28

If $a,b,c,d$ are non negatives and $a^2+b^2+c^2+d^2=3$ prove that $$abcd+3\ge a+b+c+d$$

The inequality is not as simple as it looks.The interesting part is that the equality occurs when $a=0,b=c=d=1$ upto permutation .(I don't know if there are further equality cases.)

I tried rewriting the inequality as $$a^2+a(bcd-1)+b^2+c^2+d^2-b-c-d\ge 0$$ As its a quadratic in $a$ it suffices to show the discriminant$$\Delta_a={(bcd-1)}^2-4(b^2+c^2+d^2-b-c-d)\le 0$$ which unfortunately is wrong (when $a=\sqrt{3},b=c=d=0$)

P.S ;I am not aware of using Lagranges multipliers

Its from here

Update :An answer has been posted in the link above (AOPS) similar to Dr Mathva's answer

  • 1
    I haven't figured out how to use that but in post #6 arqady (I guess a.k.a. @MichaelRozenberg at MSE) seems to give a hint... – VIVID Jan 18 '21 at 14:30
  • @VIVID nice find! thanks for your interest – Hari Ramakrishnan Sudhakar Jan 18 '21 at 14:34
  • As the problem is symmetric with respect to all variables, I tried to take $a=b=c=d=\frac{\sqrt 3}{2}$ and got $$a b c d-(a+b+c+d)\ge \frac{9}{16}-2 \sqrt{3}\approx -2.9$$ which is the exact minimum. I don't write this as an answer because without analysis I can't prove that the minimum happens when the variables are equal. – Raffaele Jan 18 '21 at 15:39
  • @Raffaele interesting ,that means this inequality is infact weaker...... – Hari Ramakrishnan Sudhakar Jan 18 '21 at 15:49
  • @AlbusDumbledore Yes. Pierrecarre has proved it, using real analysis methods (Lagrange multipliers). (Please edit to correctly write Lagrange's name, greatest Italian mathematician :) – Raffaele Jan 18 '21 at 15:56
  • 3
    -- FYI -- The statement holds more generally for three, or more than four real variables without positivity assumption: $$\text{If $,3\leqslant n,$ and};\sum^n_{d=1}a_d^2=n-1, \text{ then }:\sum^n_{d=1}a_d:\leqslant:\prod^n_{d=1}a_d+n-1,$$ cf page 57 in the GAZETA MATEMATICA . Notice that $,a_d,$ encodes Albus Dumbledore ... – Hanno Jan 21 '21 at 16:11

4 Answers4

17

Due to symmetry, assume wlog that $a\geqslant b\geqslant c\geqslant d\geqslant 0$. Since I will use this fact later, I am proving it before the cases:

Lemma: We have that $2\geqslant a+d$.

In fact, this follows from Cauchy-Schwarz: $$2= 2\sqrt{\frac{a^2+b^2+c^2+d^2}{3}}\geqslant 2\sqrt{\frac{a^2+3d^2}{3}}=\sqrt{1+\frac13}\cdot \sqrt{a^2+3d^2}\geqslant a+d$$

We will now consider some cases depending on the position of the number $1$:

  • Case 1: $1\geqslant a\geqslant b\geqslant c\geqslant d\geqslant 0$. Observe that one may rewrite the inequality as $$3+abcd-(a+b+c+d)=(1-a)(1-b)+(1-c)(1-d)+(1-ab)(1-cd) $$ Which is clearly positive.

  • Case 2: $a\geqslant 1\geqslant b\geqslant c\geqslant d\geqslant 0$. If $ab\geqslant 1$, we might proceed as in the next case. So we will work with $ab<1$. Thus \begin{align*}3+abcd-(a+b+c+d)&=(1-a)(1-b)+(1-c)(1-d)+(1-ab)(1-cd)\\&\geqslant \underbrace{(1-a)}_{<0}(1-b)+(1-c)(1-d)\\&\geqslant (1-a)(1-c)+(1-c)(1-d)\\&=(1-c)(2-(a+d))\geqslant 0 \end{align*} Where the last inequality follows from the lemma.

  • Case 3: $a\geqslant b\geqslant 1\geqslant c\geqslant d\geqslant 0$. This implies that \begin{align*}6+2abcd-2(a+b+c+d)&=a^2+b^2+c^2+d^2+3+2abcd-2(a+b+c+d)\\ &=(a-1)^2+(b-1)^2+(c+d-1)^2+2cd(ab-1)\geqslant 0\end{align*} Or, equivalently $3+abcd\geqslant a+b+c+d$.

  • Case 4: $a\geqslant b\geqslant c\geqslant 1\geqslant d\geqslant 0$. As @dezdichado noticed, this case is straightforward, since it forces directly $a=b=c=1, d=0$ due to the constraint $a^2+b^2+c^2+d^2=3$.

  • Case 5: $a\geqslant b\geqslant c\geqslant d\geqslant 1$. Clearly impossible, since this would yield $a^2+b^2+c^2+d^2\geqslant 4>3$.

