In mathematics, block matrix pseudoinverse is a formula of pseudoinverse of a partitioned matrix . This is useful for decomposing or approximating many algorithms updating parameters in signal processing , which are based on least squares method.
Derivation
Consider a column-wise partitioned matrix:
[
A
,
B
]
,
A
∈
R
m
×
n
,
B
∈
R
m
×
p
,
m
≥
n
+
p
.
{\displaystyle [\mathbf {A} ,\mathbf {B} ],\qquad \mathbf {A} \in \mathbb {R} ^{m\times n},\qquad \mathbf {B} \in \mathbb {R} ^{m\times p},\qquad m\geq n+p.}
If the above matrix is full rank, the pseudoinverse matrices of it and its transpose are as follows.
[
A
,
B
]
+
=
(
[
A
,
B
]
T
[
A
,
B
]
)
−
1
[
A
,
B
]
T
,
{\displaystyle {\begin{bmatrix}\mathbf {A} ,&\mathbf {B} \end{bmatrix}}^{+}=([\mathbf {A} ,\mathbf {B} ]^{T}[\mathbf {A} ,\mathbf {B} ])^{-1}[\mathbf {A} ,\mathbf {B} ]^{T},}
[
A
T
B
T
]
+
=
[
A
,
B
]
(
[
A
,
B
]
T
[
A
,
B
]
)
−
1
.
{\displaystyle {\begin{bmatrix}\mathbf {A} ^{T}\\\mathbf {B} ^{T}\end{bmatrix}}^{+}=[\mathbf {A} ,\mathbf {B} ]([\mathbf {A} ,\mathbf {B} ]^{T}[\mathbf {A} ,\mathbf {B} ])^{-1}.}
The pseudoinverse requires (n + p )-square matrix inversion.
To reduce complexity and introduce parallelism, we derive the following decomposed formula. From a block matrix inverse
(
[
A
,
B
]
T
[
A
,
B
]
)
−
1
{\displaystyle \mathbf {(} [\mathbf {A} ,\mathbf {B} ]^{T}[\mathbf {A} ,\mathbf {B} ])^{-1}}
[citation needed
[
A
,
B
]
+
=
[
P
B
⊥
A
(
A
T
P
B
⊥
A
)
−
1
,
P
A
⊥
B
(
B
T
P
A
⊥
B
)
−
1
]
T
,
{\displaystyle {\begin{bmatrix}\mathbf {A} ,&\mathbf {B} \end{bmatrix}}^{+}=\left[\mathbf {P} _{B}^{\perp }\mathbf {A} (\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp }\mathbf {A} )^{-1},\quad \mathbf {P} _{A}^{\perp }\mathbf {B} (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp }\mathbf {B} )^{-1}\right]^{T},}
[
A
T
B
T
]
+
=
[
P
B
⊥
A
(
A
T
P
B
⊥
A
)
−
1
,
P
A
⊥
B
(
B
T
P
A
⊥
B
)
−
1
]
,
{\displaystyle {\begin{bmatrix}\mathbf {A} ^{T}\\\mathbf {B} ^{T}\end{bmatrix}}^{+}=\left[\mathbf {P} _{B}^{\perp }\mathbf {A} (\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp }\mathbf {A} )^{-1},\quad \mathbf {P} _{A}^{\perp }\mathbf {B} (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp }\mathbf {B} )^{-1}\right],}
where orthogonal projection matrices are defined by
P
A
⊥
=
I
−
A
(
A
T
A
)
−
1
A
T
,
P
B
⊥
=
I
−
B
(
B
T
B
)
−
1
B
T
.
{\displaystyle {\begin{aligned}\mathbf {P} _{A}^{\perp }&=\mathbf {I} -\mathbf {A} (\mathbf {A} ^{T}\mathbf {A} )^{-1}\mathbf {A} ^{T},\\\mathbf {P} _{B}^{\perp }&=\mathbf {I} -\mathbf {B} (\mathbf {B} ^{T}\mathbf {B} )^{-1}\mathbf {B} ^{T}.\end{aligned}}}
Interestingly, from the idempotence of projection matrix, we can verify that the pseudoinverse of block matrix consists of pseudoinverse of projected matrices:
[
A
,
B
]
+
=
[
(
P
B
⊥
A
)
+
(
P
A
⊥
B
)
+
]
,
{\displaystyle {\begin{bmatrix}\mathbf {A} ,&\mathbf {B} \end{bmatrix}}^{+}={\begin{bmatrix}(\mathbf {P} _{B}^{\perp }\mathbf {A} )^{+}\\(\mathbf {P} _{A}^{\perp }\mathbf {B} )^{+}\end{bmatrix}},}
[
A
T
B
T
]
+
=
[
(
A
T
P
B
⊥
)
+
,
(
B
T
P
A
⊥
)
+
]
.
{\displaystyle {\begin{bmatrix}\mathbf {A} ^{T}\\\mathbf {B} ^{T}\end{bmatrix}}^{+}=[(\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp })^{+},\quad (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp })^{+}].}
Thus, we decomposed the block matrix pseudoinverse into two submatrix pseudoinverses, which cost n - and p -square matrix inversions, respectively.
