ORIGINAL PAPER
Least upper bound of truncation error of low‑rank matrix approximation algorithm using QR decomposition
with pivoting
Haruka Kawamura1 · Reiji Suda1
Received: 31 May 2020 / Revised: 24 January 2021 / Accepted: 1 February 2021 / Published online: 24 February 2021
© The Author(s) 2021
Abstract
Low-rank approximation by QR decomposition with pivoting (pivoted QR) is known to be less accurate than singular value decomposition (SVD); however, the calcula- tion amount is smaller than that of SVD. The least upper bound of the ratio of the truncation error, defined by ‖A−BC‖2 , using pivoted QR to that using SVD is proved to be √
4k−1
3 (n−k) +1 for A∈ℝm×n(m≥n) , approximated as a product of B∈ℝm×k and C∈ℝk×n in this study.
Keywords Error analysis · Pivoting · QR decomposition · Singular values Mathematics Subject Classification 65F55 · 15A45
1 Introduction
1.1 Low‑rank approximation
Low-rank matrix approximation involves approximating a matrix by a matrix whose rank is less than that of the original matrix. Let A∈ℝm×n ; then, a rank k approxima- tion of A is given by
where B∈ℝm×k and C∈ℝk×n . Low-rank matrix approximation appears in many applications such as data mining [5] and machine learning [14]. It also plays an important role in tensor decompositions [12].
A≈BC
* Reiji Suda
reiji@is.s.u-tokyo.ac.jp Haruka Kawamura kawamulahaluka@gmail.com
1 The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan
This paper discusses truncation errors of low-rank matrix approximation using QR decomposition with pivoting, or pivoted QR. In this study, rounding errors are not considered, and the norm used is basically 2-norm. A∈ℝm×n (without loss of generality, we assume that m≥n ) is approximated by a product of B∈ℝm×k and C∈ℝk×n , and the truncation error is defined by ‖A−BC‖2.
It is well-known that for any matrix A∈ℝm×n ( m≥n ), there are orthogonal matrices U∈ℝm×m and V ∈ℝn×n and a diagonal matrix 𝛴 ∈ℝn×n with nonnegative diagonal elements that satisfy
This is a singular value decomposition (SVD) of A. We define 𝜎
i(A) for i=1 , 2, ..., n, satisfying
and assume that 𝜎
1(A)≥𝜎
2(A)≥⋯≥𝜎
n(A)≥0 without loss of generality. The 𝜎 values are singular values of A. A has rank k if and only if 𝜎 i
k(A)>0=𝜎
k+1(A) . Let
Then,
holds [8]. Therefore, this is an A’s rank-k approximation whose 2-norm of truncation error is the smallest. We define the truncation error of low-rank approximation by SVD as
The amount of computation required to calculate SVD is O(nmmin(n,m)).
Pivoted QR was proposed by Golub in 1965 [7]. Because the amount of computa- tion required to calculate the low-rank approximation by pivoted QR is O(nmk), it is cheaper than SVD and hence useful in many applications such as solving rank- deficient least squares problems [2]. It consists of QR decomposition and pivoting.
For any matrix A, there exist Q∈ℝm×n and an upper triangular matrix R∈ℝn×n that satisfy A=QR and QTQ=In . This is a QR decomposition of A. We use pivoting to determine the permutation matrix 𝛱grd and apply the QR decomposition algorithm to A𝛱grd . The subscript grd signifies the greedy method, as explained previously.
Hereafter, we redefine QR as a QR decomposition of A𝛱grd=QR . Let Q and R be partitioned as
A=U (𝛴
O )
VT.
diag(𝜎
1(A),𝜎
2(A),…,𝜎n(A)) =𝛴,
𝛴k=diag(𝜎
1(A),𝜎
2(A),…,𝜎k(A)).
rank(X)min≤k‖A−X‖2=����
� A−U
�𝛴
k O
O O
� VT����
�2
=𝜎
k+1(A)
SVDk(A) =𝜎
k+1(A).
where Q1k∈ℝm×k and R1k∈ℝk×k. Then, we can approximate A to Q1k(
R1k R2k) 𝛱T
grd and
holds. We define the truncation error of low-rank approximation by pivoted QR as
In this study, the greedy method is used to make ‖R3k‖2 small in pivoting. Pivoting is performed such that the elements in R= (rij) satisfy the following inequalities [1, p.103]
Condition (1) is not used to analyze the error for l=k+1 , k+2 , ..., n−1.
