Coco notes — Algebraic graph calculus
$newcommand{1}{mathbf{1}}$
$newcommand{R}{mathbf{R}}$
We describe a graph-theoretic analogue of vector calculus. The linear
operators of vector calculus (gradient, divergence, laplacian) correspond to
the matrices naturally related to graphs (incidence matrix, adjacency
matrix). This analogy is helpful for formalizing the discretization of some
issues in picture and floor processing which are typically outlined in a steady
setting.
1. Reminder of vector calculus
Vector calculus offers with capabilities and vector fields outlined in $R^3$.
1.1. Features and vector fields
A operate (or scalar discipline) is a map $u:R^3toR$.
A vector discipline is a map $mathbf{v}:R^3toR^3$.
Vector fields are written in daring.
Allow us to repair some typical names for the coordinates.
The coordinates of some extent in $R^3$ are written as $(x,y,z)$.
If $mathbf{v}$ is a vector discipline, then $mathbf{v}=(a,b,c)$ the place $a$, $b$
and $c$ are three scalar fields referred to as the elements of $mathbf{v}$.
We denote the partial derivatives of a operate utilizing subindices, for
instance $a_y:=frac{partial a}{partial y}$.
1.2. Differential operators
The gradient of a operate $u$ is a vector discipline $nabla u$ outlined
by
[
nabla u = left(
u_x , u_y , u_z
right)
]
The divergence of a vector discipline $mathbf{u}=(a,b,c)$ is a scalar
discipline $mathrm{div}(mathbf{u})$ outlined by
[
mathrm{div}(mathbf{u}) =
a_x + b_y + c_z
]
The curl of a vector discipline $mathbf{u}=(a,b,c)$ is one other vector
discipline $mathrm{curl}(mathbf{u})$ outlined by
[
mathrm{curl}(mathbf{u}) =
left(
c_y – b_z , a_z – c_x , b_x – a_y
right)
]
Lastly, the laplacian of a scalar discipline $u$ is the scalar discipline
$Delta u$ outlined by
[
Delta u = u_{xx} + u_{yy} + u_{zz}.
]
Discover that, apart from the curl, all these operations will be outlined in
$R^N$. Nonetheless, the curl is restricted to a few dimensions. There’s a
related operator in two dimensions, which we name additionally the curl and computes
a scalar discipline $mathrm{curl}(mathbf{u})$ from a vector discipline
$mathbf{u}=(a,b):R^2toR^2$
[
mathrm{curl}(mathbf{u}) = b_x – a_y
]
Discover that it’s the final part of the 3D curl.
The curl can be outlined in dimension 7.
Let $mathbf{u}=(u^1,ldots,u^7)$ be a vector discipline in $R^7$, then
[
defcurlco#1#2#3#4#5#6{
{u^{#1}}_{#2}-{u^{#2}}_{#1}+
{u^{#3}}_{#4}-{u^{#4}}_{#3}+
{u^{#5}}_{#6}-{u^{#6}}_{#5}
}
mathrm{curl}(mathbf{u}) =
left(
begin{matrix}
curlco{2}{4}{3}{7}{5}{6}
curlco{3}{5}{4}{1}{6}{7}
curlco{4}{6}{5}{2}{7}{1}
curlco{5}{7}{6}{3}{1}{2}
curlco{6}{1}{7}{4}{2}{3}
curlco{7}{2}{1}{5}{3}{4}
curlco{1}{3}{2}{6}{4}{5}
end{matrix}
right)
]
the place a sub-index $i$ denotes a partial spinoff within the $i$-th dimension
of $R^7$. And analogously we are able to outline the 6-dimensional curl by taking the final part (leading to a scalar discipline).
1.3. Differential identities and properties
A very powerful identification is $Delta u = mathrm{div}(mathrm{grad}(u))$,
that can be utilized additionally because the definition of $Delta$.
Different identities involving the curl are
$mathrm{curl}(nabla u)=0$ and
$mathrm{div}(mathrm{curl}(mathbf{u}))=0$.
The capabilities $u$ such that $nabla u=0$ on $R^3$ are the constants.
