Linear Newton Iteration

Theorem 1   Let $ R$ be an Euclidean domain with a regular Euclidean size $ {\delta}$ . Let $ p \in R$ be a prime element. Let $ {\phi} \in R[y]$ and let $ a, a_0 \in R$ such that
  1. $ {\phi}(a) = 0$ ,
  2. $ a \equiv a_0 \mod{p}$ ,
  3. $ {\phi}'(a_0) \neq 0 \mod{p}$ .
If $ a$ is unknown, but $ {\phi}$ and $ a_0$ are known then we can compute a $ p$ -adic expansion of $ a$ . Therefore we can compute $ a$ exactly.

Proof. Let $ (a_0, a_1, \ldots, a_d)$ be a $ p$ -adic expansion of $ a$ w.r.t. $ p$ . Let $ k < d + 1$ be a positive integer. By Proposition 6, the element

$\displaystyle a^{(k)} = a_0 + a_1 p + \cdots a_{k-1} p^{k-1}$ (12)

is a $ p$ -adic approximation of $ a$ at order $ k$ . By Proposition 4 there exists a polynomial $ {\psi} \in R[y]$ such that

$\displaystyle {\phi}(a^{(k)} + a_k p^k) = {\phi}(a^{(k)}) + {\phi}'(a^{(k)}) a_k p^k + {\psi}(a^{(k)} + a_k p^k) (a_k p^k)^2.$ (13)

Since we have

$\displaystyle a \equiv a^{(k)} + a_k p^k \mod{p^{k+1}}$    

we deduce from Proposition 7

$\displaystyle {\phi}(a) \equiv {\phi}(a^{(k)} + a_k p^k) \mod{p^{k+1}}$ (14)

Since $ {\phi}(a) = 0$ this shows that $ {\phi}(a^{(k)} + a_k p^k)$ is in the ideal generated by $ p^{k+1}$ . Similarly, $ {\phi}(a^{(k)})$ is in the ideal generated by $ p^k$ . Therefore we can divide $ {\phi}(a^{(k)} + a_k p^k)$ and $ {\phi}(a^{(k)})$ by $ p^k$ , leading to

$\displaystyle \frac{{\phi}(a^{(k)} + a_k p^k)}{p^k} = \frac{{\phi}(a^{(k)})}{p^k} + {\phi}'(a^{(k)}) a_k + {\psi}(a^{(k)} + a_k p^k) (a_k )^2 p^k.$ (15)

Now observe that

$\displaystyle \frac{{\phi}(a^{(k)} + a_k p^k)}{p^k} \equiv 0 \equiv {\psi}(a^{(k)} + a_k p^k) (a_k )^2 p^k \mod{p}.$ (16)

Let us denote by $ {\Phi}_p$ the canonical homomorphism from $ R$ to $ R/\langle p \rangle$ . Then we obtain

$\displaystyle {\Phi}_p(\frac{{\phi}(a^{(k)})}{p^k}) = - {\Phi}_p({\phi}'(a^{(k)})) {\Phi}_p(a_k)$ (17)

Now, since $ a \equiv a_0 \mod{p}$ holds we have

$\displaystyle {\Phi}_p({\phi}'(a^{(k)})) \equiv {\Phi}_p({\phi}'(a_0)) \mod{p}$ (18)

Since $ {\phi}'(a_0) \neq 0 \mod{p}$ holds we can solve Equatiion 17 for $ {\Phi}_p(a_k)$ . Finally, from Theorem 1, we have $ {\deg}(a_k,p) = 0$ such that we can view $ {\Phi}_p(a_k)$ as $ a_k$ . This is straightforward if $ R = {\mbox{${\mathbb{Z}}$}}$ or if $ R = {\bf k}[x]$ where $ {\bf k}$ is a field, since we can choose for $ {\Phi}_p(x)$ the remainder of $ x$ modulo $ p$ . $ \qedsymbol$

Theorem 2   Let $ R$ be a commutative ring with identity element and let $ {\cal I}$ be a finitely generated ideal of $ R$ . Let $ r$ be a positive integer. Let $ f_1, \ldots, f_n \in R[x_1, \ldots, x_r]$ be $ n$ multivariate polynomials in the $ r$ variables $ x_1, \ldots, x_r$ . Let $ a_1, \ldots, a_r \in R$ be elements. Let $ U$ be the Jacobian matrix of $ f_1, \ldots, f_n$ evaluated at $ (a_1, \ldots, a_r)$ . That is, $ U$ is the $ n \times r$ matrix defined by

