Balancing equations

Minjae Park

Auburn University
MATH 2660 - Spring 2026

April 15, 2026

Overview

$$ % Colors

% Coordinate vectors and matrices

% Common sets

% Abstract vector symbols

% Norms / absolute value

% Optional: dot product spacing (looks nicer in slides)

% Operators $$

Main idea

• Today we explore one theme: local balance creates linear equations.
• First example: voltages on a resistor network.
• Second example: traffic flowing through a road network.
• In both cases, we write equations from conservation.
• Then MATLAB helps us solve and visualize the answer.

Plan

1. Build a voltage model on an irregular network
2. Solve and visualize the node voltages
3. Build a traffic model from flow balance
4. Solve and visualize the road flows
5. Briefly compare solver choices

Balancing models

Setting	Unknowns live on	Balance law
Voltage network	nodes	current in = current out
Traffic network	edges	flow in = flow out
Ecology	populations or masses in compartments	inflow = outflow at equilibrium
Finance	asset positions	portfolio cash flows = target obligations
Markov chain	probabilities on states	probability in = probability out at steady state
Tutte embedding	node coordinates	each interior point = average of its neighbors

Companion file

• The MATLAB script for today is scripts/balancing_equations_demo.m
• It contains:
  • an irregular resistor-network example
  • a small traffic-flow example
  • a short LU / Cholesky / QR comparison

Voltage

Voltage balance at an interior node

• Put a voltage \(u_i\) at each node.
• If edge \(i \sim j\) has resistance \(R_{ij}\), then the current from \(i\) to \(j\) is \[ \frac{u_i-u_j}{R_{ij}}. \]
• It is often cleaner to use the conductance \[ c_{ij} = \frac{1}{R_{ij}}. \]
• Fix the boundary voltages.
• At each interior node, Kirchhoff’s current law says \[ \sum_{j \sim i} c_{ij}(u_i - u_j) = 0. \]
• The unit-resistor case is just the special case \(c_{ij}=1\) for every edge.

Laplacian form

• Let \(W\) be the weighted adjacency matrix, where \[ W_{ij} = c_{ij}. \]
• Let \(D\) be the diagonal matrix of weighted degrees \[ d_i = \sum_j W_{ij}. \]
• Then the weighted Laplacian is \[ L = D - W. \]
• The balance equations are still \[ (Lu)_i = 0 \qquad \text{for interior nodes.} \]
• After splitting into boundary and interior vertices, \[ L_{II} u_I = -L_{IB} u_B. \]
• This is a linear system for the unknown interior voltages.

What do the subscripts mean?

• I means interior nodes.
• B means boundary nodes.
• We split the voltage vector into two pieces: \[ u = \begin{bmatrix} u_I \\ u_B \end{bmatrix} \]
• Here:
  • u_I = unknown interior voltages
  • u_B = fixed boundary voltages
• So the subscripts are just telling us which part of the network we are looking at.

Block matrix picture

\[ L = \begin{bmatrix} L_{II} & L_{IB} \\ L_{BI} & L_{BB} \end{bmatrix}, \qquad u = \begin{bmatrix} u_I \\ u_B \end{bmatrix} \]

• L_{II} means:
  • rows from interior equations
  • columns from interior unknowns
• L_{IB} means:
  • rows from interior equations
  • columns from boundary values
• So the interior rows of Lu = 0 become \[ L_{II}u_I + L_{IB}u_B = 0. \]
• Since u_B is already known, we move that term to the right side.

Tiny index example

• Suppose the boundary nodes are [1 2 5].

• Suppose the interior nodes are [3 4].

• Then MATLAB literally forms

  LII = L([3 4], [3 4]);
  LIB = L([3 4], [1 2 5]);

• So LII is the part of L that connects interior equations to interior unknowns.

• And LIB is the part that tells us how the boundary values influence those same interior equations.

Build an irregular network

mask = logical([
    1 1 1 1 0 0 0 0;
    1 1 1 1 1 1 0 0;
    1 1 1 1 1 1 1 0;
    1 1 1 1 1 1 1 1;
    0 1 1 1 1 1 1 1;
    0 0 1 1 1 1 1 1
]);

[m,n] = size(mask);
node_id = zeros(m,n);
coords = [];
N = 0;

for r = 1:m
    for c = 1:n
        if mask(r,c)
            N = N + 1;
            node_id(r,c) = N;
            coords(N,:) = [c, m-r+1];
        end
    end
end

• We start from a grid, then remove some nodes.
• The irregular shape makes the voltage pattern more interesting.

Assign edge resistances

s = edges(:,1);
t = edges(:,2);
x0 = coords(:,1);
y0 = coords(:,2);

resistance = zeros(length(s), 1);

for k = 1:length(s)
    if y0(s(k)) == y0(t(k))
        resistance(k) = 1;
    else
        resistance(k) = 2;
    end
end

conductance = 1 ./ resistance;

• Here we make a simple nonuniform example:
  • horizontal edges have resistance 1
  • vertical edges have resistance 2
• Lower resistance means larger conductance.
• If every resistance were 1, we would get the old unit-resistor model.

Build the weighted Laplacian

C = full(sparse(s, t, conductance, N, N));
C = C + C.';
deg = sum(C, 2);
L = diag(deg) - C;

• C(i,j) stores the conductance between nodes i and j.
• The diagonal entry deg(i) is the total conductance connected to node i.
• The formula L = D - C looks the same as before, but now the entries carry resistance information.

