Boundary value problems — shooting method, finite difference

WHAT is a BVP?

WHY do these arise? Steady-state physics: temperature along a rod with both ends fixed, deflection of a beam clamped at both ends, electrostatic potential between two plates. Nothing evolves in time — the answer must satisfy constraints everywhere at once.

METHOD 1 — Shooting Method

HOW (the algorithm):

Guess $s_0$ for the slope $y'(a)$ . Solve the IVP $\to$ get $y(b;s_0)$ .
Guess $s_1$ . Solve again $\to y(b;s_1)$ .
Use the secant method (we usually can't differentiate $\phi$ cheaply): $s_{n+1} = s_n - \phi(s_n)\,\frac{s_n - s_{n-1}}{\phi(s_n)-\phi(s_{n-1})}.$
Repeat until $|\phi(s_n)|<\text{tol}$ .

Worked example Worked shooting (linear) —

$y''=y,\ y(0)=0,\ y(1)=\sinh 1$ Exact solution is $y=\sinh x$ , true slope $y'(0)=1$ . Let's pretend we don't know it.

Shot 1: $s_0=0$ . IVP $y''=y, y(0)=0, y'(0)=0 \Rightarrow y\equiv 0$ , so $y(1;0)=0$ . Why this step? Test the lowest guess to bracket behaviour.
Shot 2: $s_1=2$ . By linearity $y=2\sinh x$ , so $y(1;2)=2\sinh 1\approx 2.3504$ . Why this step? Get a second point so we know the line $\phi(s)$ .
Target $\beta=\sinh1\approx1.1752$ . Interpolate: $s^\star = 0 + (2-0)\frac{1.1752-0}{2.3504-0}=2(0.5)=1.0.$ Why this step? Solve the straight-line root → the true slope $1$ , exactly, in 2 shots.

METHOD 2 — Finite Difference Method (FDM)

HOW — derive the difference formulas from Taylor series (first principles):

Grid: $x_i = a+ih$ , $h=\dfrac{b-a}{N}$ , $i=0,\dots,N$ . Let $y_i\approx y(x_i)$ .

Expand: $y_{i+1}=y_i+hy_i'+\tfrac{h^2}{2}y_i''+\tfrac{h^3}{6}y_i'''+\dots$ $y_{i-1}=y_i-hy_i'+\tfrac{h^2}{2}y_i''-\tfrac{h^3}{6}y_i'''+\dots$

Subtract → odd terms survive, $y_i''$ cancels: $y_{i+1}-y_{i-1}=2hy_i'+\tfrac{h^3}{3}y_i'''+\dots \;\Rightarrow\; y_i' \approx \frac{y_{i+1}-y_{i-1}}{2h}\quad (O(h^2)).$

Add → odd terms cancel, isolate $y_i''$ : $y_{i+1}+y_{i-1}=2y_i+h^2 y_i''+\tfrac{h^4}{12}y_i''''+\dots \;\Rightarrow\; y_i'' \approx \frac{y_{i+1}-2y_i+y_{i-1}}{h^2}\quad (O(h^2)).$

Apply to a linear BVP $y'' = p(x)y' + q(x)y + r(x)$ . Substitute stencils at each interior $x_i$ : $\frac{y_{i+1}-2y_i+y_{i-1}}{h^2}=p_i\frac{y_{i+1}-y_{i-1}}{2h}+q_i y_i + r_i.$ Multiply by $h^2$ and group $y_{i-1},y_i,y_{i+1}$ : $\underbrace{\left(1+\tfrac{h}{2}p_i\right)}_{\text{lower}}y_{i-1}\;\underbrace{-\,(2+h^2 q_i)}_{\text{diag}}\,y_i\;+\underbrace{\left(1-\tfrac{h}{2}p_i\right)}_{\text{upper}}y_{i+1}=h^2 r_i.$

This is a tridiagonal linear system $A\mathbf{y}=\mathbf{d}$ with $y_0=\alpha,\ y_N=\beta$ moved to the right side. Solve in $O(N)$ by the Thomas algorithm (tridiagonal Gaussian elimination).