Done!

Calvin Lin
  • 77,541
Dr. Mathva
  • 9,003
  • 4
    your case $3$ would have forced $a = b= c = 1$ and $d = 0$ anyway due to the constraint. +1 – dezdichado Jan 18 '21 at 16:06
  • Ingenious(+1) I doubt a better answer will be there – Hari Ramakrishnan Sudhakar Jan 18 '21 at 16:12
  • Thanks @AlbusDumbledore! Although I am disappointed that I could not use Michael Rozenberg's hint on aops. I thought that Turkevichi's inequality would lead somewhere, but was unable to prove that $ab+bc+cd+da+ca+bd\geqslant a+b+c+d+1$ (I am not even sure if this is true ...) – Dr. Mathva Jan 18 '21 at 16:14
  • @Dr.Mathva that was my same thought of using the turkevicus inequality :) Micheal Rozenberg has surprisingly been not answering questions for the past 3 months else i am sure he would have solved the problem using his hint:).Nevertheless this answer is worth a bounty! – Hari Ramakrishnan Sudhakar Jan 18 '21 at 16:19
  • 3
    I don't think so? E.g. $a=\sqrt{3}, b=c=d=0.$ Then L.H.S. $=0$ but R.H.S. =$\sqrt{3}+1$. – Hello Jan 18 '21 at 16:20
  • @SeeHai Thanks for pointing that out; I feel so dumb to have missed it ... – Dr. Mathva Jan 18 '21 at 16:44
  • 3
    @AlbusDumbledore I also miss Michael Rozenberg; he's able to solve (almost) every inequality, and some of them with crazy computations (such as BW or uvw ones)!! It's sad that he does not answer questions anymore on this site (does anybody now why?). Anyway, he still comes and looks around, since he's been active at least yesterday. Hope he comes back – Dr. Mathva Jan 18 '21 at 16:48
  • @Dr. Mathva please don't say so! Your solution was very nice indeed. – Hello Jan 18 '21 at 16:51
  • 2
    @Dr.Mathva :) He is still answering questions in AOPS, I am a fan of him too! – Hari Ramakrishnan Sudhakar Jan 18 '21 at 16:52
  • One way I was thinking of doing this problem (probably will get the best bound) was to extend the $uvw$ to a $uvwx$ for four variables, because this is useful for inequalities where the number of variables is important. But by the time I could figure out five to six base equalities to work with, your answer was up! +1 – Sarvesh Ravichandran Iyer Jan 19 '21 at 14:10
  • How do you say that $b^2 + c^2 + d^2 \geq 3d^2$? – Clemens Bartholdy Jan 20 '21 at 09:28
  • I assumed wlog that $a\geqslant b\geqslant c\geqslant d$. Therefore, since the variables are positive $$b^2\geqslant d^2, c^2\geqslant d^2, d^2\geqslant d^2\implies b^2+c^2+d^2\geqslant 3d^2$$ – Dr. Mathva Jan 20 '21 at 10:30
  • @Dr.Mathva An answer has been posted in AOPS forum see https://artofproblemsolving.com/community/c6t243f6h404087_inequality – Hari Ramakrishnan Sudhakar Feb 21 '21 at 14:18
7

Since $a^2+b^2+c^2+d^2 = 3$ defines a compact set in $\mathbb{R}^4$, $f(a,b,c,d)=abcd+3-a-b-c-d$ will have a global minimum and maximum over that set, that will occur (can be easily justified) in critical points of the Lagrangian $$ L(a,b,c,d,\lambda) = abcd+3-a-b-c-d -\lambda(a^2+b^2+c^2+d^2-3) $$

So, just compute the critical points, compute the value of $f$ over each of these points and get the global minimum. If the global minimum is $\ge 0$, you are done.

When computing the critical points, you will get some complex solutions, some solutions with negative components but, in the end, the relevant solutions are the ones you already mentioned and also $a=b=c=d=\frac{\sqrt{3}}{2}$. The minimum value of $f$ (even admitting negative values for $a,b,c,d$) is zero (and the maximum is $\frac{57}{16}+2\sqrt{3}$).

Edit: Missed one critical point, but the result stands.

