Let $\mathbb F$ be field with $q$ element and $\operatorname{char}(\mathbb F)\neq2$. I want to know about the number of solutions of this equation:

$$x^2_1+x^2_2+\dots+x^2_n=0 \text{ where } x_i\in \mathbb F.$$

Any suggestion?

  • $\begingroup$ Is there any context for this problem? Motivation? As is, the question appears quite difficult to answer. Even if we take the field to be $\mathbb{Z}_p$ and $n$ to be $2$ or $3$ this problem isn't easy. $\endgroup$ – Ragib Zaman Aug 16 '12 at 6:12
  • 2
    $\begingroup$ Actually, $n=2$ is very easy. If $-1$ is not a square (and that's easy to decide) then there's only 1 solution, else, $2q-1$. $\endgroup$ – Gerry Myerson Aug 16 '12 at 6:25
  • 1
    $\begingroup$ I have read another equation and I thaugh for this equation. You can look page79 form Classical group and geometric algebra by Grove $\endgroup$ – Babak Miraftab Aug 16 '12 at 6:26
  • 5
    $\begingroup$ @Ragib: actually this is one of the few families of polynomials for which a reasonably straightforward answer is possible (morally because a variety defined by a homogeneous quadratic polynomial is either empty or birational to projective space). Past the quadratic case things become much harder... $\endgroup$ – Qiaochu Yuan Aug 16 '12 at 6:33

Partial answer that works, when $q\equiv1\pmod4$. This is a minor simplification to the use of recurrences as described in the first sentence of Gerry Myerson's answer (that I somehow missed as my gaze was drawn to the character sums :/). For the cases $q\equiv-1\pmod4$ I recommend doing it his way.

When $q\equiv1\pmod4$ we know that the field contains a primitive fourth root of unity, call it $i$. Let $A(n)$ be the number of solutions for the equation in $n$ variables. We know (see also comment's by Gerry Myerson) that $A(1)=1$ and $A(2)=2q-1$.

We can write $$x_1^2+x_2^2=(x_1+ix_2)(x_1-ix_2)=z_1z_2,$$ where $z_1=x_1+ix_2$ and $z_2=x_1-ix_2$. The mapping $(x_1,x_2)\mapsto(z_1,z_2)$ is clearly a bijection, so $A(n)$ is equal to the number of solutions of $$ z_1z_2=-(x_3^2+x_4^2+\cdots+x_n^2). $$ Here the r.h.s. is equal to zero for exactly $A(n-2)$ vectors $(x_3,x_4,\ldots,x_n)$, so that gives us $(2q-1)A(n-2)$ solutions. OTOH, if the r.h.s. is non-zero, we can let $z_2$ be any non-zero element ($(q-1)$ choices), and can then uniquely solve for $z_1$. This gives us $(q-1)(q^{n-2}-A(n-2))$ solutions. Altogether we get a recurrence relation $$ A(n)=(2q-1)A(n-2)+(q-1)\left(q^{n-2}-A(n-2)\right)=(q-1)q^{n-2}+qA(n-2). $$

It is easy to prove by induction on $n$ that this has the solution $$ A(2k+1)=q^{2k} $$ for odd values of $n=2k+1$, and $$ A(2k)=q^{2k-1}+q^k-q^{k-1} $$ for even values of $n=2k$.

  • $\begingroup$ This is closely related to the birational equivalence Qiaochu is talking about in his comment. $\endgroup$ – Jyrki Lahtonen Aug 16 '12 at 7:41

Let $A_n$ be an abbreviation for $x_1^2+x_2^2+\cdots+x_n^2$. Let $a_n$ be the number of solutions of $A_n=0$, let $b_n$ be the number of solutions of $A_n=-1$, and let $c_n$ be the number of solutions of $A_n=c$, where $-c$ is some fixed nonsquare in the field. Find a system of recurrences involving $a_n$, $b_n$, and $c_n$, and solve.

