7

Let $a$, $b$, $c$ be non-negative real numbers such that $ab+bc+ca=1$. Find the minimum of $$\sqrt{\frac a{8ab+1}}+\sqrt{\frac b{8bc+1}}+\sqrt{\frac c{8ca+1}}.$$ Source: Crux Mathematicorum

I tried the cases $a=b=c$ and $a=0$, and I guess that the minimum is $\frac2{\sqrt3}$, obtained when $(a,b,c)=\left(0,3,\frac13\right)$.

I thought of using Hölder inequality, but this might be a bad idea: If we apply something like $$\left(\sum_{\rm cyc}\sqrt{\frac a{8ab+1}}\right)^2\sum a^2(8ab+1)s_a^3\ge\left(\sum_{\rm cyc}as_a\right)^3,$$ then it is difficult to homogenize because the given condition is second degree.

Also I tried rearrangement inequality. Let $x\ge y\ge z$ be a permutation of $a$, $b$, $c$, then $xy+yz+zx=1$, and $$\sum_{\rm cyc}\sqrt{\frac a{8ab+1}}\ge\sqrt{\frac x{8xy+1}}+\sqrt{\frac y{8xz+1}}+\sqrt{\frac z{8yz+1}}.$$ Now it is half symmetric with $x$ and $z$. But the usual method $x+z=p$, $xz=q$ doesn't seem to help because the condition would turn into a complicated form $pc+q=1$.

youthdoo
  • 3,663
  • The command of MMA 14.2.1 NMinimize[{p + r + s, p^2 == a/(8*a*b + 1) && r^2 == b/(8*b*c + 1) && s^2 == c/(8*c*a + 1) && a*b + b*c + c*a == 1 && a >= 0 && b >= 0 && c >= 0 && p >= 0 && r >= 0 && s >= 0}, {a, b, c, p, r, s}] results in {1.15469,{a->0.333039,b->0.,c->3.00265,p->0.57709,r->0.,s->0.577597}}. This suggests that an optimal solution is $a=\frac 1 3, b=0, c=3$. – user64494 Jun 18 '25 at 17:04

2 Answers2

2

By AM-GM $$\sum_{cyc}\sqrt{\frac{a}{8ab+1}}=\sqrt{\sum_{cyc}\frac{a}{8ab+1}+2\sum_{cyc}\sqrt{\frac{ab}{(8ab+1)(8bc+1}}}=$$ $$=\sqrt{\sum_{cyc}\frac{a}{8ab+1}+2\sqrt{\sum_{cyc}\frac{ab}{(8ab+1)(8bc+1}+2\sum_{cyc}\sqrt{\frac{ab^2c}{(8ab+1)(8bc+1)^2(8ca+1)}}}}\geq$$ $$\geq\sqrt{\sum_{cyc}\frac{a}{8ab+1}+2\sqrt{\sum_{cyc}\frac{ab}{(8ab+1)(8bc+1)}+6\sqrt[3]{\frac{a^2b^2c^2}{\prod\limits_{cyc}(8ab+1)^2}}}}$$ and it's enough to prove that $$\sum_{cyc}\frac{a}{8ab+1}+2\sqrt{\sum_{cyc}\frac{ab}{(8ab+1)(8bc+1)}+6\sqrt[3]{\frac{a^2b^2c^2}{\prod\limits_{cyc}(8ab+1)^2}}}\geq\frac{4}{3}$$ or $$\sum_{cyc}a(8bc+1)(8ca+1)+$$ $$+2\sqrt{\prod_{cyc}(8ab+1)\left(\sum_{cyc}ab(8ca+1)+6\sqrt[3]{a^2b^2c^2\prod_{cyc}(8ab+1)}\right)}\geq\frac{4}{3}\prod_{cyc}(8ab+1)$$

