2

I'm trying to fill in the details of the systematic approach of deriving the quadratic resolvent for the cubic equation according to this part of the 3Blue1Brown talk: https://youtu.be/aaW30_f2on0?t=4374. To sum up, given a cubic

$$x^3 + px + q$$

the idea is to find some polynomial expression in its roots $a, b, c$ with the orbit of size 2 under the actions of $S_3$. The suggested expressions

$$A = a^2b + b^2c + c^2a$$ $$B = a^2c + b^2a + c^2b$$

indeed constitute such an orbit. Therefore the quadratic polynomial $$(y - A)(y - B) = y^2 - (A +B)y + AB$$ has coefficients in the $p, q$. These coefficients can be expressed in the elementary symmetric polynomials ($e_1=0, e_2=p, e_3=-q$) as follows (I used Sage for the reduction): $$A+B = -3e_3 = 3q$$ $$AB = e_2^3 + 9e_3^2 = p^3 + 9q^2$$ Computing the discriminant of this quadratic equation yields: $$D = -4p^3 - 27q^2$$ This looks promising since this is up to a constant factor (and sign) the same expression as the one appearing in the final cubic formulas. Moreover, it coincides with the official discriminant of the cubic equation.

The question is, however, how to proceed from this point, i.e. how to derive the expressions for the roots $a, b, c$ of the original cubic. The only relations between the roots we have so far are the original expressions

$$A = a^2b + b^2c + c^2a$$ $$B = a^2c + b^2a + c^2b$$

We can now express the left-hand sides as (radical) expressions of $p$ and $q$ but of course this does not fully determine $a, b, c$.

  • I don't have the time to check out that video, but I suspect that you will still need to take complex cube roots at some point. Like in this old explanation of mine. But I may be wrong, also. – Jyrki Lahtonen Jan 15 '24 at 10:40
  • Cardano's formula, in a sense, relies on $(a+\omega b+\omega^2c)^3$ and $(a+\omega^2b+\omega b)^3$ as the roots of a quadratic resolvent (with the depressed cubic we have $a+b+c=0$, so if you know those three, you can solve for $a,b,c$ from a linear system). So Cardano makes the need for cube roots explicit. Hmm. I'm a bit curious if there is something I haven't seen before :-) – Jyrki Lahtonen Jan 15 '24 at 11:10
  • 1
    Correct, @OscarLanzi. There was a typo. Missus called me for lunch, so I neglected to proofread in time. Edited version above. Those are equal? Ok :-) Yes, unlimited ability to edit my comments is the one thing I miss from my diamond bearing days :-) – Jyrki Lahtonen Jan 15 '24 at 11:11
  • @JyrkiLahtonen I did check out the video. They focus on why similar algebraic combinations can't be consnstructed for all quintic equations. Basically you can't guarantee that such an expression equals exactly four different values (meaning the roots a quartic equation) because you have to satisfy invariance for both 3-cycles and 5-cycles, which is inconsistent with generating four values varying independently of the coefficient ratios. So the resolvent method that works for degree 3 or 4 breaks down. They note that the failure at 5 is connected with 5 being the second odd prime. – Oscar Lanzi Jan 15 '24 at 15:58
  • @OscarLanzi I think their argument is slightly different:
    • The action of a 5-cycle on an expression with an orbit of size 4 must be trivial (since gcd(5,4) = 1)
    • Every element of $S_5$ can be generated by a 5-cycle and a transposition.
    • Therefore such an expression cannot exist.
    – David Kubecka Jan 15 '24 at 22:24
  • @JyrkiLahtonen Only now I've noticed your original link where you actually do explain the origin of $u^3+v^3$ and $uv$. This is certainly simpler than the method involving my symmetric expressions. However, it's always nice to know a different way how to get the same result :-) – David Kubecka Jan 16 '24 at 22:56

2 Answers2

2

Following on Jyrki Lathonen's hint, render the polynomial power expansion below. The quantity $\omega$ is taken to be the unit root $(-1+i\sqrt3)/2$, thus $\omega^3=1$.

$(a+\omega b+\omega^2c)^3=a^3+b^3+c^3+3[a^2(\omega b)+a^2(\omega^2c)+a(\omega b)^2+a(\omega^2c)^2+(\omega b)^2(\omega^2c)+(\omega b)(\omega^2c)^2]+6a(\omega b)(\omega^2c)$

$=(a^3+b^3+c^3)+3\omega(a^2b+b^2c+c^2a)+3\omega^2(ab^2+bc^2+ca^2)+6abc$

$=(a^3+b^3+c^3)+3\omega A+3\omega^2 B-6q$

where we have plugged in $abc=-q$.