Figure — Boundary value problems — shooting method, finite difference

Worked example Worked FDM —

$y''=y,\ y(0)=0,\ y(1)=\sinh1$ , take $N=2$ ( $h=0.5$ ) Here $p=0,\ q=1,\ r=0$ . Only one interior node $x_1=0.5$ . Stencil: $(1+0)y_0 -(2+h^2\cdot1)y_1+(1-0)y_2 = 0$ . Why this step? Plug $p=0$ , $q=1$ into the grouped formula. $y_0 - (2+0.25)y_1 + y_2 = 0.$ With $y_0=0,\ y_2=\sinh1=1.1752$ : $0 - 2.25\,y_1 + 1.1752 = 0 \Rightarrow y_1 = \frac{1.1752}{2.25}=0.5223.$ Why this step? Boundary values are known, so the single equation gives $y_1$ directly. Check: true $\sinh(0.5)=0.5211$ . Error $\approx 0.0012$ with just 2 intervals — and it shrinks $\propto h^2$ .

Shooting vs Finite Difference

	Shooting	Finite Difference
Idea	Guess slope, root-find	Solve all nodes at once
Reuses	IVP solver (RK4)	Linear algebra (Thomas)
Nonlinear BVP	Easy (secant outside)	Needs Newton on system
Stiff/long domain	Can blow up	Stable
Accuracy	RK4 order	$O(h^2)$ (central)

Common mistake Steel-manned common errors

1. "Shooting always converges in 2 steps." Why it feels right: the linear shortcut IS exact in 2 shots. The fix: that's only for linear ODEs. For nonlinear ones $\phi(s)$ is curved → you must iterate secant/Newton until $|\phi|<$ tol.

2. Using a forward difference $y'\approx\frac{y_{i+1}-y_i}{h}$ for $y'$ in FDM "to keep it simple." Why it feels right: it's a valid derivative approximation. The fix: it's only $O(h)$ , dragging your whole scheme down to first order even though $y''$ was $O(h^2)$ . Use the central difference to match orders.

3. Forgetting to move boundary values to the RHS. Why it feels right: $y_0,y_N$ are "just numbers." The fix: at $i=1$ , the term $(1+\tfrac h2 p_1)y_0$ is known — subtract it to the right side $d_1$ . Same at $i=N-1$ . Otherwise your matrix has the wrong size.

4. Picking $N$ too small and trusting the answer. Why it feels right: you got a number. The fix: halve $h$ ; error should drop $\approx 4\times$ . If it doesn't, your stencil/setup is wrong (Forecast-then-Verify!).

Recall Feynman: explain to a 12-year-old

Imagine you must throw a ball so it lands exactly in a far-off bucket. You don't know the right throwing angle. Shooting: throw once (too short), throw again (too far), then split the difference — keep adjusting the angle until it lands in the bucket. Finite difference: instead of throwing, you draw the ball's path as lots of tiny connected dots and force the dots to obey the rule "each dot is the average-ish of its neighbours," then solve the dot-puzzle so the first and last dots sit where they must. Both find the same curve — one by aiming, one by puzzle-solving.

Flashcards

What distinguishes a BVP from an IVP?

BVP gives one condition at each endpoint (

y(a),y(b)

); IVP gives all conditions (

y,y'

) at a single point, so IVP can march but BVP cannot.

Core idea of the shooting method?

Guess the missing slope

y'(a)=s

, solve the resulting IVP, treat landing error

\phi(s)=y(b;s)-\beta

, and root-find

\phi(s)=0

(usually via secant).

Why does the linear shooting shortcut give the answer in exactly 2 shots?

For a linear ODE,

y(b;s)

is linear in

s

, so

\phi(s)

is a straight line determined exactly by two points — linear interpolation hits the root with no iteration.

