Home > Maths

The theorem that characterises those integers that are the sum of 3 squares appears not to be easily findable on the Web. I therefore present a proof here. It is based on that in [R].

Theorems referred to:

Some theory about quadratic residues and quadratic forms is needed.

Theorem (Legendre 1798): A non-negative integer m is the sum of 3 squares iff it is not of the form 4a(8k+7).

Proof: Modulo 4, every square ≡ 0 or 1. Thus the only way the sum of 3 squares can ≡ 0 is 0+0+0. Thus if 4n is the sum of 3 squares, then

4n= (2a)2 + (2b)2 + (2c)2

n= a2 + b2 + c2

Thus n is the sum of 3 squares iff 4n is.

Thus if 4 divides m, apply the theorem to m/4; m is the sum of 3 squares iff m/4 is.

Modulo 8, every square ≡ 0, 1 or 4. Thus these can give totals in the following ways:

1 ≡ 0+0+1 ≡ 4+4+1

2 ≡ 0+1+1

3 ≡ 1+1+1

5 ≡ 0+4+1

6 ≡ 4+1+1

Thus an integer of the form 8k+7 is not the sum of 3 squares. This proves the "only if" part of the theorem.

The rest of this proof follows [R].

Let

F(x, y, z) = ax2 + by2 + cz2 + 2dyz + 2exz + 2fxy

be a ternary quadratic form such that D(F) = 1.

Let C = ab−f2 ___ [1]

Lemma 1. F is positive definite iff a>0 and C>0.

Proof: [if] Suppose a>0 and C>0. Then

F(x, y, z)= ax2 + by2 + cz2 + 2dyz + 2exz + 2fxy
Ca F(x, y, z) = Ca2x2 +a2b2y2 −abf2y2 +a2bcz2 −acf2z2 +2a2bdyz −2adf2yz +2a2bexz −2aef2xz + 2Cafxy
= Ca2x2 + abf2y2f4y2 + abe2z2e2f2z2 + 2abefyz2ef3yz +2a2bexz −2aef2xz + 2Cafxy +a2b2y2 −2abf2y2 +f4y2 +2a2bdyz −2abefyz −2adf2yz +2ef3yz + a2d2z22adefz2 +e2f2z2 +a2bcz2 +2adefz2a2d2z2abe2z2 −acf2z2, where italicised terms have been added and cancel each other
= (ab−f2)(a2x2 +f2y2 +e2z2 +2efyz +2aexz +2afxy ) + (aby−f2y+adz−efz)2 + az2, using D(F) = 1 to simplify the last 5 terms
= (ab−f2)(ax+fy+ez)2 + ((ab−f2)y+adz−efz)2 + az2
= C(ax+fy+ez)2 + [Cy+(ad−ef)z]2 + az2

which is positive unless x=y=z=0.

[only if] If a≤0, F(1, 0, 0) = a ≤ 0. If a>0 and C≤0,

F(x, y, 0) = {(ax+fy)2 + Cy2}/a

which has nonpositive values if ab is large enough. []

To prove the theorem, it is sufficient to find, for each relevant m, a ternary quadratic form F and values X, Y, Z such that

F(X, Y, Z) = m ___ [2]

a > 0 ___ [3]

C > 0 ___ [4]

D(F) = 1 ___ [5]

Here is one solution: If

X = Y = 0

Z = 1

c = m

d = 0

e = 1

then

F(X, Y, Z) = F(0, 0, 1) = m

satisfying [2]. Condition [5] reduces to

D(F) = Cm−b = 1 ___ [6]

The theorem is clearly true for m=1. For m>1,

C+f2= ab(from [1])
= a(Cm−1)(from [6])
f2= a(Cm−1) − C___ [7]

which implies [3] It is sufficient to find an integer q>0 where −q is a square, modulo qm−1, giving us f from [7].

If m is even, then, by Dirichlet's theorem, there is a prime p and integer t>=0 where

p = 4mt + m − 1

Let q = 4t+1. Then

p = qm − 1

Modulo 4, p ≡ 1 and q ≡ 1, so, by the quadratic reciprocity theorem,

(−q\p)= (−1\p)(q\p)as quadratic residues are multiplicative
= (q\p)as p ≡ 1
= (p\q)by the quadratic reciprocity theorem
= (qm−1 \q)
= (−1\q)
= 1as q ≡ 1

so q is as required.

If m modulo 8 ≡ 1, then (3m−1)/2 is odd. So, by Dirichlet's theorem, there is a prime p and positive integer u where

p= 4mu + (3m−1)/2
= ((8u+3)m−1)/2

Let q = 8u+3. Then q>0 and

2p = qm − 1 ___ [8].

Then q ≡ 3, modulo 8, and p ≡ 1, modulo 4.

