected graph. One is often in-

terested in a short path between two vertices u, v. We propose a spectral
algorithm: construct the function φ : V → R≥0
2
P
(w1 ,w2 )∈E (f (w1 ) − f (w2 ))
φ = arg min P 2
.
w∈V f (w)
f :V →R
f (u)=0,f 6≡0

φ can also be understood as the smallest eigenvector of the Laplacian Matrix

L = D − A after the u−th row and column have been removed. We start in
the point v and construct a path from v to u: at each step, we move to the
neighbor for which φ is the smallest. This algorithm provably terminates and
results in a short path from v to u, often the shortest. The efficiency of this
method is due to a discrete analogue of a phenomenon in Partial Differential
Equations that is not well understood. We prove optimality for trees and
discuss a number of open questions.

1. Introduction and Algorithm

1.1. Introduction. The Shortest Path Problem asks for the shortest path connect-
ing two vertices in a graph. For simplicity, we will only deal with simple, connected
graphs throughout the entire paper. The problem is one of the foundational prob-
lems in computer science and has received a lot of attention [8, 15, 23, 28, 29, 30,
38, 40, 43, 45, 72, 68, 69, 73, 76, 83, 88, 90, 89]. In the case that we are considering
(unweighted and undirected), breadth first search is known to cost O(|V | + |E|).

Figure 1. Paths taken by the Spectral Method.

Key words and phrases. Shortest Path Problem, Spectral Theory, Hot Spots.
S.S. is supported by the NSF (DMS-1763179) and the Alfred P. Sloan Foundation.
1
2

We introduce a new algorithm based on Spectral Graph Theory; the main point
is not to find a new algorithm to compute the shortest path (though it may have
interesting applications, see §2.4) but to introduce a phenomenon in Spectral Graph
Theory. The algorithm does not always result in the shortest path but it does seem
to result in surprisingly short paths (and even on fairly strange graphs it tends to
find the shortest path for a vast majority of vertices, see §3). What is interesting
about this algorithm is that the reason for its effectiveness is closely related to
some famously unsolved problems in Partial Differential Equations (see §2.4): any
type of result that one can prove about this algorithm on graphs may lead to
a better understanding of the PDE problem in the continuous setting! There is
an understanding that ‘robust’ phenomena in PDEs also occur on graphs and,
conversely, anything that one can prove on graphs one can often also prove for
PDEs, see [12, 13, 84, 93] for recent examples. One contribution of our paper is
to show that the phenomenon that causes ‘Hot Spots’ of eigenfunctions to be on
the boundary is ‘robust’ in this sense. In light of this, any progress on the discrete
problem could be of substantial interest.

1.2. The Algorithm. We now present the algorithm. We assume that G = (V, E)
is a simple, connected graph with |V | = n vertices which, for simplicity, we label
{1, . . . , n}. We will assume that D ∈ Rn×n is a diagonal matrix where dii = deg(i)
is the degree of the vertex i. We will use A to denote the (symmetric) adjacency
n×n
matrix A = (aij ) ∈ {0, 1} where aij = 1 if and only if (i, j) ∈ E.
Algorithm.
Input: a graph G = (V, E) and two vertices i, j ∈ V .
Output: a path from j to i.
(Preprocessing)
(1) Construct the matrix L = D − A and construct Li by deleting
the i−th row and the i−th column of L.
(2) Find an eigenvector φ ∈ Rn associated to the smallest positive
eigenvalue of Li and interpret it as a function φ : V \ {i} → R.
(3) If φ has negative entries, multiply it with −1. Set φ(i) = 0.
(Path Construction)
(4) Construct a path by starting in x0 = j.
(5) While xk 6= i, look at all the neighbors of xk and pick xk+1 to
be a neighbor for which φ assumes its minimum.
In practical problems the algorithm can be applied twice: once to find a path from
i to j and then to find one from j to i. We observe in practice that this trick
leads to optimal paths in many settings (see §3). We will now give some simple
heuristic motivating the algorithm. For any function f : V → R, we can consider
the quadratic form
X
hf, Lf i = (f (w1 ) − f (w2 ))2
(w1 ,w2 )∈E

and try to minimize it. Clearly, the quadratic form has a scaling symmetry under
multiplication with constants: it is thus customary to assume
X
f (v)2 = 1.
v∈V
 3

For simple, connected graphs, this minimization procedure is quite simple: clearly,
the minimum will be given by constant functions resulting in the value 0. This
is also sometimes referred to as the ‘trivial’ eigenvector of the Graph Laplacian.
Instead, what we will do is to pick a fixed i ∈ V and consider the minimization
problem over all functions f : V → R for which f (i) = 0, i.e.
X X
φ = arg min (f (w1 ) − f (w2 ))2 subject to f (v)2 = 1.
f (i)=0
(w1 ,w2 )∈E v∈V

This minimization problem cannot result in constant functions. The way it is set
up, we expect f to be small around vertices close to i and larger when it is further
away. This simple heuristic, though meaningful and easily verified in practice, is
not really mathematically understood (see §2.4.). In terms of computation, it is
easily verified that
X
f T (D − A)f = (f (w1 ) − f (w2 ))2
(w1 ,w2 )∈E

and thus minimizing the quadratic form is akin to finding the smallest eigenvectors
of D−A (which is the constant eigenvector). If we additionally enforce that f (i) = 0,
then a quick inspection on the left-hand side shows that this the same problem
after having removed the i−th row and i−th column of D − A. Since the graph is
connected, the smallest eigenvalue will be larger than 0. The smallest eigenvector φ
is unique and does not change sign (so up to multiplication with −1 we can assume
it to have nonnegative entries). We note that if removing the vertex i disconnects
the graph (as is, for example, the case for some vertices on trees), then the problem
decouples into two or more problems and can be solved one-by-one on each graph.
1.3. Numerical Remarks. One might a priori assume that minimizing a qua-
dratic form, finding the smallest eigenvalue of a matrix, is prohibitively expensive.
However, there have been recent spectacular advances on nearly-in-time solvers
of equations involving Graph Laplacians, we refer to work of Spielman & Teng
[77, 78, 79], the papers [17, 18, 48, 49, 50, 54] and the book of Vishnoi [87]. We
point out that Spielman & Teng [79] give an algorithm that computes an approx-
imation of the first nontrivial eigenfunction in O(|V
e | + |E|). Similarly expensive
techniques have been profitably used in Computer Graphics, see §2.4.

