src/HOL/ex/Lagrange.thy
author chaieb
Wed, 27 Feb 2008 14:39:58 +0100
changeset 26161 34cb0b457dcc
parent 25475 d5a382ccb5cc
child 26480 544cef16045b
permissions -rw-r--r--
Old HOL/Dense_Linear_Order.thy and the setup in Arith_Tools for Ferrante and Rackoff's Quantifier elimination for linear arithmetic over ordered Fields.

(*  Title:      HOL/ex/Lagrange.thy
    ID:         $Id$
    Author:     Tobias Nipkow
    Copyright   1996 TU Muenchen
*)

header {* A lemma for Lagrange's theorem *}

theory Lagrange imports Main begin

text {* This theory only contains a single theorem, which is a lemma
in Lagrange's proof that every natural number is the sum of 4 squares.
Its sole purpose is to demonstrate ordered rewriting for commutative
rings.

The enterprising reader might consider proving all of Lagrange's
theorem.  *}

definition sq :: "'a::times => 'a" where "sq x == x*x"

text {* The following lemma essentially shows that every natural
number is the sum of four squares, provided all prime numbers are.
However, this is an abstract theorem about commutative rings.  It has,
a priori, nothing to do with nat. *}

(* These two simprocs are even less efficient than ordered rewriting
   and kill the second example: *)
ML_setup {*
  Delsimprocs [ab_group_add_cancel.sum_conv, ab_group_add_cancel.rel_conv]
*}

lemma Lagrange_lemma: fixes x1 :: "'a::comm_ring" shows
  "(sq x1 + sq x2 + sq x3 + sq x4) * (sq y1 + sq y2 + sq y3 + sq y4) =
   sq (x1*y1 - x2*y2 - x3*y3 - x4*y4)  +
   sq (x1*y2 + x2*y1 + x3*y4 - x4*y3)  +
   sq (x1*y3 - x2*y4 + x3*y1 + x4*y2)  +
   sq (x1*y4 + x2*y3 - x3*y2 + x4*y1)"
by (simp add: sq_def ring_simps)


text {* A challenge by John Harrison. Takes about 12s on a 1.6GHz machine. *}

lemma fixes p1 :: "'a::comm_ring" shows
  "(sq p1 + sq q1 + sq r1 + sq s1 + sq t1 + sq u1 + sq v1 + sq w1) * 
   (sq p2 + sq q2 + sq r2 + sq s2 + sq t2 + sq u2 + sq v2 + sq w2) 
    = sq (p1*p2 - q1*q2 - r1*r2 - s1*s2 - t1*t2 - u1*u2 - v1*v2 - w1*w2) + 
      sq (p1*q2 + q1*p2 + r1*s2 - s1*r2 + t1*u2 - u1*t2 - v1*w2 + w1*v2) +
      sq (p1*r2 - q1*s2 + r1*p2 + s1*q2 + t1*v2 + u1*w2 - v1*t2 - w1*u2) +
      sq (p1*s2 + q1*r2 - r1*q2 + s1*p2 + t1*w2 - u1*v2 + v1*u2 - w1*t2) +
      sq (p1*t2 - q1*u2 - r1*v2 - s1*w2 + t1*p2 + u1*q2 + v1*r2 + w1*s2) +
      sq (p1*u2 + q1*t2 - r1*w2 + s1*v2 - t1*q2 + u1*p2 - v1*s2 + w1*r2) +
      sq (p1*v2 + q1*w2 + r1*t2 - s1*u2 - t1*r2 + u1*s2 + v1*p2 - w1*q2) +
      sq (p1*w2 - q1*v2 + r1*u2 + s1*t2 - t1*s2 - u1*r2 + v1*q2 + w1*p2)"
by (simp add: sq_def ring_simps)

end