6

Consider the following function : $$\tag{1} f(v) = v^{\frac{d}{2}} \int_v^{\infty} u^{\alpha \,-\, \smash{\frac{d}{2}} \,-\, 1} \; e^{-\, \alpha \, u} \; du, $$ where $d \le 6$ and $\alpha > 0$ are two positive constants (parameters). Notice the lower limit of the integral : $v$ is a variable. The plot of this function (for $0 \le v < \infty$) shows an almost bell-shaped curve, so it has a single maximal value.

Now, I would like to find the position $v = v_0(\alpha, d)$ of the maximal value of this function. I need an analytical expression for $v_0(\alpha, d)$, probably an approximation. $$\tag{2} v_0(\alpha, d) \approx \; ? $$ From the graph of $f(v)$, I know that $v_0 \propto d$ (maybe with some exponent).

Take note that the derivative of function (1), set to 0, give this relation : $$\tag{3} f(v_0) \equiv f_{\text{max}} = \frac{2}{d} \; v_0^{\alpha} \; e^{-\, \alpha \, v_0}, $$ where $v_0 \equiv v_0(\alpha, d)$ is the position of the max value of (1).

Someone knows a method to find the function (2) ?


EDIT : Function (1) describes the "deformation" of a black body luminosity caused by the expansion of space in a cosmology model. From Wien's and Planck's laws, $\alpha$ should be around 3 (depending on the presence of gaz and dust). $d$ describes the kind of fluid contained in the cosmological model. We have $d = 3$ for a dust filled universe, and $d = 4$ for a radiation universe. $d = 0$ for an empty universe with a cosmological constant. Since $\frac{d}{2} + 1$ gives 2.5 for dust and 3 for radiation, I suspect that $\alpha$ should be close to $\frac{d}{2} + 1$ (while it is an independant parameter). In this special case, it is easy to explicitely evaluate the integral in (1) and we get the special case $$\tag{4} v_0(d) = \frac{d}{d + 2}, \quad \text{if $\alpha = \frac{d}{2} + 1$}. $$