The vector fields $mathbf{v}$ such that $mathrm{curl}(mathbf{v})=0$ are
referred to as conservative, irrotational or integrable.
They’re of the shape $mathbf{v}=nabla u$ for some operate $u$ referred to as the
potential of $mathbf{v}$.
The vector fields $mathbf{v}$ such that $mathrm{div}(mathbf{v})=0$ are referred to as
divergence-free, volume-preserving, solenoidal or
incompressible.
They’re of the shape $mathbf{v}=mathrm{curl}(mathbf{u})$ for some vector
discipline $mathbf{u}$ referred to as the
vector potential of $mathbf{v}$.
The scalar fields $u$ such that $Delta u=0$ are referred to as
harmonic capabilities.
The next identities are speedy purposes of the product rule for
derivatives:
[ nabla(fg) = fnabla g + gnabla f ]
[ mathrm{div}(fmathbf{g}) = fmathrm{div}(mathbf{g}) + mathbf{g}cdotnabla f ]
1.4. Integral calculus
The divergence theorem:
[
int_Omega mathrm{div}(mathbf{g}) =
int_{partialOmega}mathbf{g}cdotmathbf{ds}
]
Combining the divergence theorem with the product rule we get hold of the
integration by elements system.
[
int_{partialOmega} fmathbf{g}cdotmathbf{ds} =
int_Omega
fmathrm{div}(mathbf{g})
+
int_Omega
mathbf{g}cdotnabla f
]
Thus, if at the least one of many two capabilities vanishes on the boundary of
$Omega$
[
0=
int_Omega
fmathrm{div}(mathbf{g})
+
int_Omega
mathbf{g}cdotnabla f
]
or, in one other notation
[
leftlangle
f, mathrm{div}(mathbf{g})
rightrangle
=
leftlangle
-nabla f, mathbf{g}
rightrangle
]
thus that the operators $mathrm{div}$ and $-nabla$ are adjoint to every
different. Integrating by elements twice we get hold of that the operator $Delta$ is
self-adjoint.
2. Graphs and their matrices
A graph is $G=(V,E)$ the place $V$ is a set referred to as the
vertices of $G$, and $E$ is a subset of $Vtimes V$ referred to as the
edges of $G$.
We assume all the time that the set $V$ is finite, and its components are numbered
from $1$ to $n$. Thus, the set $E$ can be finite (the cardinal is at most
$n^2$) and we assume that the weather of $E$ are numbered from $1$ to $m$.
$displaystylebegin{matrix}
V = {1,2,3,4,5,6}
E = { {1,2},{1,3},{2,4},{3,4},{4,5},{5,6},{4,6} }
finish{matrix}$
2.1. The adjacency checklist
Given a graph of $n$ vertices and $m$ edges,
the adjacency checklist is a matrix
of $m$ rows and $2$ columns that accommodates the pairs of vertices related by
every edge. The entries of this matrix are integers on the set
${1,ldots,n}$. Thus, if the $okay$-th row is $(i,j)$, because of this edge
$okay$ connects vertices $i$ to $j$.
$
textrm{adjacency checklist} =
start{pmatrix}
1 & 2
1 & 3
2 & 4
3 & 4
4 & 5
5 & 6
4 & 6
finish{pmatrix}
$
The adjacency checklist is a really environment friendly illustration for sparse graphs (the place
the variety of edges is proportional to the variety of vertices).
Nonetheless, it’s not very attention-grabbing from the algebraic perspective.
We are going to see within the following three different matrices which have a really wealthy
algebraic interpretation.
2.2. The adjacency matrix $A$
Given a graph of $n$ vertices and $m$ edges,
the adjacency matrix is a sq. matrix $A=a_{ij}$ of dimension $ntimes n$.
The entries of $A$ are zeros and ones, with $a_{ij}=1$ if there may be an edge
from $i$ to $j$ and $a_{ij}=0$ in any other case.
