1/45
Vocabulary practice flashcards covering error definitions, vector and matrix norms, condition numbers, eigenproblems, numerical differentiation, Runge-Kutta IVP solvers, numerical stability, and finite difference BVPs.
True Percent Relative Error (εt)
The exact relative error calculated when the true analytical solution is known: εt=true valuetrue value−approximation×100%
Approximate Percent Relative Error (εa)
An error estimate used in iterative numerical methods when the true value is unknown: εa=present approximationpresent approximation−previous approximation×100%
Round-off Error
Numerical error arising from the computer representation of numbers with a finite number of significant digits, which decreases as the step size increases.
Truncation Error
Numerical error resulting from approximating or truncating exact mathematical operations or continuous quantities in numerical methods, which generally increases as the step size increases.

Minimum Error Point (Sweet Spot)
The optimal step size where total numerical error reaches its minimum, formed by the balance between increasing truncation error and decreasing round-off error.
Norm
A mathematical operator that takes an object (such as a vector or matrix) and maps it to a single non-negative real number representing its magnitude, length, or size.
Euclidean Norm (Vector 2-Norm)
The geometric length of an n-dimensional vector u, defined via the dot product as ∥u∥2=∑i=1nui2
Vector p-Norm (∥u∥p)
The generalized norm of an n-dimensional vector u, defined by the formula ∥u∥p=(∑i=1n∣ui∣p)1/p
Vector 1-Norm (∥u∥1)
A specific vector norm computed as the sum of the absolute values of all vector components: ∥u∥1=∑i=1n∣ui∣
Vector Infinity Norm (∥u∥∞)
A vector norm equal to the maximum absolute value among all elements of the vector: ∥u∥∞=max1≤i≤n∣ui∣
Frobenius Norm (∥A∥F)
A matrix norm quantifying the overall size of an n×n matrix A, defined as the square root of the sum of squares of all entries: ∥A∥F=∑i=1n∑j=1naij2
Column-Sum Norm (Matrix 1-Norm)
A matrix norm determined by taking the maximum of the sums of absolute values across each column: ∥A∥1=maxj∑i=1n∣aij∣
Row-Sum Norm (Matrix Infinity Norm)
A matrix norm determined by taking the maximum of the sums of absolute values across each row: ∥A∥∞=maxi∑j=1n∣aij∣
Condition Number (κ(A))
A diagnostic measure for linear systems Ax=b, defined as κ(A)=∥A∥A−1, which bounds how relative errors in b propagate to relative errors in x.
Well-Conditioned Matrix
A matrix with a condition number κ(A) that is small (close to 1), ensuring that the relative error in x is not much larger than the relative error in b.
Ill-Conditioned Matrix
A matrix characterized by a large condition number κ(A), which indicates that small perturbations in b can lead to much larger relative errors in the solution x.
Eigenvalue and Eigenvector
A scalar λ and a corresponding nonzero vector x that satisfy the matrix equation Ax=λx, or equivalently (A−λI)x=0.
Power Method
An iterative numerical method used to approximate the dominant (largest in magnitude) eigenvalue and its associated eigenvector by repeated matrix-vector multiplication and normalization using the infinity norm.
Inverse Power Method
A numerical algorithm adapted from the power method to compute the smallest eigenvalue and its corresponding eigenvector of a matrix.
Taylor Series Expansion
An infinite series approximating a function f(x) around an expansion point xi, which is mathematically exact for an nth-order polynomial when using n+1 terms and produces a truncation error of order O(Δxn+1).

Forward Finite Difference (First Derivative)
A first-order accurate (O(h)) numerical differentiation stencil estimating f′(xi) using values at xi and the forward point xi+1.

Backward Finite Difference (First Derivative)
A first-order accurate (O(h)) numerical differentiation stencil estimating f′(xi) using values at xi and the preceding point xi−1.

Centered Finite Difference (First Derivative)
A second-order accurate (O(h2)) numerical differentiation stencil estimating f′(xi) using the symmetric span between xi+1 and xi−1, which quarters its truncation error when the step size h is halved.
Centered Finite Difference (Second Derivative)
A second-order accurate numerical approximation for the second derivative defined as dx2d2T=Δx2Ti+1−2Ti+Ti−1

Forward Euler Method (RK1)
A first-order accurate (O(Δt)) explicit single-step method for IVPs that computes the solution at ti+1 by extrapolating linearly along the tangent slope evaluated at (ti,yi).

Mid-point Method (RK2)
A second-order accurate (O(Δt2)) Runge-Kutta scheme comprising 1 predictor stage (computing the slope at the half-interval ti+1/2) and 1 corrector stage (advancing from yi to yi+1 using the midpoint slope).