(−q\p)= (−1\p)(q\p)as quadratic residues are multiplicative
= (q\p)as p ≡ 1
= (p\q)by the quadratic reciprocity theorem
Also
(-1\q)= -1as q ≡ 3 modulo 4
(2\q)= -1as q ≡ 3 modulo 8
(-2\q)= 1as quadratic residues are multiplicative

If m modulo 8 ≡ 3, then (m−1)/2 is odd. So, by Dirichlet's theorem, there is a prime p and positive integer u where

p= 4mu + (m−1)/2
= ((8u+1)m−1)/2

Let q = 8u+1. Then q>0 and

2p = qm − 1 ___ [8].

Then q ≡ 1, modulo 8, and p ≡ 1, modulo 4.

(−q\p)= (−1\p)(q\p)as quadratic residues are multiplicative
= (q\p)as p ≡ 1
= (p\q)by the quadratic reciprocity theorem
Also
(-1\q)= 1as q ≡ 1 modulo 4
(2\q)= 1as q ≡ 1 modulo 8
(-2\q)= 1as quadratic residues are multiplicative

If m modulo 8 ≡ 5, then (3m−1)/2 is odd. So, by Dirichlet's theorem, there is a prime p and positive integer u where

p= 4mu + (3m−1)/2
= ((8u+3)m−1)/2

Let q = 8u+3. Then q>0 and

2p = qm − 1 ___ [8].

Then q ≡ 3, modulo 8, and p ≡ 3, modulo 4.

(−q\p)= (−1\p)(q\p)as quadratic residues are multiplicative
= −(q\p)as p ≡ 3
= (p\q)by the quadratic reciprocity theorem, and p ≡ q ≡ 3
Also
(−1\q)= −1as q ≡ 3 modulo 4
(2\q)= −1as q ≡ 3 modulo 8
(-2\q)= 1as quadratic residues are multiplicative

Thus, to summarise, in all the last three cases,

(−q\p)=(p\q) ___ [9]

(−2\q)=1 ___ [10]

Thus:

(−q\p)= (p\q)(by [9])
= (p\q)(−2\q)(by [10])
= (−2p\q)as quadratic residues are multiplicative
= (1−qm\q)(by [8])
= (1\q)
= 1

Thus −q is a square modulo p. q is odd, so −q≡12, modulo 2, so (−q\2)=1. Thus

(−q\2p)=1

⇒ (−q\qm−1)=1

i.e. −q is a square modulo qm−1 as required. []

Theorem. 8k+3 is the sum of 3 odd squares.

Proof. By the above theorem, 8k+3 is the sum of 3 squares. Modulo 8, every square ≡ 0, 1 or 4. The only way the sum of 3 squares can ≡ 3 is 1+1+1. []

Theorem (Gauss). Every non-negative integer is the sum of 3 triangular numbers.

Proof. By the above theorem,

8k+3 = (2a+1)2 + (2b+1)2 + (2c+1)2

= 8 T a + 1 + 8 T b + 1 + 8 T c + 1

⇔ k = T a + T b + T c. []

Dr. Math says:

The proof that these are the only integers not so representable [as the sum of 3 squares] was given by Legendre, in Essai sur la theorie des nombres (1798), pp. 202, 398-9, and Gauss, Disquistiones Arithmeticae (1801), Section 291. It depends on the theory of ternary quadratic forms. I do not know the details, but it is very complicated. There is an alternative proof that is more elementary, longer, and also very complicated, which is due to Liouville and Uspensky. This appears in Uspensky and Heaslet, Elementary Number Theory (1939), pp. 465-474.

References

Rose, H.E., A Course in Number Theory, 2nd ed. (1994), pub. Oxford Science Publications, chapter 9, Quadratic Forms.


Home > Maths

Commentary

I simplify [R]'s notation:

[R]me
a_11a
a_12, a_21f
a_13, a_31e
a_22b
a_23, a_32d
a_33c
first cC
second cq
t_1X
t_2Y
t_3Z
x_1x
x_2y
x_3z

I wondered if there was a Lemma. Any quadratic form F, where D(F)!=0, is equivalent to one whose matrix is diagonal. Proof. D(F)!=0, so F's matrix A has n linearly indepedent eigenvectors, so A is diagonalisable, that is, there is a matrix T such that TAT−1 is a diagonal matrix. We may take T/det T instead of T. (However, this doesn't prove that a T exists whose entries are all integers.) In fact, no. [R]'s lemma 2.1 proves this for n<6. [R] then asserts that it is true for n=6 and 7, too. However, it is false for n=8: for 8 variables s, t, u, v, w, x, y, z, 2 [ s2+t2+u2+v2+w2+x2+y2+z2 - su - tv - uv - vw - wx - xy - yz ] is positive definite, and its determinant is 1, but it is not equvalent to a sum of squares.