Proving a linear algebra equation

3 replies · 2K views · last reply
IlyaMath
Messages
3
Reaction score
0
I am having trouble proving that two multivariate formulas are equivalent. I implemented them in MATLAB and numerically they appear to be equivalent.

I would appreciate any help on this.

Prove A = B


A = (Σπ^-1 + Σy^-1)^-1 * (Σπ^-1*π + Σy^-1*y)

y = π+ X*β

Σπ =τ*Σ

Σy = X' * Σβ * X + ΣεB = (Σπ^-1 + P'*Σβ^-1*P)^-1 * (Σπ^-1*π + P'*Σβ^-1*q)

q = P*y

P = (X'*Σ^-1*X)^-1*X'*Σ^-1Assumptions

i) Σε is infinitesimally small. (If Σε is exactly zero, then Σy may not be invertible).

ii) N > F (If N = F, then the proof is trivial. If N < F then is probably not invertible and P is not defined.)Notation

A, B: Nx1

π: Nx1

y: Nx1

q: Fx1

β: Fx1

Σ: NxN

Σπ: NxN

Σy: NxN

Σβ: FxF

Σε: NxN

X: NxF

P: FxN

τ: 1x1
 
Attachments
  • Handwritten-style equation: Σ with subscript π, equals Σ with subscript y, with a bar over the second Σ.
    upload_2015-7-1_17-3-58.png
    314 bytes · Views: 598
  • upload_2015-7-1_17-3-58.png
    upload_2015-7-1_17-3-58.png
    677 bytes · Views: 596
  • Typeset equation: Σy = X Σβ X' + Σε, with Σy on the left and terms on the right involving X, Σβ, X transpose and Σε.
    upload_2015-7-1_17-3-58.png
    481 bytes · Views: 553
  • Equation y = π + Xβ, where y and π are N x 1 vectors, X is an N x F matrix and β is an F x 1 vector.
    upload_2015-7-1_17-3-58.png
    440 bytes · Views: 578
  • Equation: B = (Σ^-1 + P' Σβ^-1 P)^-1 (Σ^-1 π + P' Σβ^-1 q), with Σ^-1 the inverse of Σπ.
    upload_2015-7-1_17-3-58.png
    786 bytes · Views: 574
  • Equation q = Py written in plain text, with q on the left and P times y on the right.
    upload_2015-7-1_17-3-58.png
    337 bytes · Views: 590
"Sigma", [itex]\Sigma[/itex], here, is a matrix, not a summation symbol?
 
This problem comes from Bayesian statistics, where N is number of observations and F is number of factors. All of the Sigma (Σ) are variance-covariance matrices. The Greek characters after each Σ are meant to be subscripts.