$
A =
start{array}lllllll
Vbackslash V
& 1 & 2 & 3 & 4 & 5 & 6
hline
1 & 0 & 1 & 1 & 0 & 0 & 0
2 & 1 & 0 & 0 & 1 & 0 & 0
3 & 1 & 0 & 0 & 1 & 0 & 0
4 & 0 & 1 & 1 & 0 & 1 & 1
5 & 0 & 0 & 0 & 1 & 0 & 1
6 & 0 & 0 & 0 & 1 & 1 & 0
finish{array}
$
Discover that this matrix has considerably much less data than the adjacency
checklist, as a result of the ordering of the perimeters is misplaced. Thus, there’s a distinctive approach
to compute the adjacency matrix from the checklist, however many $m!$ other ways
to get the checklist from the matrix. We will selected an arbitrary canonical
ordering of the perimeters (for instance, in lexicographic order).
2.3. The Laplacian matrix $L$
Let $A$ be the adjacency matrix of a graph $G$.
If we sum the values of all the weather of the $i$-th row, we get hold of the
variety of edges going out of vertex $i$ (referred to as the diploma of the sting).
Allow us to put the vector with all of the levels within the diagonal of a matrix $D$; in
octave/matlab notation $mathtt{D=diag(sum(A))}$.
The Laplacian matrix of $G$ is outlined as
[
L = A – mathtt{diag}(mathtt{sum}(A))
]
Within the typical case the place $A$ is symmetric with 0 on the diagonal, the matrix
L is identical as A with minus the diploma of every vertex on the diagonal
entries.
$
L =
start{array}lllllll
Vbackslash V
& 1 & 2 & 3 & 4 & 5 & 6
hline
1 &-2 & 1 & 1 & 0 & 0 & 0
2 & 1 &-2 & 0 & 1 & 0 & 0
3 & 1 & 0 &-2 & 1 & 0 & 0
4 & 0 & 1 & 1 &-4 & 1 & 1
5 & 0 & 0 & 0 & 1 &-2 & 1
6 & 0 & 0 & 0 & 1 & 1 &-2
finish{array}
$
2.4. The incidence matrix $B$
Given a graph of $n$ vertices and $m$ edges,
the incidence matrix is an oblong matrix $B=b_{ij}$ of $m$ rows and $n$
columns. The entries of $B$ are zeros, ones and minus ones given by the
edges of the graph: if the $okay$-th edge goes from $i$ to $j$, then, on the
$okay$th row there are values $-1$ and $1$ on positions $i$ and $j$
respectively; there are zeros all over the place else.
$
B =
start{array}lllllll
Ebackslash V
& 1 & 2 & 3 & 4 & 5 & 6
hline
1 &-1 & 1 & 0 & 0 & 0 & 0
2 &-1 & 0 & 1 & 0 & 0 & 0
3 & 0 &-1 & 0 & 1 & 0 & 0
4 & 0 & 0 &-1 & 1 & 0 & 0
5 & 0 & 0 & 0 &-1 & 1 & 0
6 & 0 & 0 & 0 & 0 &-1 & 1
7 & 0 & 0 & 0 &-1 & 0 & 1
finish{array}
$
Discover that the incidence matrix accommodates the identical data because the
adjacency checklist (together with the order of the perimeters).
There’s an attention-grabbing relationship between the incidence matrix and the
Laplacian matrix, that may be checked algebraically:
[
L = -B^TB
]
This identification is the discrete analogue of $Delta=mathrm{div grad}$,
as we’ll clarify beneath.
2.5. The unsigned incidence matrix $C$
The incidence matrix $B$ outlined above is signed, on every row there are two
non-zero entries whose values are $-1$ and $1$. Thus the sum of any row is
zero.
We will write the matrix $B$ as $B=B_1-B_0$, the place the matrices
$B_0$ and $B_1$ have solely zeros and ones, with a single non-zero entry per
row.
It is going to be helpful later to think about the unsigned incidence matrix
$C$, outlined as $C=frac{1}{2}(B_0 + B_1)$, or equivalently
$C=frac{1}{2}|B|$. The rows of the matrix $C$ sum to at least one.
The next relations are speedy to confirm
[
A = 2C^TC-B^TB/2
]
[
mathrm{deg} = 2C^TC+B^TB/2
]
the place $mathrm{deg}$ is an $ntimes n$ diagonal matrix, whose values are the
levels of every vertex.
