Random Surfaces from Stacking Cubes:
A Visual Journey
Leonid Petrov
University of Virginia
Loading...
How Does Nature Make Shape?
Salt — same substance with cubic crystal structure, at different scales and growth conditions
Mathematical (random) crystal growth
phantom
How can something random have a predictable shape?
Part I
Warm-up in Flatland
Paths on a Grid
Paths may be thought of as tracing a slice of the crystal by a plane parallel to the xy-plane.
Start at (0, 0), end at (a, b)
Move only right (R) or up (U)
Move only right (R) or up (U)
How many paths are there?
\(\displaystyle\binom{a+b}{a} = \frac{(a+b)!}{a!\, b!}\) paths
A Random Path
Random path in 10 × 6 grid
As the mesh goes to zero, the random path gets close to a straight line!
Zoomed in: local segment
Coin flipping:
At each step independently, flip a biased coin: R with prob. \(\frac{a}{a+b}\), U with prob. \(\frac{b}{a+b}\).
This describes the random boxed path at the local lattice scale.
Fluctuations Around the Diagonal
Central Limit Theorem:
The random path has fluctuations of size \(O(\sqrt{N})\) (where \(N=a+b\) is the number of steps).
The random path has fluctuations of size \(O(\sqrt{N})\) (where \(N=a+b\) is the number of steps).
At the midpoint:
\(\displaystyle \frac{Y_{\text{mid}} - \mathbb{E}[Y_{\text{mid}}]}{\sqrt{N}} \to \mathcal{N}(0, \sigma^2)\) The histogram of midpoint heights converges to a Gaussian bell curve.
\(\displaystyle \frac{Y_{\text{mid}} - \mathbb{E}[Y_{\text{mid}}]}{\sqrt{N}} \to \mathcal{N}(0, \sigma^2)\) The histogram of midpoint heights converges to a Gaussian bell curve.
Brownian bridge:
The full fluctuation process (path minus diagonal) converges to the Brownian bridge — the Brownian motion pinned at both ends.
The full fluctuation process (path minus diagonal) converges to the Brownian bridge — the Brownian motion pinned at both ends.
Universality:
The Brownian motion and Brownian bridge appear as scaling limits in most of probability theory, including random walks, random permutations, queuing theory, Kolmogorov–Smirnov statistics, random matrices, random graphs, financial models, and many other settings.
The Brownian motion and Brownian bridge appear as scaling limits in most of probability theory, including random walks, random permutations, queuing theory, Kolmogorov–Smirnov statistics, random matrices, random graphs, financial models, and many other settings.
Midpoint height samples: 0
Path minus diagonal (210 × 120)
Weighted Paths
Area above a path:
Count unit squares above the path
Count unit squares above the path
Area = 8
What if some paths are more likely?
Weight by area under the path:
Weight by area under the path:
\(\displaystyle\text{Probability} = \frac{q^{\text{area}}}{\sum_{\text{paths}} q^{\text{area}}}\)
Effect of q:
- \(q = 1\): all paths equally likely (uniform)
- \(q < 1\): paths with less area are favored
- \(q > 1\): paths with more area are favored
q-binomial coefficient:
\(\displaystyle \sum_{\text{paths}} q^{\text{area}} = \binom{a+b}{a}_q = \frac{(1-q) (1-q^2)\cdots(1-q^{a+b-1})}{(1-q)\cdots(1-q^a)(1-q)\cdots (1-q^b)}\)
1.00
q = 1: ordinary binomial (35 paths)
Curved Limit Shapes
Limit shape [Vershik 1996]
As \(N \to \infty\) with \(q = e^{-\gamma/N}\), the random path concentrates around a deterministic curve
\(\displaystyle A e^{-\gamma y} + B e^{-\gamma x} = 1\), where
\(\displaystyle A = \tfrac{1-e^{-\gamma}}{1-e^{-\gamma(1+\alpha)}}\), \(\displaystyle B = \tfrac{1-e^{-\gamma \alpha}}{1-e^{-\gamma(1+\alpha)}}\), \(\alpha\) = aspect ratio
As \(N \to \infty\) with \(q = e^{-\gamma/N}\), the random path concentrates around a deterministic curve
\(\displaystyle A e^{-\gamma y} + B e^{-\gamma x} = 1\), where
\(\displaystyle A = \tfrac{1-e^{-\gamma}}{1-e^{-\gamma(1+\alpha)}}\), \(\displaystyle B = \tfrac{1-e^{-\gamma \alpha}}{1-e^{-\gamma(1+\alpha)}}\), \(\alpha\) = aspect ratio
Entropy vs gravity:
- Entropy wants the straight diagonal (straight line has the most discrete paths approximating it)
- Gravity/energy: pulls the path up for \(q<1\) (less area above is preferred)
- The compromise is a curve: the limit shape
Variational principle:
The limit shape \(y=f(x)\) maximizes
integral entropy − \(\gamma\) · area, where \(\gamma\) is the normalized area bias, \(q=e^{-\gamma/N}\to1\).