The greedy method of pivoting is not always optimal. QR decompositions of A𝛱RR , where 𝛱RR is chosen such that RRR has a small lower right block and where QRRRRR is a QR decomposition of A𝛱RR , are called rank-revealing QR (RRQR). The following theorem was shown by Hong et al. in 1992 [9].
Theorem 1 Let m≥n>k, and A∈ℝm×n. Then, there exists a permutation matrix 𝛱 ∈ℝn×n such that the diagonal blocks of R=
(R1 R2 O R3 )
, the upper triangular factor of the QR decomposition of A𝛱 with R1∈ℝk×k, satisfy the following inequality:
Finding the optimal permutation matrix is not practical from the viewpoint of computational complexity.
1.2 Truncation error of pivoted QR
Pivoted QR sometimes results in a large truncation error. A well-known example was shown by Kahan, whose work we do not reproduce here [10]. In 1968, Faddeev et al. [6] showed that
Furthermore,
Q=(
Q1k Q2k) ,R=
(R1k R2k O R3k )
‖A−Q1k�
R1k R2k� 𝛱T
grd‖2=‖R3k‖2
pivotQRk(A) =‖R3k‖2.
(1) r2ll≥
∑j i=l
r2ij (l=1, 2,…,n−1, j=l+1,l+2,…,n).
‖R3‖2≤√
k(n−k) +min(k,n−k)𝜎
k+1(A).
pivotQRn−1(A)≤
√4n+6n−1
3 SVDn−1(A).
holds [3].
However, in a survey in 2017, it was stated that “very little is known in theory about its behaviour” [13, p. 2218] with regard to pivoted QR, thus there is still room for further research on pivoted QR.
Our previous work showed that the least upper bound of the ratio of the trunca- tion error of pivoted QR to that of SVD is √
4n−1+2
3 in case an m×n ( m≥n ) matrix is approximated to a matrix whose rank is n−1 , i.e., for k=n−1 [11]. The tight upper bound for all k is proved in the rest of this paper.
We assume that all matrices and vectors in this paper are real numbers; however, we can easily extend the discussion in this paper to complex numbers, and the same results can be obtained.
2 Preliminaries
In this section, we define the notations and examine the basic properties to analyze the truncation errors. First, we introduce the concept resi.
Proposition 1 [1, p. 16] For A∈ℝm×n, there exists X∈ℝn×m that satisfies
and X is uniquely determined by the four conditions.
Definition 1 For A∈ℝm×n ( m≥n ), the generalized inverse of A is defined by X∈ℝn×m that satisfies the four conditions in Proposition 1 and is denoted by A†.
The following notation is closely related to the truncation error of pivoted QR.
Definition 2 Let A∈ℝm×n ( m≥n ) and B∈ℝm×l . We define resi(A,B) as
We denote the inner product of two vectors x and y as (x,y). Example 1 For x∈ℝn and y∈ℝn , if x≠0 , then the following holds:
The following lemma will be used to identify resi. pivotQRk(A)≤ n√
4k+6k−1
3 SVDk(A)
AXA=A,XAX=X,(AX)T =AX,(XA)T=XA
resi(A,B) =B−AA†B.
resi(x,y) =y−(x,y)
‖x‖2x.
holds.
Proof If resi(A,B) =B−AX holds, then
holds. If ATAX−ATB=O holds, then
holds. ◻
Lemma 2 [1, p. 5] Let A∈ℝm×n(m≥n), b∈ℝm, and x∈ℝn. ‖b−Ax‖≤‖b−Ay‖ holds for any y∈ℝn if and only if AT(Ax−b) =0 holds.
