Yamagishi Complexity Torus LabYamagishi Lab

[Paper 1120] The 2 in “Two Dimensions,” the Border for a Random Walk to Come Back, Is the Order of the Diffusion Equation, and the Lattice Changes Only the Value of the Return Probability

Abstract

Paper 199 counted, on the side of time, whether a point walking at random on a lattice, one step to a nearest neighbour at a time, comes back to the origin. The probability of being at the origin after 2n steps falls off as n to the power −d/2, and up to two dimensions, where the sum of these probabilities grows without bound, the walk is certain to come back (Pólya 1921). Paper 482 counted that the integral of k to the power d−3 in wavenumber space becomes finite beyond two dimensions, and Paper 1106 placed it on a map that compares the exponent of thinning with the dimension. This paper counts the expected number G of visits to the origin over shells in wavenumber space, places it on the same map, and checks what the 2 of the border is by changing the lattice and the step length. This paper claims no new theorem or law.

Written with the characteristic function φ of one step, G is the integral of 1/(1−φ) over the box of wavenumbers. For a walk stepping to nearest neighbours, 1−φ is proportional to k² at small wavenumber k, so the weight of one shell goes as k to the power d−3, and on the map the walk sits at the origin end of wavenumber space with exponent 2 (the seat of the Bose gas in Paper 1106). At k = 0.001 the shell weight agreed with 2/π, 2/π and 3/π² times k to the power d−3 to a relative difference below 10⁻⁷. Computing the expected number of visits up to N steps both as a sum over time and as an integral over wavenumbers, the two differ by 1.20×10⁻¹⁴ for the square lattice at N = 1000; each time N is multiplied by 100, the growth factor approaches 10 in one dimension, grows by (1/π) ln 100 = 1.465871 in two dimensions, and stops growing in three.

For the simple cubic lattice, two numerical routes on the side of wavenumbers (an integral of Bessel functions and a double integral left after summing one direction) agree within 10⁻¹³ with the known closed form as a product of gamma functions and give G = 1.516386059151978; a route on the side of time, adding an estimate of the tail to the sum, agrees within 1.65×10⁻⁷. The return probability F = 1−1/G is 0.340537329551, the same to the fourth decimal place as the 0.340515178 that Paper 199 extrapolated on the side of time (difference 2.215×10⁻⁵).

The lattice changes values but not whether the walk comes back. The return probabilities of the body-centred and face-centred cubic lattices are 0.282229988954 and 0.256318236505; together with the simple cubic lattice, none of the three reaches 1, and the more neighbours (6, 8, 12), the smaller it is. At small k, 1−φ is proportional to k² for all three, with only the coefficient differing: 1/6, 1/2 and 1/3. On the square, triangular and honeycomb lattices in two dimensions, the expected number of visits up to N steps grows in proportion to ln N, so the walk is certain to come back, and the coefficient of ln N approaches the area per site of the lattice with unit bonds divided by π: 0.318310, 0.275664 and 0.413497.

What moves is the exponent of the shell weight. For a walk in one dimension that chooses a step length j with probability proportional to |j| to the power −1−α, 1−φ is proportional to k to the power α (for α below 2), the weight of one shell goes as k to the power d−1−α, and the border moves to d = α. For α = 1.5, G grows without bound and the walk is certain to come back; α = 1 sits exactly on the border, where G = (1/3) ln((2π−k₀)/k₀) grows logarithmically with the lower cutoff k₀ in wavenumber. For α = 0.5, G = 1.256425 and the return probability is 0.204091, so even in one dimension the walk may fail to come back. In two dimensions too, the return probability of a walk with α = 1.5 is 0.355501. On the side of time, the exponent of the probability of being at the origin after n steps approaches 1/α (in one dimension). The α = 2 of a walk stepping to nearest neighbours is the 2 of the diffusion equation, seen coarsely, being of second order in space; both the 2 in the exponent d/2 counted in Paper 199 and the 2 of the border “two dimensions” are this α.

On the side of position, the same number becomes the resistance from the origin to infinity in a lattice of 1-ohm resistors (G divided by the number of neighbours). On the square lattice the resistance grows by ln2/(2π) = 0.110318 each time the half-width of the grounded box is doubled; on the simple cubic lattice it stops at G/6 = 0.252731 (the extrapolation from growing boxes differs by 3.39×10⁻⁷). The logarithm at the origin of wavenumber space and the logarithm at the outer end in position are one and the same infinity. The separator is whether a quantity is a value the lattice changes or the exponent of the weight of one shell.

Keywords: random walk, Pólya’s theorem, return probability, shells in wavenumber space, lattice Green function, heavy-tailed steps, separator.

