G(k(k−1)/2+1) = A(1) B(k), G(k(k−1)/2+2) = A(2) B(k−1), . . . , G(k(k−1)/2+k) = A(k) B. Moreover, we define G ̃ (v) := fl(G(v)) for all v.
Theorem 2 Assume that
k−1 k(k−1)/2+k
(i) (j) (i) (k−i+1) (k) ̃(k)
C=flAB+flAB+flAB= G. i+j≤k i=1 k=1
Then,
|C − AB| ≤ nkγn · 2(β−h)(k−1) · 2P(1) 2Q(1) T =: E. (30) If D(1) is the result of fl( k(k−1)/2+k G ̃ (k)), then
k=1
|D(1)−C|≤γ
If D(2) is the result of a summation algorithm for k(k−1)/2+k G ̃ (k) with faithful
rounding as in [11], then
|D −C|≤2u
in Algorithm 5, then |D(1)−AB|≤|C−AB|+|D(1)−C|=E+γ
k(k−1)/2+k−1
k(k−1)/2+k
|G ̃(i)|.
i=1
(31)
k(k−1)/2+k−1
k(k−1)/2+k
|G ̃(i)|
k(k−1)/2+k
(2) ̃(i)
G .
Therefore, if we apply ordinary floating-point summation in final summation
If we apply a summation algorithm with faithfully rounded result, then
k(k−1)/2+k
(2) (2) ̃(i)
i=1
Proof The error bounds of |D(1) − C| and |D(2) − C| are easily obtained by an a priori error analysis and definition of the faithful rounding. Therefore, we show only the error bound of |C − AB|. Let γn be
γn = nu forn∈N. (33) 1−nu
Then from [8], an upper bound of the rounding error of the floating-point matrix product fl(AB) is
|AB − fl(AB)| ≤ γn|A||B|. (34)
|D − AB|≤|C− AB|+|D −C|= E+2u
G . (32)
i=1
i=1
k=1
G(k(k−1)/2+1) = A(1) B(k), G(k(k−1)/2+2) = A(2) B(k−1), . . . , G(k(k−1)/2+k) = A(k) B. Moreover, we define G ̃ (v) := fl(G(v)) for all v.Theorem 2 Assume that k−1 k(k−1)/2+k (i) (j) (i) (k−i+1) (k) ̃(k)C=flAB+flAB+flAB= G. i+j≤k i=1 k=1 Then,|C − AB| ≤ nkγn · 2(β−h)(k−1) · 2P(1) 2Q(1) T =: E. (30) If D(1) is the result of fl( k(k−1)/2+k G ̃ (k)), then k=1|D(1)−C|≤γIf D(2) is the result of a summation algorithm for k(k−1)/2+k G ̃ (k) with faithfulrounding as in [11], then|D −C|≤2u in Algorithm 5, then |D(1)−AB|≤|C−AB|+|D(1)−C|=E+γk(k−1)/2+k−1k(k−1)/2+k |G ̃(i)|.i=1(31)k(k−1)/2+k−1k(k−1)/2+k |G ̃(i)|k(k−1)/2+k(2) ̃(i) G .Therefore, if we apply ordinary floating-point summation in final summationIf we apply a summation algorithm with faithfully rounded result, thenk(k−1)/2+k(2) (2) ̃(i) i=1Proof The error bounds of |D(1) − C| and |D(2) − C| are easily obtained by an a priori error analysis and definition of the faithful rounding. Therefore, we show only the error bound of |C − AB|. Let γn beγn = nu forn∈N. (33) 1−nuThen from [8], an upper bound of the rounding error of the floating-point matrix product fl(AB) is|AB − fl(AB)| ≤ γn|A||B|. (34)|D − AB|≤|C− AB|+|D −C|= E+2u G . (32)i=1i=1k=1
การแปล กรุณารอสักครู่..
