1. Nonlinear Models

ModelEquation
Exponential
Power
Saturation growth
Polynomial

The criterion is still least squares: minimize .


2. Polynomial Regression

Setup

Given points, fit with .

⚠️ With the polynomial would pass through every point, which is interpolation, not regression. Hence .

Derivation

For each :

Theorem – Normal equations ( system)


Entry is (rows/columns indexed from 0). Solve with Chapter 1’s methods.

Example (slides) – thermal expansion coefficient, 2nd order

Fit :

(°F)8040−40−120−200−280−340
(×10⁻⁶ in/in/°F)6.476.245.725.094.303.332.45

Sums ():


Thermal expansion α(T), 2nd-order polynomial fit-350-250-150-5050234567T (°F)α (×10⁻⁶ in/in/°F)α = a₀+a₁T+a₂T²data

⚠️ Slide typo: the first right-hand-side entry is printed as ; it’s .

⚠️ The matrix entries range from to , so the system is badly scaled. Keep full precision (a refit from the raw data gives , , ; the small differences come from rounding the sums).


3. Exponential Model – Direct Least Squares

Derivation



Rewritten:

Theorem – Solving for and

From the first equation:

Substituting into the second gives one nonlinear equation in :

Solve it numerically, e.g. by bisection (2.1), then compute .

Example (slides) – Technetium-99m radiation

(half of Tc-99m is gone in ≈ 6 h):

(h)013579
1.0000.8910.7080.5620.4470.355


Numerical solution (MATLAB on the slides; bisection on gives the same):


4. Transformation of Data (Linearization)

Method

Most nonlinear models lead to simultaneous nonlinear equations. Transforming the data turns them into a linear regression (4.1). For the exponential:

Let , , → (linear):

Other linearizations (extra)

ModelTransformLinear formBack to model
,
, ,
, ,

Example (slides) – Tc-99m again, by transformation

1010.000000.000000
210.891−0.11541−0.115411
330.708−0.34531−1.03599
450.562−0.57625−2.881325
570.447−0.80520−5.636449
690.355−1.0356−9.320781
25−2.8778−18.990165



Tc-99m: γ = A e^(λt)0481216202400.250.50.751half-life ≈ 6.0 ht (h)relative intensity γdirect LS: 0.9998e^(−0.1151t)ln-transform: 0.99974e^(−0.11505t)data

Half-life check (extra): h, consistent with “about 6 hours”.

⚠️ The transformed fit minimizes , not , so its constants differ slightly from the direct method ( vs , vs ).

⚠️ Slide typo: row 6 of the table lists ; it’s (consistent with and ).