Good afternoon. I have the following integral that I need help integrating;

$$ \mathrm{F}\left(x\right) = \int_{0}^{\infty}\mathrm{e}^{s\left(j - 1/x\right)}\, \left[\mathrm{T}_{N}\left(s\right)\right]^{k - j}\,\mathrm{d}s $$

where $\mathrm{T}_{N}\left(s\right)$ is the truncated exponential function $$ \mathrm{T}_{N}\left(s\right) = \sum_{n = 0}^{N}\frac{s^{n}}{n!} $$

and $j,k$ are whole numbers.

I figured that tackling this using integration by parts is the best choice, but naturally this could take a while given that my choice of $u_{1}=\mathrm{T}_{N}^{k - j}\left(s\right)$ yields a $\mathrm{d}u_{1}$ of $$ \mathrm{d}u_{1} = \left(k - j\right)\mathrm{T}_{N}^{k-j-1}\left(s\right) \mathrm{T}_{N - 1}\left(s\right)\,\mathrm{d}s $$ since $\mathrm{d}\mathrm{T}_{N}\left(s\right)/\mathrm{d}s = \mathrm{T}_{N - 1}\left(s\right)$. This does not seem to be simplifying the problem though as the next iteration would require me to then choose my $u_{2}$ to be $$ u_{2} = \frac{1}{k - j}\,\frac{\mathrm{d}u_{1}}{\mathrm{d}s} = \mathrm{T}_{N}^{k - j - 1}\left(s\right) \mathrm{T}_{N - 1}\left(s\right) $$

Are there any other approaches to this without this drawn out IBP method or without considering expanding the truncated exponential ?.

  • $\begingroup$ There was an answer here but it was deleted due to an objection I had... but I had an idea and if that person could put it back or "undelete it" it would help... I think I know how to resolve it... $\endgroup$ May 29 '19 at 14:55
  • 1
    $\begingroup$ meta.stackexchange.com/questions/126348/… $\endgroup$ May 31 '19 at 1:56
  • 1
    $\begingroup$ here it says that a moderator or 10k+ reputation account can view deleted answers. Maybe you can get one of them to help you put it up again? $\endgroup$ May 31 '19 at 1:57

I have only a partial answer: it pretty much just consists of repeated integration by parts, so maybe this isn't what you're looking for.

Write $A(n_1, \dots, n_r) = \int_0^\infty e^{-as} \prod_i T_{n_i}(s) \,ds$. Then integration by parts gives \begin{align*} A(n_1, \dots, n_r) &= -\frac{1}{a}\bigg[e^{-as}\prod_i T_{n_i}(s)\bigg]\bigg|_0^\infty + \frac{1}{a}\sum_{j : n_j > 0} \int_0^\infty e^{-as} T_{n_j - 1}(s)\prod_{i \neq j} T_{n_i}(s) \,ds \\ &= \frac{1}{a} + \frac{1}{a} \sum_{j : n_j > 0} A(n_1, \dots, n_j - 1, \dots, n_r) \end{align*} when all $n_i \geq 0$.

To make things a little easier, let's take $T_n(s) = 0$ when $n < 0$ (this is consistent with the identity $T_n'(s) = T_{n-1}(s)$), so we'll have $A(n_1, \dots, n_r) = 0$ when some $n_i < 0$, and so we can write $$A(n_1, \dots, n_r) = \frac{1}{a}\bigg(\mathbb{1}_{(n_i)} + \sum_i A(n_1, \dots, n_i - 1, \dots, n_r)\bigg)$$ for any integers $n_1, \dots, n_r$, where $\mathbb{1}_{(n_i)} = 1$ if all $n_i \geq 0$ and is $0$ otherwise.

