The Kardar-Parisi-Zhang equation and dynamic scaling

Intuition Beginner

Picture the edge of a sheet of paper that has been set alight. The burnt boundary is never a clean straight line: it bulges forward here, lags behind there, and grows rougher as the fire eats inward. Or picture sand poured slowly onto a table, building a pile whose surface is bumpy and keeps shifting. These are growing interfaces, and the surprise is that wildly different ones — burning fronts, piling sand, bacteria spreading on a dish, coffee soaking into a paper towel — all roughen in the same statistical way.

The height of the surface above each point goes up over time for two reasons. First, it spreads sideways and smooths out, the way a tall bump slumps into its neighbours. Second, and this is the new ingredient, growth happens along the local slope: a tilted patch gains material faster because it grows outward in the direction it faces, not just straight up.

On top of both effects sits randomness. Each little piece of the surface gets a noisy, unpredictable kick at every instant. The smoothing fights the noise, the slope-driven growth tilts the balance, and out of this tug-of-war comes a rough surface whose bumpiness follows exact and universal numbers.

Visual Beginner

The picture shows a jagged growing surface profile at three successive times. Early on the surface is nearly flat with small wiggles. Later the wiggles grow taller and wider, and the typical height of a bump spreads sideways as well as upward. A side panel shows the same width plotted against time on a log scale: it climbs as a straight line of a fixed slope, then flattens once the bumps span the whole sample.

The point of the image is that one set of growth rules produces a surface whose roughness grows by a power law in time and then levels off at a size-dependent ceiling. The two slopes in the plot are the universal exponents.

Worked example Beginner

Take a small interface measured by its width: the spread of its heights around the average. Start it perfectly flat, so the width is zero, and let it grow under noisy, slope-driven rules.

Early on, the width climbs with time as a power law, $w\sim t^{\beta }$. In one space dimension the measured exponent is $\beta =1/3$. So after four times as long, the width grows by a factor of $4^{1/3}\approx 1.6$, not by a factor of four — the roughening is steady but gentle.

Eventually the bumps grow as wide as the whole sample of length $L$, and the width stops climbing. It locks onto a ceiling that depends on size as $w\sim L^{\alpha }$, with $\alpha =1/2$ in one dimension. A sample twice as long saturates at a width larger by $\sqrt{2}$.

The crossover time between climbing and saturating scales as $L^z$ with $z=3/2$. What this shows: three exponents, $\alpha =1/2$, $\beta =1/3$, $z=3/2$, fully describe the roughening, and they satisfy $z=\alpha /\beta$ and $\alpha +z=2$. The same three numbers appear for sandpiles, burning fronts, and bacterial colonies.

Check your understanding Beginner

Formal definition Intermediate+

The Kardar-Parisi-Zhang (KPZ) equation governs the height field $h(\mathbf{x},t)$ of a $d$-dimensional interface growing into a $(d+1)$-dimensional space:

$$\frac{\partial h}{\partial t}=\nu {\nabla }^{2}h+\frac{\lambda }{2}(\nabla h{)}^2+\eta (\mathbf{x},t).$$

The three terms are the surface tension (diffusive smoothing with stiffness $\nu$), the lateral-growth nonlinearity $\frac{\lambda }{2}(\nabla h{)}^2$, and Gaussian white noise $\eta$ with

$$⟨\eta (\mathbf{x},t)⟩=0,⟨\eta (\mathbf{x},t)\eta ({\mathbf{x}}^{\prime },t^{\prime })⟩=2D{\delta }^d(\mathbf{x}-{\mathbf{x}}^{\prime })\delta (t-t^{\prime }).$$

Dropping the nonlinear term ($\lambda =0$) leaves the linear Edwards-Wilkinson equation, a Gaussian theory built on the same noise as 08.06.01 and 08.14.01. The nonlinearity is generated physically by growth normal to the surface: an interface advancing at speed $v$ along its local normal gains height $\sqrt{1+(\nabla h{)}^2}\approx 1+\frac{1}{2}(\nabla h{)}^2$ per unit time, and the $\frac{\lambda }{2}(\nabla h{)}^2$ piece is the leading correction ^{[Kardar-Parisi-Zhang 1986]}.