Using Lemmas 1 and 2, we can obtain the following lemma.
Lemma 3 Let A∈ℝm×n(m≥n), b∈ℝm, and x∈ℝn. ‖b−Ax‖≤‖b−Ay‖ holds for any y∈ℝn if and only if resi(A,b) =b−Ax holds.
Lemma 4 Let m≥n>k, A∈ℝm×n, and B∈ℝm×l. Let A be partitioned as
where A1k∈ℝm×k. Then,
holds.
Proof From the definition of resi , we can see that
and
hold where X=A†
1kA2k , Y =A†
1kB and Z=resi(A1k,A2k)†resi(A1k,B) . Thus, holds from (2), (3), and (4). Lemma 1 proves
ATAX−ATB=O⇔resi(A,B) =B−AX
ATAX−ATB= −ATresi(A,B) = (ATAA†−AT)B
= (AT(AA†)T−AT)B= ((AA†A)T−AT)B=O
resi(A,B) −B+AX=AX−AA†B=AA†AX−AA†B
= (AA†)TAX− (AA†)TB=A†T(ATAX−ATB) =O
A=(
A1k A2k)
resi(A,B) =resi(resi(A1k,A2k), resi(A1k,B))
(2) resi(A1k,A2k) =A2k−A1kX,
(3) resi(A1k,B) =B−A1kY
(4) resi(resi(A1k,A2k), resi(A1k,B)) =resi(A1k,B) −resi(A1k,A2k)Z
(5) resi(resi(A1k,A2k), resi(A1k,B)) =B−A1kY−A2kZ+A1kXZ
from (2),
from (3), and
from (4). We can see that
from (5), (6), and (7). We can see that
from (2), (4), (8), and (9). Then, (9) and (10) can be combined as
Next, (5) can be rewritten as
From this and (11), we have
Application of Lemma 1 to this proves the lemma. ◻
QR decomposition and resi have the following relation. Note that QR in this lemma is without pivoting.
Lemma 5 Let m≥n>l, A∈ℝm×n, and A=QR be a QR decomposition parti- tioned as
(6) AT1kA2k=AT1kA1kX,
(7) AT
1kB=AT
1kA1kY
(8) resi(A1k,A2k)T(resi(A1k,B) −resi(A1k,A2k)Z) =O
(9) AT
1kresi(resi(A1k,A2k), resi(A1k,B))
= AT1k(B−A1kY−A2kZ+A1kXZ)
= O
(10) AT
2kresi(resi(A1k,A2k), resi(A1k,B))
= (A2k−A1kX)Tresi(resi(A1k,A2k), resi(A1k,B))
= resi(A1k,A2k)T(resi(A1k,B) −resi(A1k,A2k)Z)
= O
(11) (AT1k
AT
2k
)
resi(resi(A1k,A2k), resi(A1k,B))
= ATresi(resi(A1k,A2k), resi(A1k,B)) =O.
resi(resi(A1k,A2k), resi(A1k,B)) =B−A
(Y−XZ Z
) .
AT (
B−A
(Y−XZ Z
))
=O.
where A1l∈ℝm×l,Q1l∈ℝm×l,R1l∈ℝl×l. If rank(A1l) =l holds, then
holds.
Proof We have
Let
Then, we have
Furthermore,
holds. Application of Lemma 1 to this proves the lemma. ◻ Then, we return to pivoted QR. Let
where A1k∈ℝm×k . From Lemma 5, we can see that
for l=1 , 2, ..., k and j=l+1 , l+2 , ..., n and
if rank(A1k) =k holds. The last equation suggests that, as long as rank(A1k) =k holds, the value of pivotQRk(A) is determined only from A1k and A2k , or equivalently from 𝛱
grd , and is independent of how (or in what algorithm) the QR decomposition is computed.
A=( A1l A2l)
, Q=(
Q1l Q2l)
, R=
(R1l R2l O R3l )
resi(A1l,A2l) =Q2lR3l
A1l=Q1lR1l, A2l=Q1lR2l+Q2lR3l.
X=R−11lR2l.