3. Vector calculus on graphs
Many of the constructions that we have now described on the vector calculus
reminder above have a direct correspondence within the case of graphs.
3.1. Analogies
The correspondence between vector calculus and graph idea is specified by
the next desk.
The principle thought is that scalar fields correspond to
capabilities outlined on vertices, and vector fields correspond to capabilities
outlined on edges.
Vector calculus |
Graph idea |
Base house |
Graph vertices $V$ |
Tangent house |
Graph edges $E$ |
$u:OmegatoR$ |
$u:VtoR$ |
$mathbf{v}:OmegatoR^3$ |
$mathbf{v}:EtoR$ |
Laplacian operator $Delta$ |
Laplacian matrix $Linmathcal{M}_{n,n}(R)$ |
gradient operator $nabla$ |
incidence matrix $Binmathcal{M}_{m,n}(R)$ |
divergence operator $mathrm{div}$ |
matrix $-B^Tinmathcal{M}_{n,m}(R)$ |
$Delta=mathrm{div grad}$ |
$L=-B^T B$ |
scalar discipline $u$ |
$uinR^n$ |
vector discipline $mathbf{v}$ |
$mathbf{v}inR^m$ |
vector discipline $nabla u$ |
$BuinR^m$ |
scalar discipline $Delta u$ |
$LuinR^n$ |
scalar discipline $mathrm{div}(mathbf{v})$ |
$-B^Tmathbf{v}inR^n$ |
directional spinoff $nabla |
$nabla u (a,b)$ |
$OmegasubseteqR^3$ |
$Omegasubseteq V$ |
$partialOmegasubseteqR^3$ |
$partialOmegasubseteq E$ , outlined as $partialOmega=Ecap(OmegatimesOmega^c)$ |
$displaystyleint_Omegamathrm{div}(mathbf{v}) |
$displaystylesum_{ainOmega}mathrm{div}(mathbf{v})(a) = sum_{einpartialOmega}mathbf{v}(e)$ |
Elliptic PDE $Delta u = f$ |
Linear system $Lu=f$ |
Parabolic PDE $u_t = Delta u$ |
First-order Linear ODE System $u_t=Lu$ |
$textrm{div}(Dnabla u),qquad |
$-B^TDBu,qquad Dinmathcal{M}_{m,m}$ |
$gDelta u,qquad g:OmegatoR$ |
$GLu,qquad Ginmathcal{M}_{n,n}$ |
pointwise product $u v$ |
Hadamard product $fodot g$ |
pointwise product $umathbf{v}$ |
Hadamard product $Cfodot g$ |
$nabla fg=fnabla g + gnabla f$ |
$B(fcirc g)=Cfodot Bg + Cgodot Bf$ |
(nothing) |
unsigned incidence matrix $Cinmathcal{M}_{m,n}(R)$ |
The $mathrm{curl}$ operator can’t be outlined on normal graphs, nevertheless it
will be outlined on planar graphs, and it satisfies related identities.
3.2. The graph Laplacian
The only operator of vector calculus is the Laplacian, remodeling
scalar fields into scalar fields. It’s the easiest as a result of no vector fields
are concerned, solely scalar fields.
Correspondingly, the best operator for graphs can be the Laplacian,
remodeling capabilities outlined on vertices into capabilities outlined on
vertices. It’s the easiest as a result of no capabilities outlined on edges are
concerned. As soon as we have now chosen an ordering of the vertices, a scalar discipline is
merely a vector $uinR^n$, and the Laplacian operator is outlined by a sq.
matrix of dimension $ntimes n$.
Let $G=(V,E)$ be a graph and $u:VtoR$ be a scalar discipline. The
Laplacian
of $u$ is denoted by $Delta u$ and is outlined because the scalar discipline $Delta u:VtoR$
[
Delta u(a) := sum_{(a,b)in E} u(b)-u(a)
]
Discover that the sum is carried out for a set vertex $a$, and $b$ varies
by way of all of the neighbors of $a$ within the graph.