and since by $uvw$ $$\prod_{cyc}(8ab+1)\left(\sum_{cyc}ab(8ca+1)+6\sqrt[3]{a^2b^2c^2\prod_{cyc}(8ab+1)}\right)\geq\left(3+92abc+64a^2b^2c^2\right)^2,$$ it's enough t prove $$\sum_{cyc}a(8bc+1)(8ca+1)+6+184abc+128a^2b^2c^2\geq\frac{4}{3}\prod_{cyc}(8ab+1)$$ or $$3(a+b+c)-18+24\sum_{cyc}a^2c+816abc-256abc(a+b+c)\geq1664a^2b^2c^2$$ or $$\sum_{cyc}(3a^2b+27a^2c+275abc)\geq$$ $$\geq18(ab+ac+bc)^3+256abc(a+b+c)(ab+ac+bc)+1664a^2b^2c^2$$ or $$9(ab+ac+bc)^3(a^2b+b^2c+c^2a+9(a^2c+b^2a+c^2b)+275abc)^2\geq$$ $$\geq\left(18(ab+ac+bc)^3+256abc(a+b+c)(ab+ac+bc)+1664a^2b^2c^2\right)^2,$$ which is true by BW

  • According to the result of the command of MMA Resolve[ForAll[{a, b, c}, 9 (a*b + a*c + b*c)^3 (a^2*b + b^2*c + c^2*a + 9 (a^2*c + b^2*a + c^2*b) + 275 a*b*c)^2 >= (18 (a*b + a*c + b*c)^3 + 256 a*b*c (a + b + c) (a*b + a*c + b*c) + 1664*a^2*b^2*c^2)^2], NonNegativeReals], the latest inequality is true – user64494 Jun 20 '25 at 08:25
  • @user64494 I can prove the last inequality by hand and I am ready to show(just it a bit of ugly). BW gives a possibility to see it immediately. Thank you for your interest! – Michael Rozenberg Jun 20 '25 at 08:31
  • After reaching the inequality$$\sum_{cyc}a(8bc+1)(8ca+1)+2\sqrt{\cdots}\geq\frac{4}{3}\prod_{cyc}(8ab+1)$$can we directly remove the square root by squaring? Does it lead to a correct inequality? – youthdoo Jun 20 '25 at 14:42
  • @youthdoo I think it's nothing because the right hand side may be negative. But by your question I understood that you see that the second step in my proof is hardest here. – Michael Rozenberg Jun 20 '25 at 15:02
  • I saw that $64$ and $92$ do not affect the equality case. Is there any reason why you chose these coefficients? – youthdoo Jun 21 '25 at 03:44
  • @youthdoo The equality in the second step occurs for $abc=0$. Also, I tried to chose coefficients such that this inequality would be true for closing to $a=b=c$ because it happens for the starting inequality. Also, we need that the inequality in the second step would be true. The last thing we can get by $uvw$ because this inequality it's $f(u)\geq0,$ where $f$ increases, and we obtain a polynomial inequality of one variables. After all these things we need that the last inequality would be true and it turned out. – Michael Rozenberg Jun 21 '25 at 04:39
0

Here is the solution done with Mathematica 14.2.1. Firstly, we switch to polynomials in the target function by the change $p^2=\frac a{8ab+1},r^2=\frac b{8bc+1},s^2=\frac c{8ca+1},p\ge0,r\ge 0,s\ge 0$. Then the target function becomes $p+r+s$. Secondly, we find where the restrictions hold by the MMA command (see the documentation for info)

CylindricalDecomposition[{(8*a*b + 1)*p^2 == a && (8*b*c + 1)*r^2 == 
b && (8*c*a + 1)*s^2 == c && a*b + b*c + c*a == 1 && a >= 0 && 
b >= 0 && c >= 0 && p >= 0 && r >= 0 && s >= 0}, {a, b, c, p, r, s}]

(a == 0 && b > 0 && c == 1/b && p == 0 && r == Sqrt[b/(1 + 8 b c)] && s == Sqrt[c]) || (a > 0 && ((0 <= b < 1/a && c == (1 - a b)/(a + b) && p == Sqrt[a/(1 + 8 a b)] && r == Sqrt[b/(1 + 8 b c)] && s == Sqrt[c/(1 + 8 a c)]) || (b == 1/a && c == (1 - a b)/(a + b) && p == Sqrt[a/(1 + 8 a b)] && r == Sqrt[b/(1 + 8 b c)] && s == -Sqrt[(c/(1 + 8 a c))]))) $$\left(a=0\land b>0\land c=\frac{1}{b}\land p=0\land r=\sqrt{\frac{b}{8 b c+1}}\land s=\sqrt{c}\right)\lor \left(a>0\land \left(\left(0\leq b<\frac{1}{a}\land c=\frac{1-a b}{a+b}\land p=\sqrt{\frac{a}{8 a b+1}}\land r=\sqrt{\frac{b}{8 b c+1}}\land s=\sqrt{\frac{c}{8 a c+1}}\right)\lor \left(b=\frac{1}{a}\land c=\frac{1-a b}{a+b}\land p=\sqrt{\frac{a}{8 a b+1}}\land r=\sqrt{\frac{b}{8 b c+1}}\land s=-\sqrt{\frac{c}{8 a c+1}}\right)\right)\right) $$