Q2lR3l=A2l−A1lX.
AT
1l(A2l−A1lX) =AT
1lQ2lR3l=RT
1lQT
1lQ2lR3l=RT
1lOR3l=O
A𝛱grd =(
a𝜋1 a𝜋2 … a𝜋n
)=(
A1k A2k)
(1)⇔‖‖‖
(rll rl+1l … rnl)T‖‖‖
2
≥‖‖‖
(rlj rl+1j … rnj)T‖‖‖
2
⇔‖‖‖Q2(l−1)(
rll rl+1l … rnl)T‖‖‖
2
≥‖‖‖Q2(l−1)(
rlj rl+1j … rnj)T‖‖‖
2
⇔‖
‖‖resi((
a𝜋
1 … a𝜋
l−1
),a𝜋
l
)T‖
‖‖
2
≥‖‖
‖‖resi((
a𝜋
1 … a𝜋
l−1
),a𝜋
j
)T‖‖
‖‖
2
pivotQRk(A) =‖R3k‖2=‖Q2kR3k‖2=‖resi(A1k,A2k)‖2
3 Evaluation from above We bound pivotQRSVD k(A)
k(A) from above in this section. Since pivotQRk(A) =SVDk(A) =0 holds if rank(A)≤k holds, we only consider the case rank(A)>k . Let A=U𝛴VT be one SVD. Since A𝛱
grd=U𝛴(𝛱T
grdV)T and (𝛱T
grdV)T(𝛱T
grdV) =In hold, U𝛴(𝛱T
grdV)T is one SVD of A𝛱grd . Then, we can see that
Hereafter, we change what A represents. The previous A𝛱grd is replaced by A. Let A∈ℝm×n that satisfies
be partitioned as
where A1k∈ℝm×k and rank(A1k) =k . We should compare 𝜎
k+1(A) =SVDk(A) and
‖resi(A1k,A2k)‖2=pivotQRk(A).
Lemma 6 Let m≥n, A∈ℝm×n, and B∈ℝm×l. For any v∈ℝl,
holds.
Proof From the definition of resi,
holds. Thus,
holds. ◻
We can see that
from the definition of 2-norm and Lemma 6. Now, we introduce an essential theo- rem of this paper.
SVDk(A) =𝜎
k+1(A) =𝜎
k+1(A𝛱grd).
(12)
‖‖
‖resi((
a1 … ai−1) ,ai)‖‖‖
≥ ‖‖‖resi((
a1 … ai−1) ,aj)‖
‖‖(i=1,…,k, j=i+1,…,n)
A=(
a1 a2 … an)
=(
A1k A2k)
resi(A,B)v=resi(A,Bv)
resi(A,B) =B−AA†B
resi(A,B)v=Bv−AA†Bv=resi(A,Bv)
‖resi(A1k,A2k)‖2= max
z∈ℝn−k,‖z‖=1‖resi(A1k,A2k)z‖
= max
z∈ℝn−k,‖z‖=1‖resi(A1k,A2kz)‖
Theorem 2 Let m≥n>1, A∈ℝm×n, rank(A) =n, and A be partitioned as
We define Âi as
for i=1, 2, ..., n, and di as
for i=1, 2, ..., n. Then, di≠0 for i=1, 2, ..., n and
hold.
Proof Since rank(A) =n , {a1,a2,…,an} is linearly independent. Because di is a lin- ear combination of {a1,a2,…,an} with the coefficient of ai being 1, di≠0 holds for i=1 , 2, ..., n. From the definition of resi,
holds, where x1 =Â1†a1 . Let x1=(
x12 x13 … x1n)T
. Let i be one of 2, 3, ..., n. We can see that
holds if x1i≠0 from Lemma 3. Thus,
holds. This (13) also holds if x1i=0 . We define y∈ℝm as A=(
a1 a2 … an) .