Note that the above formulae are not necessarily valid if
[
A
,
B
]
{\displaystyle [\mathbf {A} ,\mathbf {B} ]}
A
≠
0
{\displaystyle \mathbf {A} \neq 0}
[
A
,
A
]
+
=
1
2
[
A
+
A
+
]
≠
[
(
P
A
⊥
A
)
+
(
P
A
⊥
A
)
+
]
=
0
{\displaystyle {\begin{bmatrix}\mathbf {A} ,&\mathbf {A} \end{bmatrix}}^{+}={\frac {1}{2}}{\begin{bmatrix}\mathbf {A} ^{+}\\\mathbf {A} ^{+}\end{bmatrix}}\neq {\begin{bmatrix}(\mathbf {P} _{A}^{\perp }\mathbf {A} )^{+}\\(\mathbf {P} _{A}^{\perp }\mathbf {A} )^{+}\end{bmatrix}}=0}
Application to least squares problems
Given the same matrices as above, we consider the following least squares problems, which
appear as multiple objective optimizations or constrained problems in signal processing.
Eventually, we can implement a parallel algorithm for least squares based on the following results.
Column-wise partitioning in over-determined least squares
Suppose a solution
x
=
[
x
1
x
2
]
{\displaystyle \mathbf {x} ={\begin{bmatrix}\mathbf {x} _{1}\\\mathbf {x} _{2}\\\end{bmatrix}}}
[
A
,
B
]
[
x
1
x
2
]
=
d
,
d
∈
R
m
×
1
.
{\displaystyle {\begin{bmatrix}\mathbf {A} ,&\mathbf {B} \end{bmatrix}}{\begin{bmatrix}\mathbf {x} _{1}\\\mathbf {x} _{2}\\\end{bmatrix}}=\mathbf {d} ,\qquad \mathbf {d} \in \mathbb {R} ^{m\times 1}.}
Using the block matrix pseudoinverse, we have
x
=
[
A
,
B
]
+
d
=
[
(
P
B
⊥
A
)
+
(
P
A
⊥
B
)
+
]
d
.
{\displaystyle \mathbf {x} ={\begin{bmatrix}\mathbf {A} ,&\mathbf {B} \end{bmatrix}}^{+}\,\mathbf {d} ={\begin{bmatrix}(\mathbf {P} _{B}^{\perp }\mathbf {A} )^{+}\\(\mathbf {P} _{A}^{\perp }\mathbf {B} )^{+}\end{bmatrix}}\mathbf {d} .}
Therefore, we have a decomposed solution:
x
1
=
(
P
B
⊥
A
)
+
d
,
x
2
=
(
P
A
⊥
B
)
+
d
.
{\displaystyle \mathbf {x} _{1}=(\mathbf {P} _{B}^{\perp }\mathbf {A} )^{+}\,\mathbf {d} ,\qquad \mathbf {x} _{2}=(\mathbf {P} _{A}^{\perp }\mathbf {B} )^{+}\,\mathbf {d} .}
Row-wise partitioning in under-determined least squares
[Comment re. below: According to section "Derivation" above this method of calculating [A, B]+ for m>=n+p. Can it then be used for an underdetermined system where by definition m (size of x and equal to number of variables) > n+p (number of equations)?]
Suppose a solution
x
{\displaystyle \mathbf {x} }
[
A
T
B
T
]
x
=
[
e
f
]
,
e
∈
R
n
×
1
,
f
∈
R
p
×
1
.
{\displaystyle {\begin{bmatrix}\mathbf {A} ^{T}\\\mathbf {B} ^{T}\end{bmatrix}}\mathbf {x} ={\begin{bmatrix}\mathbf {e} \\\mathbf {f} \end{bmatrix}},\qquad \mathbf {e} \in \mathbb {R} ^{n\times 1},\qquad \mathbf {f} \in \mathbb {R} ^{p\times 1}.}
The minimum-norm solution is given by
x
=
[
A
T
B
T
]
+
[
e
f
]
.
{\displaystyle \mathbf {x} ={\begin{bmatrix}\mathbf {A} ^{T}\\\mathbf {B} ^{T}\end{bmatrix}}^{+}\,{\begin{bmatrix}\mathbf {e} \\\mathbf {f} \end{bmatrix}}.}
Using the block matrix pseudoinverse, we have
x
=
[
(
A
T
P
B
⊥
)
+
,
(
B
T
P
A
⊥
)
+
]
[
e
f
]
=
(
A
T
P
B
⊥
)
+
e
+
(
B
T
P
A
⊥
)
+
f
.
{\displaystyle \mathbf {x} =[(\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp })^{+},\quad (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp })^{+}]{\begin{bmatrix}\mathbf {e} \\\mathbf {f} \end{bmatrix}}=(\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp })^{+}\,\mathbf {e} +(\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp })^{+}\,\mathbf {f} .}
Instead of
(
[
A
,
B
]
T
[
A
,
B
]
)
−
1
{\displaystyle \mathbf {(} [\mathbf {A} ,\mathbf {B} ]^{T}[\mathbf {A} ,\mathbf {B} ])^{-1}}
[citation needed [original research?