2. The Results
2.1. The Algorithm terminates. Our first result is that the algorithm provably
terminates. This requires us to prove several properties about the solution of the
minimization problem (in particular, uniqueness, that it only vanishes in one point
and that there are no local minima).
Theorem 1. The Algorithm produces a path between the vertices.
The argument follows classical arguments from the continuous setting in spirit, we
refer to Courant & Hilbert [19, 20]. One downside is that the argument does not
seem to result in quantitative estimates which would be desirable: as emphasized
above, and in other parts of the paper, this is not merely an interesting question
in terms of the algorithm proposed here. A better understanding of the algorithm
would likely to lead progress on some notoriously difficult questions in partial dif-
ferential equations (see §2.4).
4

2.2. Trees. We can prove that the algorithm produces the optimal path on graphs
that are connected trees.
Theorem 2. The Algorithm produces the shortest path on connected trees.
The argument makes use of a reflection trick that allows us to reduce the problem
to one the has been studied previously (with a recent and particularly short solution
by Lederman and the author [52]). Needless to say, it would be very nice to have
an extension of the result to broader classes of graphs.
2.3. A Continuous Analogue. In this section, we describe a continuous analogue
of the statements above (which, indeed, was motivated by the results above). We
will consider a simply connected domain Ω ⊂ R2 with smooth boundary. We assume
the boundary is comprised of two parts: a simply connected set ∂Ω0 (homeomorphic
to an interval) and the complement ∂Ωn (see Fig. 2 for a sketch of how such a
domain could look like).

∂Ωn

∂Ω0

Figure 2. Theorem 3 illustrated: the gradient flow moves towards

the boundary and impacts on the part of the boundary where the
function vanishes.

We will consider the differential inequality

∆u < 0 inside Ω
u=0 on ∂Ω0
∂u
=0 on ∂Ωn ,
∂n
where n is the normal vector on the boundary. This is, in some sense, the canonical
analogue of the structure we encounter on graphs: there is a vertex pinned down to
0 on which φ vanishes (and it vanishes nowhere else), Neumann boundary conditions
are the natural analogue to the absence of boundary conditions on the graph. A
particularly canonical way of obtaining this continuous setup is to consider the first
nontrivial Laplacian eigenfunction −∆u = µ2 u with Neumann boundary conditions
and to take a connected component of {x : u(x) > 0}. We will now study gradient
flows on this space: at each point x ∈ Ω where the gradient ∇u does not vanish, we
can define a local vector field pointing in the direction where u decays the fastest
∇u(x)
V (x) = − .
|∇u(x)|
 5

We can now, correspondingly, define a gradient flow started in a point x0 to be the

solution of the ODE γ(0) = x0 and
d
γ(t) = V (γ(t)).
dt
This is the continuous analogue of our way of constructing the path: at each step,
we add the adjacent vertex with the smallest value on it.
Theorem 3. Let Ω be a simply connected domain and u be as above. If every local
maximum of u is on the boundary and the critical points are non-degenerate, then
all the critical points are on the boundary. Moreover, the gradient flow started in
any point inside Ω converges to ∂Ω0 .
Very few results in this direction seem to be known, we refer to [2]. It could be
quite interesting to make this statement quantitative. Is there an upper bound on
the length of the gradient flow in terms of geometric quantities of Ω?

Figure 3. Paths taken by the Spectral Method.

2.4. Related Results and Problems. We first discuss some explicitly related ap-
proaches in Computer Graphics and then relate these questions to classical problem
for Partial Differential Equations.

Geodesics in Heat. To the best of our knowledge, the philosophically most re-
lated approach that we were able to find in the literature is the heat construction
of Crane, Weischedel & Wardetzky [21, 22] to obtain a natural notion of distance.
Their idea consists of solving the heat equation with an initial point source δx for a
short amount of time t to obtain et∆ δx . Varadhan’s Lemma [85, 86] suggests that
this solution behaves roughly like a Gaussian in the geodesic distance. This infor-
mation is then used by solving a second elliptic equation to obtain a normalized
distance. The method is shown to be effective in obtaining a notion of intrinsic
distance on a variety of domains and has inspired a lot of subsequent work, we
mention [9, 32, 55, 58, 75, 82, 91, 92]. Crane, Weischedel & Wardetzky [21, 22]
also establish that the method can, in principle, recover the exact distances on
graphs as t → 0 (the implicit constant in their result was shown by the author
to have combinatorial significance [80]). This may not be reasonable in practice
6

since an accurate computation of t → 0 requires the use of all eigenvectors and

very high accuracy. Other PDE-type methods have been used, most notably the
eikonal equation |∇u| = 1 [14, 47, 60, 61, 74], the Schrodinger equation [71], the
biharmonic equation [57] and smoothed distances using spectral geometry [16]. Our
method is quite different since we do not work with the heat equation at all and
instead used a pinned ground state of the Laplacian. In particular, our method is
not directly inspired by a continuous method (where imposing that the solution of
an elliptic equation vanishes in a single point comes with some difficulty in d = 2
and significant issues stemming from potential-theoretic obstructions when d ≥ 3).