Âi=(
a1 … ai−1 ai+1 … an)
di=resi(Âi,ai)
‖a1‖
‖d1‖ ≤
�n i=2
‖ai‖
‖di‖
d1=a1−Â1x1
‖di‖≤
��
��
��
��
��
��
��
��
�� ai−Âi
⎛⎜
⎜⎜
⎜⎜
⎜⎜
⎜⎜
⎝
1 x1i
−x12
x1i
⋮
−x1i−1
x1i
−x1i+1
x1i
⋮
−x1n
x1i
⎞⎟
⎟⎟
⎟⎟
⎟⎟
⎟⎟
⎠
��
��
��
��
��
��
��
��
��
= ‖d1‖
�x1i�
(13)
�x1i�≤ ‖d1‖
‖di‖
y=d1+x1nan=a1−
∑n−1 i=2
x1iai.
Since {a1,a2,…,an−1} is linearly independent, y≠0 holds. As Lemma 1 gives Â1Td1=0 , we have (an,d1) =0 . Thus,
holds. We can see that
holds from Lemma 3 because y is a linear combination of ai(i=1, 2,…,n−1) . Since
and ‖dn‖>0 hold,
holds. Furthermore, since
holds from (13),
holds, and the theorem has been proved. ◻
We refer to an essential theorem by Hong et al.
Theorem 3 [9, p. 218] Let m≥n>l , A∈ℝm×n and A=QR=U𝛴VT be a QR decomposition and an SVD, respectively. Let R and V be partitioned as
x1n= (an,y)
‖an‖2
‖dn‖≤�
��
�an−(y,an)
‖y‖2 y�
��
�
‖dn‖2=‖dn‖2
‖d1‖2
��
��
�
y−(an,y)
‖an‖2an���
��
2
= ‖dn‖2
‖d1‖2‖y‖2
�
1− (an,y)2
‖y‖2‖an‖2
� ,
��
��an−(y,an)
‖y‖2 y�
��
�
2
=‖an‖2
�
1− (an,y)2
‖y‖2‖an‖2
�
‖an‖
‖dn‖≥ ‖y‖
‖d1‖
‖y‖≥‖a1‖−
�n−1 i=2
�x1i�‖ai‖
≥‖a1‖−
�n−1 i=2
‖d1‖
‖di‖‖ai‖
‖an‖
‖dn‖ ≥ ‖a1‖
‖d1‖−
�n−1 i=2
‖ai‖
‖di‖
R=
(R1l R2l O R3l )
, V =
(V1l V2l V3l V4l )
where R1l∈ℝl×l and V1l∈ℝl×l.
holds.
In the present study, this theorem is only used for l=n−1 . The following lemma provides an inequality between resi and the singular value.
Lemma 7 Under the same assumptions as Theorem 2,
holds.
Proof Let A=U𝛴VT be an SVD partitioned as
where V1∈ℝn×(n−1) . Let ei be the ith column of In for i=1 , 2, ..., n. Define a permu- tation matrix 𝛱
i as
for i=1 , 2, ..., n. Since
and (𝛱T
i V)T(𝛱T
i V) =In , U𝛴(𝛱T
i V)T is one SVD of (Âi ai)
. Let A𝛱
i=QiRi be a QR decomposition partitioned as
where Qi1∈ℝm×(n−1),Ri1∈ℝ(n−1)×(n−1) . Using Theorem 3,
holds. We can see that
holds from Lemma 5. Thus,
holds. Then,
‖R3l‖2𝜎n−l(V4l)≤𝜎
l+1(A)
1≤(𝜎
n(A))2
�n i=1
1
‖di‖2
V=( V1 v2)
, v2=(
v21 v22 … v2n)T
𝛱i=(
e1 … ei−1 ei+1 … en ei)
(Âi ai)
=A𝛱i=U𝛴(𝛱T
i V)T
Qi=( Qi1 qi2)
,Ri=
(Ri1 ri2 O ri3 )
𝜎n(A) =𝜎n(A𝛱i)≥|v2i| |ri3|
‖di‖=‖resi(Âi,ai)‖=‖ri3qi2‖=�ri3�
𝜎n(A)≥�v2i� ‖di‖
holds. ◻ Proposition 2 Let m≥n>k and A∈ℝm×n satisfy (12) with being partitioned as where A1k∈ℝm×k. Let A satisfy rank(A1k) =k. Then, for all z∈ℝn−k with ‖z‖=1,
holds.
