Jtdylktuy

Let $X_1, X_2, dots, X_k$ each be random digits. That is, they are independent random variables each uniformly distributed over the finite set ${0, 1, 2, 3, 4, 5, 6, 7, 8, 9}$. Let $S = X_1 + X_2 + dots + X_k$. Given some large integer $n$, what is the probability that $S = n$?

When $n$ or $k$ is small, the exact number can be computed as

$$[z^n]left(frac{1-z^{10}}{1-z}right)^k = frac{1}{10^k} sum_{10r+s=n} (-1)^{r+s} binom{k}{r} binom{-k}{s} = frac{1}{10^k} sum_{r} (-1)^{r} binom{k}{r} binom{k+n-10r-1}{n-10r}$$

(see the long list of questions below to see derivations) but what I need is an asymptotic expression useful for large $n$ and $k$, and I don't know how to either derive one from this formula, or arrive at one independently.

(For now, I'm not worrying about the complication of insisting that $X_1$ be nonzero, though feel free to consider it if it actually helps.)

What I have tried, part 1: As the $X_i$s are IID random variables (with mean $mu = 4.5$ and variance $sigma^2 = 8.25$), the central limit theorem applies, so we expect $Pr(S = n)$ to be highest for $n$ around $4.5k$, and the probability distribution of $S$ to be bell-curve-shaped around that value (and most of the probability will be distributed for $n$ about $O(sqrt{k})$ from that value).

Trying to make this more precise, the central limit theorem gives us
$$sqrt{k}(S/k - 4.5) xrightarrow{d} N(0,8.25) quad text{i.e.} quad lim _{kto infty}Pr left[sqrt{k}(S_{k}/k- 4.5)le zright]= Phileft(frac {z}{sqrt{8.25}}right)$$
where $Phi(x) = frac12 left[1+operatorname{erf} left(frac{x}{sqrt {2}}right)right]$ is the CDF of the standard normal distribution $N(0,1)$ (and erf is a special function) and therefore
$$Pr(S le x) = Pr(S/sqrt{k} - 4.5sqrt{k} le x/sqrt{k} - 4.5sqrt{k}) to Phileft(frac{x - 4.5k}{sqrt{8.25k}}right)$$
and so, applying a continuity correction,
$$begin{align}Pr(S = n) &approx Pr(n - 0.5 < S le n + 0.5) \
&to Phileft(frac{n + 0.5 - 4.5k}{sqrt{8.25k}}right) - Phileft(frac{n - 0.5 - 4.5k}{sqrt{8.25k}}right) \
&= frac12operatorname{erf}left(frac{2n+1-9k}{sqrt{66k}}right) - frac12operatorname{erf}left(frac{2n-1-9k}{sqrt{66k}}right) \
&overset{?}{approx} frac{2}{sqrt{pi}}frac{1}{sqrt{66k}}expleft(-left(frac{2n-9k}{sqrt{66k}}right)^2right)
end{align}$$
but I neither know whether this is correct and rigorous, nor what to do with this next.

What I have tried, part 2: There are many questions on this site about calculating this number exactly:

• Probability of random integer's digits summing to 12




• how many integers between one and 100000 have the sum equal to fifteen?




• How many numbers between 100 and 900 have sum of their digits equal to 15?




• Stars and bars to find "how many $x$ digit numbers are there with sum of digits $y$"?




• Find numbers whose sum of digits equals a value




• Counting the numbers with certain sum of digits.




• How many natural numbers are there less than $90000$ that have the sum of digits equal to $8$?




• How many integers between [3,000, 8,000] have digit sum 20?




• Counting $4$-digits numbers whose digits sum is $9$




• How many numbers between $1$ and $9999$ have sum of their digits equal to $8$? $16$?




• How many numbers between $0$ and $999,999$ are there whose digits sum to $r$




• How many positive integers less than 1,000,000 have the sum of their digits equal to 19?




• How many positive integers < 10^6 have sum of digits equal to 19




• For how many integers from 1 through 99,999 is the sum of their digits equal to 10?




• If I have an integer, how many numbers are there whose digits sum to the integer?




• Determine the number of positive integer x where x≤9,999,999 and the sum of the digits in x equals 31.




