Uniqueness and Existence of QR Decomposition
You'll learn when a QR decomposition exists for a matrix and what makes it unique (or not).
The Core Idea
When we decompose a matrix A into QR (where Q is orthogonal and R is upper triangular), two natural questions arise: Does this decomposition always exist? And if it does, is there only one way to do it?
The good news: QR decomposition exists for any m×n matrix where m ≥ n. We can always apply Gram-Schmidt to the columns of A to construct it.
The uniqueness question is more subtle.
Existence
Given an m×n matrix A with m ≥ n, QR decomposition always exists. The Gram-Schmidt process converts the columns of A into an orthonormal set (accounting for linear dependencies), which becomes the columns of Q. The coefficients from this process populate R.
If A has full column rank (its columns are linearly independent), then R will be invertible with positive diagonal entries.
Uniqueness (With a Catch)
If A has full column rank and we require R to have positive diagonal entries, then QR decomposition is unique.
Without this constraint, there's ambiguity: you can flip the sign of any column of Q and the corresponding row of R simultaneously, producing different valid QR decompositions:
- A = QR = (Q · D)(D · R) where D is diagonal with ±1 entries
Think of it like writing 6 as 2×3 or (-2)×(-3) — both are valid factorizations.
If A lacks full column rank, even with the positive diagonal convention, uniqueness can fail due to zero diagonal entries in R.
Key Takeaway: QR decomposition exists for any matrix, but is unique (up to orthogonal sign choices) only when the matrix has full column rank and we enforce positive diagonal entries in R.