Scope (what is not claimed)

This paper claims no new theorem or law. That the walk on a simple lattice comes back to the origin with probability 1 up to two dimensions and may fail to come back from three dimensions on (Pólya 1921), the relation F = 1−1/G between the return probability and the expected number of visits to the origin (Feller 1968), deciding this by an integral over wavenumbers (a revised form of the criterion of Chung and Fuchs 1951 is said to be in Chapter 2, Section 8 of Spitzer 1964; the book has not been opened), that a walk whose step has finite variance and spreads in every direction comes back up to two dimensions whatever the lattice (said to follow from the same criterion; this paper counted only six lattices), that Watson 1939 evaluated G for the three cubic lattices in closed form (Watson’s integrals are one third of G for the simple and face-centred cubic lattices and G itself for the body-centred one), and the correspondence between a lattice of resistors and a random walk (Doyle and Snell 1984) are all known. The form of G for the simple cubic lattice as a product of gamma functions is due to Glasser and Zucker 1977 (checked by title and abstract; the paper itself has not been read). Here the closed forms from the literature were used to check the numerical routes. That a walk with heavy-tailed step lengths comes back for d ≤ α also appears to be known (the original has not been checked here). Pólya’s theorem is not proved. What is added is placing the number of visits of a random walk on the map of Paper 1106 (the origin end of wavenumber space, with exponent α); computing the return probability in three dimensions by two numerical routes on the side of wavenumbers, bringing it into agreement with the closed form, and comparing it with the extrapolation of Paper 199; counting, on three cubic lattices and three flat lattices, that the lattice changes only values (the return probability and the coefficient of the logarithm); counting that the tail exponent of the step length moves the border to d = α, so that a walk may fail to come back even in one or two dimensions; and counting the same logarithm as a resistance on the side of position. The scope note of Paper 199 wrote that it treats neither general lattices nor weighted walks and does not derive the exact return probability in three dimensions (0.340515 is an extrapolation). The scope note of Paper 482 wrote that other situations share the same integrand but numbers were given for only two, and Paper 1106 sat in that seat and put numbers on ten infinities. This paper sits in the seats of these two scope notes. Paper 199 counted, on the side of time, the exponent d/2 of the probability of being at the origin, the border 2, and that the mean time until return is infinite. The abstract and closing of Paper 199 wrote that whether one can come home is decided by the dimension alone, not by the terrain or the step length, and that what decides it is the exponent d/2 standing in the denominator. Paper 199 counted the walk to nearest neighbours, and for any walk whose step has finite variance this statement holds whatever the lattice or the size of the step (Section 3). When the tail of the step length is heavy, the exponent in the denominator becomes d/α (Section 4), and whether the walk comes back is decided by comparing the dimension with α. The core of Paper 199, that the exponent in the denominator decides, remains as it is. Paper 204 lined up the same d/2 in the heat kernel and in counting eigenvalues. Paper 249 counted, for continuous diffusion, that there is no steady absorption in two dimensions (the flux falls off as 1/log). Paper 264 lined up, for percolation, that the threshold moves with the lattice while the exponents do not, and Paper 1033 counted that square and triangular nets share an exponent and differ only in the coefficient. The mean time until return (Paper 199), biased walks, self-avoiding walks (Paper 279), the finer properties of Brownian motion in continuous time, the side of counting eigenvalues (Paper 204), the coarse equation of heavy-tailed walks itself (a diffusion equation of fractional order), and counting heavy-tailed walks by a network of resistors on the side of position are not treated.

1. Counting the visits over shells in wavenumber space

A point on a lattice moves at each step with fixed probabilities. It starts at the origin, and $p_n$ is the probability of being at the origin after $n$ steps. With the expected number of visits to the origin (including step 0) written as

$$G=\sum_{n=0}^{\infty}p_n,$$

the probability of ever coming back to the origin is $F=1-1/G$ (Feller 1968). If $G$ grows without bound, $F=1$ and the walk is certain to come back; if it is finite, the walk may fail to come back. Paper 199 counted that, for a walk stepping to nearest neighbours, $p_{2n}$ falls off in proportion to $n^{-d/2}$, and looked at the border through this sum.

Let the characteristic function of one step be $\varphi(\mathbf k)=\sum_{\mathbf j}P(\mathbf j)\cos(\mathbf k\cdot\mathbf j)$ ($P(\mathbf j)$ is the probability of moving by $\mathbf j$ in one step; left and right are equally likely, so only cosines remain). Since $p_n$ is the average of $\varphi^n$ over the box of wavenumbers $[-\pi,\pi]^d$,

$$G=\frac{1}{(2\pi)^d}\int_{[-\pi,\pi]^d}\frac{d^dk}{1-\varphi(\mathbf k)}.$$

On the simple lattice in $d$ dimensions (one step to a nearest neighbour, choosing among $2d$ directions with equal probability), $\varphi=(\cos k_1+\cdots+\cos k_d)/d$, and at small $k$, $1-\varphi\approx k^2/(2d)$. The weight of one shell of radius $k$ is

$$w(k)=\frac{|S^{d-1}|\,k^{d-1}}{(2\pi)^d}\cdot\frac{2d}{k^2}\propto k^{d-3},$$

the same form as the integrand of $\int k^{d-3}dk$ counted in Paper 482 ($|S^{d-1}|$ is the area of the unit sphere, two points in one dimension). Adding up $1/(1-\varphi)$ over the angles of the shell numerically and multiplying by $k^{3-d}$, the values approached $2/\pi$, $2/\pi$ and $3/\pi^2$ as in the table (the relative difference at $k=0.001$ was $5.00\times10^{-8}$ to $8.33\times10^{-8}$).

