4

Currently I am reading the paper 'Excited Random Walk in One Dimension.'

At page $8$ left column, the authors obtain the following:

Probability that the walk eats precisely $r > 0$ consecutive cookies (we term this event a single “meal”) from the right edge of the cookie-free region is $$P(r) = 2q \frac{\Gamma(L)}{\Gamma(L-2q)} \frac{\Gamma(L+r-1-2q)}{\Gamma(L+r)}$$ where $L-2$ refers to cookie-free gap and $p$ refers to probability of the walk moving to the right and $q$ is the probability of the walk moving to the left.

However, when they calculate the average relative number of consecutive cookies eaten from the right side of the gap, they compute $$\int_0^\infty \tilde{r} \tilde{P}(\tilde{r})\,d\tilde{r}$$ where $\tilde{r} = \frac{r}{L}$ and $\tilde{P} = LP(r).$

Question: Why do they integrate with respect to $\tilde{r}$ with integrand $\tilde{P}?$ I thought to find the average number of cookie eaten, one just needs to compute $$\int_0^\infty r P(r)\, dr$$ instead of the above.

Idonknow
  • 15,643
  • Reading the paper this seems like they make the assumption that the first cooke free gap is large in which case the probability of $P(r)$ simplifies, as they state, and then this is normalized to get a probability again. To me this seems like they use asymptotic results. – Jan Oct 18 '18 at 12:24
  • Yes, they use asymptotic results. However, I do not understand why their calculation gives average number of cookies eaten. – Idonknow Oct 18 '18 at 13:14

1 Answers1

2

The two expressions denote a change of variable (change of scale): they apply $$ \eqalign{ & 1 = \int_{\,r\, = \,0}^{\;\infty } {P(r)dr} = \int_{\,r\, = \,0}^{\;\infty } {LP(r)d\left( {{r \over L}} \right)} = \cr & = \int_{\,r\, = \,0}^{\;\infty } {LP\left( {L{r \over L}} \right)d\left( {{r \over L}} \right)} = \int_{\,\tilde r\, = \,0}^{\;\infty } {LP\left( {L\tilde r} \right)d\tilde r} = \cr & = \int_{\,\tilde r\, = \,0}^{\;\infty } {\tilde P\left( {\tilde r} \right)d\tilde r} \cr} $$ to pass from $r,P(r)$ to $\tilde r,\tilde P\left( {\tilde r} \right)$ by putting $$ \left\{ \matrix{ \tilde r = r/L \hfill \cr \tilde P\left( {\tilde r} \right) = LP\left( {L\tilde r} \right) = LP\left( r \right) \hfill \cr} \right. $$

This is a totally licit and very common operation done in probability, for example when reconducing a Normal distribution with a given $\sigma$ to the standard one.

They explain that such a "standardization" allows to simplify (in some cases) the expressions by "absorbing" the $L$ parameter, which is in fact a scale parameter. The parallel with the Normal helps to understand why.

That premised, concerning your doubt on the average, $$ \int_{\,\tilde r\, = \,0}^{\;\infty } {\tilde r\,\tilde P\left( {\tilde r} \right)d\tilde r} $$ gives of course the average of $\tilde r$ , denoted as $ \left\langle {\tilde r} \right\rangle$ which tied to $ \left\langle {r} \right\rangle$ by $$ \left\langle {\tilde r} \right\rangle = \left\langle {r/L} \right\rangle = \left\langle r \right\rangle /L $$

In fact, soon after eq. (31) they speak of avg.$\tilde r$ as the "average relative number of consecutive cookies ..": relative is understood to refer to $/L$, and actually immediately below they give $\left\langle {\tilde r} \right\rangle = \left\langle {r/L} \right\rangle = \cdots $.

Addendum

Going back to eq.(30) reported at the beginning of your post $$ P(r) = 2q{{\Gamma (L)} \over {\Gamma (L - 2q)}}{{\Gamma (L + r - 1 - 2q)} \over {\Gamma (L + r)}} $$

The average number of $r$ would be given by $$ \left\langle r \right\rangle = \sum\limits_{0\, < \,r} {r\,P(r)} = 2q{{\Gamma (L)} \over {\Gamma (L - 2q)}}\sum\limits_{0\, \le \,r} {\left( {r + 1} \right)\,{{\Gamma (L + r - 2q)} \over {\Gamma (L + r + 1)}}} $$

In the above $q$ is a real number in the range $(0,1)$; the sum above can be expressed by means of the Gaussian Hypergeometric Function as $$ \left\langle r \right\rangle = {{2q} \over L}\;{}_2F_{\,1} \left( {2,\,L - 2q\,;\;L + 1\,;1} \right) $$ which, in virtue of the Gaussian theorem gives simply $$ \eqalign{ & \left\langle r \right\rangle = {{2q} \over L}{{\Gamma (L + 1)\Gamma ( - 1 + 2q)} \over {\Gamma (L - 1)\Gamma (1 + 2q)}} \quad \left| {\,0 < {\mathop{\rm Re}\nolimits} \left( { - 1 + 2q} \right)} \right.\quad = \cr & = \left\{ {\matrix{ {{{\left( {L - 1} \right)} \over {\left( {2q - 1} \right)}}} & {\left| \matrix{ \;1 \le L \hfill \cr \;1/2 < q \hfill \cr} \right.} \cr \infty & {\left| \matrix{\;1 \le L \hfill \cr \;q \le 1/2 \hfill \cr} \right.} \cr } } \right. \cr } $$ which, for large $L$, correspond to eq.(32).