Hot Spots Conjecture. The Hot Spots conjecture, first posed by Rauch at a 1974
conference in Tulane, asks whether, for a domain Ω ⊂ R2 , the first nontrivial eigen-
function of the Laplacian with Neumann boundary conditions assumes maximum
and minimum on the boundary. It is not quite clear what to assume on the domain
Ω. Burdzy & Werner [11] constructed a counterexample for a domain that is not
simply connected, so the conjecture is now commonly stated either for simply con-
nected domains or convex domains. The Hot Spots conjecture has inspired a lot of
work, we refer to [3, 4, 5, 6, 7, 8, 10, 11, 31, 35, 36, 41, 42, 44, 46, 65, 62, 63, 64, 70, 81]
and references therein. It is expected that some form of the statement is also true in
higher dimensions and on manifolds. The relationship to the problem we consider
is as follows: let −∆φ2 = µ2 φ2 denote the second Neumann eigenfunction. It has
mean value 0 and, generically, we expect the nodal line {x : φ2 (x) = 0} to have one
connected component that touches the boundary ∂Ω in two points and to separate
Ω into two connected components Ω = Ω1 ∪ Ω2 . This is Payne’s nodal domain
conjecture [67] which is known to hold in convex domains [1, 65] and known to fail
in some other settings [31, 39]. Let us now take a simply connected component of
{x : φ2 (x) > 0}. This domain is partially bounded by Dirichlet conditions φ2 (x) = 0
and, on the boundary of ∂Ω, by Neumann boundary conditions ∂u/∂n = 0. This is
the natural continuous analogue of the equation we consider graphs. What one ob-
serves in the continuous setting is that φ2 is a remarkably simple object that seems
to, more or less, nicely increase with distance from the nodal set; the maxima being
on the boundary being merely one instance of this phenomenon. This phenomenon
is poorly understood.

One of the contributions of our paper is a simple algorithm on graphs whose ef-
ficiency is rooted in the very same phenomenon. In particular, it seems that the
underlying reason why the Hot Spots Conjecture is true, is robust enough to survive
the transition to graphs. This is well in line with the assumption that versions of
the Hot Spots conjecture will also hold in higher dimensions. We point out that
Hot Spots problems are not usually considered on graphs due to the difficulty of
defining the boundary; we believe our approach to be a suitably adapted analogue
that avoids this problem altogether. It is quite conceivable that the discrete setting,
allowing for a different set of tools, may give rise to new ideas for the continuous
problem and this is part of the motivation underlying this paper. While the exis-
tence of the phenomenon, the solution of the PDE behaving nicely with respect to
the distance from the nodal set, is not surprising, perhaps its strength, mirrored in
these surprisingly accurate results on topologically highly nontrivial graphs, is.
 7

2.5. Other Laplacians. There are various notions of a Laplacian on a graph.

Throughout this paper, we will only be concerned with the Laplacian matrix L =
D − A. However, there are at least two other canonical notions: the random
walk normalized Laplacian given by Idn×n − D−1 A and the symmetric normalized
Laplacian Idn×n −D−1/2 AD−1/2 . It is not difficult to see that our proof of Theorem
1 has a natural analogue for these notions and it is possible to use the Algorithm
with these other eigenfunctions as underlying geometry. We note that one has to be
careful in using the symmetric normalized Laplacian: evaluating the corresponding
eigenfunction φ of Idn×n − D−1/2 AD−1/2 at a vertex v results in
X φ(w)
φ(v) − p = µ2 φ(v) > 0.
(v,w)∈E
deg(v)deg(w)

Multiplying the equation with deg(v)−1/2 results in

φ(v) 1 X φ(w)
p − p >0
deg(v) deg(v) (v,w)∈E deg(w)

which shows that every vertex v 6= i has a neighboring vertex w such that
φ(w)deg(w)−1/2 < φ(v)deg(v)−1/2
and the algorithm has to account for this change and minimize this expression
instead. What we observe in basic numerical experiments is that in many cases
that we considered, these three Laplacians perform in a very similar way but that
D−A seems, generically, to yield slightly better results. It is not clear to us whether
there are other differential operators that might end up yielding even better results.

2.6. Open Problems. We conclude with a list of open questions; we will only
ask them in the discrete setting but they are (with the proper modifications) also
interesting in the continuous setting.
(1) Rigorous results. Which rigorous results can be proven about the algo-
rithm? We emphasize once more that any such results might have interest-
ing counterparts in the continuous setting.
(2) Upper Bounds on the Length. Is it possible to obtain upper bounds on
the length of such an arising path? This would be of interest both in the
discrete case but also, for gradient flows, in the continuous setting where it
seems to be related to upper bounds on the size of level sets: this questions
is notoriously difficult: the main result in this direction were given by Dong
[24], Donnelly & Fefferman [25, 26], Hardt & Simon [37] and Lin [56]. A
natural conjecture in convex domains is that the length of the gradient flow
is at most c · diam(Ω). It is less clear what to expect on graphs.
(3) Average Guarantees. Even though the Algorithm can fail, it seems to
fail by very little – moreover, in the rare case when it fails by a lot, this
seems to occur for a very small number of vertices (see §3). Is it possible
to obtain upper bounds for a ‘typical’ pair of points?
(4) Fast Computation. How fast can this algorithm be carried out? In
practice, we are not interested in an arbitrary eigenfunction but the smallest
one. Is it important that it is exactly the smallest eigenfunction or are, in
practice, small errors permissible?
8

(5) Symmetrization. The algorithm is not symmetric: the path it finds from
i to j will not necessarily be the reverse path that it produces from j to i.
This allows us to take the better of the two for a more stable estimate.
(6) Effective Variants. Are there effective variations on this idea?