The roughening is quantified by the width $w(L,t)=⟨[h-\bar{h}{]}^2{⟩}^{1/2}$, which obeys the dynamic-scaling (Family-Vicsek) ansatz

$$w(L,t)\sim L^{\alpha }f\left(\frac{t}{L^z}\right),f(u)\sim \{\begin{array}{ll}{u}^{\beta },& u\ll 1,\\ \text{const},& u\gg 1,\end{array}$$

with roughness exponent $\alpha$, growth exponent $\beta$, and dynamic exponent $z=\alpha /\beta$. Matching the two regimes at the crossover forces $z=\alpha /\beta$, so two of the three exponents are independent.

Key theorem with proof Intermediate+

Theorem (exponent identities and the $1+1$d exact values). The KPZ equation possesses a Galilean invariance that fixes $\lambda$ under renormalisation and yields the exact Ward identity

$$\alpha +z=2.$$

In $d=1$ a second, fluctuation-dissipation relation pins the exponents to the exact values

$$\alpha =\frac{1}{2},\beta =\frac{1}{3},z=\frac{3}{2}.$$

Proof. Galilean invariance is the statement that the equation is unchanged under the tilt $h(\mathbf{x},t)\mapsto h(\mathbf{x}+\lambda \mathbf{v}t,t)+\mathbf{v}\cdot \mathbf{x}-\frac{\lambda }{2}{v}^{2}t$ for any constant slope $\mathbf{v}$. Substituting, the convective shift exactly cancels the change in the nonlinear term because the same $\lambda$ multiplies the tilt and the $(\nabla h{)}^2$ coupling. Hence $\lambda$ receives no independent renormalisation: the renormalised coupling $\lambda$ is fixed, which ties the scaling of space, time, and height together. Demanding that the rescaling $\mathbf{x}\to b\mathbf{x}$, $t\to b^{z}t$, $h\to b^{\alpha }h$ leave the nonlinear term's dimension equal to the time-derivative term's gives $\alpha +z=2$, the stated Ward identity.

In $d=1$ the stationary measure of the KPZ equation is known exactly: it coincides with the Edwards-Wilkinson Gaussian measure, in which the height increments perform a Brownian motion in space. That stationary measure has $⟨[h(\mathbf{x})-h({\mathbf{x}}^{\prime }){]}^2⟩\sim |\mathbf{x}-{\mathbf{x}}^{\prime }|$, so the roughness exponent is $\alpha =1/2$. Combining $\alpha =1/2$ with the Ward identity $\alpha +z=2$ gives $z=3/2$, and then $\beta =\alpha /z=(1/2)/(3/2)=1/3$. $\square$

Bridge. These exponents build toward the entire modern theory of the KPZ universality class, and the foundational reason they are exact rather than fitted is the pairing of a symmetry Ward identity with the known $1+1$d stationary measure. This is exactly the strategy that recurs whenever a nonlinear stochastic theory becomes solvable: a symmetry removes one coupling's flow, and an exact stationary law supplies the second relation. The Galilean identity $\alpha +z=2$ generalises the naive dimensional analysis of the linear Edwards-Wilkinson theory built on the Gaussian field of 08.06.01, and the central insight is that the same white noise of 08.14.01 drives both, with the nonlinearity merely reorganising its dressed exponents. Putting these together, the bridge is that one universal triple $(\alpha ,\beta ,z)$ governs surfaces as different as flames and bacterial colonies, and this pattern appears again in the Cole-Hopf map and the Tracy-Widom statistics of the Advanced section.

Exercises Intermediate+

Advanced results Master

The Cole-Hopf transformation $Z=\exp (\lambda h/2\nu )$ is the analytic backbone of the whole subject. It maps the nonlinear KPZ height onto the linear stochastic heat equation with multiplicative noise, ${\partial }_{t}Z=\nu {\nabla }^{2}Z+(\lambda /2\nu )\eta Z$, whose Feynman-Kac solution is the partition function of a directed polymer in a random potential ^{[Halpin-Healy-Zhang 1995]}. Reading the same object as a Langevin process with a response (conjugate) field places KPZ squarely inside the Martin-Siggia-Rose / Janssen-De Dominicis response-field formalism — the doubled-field generating functional that turns any Langevin stochastic-PDE into a Euclidean field theory, and the same machinery that underlies stochastic quantisation. Differentiating KPZ in space yields the noisy Burgers equation for the slope field, so interface growth, randomly stirred Burgers turbulence, and directed polymers are one universality problem viewed three ways.