Cham
  • 322
  • If you want a Gaussian, approximate $\log f(v)$ by its order $2$ Taylor expansion around $v_0$ – reuns Nov 05 '16 at 01:30
  • According to Mathematica the value of (1) is $$f(v) = \alpha ^{-\alpha } (\alpha v)^{d/2} \Gamma \left(\alpha -\frac{d}{2},v \alpha \right)$$ and its derivative is $$f'(v) = -\frac{2 e^{\alpha (-v)} v^{\alpha }-\alpha ^{-\alpha } d (\alpha v)^{d/2} \Gamma\left(\alpha -\frac{d}{2},v \alpha \right)}{2 v}.$$ So you want to solve $$2 (\alpha v)^{\alpha }=d e^{\alpha v} (\alpha v)^{d/2} \Gamma \left(\alpha-\frac{d}{2},v \alpha \right)$$ for $v$. –  Nov 07 '16 at 01:13
  • From numerical tests it looks like the optimal $v_0$ does not exceed $1$, so you cannot have $v_0\propto d$. –  Nov 07 '16 at 01:22
  • there's an exponent to d. If d = 0, then $v_0 = 0$. If d increases, then $v_0$ increases too. There's another number in front of d, maybe something dependant on $\alpha$. – Cham Nov 07 '16 at 01:45
  • 1
    Have you investigated with $d \approx 2 (\alpha - 1)$? The behaviour seems to change around there. – Eric Towers Nov 07 '16 at 02:36
  • @EricTowers, yes, I played a bit with the special value $\alpha = d/2 + 1$. In my case, $d$ is constrained by these values : $0 \le d \le 6$. If $\alpha = d/2 + 1$, then $$v_0(d) = \frac{d}{d + 2}.$$ – Cham Nov 07 '16 at 03:04
  • @Rahul see incomplete gamma function what you wrote is a tautology – reuns Nov 07 '16 at 03:15
  • @Cham you should say why you need such an approximation, what you really want to do – reuns Nov 07 '16 at 03:17
  • @user1952009, function (1) is from cosmology. In short : it gives the total apparent luminosity (per frequency unit) of the whole sky in a "flat" universe, of all the visible galaxies. Each galaxy is emitting light according to Wien's thermal law (approximation of a black body). The total luminosity is modified by the expansion of space. The maximal value of (1) is interesting because $v_0$ corresponds to the light frequency of the maximal luminosity. Actually, $v \equiv \omega / \omega_0$ where $\omega_0$ is the maximal output of a single galaxy. – Cham Nov 07 '16 at 03:21
  • Ok but then $\alpha,d$ are fixed constants (in that can you can evaluate numerically everything) ? Or do you want the asymptotic as $\alpha,d \to\ ?$ (maybe $\alpha$ depends on the universe expansion and you want to know the limiting case ?) – reuns Nov 07 '16 at 03:23
  • $\alpha$ is a constant, yes. But it's an adjustable parameter which describes the galaxy's light output. Usually, $\alpha \approx 3$ (from Wien's and Planck's law). $d \le 6$ is describing the matter distribution in that universe. $d = 3$ for dust. $d = 4$ for radiation. – Cham Nov 07 '16 at 03:25
  • Once you know $v_0(\alpha,d)$ you can Taylor expand and approximate $v_0(\alpha',d')$ for $\alpha',d'$ close to $\alpha,d$, is it what you really want ? – reuns Nov 07 '16 at 03:27
  • I'm not very interested in numerical evaluation, since it's very easy to find $v_0$ (numerically) from the curve plot. I would like to find a general formula, even if it's approximate. – Cham Nov 07 '16 at 03:29
  • For dust and radiation, we get $d/2 + 1 = 2.5$ or $3$. And usually $\alpha \approx 3$, so this is a clue. – Cham Nov 07 '16 at 03:30
  • You won't get a general formula valid for any magnitude of $\alpha,d$, that's why constraining $(\alpha ,d) \in [2,5]\times [2,5]$ and approximating numerically is much easier. Is it what you want ? (in that case can you say it in your question ?) – reuns Nov 07 '16 at 03:31
  • I had thought $d>0$ from the problem statement. It is far more useful to know $0 \leq d \leq 6$. Any chance $\alpha$ is similarly upper bounded? – Eric Towers Nov 07 '16 at 03:33
  • $\alpha$ is an arbitrary positive number. From Wien's and Planck's law, it should be around 3, though. I guess that $\alpha$ should be close to $d/2 + 1$. – Cham Nov 07 '16 at 03:35
  • I've edited the question to add some information. – Cham Nov 07 '16 at 03:43
  • @EricTowers, I'm not sure to understand. Yes, $v_0 = 0$ and $v_0 \rightarrow \infty$ give the minimal value of (1) (i.e. 0). Between both of these, there's a maximal value at $$v_0 = \text{function of $\alpha$ and $d$} \approx \frac{d}{d + 2} ?$$ – Cham Nov 07 '16 at 03:49
  • @EricTowers, ??. Try this : derive (1), and it will give you (3), if $v \ne 0$. There's some magic occuring there ! ;-) – Cham Nov 07 '16 at 03:56
  • @EricTowers, I'm getting this :$$f'(v) = \frac{d}{2} ; v^{-1} , f(v) - v^{\alpha ,-, 1} ; e^{-, \alpha , v}.$$ Setting this to 0 for the maximal value, at $v_0 \ne 0$, gives $$\frac{d}{2} ; v_0^{- 1} ; f_{\text{max}} - v_0^{\alpha ,-, 1} ; e^{-, \alpha , v_0} = 0,$$ which gives (3). Is there a mistake here ? – Cham Nov 07 '16 at 04:06
  • @EricTowers, (3) is not about zeros of (1), it's a relation for the maximal value of $f(v)$. I don't see any difference in Rahul's derivative and mine, except that Rahul's has a global sign mistake (copied from Mathematica ?). That sign mistake is of no importance. – Cham Nov 07 '16 at 04:17
  • Working back through my scratch work, I may have found my error. However, I do get Rahul's minus sign. – Eric Towers Nov 07 '16 at 04:24
  • Oops, yes, Rahul's derivative is correct. I'm getting the same. So we're back at (3) beeing correct too. – Cham Nov 07 '16 at 04:32
  • Agreed. (and enough filler to allow submission...) – Eric Towers Nov 07 '16 at 04:33

