1

Can anyone see a trick to get large-$n$ formula for the maximum for the following for integer $n>0$ and $t\in \left( 0, \frac{\pi}{n+1}\right)$?

$$\frac{\sin (t) \sin ((n+1) t) \csc (n t)}{\pi }$$

Ideally I could use asymptotic + maximize, but that returns unevaluated

Asymptotic[
 Maximize[{(Csc[n t] Sin[t] Sin[(1 + n) t])/\[Pi], 
   0 < t < Pi/(n + 1)}, t], n -> \[Infinity]]

Plotting on log-log plot it appears to approach power law

getVal[n_] := 
  First@Maximize[{(Csc[n t] Sin[t] Sin[(1 + n) t])/\[Pi], 
      0 < t < Pi/(n + 1)}, t];
ListPlot[Table[getVal[n], {n, 1, 6}], 
 ScalingFunctions -> {"Log", "Log"}]

enter image description here

Background: this is weighted spectral density of a product of $n$ independent Gaussian matrices, ie getVal[1] is equivalent to Maximize[x PDF[MarchenkoPasturDistribution[1], x], x]

user64494
  • 26,149
  • 4
  • 27
  • 56
Yaroslav Bulatov
  • 7,793
  • 1
  • 19
  • 44
  • 1
    Honestly this reads like a math question more than a Mathematica one. – MarcoB Jul 29 '23 at 23:58
  • 1
    This terrifies me. f=Map[First,Table[Maximize[{Sin[t]Sin[(n+1)t]Csc[n t]/Pi,0<t<Pi/(n+1)} ,t],{n,1,60}]//N]; g=Transpose[{Range[60],f}]; h=Fit[g,{1,1/x^.75},x]; Show[ListPlot[g,PlotRange->All],Plot[h,{x,1,60},PlotRange->All]] Use at your own risk ;} – Bill Jul 30 '23 at 00:14
  • @Bill Why the terror? – JimB Jul 30 '23 at 01:45
  • @JimB "large-n" when we have no idea how big his n is going to be. Didn't spend the time and effort to see if it was possible to derive an exact analytic solution. Tried 10 times as many data points as he did and fiddled with the exponent to eyeball what looks like an ok fit for those points. Turning him loose with this and have no idea what he is about to do with this. I try to be MUCH more careful than that, and even then ask people to carefully check every detail to see if they can find and fix my mistakes. That is all. – Bill Jul 30 '23 at 03:31
  • @Bill Got it. Thanks. – JimB Jul 30 '23 at 04:35
  • @JimB I looked at the residuals out to n==240. They are very small. That and the graph was what fooled me into believing I had chosen a good enough and perhaps even the correct model. But they are not normally distributed about zero. So I did not choose the correct model. Fortunately there is an answer and hopefully it is correct. – Bill Jul 30 '23 at 22:42

1 Answers1

1

A typical case is as follows.

n = 26; Plot[(Csc[n t] Sin[t] Sin[(1 + n) t])/Pi, {t, 0, Pi/(n + 1)}]

enter image description here

We see the maximum of the differentiable in $t$ function under consideration is reached near $t=\frac \pi {n+1}$. In view of it we substitute t -> Pi/(n + 1) - a in its derivative and expand the derivative by powers of a near zero.

ClearAll[n];Series[D[(Csc[n t] Sin[t] Sin[(1 + n) t])/Pi, t] /. 
t -> Pi/(n + 1) - a, {a, 0, 1}] // Normal

-(((1 + n) Csc[(n \[Pi])/(1 + n)] Sin[\[Pi]/( 1 + n)])/\[Pi]) - (1/\[Pi]) 2 a (-Cos[\[Pi]/(1 + n)] Csc[(n \[Pi])/(1 + n)] - n Cos[\[Pi]/(1 + n)] Csc[(n \[Pi])/(1 + n)] + n Cot[(n \[Pi])/(1 + n)] Csc[(n \[Pi])/(1 + n)] Sin[\[Pi]/( 1 + n)] + n^2 Cot[(n \[Pi])/(1 + n)] Csc[(n \[Pi])/(1 + n)] Sin[\[Pi]/( 1 + n)])

({a, 0, 2} does not make things better.) Then we equate the above to zero and solve in a.

sol=Solve[% == 0, a]

{{a -> Sin[\[Pi]/(1 + n)]/( 2 (Cos[\[Pi]/(1 + n)] - n Cot[(n \[Pi])/(1 + n)] Sin[\[Pi]/(1 + n)]))}}

We are interested in the maximum value so we substitute that value of a in the function.