$d$weight of one shell$k=0.1$$k=0.01$$k=0.001$limit
1$\propto k^{-2}$0.63715055420.63662507760.6366198254$2/\pi=0.6366197724$
2$\propto k^{-1}$0.63701781170.63662375130.6366198122$2/\pi$
3$\propto k^{0}$0.30411557900.30396507070.3039635661$3/\pi^2=0.3039635509$

In the language of the map of Paper 1106, the density $1/(1-\varphi)$ is proportional to $k^{-2}$ at the origin of wavenumber space, and the exponent of thinning is $n=2$ ($n$ is the symbol of 1106, not the number of steps). The sum is finite at the origin end when $n<d$, so $G$ is finite from three dimensions on. On the map of 1106 this is the same place as the seat of the uncondensed particles of the Bose gas in two and three dimensions (the same $1/k^2$).

The sum over time and the integral over wavenumbers are two ways of writing the same number. The sum $G_N$ up to $N$ steps (from step 0 to step $N-1$) is the average of $(1-\varphi^N)/(1-\varphi)$ over the box of wavenumbers. For the square lattice at $N=1000$, the sum over time and the average over $1024\times1024$ points placed evenly in the box of wavenumbers (the integrand is a trigonometric polynomial, so with more points than its degree the average equals the integral exactly) were both 3.0442879279684, differing by $1.20\times10^{-14}$. The factor $\varphi^N$ acts as a cutoff that removes $k$ below about $1/\sqrt N$. Reading $n\leftrightarrow1/k^2$, the sum $\sum n^{-d/2}$ becomes the integral $\int k^{d-3}dk$. The $p$ in the table of Paper 199 (four of them, for $d=1$, 2, 3) were reproduced by the computation here to the seven digits of that table.

$N$one dimensiontwo dimensionsthree dimensions (simple cubic)
1007.958923742.30992327661.4502368696
$10^4$79.786461393.77736673341.5097874420
$10^6$797.884361335.2432536872(not counted)
$N\to\infty$without boundwithout bound1.516386059151978

Each time $N$ is multiplied by 100, $G_N$ in one dimension grows by factors of 10.02478024 and 10.00024750, approaching 10 (it grows as $\sqrt N$); in two dimensions it grows by 1.4674434568 and 1.4658869538, approaching $(1/\pi)\ln100=1.4658711978$ (it grows logarithmically). In three dimensions it stops; the remainder $G-G_N$ times $\sqrt N$ is 0.660010 at $N=10^3$ and 0.659862 at $N=10^4$, approaching the tail estimate $4\sqrt2 (3/4\pi)^{3/2}=0.659845$.

2. The return probability in three dimensions from the side of wavenumbers

$G$ for the simple cubic lattice was computed by three routes and compared with a closed form from the literature. (a) Writing $1/(1-\varphi)=\int_0^\infty e^{-t(1-\varphi)}dt$, the average over each of the three directions becomes a modified Bessel function $I_0(t/3)$, and $G=\int_0^\infty e^{-t}I_0(t/3)^3 dt$ becomes a single integral (computed to 30 digits). (b) The average over $k_3$ can be done in closed form, and the remaining double integral was summed by Gaussian quadrature with points crowded towards both ends. (c) On the side of time, the probability of being at the origin after $2n$ steps was summed exactly by splitting the steps into $2j$ steps along $x$ and the rest on the square lattice of the plane, and an estimate of the tail was added to the sum up to $N=10^4$. (d) As a check, the closed form from the literature as a product of gamma functions (Glasser and Zucker 1977),

$$G=\frac{\sqrt6}{32\pi^3}\,\Gamma\Bigl(\frac{1}{24}\Bigr)\Gamma\Bigl(\frac{5}{24}\Bigr)\Gamma\Bigl(\frac{7}{24}\Bigr)\Gamma\Bigl(\frac{11}{24}\Bigr),$$

was computed to 30 digits.