We see that the restrictions split in two ones.

Thirdly, we consider the first case

Minimize[{p + r + s, (a == 0 && b > 0 && c == 1/b && p == 0 && 
r == Sqrt[b/(1 + 8 b c)] && s == Sqrt[c])}, {a, b, c, p, r, s}]

{2/Sqrt[3], {a -> 0, b -> 3, c -> 1/3, p -> 0, r -> 1/Sqrt[3], s -> 1/Sqrt[3]}} $$\left\{\frac{2}{\sqrt{3}},\left\{a\to 0,b\to 3,c\to \frac{1}{3},p\to 0,r\to \frac{1}{\sqrt{3}},s\to \frac{1}{\sqrt{3}}\right\}\right\} $$

Fourthly, we consider the second case by

NMinimize[{p + r +  s, (a > 0 && ((0 <= b < 1/a && c == (1 - a b)/(a + b) && 
    p == Sqrt[a/(1 + 8 a b)] && r == Sqrt[b/(1 + 8 b c)] && 
    s == Sqrt[c/(1 + 8 a c)]) || (b == 1/a && 
    c == (1 - a b)/(a + b) && p == Sqrt[a/(1 + 8 a b)] && 
    r == Sqrt[b/(1 + 8 b c)] && 
    s == -Sqrt[(c/(1 + 8 a c))])))}, {a, b, c, p, r, s}]

{1.19043, {a -> 0.57735, b -> 0.57735, c -> 0.57735, p -> 0.396811, r -> 0.396811, s -> 0.396811}}

In the above we use the numeric command NMinimize since Minimize is running without any response for a long time. Because $\frac{2}{\sqrt{3}} < 1.19043$ we can draw the conclusion that one of 6 optimal solutions is $a=0,, b=3, c=\frac 1 3$. I think all that can be mimicked by hand.

Addition. The doubts of @RiverLi concerning a local minimum instead of a global one produced by

NMinimize[{p + r +  s, (a > 0 && ((0 <= b < 1/a && c == (1 - a b)/(a + b) && 
p == Sqrt[a/(1 + 8 a b)] && r == Sqrt[b/(1 + 8 b c)] && 
s == Sqrt[c/(1 + 8 a c)]) || (b == 1/a && 
c == (1 - a b)/(a + b) && p == Sqrt[a/(1 + 8 a b)] && 
r == Sqrt[b/(1 + 8 b c)] && 
s == -Sqrt[(c/(1 + 8 a c))])))}, {a, b, c, p, r, s}]

came true. Indeed, the result of the command

NMinimize[{p + r + s, a > 0 && r >= 0 && p >= 0 && 
  s >= 0 && (0 <= b < 1/a && c == (1 - a b)/(a + b) && 
 p^2 == a/(1 + 8 a b) && r^2 == b/(1 + 8 b c) && 
 s^2 == c/(1 + 8 a c))}, {a, b, c, p, r, s}]

{1.1547, {a -> 0.333334, b -> 6.04174*10^-12, c -> 2.99999, p -> 0.577351, r -> 2.66924*10^-9, s -> 0.57735}}

(which leads to the exact solution $a= \frac 1 3, b=0, c=3$) shows the minimum in the second case is not greater than the one in the first case. A happy end is achieved through the Mathematica command FindInstance(see its documentation), where we exclude c and return to the target function with square roots,

FindInstance[a > 0 && 0 <= b < 1/a &&    2/Sqrt[3] > Sqrt[a/(1 + 8 a b)] + Sqrt[b/(1 + 8 b c)] + 
 Sqrt[c/(1 + 8 a c)] /. c -> (1 - a b)/(a + b), {a, b}, Reals]