Figure 4. Very different types of graphs on which all pairwise

shortest paths are reconstructed exactly by the spectral method.
The graphs are (1) the Folkman Graph, (2) the Gray graph, (3)
the Harries-Wong Graph, (4) the Ljubljana Graph, (5) the Petersen
Graph, (6) the Line Petersen Graph, (7) the Tutte 12 Cage, (8)
the Wells Graph and (9) the Zamfirescu Graph 36.

3. Some Examples
We describe some explicit examples. Our observations can be roughly summarized
as follows: (1) on graphs with symmetries or some overall homogeneity (see Fig. 4
for some examples), the algorithm tends to produce the shortest path for all pairs of
vertices, (2) on graphs that have different densities and bottlenecks, the algorithm
still produces the shortest path for most pairs of vertices and very short paths on
averages and (3) employing the trick of computing both the optimal path from i
to j and from j to i leads to near-optimal results. We see (see §3.3 and §3.4.) that
even in graphs designed to break the algorithm, it yields optimal results for ∼ 95%
 9

of all pairs of vertices; moreover, the average path produced by the spectral method
is merely ∼ 0.1 steps longer than the optimal one.

3.1. Some Examples that work. We emphasize that the algorithm does, on
average, perform extraordinarily well. Even when it seems to fail, it is often on a
few pairs of vertices by a little bit. Before discussing some of its failures, we give a
list of graphs where it perfectly reconstructs all pairwise distances in Fig. 4.

3.2. Erdős-Renyi Graph. We continue with the Erdős-Renyi model G(n, p). We
pick n vertices and connect each of them with likelihood p. Numerical Experiments
for 100 ≤ n ≤ 500 and p ≥ 0.1 seem to indicate that the method performs perfectly
accurately (with high likelihood). Indeed, in this regime we did not find a single
example where it fails (though it does fail by small amounts for small p). This
could be an interesting starting point for a theoretical investigation.

3.3. Deterministic Graphs. One of the more extreme examples of failure of the
algorithm that we found occurs on an amusing graph. On April 1 (!), 1975, a year
before the four color theorem was proven, Martin Gardner announced in his column
in the Scientific American that the ‘most sensational of last year’s discoveries in
pure mathematics was surely the finding of a counterexample to the notorious four-
color-map conjecture’ [33] and provided such a map that is now known as Gardner’s
Map or the April Fools’ map. It contains a pair of vertices such that the path arising
from the Spectral Method requires 22 steps, 5 steps more than the maximum. The
structure of this example is most intriguing, see Figure 5.

Figure 5. An extreme example we found: the shortest path

(shown in green) has length 17, the Spectral Method leads to a
path (shown in red) of length 22. Nonetheless, for a randomly cho-
sen pair of vertices in the graph, the spectral method produces a
path that is only 0.08 steps longer than the optimal length.
10

Even in this remarkable example, the average case is very good: the path obtained
by the spectral method for a random pair of vertices is only 0.08 steps longer than
the shortest path. We give a second example that occurs on another extreme graph:
the Barnette-Bosak-Lederberg graph [51] which is the smallest known 3-regular
graph without a Hamiltonian cycle, see Fig. 6. The picture that emerges from
these experiments is that counterexamples have to have rather peculiar structure
and even then, the failure happens to be concentrated on relatively few points.

Figure 6. The Barnette-Bosak-Lederberg graph: the shortest

path (shown in green) has length 4, the Spectral Method leads
to a path (shown in red) of length 6. Nonetheless, for a random
pair of vertices in the graph, the spectral method produces a path
that is only 0.09 steps longer than the optimal length.

We pose two questions (that are also of interest for non-planar graphs):
(1) Question 1. Are there planar graphs such that, for some pair of vertices
u 6= v, the path produced by the spectral method is c−times as long as the
shortest path for any c > 1?
(2) Question 2. Is there a universal constant c0 such that for all planar
graphs the path produced by the spectral method between a random pair
of vertices is at most c0 −times as long as the shortest path?

Figure 7. The shortest path (shown in green) has length 8, the

Spectral Method leads to a path (shown in red) of length 12. For a
random pair of vertices in the graph, the spectral method produces
a path that is only 0.05 steps longer than the optimal length.
 11

3.4. Random Geometric Graph. Consider a random geometric graph: we pick

300 points uniformly at random in the rectangle [0, 3] × [0, 1] and connect any two
points with an edge if they are distance at most 0.3. We find that the Spectral
Method (run twice: to find a path from i to j and a path from j to i) performs
exceedingly well on average: the average path between two vertices as found by
the spectral method is 0.05 steps longer than the shortest path. The most extreme
case among all pairs of vertices is shown in the figure: the Spectral Method finds a
path of length 12 while the optimal path has length 8.

4. Proofs
4.1. Proof of Theorem 1.

Proof. Proving feasability of the algorithm requires us to show several statements.

