0

I want to maximize

Abs[a1 b1] + Abs[a2 b2] + Abs[a3 b3]

subject to the joint constraints

9 (a1^2+a2^2)<=4&&18 (a1^2+a2^2)+9 (2+3 a2) a3^2<=8&&4 (b1^2+b2^2+b3^2)<=1

(The first constraint is circular and the last constraint, spherical in nature. The middle one is independent of the $b$'s.)

In lieu of an exact solution, a high-precision numerical one would be desired.

Conjecturally, the exact solution (to this quantum-information-related problem) has a denominator that is the product of powers of 2 and/or of 3.

To further expand, the constraints were obtained by requiring the joint positive-semidefiniteness of the $3 \times 3$ and $4 \times 4$ ("density") matrices

{{1/3 - a2/2, -((I a1)/2), (I a3)/2}, {(I a1)/2, 1/3 + a2/2, 0}, {-((I a3)/2), 0, 1/3}}

and

{{1/4, 0, b1/2, 0}, {0, 1/4, 1/2 (I b2 - b3), 0}, {b1/2, O1/2 (-I b2 - b3), 1/4, 0}, {0, 0, 0, 1/4}}
J. M.'s missing motivation
  • 124,525
  • 11
  • 401
  • 574
Paul B. Slater
  • 2,335
  • 10
  • 16
  • 1
    Have you stated the problem correctly? If a2==-2/3, I think a3 is unbounded. NMaximize should handle problems of this sort. – mikado Feb 01 '20 at 13:33
  • I did try NMaximize rather extensively--couldn't obtain any convincing convergence--despite various WorkingPrecision and MaxIterations choices. Not sure about your a2==-2/3 observation at this point. Oops--the first constraint is "circular" not "spherical". Will correct. – Paul B. Slater Feb 01 '20 at 14:11
  • 1
    As two others have observed, it is unbounded. Is one of the constraints slightly off? – Daniel Lichtblau Feb 01 '20 at 16:35
  • For example, should a2 be constrained to be nonnegative? – Daniel Lichtblau Feb 01 '20 at 17:47
  • 1
    Do you have a typo in your matrices, specifically, the variable O1? – mikado Feb 01 '20 at 19:06
  • 1
    I don't think your conditions ensure that your matrices are positive semi-definite. E.g. with b1==b2==b3==1/Sqrt[3] – mikado Feb 01 '20 at 19:17
  • The comments led me to post https://mathematica.stackexchange.com/questions/214002/is-this-a-correct-implementation-of-the-principal-leading-minors-test-for-positi – Paul B. Slater Feb 02 '20 at 15:46

1 Answers1

1
$Version

(* "12.0.0 for Mac OS X x86 (64-bit) (April 7, 2019)" *)

Clear["Global`*"]

Assuming that all of the variables are real

sys1 = {Abs[a1 b1] + Abs[a2 b2] + Abs[a3 b3],
    9 (a1^2 + a2^2) <= 4, 
    18 (a1^2 + a2^2) + 9 (2 + 3 a2) a3^2 <= 8, 
    4 (b1^2 + b2^2 + b3^2) <= 1} /.
   Abs[z_] :> Sqrt[z^2];

var1 = Variables[Level[sys1, {-1}]]

(* {a1, a2, a3, b1, b2, b3} *)

NMaximize[sys1, var1, WorkingPrecision -> 20]

(* {0.33333333333333343737, {a1 -> -0.63260132112986447330, 
  a2 -> 0.21038063824695148785, a3 -> 4.7334038188327067403*10^-10, 
  b1 -> -0.47445099084739835497, b2 -> 0.15778547868521361589, 
  b3 -> -3.5500528641245300550*10^-10}} *)

Let a3 == 0 and b3 == 0

sol1 = {a3 -> 0, b3 -> 0};

sys2 = sys1 /. sol1

(* {Sqrt[a1^2 b1^2] + Sqrt[a2^2 b2^2], 9 (a1^2 + a2^2) <= 4, 
 18 (a1^2 + a2^2) <= 8, 4 (b1^2 + b2^2) <= 1} *)

var2 = Variables[Level[sys2, {-1}]]

(* {a1, a2, b1, b2} *)

sol2 = Maximize[sys2, var2]

(* {1/3, {a1 -> -(5/16), a2 -> -(Sqrt[799]/48), b1 -> -(15/64), 
  b2 -> -(Sqrt[799]/64)}} *)

sol = Join[sol1, sol2[[2]]] // Sort

(* {a1 -> -(5/16), a2 -> -(Sqrt[799]/48), a3 -> 0, b1 -> -(15/64), 
 b2 -> -(Sqrt[799]/64), b3 -> 0} *)

Abs[a1 b1] + Abs[a2 b2] + Abs[a3 b3] /. sol

(* 1/3 *)

EDIT: As pointed out in a comment by user64494, this is a local maximum.

Bob Hanlon
  • 157,611
  • 7
  • 77
  • 198
  • Wow--terrific looking! Seems like the denominator-power conjecture of mine is substantiated--that is, given the 1/3 answer. – Paul B. Slater Feb 01 '20 at 15:07
  • 2
    How about 9 (a1^2 + a2^2) <= 4 && 18 (a1^2 + a2^2) + 9 (2 + 3 a2) a3^2 <= 8 && 4 (b1^2 + b2^2 + b3^2) <= 1 /. {a1 -> 0, a2 -> -2/3, a3 -> 100, b1 -> 0, b2 -> 0, b3 -> 1/2} which produces True and Abs[a1 b1] + Abs[a2 b2] + Abs[a3 b3] /. {a1 -> 0, a2 -> -2/3, a3 -> 100, b1 -> 0, b2 -> 0, b3 -> 1/2} which produces 50? – user64494 Feb 01 '20 at 15:18