DIFFERENTIAL EQUATIONS ON GRAPHS

OLIVER KNILL, HCRP PROJECT WITH ANNIE RAK

Abstract. We look at examples of dynamical systems on finite

simple graphs. These systems correspond to partial differential
equations in the continuum. This documents contains some notes
related to a project supported by the HCRP Summer 2016 program
with Annie Rak who submitted in November 2016 a senior thesis
about advection, a particular aspect of this PDE topic.

1. Introduction
1.1. Most partial differential equations on a compact Riemannian
manifold M are formulated in terms of an exterior derivative d given on
M . A Riemannian metric on M then defines inner products on k-forms
and so dual notions d∗ . Of particular interest are the Laplacians Lk =
d∗k dk + dk−1 d∗k−1 which define Laplace equations Lk u = 0, Poisson
equations Lk u = g, heat flows u0 = −Lk u or wave equations u00 =
−Lu all defined on k-forms. Examples of nonlinear equations are the
sin-Gordon equation u00 = −Lu+sin(u) or the eiconal equation |df | =
1 on k-forms. Among linear systems, also the advection transport
equations like u0 = d∗v du or u0 = d∗ dv u are of interest, where dv rsp.
d∗v are modifications of d, d∗ playing the role of a generalized gradient
or divergence and where v is a vector field, defined by a weighted edge
attached to a vertex.

1.2. Since classical PDEs only refer to an exterior derivative, a trans-

lation to the discrete should be straight forward. In particular, an
adaptation to graphs where a natural exterior derivative and calculus
exists using incidence matrices as defined by Poincaré. One should
note however that discrete differential calculus has been initiated half
a century before Poincaré already starting with Kirchhoff who is com-
memorated in the Laplacian L on a graph: one often calls this matrix
the Kirchhoff operator. Some of the discretizations, especially in

Date: Summer, 2016.

1991 Mathematics Subject Classification. 05C21,94C15,37M99.
Key words and phrases. Graph theory, dynamical systems.
1
2 OLIVER KNILL, HCRP PROJECT WITH ANNIE RAK

the nonlinear realm are not canonical at all. One would like to em-
ulate the continuum as well as possible. Discretizations of nonlinear
systems in particular should have global existence results if the con-
tinuum version has global solutions. The discrete versions also should
preserve integrability. The discretization should also work for gen-
eral networks and not only for numerical discretisations of continuum
spaces. Still, already in the linear case we would like to have a match
between the behavior on discrete spaces and the continuum limit with-
out just doing numerical discretizations in the form of partial difference
equations. In the discrete, on a general network, functions on the set of
k-dimensional simplices play the role of k-tensors. After a pre-ordering
of the simplices, one can define anti-symmetric tensors, discrete ana-
logue of k-forms. Following Poincaré, the exterior derivatives are then
given by incidence matrices of the graph. While the Laplacians are
independent of the choice of directions, the prescribed orientations in
general matter. The classical transport equation ut = ux for example
on k-forms, needs to be replaced by ut = F du, where F is a suitable op-
erator from (k + 1)-forms to k-forms. If F = −d∗ , transport happens in
all directions, leading to diffusion. For F (u)(v) = u(G(v)) with a vector
field G assigning to a vertex v an edge G(v), then this is deterministic
transport and a flow analogue to a differential equation u0 = G(u) on
a manifold M , where G is a vector field, a map x → G(x) ∈ Tx M from
M to the tangent bundle of M .

1.3. When looking at evolution equations in a discrete set-up, one has

to make some choices: one can discretize space, time or the tar-
get space of the functions. In this project, we assume space to be
discrete. Rather than quantizing time also, we consider difference-
differential equations of the form ut = A(u) with a discrete differ-
ence operator defined by possibly modified incidence matrices, leading
to exterior derivative d and operators D = d + d∗ or L = D2 . If
a time derivative is involved, these equations become ordinary dif-
ferential equations with finitely many variables. What about time
and target space discretisations? A time discretisation in general is
no problem, as on a fundamental level, one can just deform the poly-
nomial algebra to get the same formalism as in the continuum. We
don’t go into this quantum calculus frame work here (known to New-
ton, Gregory and Taylor already) but roughly, it works as follows:
the time derivative d/dt is replaced with a difference operator like
Df (t) = f (t + 1) − f (t). Now, if the polynomials are deformed such
that Dxn = nxn−1 and the exponential function is redefined so that
Deax = aeax (there is canonical way to do that like [x]2 = x(x − 1)
 DIFFERENTIAL EQUATIONS ON GRAPHS 3