The dynamic renormalisation group of Forster, Nelson, and Stephen runs the couplings $(\nu ,D,\lambda )$ under coarse-graining ^{[Forster-Nelson-Stephen 1977]}. The effective dimensionless coupling is $g\sim {\lambda }^{2}D/{\nu }^3$, and its one-loop flow has the structure $dg/d\ell =(2-d)g+c{g}^2$. For $d<2$ the linear term is positive: the nonlinearity is **relevant**, the Gaussian Edwards-Wilkinson fixed point is unstable, and the flow runs to a **strong-coupling fixed point** that controls the rough phase with the exact $d=1$ exponents above. For $d>2$ there is a roughening transition: weak coupling flows back to the smooth Edwards-Wilkinson fixed point, while strong coupling flows to the rough phase, with a perturbatively inaccessible strong-coupling fixed point separating them. The upper critical dimension of KPZ — whether the strong-coupling exponents ever reach mean-field values — remains an open question, with numerical and functional-RG evidence but no consensus closed form for $d\ge 2$.

The deepest modern development is exact solvability. For the $1+1$d KPZ equation with sharp-wedge initial data, the height fluctuations are not Gaussian: rescaled as $h(t)\simeq v_{\infty }t+(\Gamma t{)}^{1/3}\chi$, the random variable $\chi$ follows the Tracy-Widom distribution of the largest eigenvalue of a random matrix — the GUE Tracy-Widom law for curved (droplet) geometry and the GOE law for flat geometry ^{[Corwin 2012]}. This places KPZ growth in the same statistical universe as the edge of the spectrum of a random Hermitian matrix, the link being the directed polymer / last-passage percolation picture solved by the RSK correspondence and determinantal point processes. The scaling limit is conjectured to be governed by a single universal object, the KPZ fixed point, whose construction by the integrable-probability program (Matetski-Quastel-Remenik and others) is one of the landmark results of twenty-first-century mathematical physics.

Synthesis. Putting these together, KPZ is the foundational reason a single nonlinear stochastic PDE can organise a vast class of out-of-equilibrium phenomena: the central insight is that the lateral-growth term, the Galilean Ward identity, and the white noise of 08.14.01 combine to fix universal exponents that are dual to the largest-eigenvalue statistics of a random matrix. This is exactly the dynamic-scaling pattern that recurs across the universality class — ballistic deposition, the asymmetric simple exclusion process, the Eden model, directed polymers, burning and bacterial fronts — and the Cole-Hopf map generalises the linear Edwards-Wilkinson Gaussian theory of 08.06.01 into the solvable stochastic heat equation. The dynamic RG of 08.04.05 supplies the strong-coupling fixed point, while the Martin-Siggia-Rose response-field route is dual to the operator Fokker-Planck description of 08.10.02, and the bridge across all of these is one principle: the noisy growth of a rough surface, the free energy of a directed polymer, and the edge of a random spectrum are three readings of a single fixed point, and identifying that fixed point is the substance of the KPZ universality program.

Full proof set Master

Proposition (Cole-Hopf linearisation). Let $h$ solve the KPZ equation ${\partial }_{t}h=\nu {\nabla }^{2}h+\frac{\lambda }{2}(\nabla h{)}^2+\eta$ and set $Z=\exp (\lambda h/2\nu )$. Then, interpreting the products in the Stratonovich (smooth-noise) sense, $Z$ solves the multiplicative-noise stochastic heat equation ${\partial }_{t}Z=\nu {\nabla }^{2}Z+\frac{\lambda }{2\nu }\eta Z$.

Write $a=\lambda /2\nu$, so $Z=e^{ah}$ and $h=a^{-1}\ln Z$. Then ${\partial }_{t}Z=aZ{\partial }_{t}h$ and $\nabla Z=aZ\nabla h$, whence ${\nabla }^{2}Z=aZ{\nabla }^{2}h+a^{2}Z(\nabla h{)}^2$. Solve the last identity for $\nu {\nabla }^{2}h$:

$$\nu {\nabla }^{2}h=\frac{\nu }{aZ}{\nabla }^{2}Z-\nu a(\nabla h{)}^2.$$

Substitute into the KPZ right-hand side. The nonlinear term is $\frac{\lambda }{2}(\nabla h{)}^2=\nu a(\nabla h{)}^2$, which cancels the $-\nu a(\nabla h{)}^2$ produced above exactly because $\lambda /2=\nu a$. Hence

$${\partial }_{t}h=\frac{\nu }{aZ}{\nabla }^{2}Z+\eta .$$

Multiplying by $aZ=a{e}^{ah}$ and using ${\partial }_{t}Z=aZ{\partial }_{t}h$ gives ${\partial }_{t}Z=\nu {\nabla }^{2}Z+a\eta Z$, i.e. ${\partial }_{t}Z=\nu {\nabla }^{2}Z+\frac{\lambda }{2\nu }\eta Z$, the stochastic heat equation. The cancellation of the quadratic term is the entire content of the map; it works precisely because the same constant $\lambda$ sits in the nonlinearity and in the exponent. $\square$