2 Answers2

1

This is a partial solution only and needs a correction for more precision in the result.

I wrote a Mathematica code to show the curve of function (1), with sliders to interactively change the two parameters $\alpha$ and $d$. When $\alpha = \frac{d}{2} + 1$, the exact solution to the problem is $$\tag{1} v_0(d) = \frac{d}{d + 2}, \quad \text{if $\alpha = \frac{d}{2} + 1$}. $$ When $\alpha \ne \frac{d}{2} + 1$, the Mathematica graphics shows that we still have $v_0 \approx \frac{d}{d \,+\, 2}$, for all values in the ranges $0 \le \alpha \le 10$ and $0 \le d \le 6$.

When $\alpha$ is very different than $\frac{d}{2} + 1$, formula (1) misses the true value by a small amount only. So it needs a small correction like $$\tag{2} v_0(\alpha, d) \approx \frac{d}{d + 2} \; g(\alpha, d), $$ where $g(\alpha, d) \approx 1 + \delta(\alpha, d)$, or maybe something like $$\tag{3} v_0(\alpha, d) \approx \frac{d}{d + 2} + h(\alpha, d), $$ where $h(\alpha, d)$ is a small correction.

Any idea about how to find the correction $\delta(\alpha, d)$ or $h(\alpha, d)$?

Here's the Mathematica code for you to see how good is the partial solution (1) above :

Clear["Global`*"]

function[v_, a_, d_] := v^(d/2) NIntegrate[u^(a - d/2 - 1) Exp[-a u], {u, v, Infinity}]

Xpic[a_, d_] := d/(d + 2)
Ypic[a_, d_] := function[d/(d + 2), a, d]

Pic[a_, d_] := Graphics@{Black, PointSize -> 0.01, Point[{Xpic[a, d], Ypic[a, d]}]}

Manipulate[
    Show[{Plot[{function[v, a, d]}, {v, 0, 4}, PlotStyle -> Directive[Red, Thick]], Pic[a, d]},
    Frame -> True,
    AspectRatio -> 1,
    PlotRange -> Automatic,
    GridLines ->  Automatic,
    GridLinesStyle -> Directive[Dashed, GrayLevel[0.75]],
    ImageSize -> 500
    ],
    {{a, 3, Style["\[Alpha]", 10]}, 0.5, 10, 0.01, ImageSize -> Large},
    {{d, 3, Style["d", 10]}, 0, 6, 0.01, ImageSize -> Large},
    ControlPlacement -> Bottom,
    FrameMargins -> None
]
Cham
  • 322
  • 1
    Here's some code to directly plot $v_0(\alpha,d)$: f[v_] := Evaluate@Simplify[v^(d/2) Integrate[u^(a - d/2 - 1) Exp[-a u], {u, v, Infinity}], Assumptions -> v > 0 && a > 0 && d > 0]; df[v_] := Evaluate@Simplify[f'[v], Assumptions -> v > 0 && a > 0 && d > 0]; ContourPlot3D[df[v] == 0, {a, 0, 10}, {d, 0, 10}, {v, 0, 1}, RegionFunction -> Function[{a, d}, a > d/2]]. You can also test what the corrections $g(\alpha,d)$ and $h(\alpha,d)$ must look like, by replacing the ContourPlot call with e.g. ContourPlot3D[df[d/(d + 2) g] == 0, {a, 0, 10}, {d, 0, 10}, {g, 0, 2}, ... –  Nov 07 '16 at 13:41
  • I'll try your code later today, when I'll be back home. Thanks. – Cham Nov 07 '16 at 15:47
  • @Rahul, your code isn't obvious. Could you add labels on the three axes to show what this is all about ? – Cham Nov 07 '16 at 22:22
  • Using my code above, I noticed that the approxmation (1) gets better and better when we increase d, wathever what value is selected for $\alpha$. The approximation is not excellent when $d \le 2.5$. – Cham Nov 07 '16 at 22:35
  • The axes are $\alpha$, $d$, and $v$ respectively. It plots the solutions to $\mathrm df/\mathrm dv = 0$ as an implicit function in $(\alpha,d,v)$ space. –  Nov 07 '16 at 22:44
  • Can you use it to suggest a modification of the approximation (1) ? – Cham Nov 07 '16 at 22:47
  • Not yet, I just thought it would help you in your investigations. Since you wrote a manipulator to explore $v_0$ for different values of $\alpha$ and $d$, I thought you might find it useful to have a single graph that plots $v_0$ for a whole range of $\alpha$ and $d$. –  Nov 08 '16 at 00:01
0