$\displaystyle U = (u_{ij}) \ \ {\rm where} \ \ u_{ij} = \frac{{\partial} f_i}{{\partial} x_j}(a_1, \ldots, a_r)$ (19)

We assume that the following properties hold
  1. for every $ i = 1 \cdots n$ we have $ f_i(a_1, \ldots, a_r) \equiv 0 \mod {{\cal I}}$ .
  2. the Jacobian matrix $ U$ is left-invertible modulo $ {{\cal I}}$ .
Then, for every positive integer $ {\ell}$ we can compute $ a^{({\ell})}_1, \ldots, a^{({\ell})}_r \in R$ such that
  1. for every $ i = 1, \ldots, n$ we have $ f_i(a^{({\ell})}_1, \ldots, a^{({\ell})}_r) \equiv 0 \mod {{\cal I}^{\ell}}$ ,
  2. for every $ j = 1, \ldots, r$ we have $ a^{({\ell})}_j \equiv a_j \mod {{\cal I}}$ .

Proof. We proceed by induction on $ {\ell} \geq 1$ . For $ {\ell} = 1$ the claim follows from the hypothesis of the theorem. So let $ {\ell} \geq 1$ be such that the claim is true. Hence there exist $ a^{({\ell})}_1, \ldots, a^{({\ell})}_r \in R$ such that

$\displaystyle f_i(a^{({\ell})}_1, \ldots, a^{({\ell})}_r) \equiv 0 \mod {{\cal I}^{\ell}}, \ \ i = 1, \ldots, n$ (20)

and

$\displaystyle a^{({\ell})}_j \equiv a_j \mod {{\cal I}}, \ \ i = 1, \ldots, r$ (21)

Since $ {\cal I}$ is finitely generated, then so is $ {\cal I}^{\ell}$ and let $ g_1, \ldots, g_s \in R$ such that

$\displaystyle {\cal I}^{\ell} = \langle g_1, \ldots, g_s \rangle$ (22)

Therefore, for every $ i = 1, \ldots, r$ , there exist $ q_{i1}, \ldots q_{is} \in R$ such that

$\displaystyle f_i(a^{({\ell})}_1, \ldots, a^{({\ell})}_r) \ = \ {\Sigma}_{k=1}^{k=s} q_{ik} g_k$ (23)

For each $ j = 1, \ldots, r$ we want to compute $ B_j \in R$ such that

$\displaystyle a^{({\ell}+1)}_j = a^{({\ell})}_j + B_j$ (24)

is the desired next approximation. We impose $ B_j \in {\cal I}^{\ell}$ so let $ b_{j1}, \ldots, b_{js} \in R$ be such that

$\displaystyle B_j \ = \ {\Sigma}_{k=1}^{k=s} b_{jk} g_k$ (25)

Using Proposition 5 we obtain

\begin{displaymath}\begin{array}{rcl} f_i(a^{({\ell}+1)}_1, \ldots, a^{({\ell}+1...
... b_{jk} \right) g_k \mod{{ {\cal I}^{{\ell}+1}}} \\ \end{array}\end{displaymath} (26)

where $ \left( u^{({\ell})}_{ij} \right)$ is the Jacobian matrix of $ (f_1, \ldots, f_n)$ at $ (a^{({\ell})}_1, \ldots, a^{({\ell})}_r)$ .

Hence, solving for $ a^{({\ell}+1)}_1, \ldots, a^{({\ell}+1)}_r \in R$ such that

$\displaystyle f_i(a^{({\ell}+1)}_1, \ldots, a^{({\ell}+1)}_r) \equiv 0 \mod{{ {\cal I}^{{\ell}+1}}}$    

leads to solving the system of linear equations:

$\displaystyle q_{ik} + {\Sigma}_{j=1}^{j=r} u^{({\ell})}_{ij} b_{jk} \equiv 0 \mod{{ {\cal I} }}$ (27)

for $ k = 1, \ldots, s$ and $ i = 1, \ldots, n$ .

Now using $ a^{({\ell})}_j \equiv a_j \mod {{\cal I}}$ for $ j = 1, \ldots, r$ we obtain

$\displaystyle u^{({\ell})}_{ij} \equiv u_{ij} \mod{{\cal I}}$ (28)

Therefore the linear system equations given by Relation (27) has solutions. $ \qedsymbol$

Marc Moreno Maza
2008-01-07