(1) The smallest eigenvector φ of Li does not change sign (validity of step 2)
and can thus be made positive (validity of step 3).
(2) If φ(v) 6= 0, then there exists a neighbor (w, v) ∈ E of v with φ(w) < φ(v).
(3) φ only vanishes in the vertex i.
We start by analyzing the variational problem
X X
φ = arg min (f (w1 ) − f (w2 ))2 subject to f (v)2 = 1.
f (i)=0
(w1 ,w2 )∈E v∈V

Clearly, a minimizer exists by compactness. We observe two cases: either the

graph becomes disconnected when removing i or it stays connected. If it becomes
disconnected (which, for example, would be the case for trees), then we can think of
the problem as being comprised of two or more graphs being connected via i. It thus
suffices to prove the statement for each component individually and we can assume
that G remains connected after removing i. We shall assume this henceforth. Since
X
(f (w1 ) − f (w2 ))2 = hf, Li f i
(w1 ,w2 )∈E

and Li is a symmetric matrix, we know that a minimizer corresponds to the smallest

eigenvalue of Li . This eigenvalue is easily seen to be bigger than 0: if it were 0,
connectedness of the graph implies that f is constant which creates a contradiction
between its `2 −norm being 1 and f (i) = 0. Let us now suppose that f is such a
minimizer. We define g(·) = |f (·)| and note that g(i) = 0, the sum of the squares
of g is still 1 and
X X
(g(w1 ) − g(w2 ))2 = (|f (w1 )| − |f (w2 )|)2
(w1 ,w2 )∈E (w1 ,w2 )∈E
X
≤ (f (w1 ) − f (w2 ))2 .
(w1 ,w2 )∈E

However, since f was a minimizer, we have to have equality: for all (w1 , w2 ) ∈ E
we have to have
|f (w1 ) − f (w2 )| = ||f (w1 )| − |f (w2 )||.
This, in turn, requires that if (w1 , w2 ) ∈ E are two connected vertices, then f (w1 )
and f (w2 ) have the same sign (though it is conceivable that one of the values is 0).
In particular, if f were to change sign, then, since the Graph remains connected
12

after removing i, there would have to be at least one vertex k where f and thus g
vanishes. In summary, we have obtained a function g such that (1)
X X
(g(w1 ) − g(w2 ))2 = (f (w1 ) − f (w2 ))2
(w1 ,w2 )∈E (w1 ,w2 )∈E

is minimal, (2) g ≥ 0 and (3) g vanishes in at least one additional vertex that is not
i. Moreover, again by connectedness, there exists one vertex ` such that g vanishes
in ` and ` has at least one neighbor where g does not vanish. We conclude with one
elementary fact about real numbers: for any {x1 , . . . , xn } ⊂ R and x ∈ R, we have
n n n
!2
X X 1 X
(xk − x)2 ≥ xk − xi
n i=1
k=1 k=1

with equality if and only if x is indeed the mean of the n numbers. This means
that if we replace the value of g in ` by the mean of the neighbors, then the
Dirichlet energy decreases and the L2 −norm increases which is a contradiction to g
minimizing the functional. Therefore f does not change sign and only vanishes in
i. This implies uniqueness: suppose there was a second minimizer, g, distinct from
f . Since both f and g can be interpreted as eigenvectors of a symmetric matrix,
we know that they can be chosen to be orthogonal. However, since a minimizer
does not change sign, the inner product of any two minimizers cannot vanish which
establishes uniqueness. It remains to show that the algorithm cannot get ‘stuck’.
Let v 6= i be a vertex and suppose the algorithm finds itself in v. We take the v−th
row of the equation Li f = µf and obtain that
X X
0 < µf (v) = (f (v) − f (w)) = deg(v)f (v) − f (w).
(v,w)∈E (v,w)∈E

Therefore
X
f (w) < deg(v)f (v)
(v,w)∈E

and there exists at least one neighbor of v where f assumes a smaller value. This
establishes the desired result. 

G2
G1 i

Figure 8. The Duplication Procedure from the Proof of Theorem 2.

 13

4.2. Proof of Theorem 2.

Proof. The proof is based on a duplication trick. Let i ∈ V be given and let
2
P
(w1 ,w2 )∈E (f (w1 ) − f (w2 ))
φ = arg min P 2
.
w∈V f (w)
f :V →R
f (u)=0,f 6≡0

We know from the proof of Theorem 1 that φ does not change sign and can w.l.o.g.
be taken positive. Since the graph is a tree, we know that removal of the vertex
i leads to two connected graphs G1 and G2 (one of them might be empty if i is a
leaf). It remains to prove the desired result for G1 and G2 separately. We instead
argue with a separation trick (see Fig. 8): we mirror one of the graphs and define
its values at a mirrored vertex to be the negative value of φ at the original vertex
and call the arising function ψ. We see that ψ satisfies (D − A)ψ = µ2 ψ since
the equation is valid on each side of the graph and is easily verified in i since
ψ(i) = 0. By fairly standard arguments, we obtain that ψ is the smallest nontrivial
eigenfunction on the new graph and we then obtained the desired property from a
result of Lefevre [53] (see als Gernandt & Pade [34]). An alternative proof could be
given as follows: we can interpret the function φ on G2 ∪ {i} as a function satisfying
(D − A)φ = µ2 φ in all vertices except i. This allows for φ to be reinterpreted as
the expected outcome of a game that was described by Lederman and the author
in [52]. The game is obviously strictly monotonic in the distance from i and this
yields the desired result. We refer to [52] for the details. 