Secant update formula for shooting?

s_{n+1}=s_n-\phi(s_n)\dfrac{s_n-s_{n-1}}{\phi(s_n)-\phi(s_{n-1})}

Central difference for

y'

and its order?

y_i'\approx\dfrac{y_{i+1}-y_{i-1}}{2h}

, accurate to

O(h^2)

Central difference for

y''

and its order?

y_i''\approx\dfrac{y_{i+1}-2y_i+y_{i-1}}{h^2}

, accurate to

O(h^2)

(the

+1,-2,+1

stencil).

How do you derive the

y''

stencil?

Add the Taylor expansions of

y_{i+1}

and

y_{i-1}

; odd-power terms cancel, leaving

y_{i+1}+y_{i-1}=2y_i+h^2y_i''+O(h^4)

, then solve for

y_i''

Tridiagonal FDM coefficients for

y''=p y'+qy+r

Lower

1+\tfrac h2 p_i

, diagonal

-(2+h^2q_i)

, upper

1-\tfrac h2 p_i

, RHS

h^2 r_i

Which algorithm efficiently solves the FDM system and at what cost?

The Thomas algorithm (tridiagonal Gaussian elimination),

O(N)

When prefer FDM over shooting?

Stiff problems or long domains where small slope errors in shooting blow up; FDM keeps errors local and is more stable.

Why is using a forward difference for

y'

a bad idea in a central-difference FDM?

It's only

O(h)

, degrading the whole scheme from second order to first order.

What must you do with

y_0

and

y_N

when assembling the FDM matrix?

Move their (known) contributions to the right-hand side vector at the first and last interior equations.

Connections

Initial Value Problems — shooting reduces a BVP to repeated IVPs.
Runge-Kutta methods — the inner IVP solver for shooting.
Root finding — Secant and Newton — the outer loop of shooting.
Taylor series — source of every finite-difference stencil and its error order.
Tridiagonal systems — Thomas algorithm — solves the FDM linear system.
Truncation error and order of accuracy — why $O(h^2)$ matters.
Partial Differential Equations — finite differences — same stencils in 2D (Laplace/Poisson).

Concept Map

Hinglish (regional understanding)

Intuition Hinglish mein samjho

Dekho, BVP (boundary value problem) ka matlab hai ki tumhe do alag-alag ends par condition di gayi hai — jaise rod ke dono sire ka temperature, ya $y(0)$ aur $y(1)$ . IVP me sab kuch ek hi point par milta hai isliye tum aage march kar sakte ho, lekin BVP me start ka slope $y'(a)$ pata hi nahi hota, to direct march nahi chalega. Yahin do techniques kaam aati hain.

Shooting method ekdum cannon aim karne jaisa hai. Tum missing slope $s=y'(a)$ ka ek guess maaro, us guess se IVP solve karo (RK4 vagaira se), aur dekho ki $y(b)$ kahan gira. Agar target $\beta$ se kam aaya to slope badhao, zyada aaya to ghatao. Yeh adjusting kaam secant method se hota hai. Mast baat: agar ODE linear hai, to sirf do shots me exact answer aa jata hai, kyunki landing $y(b)$ slope ka straight-line function hota hai.

Finite difference thoda alag soch hai — yahan tum aim nahi karte, balki poore domain ko chhote-chhote grid points me kaat dete ho. Har point par derivative ko Taylor series se nikla hua formula laga dete ho: $y''\approx\frac{y_{i+1}-2y_i+y_{i-1}}{h^2}$ (yaad rakho: beech me minus two!). Isse ek tridiagonal linear system ban jata hai jo Thomas algorithm se fatafat solve hota hai. Saare points ek saath solve hote hain, isliye yeh stiff problems me shooting se zyada stable rehta hai.

Exam tip: stencil ka sign mat bhulo, boundary values $y_0,y_N$ ko RHS me shift karo, aur central difference hi use karo (forward difference order kam kar deta hai). Accuracy $O

Boundary value problems — shooting method, finite difference

WHAT is a BVP?

METHOD 1 — Shooting Method

METHOD 2 — Finite Difference Method (FDM)

Shooting vs Finite Difference

Flashcards

Connections

Concept Map

Hinglish (regional understanding)

Go deeper — visual, from zero

Test yourself — Numerical Methods

Connections