Alternatively, use exponential sums. This is easiest to do when $q$ is prime. $$\sum_{x_1}\sum_{x_2}\cdots\sum_{x_n}\sum_te^{2\pi it(x_1^2+x_2^2+\cdots+x_n^2)/p}$$ gives $p$ times the number of solutions, where the sums are over all $x_1,x_2,\dots,x_n,t$ in the field. Bring the sum on $t$ to the front, separate out the term with $t=0$ to get $$p^n+\sum_{t=1}^{p-1}\sum_{x_1}\sum_{x_2}\cdots\sum_{x_n}e^{2\pi it(x_1^2+x_2^2+\cdots+x_n^2)/p}=p^n+\sum_{t=1}^{p-1}\left(\sum_xe^{2\pi itx^2/p}\right)^n$$ The sum on the inside is a quadratic Gauss sum whose value is recorded in many texts and on many websites.

Similar methods apply when $q=p^r$, $r\gt1$, but I disremember the trick.

Note the problem $x_1^2+x_2^2+\cdots+x_n^2=1$ (over the field of $p$ elements) is done in detail in Ireland and Rosen, A Classical Introduction to Modern Number Theory, Chapter 8, Section 6, pages 101-102. $x_1^2+x_2^2+\cdots+x_n^2=0$ is left as Exercise 19 in that chapter. The tools for extending the results to finite fields more generally are developed in Chapter 10.

  • $\begingroup$ thanks, I am looking for simple formula for this like the number of k-dimensional subspace of n-dimensional space $\endgroup$ – Babak Miraftab Aug 16 '12 at 6:34
  • 2
    $\begingroup$ Well, I've shown you two ways to find a formula, and a book that does it for you. I don't know what you consider to be a simple formula, and there's no point in looking for one if it doesn't exist. $\endgroup$ – Gerry Myerson Aug 16 '12 at 6:41
  • $\begingroup$ @Garry Myerson I saw this exercise I don't have information about character of order 2. I was wondering if you could tell me explicit formula by p and n. $\endgroup$ – Babak Miraftab Aug 16 '12 at 7:14
  • 1
    $\begingroup$ Sure, if you can wait a couple of weeks while I try to find the time. In the meanwhile, you could always try to work it out on your own from the information I've already given. $\endgroup$ – Gerry Myerson Aug 16 '12 at 13:06
  • $\begingroup$ :I appreciate you if you calculate for me. I looked at page 101 and 102 Modern Number Theory. they use character of order 2 $\endgroup$ – Babak Miraftab Aug 16 '12 at 13:32

I'll try to answer this in a self-contained manner - I will repeat ideas by Paul and Gerry but it will be more detailed. I'll restrict myself to $q = p$ (a prime), but the same arguments are valid in the general case.

We'll generalize, because the problem calls for it: you want to count solutions to $\sum_{i=1}^{n} x_i^2 = 0$ - why not understand the distribution of $f(x_1,\cdots,x_n) = \sum_{i=1}^{n} x_i^2$, i.e. how many times each value in $\mathbb{F}_{p}$ appears in the multiset $\{ f(\vec{x}) | \vec{x} \in \mathbb{F}_{p}^{n} \}$.

This problem is a convolution problem - let $v$ be the vector that counts solutions to $x^2 = i$ - its i'th coordinate is the number of square roots of $i$ in $\mathbb{F}_{p}$. If we convolve $v$ with itself $n$ times, you'll get the distribution of our function $f$, and you need the 0'th coordinate: $\underbrace{v * \cdots * v}_\textrm{n times}$.

You can approach this via Fourier Transform (so-called "Gauss sums"), since Fourier turns convolution into pointwise multiplication which is easier, but we can avoid that.

The key step is noting that $v*v$ is almost constant - it has a certain value on 0, and another value on all $\mathbb{F}_{p}^{\times}$. Why is that?

  1. If $p \equiv 1 (4)$, there's a square root of -1 in $\mathbb{F}_{p}$, call it $i$. Then the equation $x^2 + y^2 = j$ is equivalent to $(x+iy)(x-iy) = j$. We can make the substitution $x+iy=t, x-iy=s$ (since $char(F) \neq 2$) and this turns to $st = j$. If $j=0$ there are $2p-1$ solution, else - there are $p-1$ solutions ($(s,js^{-1})$).