(a) gave 1.5163860591519780144 and (d) gave 1.5163860591519780182, a difference of $3.78\times10^{-18}$. (b) with 800 points gave 1.516386059152024, differing from (d) by $4.57\times10^{-14}$, and (c) gave 1.5163858942, differing by $1.65\times10^{-7}$. The two numerical routes (a) and (b) on the side of wavenumbers agree with the closed form (d) within $10^{-13}$, $G=1.516386059151978$, and the return probability is

$$F=1-\frac{1}{G}=\mathbf{0.340537329551}.$$

This $G$ corresponds to one of the three triple integrals that Watson (1939) evaluated in closed form (three times it).

Paper 199 extrapolated 0.340515178 from the sum over time for $2n\le600$ and wrote that this is an extrapolation and not a derivation. The difference from the value here is $2.215\times10^{-5}$ (relatively $6.505\times10^{-5}$), the same to the fourth decimal place. For $G$ too, the 1.516335126 of 199 and the 1.516386059 here differ by $5.09\times10^{-5}$. On the side of time the remainder shrinks only as $1/\sqrt N$, so even the sum up to $N=10^4$ falls short of $G$ by $6.6\times10^{-3}$ (the table of Section 1). On the side of wavenumbers the same number comes out as one finite integral.

3. The lattice changes the value

The simple cubic lattice (6 neighbours), the body-centred cubic lattice (8 neighbours, one step being $(\pm1,\pm1,\pm1)$) and the face-centred cubic lattice (12 neighbours, one step being $(\pm1,\pm1,0)$ and its permutations) are compared. Their $\varphi$ are $(\cos k_1+\cos k_2+\cos k_3)/3$, $\cos k_1\cos k_2\cos k_3$ and $(\cos k_1\cos k_2+\cos k_2\cos k_3+\cos k_3\cos k_1)/3$.

latticeneighbours $z$$1-\varphi$ at small $k$$G$return probability $F$resistance to infinity $G/z$
simple cubic6$k^2/6$1.5163860591520.3405373295510.2527310099
body-centred cubic8$k^2/2$1.3932039296860.2822299889540.1741504912
face-centred cubic12$k^2/3$1.3446611831650.2563182365050.1120550986

For the body-centred cubic lattice the three coordinates move by $\pm1$ independently, so $G$ is also $\sum_n\bigl(\binom{2n}{n}/4^n\bigr)^3$; this sum, added up as a hypergeometric series to 30 digits, and the closed form $\Gamma(1/4)^4/(4\pi^3)$ are the same in the 18 digits shown, and the double integral agrees with them within $5.93\times10^{-14}$. For the face-centred cubic lattice the closed form $9 \Gamma(1/3)^6/(2^{14/3}\pi^4)$ and the double integral agree within $4.53\times10^{-14}$. For all three the return probability does not reach 1, and the more neighbours, the smaller it is. At small $k$, $1-\varphi$ is proportional to $k^2$ for all three (three directions compared at $k=10^{-4}$ gave no dependence on direction), with coefficients $1/6$, $1/2$ and $1/3$ that are the only difference, and the weight of one shell is $k^0$ for all three. In the body-centred and face-centred cubic lattices, however, the box of wavenumbers contains 4 and 2 points where $\varphi=1$ (for the body-centred lattice $(0,0,0)$, $(\pi,\pi,0)$ and two more; for the face-centred lattice $(0,0,0)$ and $(\pi,\pi,\pi)$). The coefficient of the shell weight summed over all of them is $3/\pi^2$ for the simple, $4/\pi^2$ for the body-centred and $3/\pi^2$ for the face-centred cubic lattice. The simple and face-centred cubic lattices have the same coefficient, yet $G$ is 11.3 percent smaller for the face-centred one, and the body-centred lattice has a larger coefficient but a smaller $G$. The value of $G$ is decided not only by the coefficient of the shell weight but by the shape of the whole box of wavenumbers. This value is what the lattice changes; it does not change whether $G$ is finite or infinite.

In two dimensions, the square (4 neighbours), triangular (6 neighbours) and honeycomb (3 neighbours) lattices were compared. The honeycomb lattice has two kinds of sites alternating, so it was viewed two steps at a time, as a walk on the sites of one kind. $G_N$ up to $N$ steps was found for $N=256$, 1024 and 4096 as the average over evenly placed points in the box of wavenumbers (with more points than the degree, so that the average equals the integral exactly).

latticeneighboursarea per sitegrowth 256 to 1024growth 1024 to 4096(area$/\pi$)$\times\ln4$coefficient of $\ln N$
square410.441737000.441387740.441271200.318310
triangular6$\sqrt3/2$0.382151580.382152040.382152070.275664
honeycomb3$3\sqrt3/4$0.573833570.573379520.573228100.413497

For all three, $G_N$ grows by nearly the same amount each time $N$ is multiplied by 4, so it grows logarithmically and the walk is certain to come back. The growth approaches (area per site$/\pi$)$\times\ln4$, and the coefficient of $\ln N$ is $1/\pi=0.318310$, $\sqrt3/(2\pi)=0.275664$ and $3\sqrt3/(4\pi)=0.413497$ (with unit bonds, the area per site is $1$, $\sqrt3/2=0.866025$ and $3\sqrt3/4=1.299038$). In all three lattices one step has variance $1/2$ per direction, and the only difference that acts on the coefficient of the logarithm is the density of sites. The coefficient of the logarithm changes with the density of sites; that it is a logarithm does not.