((Csc[n t] Sin[t] Sin[(1 + n) t])/Pi /. t -> Pi/(n + 1) - a) /. sol

{(1/\[Pi]) Csc[n (\[Pi]/(1 + n) - Sin[\[Pi]/(1 + n)]/( 2 (Cos[\[Pi]/(1 + n)] - n Cot[(n \[Pi])/(1 + n)] Sin[\[Pi]/(1 + n)])))] Sin[\[Pi]/( 1 + n) - Sin[\[Pi]/(1 + n)]/( 2 (Cos[\[Pi]/(1 + n)] - n Cot[(n \[Pi])/(1 + n)] Sin[\[Pi]/(1 + n)]))] Sin[(1 + n) (\[Pi]/(1 + n) - Sin[\[Pi]/(1 + n)]/( 2 (Cos[\[Pi]/(1 + n)] - n Cot[(n \[Pi])/(1 + n)] Sin[\[Pi]/(1 + n)])))]}

At last, in order to find the asymptotic,

Series[%, {n, Infinity, 2}]

{1/(3 n)-7/(18 n^2)+O[1/n]^3}

Calculations confirm the exponent -1.

getVal[n_] := First@NMaximize[{(Csc[n t] Sin[t] Sin[(1 + n) t])/\[Pi],
 0 < t < Pi/(n + 1)}, t];Table[Log[getVal[n]]/Log[n], {n, 200, 260}]

{-1.02648, -1.02639, ... , -1.02222, -1.02217}

Addition. Here is my successful attempt. Let us write down $\sin((n+1)t)=\sin(nt)\cos(t)+\cos(nt)\sin(t)$. Because $t\in \left[0,\frac \pi {n+1}\right]$ and $n$ tends to $\infty$, I replace $\sin(t)$ by $t$ and $\cos(t)$ by $1- \frac {t^2} 2$. It's summer now and I'm not in the mood to give error estimates. Let us look at a typical case

n = 26; Plot[{(Sin[n*t]*(1 - t^2/2) + Cos[n*t]*t)*t*Csc[n*t]/Pi, 
(Sin[t]* Csc[n t] Sin[(1 + n) t])/\[Pi]}, {t, 0, Pi/n}, PlotStyle -> {Red, {Blue, Dashed}}]

enter image description here

It is convenient to make a replacement

ClearAll[n]; (Sin[n*t]*(1 - t^2/2) + Cos[n*t]*t)*t*Csc[n*t]/Pi /. t -> x/n

(x Csc[x] ((x Cos[x])/n + (1 - x^2/(2 n^2)) Sin[x]))/(n \[Pi])

Now we expand it around Pi

b = Series[(x Csc[x] ((x Cos[x])/n + (1 - x^2/(2 n^2)) Sin[x]))/(n*Pi), {x, Pi, 4}] // Normal

(4 n + 2 n^2 - \[Pi]^2)/(2 n^3) + \[Pi]/( n^2 (-\[Pi] + x)) + ((6 n + 6 n^2 - 9 \[Pi]^2 - 2 n \[Pi]^2) (-\[Pi] + x))/( 6 n^3 \[Pi]) + ((-9 - 4 n) (-\[Pi] + x)^2)/( 6 n^3) + ((-45 - 30 n - 2 n \[Pi]^2) (-\[Pi] + x)^3)/( 90 n^3 \[Pi]) - (2 (-\[Pi] + x)^4)/(45 n^2)

I repeat it's summer now and I'm not in the mood to give error estimates.

Maximize[{b, x >= Pi/2 && x <= Pi && n >= 35}, x] // First

produces a huge result in terms of a Root of a polynomial of 5-th degree. Now

Series[%, {n, Infinity, 1}]

1/n+O[1/n]^(3/2)

Calculations confirm it by

getVal[n_] := First@NMaximize[{(Csc[n t] Sin[t] Sin[(1 + n) t])/\[Pi], 
0 < t < Pi/(n + 1)}, t]; Table[getVal[n]*n, {n, 3000000, 3000020}]

{0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846, 0.998846}

Moreover, the result of

getVal[n_] := First@NMaximize[{(Csc[n t] Sin[t] Sin[(1 + n) t])/\[Pi],  0 < t < Pi/(n + 1)}, t]; 
Table[(getVal[n] - 1/n)*n^(3/2), {n, 3000000, 5000000, 500000}]

{-1.99884, -1.99893, -1.99902, -1.99906, -1.99911}

confirms $O(n^{-3/2}),\,n\to \infty.$

user64494
  • 26,149
  • 4
  • 27
  • 56