Quadratic Newton Iteration

Let $ R$ be a commutative ring with units. Let $ p \in R$ and $ {\phi} \in R[y]$ . We are concerned with solving the equation

$\displaystyle {\phi}(g) \ = \ 0.$ (29)

The following Proposition 8 states that Relation (31) computes better and better approximations of solutions of Relation (29) that is, approximations modulo higher and higher powers of $ p$ . Comparing to the results of the previous section, this proposition offers a significant improvement: the accuracy doubles at each step, leading to the so-called Quadratic Symbolic Newton Iteration.

Proposition 8   Let $ {\phi} \in R[y]$ and let $ g \in R$ be such that

$\displaystyle {\phi}(g) \ {\equiv} \ 0 \mod{p}$ (30)

Assume that $ {\phi}'(g)$ is invertible modulo $ p$ and let denote by $ {{\phi}'(g)}^{-1}$ its inverse. Consider an element $ h \in R$ such that

$\displaystyle h \ {\equiv} \ g - {\phi}(g) \, {{\phi}'(g)}^{-1} \mod{ p^{2}}$ (31)

Then the following assertions hold:
  1. $ {\phi}(h) \ {\equiv} \ 0 \mod{ p^{2}}$ ,
  2. $ h \ {\equiv} \ g \mod{ p}$ ,
  3. $ {\phi}'(h)$ is invertible modulo $ p$ , and thus modulo $ p^2$ .

Proof. Since $ {\phi}'(g)$ is invertible modulo $ p$ then $ {\phi}'(g)$ is invertible modulo $ p^{2}$ . (This is easy to check: Exercise!) Hence Relation 31 makes sense. In fact, Algorithm [*] computes $ {{\phi}'(g)}^{-1} \mod{ p^{2}}$ given $ {{\phi}'(g)}^{-1} \mod{ p }$ .

Since $ p$ divides $ p^2$ , the following congruence holds

$\displaystyle h \ {\equiv} \ g - {\phi}(g) \, {{\phi}'(g)}^{-1} \mod{ p}$ (32)

Since $ {\phi}(g) \ {\equiv} \ 0 \mod{p}$ , this leads to

$\displaystyle h \ {\equiv} \ g \mod{ p}$ (33)

and we have proved the second claim. Let us prove the first one. Using the Taylor expansion of $ {\phi}$ at $ h$ and the fact that $ p^2$ divides $ (h -g)^2$ we obtain

\begin{displaymath}\begin{array}{rcl} {\phi}(h) & {\equiv} & {\phi}(g) + {{\phi}...
...g)}^{-1} \mod{ p^2} \\ & {\equiv} & 0 \mod{ p^2} \\ \end{array}\end{displaymath} (34)

proving the first claim. Finally, observe that $ h \ {\equiv} \ g \mod{ p}$ implies $ {\phi}'(h) \ {\equiv} \ {\phi}'(g) \mod{ p}$ . Indeed $ {\phi}'$ is a polynomial in $ R[x]$ . Therefore, $ {\phi}'(h)$ is invertible modulo $ p$ , since $ {\phi}'(g)$ is invertible modulo $ p$ , by hypothesis. $ \qedsymbol$

Remark 7   Intuitively, Proposition 8 says that if $ g$ is a good approximation of a zero of $ {\phi}$ then $ h$ is a better approximation, at least twice as accurate.

Remark 8   Proposition 8 leads to an algorithm provided that we can Then for computing the inverse of $ {\phi}'(g)$ modulo $ p$ the idea is to apply the algorithm to itself! That is, to the particular case of

$\displaystyle {\phi}(g) \ = \ 1/g -f$ (35)

This leads to Algorithm 1. Finally recall that for a prime element $ p$ in an Euclidean domain $ R$ , the computation of $ {{\phi}'(g)}^{-1}$ modulo $ p$ can be done by the Extended Euclidean Algorithm.

Algorithm 1  

\fbox{
\begin{minipage}{12 cm}
\begin{description}
\item[{\bf Input:}] ${\ell} \...
...) s_{r-1} \mod{ p^{\ell}}$\ \\
\> {\bf return} $g$
\end{tabbing}\end{minipage}}

Theorem 3   Algorithm 1 is correct.

Proof. Let $ g_r \ {\equiv} \ g_{r-1} - {\phi}(g_{r-1}) s_{r-1} \mod{p^{2^r}}$ . Then we have

$\displaystyle g \ {\equiv} \ g_r \ \mod{ p^{\ell}}$ (36)

In other words $ g_r$ is a solution of higher precision. Then proving the theorem turns into proving the invariants for $ i = 0 \cdots r$
  1. $ g_i \ {\equiv} \ g_0 \mod{ p}$ ,
  2. $ {\phi}(g_i) \ {\equiv} \ 0 \ \mod{ p^{2^i}}$ ,
  3. if $ i < r$ then $ s_i \, {\phi}'(g_i) \ {\equiv} \ 1 \ \mod{ p^{2^i}}$ .
The case $ i = 0$ is clear and we assume that $ i > 0$ holds. Then by induction hypothesis $ p^{2^{i-1}}$ divides $ {\phi}(g_{i-1})$ and $ s_{i-1} - {{\phi}'(g_{i-1})}^{-1}$ . Then $ p^{2^i}$ divides their product, that is

$\displaystyle {\phi}(g_{i-1}) s_{i-1} \ {\equiv} \ {\phi}(g_{i-1}) {{\phi}'(g_{i-1})}^{-1} \ \mod{ p^{2^i}}.$ (37)

Now we calculate

\begin{displaymath}\begin{array}{rcl} g_i & {\equiv} & g_{i-1} - {\phi}(g_{i-1})...
...g_{i-1}) {{\phi}'(g_{i-1})}^{-1} \ \mod{p^{2^i}} \\ \end{array}\end{displaymath} (38)

Now applying Proposition 8 with $ g = g_{i-1}$ , $ h = g_i$ and $ m = p^{2^{i-1}}$ we obtain the first two invariants. In addition, this leads to

$\displaystyle g_i \ \equiv \ g_{i-1} \ \mod{ p^{2^{i-1}} }$ (39)

for $ i < r$ . This implies

$\displaystyle {{\phi}'(g_i)}^{-1} \ \equiv \ {{\phi}'(g_{i-1})}^{-1} \ \equiv \ s_{i-1} \ \mod{ p^{2^{i-1}} }$ (40)

Now, $ s_i$ is obtained by one Newton step for inversion (see Algorithm [*]). Therefore the third invariant follows. $ \qedsymbol$

Marc Moreno Maza
2008-01-07