Paper 264 lined up, for percolation, that the concentration at which things begin to connect moves with the lattice while the critical exponents are fixed by the dimension alone, and Paper 1033 counted that square and triangular nets describe the world with the same exponent and differ only in the coefficient. The return probability of a random walk splits in the same way.

4. When the tail of the step length is heavy, the border moves

In one dimension, consider a walk that chooses a step length $j$ (a nonzero integer) with probability proportional to $|j|^{-1-\alpha}$: $P(j)=|j|^{-1-\alpha}/(2\zeta(1+\alpha))$, with left and right equally likely. The smaller $\alpha$, the more often it jumps far. $\varphi(k)$ is the real part of the polylogarithm $\mathrm{Li}_{1+\alpha}(e^{ik})$ divided by $\zeta(1+\alpha)$, and it agreed with the power-series expansion (a closed polynomial for $\alpha=1$ and 3) within a relative $7.63\times10^{-16}$. At small $k$, for $\alpha<2$,

$$1-\varphi(k)\approx c_\alpha\,|k|^{\alpha},\qquad c_\alpha=-\frac{\Gamma(-\alpha)\cos(\pi\alpha/2)}{\zeta(1+\alpha)};$$

for $\alpha=1$ exactly $1-\varphi=(6/\pi^2)(\pi k/2-k^2/4)$ ($0\le k\le2\pi$), and for $\alpha=3$ the variance of one step is finite, so it is proportional to $k^2$. The weight of one shell is $k^{d-1-\alpha}$ (for $\alpha$ larger than 2, read 2, as for the walk to nearest neighbours), the exponent on the map of Paper 1106 (the $n$ of 1106) is $\alpha$, and the border is at $d=\alpha$.

$\alpha$$1-\varphi$ at small $k$weight of one shellorigin endcomes back?
0.5$0.959520719675 |k|^{0.5}$$k^{-0.5}$finite ($G=1.25642453885894$)return probability 0.204090680282
1$(3/\pi) |k|$$k^{-1}$grows logarithmicallyyes (with probability 1)
1.5$1.24569615357 |k|^{1.5}$$k^{-1.5}$grows as $k_0^{-1/2}$yes (with probability 1)
3$0.759908877318 k^{2}$$k^{-2}$grows as $1/k_0$yes (with probability 1)

For $\alpha=0.5$, $G$ summed with two different substitutions agreed within $10^{-12}$: $G=1.25642453885894$, and the return probability is $\mathbf{0.204090680282}$. Even in one dimension, about four walks out of five never come back. $\alpha=1$ is exactly on the border; with the lower cutoff in wavenumber written $k_0$,

$$G(k_0)=\frac{1}{\pi}\int_{k_0}^{\pi}\frac{dk}{1-\varphi}=\frac13\ln\frac{2\pi-k_0}{k_0},$$

and the numerically summed values and the closed form agreed at $k_0=10^{-3}$, $10^{-6}$ and $10^{-9}$: 2.9151577259273, 5.2177958217396 and 7.5203809677322. The growth on dividing $k_0$ by 1000 was 2.30263809581 and 2.30258514599, approaching $\ln10=2.30258509299$. For $\alpha=1.5$, $G(k_0)$ at $k_0=10^{-2}$, $10^{-4}$ and $10^{-6}$ is 5.779436982, 52.29760164 and 512.7625387, growing by factors of 9.0489094 and 9.8047047 each time $k_0$ is divided by 100, approaching 10 ($k_0^{-1/2}$). For $\alpha=3$ it is 42.6799897, 4190.196631 and 418881.0409, with factors 98.177077 and 99.966918 approaching 100 ($1/k_0$), the same growth as for the walk to nearest neighbours.

On the side of time, the slope on logarithmic axes of the probability $p_n=(1/\pi)\int_0^\pi\varphi(k)^n dk$ of being at the origin after $n$ steps approaches $-1/\alpha$ ($-1/2$ for $\alpha$ larger than 2).

$\alpha$$n=100$ to 1000$n=1000$ to 10000$-1/\alpha$
0.5$-1.9883824$$-1.9988285$$-2$
1$-0.99739902$$-0.99973947$$-1$
1.5$-0.68173444$$-0.67387956$$-0.66666667$
3$-0.50493009$$-0.50181780$$-1/2$ (finite variance, read as 2)

For $\alpha=1.5$ the approach is slow; this appears to be because the next term of $1-\varphi$ is $k^2$, whose relative size shrinks only as $k^{0.5}$. In $d$ dimensions the exponent is $d/\alpha$, and $\sum n^{-d/\alpha}$ grows without bound when $d\le\alpha$. Reading $n\leftrightarrow k^{-\alpha}$, this is the same comparison as the shell weight $k^{d-1-\alpha}$.