∂Ωn

x0
∂Ω0

Figure 9. Sketch of the Proof of Theorem 3

4.3. Proof of Theorem 3.

Proof. It follows from standard arguments that u > 0 inside the domain. Let us
now suppose that there is a critical point x0 inside the domain. Since we are in two
dimensions and ∆u < 0, the critical point has to be either a local maximum (in
which case it lies on the boundary by assumption) or it is an isolated critical point
x0 . By the Morse Lemma, the function u has a local saddle structure. In particular,
there exist two directions in which u is increasing. We start two increasing gradient
(ascent) flows γ1 and γ2 , one in each direction. These gradient ascent flows can
never intersect themselves or each other (we explain below why). This means they
ascend until they hit a critical point. This critical point can be a maximum (in
14

which case it is at the boundary and we stop) or it is another critical point: in that
case, we use the local saddle structure, rotate the vector by 90◦ and continue the
gradient ascent until we eventually hit a maximum on the boundary. These flows
can never hit ∂Ω0 because the function is increasing along them – this means they
cut the domain into two parts, one containing ∂Ω0 , the other only containing ∂Ωn .
We pick the second part, Ω2 , and note that ∂u/∂n = 0 on the entire domain: on
∂Ωn this is merely the imposed boundary condition, on the part given by γ1 and
γ2 , this follows from the fact that the tangent line of γi is parallel to the gradient
of u in the point. Then, however,
Z Z
∂u
0= dσ = ∆u < 0
∂Ω2 ∂n Ω2

which is a contradiction. The same argument could have also been applied if the
gradient ascent curves were to intersect themselves or each other. It remains to
show that the Gradient flow converges to Ω0 . We pick a point x0 and start the
gradient flow γ1 and, simultaneously, the gradient ascent flow γ2 . We know that γ2
terminates in a critical point and these are all on the boundary. If γ1 were to not
terminate in ∂Ω0 but on ∂Ωn , then we could again extract a domain for which u
has Neumann boundary conditions and obtain a contradiction. This concludes the
argument. 

References
[1] G. Alessandrini. Nodal lines of eigenfunctions of the fixed membrane problem in general
convex domains. Comment. Math. Helvetici, 69:142–154, 1994.
[2] L. Alon, R. Band, M. Bersudsky, S. Egger, Neumann Domains on Graphs and Manifolds,
arXiv:1805.07612
[3] R. Atar and K. Burdzy, On nodal lines of Neumann eigenfunctions, Electronic Communica-
tions in Probability 7 (2002), p. 129–139
[4] R. Atar and K. Burdzy, On Neumann eigenfunctions in lip domains, Journal of the AMS 17
(2004), p. 243–262
[5] R. Bañuelos and K. Burdzy, On the ”hot spots” conjecture of J. Rauch, J. Func. Anal. 164
(1999), p. 1–33
[6] R. Bañuelos and M. Pang, An Inequality for Potentials and the ”Hot-spots” Conjecture,
Indiana University Mathematics Journal 53 (2004), p. 35–47.
[7] R. Bass and K. Burdzy, Fiber Brownian motion and the ”hot spots” problem, Duke Math.
J. Volume 105, Number 1 (2000), p. 25–58.
[8] R. Bellman, On a routing problem. Quarterly of Applied Mathematics. 16 (1960), p. 87–90.
[9] A. Belyaev and P. A. Fayolle. On Variational and PDEBased Distance Function Approxima-
tions. Computer Graphics Forum 34 (2015) pp. 104–118.
[10] K. Burdzy, The hot spots problem in planar domains with one hole, Duke Math. J. 129
(2005), 481–502.
[11] K. Burdzy and W. Werner, A counterexample to the ”hot spots” conjecture, Ann. Math. 149
(1999), p. 309–317
[12] J. Calder, The game theoretic p-Laplacian and semi-supervised learning with few labels,
Nonlinearity 32 (2018), p. 301-330
[13] J. Calder and D. Slepcev, Properly-Weighted Graph Laplacian for Semi-supervised Learning.
Appl. Math Optim. (2019), to appear.
[14] M. Campen and L. Kobbelt. Walking On Broken Mesh: DefectTolerant Geodesic Distances
and Parameterizations. Computer Graphics Forum 30 (2011), 623–632.
[15] B. Cherkassky, A. Goldberg, T. Radzik, Shortest paths algorithms: theory and experimental
evaluation. Mathematical Programming. Ser. A. 73 (2): p. 129–174.
[16] R. Coifman and S. Lafon. 2006. Diffusion Maps, ACHA 21 (2006), p. 5–30.
[17] M. Cohen, R. Kyng, J. Pachocki, R. Peng, A. Rao, Preconditioning in Expectation, STOC
2014
 15

[18] M. Cohen, G. Miller, J. Pachocki, R. Peng, S. C. Xu, Stretching Stretch, arXiv:1401.2454