Proof From (12) and Lemma 6, the following holds for i=1 , 2, ..., k:
Define A′ as
If rank(A�)≠k+1 , then {a1,a2,…,ak,A2kz} is linearly dependent. Since
rank(A1k) =k , {a1,a2,…,ak} is linearly independent, and A2kz can be expressed as a linear combination of {a1,a2,…,ak} . Then, we have
resi(A1k,A2kz) =0 from Lemma 3, and the conclusion holds. Therefore, we only consider the case rank(A�) =k+1 in the remainder of this proof. We define d′i as
From Lemma 4, we can see that
holds for i=1 , 2, ..., k and j=i , i+1 , ..., k, where A�ijk =(
ai … aj−1 aj+1 … ak A2kz) , and 1=
�n i=1
(v2i)2≤(𝜎
n(A))2
�n i=1
1
‖di‖2
A=(
a1 a2 … an)
=(
A1k A2k)
‖resi(A1k,A2kz)‖≤
�4k−1
3 (n−k) +1𝜎
k+1(A)
(14) (n−k)‖
‖‖resi((
a1 … ai−1) ,ai)‖
‖‖
2
≥
∑n j=k+1
‖‖
‖resi((
a1 … ai−1) ,aj)‖
‖‖
2
=‖‖‖resi((
a1 … ai−1) ,A2k)‖‖‖
2 F
≥‖‖
‖resi((
a1 … ai−1) ,A2k)‖‖
‖
2 2
≥‖‖‖resi((
a1 … ai−1) ,A2kz)‖
‖‖
2
.
A�=(
a1 a2 … ak A2kz) .
d�i =resi((
a1 … ai−1 ai+1 … ak A2kz) ,ai)
(i=1, 2,…,k).
d�j =resi( resi((
a1 … ai−1) ,A�ijk)
, resi((
a1 … ai−1) ,aj))
holds for i=1, 2, ..., k. Using Theorem 2 on resi((
a1 a2 … ai−1) ,(
ai ai+1 … ak A2kz))
, we can see that
holds. Thus,
holds for i=1 , 2, ..., k from (12) and (14). Thus,
holds. We want to show that
and prove this using induction in the order of i=k , k−1 , ..., 1. Applying (15) for i=k gives
resi(A1k,A2kz)
= resi( resi((
a1 … ai−1) ,(
ai … ak)) , resi((
a1 … ai−1) ,A2kz))
‖‖
‖resi((
a1 a2 … ai−1) ,ai)‖
‖‖
‖‖
‖resi( resi((
a1 … ai−1) ,A�iik)
, resi((
a1 … ai−1) ,ai))‖‖
‖
≤
∑k j=i+1
‖‖
‖resi((
a1 a2 … ai−1) ,aj)‖‖‖
‖‖
‖‖resi (
resi((
a1 … ai−1) ,A�ijk
) , resi((
a1 … ai−1) ,aj))‖‖‖‖
+
‖‖
‖resi((
a1 a2 … ai−1) ,A2kz)‖
‖‖
‖‖
‖resi( resi((
a1 … ai−1) ,(
ai … ak)) , resi((
a1 … ai−1)
,A2kz))‖‖
‖
��
�resi��
a1 a2 … ai−1� ,ai����
‖d�i‖
≤
�k j=i+1
��
�resi��
a1 a2 … ai−1� ,aj���
�
‖d�j‖ +
��
�resi��
a1 a2 … ai−1� ,A2kz���
�
‖resi(A1k,A2kz)‖
≤ ���resi��
a1 a2 … ai−1� ,ai����
� k
�
j=i+1
1
‖d�j‖+
√n−k
‖resi(A1k,A2kz)‖
�
1 (15)
‖d�i‖ ≤
�k j=i+1
1
‖d�j‖+
√n−k
‖resi(A1k,A2kz)‖ (i=1, 2,…,k)
1 (16)
‖d�i‖ ≤ 2k−i√ n−k
‖resi(A1k,A2kz)‖ (i=1, 2,…,k)
1
‖d�k‖ ≤
√n−k
‖resi(A1k,A2kz)‖ = 2k−k√ n−k
‖resi(A1k,A2kz)‖.