(
A
T
A
)
−
1
,
(
B
T
B
)
−
1
,
(
A
T
P
B
⊥
A
)
−
1
,
(
B
T
P
A
⊥
B
)
−
1
.
{\displaystyle \quad (\mathbf {A} ^{T}\mathbf {A} )^{-1},\quad (\mathbf {B} ^{T}\mathbf {B} )^{-1},\quad (\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp }\mathbf {A} )^{-1},\quad (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp }\mathbf {B} )^{-1}.}
In a dense and small system, we can use singular value decomposition , QR decomposition , or Cholesky decomposition to replace the matrix inversions with numerical routines. In a large system, we may employ iterative methods such as Krylov subspace methods.
Considering parallel algorithms , we can compute
(
A
T
A
)
−
1
{\displaystyle (\mathbf {A} ^{T}\mathbf {A} )^{-1}}
(
B
T
B
)
−
1
{\displaystyle (\mathbf {B} ^{T}\mathbf {B} )^{-1}}
(
A
T
P
B
⊥
A
)
−
1
{\displaystyle (\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp }\mathbf {A} )^{-1}}
(
B
T
P
A
⊥
B
)
−
1
{\displaystyle (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp }\mathbf {B} )^{-1}}
See also
References
External links
Key concepts Problems Hardware Software
![{\displaystyle [\mathbf {A} ,\mathbf {B} ],\qquad \mathbf {A} \in \mathbb {R} ^{m\times n},\qquad \mathbf {B} \in \mathbb {R} ^{m\times p},\qquad m\geq n+p.}](/media/api/rest_v1/media/math/render/svg/3a030f9182d9cca0ac4763e79119e68ef3a40d73)
![{\displaystyle {\begin{bmatrix}\mathbf {A} ,&\mathbf {B} \end{bmatrix}}^{+}=([\mathbf {A} ,\mathbf {B} ]^{T}[\mathbf {A} ,\mathbf {B} ])^{-1}[\mathbf {A} ,\mathbf {B} ]^{T},}](/media/api/rest_v1/media/math/render/svg/0483d77b5ab79a969b9a757fa514bd6f36c7b137)
^{-1}.}](/media/api/rest_v1/media/math/render/svg/a27a0849a2ea247f3ca9dd8a76843967e3ce4e41)
![{\displaystyle \mathbf {(} [\mathbf {A} ,\mathbf {B} ]^{T}[\mathbf {A} ,\mathbf {B} ])^{-1}}](/media/api/rest_v1/media/math/render/svg/a49e5b1856cd8343f44ab5b927689bc2709a2522)
![{\displaystyle {\begin{bmatrix}\mathbf {A} ,&\mathbf {B} \end{bmatrix}}^{+}=\left[\mathbf {P} _{B}^{\perp }\mathbf {A} (\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp }\mathbf {A} )^{-1},\quad \mathbf {P} _{A}^{\perp }\mathbf {B} (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp }\mathbf {B} )^{-1}\right]^{T},}](/media/api/rest_v1/media/math/render/svg/22698cf8b77740c8140213a20c738d093dbb1347)
![{\displaystyle {\begin{bmatrix}\mathbf {A} ^{T}\\\mathbf {B} ^{T}\end{bmatrix}}^{+}=\left[\mathbf {P} _{B}^{\perp }\mathbf {A} (\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp }\mathbf {A} )^{-1},\quad \mathbf {P} _{A}^{\perp }\mathbf {B} (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp }\mathbf {B} )^{-1}\right],}](/media/api/rest_v1/media/math/render/svg/f3fd897104afd53aba6bf65e1ffcc6cd10b7e9a4)
![{\displaystyle {\begin{bmatrix}\mathbf {A} ^{T}\\\mathbf {B} ^{T}\end{bmatrix}}^{+}=[(\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp })^{+},\quad (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp })^{+}].}](/media/api/rest_v1/media/math/render/svg/0c48dc148a26040931dfa82bdc610b7a5bb63dad)
![{\displaystyle [\mathbf {A} ,\mathbf {B} ]}](/media/api/rest_v1/media/math/render/svg/02de22a76b03fc739a88a7b3ba82fd6c66f68521)
![{\displaystyle \mathbf {x} =[(\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp })^{+},\quad (\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp })^{+}]{\begin{bmatrix}\mathbf {e} \\\mathbf {f} \end{bmatrix}}=(\mathbf {A} ^{T}\mathbf {P} _{B}^{\perp })^{+}\,\mathbf {e} +(\mathbf {B} ^{T}\mathbf {P} _{A}^{\perp })^{+}\,\mathbf {f} .}](/media/api/rest_v1/media/math/render/svg/5ce5932ef6a113f52465dad3030056664b9877dc)