  2. If $p \equiv 3 (4)$, there's no square root of -1 living in our field, but we can construct a quadratic extension of our field, which will contain it - $i \in \mathbb{F}_{p}[x]/(x^2+1) \cong \mathbb{F}_{p^2}$ (as there's a unique quadratic extension of any finite field). The equation $x^2 + y^2 = j$ becomes : $$x^2 + y^2 = (x+iy)(x-iy) = (x+iy)^{p+1} = j$$ because $x-iy$ is the conjugate of $x+iy$, and Frobenius Automorphism $t \to t^p$ takes elements to their conjugates. (Since it's linear, you can just verify $1^p = 1, i^p = -i$). We can write this equation as $t^{p+1} = j$ where $j \in \mathbb{F}_{p}, t \in \mathbb{F}_{p^2}$. Note that $\mathbb{F}_{p} = (\mathbb{F}_{p^2})^{p+1}$ (one inclusion follows since for $t \neq 0$, $(t^{p+1})^{p-1} = t^{p^2 - 1} = 1$, so $t^{p+1} \in \mathbb{F}_{p}$). Since $(\mathbb{F}_{p^2})^{\times}$ is cyclic, for $j \neq 0$ this reduces to the equation $$(p+1)x \equiv (p+1)a (\mod p^2-1)$$ which has $p+1$ solutions in $\mathbb{Z}/(p^2-1)\mathbb{Z}$. For $j=0$, this clearly has one solution - $t=0$, i.e. $x=y=0$.

So $v*v$ is nice - as a function of $i$, it depends only on whether $i =0$ or not.

Now, a nice exercise: Let $u = (a,\underbrace{b, \cdots , b}_\textrm{p-1 times})$. $v = (c,\underbrace{d, \cdots , d}_\textrm{p-1 times})$. Then $u*v = (e,\underbrace{f, \cdots , f}_\textrm{p-1 times})$, where $(e-f) = (a-b)(c-d)$ and $e + (p-1)f = (a + (p-1)b)(c + (p-1)d)$. If $a=c, b=d$, we get: $(e-f)=(a-b)^2, e+(p-1)f = (a+(p-1)b)^2$. This is where the Fourier aspect is hidden - convolution of vectors multiplies the frequencies.

We're really close.

  1. If $n$ is even, then $\underbrace{v * \cdots * v}_\textrm{n times} = \underbrace{(v * v) * \cdots * (v*v)}_\textrm{n/2 times}$, and this is easy by the exercise - if we apply the exercise $\frac{n}{2}$ times, our final vector $w$ will be of the form $(a,\underbrace{b, \cdots , b}_\textrm{p-1 times})$, where $a-b = \begin{cases} \ ((2p-1) - (p-1))^{n/2} &\text{if } p \equiv 1 \pmod{4}\\ ((p+1) - (1))^{n/2} & \text{if } p\equiv 3 \pmod{4} \end{cases} $ and $a+b(p-1) = (p^2)^{n/2}=p^n$. Solving these 2 equation gives $a$, which is the number you're seeking.

  2. If $n$ is odd, we first calculate $\underbrace{v * \cdots * v}_\textrm{n-1 times}$ similarly to the first case. Then we need to convolve $(a,\underbrace{b, \cdots , b}_\textrm{p-1 times})$ with $v$, which turns out to be a relatively easy problem: the i'th coordinate will be $(a-b)v_i + pb$ (do it yourself). For $i=0$, $v_i = 1$ and the answer is $a-b+pb = (p-1)b+a$.

So there are 4 cases, depending on whether $i \in \mathbb{F}_{p}$ ($p \equiv 1(4)$) and on the parity of $n$.


Another way to get a recursion is by first noting that, whether or not $\sqrt{-1}$ lies in the ground field $\mathbb F_q$, every element of $\mathbb F_q$ is representable as $x^2+y^2$: when there is a $\sqrt{-1}$, this factors, and, when there is no $\sqrt{-1}$, this is the norm from the quadratic extension $\mathbb F_{q^2}$, which is surjective.

To hit $0$, there is a unique solution of $x^2+y^2=0$ without $\sqrt{-1}$, while there are $2q-1$ solutions with $\sqrt{-1}$. For non-zero values $z$, there are $q-1$ solutions of $x^2+y^2=z$ with $\sqrt{-1}$, and $q+1$ without.

Thus, there are two families of cases, depending on presence or absence of $\sqrt{-1}$, and incorporating whether one is hitting $0$ or not. That is, looking at $x^2+y^2=1-(x_3^2+\ldots+x_n^2)$, treat the cases that the right-hand side is $0$, or not, etc.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.