Thus, (16) is shown in case i=k . Then, we prove that (16) holds for i=l , assuming that (16) holds for i=l+1 , l+2 , ..., k. We can see that
holds from (15) and the assumption of induction. Thus, (16) has been shown in case i=1 , 2, ..., k. Using Lemma 7 on A′,
holds. Thus,
holds. Now, if we can show that
then the proof is complete. Considering the fact that
we want a subspace 𝛩 that satisfies
Let
1
‖d�l‖ ≤
�k j=l+1
1
‖d�j‖+
√n−k
‖resi(A1k,A2kz)‖
≤
√n−k
‖resi(A1k,A2kz)‖
� k
�
j=l+1
2k−j+1
�
= 2k−l√ n−k
‖resi(A1k,A2kz)‖
1≤(𝜎
k+1(A�))2
� k
�
i=1
1
‖d�i‖2 + 1
‖resi(A1k,A2kz)‖2
�
≤ (𝜎
k+1(A�))2
‖resi(A1k,A2kz)‖2
� (n−k)
�k i=1
4k−i+1
�
= (𝜎
k+1(A�))2
‖resi(A1k,A2kz)‖2
�4k−1
3 (n−k) +1
�
‖resi(A1k,A2kz)‖≤
�4k−1
3 (n−k) +1𝜎
k+1(A�)
𝜎k+1(A�)≤𝜎
k+1(A),
𝜎k+1(A) = max
𝛩,dim𝛩=k+1 min
x∈𝛩,‖x‖=1‖Ax‖,
x∈𝛩min,‖x‖=1‖Ax‖≥𝜎
k+1(A�).
𝛩�=span {
e1,e2,…,ek, (0
z )}
.
Then, we have dim(𝛩�) =k+1 since {
e1,e2,…,ek, (0
z )}
is linearly independent.
Let y= (yi) ∈ℝk+1 . Since (
e1 e2 … ek (0
z ))T(
e1 e2 … ek (0
z ))
=Ik+1 holds,
holds. For all y∈ℝk+1 that satisfies the right-hand side of (17),
holds. Then,
holds. ◻
Thus, we have proved that
4 Evaluation from below
In this section, we show that the inequality proved in the previous section is tight. An example of matrix Rh with real-valued parameter h that satisfies
is shown. Rh is as follows:
The Kahan matrix is [10]
(17)
‖y‖=1⇔
��
��
��
�k i=1
yiei+yk+1
�0 z
������
�
=1
��
��
�� A
� k
�
i=1
yiei+yk+1
�0 z
�����
��
�
=‖A�y‖≥𝜎
k+1(A�)
𝜎k+1(A)≥ min
x∈𝛩�,‖x‖=1‖Ax‖≥𝜎
k+1(A�)
pivotQRk(A)≤
√4k−1
3 (n−k) +1SVDk(A).
pivotQRk(Rh) SVDk(Rh)
�
���������������→
h→0
√4k−1
3 (n−k) +1
Rh=
⎛⎜
⎜⎜
⎜⎜
⎜⎝
1 0 … 0
0 h ⋱ ⋮
⋮ ⋱ ⋱ 0
0 … 0 hk
O
O O
⎞⎟
⎟⎟
⎟⎟
⎟⎠
⎛⎜
⎜⎜
⎜⎜
⎜⎝ 1 −√
1−h2 … −√
1−h2 … −√ 1−h2
0 1 ⋱ ⋱ ⋱ ⋮
⋮ ⋱ ⋱ −√
1−h2 … −√ 1−h2
0 … 0 1 … 1
O
⎞⎟
⎟⎟
⎟⎟
⎟⎠ .
Therefore, Rh is the same as the Kahan matrix in case m=n=k+1 and is an exten- sion of the Kahan matrix otherwise.
