Curve Fitting

What is Curve Fitting?

Curve Fitting is the process of finding a mathematical function that best represents a set of measured data points. In science and engineering, real-world measurements are always scattered — a ball dropped from a height does not follow a textbook parabola exactly because of air resistance, timing errors, and measurement noise. Curve fitting lets you cut through that scatter to reveal the underlying relationship. This simulation lets you plot data points, choose a model type (linear, quadratic, cubic, or sinusoidal), and observe how the algorithm adjusts the model's coefficients to minimize the sum of squared residuals — the vertical gaps between each data point and the fitted curve. The quality of fit is summarized by R², the coefficient of determination. In this simulation R² is displayed from 0 to 1; mathematically, un-clamped R² can be negative for fits worse than the mean-only model. Alongside R², inspecting the residual plot — a graph of those vertical gaps — reveals whether the model captures the true structure of the data or whether a systematic pattern remains. These skills transfer directly to lab reports, AP exam free-response questions, and real scientific practice.

Parameters explained

Fit Type

Selects the type of mathematical model applied to the data (integer index 0 through 4, corresponding to models such as linear y = mx + b, quadratic y = ax² + bx + c, cubic, and sinusoidal). Choosing the right model requires understanding what physical or mathematical relationship you expect in the data — fitting a quadratic to genuinely linear data will often improve R² slightly but wastes a parameter and may produce misleading predictions outside the measured range.

Data Noise(%)

Sets the magnitude of random scatter added to the data points, from 0% (perfectly noiseless, on-curve data) to 50% (heavily scattered). Real experimental data typically falls somewhere in between. Higher noise makes it harder to distinguish a good fit from a mediocre one using R² alone, and it makes the residual plot more important for diagnosing whether the chosen model is appropriate.

Data Points

Controls how many data points are plotted, from 5 to 50 (default 15). Fewer points make individual outliers more influential and R² less reliable. More points give the fitting algorithm more information, typically producing a more stable and trustworthy estimate of the model parameters. This parameter is available in the pro tier and illustrates why scientists collect multiple trials rather than relying on a single measurement.

Common misconceptions

• MisconceptionA high R² value means the model is correct and will make good predictions.

  CorrectR² measures how much variance in the data the model explains within the range of measured points. It does not guarantee the model is physically correct, does not protect against overfitting with unnecessary parameters, and says nothing about how well the model predicts outside the measured range (extrapolation). A cubic polynomial often achieves R² > 0.99 on noisy data that is genuinely linear simply because extra parameters absorb random scatter.

• MisconceptionCorrelation between two variables means one causes the other.

  CorrectA high R² or strong correlation tells you two variables move together in the data — it says nothing about why. Ice cream sales and drowning rates are historically correlated because both rise in summer; neither causes the other. Establishing causation requires controlled experiments, not just correlation. The residual plot and R² describe the mathematical relationship in the data, not the physical mechanism.

• MisconceptionAdding more parameters (higher polynomial degree) always gives a better model.

  CorrectMore parameters always improve R² on the training data — a degree-n polynomial can pass through n+1 points exactly. But overfitting produces a curve that wiggles through the noise rather than capturing the real trend, and its predictions between or beyond the measured points can be wildly wrong. Good model selection balances fit quality against model complexity, often using physical reasoning to decide how many parameters are justified.

• MisconceptionRandomly scattered residuals mean the fit is bad.

  CorrectRandomly scattered residuals are actually the hallmark of a good fit. They indicate that the model has captured the systematic trend in the data and that what remains is unpredictable measurement noise. The warning sign is systematic residuals — a U-shape, a sinusoidal wave, or a diagonal drift — which tells you the model is missing a real structural feature of the data.

How teachers use this lab

1. Linear vs. quadratic comparison: set fit_type to linear (index 0), data_noise to 10%, num_points to 15 and record R². Then switch fit_type to quadratic and record R² again. Students calculate the gain and debate whether the improvement justifies the extra parameter, practicing the principle of parsimony (Occam's Razor) in model selection.
2. Noise sensitivity investigation: fix fit_type to linear and num_points to 20. Sweep data_noise from 0% to 10% to 30% to 50%, recording R² at each level. Students plot noise vs. R² and observe how scatter degrades fit quality, building intuition about why repeated trials and careful measurement technique matter in real experiments.
3. Residual pattern diagnosis: set fit_type to linear but use a dataset generated from a quadratic relationship (set noise low). Students observe the U-shaped residual pattern and diagnose that a linear model is missing a quadratic term — then switch to quadratic and watch residuals randomize. Teaches the residual plot as a diagnostic tool, not just R².
4. Sample size and stability: fix fit_type to linear and data_noise to 20%. Compare num_points at 5 vs. 15 vs. 40 by noting how much the fitted line shifts between runs. Students recognize that small samples produce unreliable parameter estimates, connecting to the statistical concept of standard error and why scientists replicate measurements.
5. Sinusoidal fit for periodic data: switch fit_type to sinusoidal and increase data_noise to 15%. Ask students to predict what R² they expect when the underlying data was generated with periodic structure. Compare to a linear fit on the same data. This demonstrates that model choice matters as much as parameter tuning, and connects to Fourier analysis concepts in physics.