The limit shape \(y=f(x)\) maximizes
integral entropy − \(\gamma\) · area, where \(\gamma\) is the normalized area bias, \(q=e^{-\gamma/N}\to1\).
\(\displaystyle \int \sigma(f'(x))\,dx - \gamma \int f(x)\,dx\),
\(\qquad \qquad \qquad \sigma(s)\coloneqq (s+1)\log(s+1) - s\log s\)
\(\qquad \qquad \qquad \sigma(s)\coloneqq (s+1)\log(s+1) - s\log s\)
3.6
(q = 0.990)
a = 240, b = 180 · Loading...
Fluctuations are still Brownian bridge:
Random curve minus limit shape
Summary of Flatland Exploration
Local rules —
paths go up or right
Boundary conditions —
paths live in a box
Limit shape —
from entropy/energy balance;
maximizes \(\displaystyle\int \sigma(f') \, dx - \gamma \int f \, dx\)
maximizes \(\displaystyle\int \sigma(f') \, dx - \gamma \int f \, dx\)
Local universality —
coin flips with \(p\) = slope of limit shape
Universal global Gaussian fluctuations —
Brownian bridge (time-changed) for our 2D situation
a=240, b=180, q=0.99 · Loading...
Local zoom — IID Bernoulli
Path − limit shape ≈ time-changed BB
3.6
(q = 0.990)
a = 180, b = 135 · Loading...
Zoomed in: local segment
Locally, paths don't know about the global weight — still IID Bernoulli!
Why?
The limit shape maximizes entropy minus γ × area.
This yields IID Bernoulli with position-dependent \(p\) = slope of the limit shape.
The weight only affects the global shape, not the local behavior.
-->
Part II
Random Surfaces
From Paths to 3D
surface in 1 × 12 × 9 box = 2D path in 12 × 9 rectangle
Theorem [MacMahon, c. 1900]:
Number of surfaces in \(a \times b \times c\) box is \(\displaystyle\prod_{i=1}^{a}\prod_{j=1}^{b}\prod_{k=1}^{c}\frac{i+j+k-1}{i+j+k-2}\)
Number of surfaces in \(a \times b \times c\) box is \(\displaystyle\prod_{i=1}^{a}\prod_{j=1}^{b}\prod_{k=1}^{c}\frac{i+j+k-1}{i+j+k-2}\)
For 9 × 12 × 9: 1,340,992,301,315,806,672,824,460,528,012,500
(~1034, about 1000× more than virus particles on Earth)
... and pick one at random from this collection
(how? — end of the talk)
More complex boxes:
Limit Shape in 3D
pre-sampled
Concentration:
[Cohn–Kenyon–Propp 2000]
As the mesh goes to zero, the random surface concentrates around a deterministic limit shape surface.
As the mesh goes to zero, the random surface concentrates around a deterministic limit shape surface.
Entropy vs boundary conditions leads to curving — even for uniformly random tilings, the limit shape is not flat!