Solve the voltage system

G = graph(s, t);

% rows = interior equations, columns = interior unknowns
LII = L(interior, interior);

% rows = interior equations, columns = boundary values
LIB = L(interior, boundary);

uB = zeros(length(boundary), 1);
source_nodes = boundary(y0(boundary) == max(y0(boundary)));
uB(ismember(boundary, source_nodes)) = 1;

uI = LII \ (-LIB * uB);

u = zeros(N,1);
u(boundary) = uB;
u(interior) = uI;

• Top boundary segment: voltage \(1\)
• Rest of the boundary: voltage \(0\)
• MATLAB solves for the unknown interior values
• The only difference from the unit-resistor case is that L is now weighted.

Visualize the solution

figure;
plot(G, "XData", x0, "YData", y0, ...
        "NodeCData", u, "MarkerSize", 8, "LineWidth", 1.5);
title("Voltage on an irregular network");
colormap turbo;
colorbar;
axis equal off;

• The node colors show the solved voltage.
• The shape matters, and the edge resistances matter too.

Traffic

Traffic balance at an intersection

• Let \(x_k\) be the flow on road \(k\).
• At each intersection: \[ \text{flow in} = \text{flow out}. \]
• These are also linear equations.
• We may also use a few known road counts to determine the unknown flows.

A small road network

• Roads:
  • \(x_1\): In \(\to\) North
  • \(x_2\): In \(\to\) South
  • \(x_3\): North \(\to\) East
  • \(x_4\): South \(\to\) East
  • \(x_5\): North \(\to\) South
  • \(x_6\): East \(\to\) Out
• Conservation gives the interior balance equations.
• Then a couple of measured roads pin down the remaining freedom.

Traffic equations

\[ \begin{aligned} x_1 - x_3 - x_5 &= 0 \\ x_2 + x_5 - x_4 &= 0 \\ x_3 + x_4 - x_6 &= 0 \\ x_6 &= 120 \\ x_5 &= 20 \\ x_1 &= 70 \end{aligned} \]

• First three rows: conservation at intersections
• Last three rows: measured or prescribed road values

Matrix form and solve

Atraffic = [
    1 0 -1 0 -1 0;
    0 1  0 -1 1 0;
    0 0  1 1  0 -1;
    0 0  0 0  0  1;
    0 0  0 0  1  0;
    1 0  0 0  0  0
];

btraffic = [0; 0; 0; 120; 20; 70];
xtraffic = Atraffic \ btraffic;

• The unknown vector is \[ x = \begin{bmatrix} x_1 & x_2 & x_3 & x_4 & x_5 & x_6 \end{bmatrix}^T. \]
• This is the same Ax = b story in a different application.

Visualize the road flows

Gtraffic = digraph([1 1 2 3 2 4], [2 3 4 4 3 5]);
xy = [0 0; 2 1; 2 -1; 4 0; 6 0];

figure;
p = plot(Gtraffic, "XData", xy(:,1), "YData", xy(:,2), ...
         "LineWidth", 2, "ArrowSize", 14);
labeledge(p, Gtraffic.Edges.EndNodes(:,1), ...
             Gtraffic.Edges.EndNodes(:,2), ...
             compose("%.0f", xtraffic));
title("Solved traffic flows");
axis equal off;

• Now the unknowns live on the edges instead of the nodes.
• The edge labels show the solved traffic flow on each road.

Digression: more roads can make traffic worse

• A famous traffic-network surprise is the Braess paradox.
• Sometimes adding a new road makes congestion worse, not better.
• Why?
  • drivers choose selfishly
  • travel time depends on how many drivers use each road
• That is an interesting network idea, but it is not the same as today’s linear Ax = b traffic model.
• Let’s watch this nice explanation by Veritasium

Solvers

Same modeling habit, different matrix structure

Problem	Matrix feature	Good practical choice
Voltage on a graph	square, symmetric positive definite	\ or Cholesky
Repeated square solve	same matrix, new right-hand sides	LU
Noisy traffic counts	overdetermined least squares	QR
Nearly dependent equations	rank / conditioning concerns	SVD

First MATLAB habit

• Start with

  x = A \ b;

• This is much better than using inv(A).

• MATLAB inspects the matrix and chooses a good strategy internally.

• But it still helps to know what kind of matrix we have.

What solver makes sense here?

• For the voltage problem, \(L_{II}\) is symmetric positive definite as long as the resistances are positive.
• That makes Cholesky especially attractive.
• If the same square matrix is reused many times, LU is also a natural idea.
• If traffic data are noisy and we solve a least-squares problem instead, QR is the short answer.
• SVD is the tool to remember when rank or sensitivity becomes the issue.

Conclusion

Why this matters

• These are not abstract exercises.
• Engineers write linear systems from balance laws all the time.
• The physical meaning changes, but the linear algebra habit stays the same.
• The next step is a different question:
  • what if we do not want an exact solve?
  • what if we want to minimize prediction error instead?
• That leads naturally to least squares and gradient descent.

Summary

• Voltage and traffic both come from local conservation.
• In the voltage problem, the unknowns are node voltages.
• In the traffic problem, the unknowns are edge flows.
• Solver choice depends on the matrix structure.
• The next lecture shifts from solving equations to minimizing error.