so that D[x]2 = (x + 1)x − x(x − 1) = x2 = 2[x] etc or by defining

eax = (1+ax)n so that ea(x+1) −eax = aeax ), we don’t have to change the
content of calculus books. The original intentions of the calculus books
is then just the limiting idealized case, where one deals with smooth
functions). For more details, see [5]. Like Nelsons non-standard anal-
ysis [9], the discrete setup is an extension of the traditional calculus
as we know and teach it. The results remain true verbatim: its just
that the meaning of the formulas has changed. So, discretizing time
is no problem on a fundamental level. Discretizing the target space is
more subtle and serious, as it leads to cellular automata [12], where the
adaptations of the models is often not obvious: the task for example to
build cellular automaton for which solutions behave like the wave equa-
tion is more subtle if the target space is just a finite set but it is done
already on a computer as a computer always deals with functions tak-
ing finitely many values, but where the computer also produces some
unpredictable noise due to rounding arithmetic. The investigation of
finite dynamics obtained by simulating continuum systems has been
pioneered in the 70ies already [10] and was studied since. Examples
are [13, 7]. So, in the following, we assume space is a graph, time is
the real axes and the functions are real-valued.

2. Examples
2.1. The Heat equation on k-forms is the dynamical system

ut = −Lu

on k-forms, where L is the Laplacian on k forms. It produces the heat

flow. Since L is just a finite matrix, it can be solved Pdirectly−λwith
−Lt
e u(0) or by doing an eigenvalue expansion u(t) = j∈Z ûj e j t fj .
The set Ωk of k-harmonic functions is the attractor on Ωk . In the
case k = 0, these are the functions which are constant on connected
on Ω = Ωk as
S
components. The flow is geometrically Pinteresting
k
in general the super trace str(L) = k (−1) tr(Lk ) stays constant
by
P McKean-Singer. P This implies that the super trace str(e−Lt ) =
−Lk t −Lk t
Pkeven tr(e2 ) − kodd tr(e ) is constant. The energy E(u) =
x∈V u(x) = hu, ui is always a Lyapunov function on any space
Ω of differential forms: it decreases as E 0 (u) = h−Lu, ui = −hdd∗ +
k

d∗ du, ui which is −hd∗ u, d∗ ui − hdu, dui = −|d∗ u|2 − |du|2 ≤ 0.

2.2. Given an initial position and velocity, the wave equation

utt = −Lu
4 OLIVER KNILL, HCRP PROJECT WITH ANNIE RAK

has the explicit d’Alembert solution cos(Dt)u(0)+sin(Dt)D−1 u0 (0),

where D = d + d∗ and D−1 is the inverse matrix restricted to the ortho-
complement of the kernel. We see that it is really useful to look at the
wave equation simultaneously on all forms Ωk since there, the Laplacian
L has a factorization L = D2 . Because D is a non invertible matrix,
the initial velocity u0 (0) must be chosen perpendicularly to the kernel
of D. The inverse D−1 can then be understood as the Moore-Penrose
inverse of D: we only invert on the subspace spanned by eigenvectors
to non-zero eigenvalues. In the case of scalar functions, we only need
the velocity to average to zero. The solution can also P be obtained
by going into aPbasis in which L is diagonal. If u(0) = k ûk fk and
u0 (0) = v(0) = v̂k fk and Lfk = λk fk , then
X p p p
u(t) = ûk cos( λk t)fk + v̂k sin( λk t)/ λk fk .