PierreCarre
  • 23,092
  • 1
  • 20
  • 36
  • sorry but what is a compact set and langrangian thing i am not at all aware – Hari Ramakrishnan Sudhakar Jan 18 '21 at 14:11
  • @AlbusDumbledore A compact set in $\mathbb{R}^4$ is a bounded and closed set. THis is relevant because a continuous function ($f$) always attains a global maximum and minimum values over a compact set. – PierreCarre Jan 18 '21 at 14:14
  • @AlbusDumbledore Most likely, that is the Lagrange Multiplier method – VIVID Jan 18 '21 at 14:15
  • The Lagrangian is the the function $L$ I defined in the post and is used as a tool to compute candidates to min/max with restrictions. – PierreCarre Jan 18 '21 at 14:15
  • The critical points of $L$ are the vectors $(a,b,c,d,\lambda)$ that make all the partial derivatives be zero simultaneously. – PierreCarre Jan 18 '21 at 14:17
  • 1
    IHMO, the bulk o f Lagrange multiplier approach is always working on the derivatives so as to find the global extrema, which your answer didn't shed any light on. Your answer might just be a comment as VIVID posted. – Quang Hoang Jan 18 '21 at 14:24
  • @QuangHoang Providing an answer does not mean making all the calculations. I think I outlined the milestones. – PierreCarre Jan 18 '21 at 14:28
  • 3
    My point is, yes, one doesn't need to provide all the calculations, but your answer is simply try Lagrangian Multiplier. It doesn't provide any further information specific to this question other than the generic method. E.g. Is it a straight forward process as solving a quadratic polynomial or what should OP expect from using this method to this question? – Quang Hoang Jan 18 '21 at 14:32
  • @PierreCarre This is supposed to be a contest question and such methods i am not aware even if one is able to solve those equaions,I had also mentioned i am not ware of langranges multipliers .Any way maybe i van fathom your idea after i learn more mathematics – Hari Ramakrishnan Sudhakar Jan 18 '21 at 14:47
  • @AlbusDumbledore I did not interpret your comment on not being aware of the method as a way of dismissing it as a solution process. I'm sure there are purely algebraic methods to solve the question and, I would gladly provide one if I had found it! – PierreCarre Jan 18 '21 at 15:03
  • @PierreCarre OK ,thanks for your effort grateful for that! – Hari Ramakrishnan Sudhakar Jan 18 '21 at 15:05
  • @dezdichado No. The restriction places us in the surface of the ball. The process I proposed allows the calculation of the global max/min on the surface of the ball. – PierreCarre Jan 18 '21 at 15:06
  • Yeaj my bad lol i confuser myself with the second derivative test – dezdichado Jan 18 '21 at 15:07
  • global extrema attains at either critical points or on boundary. Your solution did not consider boundary point, and also did not show that the critical point you (sort of guess) is a maximum or a minimum. The computed value is obviously larger than the case when $a=b=c=1, d=0$, which gives the minimum stated in the problem. – Quang Hoang Jan 18 '21 at 15:09
  • @QuangHoang I think you are confused about how Lagrange multipliers work. What do you think is the boundary in this case? We are computing the max/min over a set that is the surface of a sphere. Considerations about the rank of the Jacobian matrix of the restrictions allow you to say that in this case the max/min will occur on critical points. And the global max/min will be one of those. – PierreCarre Jan 18 '21 at 15:23
  • @QuangHoang And of coarse the value on this other point will be greater than the global minumum! But it is necessary to look at this additional point to make sure that the global mimumun is attained at the other critical points. – PierreCarre Jan 18 '21 at 15:26
  • 1
    @QuangHoang the solution is fine. Smooth function on a compact set attains its maximum and minimum in that set and they can only be achieved at a critical point. Then, you use Lagrange to find all the critical points and then find the maximum and minimum and by just looking at their values. Your concern would be valid if there happen to be an infinite number of critical points, so inspecting them individually would be hard, but it's not the case here assuming his computation is right. – dezdichado Jan 18 '21 at 15:44
  • $a=b=c\approx0.721545, d\approx1.19922$ and $a=b\approx0.556325, c=d\approx1.09110$ are other critical points – didgogns Sep 14 '22 at 16:56
4

This is a remark on Dr. Mathva's answer, which may seem very mysterious at first but becomes more transparent when you consider a simpler problem:

If $a, b \geq 0$ with $a^2 + b^2 = 1$, prove that $$ ab+1 \geq a+b \,.$$

(It is the problem from the question if you impose $c = d = 1$.)

Here it is clear that you can factorize it as $(a-1)(b-1) \geq 0$, so you are naturally led to distinguish cases according to whether $a, b \geq 1$ or $\leq 1$. If both are $\geq 1$ or both are $\leq 1$, we are done. So suppose wlog $a \geq 1 \geq b$. Then plug in the constraint to get $$ab + a^2+b^2 \geq a+b \,.$$ Now play a bit to recover every term in the RHS from the LHS. Because $a \geq 1$, we can just remove factors $a$ from the LHS. That is, $ab \geq b$ and $a^2 \geq a$, and we are done.

By considering this simpler problem, you can more easily get the idea of assuming wlog $a \geq b \geq c \geq d$, and of distinguishing cases according to the position of $1$.

Bart Michels
  • 26,985
  • 6
  • 59
  • 123
3

See, in the link below, our extension to n real variables (GMA Problem 494, page 57, with the our solution): https://ssmr.ro/gazeta/gma/2020/gma3-4-2020-continut.pdf


Screenshots from the link provided:

enter image description here

enter image description here

enter image description here

enter image description here

enter image description here

enter image description here

VIVID
  • 11,667