In two dimensions, a walk was considered that chooses one of the two directions with equal probability and jumps along it with the same heavy-tailed step length ($\varphi=(\varphi_1(k_1)+\varphi_1(k_2))/2$). For $\alpha=1.5$ the shell weight is $k^{-0.5}$ and the origin end is finite. Summing over the singularity at the origin with a substitution that splits the square of wavenumbers into two triangles and spreads the origin into one side, doubling the points from 160 to 320 changed the value by $3.69\times10^{-12}$; written to digits coarser than that difference, $G=1.5515920424$ and the return probability is $\mathbf{0.35550069046}$. Even in two dimensions, a walk with a heavy tail of step length may fail to come back.

walk$d$$\alpha$shell exponent $d-1-\alpha$$d-\alpha$origin end
1D, nearest neighbour12$-2$$-1$grows as a power (certain return)
2D, nearest neighbour22$-1$0grows logarithmically (certain return)
3D, nearest neighbour3201finite ($F=0.340537$)
1D, $\alpha=1.5$11.5$-1.5$$-0.5$grows as a power (certain return)
1D, $\alpha=1$11$-1$0grows logarithmically (certain return)
1D, $\alpha=0.5$10.5$-0.5$0.5finite ($F=0.204091$)
2D, $\alpha=1.5$21.5$-0.5$0.5finite ($F=0.355501$)

For the walk to nearest neighbours (and any walk whose step has finite variance), $1-\varphi$ begins with $k^2$. Seen coarsely, the motion of $p$ obeys the diffusion equation $\partial p/\partial t=D\nabla^2p$, and the form of $\nabla^2$ in wavenumbers is $-k^2$. That $1-\varphi$ begins with $k^2$ is the same thing as this equation being of second order in space. When the tail of the step length is heavy, the right-hand side of the coarse equation becomes a derivative of fractional order corresponding to $|k|^\alpha$ (the equation itself is not treated here; only the exponent of $1-\varphi$ was counted). The border “two dimensions” is where the dimension $d$ equals the order $\alpha=2$. The $d/2$ of Paper 199 is the $d/\alpha$ on the side of time seen at $\alpha=2$, and the same $d/2$ lined up by Paper 204 in the trace of the heat kernel also comes from the eigenvalues being arranged by the square of the wavenumber (the side of 204 is not computed here). Paper 1107 counted that the field equation being of second order fixes how many numbers can be read from outside a closed skin, and separated being of second order and the field being massless as distinct roots. A random walk needs both. For a walk that vanishes with probability $q$ at each step (corresponding to a mass), $G$ is finite even on the square lattice: 2.1368782611111, 2.8619855541775 and 3.5938129727794 for $q=0.01$, 0.001 and 0.0001 (the closed form $(2/\pi)K((1-q)^2)$ and the sum over time agree within $10^{-17}$; $K$ is the complete elliptic integral of the first kind, written with the squared modulus as its argument). What fixes the border is the second order when the walk does not vanish ($\varphi(0)=1$, massless in the sense of 1107). Paper 190 counted that in sums with heavy tails one big jump matters. That a walk may fail to come back even in one dimension can be read as the occasional big jump carrying it far from the origin.

The abstract and closing of Paper 199 wrote that whether one can come home is decided by the dimension alone, not by the terrain, the step length or luck, and that what decides it is the exponent $d/2$ standing in the denominator. Paper 199 counted the walk to nearest neighbours, and for any walk whose step has finite variance this statement holds whatever the lattice or the size of the step (Section 3). When the tail of the step length is heavy, the exponent in the denominator changes from $d/2$ to $d/\alpha$, and whether the walk comes back is decided by comparing the dimension with $\alpha$. The core of Paper 199, that the exponent in the denominator decides, remains as it is.

5. On the side of position it becomes the resistance to infinity

Connect nearest neighbours of the lattice by 1-ohm resistors, feed a current of 1 into the origin, and ground the lattice far from the origin. The voltage at the origin is the resistance $R$ from the origin to infinity. By the correspondence between random walks and networks of resistors (Doyle and Snell 1984), the probability $1/G$ that a walk leaving the origin never comes back equals $1/(zR)$ ($z$ is the number of neighbours). That is, $R=G/z$, and whether the walk comes back is the same as whether the resistance to infinity grows without bound. The right-hand column of the table of Section 3 is this $G/z$.

The outside of a box of half-width $L$ was grounded and the voltage at the origin was solved numerically (a direct method in two dimensions and the conjugate gradient method in three; the residuals are smaller than $2\times10^{-13}$).

