Cholesky Decomposition

Cholesky Decomposition #

Cholesky decomposition is a special matrix factorisation used for symmetric positive definite matrices.

From lecture discussions, this decomposition is powerful because it reduces a matrix into a triangular form, making computations easier and more stable.

Key Idea: Cholesky decomposition expresses a matrix as a product of a lower triangular matrix and its transpose. It is efficient and numerically stable.


Definition #

For a symmetric positive definite matrix:

\[ A = L L^T \]

Where:

  • L is a lower triangular matrix
  • diagonal entries of L are positive

Conditions for Cholesky #

From lectures:

Matrix must be:

  1. Symmetric
\[ A = A^T \]
  1. Positive definite
\[ x^T A x > 0 \quad \forall x \neq 0 \]

If these conditions are not satisfied → Cholesky cannot be applied.


Why It Works (Lecture Insight) #

From lecture explanation:

  • Symmetry ensures consistent structure
  • Positive definiteness ensures square roots are valid
  • Triangular form simplifies solving systems

So instead of solving Ax = b directly, we solve:

\[ L y = b \] \[ L^T x = y \]

Step-by-Step Computation #

From slides:

To compute L:

  1. Fill entries row by row
  2. Use previously computed values
  3. Diagonal entries involve square roots

Example formula:

\[ l_{ii} = \sqrt{a_{ii} - \sum_{k=1}^{i-1} l_{ik}^2} \]

Example #

Given:

\[ A = \begin{bmatrix} 4 & 2 \\ 2 & 3 \end{bmatrix} \]

We compute L such that:

\[ A = L L^T \]

Geometric Interpretation #

From lecture:

  • Matrix defines quadratic form
  • Cholesky expresses it as squared transformation

This simplifies analysis of shapes like ellipses.


Why Cholesky is Useful #

From lecture and webinar:

1. Solving Linear Systems #

Efficient alternative to Gaussian elimination.


2. Numerical Stability #

  • Fewer operations
  • Reduced rounding errors

3. Machine Learning Applications #

  • Multivariate Gaussian sampling
  • Covariance matrices
  • Optimisation algorithms

Connection to Other Decompositions #

  • Similar to LU decomposition but specialised
  • More efficient when conditions are met
  • Used before SVD in some pipelines

Hidden Exam Pattern #

From lectures:

  • Questions focus on:
    • checking conditions
    • constructing L
    • solving systems

Common Mistakes #

  • Applying to non-symmetric matrix
  • Ignoring positive definiteness
  • Incorrect square root calculation
  • Mixing L and U forms

Strategy to Prepare #

  1. Check symmetry and positivity first
  2. Practice small matrix examples
  3. Memorise computation steps
  4. Understand application in solving systems

Quick Summary Table #

ConceptMeaning
A = LL^TCholesky decomposition
Conditionsymmetric + positive definite
Llower triangular
Useefficient solving

References #


Home | Matrix Decompositions