Now comes the subtle part. We have the factorization

$a^3+b^3+c^3-3abc=\color{blue}{(a+b+c)}(a^2+b^2+c^2-ab-ac-bc)$

and $a+b+c=0$! So $a^3+b^3+c^3=3abc=-3q$ and we finally have

$(a+\omega b+\omega^2c)^3=3\omega A+3\omega^2 B-9q$

We repeat the above with $\omega^2$ in place of $\omega$, so that $(\omega^2)^2=\omega^4=\omega$; thus

$(a+\omega^2b+\omega c)^3=3\omega^2 A+3\omega B-9q.$

There is, actually, a better way to render $(a+\omega^2b+\omega c)$. Form the product

$(a+\omega b+\omega^2c)(a+\omega^2b+\omega c)=(a^2+b^2+c^2)-(ab+ac+bc).$

Squaring the relation $a+b+c=0$ for the depressed cubic leads to $a^2+b^2+c^2=-2(ab+ac+bc)$, and we also have the symmetric sum $(ab+ac+bc)=p$. Thus

$(a+\omega b+\omega^2c)(a+\omega^2b+\omega c)=-3p.$

This avoids introducing a second cube root, and allows the roots $a,b,c$ to be rendered in terms of branches of a single cube root.

So

$a+b+c=0$

$a+\omega b+\omega^2c=\sqrt[3]{3\omega A+3\omega^2 B-9q}$

$a+\omega^2b+\omega c=-3p/\sqrt[3]{3\omega A+3\omega^2 B-9q}$

We now have a linear system with a coefficient matrix having a nonzero determinant (we compute the determinant as $3i\sqrt3$), so we have a unique solution for $a,b,c$. Dfferen branches of the single cube root $\sqrt[3]{3\omega A+3\omega^2 B-9q}$ correspond to different orderings of the same roots $,b,c$. When $A$and $B$ are real and distinct, the solution will contain complex cube roots, whose real and imaginary parts are not accessible by algebraic methods.

Oscar Lanzi
  • 48,208
  • This is great! To sum up, the overall approach is to find a linear system with coefficients expressed in p, g. One way to obtain such coefficients is to expand (++^2)^3 and observe that symmetric polynomials of orbit 2 are lurking there. This allows me to solve an equation of order 2 which is what I did in my post and that proves that such an approach is indeed feasible.

    Also, I hope (and will try) that the same approach for the quartic will reveal a polynomial with orbit 3 and that for the quintic there won't be such nice symmetries in the expansion.

    – David Kubecka Jan 15 '24 at 16:00
  • The question is how do you come up with the (++^2)^3 idea in the first place :-) Are there possibly different ways how to derive a linear system where symmetric polynomials appear? – David Kubecka Jan 15 '24 at 16:01
  • Regarding the key step 3+3+3−3=(++)(2+2+2−−−). I believe that this can also be "mechanically" derived using the elementary symmetric polynomials algorithm. – David Kubecka Jan 15 '24 at 16:02
  • 2
    I now probably see the motivation. It's analogous to the quadratic case where you can derive the solutions from a + b = r_1 and (a - b)^2 = ... where -1 is the primitive second root of unity. – David Kubecka Jan 15 '24 at 16:16
  • You are on target for the motivation for the $(a+\omega b+\omega^2c)^3$ expression. You might also want to render the product $(a+b+c)(a+\omega b+\omega^2c)(a+\omega^2b+\omega c)$ in terms of real variables. – Oscar Lanzi Jan 15 '24 at 16:19
  • I have finally managed to work out the method in detail for the cubic case and there are a couple of observations. First, it's definitely better to use the inverse Fourier transform the solve the final system--the $u^{1/3} + v^{1/3}$ nature of the solutions is then immediately clear. – David Kubecka Jan 16 '24 at 22:01
  • Second, the solution is in a certain sense nicer than the usual Cardano's formula because it naturally involves a square root of the discriminant $-4p^3 - 27q^2$ instead of its multiple. A slight disadvantage is that the $i$ appears in the formula directly whereas in the Cardano's formula it's "hidden". – David Kubecka Jan 16 '24 at 22:05
  • Third, this is the only method I know where my original symmetric expressions A and B for the quadratic resolvent come naturally. In the other methods (Cardano and the "usual" Lagrange), the alternative expressions $u^3 + v^3$ and $uv$ come seemingly out of the blue. This magical aspect was in fact the main motivation why I asked the question in the first place! – David Kubecka Jan 16 '24 at 22:11
0

One standard approach to solving the cubic is to write $x=u+v$, in which case we have a solution provided $uv=-p/3$ and $u^3+v^3=-q$. Then $u^3,v^3$ are the solutions of the auxiliary quadratic $t^2+qt-p^3/27$, and one root of the original cubic is of course $\sqrt[3]{u^3}+\sqrt[3]{v^3}$.

Comparing this quadratic to yours, we have $A=\lambda u^3+\frac12q(3+\lambda)$ and $B=\lambda v^3+\frac12q(3+\lambda)$, where $\lambda^2=-27$. It is then clear that the choices for $A$ and $B$ are not necessarily the most appropriate ones for obtaining the roots for the cubic.

Note also that one often takes $\delta=(a-b)(a-c)(b-c)$, so that $\delta$ is fixed by all even permutations of $a,b,c$, and sent to its negative under odd permutations. Expanding out we have $\delta=B-A=\lambda(v^3-u^3)$.