Unit 8: Boundary Value Problems for Ordinary Differential Equations: Shooting & Finite Differences

1. Two-Point Boundary Value Problems (BVPs)

In contrast to Initial Value Problems (IVPs) where all auxiliary conditions are specified at a single point $t_0$, Boundary Value Problems (BVPs) specify physical constraints at two or more distinct spatial endpoints $x = a$ and $x = b$.

A general second-order ODE on $x \in [a, b]$ is given by:

subject to boundary conditions at both endpoints.

2. Classification of Boundary Conditions

Boundary conditions are categorized based on function values and spatial derivatives:

1. Dirichlet Boundary Conditions (First Kind):

Fixed values of the state variable at the endpoints:

Physical Example: Prescribed temperatures at the ends of a conducting metal rod.

2. Neumann Boundary Conditions (Second Kind):

Fixed values of the outward flux or derivative:

Physical Example: Insulated adiabatic rod ends (zero heat flux: $y'(a) = 0$).

3. Robin / Mixed Boundary Conditions (Third Kind):

Linear combinations of state values and derivatives:

with $a_0, a_1, b_0, b_1 \ge 0$ and $a_0 + a_1 > 0, b_0 + b_1 > 0$. Physical Example: Newton's law of cooling / convective heat transfer at the boundary.

3. Existence and Uniqueness for Second-Order BVPs

Unlike IVPs where Picard's theorem ensures existence locally, BVPs can have no solutions, a unique solution, or infinitely many solutions (e.g., resonance in vibrating beams).

Theorem (Existence and Uniqueness for BVPs): Consider the Dirichlet BVP $y'' = f(x, y, y'), y(a) = \alpha, y(b) = \beta$ on $[a, b]$. Suppose $f$ is continuous on $D = [a, b] \times \mathbb{R}^2$ and satisfies:

1. $\dfrac{\partial f}{\partial y}(x, y, y') > 0$ for all $(x, y, y') \in D$.
2. There exists a constant $M > 0$ such that $\left|\dfrac{\partial f}{\partial y'}(x, y, y')\right| \le M$ on $D$.

Then the boundary value problem has a unique solution $y \in C^2([a, b])$. Significance of Condition 1: Positivity $\partial f / \partial y > 0$ prevents eigenvalues from crossing zero, ruling out resonant standing wave states.

1. Transformation into Two Initial Value Problems

Consider a general linear second-order two-point BVP:

with Dirichlet boundary conditions $y(a) = \alpha, y(b) = \beta$.

By the principle of linear superposition, the general solution can be written as:

where:

1. $y_1(x)$ solves the non-homogeneous IVP:

1. $y_2(x)$ solves the homogeneous IVP:

2. Determining the Superposition Coefficient

By construction, at the initial boundary $x = a$:

which matches the left boundary condition automatically for any choice of $c \in \mathbb{R}$!

To satisfy the right boundary condition $y(b) = \beta$:

Assuming $y_2(b) \ne 0$ (which holds whenever $q(x) > 0$), we solve directly for $c$:

Computational Algorithm:

1. Convert the two second-order IVPs into two 2D first-order IVP systems.
2. Integrate both systems from $x = a$ to $x = b$ using high-order single-step methods (e.g. classical RK4).
3. Extract terminal values $y_1(b)$ and $y_2(b)$.
4. Compute $c = \frac{\beta - y_1(b)}{y_2(b)}$.
5. The initial slope is given directly by:

1. Synthesize the complete spatial solution:

1. Conceptual Framework of Nonlinear Shooting

For a nonlinear BVP:

superposition fails because the governing equation is nonlinear.

Instead, introduce an unknown initial slope parameter $s \in \mathbb{R}$ and define the parameter-dependent IVP:

Let $y(x; s)$ denote the solution to this IVP. We seek a value of $s$ such that the trajectory hits the target $\beta$ at the right boundary:

This is a scalar root-finding problem for $s$!

2. Newton-Raphson Iteration & Variational Sensitivity Equations

Applying Newton-Raphson to $F(s) = 0$:

To compute the derivative $\frac{\partial y}{\partial s}(x; s)$, define the variational sensitivity function:

Differentiating the governing ODE $y'' = f(x, y, y')$ with respect to the initial slope $s$:

Assuming smooth partial derivatives, interchange order of differentiation:

with initial conditions obtained by differentiating the boundary constraints:

Coupled 4D System for Nonlinear Shooting:

At each Newton iteration, integrate the coupled system of 4 first-order ODEs simultaneously from $x = a$ to $x = b$:

Then update:

Repeat until $|y_1(b) - \beta| < \text{TOL}$.

1. Discretization of the Domain

Instead of integrating forward in space like shooting methods, Finite Difference Methods (FDM) discretize the spatial domain $[a, b]$ simultaneously into $N$ subintervals of width $h = \frac{b - a}{N}$:

Let $w_i \approx y(x_i)$ denote the discrete numerical approximations.

Using second-order centered difference approximations:

2. The Linear Tridiagonal System

Consider the linear BVP:

Substituting centered differences at interior grid points $i = 1, 2, \dots, N-1$:

Multiplying through by $-h^2$ and collecting terms:

Set:

This forms an $(N-1) \times (N-1)$ tridiagonal linear system:

3. The Thomas Algorithm (TDMA)

The Tridiagonal Matrix Algorithm (TDMA) is a specialized $\mathcal{O}(N)$ Gaussian elimination method that bypasses the general $\mathcal{O}(N^3)$ complexity:

1. Forward Sweep (Eliminate lower diagonal $a_i$):

2. Backward Substitution:

Total Operation Count: Exactly $5(N-1)$ additions and multiplications—massively faster and numerically stable whenever $q(x) > 0$ (strict diagonal dominance).