Simplify a model while tracking the error regime

Try

Suppose x+εx²=1 with a small nonnegative dimensionless ε. At ε=0 the solution is x=1. Rather than immediately using a quadratic formula, seek a correction x=1+aε+O(ε²). The expansion aims to explain how the nonlinear term changes the solution, not merely to produce a decimal.

Analytical simplification is part of the breadth illustrated by Duke MATH577, Spring 2026. That advanced course is a methods comparator, not a claim that this lesson reproduces its full perturbation curriculum.

Predict

For positive ε, should the root near 1 move upward or downward? Since the extra term is positive, the linear part must become smaller. Predict the sign of a before matching coefficients.

Experiment

Substitution gives 1+(a+1)ε+O(ε²)=1, so a=−1. The approximation is x≈1−ε.

ε	Approximation	Residual x+εx²−1
1/10	9/10	−19/1000
1/20	19/20	−39/8000

The exact residual formula is −2ε²+ε³. Halving ε approximately quarters its leading magnitude. A residual is not automatically the root error: use the derivative 1+2εx, bounded away from zero near the positive root, to relate the two by a mean-value argument.

For 0≤ε≤1/2, the approximation 1−ε is positive, and the exact positive root lies between zero and one. Throughout the interval joining them, the derivative of x+εx²−1 is at least one. The mean-value theorem therefore bounds the absolute root error by the absolute residual, at most 2ε². This is a genuine uniform bound on the stated parameter interval, stronger than observing two favorable rows in a table. It does not justify using the approximation for arbitrary ε.

Understand

A regular perturbation remains informative throughout its stated region. A singular perturbation can change the effective equation or lose boundary information. For εy′+y=0 with y(0)=1, the exact solution is e^(−x/ε). Setting ε=0 suggests y=0, which is correct at each fixed positive x in the limit but fails at the boundary. The convergence is not uniform on an interval including zero.

The inner variable X=x/ε resolves the boundary layer: y=e^(−X). Its width is proportional to ε. This calculation explains why a small coefficient multiplying a derivative cannot always be discarded; the derivative itself can become large in a narrow region.

Similarity seeks a scale-invariant profile. For one-dimensional diffusion, try u=t^(−α)F(x/t^β) at positive time. Matching powers in uₜ=κuₓₓ gives 2β=1. If total mass is finite and conserved, its scaling t^(β−α) must be constant, so α=β=1/2. The resulting profile equation still needs solving and boundary or integrability conditions. Scaling exponents alone do not establish existence or uniqueness of the profile.

Apply

For one approximation, specify the small parameter, output, norm and domain where you expect accuracy. Calculate a residual and compare with an exact or independently refined reference. Include a boundary region where the approximation may fail. Separate an observed error trend from a proved uniform error bound.

Continue

A model’s state may differ from what a sensor records. Observation design comes next, because parameter estimation can be misleading when bias, aggregation or sampling is left outside the mathematical specification.