The wave equation has the integrals of motion û2k +v̂k2 and is an example
of a Hamiltonian system. On each eigen-mode, we have a harmonic
oscillator. The wave equation therefore is equivalent to a product of
finitely many harmonic oscillators. This is the same in the continuum,
but only that the number of oscillators is now countable. A Fourier ex-
pansion renders the wave equation equivalent to a sequence of harmonic
oscillators. Now, the d’Alembert solution can be implemented fast nu-
merically if the matrix exponential has been built in. Computer algebra
systems have procedures like matrix exponentiation already hardwired
so that solving the wave equation on a general network needs only a
few lines of code.

2.3. The Laplace equation

Lu = λu

is just an eigenvalue problem for the form Laplacian L. For λ = 0,

the solutions are called harmonic forms. There are solutions to the
Laplace equation if and only if λ is an eigenvalue of L. This motivates
to study the eigenvalues of the operators Lk on k-forms. As mentioned
already, these spectra have some Mc-Kean Singer symmetry: the union
of the eigenvalues on even forms is the union of the eigenvalues on odd
forms. [8, 4] The inverse spectral theorem is to find the set of graphs for
which the form spectra are the same. There are many examples known
of this type. The structure of the solutions is of interest too. For any
eigenfunction, one can look the nodal surfaces, graphs formed by
complete subgraphs on which function changes sign [6].
 DIFFERENTIAL EQUATIONS ON GRAPHS 5

2.4. Also the Poisson equation

Lu = g

is in finite dimensions just a linear algebra problem. For g perpendic-

ular to the kernel of L, this has a solution u = L−1 g, where L−1 is the
pseudo inverse. Many problems in physics can be formulated as Pois-
son equations. Here are two examples: the Poisson equation LA = j
on 1-forms for example has an interpretation as a Maxwell equation:
think of j as the current and of A as the electro magnetic potential.
If A is gauged (replaced by A + df ) so that d∗ A = 0, then F = dA
and j = LA = (d∗ d + d∗ d)A = d∗ dA = d∗ F it gives the Maxwell
equations dF = 0, d∗ F = j. The Poisson equations also give the force
F from the mass density ρ as LF = ρ is the Gaussian formulation
of gravity. In R3 with a spherically symmetric spherically symmetric
charge distribution of compact support, it produces the Newton force
C/r2 away from that support, where r is the distance to the center of
the charge and C is a constant which depends on the mass of the body.

2.5. Like the Laplace or Poisson equation, the eiconal equation is

not a time evolution. It is the discrete difference equation

|du| = 1

where |du(x)| is the maximal absolute value of the derivative du on

k + 1 forms attached to the simplex x. For k = 0, the Eiconal equation
tells that all gradients are constant 1. It is nonlinear in nature as in
the continuum; in Euclidean space R3 for example it is fx2 + fy2 + fz2 =
1. Already in the continuum, on simple surfaces like ellipsoids, the
solutions fail uniqueness on caustics defined by the initial condition.
The structure of these caustics is still in the dark even for ellipsoids
and initial Dirac initial conditions located on a point, where Jacobi’s
last statement on caustic is an open problem. In the discrete, the
solutions of the eiconal equations are k-forms with the property that
the maximal steepness at every k-simplex x is 1. This interpretation is
motivated by the fact that in the continuum, the maximal directional
derivative at a point is the length of the gradient. Let a graph be
flat if the distance function to any point satisfies the eiconal equation
globally. Examples of flat graphs are wheel graphs with less than 7
vertices or trees or the hexagonal tiling graph in the plane, where each
vertex has 6 neighbors. Can we characterize graphs which admit a
global solution to the eiconal equation?
6 OLIVER KNILL, HCRP PROJECT WITH ANNIE RAK

2.6. Convection. Replace the adjoint d∗ in ut = −d∗ du or ut =

−dd∗ u or ut = −Lu with an other operator from (k + 1)-forms to k
forms. An example for k = 1 is

ut = d(v · u)

which is a differential equation on 1-forms. Here vP is a fixed 1-form

and v · u is the dot product defined as v · u(x) = (y,x)∈E v(y)u(y).
One can think of v as a background field which guides the diffusion. It
for example applies if a graph is weighted.
P Think of v(e) as how much
can get through the edge e. The sum x u(x) is conserved under the
time evolution simply because its derivative is a sum over gradients.
2.7. Advection. We could also look at the system

ut = v · du

on 0 forms. Here again, v is a fixed background direction 1-form.

