This is a model-selection / approximation trade-off question. A clean way to think about it is:
Do I pay complexity once (single higher-order curve), or pay complexity repeatedly (many simple pieces)?
Below is a decision framework you can actually use.
1. Two competing models
A. Higher-order quadratic (global polynomial)
Example:
[
y = a_0 + a_1 x + a_2 x^2 + a_3 x^3 + \dots
]
Characteristics
- Single global function
- Smooth, differentiable everywhere
- Error spreads globally
- Sensitive to outliers (Runge phenomenon at higher degrees)
B. Series of lines (piecewise linear / spline-like)
Example:
[
y =
\begin{cases}
m_1 x + c_1 & x \in [x_0, x_1] \
m_2 x + c_2 & x \in [x_1, x_2] \
\vdots
\end{cases}
]
Characteristics
- Local control
- Errors are local
- Discontinuous derivatives (unless constrained)
- Very robust to noise
2. Cost model (this is the key)
Define total cost as:
[
\text{Total Cost} = \text{Model Cost} + \text{Error Cost}
]
Model Cost
| Aspect | Higher-order quadratic | Series of lines |
|---|---|---|
| Parameters | (O(d)) | (O(k)) (segments) |
| Training | (O(d^2)) | (O(k)) |
| Inference | constant | conditional branching |
| Maintenance | hard | easy |
| Interpretability | low | high |
Rule: cost grows quadratically with degree, linearly with number of segments.
Error Cost
| Error behavior | Polynomial | Piecewise linear |
|---|---|---|
| Local spikes | ❌ affects whole curve | ✅ contained |
| Boundary behavior | ❌ unstable | ✅ stable |
| Noise tolerance | ❌ poor | ✅ strong |
3. Error allowance → model choice
Let ε be acceptable max error.
Case 1: Low ε, high smoothness required
✔ Use higher-order quadratic if:
- Function is inherently smooth (physics, geometry)
- Error tolerance is uniform
- Derivatives matter
Examples:
- Motion equations
- Trajectories
- Sensor calibration curves
Case 2: Low ε, but structure is unknown or noisy
✔ Use series of lines if:
- Data is empirical
- Regime changes exist
- Outliers present
Examples:
- Pricing slabs
- Demand curves
- Supply–cost relations
Case 3: Medium ε, low cost preference
✔ Series of lines wins
- Easier to tune
- Easier to explain
- Lower risk
4. Bias–Variance lens (important)
| Model | Bias | Variance |
|---|---|---|
| High-order polynomial | Low bias | High variance |
| Piecewise linear | Moderate bias | Low variance |
If you fear overfitting → choose lines
If you fear underfitting → choose polynomial
5. Decision rule (practical)
Use higher-order quadratic when:
- You can guarantee smoothness
- Global behavior matters
- You have low noise
- You need derivatives
Use series of lines when:
- Data is empirical / operational
- Interpretability matters
- Errors must stay local
- Cost of mistakes is high
6. A very strong heuristic (engineering rule)
If you can explain the breakpoints → use lines
If you can explain the equation → use polynomial
7. Hybrid (often best)
Spline / quadratic per segment
- Low-degree polynomial per piece
- Continuous at joins
- Bounded complexity
This dominates both in many real systems.
8. One-line takeaway
Global smooth truth → polynomial
Local uncertain reality → piecewise linear
If you want, tell me:
- noise level
- interpretability importance
- whether derivatives matter
- cost of a bad prediction
and I’ll give you a numerical decision rule (even a formula).