The scalar discipline $u$ | The scalar discipline $Delta u$ |
Simply from the definition, we are able to deduce a number of properties of the laplacian
- The sum of all of the values of $Delta u$ is all the time zero
- If $u(a)$ is an area most, then $Delta u(a)<0$
- If $u(a)$ is an area minimal, then $Delta u(a)>0$
- If $u$ is fixed, then $Delta u$ is zero
If we repair an ordering of the vertices, then the scalar fields $u$ and $Delta
u$ are two vectors in $R^n$, and the linear operator $umapstoDelta u$ is
given by the matrix $L=A-mathtt{diag}(mathtt{sum}(A))$. This follows
straight by decomposing the definition of $Delta$ into two sums:
[
Delta u(a)
=
sum_{(a,b)in E}
u(b)
–
sum_{(a,b)in E}
u(a)
=
–
u(a)mathrm{degree}(a)
+sum_{(a,b)in E}
u(b)
]
Discover that the Laplacian has a pleasant interpretation.
If we regard the values of $u$ as a amount distributed on the vertices of
the graph, the situation $Delta u = 0$ says that the amount is distributed
evenly, or in equilibrium: the quantity of amount at every vertex equals the
common quantity over its neighbours. Particularly, if $u$ is fixed then
$Delta u = 0$.
Discover that the matrix $L$ is all the time singular: a relentless vector is an
eigenvector of eigenvalue 0. If the graph has $okay$ related elements, then
we have now null vectors which are fixed on every related part, thus the
matrix $L$ has rank $n-k$.
3.3. Graph gradient and graph divergence
Recall that scalar fields are capabilities outlined on vertices and vector fields
are capabilities outlined on edges. Thus, the gradient transforms a operate
outlined on vertices right into a operate outlined on edges. There’s a very
pure approach of doing that: the worth at every edge is obtained because the
distinction between the values at either side of the sting.
The scalar discipline $u$ | The vector discipline $nabla u$ |
Extra formally,
let $G=(V,E)$ be a graph and $u:VtoR$ be a scalar discipline.
The gradient of $u$ is the vector discipline $nabla u:EtoR$ outlined by
[
nabla u(a,b) := u(b) – u(a)
qquad mathrm{for} (a,b)in E
]
The matrix of this linear map is the incidence matrix $B$ of the graph.
Consider the gradient $nabla u(a,b)$ because the directional spinoff of $u$
at level $a$ within the route of the vector from $a$ to $b$.
Now let $mathbf{v}:EtoR$ be a vector discipline.
The divergence of $mathbf{v}$ is the scalar discipline
$mathrm{div}(mathbf{v}):VtoR$ outlined by:
[
mathrm{div}(mathbf{v})(a)
:=
sum_{(a,b)in E}mathbf{v}(a,b)
-sum_{(b,a)in E}mathbf{v}(b,a)
qquad mathrm{for} ain V
]
The matrix of this linear map is minus the transposed incidence matrix of the
graph $-B^T$.
Discover that the identification $Delta=mathrm{div grad}$ is trivial from the
definitions above, since each side are precisely $sum_{(a,b)in E}u(b)-u(a)$.
Thus, $L=-B^TB$.
3.4. Graph curl
We don’t want curls for our software, however allow us to say some phrases about
them.
These graph-theoretic analogues are simpler to grasp once we use
differential geometry as an alternative of plain vector calculus. In that case, the
discrete analogue of $okay$-forms are capabilities outlined over the $okay$-cliques of the
graph. Then the outside spinoff is instantly constructed for all values of $okay$,
and it accommodates the gradient, curl and divergence as explicit instances.
The particularity of third-dimensional manifolds comes from the truth that in
that in that case 1-forms and 2-forms have
the identical dimension and might each be interpreted as “vector fields”, thus the
curl operator is outlined from the outside spinoff
$d:Omega^1toOmega^2$. Within the case of graphs, we can’t basically determine capabilities
outlined on edges to capabilities outlined on triangles, besides in a single explicit
case: when the graph is a triangulation. In that case, there’s a
building that permits to outline the curl of a vector discipline as a vector
discipline, by traversing the 2 triangles at either side of an edge. The identification
$mathrm{curl grad}=0$ is then the sum of 6 values that cancel pairwise, and
so on. See the attractive papers of Oliver Knill for a complete protection
of this.