This allows to modelP the flow of a “vector field” v. This is a modi-
fication of u0 (i) = j F (e)du(e). If L+ = B+ − A, where B+ is the
out vertex degree matrix and A is the adjacency matrix. Now,
u0 = −L+ u. There is a dual consensus dynamics u0 = −L− u and
0
P dynamics u = −Lu which is0 the heat
of course diffusion P equa-
tion. The sum x∈V u(x) is conserved. [1] look at u (i) = j v(j →
i)u(j) − j v(i → j)u(i) = (Dv − Av )u. Here, v(i → j) is a function
P
attached to the directed edge i → j. The right hand side can be writ-
ten as d∗ vd0 , where d0 = max(d, 0) and d is the exterior derivative.
Compare that d∗ vd is a weighted graph Laplacian.

2.8. The Sine-Gordon equation is

utt = −Lu − c sin(u)

In the simplest case of a one-point graph G = K1 we have the physics

pendulum utt = −c sin(u), one of the simplest nonlinear Hamiltonian
systems. In the case of the complete graph with two vertices K2 , we
have the Hamiltonian differential equation
x00 (t) = −x(t) + y(t) − c sin(x(t)), y 00 (t) = −y(t) + x(t) − c sin(y(t)) .
It defines a flow on a 3-dimensional energy surface which we measure to
be integrable: the orbits appear to lie on tori. For c = 0, it is a flow on
a three sphere. Also for small c, the energy surface is still a topological
 DIFFERENTIAL EQUATIONS ON GRAPHS 7

sphere and we expect the flow to remain integrable in the sense that
all orbits are located on points or 1D or 2D tori. Numerics confirms
that. Can this be written as a Lax pair as in the case of the pendulum?
The three dimensional energy surface H(x, y, x0 , y 0 ) = (x02 + y 02 )/2 +
(x − y)2 /2 + cos(x) + cos(y) is constant. For a circular graph G = Cn ,
the stationary points of the sin-Gordon equation are critical points
of the Frenkel-Kontorova model and correspond to periodic solutions
of a symplectic map, the Standard map given by the second order
recursion equation xn+1 − 2xn + xn−1 = c sin(xn ) leading to the map
(x, y) → (2x − y + c sin(x), x) on the 2-torus R2 /(2πZ)2 .
2.9. Burgers equation. The classical friction-free Burgers equation
on the real line ut = 2uux can be written as ut = (u2 )0 , where u0 is the
gradient. If u is a function which vanishes at infinity or is defined on
Ra circle, this differential
0
Requation classically
R has the integral F (u) =
R
log(u) dx as F (u) = u
R t
/u dx = u
R x
dx = 0. In the discrete, we
k k+1
can define an operator P F :Ω →Ω which assigns to a form u the
form F (u)(x) = S( y,z⊂x u(y)u(z)), here S anti-symmetrizes, which is
only necessary for k > 0. Now we can look at the differential equation

u0 = d∗ F (u) .

As an example, lets consider at the case of a circular graph Cn =

(V, E) and a 0-form u, for which F (u)((a, b) ∈ E) = u(b)u(a). Now,
d∗ F (u)(n) = u(n)(u(n+1)−u(n−1)). The differential equation is well
studied and called the Volterra-Kac-Moerbeke Langmuir lattice
ut (n) = u(n)[u(n + 1) − u(n − 1)] .
It is integrable and equivalent to the Toda lattice [11]. What happens
on a general graph? Is it still integrable?
2.10. Navier Stokes type equations. Assume u is a function on
edges modeling a velocity field and that we have a pressure 2-form
p(u), a stress tensor T (u) and a tensor product u ⊕ u and a force
term F which is a 1-form, a discrete Cauchy momentum equations is
u0 + d∗ (u ⊕ u + p) = d∗ T + F . To make sense of this in the discrete, we
need a cross product on graphs and a notion of pressure. Since pressure
is defined as F = pA,P where A is the area element and F the force, we
could define u ⊕ v as i,j ui vj , where i, j runs over all pairs of edges in
the triangle t. If we think about T as the adjacent pressure functions,
then (u ⊕ u + p − T )(t) could be modeled more easily as a function on
triangles given by summing u over all edges hitting the triangles but not
8 OLIVER KNILL, HCRP PROJECT WITH ANNIE RAK