• Find number of positive integers less than $10^8$ with digit sum of $24$




• Find the number of positive integers whose digits add up to 42




• Enumerating number of solutions to an equation

The best that can be got from reading all of them is the exact formula mentioned at the top of the question (done with generating functions or inclusion-exclusion), not an asymptotic one. In particular I'd like to be able to get something which can be summed over $k$, to answer the following question:

Question 2: Let $X_{i,j}$ each be IID random digits as before, for $i = 1, 2, dots$ and $j = 1, 2, dots, i$. Let $S_k = X_{k,1} + X_{k,2} + dots + X_{k,k}$ be the sum of $k$ random digits. So we have an infinite sequence of random variables (sums) $S_1, S_2, S_3, dots$ each obtained by adding the digits of a random number of a different length. Given an integer $n$, what is the probability that some element of this sequence is exactly equal to $n$?

In other words, for each $k$ there is a probability distribution over $n$, and we want to know the total probability that falls on a single integer $n$. (Basically for each $n$ there will be some significant probability for $k$ around $n/4.5$ and the probability will fall off significantly for $k$ further away from this.)

(Again, feel free to add or remove the condition that $X_{k,1}$ is actually distributed over the set ${1, 2, dots, 9}$, i.e. is nonzero.)

What I have tried, part 3: I tried to read Distribution of the sum-of-digits function of random integers: A survey (which I found by searching for some relevant terms), and many of the papers it references. But I got pretty lost trying to figure out what is true for base-$2$ versus base-$10$, and things like that. Perhaps the answer to my question is buried somewhere in there, but I'm not sure.

Here we show the main term $frac{2}{sqrt{pi}}frac{1}{sqrt{66k}}$ in OPs asymptotic expansion is correct using the saddle-point method.

The coefficients in the expansion of the polynomials
begin{align*}
left(frac{1-z^{10}}{1-z}right)^{k}=left(1+z+z^2+z^3+z^4+z^5+z^6+z^7+z^8+z^9right)^{k}tag{1}
end{align*}
form a unimodal sequence. Taking even values $2k$ the maximum is the coefficient of the middle term
begin{align*}
[z^{9k}]left(frac{1-z^{10}}{1-z}right)^{2k}qquad kgeq 0tag{2}
end{align*}

The coefficients in (2) form a sequence starting with $(1,10,670, 55,252,4,816,030,ldots)$. These numbers are called Lucky ticket numbers and are stored in OEIS. In fact they are the even central decanomial coefficients which are the largest coefficients in the expansion of (1).

They also cite an asymptotic formula for the largest coefficient of $left(1+z+cdots+z^qright)^k$ which gives in the case (2)
begin{align*}
color{blue}{[z^{9k}]left(frac{1-z^{10}}{1-z}right)^{2k}sim 10^{2k}frac{1}{sqrt{33pi k}}}tag{3}
end{align*}
corresponding to OPs main term in his expansion ($k$ even).

We can prove the asymptotic expansion (3) using the saddle-point method. In order to do so we closely follow section VIII.8 Large Powers in Analytic Combinatorics by P. Flajolet and R. Sedgewick. We also give a little bit of surrounding information to ease readability.

VIII.8 Large powers:

• 
  (p. 585): The extraction of coefficients in powers of a fixed function and more generally in functions of the form $A(z)B(z)^k$ constitutes a prototypical and easy application of the saddle-point method. We will accordingly be concerned here with the problem of estimating
  begin{align*}
  [z^K]A(z)cdot B(z)^k=frac{1}{2pi i}oint A(z)B(z)^kfrac{dz}{z^{K+1}}
  end{align*}
  as both $k$ and $K$ get large.

In our situation (3) we have $A(z)=1$ and $B(z)=left(1+z+z^2+cdots+z^9right)^2$ with $K=9k$. What follows are conditions which need to be fulfilled by $A(z)$ and $B(z)$ in order to apply the saddle-point method.

VIII. 8.1. Large powers: saddle-point bounds: We consider throughout this section two fixed functions, $A(z)$ and $B(z)$ satisfying the following conditions.

