Skip to content
Tutorials Center
My learning

Topics

Lesson 46 of 62 ยท Eigenvalues and Decompositions

Polar Decomposition

Split any matrix into a rotation (or reflection) and a stretch with the polar decomposition A = UP, computed from the SVD and used for the nearest orthogonal matrix.

Every nonzero complex number can be written z = r ยท e^(iฮธ): a positive length r and a rotation e^(iฮธ). The polar decomposition says that every square matrix splits the same way. Any matrix A is a rotation (or reflection) followed by a stretch along perpendicular axes:

TEXT
A = U P

Here U is orthogonal (unitary if complex), playing the role of e^(iฮธ), and P is symmetric positive semi-definite, playing the role of r. When A is invertible, P is positive definite and the factorization is unique.

It is the cleanest answer to the question "what does this matrix do to space?", and it solves a practical problem: finding the nearest orthogonal matrix to a noisy or drifting rotation. It follows directly from the singular value decomposition, and uses the ideas of symmetric matrices and orthogonal matrices.

Computing it from the SVD

Start from A = W ฮฃ Vแต€, with W and V orthogonal and ฮฃ diagonal with the singular values. Insert Vแต€ V = I:

TEXT
A = W ฮฃ Vแต€ = (W Vแต€) (V ฮฃ Vแต€) = U P
U = W Vแต€        (a product of orthogonal matrices, hence orthogonal)
P = V ฮฃ Vแต€      (symmetric, eigenvalues = singular values >= 0)

Also P = sqrt(Aแต€A), the unique positive semi-definite square root of Aแต€A, because Aแต€A = V ฮฃยฒ Vแต€. That is the "length" part of the factorization, and its eigenvalues are the stretch factors along the perpendicular axes (the columns of V).

There is a mirrored left polar decomposition A = K U with K = W ฮฃ Wแต€ = sqrt(AAแต€), where the stretch comes after the rotation. Both use the same U.

The sign of the determinant. det A = det U ยท det P, and det P โ‰ฅ 0, so det U has the sign of det A. If det A > 0, U is a proper rotation; if det A < 0, U includes a reflection.

Polar decomposition at a glance

U Rotation or reflection. Lengths unchanged.
P Stretch along perpendicular axes.
From the SVD U = W V^T and P = V S V^T.
1 / 2
Python Polar decomposition of a 2 x 2 matrix, step by step

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

U here is a small clockwise rotation of about 11.3 degrees, the stretch P has eigenvalues 3.26 and 1.84 (the singular values of A), and the circle-to-ellipse test confirms that the extreme stretches are the singular values.

The nearest orthogonal matrix

If you have a matrix that should be a rotation but has been spoiled by noise or accumulated rounding error (a camera pose estimate, a rotation that drifted after thousands of multiplications), you can fix it by replacing it with its polar factor U. This U is the closest orthogonal matrix to A in the Frobenius norm: no other orthogonal Q makes โ€–A - Qโ€– smaller. It is the solution of the orthogonal Procrustes problem, used to align two point clouds in shape analysis, robotics and 3D scanning.

If you also need a proper rotation (det = +1), flip the sign of the last column of W when the determinant of U comes out -1 (the reflection case).

Python Repairing a drifted rotation matrix with the polar factor

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The polar factor removes all of the non-orthogonal error, including the large error built up by the repeated products, and it sits closer to the noisy matrix than the true rotation does (it is the nearest orthogonal matrix to the noisy input, not necessarily the original). The last line shows that the sign of det U matches det A.

Check your understanding

Quick check ยท 5 questions

of 5 answered You scored / 5 ยท perfect!

  1. 1. What are the two factors in the polar decomposition A = U P?

    U is orthogonal (rotation or reflection) and P is symmetric positive semi-definite (a stretch along perpendicular axes).
  2. 2. If A = W S V^T is an SVD, which formula gives the orthogonal factor U?

    A = (W V^T)(V S V^T), so U = W V^T and P = V S V^T.
  3. 3. What are the eigenvalues of P in the polar decomposition?

    P = V S V^T has the singular values of A as its eigenvalues.
  4. 4. A matrix has negative determinant. What does that say about U in its polar decomposition?

    det A = det U * det P, and det P is nonnegative, so det U has the sign of det A.
  5. 5. Which problem does the polar factor solve?

    Among all orthogonal matrices, the polar factor U minimizes the Frobenius distance to A (the orthogonal Procrustes problem).

Summary

  • Every square matrix factors as A = U P, a rotation or reflection U times a symmetric positive semi-definite stretch P, just as a complex number is r e^(iฮธ).
  • From the SVD A = W ฮฃ Vแต€: U = W Vแต€ and P = V ฮฃ Vแต€ = sqrt(Aแต€A); the left version is A = K U with K = sqrt(AAแต€).
  • The eigenvalues of P are the singular values, the circle maps to an ellipse with those semi-axes, and det U has the sign of det A.
  • U is the nearest orthogonal matrix to A: use it to repair drifting rotations and to align point clouds.

Next: Low-Rank Approximation, another application of the SVD.