being in the triangle, then summing over all pairs (e, f ) of edges in the
triangle t and then over all pairs of edges not in the triangle. The most
0 ∗
P
natural simplification is u = d F (u), where F (u)(t) = e,f ⊂t u(e)u(f )
is a nonlinear 2-form, a function on triangles. One can generalize this
to differential equations on k-forms:

u0 = d∗
P
e,f u(e)u(f )
,

where the sum F (u)(x) on a (k + 1) simplex x sums over all pairs of

k-simplices which intersect. This is not yet explored at all, but it could
be that ”fluid dynamics” in a connection calculus setup on a network
can be set naturally. In the continuum, the difficulty is illustrated by
the struggle to see whether global solutions exist.

3. Wave speeds
Since L = D2 , the wave equation (d2 /dt2 + L)ψ = 0 can be factored
(d/dt + iD)(d/dt − iD)ψ = 0 leading to the Schrödinger equations
ψ 0 = ±iDψ for the Dirac operator D and a complex-valued function
ψ(t) = u(t) + iD−1 u0 (t) encoding position u(t) = Re(ψ(t)) and velocity
u0 (t) = DIm(ψ(t)) of the classical wave. We write expx (v) = ψ(t) if
ψ(0) = x + iD−1 v. One of the basic questions we would like to know
is how fast signals propagate on the space of k-forms. We have not yet
looked at this systematically.

4. Integrability
A differential equation is DS integrable, if every invariant measure
defines an almost periodic systems in the sense that the spectrum of
the Koopman operator has pure discrete spectrum [2]. In smaller di-
mensional systems, this can be the case if every invariant measure of
the flow is located either on a fixed point, a fixed circle, a fixed 2-torus
or higher dimensional torus and the dynamics is conjugated there to a
translation. The pendulum is an example the flow is integrable as ev-
ery invariant measure is either located on an equilibrium point or a one
dimensional circle (homotopically trivial for small energies and wrap-
ping around the cylinder phase space for larger energies). The heat
and wave equations are examples of integrable systems when consid-
ered on graphs: For the heat flow, every invariant measure is located on
a point, a constant function, for the wave flow, every invariant measure
is located on a torus. The dimension depends on rational dependencies
of the scaled eigenvalues 2πλk of the Dirac operator D = d + d∗ of the
 DIFFERENTIAL EQUATIONS ON GRAPHS 9

graph.

A special case of the just given notion of integrability is Liouville in-

tegrability. Recursively, a system can be called Liouville integrable
if it is defined on a finite dimensional manifold M , for which there ex-
ists a smooth invariant function (Liouville integral) such that on every
level set {f = c} the system is again Liouville integrable. Liouville
integrability implies integrability in the above sense but the reverse is
not known. One could imagine invariant sets on which the dynamics
is conjugated to a group translation on an other compact group than a
torus. But no such example seems to be known. The problem is related
to the problem to determine whether there are open sets of Hamilton-
ian systems of dimension d ≥ 4 on which no weak mixing tori exist;
it is not known whether open sets of smooth integrable Hamiltonian
systems exist in dimension larger than 2, if integrability is understood
in the above sense (using invariant measures).

There are many notions of “integrable”. Birkhoff for example suggested

an analytic version in that that the phase space is partitioned into
regions, in which the Birkhoff normal form at an equilibrium point
converges. This is probably not equivalent since it can happen that
in touching boundaries of different regions of analyticity the dynamics
is chaotic. Such phenomena happen in complex dynamics, where on
the boundary of a Siegel disk or Herman ring the dynamics can be
chaotic. No such phenomena are known in say more narrow classes
of systems like Hamiltonian systems. Even the classic Kolmogorov
problem on mixing invariant tori (which initiated research on KAM
theory initially) is open. See [3].