• L1: The functions $A(z)=sum_{jgeq 0}a_jz^j$ and $B(z)=sum_{jgeq 0}a_jz^j$ are analytic at $0$ and have non-negative coefficients; furthermore it is assumed (without loss of generality) that $B(0) ne 0$.




• L2: The function $B(z)$ is aperiodic in the sense that $gcd{j|b_j>0}=1$. (Thus $B(z)$ is not a function of the form $beta(z^p)$ for some integer $pgeq 2$ and some $beta$ analytic at $0$.)




• L3: Let $Rleq infty$ be the radius of convergence of $B(z)$; the radius of convergence of $A(z)$ is at least as large as $R$.

We observe conditions L1 to L3 are fulfilled for $A(z)=1$ and $B(z)=left(1+z+z^2+cdots+z^9right)^2$ with radius of convergence $R=infty$. In the following we need the quantity $T$ called spread which is defined as

begin{align*}
color{blue}{T}&:=lim_{zto R^{-}}frac{zB^{prime}(z)}{B(z)}\
&=lim_{zto infty}frac{2zleft(1+z+cdots+z^9right)left(1+2z+cdots+9z^8right)}{left(1+z+cdots+z^9right)^2}\
&=lim_{ztoinfty}frac{2zleft(1+2z+cdots+9z^8right)}{1+z+cdots+z^9}\
&,,color{blue}{=18}
end{align*}

The purpose is to analyse the coefficients
begin{align*}
[z^K]A(z)cdot B(z)^k
end{align*}
when $K$ and $k$ are linearly related. In order to do so the condition $K<Tk$ will be imposed which is inherent in the nature of the problem. Note that in our case we have $K=9k$ and with $T=18$ the condition $K<Tk$ evaluates to $9k<18k$ which is valid.

We also need a quantity $zeta$ which is introduced in Proposition VIII.7 and since this is a useful and easily derived upper bound for the coefficients we note

Proposition VIII.7 (Saddle-point bounds for large powers):

• Consider functions $A(z)$ and $B(z)$ satisfying the conditions L1,L2,L3 above. Let $lambda$ be a positive number with $0<lambda <T$ and let $zeta$ be the unique positive root of the equation

begin{align*}
zetafrac{B^{prime}{(zeta)}}{B(zeta)}=lambda tag{4}
end{align*}

Then, for $K=lambda k$ an integer; one has
begin{align*}
[z^K]A(z)cdot B(z)^kleq A(zeta)B(zeta)^kzeta^{-K}tag{5}
end{align*}

We start with calculating the roots of (4). We set $lambda =9$ and obtain according to (4)
begin{align*}
zfrac{B^{prime}(z)}{B(z)}&=9\
end{align*}
which gives with $B(z)=left(1+z+cdots+z^9right)^2$:
begin{align*}
color{blue}{9z^9+7z^8+5z^7+3z^6+z^5-z^4-3z^3-5z^2-7z-9=0}
end{align*}
from which we easily derive the positive root $color{blue}{zeta =1}$.

We find as upper bound according to (5)
begin{align*}
[z^{9k}]left(frac{1-z^{10}}{1-z}right)^{2k}leq left(sum_{k=0}^91right)^{2k}cdot 1^{-9k}=10^{2k}
end{align*}
This upper bound isn't really sharp but it may be useful whenever only rough order of magnitude estimates are sought.

Now we are well prepared for the main theorem.

Theorem VIII.8 (Saddle-point estimates of large powers).

• 
  

  Under the conditions of Proposition VIII.7, with $lambda = K/k$, one has
  begin{align*}
  color{blue}{[z^K]A(z)cdot B(z)^k=A(zeta)frac{B(zeta)^k}{zeta^{K+1}sqrt{2pi n xi}}left(1+o(1)right)}tag{6}
  end{align*}

  
  
  

  where $zeta$ is the unique root of $zeta B^{prime}(zeta)B(zeta)=lambda$ and

  
  

begin{align*}
xi=frac{d^2}{dzeta^2}left(log B(zeta)-lambdalog zetaright).tag{7}
end{align*}

• In addition, a full expansion in descending powers of $k$ exists.

These estimates hold uniformly for $lambda$ in any compact interval of $(0,T)$, i.e., any interval $[lambda^{prime},lambda^{primeprime}]$
with $0<lambda^{prime}<lambda^{primeprime}<T$, where $T$ is the spread.

