Mathematical Foundations

Metric

  • Function \( d:K\times K \rightarrow \mathbb{R} \)
  • Characteristics:
    • Non-negativity: \( d(x,y)\geq 0 \)
    • Identity: \( x=y \Rightarrow d(x,y) = 0 \)
    • Symmetry: \( d(x,y) = d(y,x) \)
    • Triangular inequality: \( d(x,z)\leq d(x,y)+d(y,z) \)

Distance Functions

Minkowski Distance

  • Generalization of Manhatten/Euclidean distance
  • \( x, y \in \mathbb{R}^n \)
  • Parameter \( r \)
  • \( d_r(x, y) = (\sum_{1\leq i\leq n}|x_i-y_i|^r)^\frac{1}{r} \)
  • \( r=1 \): Manhattan distance
  • \( r=2 \): Euclidean distance

Mahalanobis distance

  • Distance between vectors
  • Often used in statistics
  • Covariance matrix \( \Sigma \)
  • \( D(x,y)=\sqrt{(x-y)^T * \Sigma^{-1} * (x-y)} \)

Transformation

  • Idea:
    • Transform from one domain into another (keeping all information)
    • See other things
  • Properties
    • Reversible
    • Information preserving (no loss)
  • Using input signal
    • Discrete function

Fourier Transformation

  • High-level feature
  • Transformation from original domain to frequency domain
  • Original domain
    • Images: Position space (image row)
    • Audio: Time domain
  • Decompose periodic function into sum of simple oscillating functions (sin, cos)
  • Lower frequency contain most information
  • Discrete Fourier Transform
    • For discrete functions (i.e., sequence of real values)
    • Sequence of \( n \) real numbers \( x_0,...,x_{n-1} \)
    • Transform to sequence of coefficients
      • \( a_0,...,a_{n/2} \) (for cos)
      • \( b_0,...,b_{n/2} \) (for sin)
      • \( a_k = \sum_{j=0}^{n/2} f(j)cos(kj) \)
    • Coefficients are amplitudes of sine and cosine waves (project signal into sine/cosive waves)
    • \( x_k = \sum_{i=0}^{n/2}a_l cos(2\pi\frac{lk}{n}) + \sum_{i=0}^{n/2}b_l sin(2\pi\frac{lk}{n}) \)