A nonparametric method to estimate an unknown probability density function by placing smooth, scaled kernel functions at observed sample points and summing them.

Principle

Replace each sample point by a localized bump (kernel) with a bandwidth that controls smoothness; the sum approximates the underlying continuous density as sample size grows.

Demonstration

Given samples x_i and kernel K, the estimator at x is (1/nh) Σ_i K((x−x_i)/h); choices of K and bandwidth h determine bias–variance tradeoff.

Misapplication

Using an overly small bandwidth produces spurious multimodality and overfitting, while too large a bandwidth erases genuine structure and underfits.

Consequence

Provides a flexible, smooth estimate of density useful for visualization, mode detection, and as a building block in nonparametric inference and classification.

Reversal

Instead of smoothing discrete samples to obtain a density, decompose a smooth target density into kernel-weighted contributions to recover sample-generating components.

Boundary

Appropriate when data are samples from a continuous support and smoothing assumptions hold; problematic for mixed discrete–continuous data, high-dimensional spaces without dimensionality reduction, or boundary bias at support edges.

Semantic Tension

Competes with parametric density fitting when a specific model family is justified; KDE trades model parsimony for flexibility and requires careful bandwidth selection.

Synthesis

A simple, data-driven smoothing technique that reconstructs a continuous approximation of an unknown density by aggregating localized kernel contributions centered at sample observations.