Moreover, it contains flat facets which are parallel to the coordinate planes.
Moreover, it contains flat facets which are parallel to the coordinate planes.
Local Patches
Local patches:
- Approximate the surface by piecewise-planar patches
- Each patch is a flat random surface with a given slope \((s_1, s_2)\)
- For a given slope, the number of approximating 3D surfaces is \(\approx e^{N^2\cdot \sigma(s_1,s_2) }\)
- Not IID Bernoulli! Unlike 1D coin flips, the 2D patches are determinantal point processes — joint occupation probabilities have determinantal form
Variational principle [Cohn–Kenyon–Propp 2000]:
Limit shape \(z=h(x,y)\) maximizes \[\displaystyle\iint \sigma(\nabla h)\, dxdy\] subject to boundary conditions on \(h(x,y)\)
Limit shape \(z=h(x,y)\) maximizes \[\displaystyle\iint \sigma(\nabla h)\, dxdy\] subject to boundary conditions on \(h(x,y)\)
\(\sigma(s_1,s_2) = \tfrac{1}{\pi}\bigl[\Lambda(\pi s_1) + \Lambda(\pi s_2) + \Lambda(\pi (1\!-\!s_1\!-\!s_2))\bigr]\)
\(\Lambda(\theta)=-\!\int_0^\theta \ln|2\sin t|\,dt\) (Lobachevsky function)
Other 3D building blocks besides squares of three directions: [Kenyon–Okounkov–Sheffield 2006]
plane z=Ax+By
Lozenge Tilings
3 types of lozenges:
Rhombi of 3 orientations, projections of cube faces onto the \(x+y+z=0\) plane.
Rhombi of 3 orientations, projections of cube faces onto the \(x+y+z=0\) plane.
Frozen vs Liquid Regions
Frozen regions:
Near the edges, the tiling is forced — no randomness.
Near the edges, the tiling is forced — no randomness.
Liquid region:
In the middle, the tiling is random — all lozenge types appear.
In the middle, the tiling is random — all lozenge types appear.
Arctic boundary:
A sharp algebraic curve separates frozen from liquid (simply connected case). [Kenyon–Okounkov 2007]
A sharp algebraic curve separates frozen from liquid (simply connected case). [Kenyon–Okounkov 2007]
Hexagon
Sawtooth
General polygon
Ellipse example
\(\displaystyle \frac{x^2}{a^2} + \frac{y^2}{b^2} = 1\)
\(\displaystyle \frac{x^2}{a^2} + \frac{y^2}{b^2} = 1\)
Cardioid example
\((x^2\!+\!y^2)^2\!+\!4x(x^2\!+\!y^2)\!-\!4 y^2\!=\!0\)
\((x^2\!+\!y^2)^2\!+\!4x(x^2\!+\!y^2)\!-\!4 y^2\!=\!0\)
Height Fluctuations in 3D
Gaussian Free Field
Like the Brownian bridge, but with two-dimensional time. Take two independent random tilings \(h_1, h_2\). Each one:
Like the Brownian bridge, but with two-dimensional time. Take two independent random tilings \(h_1, h_2\). Each one:
\(\displaystyle h_1 = \underbrace{\bar{h}}_{\text{limit shape}} + \underbrace{\text{fluctuations}_1}_{??\text{ size}}\)
\(\displaystyle h_2 = \underbrace{\bar{h}}_{\text{limit shape}} + \underbrace{\text{fluctuations}_2}_{??\text{ size}}\)
Subtract: limit shapes cancel!
\(\displaystyle {h_1 - h_2} = {\text{fluct}_1 - \text{fluct}_2} \approx \text{Gaussian}\)
How big are the fluctuations?
- In 2D (paths): \(\sqrt{N}\)
- In 3D (surfaces): only \(\sqrt{\log N}\) — surprisingly flat!
GFF fluctuations
Governed by the Gaussian Free Field (GFF) — the analog of a Brownian bridge with 2D "time". Fluctuations grow to infinity at each point, but slowly — we can make sense of a random distribution (generalized function) in the sense of Schwartz.
