where Tm is the m-th Chebyshev polynomial of the first kind. Then g is 2ε-close to 1/x on Dκ=[−1,−1/κ]∪[1/κ,1]:
g(x)−x1≤2εfor all x∈Dκ.
This is the decomposition of 1/x into odd Chebyshev polynomials of degree O(κlog(κ/ε)) on which the Chebyshev-based quantum linear systems algorithm (Theorem 4 of the paper) is built: applied to a Hermitian matrix with spectrum in Dκ, the same combination of Chebyshev polynomials of the matrix approximates its inverse.
Formalization Noteκ≥1 is a condition number (ratio of the largest to the smallest singular value, p. 3); it makes Dκ nonempty and keeps 0 out of it. The page's real b is rounded up, b=⌈κ2log(κ/ε)⌉ (it is an exponent, and this keeps the hypothesis of Lemma 17), and the sum runs over j=0,…,⌊j0⌋. No upper bound on ε is assumed: if ε≥κ then b=0, g≡0 and ∣1/x∣≤κ≤ε on Dκ, so the statement is still true there.
Childs, Kothari and Somma, Quantum algorithm for systems of linear equations with exponentially improved dependence on precision, arXiv:1511.02306v2, p. 16, Lemma 14 (eq. (55))
Confirmed by the moderator at approval.