{}

i.e. the value of the target function less than $\frac 2 {\sqrt{3}}$ cannot be obtained.

user64494
  • 5,954
  • Is there evidence that CylindricalDecomposition makes no mistake at all? Anyway this is not a readable proof. I accept software-assisted proofs but probably not this one... – youthdoo Jun 19 '25 at 11:21
  • @youthdoo: Thank you for your interest to the answer and valuable comment. The command of Maple 2025 SolveTools:-SemiAlgebraic([(8*b*a + 1)*p^2 = a, (8*c*b + 1)*r^2 = b, (8*c*a + 1)*s^2 = c, a*b + a*c + b*c = 1, 0 <= a, 0 <= b, 0 <= c, 0 <= p, 0 <= r, 0 <= s], [a, b, c, p, r, s]) results in – user64494 Jun 19 '25 at 11:39
  • $[\left[a {=} 0,0 {<} b,c {=} \frac{1}{b},p {=} 0,r{=} \frac{\sqrt{b}}{3},s {=} \frac{1}{\sqrt{b}}\right],\left[0{<} a,a {<} \frac{\sqrt{5}}{20},b{=} 0,c {=} \frac{1}{a},p {=} \mathrm{\sqrt{a}},r {=} 0,s {=} \frac{1}{3 \mathrm{\sqrt{a}}}\right],\dots]$. Maple considers all the possible cases whereas the one case is sufficient as MMA does. – user64494 Jun 19 '25 at 11:45
  • As I understand it, all 6 optimal solutions can be derived from the above Maple's answer. MMA leads only to one optimal solution – user64494 Jun 19 '25 at 12:07
  • 3
    @user64494 Usually we want a human verifiable, readable, step-by-step solutions (we can use Computer Algebra Systems to motivate our solutions). However, I think that your answer is to verify the inequality by computer algebra system. The step-by-step solution in Mathematica's innards can be very, very long. – River Li Jun 19 '25 at 12:47
  • @user64494 I said your answer is to verify the inequality by the Computer Algebra Systems (CAS). It told us that the inequality is true. Alternatively, we can use Resolve in Mathematica which outputs True. About Resolve, see my answer for this question in 2019: https://math.stackexchange.com/questions/2016364/inequality-a-b-c-d-subset0-1-rightarrow-sum-limits-cyca4a2b28-pr/3446031#3446031 – River Li Jun 20 '25 at 03:15
  • @RiverLi: You wrote "However, if you only use NMinimize, then you don't verify the inequality. NMinimize verify nothing". Sorry, you are not right. The result of NMinimize in my answer {1.19043, {a -> 0.57735, b -> 0.57735, c -> 0.57735, p -> 0.396811, r -> 0.396811, s -> 0.396811}} means that the minimal value of p+q+r equals 1.19043 with relative error at most $10^{-6}$ and this value can be verified by the substitution, e.g – user64494 Jun 20 '25 at 05:19
  • c - (1 - a b)/(a + b) /. {a -> 0.5773502876491612, b -> 0.5773502031566574, c -> 0.5773501981392687, p -> 0.3968112580038495, r -> 0.39681125829089187, s -> 0.39681119464976233} which results in $−1.1862379301064152 ,,10^{−7}$. Briefly, the obtained result of NMinimize means there exists an optimal solution of the optimization problem under consideration which is close to {a -> 0.57735, b -> 0.57735, c -> 0.57735, p -> 0.396811, r -> 0.396811, s -> 0.396811} with relative error at most $10^{-6}$. – user64494 Jun 20 '25 at 05:37
  • 1
    @RiverLi: Indeed, there are pathological examples (e.g. Clear[f]; f = E^Sin[50 x] + Sin[60 E^y] + Sin[70 Sin[x]] + Sin[Sin[80 y]] - Sin[10 (x + y)] + 1/4 (x^2 + y^2); Plot3D[f, {x, -1, 1}, {y, -1, 1}, PlotPoints -> 50, Mesh -> False], see the documentation, where NMinimize without any options finds only a local minimum. The problem under cosideration is not of such a sort. Its result can be verified in Maple . I'll try it later. – user64494 Jun 20 '25 at 06:33