Governed by the Gaussian Free Field (GFF) — the analog of a Brownian bridge with 2D "time". Fluctuations grow to infinity at each point, but slowly — we can make sense of a random distribution (generalized function) in the sense of Schwartz.
Universal: same GFF for all simply-connected domains.
[Kenyon 2001], [Petrov 2014], [Bufetov–Gorin 2018]
What Else Can We Tile by Lozenges?
Non-simply-connected
Non-planar orientable
Non-orientable (Möbius strip)
State of the art
[Borot–Gorin–Guionnet 2026]: Universal Gaussian fluctuations for some non-simply-connected, non-planar, or non-orientable domains.
Fluctuations for generic simply-connected domains: still open!
Fluctuations for generic simply-connected domains: still open!
q-volume tilt
q =
q =
q =
q =
q =
\(\displaystyle \text{Probability} = \frac{q^{\text{volume}}}{Z}\)
Same idea as 2D: weight each surface by \(q^{\text{volume}}\)
Entropy vs Energy:
- \(q < 1\): less volume preferred (gravity pulls down)
- \(q > 1\): more volume preferred (buoyancy pushes up)
New phenomena in weighted models
[Knizel–P. 2025]:
In an interpolation between \(q^{\text{volume}}\) and \(q^{-\text{volume}}\), keeping \(q\) fixed, one sees a collapse of 2D random behavior into a new 1D random interface, the waterfall.
In an interpolation between \(q^{\text{volume}}\) and \(q^{-\text{volume}}\), keeping \(q\) fixed, one sees a collapse of 2D random behavior into a new 1D random interface, the waterfall.
Waterfall eureka: Formulas teach us [Knizel–P. 2025]
Dimensional collapse:
Horizontal lozenges vanish → 2D surface becomes a 1D random sequence of two lozenge types
What does the 1D sequence look like?
Put the physics hat on: write a formal series for the density of lozenges:
\(\displaystyle \rho(t) = \sum_{n \in \frac{1}{2}\mathbb{Z}_{\geq 0}} \frac{(-q^{-n})\mathcal{F}_n(-\frac{t}{2})\mathcal{F}_n(1-\frac{t}{2})}{\|\mathcal{F}_n\|^2_{\ell^2(\mathbb{Z})}}\)
But... this series does not converge!
Diverges "mildly": \(\quad +c\;-c\;+c\;-c\;+c\;-\;\cdots\)
The divergence IS the answer!
\(\rho_0\) := lim (even partial sums) \(\approx 0.5611\)
\(\rho_1\) := lim (odd partial sums) \(\approx 0.4389\)
\(\rho_0+\rho_1 = 1\), both independent of \(t\)
The 1D sequence is two-periodic (!)
The 1D sequence is two-periodic (!)
Part III
How to Make Pictures
Glauber Dynamics
Elementary move:
Three lozenges around a hexagon can be flipped — rotated 180°. This is the only local move that preserves the tiling.
Three lozenges around a hexagon can be flipped — rotated 180°. This is the only local move that preserves the tiling.
Markov chain:
- Pick a random site with a \(1\times 1\times 1\) hexagon on it
- If it's flippable, flip it
- Repeat many times
How fast?
For the hexagon, Glauber dynamics mixes in \(O(N^{4+\epsilon})\) rotations [Laslier–Toninelli 2015], [Caputo–Martinelli–Toninelli 2012]. Mixing means that the two distributions become statistically indistinguishable.
For the hexagon, Glauber dynamics mixes in \(O(N^{4+\epsilon})\) rotations [Laslier–Toninelli 2015], [Caputo–Martinelli–Toninelli 2012]. Mixing means that the two distributions become statistically indistinguishable.
Exact Sampling: We need to know when to stop waiting
Coupling From The Past (CFTP)
[Propp–Wilson 1996]
Surfaces are partially ordered by height function. The Markov chain is monotone, so it suffices to track just the maximal and minimal chains.