3.5. Graph subsets and their boundaries
It’s typically essential to cope with subset of graphs (for instance, once we
wish to interpolate a operate which is understood solely over some vertices).
In an effort to do algebra with them, we mannequin subsets as diagonal operators that
include the indicator operate of the subset because the diagonal entries. This
mannequin is used for subsets of vertices and subsets of edges.
Notations:
Let $X={1,ldots,n}$ (or any finite ordered set) and $Ysubseteq X$.
Let $a$ be a vector of size $n$ and $A$ a
matrix of dimension $ntimes n$ . We use the next, considerably ambiguous,
abuses of notation:
$mathrm{diag}(A)inR^n$:
the vector with the weather on the diagonal of $A$
$mathrm{diag}(a)inR^{ntimes n}$:
the diagonal matrix whose diagonal is $a$.
$1_YinR^{n}$: the indicator vector of the subset $Y$
$Y=mathrm{diag}(1_Y)inR^{ntimes n}$: the diagonal operator of $Y$
This final notation may be very handy in picture processing, as a result of it
represents point-wise multiplication by a binary picture as a linear operator
(with the identical identify because the binary picture). The $mathrm{diag}$ operator has
the identical semantics as that of octave/matlab.
Let $G=(V,E)$ be a graph with $n$ vertices and $m$ edges, and let
$Omegasubseteq V$.
To keep away from introducing new letters, we denote additionally by
$Omega=omega_{ij}$ the $ntimes n$ matrix that accommodates the indicator
operate of this set in its diagonal: $w_{ii}=1$ if $iin V$ and $w_{jj}=0$
in any other case.
Discover that the matrix $I-Omega$ corresponds to the
complementary set $Omega^c$.
We outline the boundary of a subset of vertices $Omegasubseteq V$ as
the subset of edges $partialOmegasubseteq E$ that go from $Omega$ to
$Omega^c$. Discover that $partialOmega=Ecap(OmegatimesOmega)$ in set
notation. Since $partialOmega$ is a subset of edges, it corresponds to a
diagonal matrix, additionally named $partialOmega$, of dimension $mtimes m$.
In matrix notation we have now
[partialOmega=mathrm{diag}(Bmathrm{diag}(Omega))] the place $B$ is the
incidence matrix of the graph. We will additionally write
$displaystyle1_{partialOmega}=B1_Omega$.
3.6. Equations on graphs
Now that we have now described the differential and boundary operators operator
in matrix type, it’s speedy to put in writing the discrete analogues of a number of
linear PDE. That is very lovely as a result of the analytic properties of the
corresponding PDE are recovered by elementary linear algebra.
3.6.1.
Laplace equation on the entire graph:
[
Lu=0
]
If the graph is related, the matrix $L$ has rank $n-1$ thus its kernel is
one-dimensional, equivalent to the fixed options $u=c$.
3.6.2.
Poisson equation on the entire graph, with information $f:VtoR$:
[
Lu=f
]
has a novel resolution except $f$ is fixed.
3.6.3.
Laplace equation on a subset $Omegasubseteq V$, with Dirichlet
boundary situations
$f:Omega^ctoR$:
[
Omega Lu + (I-Omega)(u-f)=0
]
Discover that that is written as an $ntimes n$ linear system, nevertheless it has a
diagonal half equivalent to the values of $u$ outdoors of $Omega$.
Discover additionally that the values of $f$ on the vertices that don’t have any neighbors in
$Omega$ solely seem within the diagonal half. The values of $f$ inside $Omega$
don’t seem in any respect (are cancelled out).
3.6.4.
Laplace equation on a subset $Omegasubseteq V$,
with Neumann boundary situations
$g:partialOmegatoR$:
[
Omega Lu + (partialOmega)(nabla u – g)=0
]
Or equivalently, by creating the boundary and gradient operators,
[
left[Omega L + mathrm{diag}(Bmathrm{diag}(Omega))Bright]u =mathrm{diag}(Bmathrm{diag}(Omega)) g
]
or, in another notation
[
(mathrm{diag}(1_Omega) L + mathrm{diag}(B1_Omega))B)u
=mathrm{diag}(B1_Omega) g
]
3.6.5.