Taking logarithmic derivatives, the maximum is attained when $$ \frac{d}{2v} - \frac{\mathrm{e}^{-v \alpha} \alpha (v \alpha)^{\alpha - (d/2 + 1)}}{\Gamma(\alpha - d/2, v\alpha)} = 0 $$

The form of the first summand suggests finding the (Laurent) series expansion of the second term around $v = \infty$. (I.e., find a Taylor series in powers of $1/v$.) This series is $$ \alpha - \frac{\alpha - (d/2-1)}{v} + \frac{2\alpha -(d+2)}{2 \alpha v^2} + ... $$

Solving $\dfrac{d}{2v} - \left( \alpha - \frac{\alpha - (d/2-1)}{v} \right) = 0$ for $v$, we get $\frac{\alpha - 1}{\alpha}$. If we substitute $\frac{d}{2}+1$ for $\alpha$ in this, we get $\frac{d}{d+2}$, as expected from your prior work.

Solving $\dfrac{d}{2v} - \left( \alpha - \frac{\alpha - (d/2-1)}{v} + \frac{2\alpha -(d+2)}{2 \alpha v^2} \right) = 0$ for $v$, we get two roots: $$ v_\pm = \frac{\alpha - 1 \pm \sqrt{2d + (\alpha - 5)(\alpha-1)}}{2\alpha} \text{.} $$ If we substitute $\frac{d}{2}+1$ for $\alpha$ in these, we get $0$ for $v_-$ (corresponding to the zero at $v=0$ for certain parameter values) and $\frac{d}{d+2}$ for $v_+$, as expected. This suggests $v_+$ is an improved estimate for your $v_0(\alpha,d)$.

(Note: I haven't graphed, bounded, or sanity checked the above, and it looks like I won't have time for several days. I wouldn't be very surprised to find I have a sign error (or worse) in transcription. I'm considering making this Community Wiki so others can improve it while I'm otherwise engaged. If someone finds a meaningful improvement, correction, or unrecoverable defect, say so in comments and I'll flip the switch (or withdraw the answer).)

Eric Towers
  • 70,953
  • I just tried your $v_{+}$ with my code of Mathematica, and it doesn't work well. For some values of d and $\alpha$, $v_{+}$ may turn into a complex number, which doesn't make sense. Why a Taylor series around $v = \infty$ ? I suspect it should be around $v = 1$ instead. And you should identify the series above : $?? \approx \alpha - \ldots$ – Cham Nov 16 '16 at 14:45
  • Why expand around $v = \infty$? Because the first term, $\dfrac{d}{2v}$ is a (very short) Laurent series. – Eric Towers Nov 16 '16 at 23:59