$L$voltage, two dimensions (square)growth on doubling $L$
250.7816746093
500.89201237940.1103377701
1001.00233516920.1103227898
2001.11265421650.1103190473
4001.22297232840.1103181119

On the square lattice in two dimensions the resistance grows each time $L$ is doubled, and the growth approaches $\ln2/(2\pi)=0.1103178001$. It grows without bound, but logarithmically. On the side of wavenumbers, the growth of $G_N$ when $N$ is multiplied by 4, divided by the number of neighbours 4, is 0.1103469348 ($N=1024$ to 4096), approaching the same number. Doubling $L$ corresponds to multiplying by 4 the number of steps, which is proportional to $L^2$. On the simple cubic lattice in three dimensions the resistance approaches its limit as $1/L$; extrapolating from $L=32$ and 64 gives 0.25273135, agreeing with $G/6=0.2527310099$ from the side of wavenumbers within $3.39\times10^{-7}$. In one dimension, with both ends grounded, the resistance is $L/2$ and grows in proportion to $L$.

Section 5 of Paper 1106 confirmed that an infinity overflowing at the centre in position overflows at the outside in wavenumbers, with the exponent of a squared quantity folding about the border $n=d$. The same section named in one sentence that a divergence at long distances lies at the outside in position and at the origin in wavenumbers; here numbers are put on that direction. The potential of a current fed in at a point is proportional to $r^{2-d}$, and the density of dissipated power to $r^{2-2d}$. The exponent in position is $2d-2$, which added to the exponent 2 in wavenumbers gives $2d$. In two dimensions $2d-2=2=d$, a logarithm at the outer end; in three dimensions $4>3$, finite at the outer end; in one dimension $0<1$, growing in proportion to $L$. Paper 249 counted that, for continuous diffusion, there is no steady absorption in two dimensions and the flux falls off as $1/\log$. This too is the same logarithm.

6. Same root or distinct roots

The quantities of this paper are sorted with the test of Paper 300.

(1) The logarithm of the random walk in two dimensions, the logarithm of the Bose gas in two dimensions (Paper 482), the logarithm of the network of resistors in two dimensions, and the absence of steady absorption in two dimensions in Paper 249: all come from the $1/k^2$ at the origin of wavenumber space making the shell weight $1/k$ in two dimensions, so they share a root. Counting in position or in wavenumbers is two ways of writing the same infinity.

(2) The differences between lattices (return probabilities 0.340537, 0.282230 and 0.256318; coefficients of $\ln N$ 0.318310, 0.275664 and 0.413497) and whether the walk comes back: the former is fixed by everything outside the exponent (the coefficient of the shell weight and the shape of the whole box of wavenumbers) and the latter by the sign of the exponent of the shell weight, so their roots are distinct. This is the same form as the numbers that move and do not move with the lattice, separated by Paper 264 for percolation.

(3) The 2 in the exponent $d/2$ of Paper 199 and the 2 of the border “two dimensions”: both are $\alpha=2$, so they share a root. When the tail exponent of the step length is moved, the two move together to $d/\alpha$ and $d=\alpha$. The border “two dimensions” is where two 2s meet, the dimension of space and the order of the walk, in the same form as Paper 1092 separating the 2 of the density $1/r^2$ into two.

(4) That a walk may fail to come back in three dimensions, and that it may fail to come back for $\alpha=0.5$ in one dimension or $\alpha=1.5$ in two: all have $d-\alpha>0$, so they share a root.

7. The separator

The separator is whether a quantity is a value the lattice changes or the exponent of the weight of one shell. This is the separator of Paper 1106 (whether a coefficient multiplying the whole or the exponent of the weight of one shell) applied to random walks; what corresponds to the overall coefficient of 1106 is here the value the lattice changes, though the lattice changes not only a multiplying coefficient but also the shape of the whole box of wavenumbers (Section 3). Changing the lattice (the number of neighbours, the arrangement of sites) changes the value of the return probability and the coefficient of $\ln N$, but not whether the walk comes back. Changing the tail exponent $\alpha$ of the step length changes the exponent $d-1-\alpha$ of the weight of one shell, and changes whether the walk comes back. For the walk to nearest neighbours $\alpha=2$, and the border is at two dimensions.

Closing

Whether a random walk comes back to the origin is decided by whether the expected number G of visits to the origin is finite. Counted over shells in wavenumber space, the weight of one shell goes as k to the power d−1−α; for the walk to nearest neighbours α = 2 and the border is at two dimensions. This 2 is the 2 of the diffusion equation, seen coarsely, being of second order; when the tail of the step length is heavy, α falls below 2, the border moves to d = α, and the walk may fail to come back even in one or two dimensions. The lattice changes the value of the return probability (0.340537, 0.282230 and 0.256318 for the simple, body-centred and face-centred cubic lattices in three dimensions) and the coefficient multiplying ln N in two dimensions, but not whether the walk comes back. On the side of position, the same number becomes the resistance from the origin to infinity in a lattice of resistors, which grows logarithmically in two dimensions and stops in three.