[19] R. Courant and D, Hilbert, Methods of mathematical physics, I, New York, N.Y.: Interscience
Publishers, Inc. (1953)
[20] R. Courant and D, Hilbert, Methods of mathematical physics, II, New York, N.Y.: Inter-
science Publishers, Inc. (1953(1962)
[21] K. Crane, C. Weischedel and M. Wardetzky, Geodesics in heat: A new approach to computing
distance based on heat flow, ACM Transactions on Graphics (2013), Article No. 152
[22] K. Crane, C. Weischedel and M. Wardetzky, The Heat Method for Distance Computation,
Commun. ACM 60 (2017), p. 90–99.
[23] E. W. Dijkstra, A note on two problems in connexion with graphs. Numerische Mathematik.
1 (1959): 269–271.
[24] R.-T. Dong, Nodal sets of eigenfunctions on Riemann surfaces, J. Differential Geom. 36:2
(1992), 493–506
[25] H. Donnelly and C. Fefferman, Nodal sets of eigenfunctions on Riemannian manifolds, Invent.
Math. 93 (1988), p. 161–183.
[26] H. Donnelly and C. Fefferman, Nodal sets for eigenfunctions of the Laplacian on surfaces, J.
Amer. Math. Soc. 3 (1990), p. 333–353
[27] Robert W. Floyd, Algorithm 97: Shortest Path. Comm. ACM. 5 (1962): p. 345.
[28] L. R. Ford, Network Flow Theory. Rand Corporation. P-923.
[29] M. Fredman and R. Tarjan, Fibonacci heaps and their uses in improved network optimization
algorithms. 25th Annual Symposium on Foundations of Computer Science. IEEE (1984). p.
338–346.
[30] M. Fredman and R. Tarjan, Fibonacci heaps and their uses in improved network optimization
algorithms. Journal of the Association for Computing Machinery. 34 (1987): p. 596–615.
[31] P. Freitas, Closed Nodal Lines and Interior Hot Spots of the Second Eigenfunctions of the
Laplacian on Surfaces, Indiana Univ. Math. J. 51 (2002), 305–316.
[32] F. Gao, G. Wei, S. Xin, S. Gao, S. and Y. Zhou, 2D skeleton extraction based on heat
equation. Computers & Graphics 74 (2018), p.99–108.
[33] M. Gardner, Mathematical Gems, Scientific American 232 (1975), pp. 126-133.
[34] H. Gernandt and J. Pade, Schur reduction of trees and extremal entries of the Fiedler vector,
Linear Algebra and its Applications 570 (2019), p. 93–122
[35] D. Grieser and D. Jerison, Asymptotics of the first nodal line of a convex domain, Invent.
Math., 125 (1996), p. 197–219.
[36] D. Grieser and D. Jerison, The size of the first eigenfunction of a convex planar domain, J.
Amer. Math. Soc. 11 no. 1 (1998), 41–72.
[37] R. Hardt and L. Simon, Nodal sets for solutions of elliptic equations, J. Differential Geom.
30 (1989), 505–522.
[38] P. E. Hart, N. J. Nilsson and B. Raphael, A Formal Basis for the Heuristic Determination
of Minimum Cost Paths. IEEE Transactions on Systems Science and Cybernetics. 4 (2): p.
100–107.
[39] M. Hoffmann-Ostenhof, T. Hoffmann-Ostenhof, and N. Nadirashvili. On the nodal line con-
jecture. In Advances in differential equations and mathematical physics (Atlanta, GA, 1997),
volume 217 of Contemp. Math, p. 33–48.
[40] P. Ingerman, Algorithm 141: Path Matrix, Communications of the ACM. 5 (1962): p. 556.
[41] D. Jerison Locating the first nodal line in the Neumann problem. Trans. Amer. Math. Soc.
352 (2000), p. 2301–2317.
[42] D. Jerison and N. Nadirashvili, The hot spots conjecture for domains with two axes of sym-
metry. J. Amer. Math. Soc. 13 (2000), p. 741–772.
[43] D. Johnson, Efficient algorithms for shortest paths in sparse networks, Journal of the ACM,
24 (1977): p. 1–13,
[44] C. Judge and S. Mondal, Euclidean triangles have no hot spots, Annals of Math. 191 (2020),
p. 167–211.
[45] R. Karlsson, P. Poblete, An O(m log log D) algorithm for shortest paths”. Discrete Applied
Mathematics. 6 (1983): p. 91–93.
[46] B. Kawohl, Rearrangements and Convexity of Level Sets in PDE, Lecture Notes in Mathe-
matics 1150, Springer, Berlin, 1985.
[47] R. Kimmel and J.A. Sethian. Fast Marching Methods on Triangulated Domains. Proc. Nat.
Acad. Sci. 95 (1998), 8341–8435.
16

[48] I. Koutis, G. Miller, R. Peng A nearly-m log n time solver for SDD linear systems, FOCS
2011
[49] R. Kyng, S. Sachdeva, Approximate Gaussian Elimination for Laplacians: Fast, Sparse, and
Simple, arXiv:1605.02353
[50] R. Kyng, Y.-T. Lee, R. Peng, S. Sachdeva, D. Spielman, Sparsified Cholesky and Multigrid
Solvers for Connection Laplacians, STOC 2016
[51] J. Lederberg, Hamilton Circuits of Convex Trivalent Polyhedra (Up to 18 Vertices), The
American Mathematical Monthly 74, (1967), pp. 522–527
[52] R. Lederman and S. Steinerberger, Extreme Values of the Fiedler Vector on Trees,
arXiv:1912.08327
[53] J. Lefevre, Fiedler Vectors and Elongation of Graphs: A Threshold Phenomenon on a Par-
ticular Class of Trees, arXiv:1302.1266
[54] Y.-T. Lee, A. Sidford, Efficient Accelerated Coordinate Descent Methods and Faster Algo-
rithms for Solving Linear Systems. 54th Annual IEEE Symposium on Foundations of Com-
puter Science, FOCS 2013
[55] B. Lin, J. Yang, X. He, and J. Ye. Geodesic distance function learning via heat flow on vector
fields. In International Conference on Machine Learning (2014), pp. 145–153.
[56] F.-H. Lin, Nodal sets of solutions of elliptic and parabolic equations. Comm. Pure Appl.
Math. 44 (1991): p. 287–308
[57] Y. Lipman, R. Rustamov, and T. Funkhouser. Biharmonic Distance. ACM Trans. Graph. 29
(2010), Article 27
[58] R. Litman and Alex M. Bronstein. Spectrometer: Amortized sublinear spectral approximation
of distance on graphs. In 2016 Fourth International Conference on 3D Vision (3DV), pp. 499–
508. IEEE, 2016.
[59] A. Melas, On the nodal line of the second eigenfunction of the Laplacian in R2 , J. Differential
Geom. 35 (1992), p. 255–263.
[60] F. Memoli and G. Sapiro, Fast Computation of Weighted Distance Functions and Geodesics
on Implicit Hyper-Surfaces. Journal of Comp. Physics 173 (2001), p. 730–764.
[61] F. Memoli and G. Sapiro, Distance Functions and Geodesics on Submanifolds of Rd and Point
Clouds. SIAM J. Appl. Math. 65 (2005), pp. 1227–1260.
[62] Y. Miyamoto, The hot spots conjecture for a certain class of planar convex domains Journal
of Mathematical Physics 50 (2009), 103530.
[63] Y. Miyamoto, A planar convex domain with many isolated hot spots on the boundary, Japan
Journal of Industrial and Applied Mathematics 30 (2013),p. 145–164.
[64] M. Pascu, Scaling coupling of reflecting Brownian motions and the hot spots problem, Trans.
Amer. Math. Soc. 354 (2002), p. 4681–4702.
[65] A.D. Melas. On the nodal line of the second eigenfunction of the Laplacian in R2 , J. Diff.
Geom., 35:255–263, 1992.
[66] E. F. Moore, The shortest path through a maze. Proceedings of an International Symposium
on the Theory of Switching (Cambridge, Massachusetts, 2-5 April 1957). Cambridge: Harvard
University Press, 1959, p. 285–292.
[67] L. Payne, Isoperimetric inequalities and their applications. Siam Review 9 (1967):453–488.
[68] S. Pettie, A new approach to all-pairs shortest paths on real-weighted graphs. Theoretical
Computer Science. 312 (2004): p. 47–74.
[69] M. Pollack, W. Wiebenson, Solution of the Shortest-Route ProblemA Review. Oper. Res. 8
(1960): p. 224–230
[70] M. Rachh and S. Steinerberger, On the location of maxima of solutions of Schroedinger’s
equation, Comm. Pure. Appl. Math. 71 (2018), 1109–1122.
[71] A. Rangarajan and K. Gurumoorthy. A Fast Eikonal Equation Solver using the Schrodinger
Wave Equation. Technical Report REP-2011- 512. CISE, University of Florida
[72] B. Roy, Transitivite et connexite. C. R. Acad. Sci. Paris 249 (1959): p. 216–218.
[73] R. Seidel, On the All-Pairs-Shortest-Path Problem in Unweighted Undirected Graphs. Journal
of Computer and System Sciences. 51 (1995): p. 400–403.
[74] J.A. Sethian. 1996. Level Set Methods and Fast Marching Methods: Evolving Interfaces in
Computational Geometry, Fluid Mechanics, Computer Vision and Materials Science. Cam-
bridge University Press.
[75] N. Sharp, Y. Soliman and K. Crane. The vector heat method. ACM Transactions on Graphics
(TOG) 38, no. 3 (2019): pp. 1–19
 17

