“…We can then approximate A by a series expansion of order n : A n = I + K + K 2 + K 3 + K n .The absolute truncation error ε abs introduced by this approximation is then‖ A − A n ‖ ≤ ε abs ,where ‖ ‖ represents any matrix norm. Using eqn (38), the previous inequality can be written as‖( K ) n +1 ( I − K ) −1 ‖ ≤ ε abs .We can then use the property‖ XY ‖ ≤ ‖ X ‖·‖ Y ‖,which is valid for any matrix norm, as well as 18 ρ ( X n ) = ρ ( X ) n ≤ ‖ X ‖ n .Choosing the minimum norm for which 21
the inequality (39) will always be satisfied if ρ ( K n +1 )(1 − ρ ( K )) −1 ≤ ε abs .This leads to the condition
For instance, choosing an accuracy ε abs = 0.1, we find n ∼ 43 for ρ ( K ) = 0.9, n ∼ 9 for ρ ( K ) = 0.7, while for ρ ( K ) = 0.5 n drops to ∼3.Defining the relative truncation error ε rel by
and using similar arguments, we obtain an estimate for the truncation order required for convergence to within ε rel :
…”