Letting $f(x_1, \dots, x_r) = \sum A(n_1, \dots, n_r) x_1^{n_1} \cdots x_r^{n_r}$ be the generating function of $A$, this identity gives $$af = \frac{1}{(1-x_1) \cdots (1-x_r)} + (x_1 + \cdots + x_r)f$$ hence \begin{align*} f &= \frac{1}{(1 - x_1) \cdots (1 - x_r)(a - x_1 - \cdots - x_r)} \\ &= \frac{1}{a} \sum_{l_1, \dots, l_r, m_1, \dots, m_r} a^{-m_1 - \cdots - m_r}\binom{m_1 + \cdots + m_r}{m_1, \dots, m_r} x_1^{l_1 + m_1} \cdots x_r^{l_r + m_r} \\ &= \frac{1}{a} \sum_{n_1, \dots, n_r} \sum_{0 \leq m_i \leq n_i} a^{-m_1 - \cdots - m_r}\binom{m_1 + \cdots + m_r}{m_1, \dots, m_r} x_1^{n_1} \cdots x_r^{n_r} \\ \end{align*} and finally, $$A(n_1, \dots, n_r) = \frac{1}{a} \sum_{0 \leq m_i \leq n_i} a^{-m_1 - \cdots - m_r}\binom{m_1 + \cdots + m_r}{m_1, \dots, m_r} $$ (this could alternatively be found by repeatedly expanding the recurrence).

The case relevant to the question is the one where $n_1 = \cdots = n_r = N$, $r = k-j$, $a = 1/x - j$, but I'm not sure how to simplify this sum. I've been thinking about it in terms of lattice paths in $r$ dimensions: in particular the coefficient of $a^{-t}$ is the number of lattice paths of length $t$ where we have at most $n_i$ steps in the $i$-th direction for each $i$, but that doesn't seem to help simplify any further.

As noted by DinosaurEgg, we're only interested in $$A(N, \dots, N) = \frac{1}{a} \sum_{0 \leq m_i \leq N} a^{-m_1 - \cdots - m_r}\binom{m_1 + \cdots + m_r}{m_1, \dots, m_r}$$ but this sum still seems difficult to simplify to me, because of the $m_i \leq N$ constraints.

  • $\begingroup$ I love what you did here, but I am giving the bounty to @ben, only because his work is closer to all that I have done with my calculations over the last couple of weeks. I definitely liked your approach though and will be referring to it as I continue. Thank you for your effort. $\endgroup$ May 31 '19 at 15:29

So long as $j < 1/x$ it is possible to write the function as a finite sum of terms (i.e., in closed form) by writing the power of the truncated exponential as a polynomial, and then turning the integral into a polynomial sum of gamma functions. If we apply the multinomial theorem we can write the power of the truncated exponential as a polynomial of degree $N(k-j)$ as follow:

$$\begin{equation} \begin{aligned} \left[ \sum_{n=0}^N \frac{s^n}{n!} \right]^{k-j} &= \sum_{r_0 + \cdots + r_N = k-j} {k-j \choose \mathbf{r}} \prod_{n=0}^N \Big( \frac{s^n}{n!} \Big)^{r_n} \\[6pt] &= \sum_{r_0 + \cdots + r_N = k-j} \frac{(k-j)!}{\prod_{n=0}^N (r_n!) (n!)^{r_n}} \prod_{n=0}^N s^{n \cdot r_n} \\[6pt] &= \sum_{r_0 + \cdots + r_N = k-j} \frac{(k-j)!}{\prod_{n=0}^N (r_n!) (n!)^{r_n}} s^{\sum_{n=0}^N n \cdot r_n} \\[6pt] &= \sum_{i=0}^{N(k-j)} a_i s^i, \\[6pt] \end{aligned} \end{equation}$$

where the coefficients $a_0,a_1,...,a_{N(k-j)}$ depend on $N$ and $k-j$, and are obtained from this multinomial summation. We then have:

$$\begin{equation} \begin{aligned} F(x) &= \int \limits_0^\infty e^{s (j-\frac{1}{x})} \left[ \sum_{n=0}^N \frac{s^n}{n!} \right]^{k-j} ds \\[6pt] &= \int \limits_0^\infty e^{s (j-\frac{1}{x})} \sum_{i=0}^{N(k-j)} a_i s^i \ ds \\[6pt] &= \sum_{i=0}^{N(k-j)} a_i \int \limits_0^\infty e^{s (j-\frac{1}{x})} s^i \ ds \\[6pt] \end{aligned} \end{equation}$$