[76] A. Shimbel, Structure in communication nets. Proceedings of the Symposium on Information

Networks. New York, New York: Polytechnic Press of the Polytechnic Institute of Brooklyn,
1955, pp. 199–203.
[77] D. Spielman and S.-H. Teng, A Local Clustering Algorithm for Massive Graphs and its
Application to Nearly-Linear Time Graph Partitioning, SIAM Journal on Computing 42
(2013), p. 1–26
[78] D. Spielman and S.-H. Teng, Spectral Sparsification of Graphs, SIAM Journal on Computing
40 (2011), p. 981–1025
[79] D. Spielman and S.-H. Teng, Nearly-Linear Time Algorithms for Preconditioning and Solving
Symmetric, Diagonally Dominant Linear Systems, SIAM Journal on Matrix Analysis and
Applications 35 (2014), p. 835–885.
[80] S. Steinerberger, Varadhan Asymptotics for the Heat Kernel on Finite Graphs,
arXiv:1801.02183.
[81] S. Steinerberger, Hot Spots in Convex Domains are in the Tips (up to an Inradius), Comm.
PDE, to appear.
[82] J. Tao, J. Zhang, B. Deng, Z. Fang, Y. Peng and Y. He, Parallel and scalable heat meth-
ods for geodesic distance computation. IEEE transactions on pattern analysis and machine
intelligence (2019).
[83] M. Thorup, Undirected single-source shortest paths with positive integer weights in linear
time, Journal of the ACM. 46 (1999): p. 362–394.
[84] N. Garcia Trillos and D. Slepvev, A variational approach to the consistency of spectral clus-
tering, Applied and Computational Harmonic Analysis 45 (2018), p. 239–281.
[85] S. R. S. Varadhan. On the behavior of the fundamental solution of the heat equation with
variable coefficients. Comm. Pure Appl. Math. 20 (1967): p. 431–455,
[86] S. R. S. Varadhan. Diffusion processes in a small time interval. Comm. Pure Appl. Math 20
(1967): p. 659–685, 1967.
[87] N. Vishnoi, Lx = b – Laplacian Solvers and Their Algorithmic Applications, Now Publishers
(2013)
[88] S. Warshall, A theorem on Boolean matrices. Journal of the ACM. 9 (1962): p. 11–12.
[89] P. D. Whiting, J. A. Hillier, A Method for Finding the Shortest Route through a Road
Network. Operational Research Quarterly. 11 (1960): p. 37–40
[90] R. Williams, Faster all-pairs shortest paths via circuit complexity. Proceedings of the 46th
Annual ACM Symposium on Theory of Computing (STOC ’14). New York: ACM. pp. 664–
673.
[91] F. Yang and L. Cohen, Geodesic distance and curves through isotropic and anisotropic heat
equations on images and surfaces. Journal of Mathematical Imaging and Vision 55 (2016), p.
210–228.
[92] F. Yang, L. Chai, D. Chen, D. and L. Cohen, Geodesic via Asymmetric Heat Diffusion Based
on Finsler Metric. In Asian Conference on Computer Vision 2018, pp. 371–386.
[93] A. Yuan, J. Calder, B. Osting, A continuum limit for the PageRank algorithm,
arXiv:2001.08973

Department of Mathematics, Yale University, New Haven, CT 06511, USA

E-mail address: stefan.steinerberger@yale.edu