Fourth-Order Runge-Kutta Method (RK4)
A fourth-order accurate (O(Δt4)) numerical method for initial value problems that computes 4 slope evaluations per time step (consisting of 3 predictors and 1 corrector) combined in a weighted average.
Numerical Accuracy vs. Numerical Stability
Accuracy refers to how closely a numerical solution approximates the exact solution at each step, while stability describes whether numerical errors decay or amplify over successive computational steps.
Amplification Factor (A)
A scalar multiplier relating the error or solution amplitude between consecutive steps (yi+1=Ayi); stability requires ∣A∣<1 to ensure errors diminish over time.
Conditional Stability (Explicit/Forward Euler)
A stability property where explicit methods require the step size to be strictly bounded, specifically Δt<∣λ∣2 for dtdy=λy, to keep ∣A∣=∣1+Δtλ∣<1.
Unconditional Stability (Implicit/Backward Euler)
A stability property where a method satisfies the error decay criterion (∣A∣<1) for all positive step sizes Δt, as seen in Backward Euler for dtdy=λy with λ<0.
Ghost Point
A fictitious node (such as T0) introduced outside a physical boundary to allow centered finite difference approximation of a derivative boundary condition while maintaining second-order accuracy.
Neumann Boundary Condition
A boundary condition in a differential equation problem that prescribes the derivative of the dependent variable at the boundary (e.g., T′(0)=0).
Dirichlet Boundary Condition
A boundary condition in a differential equation problem that fixes the exact value of the dependent variable at the boundary (e.g., T(10)=400K ).
Step size error
(O(h)→O(21h) )
(O(h2)→O(41h2) )
Order of accuracy is not the same as size of error
A (second-order) central difference has error O(Δx²), which you can write out as O((L/(N−1))²). The “trick” is that the error depends on the step size, not on N alone. N and L both feed into it.
How someone could be tricked
Confusing “order” with “N.” A question might say “the method is O(Δx²)” and ask for the order of accuracy in terms of N. The answer is not “order N.” In terms of grid points it’s O(1/N²) for fixed L, but the order of accuracy itself is still 2. The order is the exponent on Δx, nothing else.
Forgetting the domain length. Two people use the same N = 101 points, one on L = 1 and one on L = 100. Same N, same “second-order” method, but Δx is 0.01 versus 1. The second error is roughly 10,000 times larger (since 100² = 10⁴). Same method, same N, wildly different accuracy. Someone who thinks “more points = more accurate” without checking L gets burned.
The N vs. N − 1 trap. N points give N − 1 intervals, so Δx = L/(N−1), not L/N. It’s an easy off-by-one mistake, and it matters when N is small.
Thinking a higher-order method is automatically better. An O(Δx²) scheme with a large Δx can easily be worse than an O(Δx) scheme with a tiny Δx. Order tells you how fast the error shrinks as you refine, not how small it is at a given grid.
Doubling N doesn’t exactly halve Δx. Going from N to 2N gives Δx = L/(2N−1), not L/(2N−2). To cleanly halve Δx you need to go from N to 2N − 1 points. This matters if you’re doing a convergence study and expecting the error to drop by exactly 4×.
Accuracy
how closely a numerical solution approximates the exact solution at each step
Stability
describes whether numerical errors decay or amplify over successive computational steps
Stable
errors diminish over time
Unstable
errors grow
Central difference, second derivative
(Δx)2(f(xi+1)−2f(xi)+f(xi−1))≈dx2d2f
O(Δx2) (second order accuracy)
Central difference, first derivative
(2Δx)(f(xi+1)−f(xi−1))≈dxdf
O(Δx2) (second order accuracy)
Forward difference, first derivative
Δx(f(xi+1)−f(xi))≈dxdf
O(Δx) (first order accuracy)
Backward difference, first derivative
Δx(f(xi)−f(xi−1))≈dxdf
O(Δx) (first order accuracy)
Halving Error (RK1 euler, RK2 midpoint, RK4)
For error of order p, halving Δx reduces the error by a factor of about 2ᵖ
RK1: error ∝Δt → ½ t → (1/2) error (drops by factor of 2)
RK2: error ∝Δt2 → ½ t → (1/2)² = ¼ error (drops by factor of 4)
RK4: error ∝Δt4 → ½ t → (1/2)^4 = 1/16 error (drops by factor of 16)
Ghost Point Example (Neumann BC)
Use central finite difference to approximate
Rearrange equation for unknown node
\frac{dy_{ghost}}{\differentialD x}=\frac{y_{i+1}-y_{i-1}}{2\Delta x}=BC\to y_{i+1}=y_{i-1}+2BC\Delta x
substitute into discretized equation