Proposition 3 Let m≥n>k. Define 𝛴h∈ℝm×n, (whij) =Wh∈ℝn×n, and Rh∈ℝm×n as follows:
and Rh=𝛴
hWh where 0<h<1. Then,
holds.
Proof Let Q= (In
O )
∈ℝm×n and R=diag(1,h,…,hk, 0, 0,…, 0)Wh∈ℝn×n . Since R is an upper triangular matrix and QTQ=In holds, Rh=QR is one QR decomposi- tion. We check (1) for this R. Since
(1) holds for l=1 , 2, ..., k+1 , j=l+1 , l+2 , ..., n. Obviously (1) also holds for l=k+2 , k+3 , ..., n−1 , j=l+1 , l+2 , ..., n. As in Sect. 2, let R be partitioned as
where R1k∈ℝk×k . Then, Kn =
⎛⎜
⎜⎜
⎝
1 0 … 0
0 h ⋱ ⋮
⋮ ⋱ ⋱ 0
0 … 0 hn−1
⎞⎟
⎟⎟
⎠
⎛⎜
⎜⎜
⎜⎝
1 −√
1−h2 … −√ 1−h2
0 1 ⋱ ⋮
⋮ ⋱ ⋱ −√
1−h2
0 … 0 1
⎞⎟
⎟⎟
⎟⎠ .
𝛴h=
�diag(1,h,…,hk, 0, 0,…, 0) O
� ,
whij=
⎧⎪
⎨⎪
⎩
1 (i=jand 1≤i≤k)or(i=k+1 andk+1≤j≤n),
−√
1−h2 (i<jand 1≤i≤k),
0 otherwise
limh→0
pivotQRk(Rh) SVDk(Rh) =
√4k−1
3 (n−k) +1
(left side of (1)) =h2l−2= (1−h2)
min(j,k+1)−1∑
i=l
h2i−2+h2 min(j,k+1)−2
= (right side of (1)),
R=
(R1k R2k O R3k )
pivotQRk(Rh) =‖R3k‖2
holds. Define V ∈ℝ(n−k)×(n−k) and v1∈ℝn−k as follows:
where v2 , v3 , ..., vn−k are chosen such that VTV =In−k holds. We can choose them freely as long as this is satisfied. Since
holds, ‖R3k‖2=hk√
n−k holds. We consider the value of SVDk(Rh) =𝜎
k+1(Rh) . Considering the fact that
we want a subspace 𝛩 whose maxx∈𝛩,‖x‖=1‖Rhx‖ is small. Since vT
1vi=0 holds for i=2 , 3, ..., n−k,
holds for i=2 , 3, ..., n−k . We define yj=1 for j=k+1 , ..., n and define yj from j=k down to j=1 as yj=√
1−h2∑n
i=j+1yi . We define y∈ℝn as (
y1 y2 …yn)T
Then, .
holds. Since
holds,
V =�
v1 v2 … vn−k�
, v1= 1
√n−k
�1 1 … 1�T
R3k=hk√ n−k�
v1 0 … 0�T
𝜎k+1(Rh) = min
𝛩,dim𝛩=n−k max
x∈𝛩,‖x‖=1‖Rhx‖,
Rh
�0 vi
�
=
⎛⎜
⎜⎝ R2k R3k O
⎞⎟
⎟⎠ vi=
⎛⎜
⎜⎜
⎜⎜
⎜⎜
⎜⎜
⎜⎜
⎝
−√
(n−k)(1−h2)h0
−√
(n−k)(1−h2)h1
⋮
−√
(n−k)(1−h2)hk−1
√n−k hk 0 0
⋮ 0
⎞⎟
⎟⎟
⎟⎟
⎟⎟
⎟⎟
⎟⎟
⎠
vT1vi=0
Rhy=(
0 0 … 0(n−k)hk 0 0 … 0)T
limh→0‖y‖=���
�(n−k)2k−1 (n−k)2k−2 … (n−k)20 1 1 … 1����
=
�4k−1
3 (n−k)2+n−k