Warmth equation on the entire graph with preliminary situation $u_0:VtoR$:
[
begin{cases}
u_t & =Lu
u(0) & = u_0
end{cases}
]
It is a system of $n$ first-order linear differential equations with
fixed coefficients. It has a closed-form resolution utilizing the matrix
exponential $u=e^{tL}u_0$.
3.6.6.
Warmth equation with supply time period $f:VtoR$ and preliminary situation
$u_0:VtoR$
[
begin{cases}
u_t & =Lu+f
u(0) & = u_0
end{cases}
]
It has likewise a closed-form resolution $u=e^{tL}(u_0-L^{-1}f)-L^{-1}f$.
Discover that $L^{-1}f$ solely is smart when $f$ will not be a relentless vector.
3.6.7.
Different mixtures are attainable, and simple to infer from the less complicated instances:
Poisson and Warmth equation on subsets with varied boundary situations, and so on.
3.7. Riemannian graph geometry
The isotropic case of “anisotropic” diffusion in picture processing is
modelled by phrases of the shape $gDelta u$, the place $g$ is a positive-real valued
operate on $Omega$. Within the case of graphs, the operate $g$ corresponds to
a scalar discipline $g:VtoR$, which we affiliate to a diagonal $ntimes n$
matrix $tilde g$ with the values of $g$. Then these phrases turn out to be $tilde gL
u$ within the corresponding discrete mannequin.
Really anisotropic diffusion comes from phrases of the shape
$mathrm{div}(Dnabla u)$, the place the diffusivity $D$ is a discipline of
positive-definite symmetric matrices outlined over $Omega$. Within the case of
graphs, we have now a matrix $tilde D$, which can be diagonal, however now of
dimension $mtimes m$. Then these phrases turn out to be $mathrm{div}(Dnabla u)$ within the
discrete mannequin. Or, in matrix type, $B^TDBu$.
3.8. Algebraic graph integral calculus
Integral calculus will be generalized readily to graphs. Integrals are
changed by sums over a finite area, and the varied identities of integral
calculus (e.g., the divergence theorem) turn out to be speedy matrix identities.
Let $G=(V,E)$ be a graph with $V={1,ldots,n}$ and $E={1,ldots,m}$
Let $Omegasubseteq V$ and let $f:VtoR$ be a scalar discipline.
The integral of $f$ over $Omega$ is outlined as
[
int_Omega f=sum_{pin Omega}f(p)
]
in matrix notation we have now
(
int_Omega f := mathrm{sum}(Omega f).
)
Discover that right here $f$ is a vector of size $n$, $Omega$ is an $ntimes n$
matrix, and we’re computing the sum of all of the elements of the vector
$Omega f$ to acquire a single quantity.
Discover that $f$ should be outlined all over the place, however solely the values inside
$Omega$ are used; thus, we may have outlined $f$ solely inside $Omega$.
An interface inside a graph is outlined as a set of edges $Ssubseteq
E$. Given a vector discipline $mathbf{v}:EtoR$ we outline the circulation
of $mathbf{v}$ by way of $S$
[
int_S mathbf{vcdot ds} := sum_{ein S}mathbf{v}(e)
]
or, in matrix notation, $int_S mathbf{vcdot ds}=mathrm{sum}(tilde S
mathbf{v})$ the place $tilde S$ is the diagonal matrix containing the indicator
operate of $S$.
An attention-grabbing explicit case occurs when $S$ is the boundary of some
area $Omega$. We’ve seen above that the matrix $tilde S$ is then
equal to $mathrm{diag}(Bmathrm{diag}(Omega))$. This statement results in the graph divergence
theorem that claims that
[
int_{partialOmega} mathbf{vcdot ds} =int_Omegamathrm{div}(mathbf{v})
]
or, in matrix notation,
[
1_Omegacdot(B^Tmathbf{v})
=
(B1_Omega)cdotmathbf{v}
]
which is strictly the identical factor, written in a different way.