5. Topology
Differential equations on graphs help to prove theorems. Here is an ex-
ample, where the heat flow is used to prove the Brouwer-Lefschetz fixed
point theorem. A special case of this theorem tells that if a graph G is
homotopic to a point, then any automorphism of G necessarily has a
fixed simplex. Since a simplex is a vertex of the Barycentric refinement
and the Barycentric refinement of a graph has the same cohomology,
this easily implies the classical Brouwer-Lefschetz fixed point theorem.

Lets first look at the cohmology. Since the exterior derivative satisfies
d2 = 0, the kernel of dp : Ωp → Ωp+1 contains the image of dp−1 . The
10 OLIVER KNILL, HCRP PROJECT WITH ANNIE RAK

vector space
H p (G) = ker(dp )/im(dp−1 )
is the p’th simplicial cohomology of G.
The Betti numbers bp (G) = dim(H p (G)) define p (−1)p bp which is
P

Euler characteristic χ(G) = x (−1)dim(x) , summing over all com-

P
plete subgraphs x of G. If T is an automorphism of the graph G, the
Lefschetz number, the super
P trace χT (G) of the induced map UT on
p dim(x)
H (G) is equal to the sum T (x)=x iT (x), where iT (x) = (−1) sign(T |x)
is the Brouwer index. This is the Lefschetz fixed point theorem:

For T = Id, the Lefschetz formula is Euler-Poincaré. By Hodge

theory, bp (G) is the nullity of L restricted to p-forms. By McKean
Singer super symmetry, the positive Laplace spectrum on even-
forms is the positive Laplace spectrum on odd-forms. The super trace
str(Lk ) is therefore zero for k > 0 and l(t) = str(exp(−tL)UT ) with
UT f = f (T ) is t-invariant.PThis heat flow argument proves Lefschetz
because l(0) = str(UT ) is T (x)=x iT (x) and limt→∞ l(t) = χT (G) by
Hodge.

References
[1] A. Chapman and M. Mesbahi. Advection on graphs. In 2011 IEE Conference
on Decision and Control of European Control Conference, 2011.
[2] O. Knill. On a spectral notion of integrability.
Univ. of Alabama in Birmingham talk, 3/28/2002,
http://www.math.harvard.edu/˜knill/seminars/integrable/birmingham.pdf.
[3] O. Knill. Weakly mixing invariant tori of hamiltonian systems. Commun. Math.
Phys., 205:85–88, 1999.
[4] O. Knill. The McKean-Singer Formula in Graph Theory.
http://arxiv.org/abs/1301.1408, 2012.
[5] O. Knill. If Archimedes would have known functions .
http://arxiv.org/abs/1403.5821, 2014.
[6] O. Knill. A Sard theorem for graph theory.
http://arxiv.org/abs/1508.05657, 2015.
[7] Oscar E. Lanford, III. Informal remarks on the orbit structure of discrete
approximations to chaotic maps. Experiment. Math., 7(4):317–324, 1998.
[8] H.P. McKean and I.M. Singer. Curvature and the eigenvalues of the Laplacian.
J. Differential Geometry, 1(1):43–69, 1967.
[9] E. Nelson. Internal set theory: A new approach to nonstandard analysis. Bull.
Amer. Math. Soc, 83:1165–1198, 1977.
[10] F. Rannou. Numerical study of discrete area-preserving mappings. 31:289–301,
1974.
[11] H. Toda. Theory of nonlinear lattices. Springer-Verlag, Berlin, 1981.
[12] S. Wolfram. Theory and Applications of Cellular Automata. World Scientific,
1986.
 DIFFERENTIAL EQUATIONS ON GRAPHS 11

[13] X-S. Zhang and F. Vivaldi. Small perturbations of a discrete twist map. Ann.
Inst. H. Poincaré Phys. Théor., 68:507–523, 1998.

Department of Mathematics, Harvard University, Cambridge, MA, 02138,

USA