Now it's time to harvest. At first we calculate $xi$ according to (7). We obtain with $B(z)=left(1+z+cdots+z^9right)^{2}$ and $lambda=9$:
begin{align*}
color{blue}{xi}&=left.left(frac{d^2}{dz^2}left(log (B(z)-9log zright)right)right|_{z=1}\
&=left.left(frac{d}{dz}left(frac{B^{prime}(z)}{B(z)}-frac{9}{z}right)right)right|_{z=1}\
&=frac{B^{primeprime}(1)}{B(1)}-left(frac{B^{prime}(1)}{B(1)}right)^2+9\
&=frac{177}{2}-81+9\
&,,color{blue}{=frac{33}{2}}tag{8}
end{align*}

Putting all together into (6) we finally obtain with $B(zeta)=B(1)=left.left(1+z+cdots+z^9right)^{2k}right|_{z=1}=10^{2k}$:
begin{align*}
color{blue}{[z^{9k}]left(frac{1-z^{10}}{1-z}right)^{2k}}
&=frac{10^{2k}}{1^{9k+1}sqrt{2pi k frac{33}{2}}}(1+o(1))\
&,,color{blue}{=10^{2k}frac{1}{sqrt{33pi k}}(1+o(1))}
end{align*}

in accordance with the claim (3).

Note: We have according to theorem VIII.8 the possibility to calculate a full expansion in descending powers of $k$. We can also study asymptotic expansions of $[z^K]B(z)^{k}$ for other quantities of $K$ as long as we fulfill the spread condition $K<Tk$.

You can prove that your asymptotic expression is correct using the Edgeworth series.

Let $F_k$ be the cdf for $sqrt{frac{k}{8.25}}(S_k/k-4.5)$. By the Central Limit theorem, $F_k(x)$ is approximately equal to $Phi(x)$. Specifically, the Edgeworth series shows that
$$
F_k(x)=Phi(x) -frac{lambda_3}{6sqrt k}Phi'''(x)+O(k^{-1})
$$
Here, $lambda_3$ is the skewness of a single random digit. Since this skewness is zero, as the distribution is symmetric about $[0,9]$, we get
$$
F_k(x)=Phi(x)+O(k^{-1}).
$$
This shows the error in the central limit approximation is linear in $k$. Therefore,
begin{align}
P(S_k=n)
&=P(n-tfrac12<S_kle n+tfrac12)
\&=F_kleft(frac{n+frac12-4.5k}{sqrt{8.25 k}}right)-F_kleft(frac{n-frac12-4.5k}{sqrt{8.25 k}}right)
\&=Phileft(frac{n+frac12-4.5k}{sqrt{8.25 k}}right)-Phileft(frac{n-frac12-4.5k}{sqrt{8.25 k}}right)+O(k^{-1})
\&stackrel1= Phi'(xi)cdotfrac1{sqrt{8.25k}}+O(k^{-1})
\&= frac1{sqrt{8.25k}}left(Phi'(xi)-Phi'left(frac{n-4.5k}{sqrt{8.25k}}right)+Phi'left(frac{n-4.5k}{sqrt{8.25k}}right)right)+O(k^{-1})
\&stackrel2= frac1{sqrt{8.25k}}Phi'left(frac{n-4.5k}{sqrt{8.25k}}right)+frac1{8.25k}Phi''(eta)+O(k^{-1})
\&stackrel3= frac1{sqrt{8.25k}}Phi'left(frac{n-4.5k}{sqrt{8.25k}}right)+O(k^{-1})
end{align}

Explanations:

1. Here, we are applying the mean value theorem. $xi$ is a number between $frac{npmfrac12-4.5k}{sqrt{8.25 k}}$.




2. Now we apply the mean value theorem to $Phi'$. Here, $eta$ is a number between $xi$ and $frac{n-4.5k}{sqrt{8.25 k}}$




3. We can absorb the $frac1{8.25k}Phi''(eta)$ into the $O(k^{-1})$ because $Phi''$ is bounded.

This is exactly the asymptotic expansion you guessed; however, the above rigorously shows that your answer is correct, with the error decreasing linearly as $ktoinfty$.