- The stationary chain is somewhere in between the maximal and minimal surfaces
- Run coupled dynamics — when max and min coalesce, we've caught the stationary chain
- To ensure exactness, use the backward-doubling protocol: start chains at times \(-1, -2, -4, -8, \ldots\) and check coupling at time \(0\)
- Output is an exact sample from the stationary distribution
Not approximately random — EXACTLY random.
Compare with calculus: we know \(1/n \to 0\), but for which \(n\) is it equal to zero?
Here we reach the limit exactly — but at a random time.
Here we reach the limit exactly — but at a random time.
Hexagon 35 × 35 × 35
When Sampling is Hard
CFTP can be slow:
Exponentially slow in \(N\) — prohibitively long to get an exact sample: the monotone sandwich may take \(\sim e^{cN}\) steps to collapse.
This happens in the waterfall model [Knizel–P. 2025] and also in multi-\(q\)-weighted models.
Exponentially slow in \(N\) — prohibitively long to get an exact sample: the monotone sandwich may take \(\sim e^{cN}\) steps to collapse.
This happens in the waterfall model [Knizel–P. 2025] and also in multi-\(q\)-weighted models.
Why?
The CFTP needs to go through local moves, and some local configurations are exponentially unlikely. For example, in the waterfall, we do not have horizontal pieces in the middle — but to mix, horizontal pieces need to cross this gap.
The CFTP needs to go through local moves, and some local configurations are exponentially unlikely. For example, in the waterfall, we do not have horizontal pieces in the middle — but to mix, horizontal pieces need to cross this gap.
Alternatives come from algebra in specific cases:
- Commuting Markov matrices for waterfalls
[Borodin–Gorin–Rains 2009] - Powerful local equivalence relations that transform the lattice and weights in algebraically tractable manner (Yang–Baxter equation / urban renewal / spider move)
Hexagon 35 × 35 × 35
What Algebraic Structures Can Do for You
Commuting transfer matrices
[Borodin-Gorin 2008], [Borodin–Gorin–Rains 2009]
The algebraic structure of the weights (highly specific to the hexagon) gives commuting Markov operators that map the desired distribution to the desired distribution on a different hexagon.
The algebraic structure of the weights (highly specific to the hexagon) gives commuting Markov operators that map the desired distribution to the desired distribution on a different hexagon.
Growing the hexagon:
Start from a parallelogram — it has exactly one tiling, so it is already an exact sample.
Then apply the Markov operators one by one, changing the hexagon step by step (increasing one side, decreasing the other).
Each step preserves exactness!
Start from a parallelogram — it has exactly one tiling, so it is already an exact sample.
Then apply the Markov operators one by one, changing the hexagon step by step (increasing one side, decreasing the other).
Each step preserves exactness!
Polynomial time:
Each step costs \(\sim N^2\) (area), and we need \(\sim N\) steps — exact sample in \(O(N^3)\) time.
Each step costs \(\sim N^2\) (area), and we need \(\sim N\) steps — exact sample in \(O(N^3)\) time.
Compare: Glauber CFTP needed \(\sim e^{cN}\) steps on the previous slide.
Step: 0 / 80
Summary
Limit shapes
Random paths and surfaces concentrate around deterministic shapes as mesh shrinks (\(=\) size grows).
Random paths and surfaces concentrate around deterministic shapes as mesh shrinks (\(=\) size grows).
Universality
- Local statistics everywhere share the same nature, independent of boundary shape.
- Gaussian fluctuations on the global scale.
Exact sampling
Almost every picture in this talk was an exact random sample, computed live as we went along.
Almost every picture in this talk was an exact random sample, computed live as we went along.
How can something random have a predictable shape?
— Through a careful balance of entropy and energy: Local rules create global structures.
— Through a careful balance of entropy and energy: Local rules create global structures.
draw any border and lozenge tile! (plus 3D printing exports)
lpetrov.cc/lozenge-draw (opens in new tab)