Now, assume that $j < 1/x$ so that $(j-\tfrac{1}{x}) <0$. In this case we can use the change-of-variable $m = s |j-\frac{1}{x}|$ to get $dm = -j ds$ and we have:

$$\begin{equation} \begin{aligned} F(x) &= \sum_{i=0}^{N(k-j)} a_i \int \limits_0^\infty e^{-s |j-\frac{1}{x}|} s^i \ ds \\[6pt] &= \sum_{i=0}^{N(k-j)} a_i \cdot \frac{1}{(-j) |j - \frac{1}{x}|^i} \int \limits_0^\infty e^{-s |j-\frac{1}{x}|} (s |j-\tfrac{1}{x}|)^i \ (-j) \ ds \\[6pt] &= \sum_{i=0}^{N(k-j)} a_i \cdot \frac{1}{(-j) |j - \frac{1}{x}|^i} \int \limits_0^\infty e^{-m} m^i \ dm \\[6pt] &= \sum_{i=0}^{N(k-j)} a_i \cdot \frac{i!}{(-j) |j - \frac{1}{x}|^i}. \\[6pt] \end{aligned} \end{equation}$$

This form of the function is a finite sum, where the only serious difficulty (for large $N$ or $k-j$) is to calculate the polynomial coefficients $a_0,a_1,...,a_{N(k-j)}$. Calculation of these coefficients may be cumbersome for large degree, but once they are calculated the function is then amenable to calculation.

  • $\begingroup$ The answer I mentioned that was deleted had a similar expansion for the powered truncated exponential... I felt that it could be written as a sum of powers of $s$ better than that answerer did...I like where this is going... I suppose on first glance, those polynomial coefficients are hard to find a nice form for... Perhaps looking at small values of $N$ first and looking for pattern using computational methods could help here? $\endgroup$ May 31 '19 at 11:36
  • 1
    $\begingroup$ Yes; I'm not sure of any simple recursion for the coefficients, but maybe some pattern could be found to assist this. $\endgroup$
    – Ben
    May 31 '19 at 18:01

As mentioned in another answer deriving an exact expression is definitely hard, but we can bound this sum pretty easily from above:

$$A(N,..., N)<\frac{1}{a}\sum_{n=0}^{rN}\sum_{\sum m_i=n}a^{-n}\frac{n!}{m_1!...m_r!}=\frac{1}{a}\sum_{n=0}^{rN}\Big(\frac{r}{a}\Big)^n=\frac{\Big(\frac{r}{a}\Big)^{rN+1}-1}{r-a}=\frac{\Big(\frac{k-j}{1/x-j}\Big)^{rN+1}-1}{k-1/x}$$

since as user125932 mentions, the $m_i$'s are ranging up to $N$, and not $rN$ like the multinomial distribution would demand.

Also it is obvious that due to the fact that if we only sum up to $n\leq N$ we get an exact multinomial estimate and since all the summands are positive:

$$A(N,..., N)>\frac{1}{a}\sum_{n=0}^{N}\sum_{\sum m_i=n}a^{-n}\frac{n!}{m_1!...m_r!}=\frac{1}{a}\sum_{n=0}^{N}\Big(\frac{r}{a}\Big)^n=\frac{\Big(\frac{k-j}{1/x-j}\Big)^{N+1}-1}{k-1/x}$$

which by sandwiching reduces to the correct expression of the Laplace transform of the exponential in the limit $N\rightarrow \infty$:

$$\lim_{N\to\infty}A(N,..., N)=\frac{x}{1-kx}$$

To appreciate the complexity of this sum, let us attempt an evaluation for $r=2$. The sum is close to a binomial expansion but not exactly. We can break it up as follows (here $\Delta=m+n$): $$S_2=\sum_{m,n=0}^{N}{m+n\choose m}a^{-m-n}=\Big(\sum_{\Delta=0}^{N}\sum_{m=0}^{\Delta}+\sum_{\Delta=N+1}^{2N}\sum_{m=0}^{\Delta}-\sum_{\Delta=N+1}^{2N}\sum_{m=N+1}^{\Delta}\Big){\Delta\choose m}a^{-\Delta}$$

The first two terms in the sum can be summed from the binomial theorem and the third term is a remainder.


By changing the order of summation we find for $R_2$


(To be continued)

  • 2
    $\begingroup$ This isn't quite the same though, right? My sum also constrains $m_i \leq N$, which is missing here -- I think this is what makes the sum difficult to work with. $\endgroup$
    – user125932
    May 31 '19 at 3:22
  • 1
    $\begingroup$ I agree with what you said. The approximation is good however for $N\to\infty$, so I'll try to do an analysis that explains why. Thanks for pointing it out! $\endgroup$ May 31 '19 at 4:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.