To this regard we shall note that:
- the summand $\left( {r + 1} \right)\,{{\Gamma (L + r - 2q)} \over {\Gamma (L + r + 1)}}$ has a series expansion at $r=\infty$ which is $1/r^{2q} + O(1/r^{2q+1})$ and the sum is therefore convergent for $1<2q$;
- the Hypergeometric $ {}_2F_{\,1} \left( {a,\,b\,;\;c\,;z} \right)$ has a singularity at $z=1$, so that its value there shall be taken in the limit with due restrictions;
- the restrictions are those provided for the validity of its conversion into the fraction with Gammas, i.e. $0<1-2q$.

G Cab
  • 35,272
  • Yes. I understand that $\tilde{P}$ is a probability density function of $\tilde{r}$. However, I still couldn't understand why integrate $\tilde{r} \tilde{P}$ gives us average number. – Idonknow Oct 18 '18 at 13:43
  • it gives the average or $\tilde r$ of course, which means average of $r/L$, i.e. average of $r$ divided $L$. And you can always recover the "un-standard" quantity from the "standard" treatment. Are you familiar with scale parameters in the various important distributions ? – G Cab Oct 18 '18 at 14:05
  • I "absorbed" my previous comment in the answer. – G Cab Oct 18 '18 at 14:42
  • I am not very familiar with scale parameters in various distributions... – Idonknow Oct 18 '18 at 22:38
  • Regarding the newly added paragraph in your answer, so $\langle r\rangle$ is the average that we want right? They calculated $\langle\tilde{r}\rangle$ because the scaled parameter is easier to calculuate? – Idonknow Oct 18 '18 at 22:41
  • As for scale parameters, you might consider and have a look at this. It also somehow enlightens about your second comment: I will come back on that tomorrow. – G Cab Oct 19 '18 at 00:14
  • Regarding your last sentence in your answer, I notice that authors give a formula for $\langle \tilde{r} \rangle.$ However, it seems to me that the equal sign is not equal sign as numbers. – Idonknow Oct 19 '18 at 14:50
  • @Idonknow: concerning your question on why they pass to $\tilde r$, besides the fact that the scaled $r$ has an easier formulation, also the probability has. But the practice to "standardize" the formulas, specially in mathematics applied to physics, engineering, etc. is that by appropriately standardizing the formulas you get a better insight on the behaviour of the model that they represents (what is influencing what) and simplify their study (analysis, graphing,...) by working on a more concise representation. – G Cab Oct 19 '18 at 23:07
  • @Idonknow: concerning your question about the formula for the avg. of $\tilde r$, do you refer to eq. (32) ? if so, what is exactly your doubt ? – G Cab Oct 19 '18 at 23:16
  • I am referring to eq. (31), second equality. The authors derived second equality by using asymptotic approximation. – Idonknow Oct 19 '18 at 23:44
  • @G Cab So to calculate average number of cookie eaten, we just need to calculate $\langle r \rangle.$ To calculate average relative number of cookie eaten, it suffices to calculate $\left\langle {\tilde r} \right\rangle = \langle \frac{r}{L} \rangle = \frac{\langle r \rangle}{L}.$ Am I right? If yes, which one is the authors' intention? – Idonknow Oct 21 '18 at 11:55
  • @Idonknow: i posted in chat: better we continue our discussion there. – G Cab Oct 21 '18 at 18:32
  • @Idonknow: put and addendum which should answer to your doubts. – G Cab Oct 22 '18 at 00:55
  • @G Cab I understand better now due to your addendum. Thanks. By the way, I suppose $L\neq 1$ because the authors started with $L-2$ cookie-free gap. Also, in your addendum, how do you obtain the last piecewise function for $\langle r \rangle?$ From Gauss's Hypergeometric Series, we know that Gauss's formula holds if $L+1>(L-2q)+2,$ which is equivalent to $q>\frac{1}{2}.$ However, if the inequality is not satisfied, can we conclude that Gauss's formula is $\infty?$ – Idonknow Oct 22 '18 at 02:47
  • @Idonknow: range of convergence explicitated. – G Cab Oct 22 '18 at 17:56
  • @G Cab Thanks. I have another question based on the same paper. Would you be able to help? https://math.stackexchange.com/questions/2965575/random-walk-compute-average-ratio-for-the-number-of-right-cookies-to-total-coo – Idonknow Oct 24 '18 at 14:42