References

  1. Chung, K. L., and Fuchs, W. H. J. (1951). On the distribution of values of sums of random variables. Memoirs of the American Mathematical Society.
  2. Doyle, P. G., and Snell, J. L. (1984). Random Walks and Electric Networks. Mathematical Association of America.
  3. Feller, W. (1968). An Introduction to Probability Theory and Its Applications (Volume 1, 3rd ed.). Wiley.
  4. Glasser, M. L., and Zucker, I. J. (1977). Extended Watson integrals for the cubic lattices. Proceedings of the National Academy of Sciences of the United States of America.
  5. Pólya, G. (1921). Über eine Aufgabe der Wahrscheinlichkeitsrechnung betreffend die Irrfahrt im Straßennetz. Mathematische Annalen.
  6. Spitzer, F. (1964). Principles of Random Walk. Van Nostrand.
  7. Watson, G. N. (1939). Three triple integrals. The Quarterly Journal of Mathematics.
  8. Yamagishi, Y. (2026). “Rare” Is Measured on a Logarithmic Scale [Paper 190]. YOI: https://www.yamagishiyuuki.com/yoi/190; DOI: 10.5281/zenodo.22106182
  9. Yamagishi, Y. (2026). What Decides Whether You Come Home Is the Exponent in the Denominator [Paper 199]. YOI: https://www.yamagishiyuuki.com/yoi/199; DOI: 10.5281/zenodo.22108396
  10. Yamagishi, Y. (2026). The Power in the Eigenvalue Count Is Half the Dimension [Paper 204]. YOI: https://www.yamagishiyuuki.com/yoi/204; DOI: 10.5281/zenodo.22110032
  11. Yamagishi, Y. (2026). The Four Pi That Catches Molecules Is the Four Pi That Spreads a Field [Paper 249]. YOI: https://www.yamagishiyuuki.com/yoi/249; DOI: 10.5281/zenodo.22160409
  12. Yamagishi, Y. (2026). Inside One Phenomenon Live a Number That Moves with the Lattice and a Number That Does Not [Paper 264]. YOI: https://www.yamagishiyuuki.com/yoi/264; DOI: 10.5281/zenodo.22169720
  13. Yamagishi, Y. (2026). What Changed the Exponent of a Chain Was Self-Avoidance Alone, and above Four Dimensions That Cost Disappears [Paper 279]. YOI: https://www.yamagishiyuuki.com/yoi/279; DOI: 10.5281/zenodo.22170222
  14. Yamagishi, Y. (2026). The Three Hundredth Census: Whether Two Things Share a Root Is Decidable [Paper 300]. YOI: https://www.yamagishiyuuki.com/yoi/300; DOI: 10.5281/zenodo.22171914
  15. Yamagishi, Y. (2026). The Same Words “Critical Dimension” Name Two Different Numbers, Above and Below [Paper 482]. YOI: https://www.yamagishiyuuki.com/yoi/482; DOI: 10.5281/zenodo.22759625
  16. Yamagishi, Y. (2026). A Square Net and a Triangular Net Describe the World With the Same Exponent, and Only the Coefficient Differs [Paper 1033]. YOI: https://www.yamagishiyuuki.com/yoi/1033; DOI: 10.5281/zenodo.22860565
  17. Yamagishi, Y. (2026). The “2” of the Inverse-Square Density Is Two 2s, and the 2 of the Shell Mass Meets the 2 of the Horizon Balance Only in Three Dimensions [Paper 1092]. YOI: https://www.yamagishiyuuki.com/yoi/1092
  18. Yamagishi, Y. (2026). A Physical Infinity Is Set by How Fast Things Thin Out against the Dimension, Not by the Mass [Paper 1106]. YOI: https://www.yamagishiyuuki.com/yoi/1106
  19. Yamagishi, Y. (2026). When the Field Equation Is of Second Order, the Interior Can Be Read from Outside a Closed Skin Only as One Number per Angular Pattern [Paper 1107]. YOI: https://www.yamagishiyuuki.com/yoi/1107

Reference

Yamagishi, Y. (2026). The 2 in “Two Dimensions,” the Border for a Random Walk to Come Back, Is the Order of the Diffusion Equation, and the Lattice Changes Only the Value of the Return Probability [Paper 1120]. YOI: https://www.yamagishiyuuki.com/yoi/1120
ORCID: 0009-0006-8587-3450

PDF (English) · PDF (Japanese) · Date proof of the English PDF (OpenTimestamps .ots) · SHA256SUMS

The record page for this paper (in Japanese, with the PDFs in each language and the date proof): YOI 1120. This paper is a preprint and has not been peer reviewed.