Proposition (Galilean Ward identity $\alpha +z=2$). The KPZ equation is invariant under the infinitesimal tilt $h(\mathbf{x},t)\mapsto h(\mathbf{x}+\lambda \mathbf{v}t,t)+\mathbf{v}\cdot \mathbf{x}$, and this symmetry forces the scaling relation $\alpha +z=2$ at any fixed point with $\lambda \ne 0$.

Let $h^{\prime }(\mathbf{x},t)=h(\mathbf{x}+\lambda \mathbf{v}t,t)+\mathbf{v}\cdot \mathbf{x}-\frac{\lambda }{2}{v}^{2}t$. Compute the transformed derivatives: $\nabla h^{\prime }=\nabla h+\mathbf{v}$ and ${\partial }_t{h}^{\prime }={\partial }_{t}h+\lambda \mathbf{v}\cdot \nabla h-\frac{\lambda }{2}{v}^2$, with ${\nabla }^2{h}^{\prime }={\nabla }^{2}h$. Then the nonlinear term becomes $\frac{\lambda }{2}(\nabla h^{\prime }{)}^2=\frac{\lambda }{2}(\nabla h{)}^2+\lambda \mathbf{v}\cdot \nabla h+\frac{\lambda }{2}{v}^2$. Adding $\nu {\nabla }^2{h}^{\prime }+\eta$ and collecting, the $\lambda \mathbf{v}\cdot \nabla h$ and $\frac{\lambda }{2}{v}^2$ terms cancel exactly against those generated in ${\partial }_t{h}^{\prime }$, so $h^{\prime }$ satisfies the same equation. Because the symmetry transformation carries the single parameter $\lambda$ in both the tilt rate and the vertex, no renormalisation can shift $\lambda$ without breaking the invariance; thus the renormalised $\lambda$ is a fixed pure number. Under the anisotropic rescaling $\mathbf{x}\to b\mathbf{x}$, $t\to b^{z}t$, $h\to b^{\alpha }h$, the nonlinear term scales as $b^{2\alpha -2}$ relative to space while ${\partial }_{t}h$ scales as $b^{\alpha -z}$; requiring the dimensionless $\lambda$ to be scale-invariant equates these, $2\alpha -2=\alpha -z$, i.e. $\alpha +z=2$. $\square$

Connections Master

This unit builds directly on the Gaussian white noise and Wiener measure of 08.14.01: the noise $\eta$ driving the interface is the spacetime white noise whose single-path version is Brownian motion, and the linear (Edwards-Wilkinson) limit of KPZ is precisely a noise-driven Gaussian field of the kind in 08.06.01, so KPZ is the minimal nonlinear deformation of that Gaussian theory.

The Cole-Hopf map exhibits KPZ as a Langevin / stochastic-PDE problem, connecting it to the operator Fokker-Planck and equilibrium machinery of 08.10.02 and the lattice Langevin numerics of 08.10.08: the stochastic heat equation is a multiplicative-noise Langevin equation, and the same response-field (Martin-Siggia-Rose / Janssen-De Dominicis) functional that doubles the fields in stochastic field theory is the perturbative engine for KPZ, linking it to the stochastic-quantisation thread of chapter 08.10.

The dynamic renormalisation group of KPZ is the time-dependent counterpart of the momentum-shell Wilson RG of 08.04.05: coarse-graining flows the couplings $(\nu ,D,\lambda )$ to a strong-coupling fixed point below two dimensions, exactly as static critical phenomena flow to the Wilson-Fisher fixed point, and the roughness, growth, and dynamic exponents play the role of the static critical exponents. Laterally, the Tracy-Widom height statistics tie KPZ to random-matrix theory and to the directed-polymer and last-passage-percolation problems of integrable probability.

Historical & philosophical context Master

In 1986 Mehran Kardar, Giorgio Parisi, and Yi-Cheng Zhang published a three-page Letter proposing that the roughening of a growing interface is governed by the single nonlinear stochastic equation now bearing their initials, and they extracted the $1+1$d exponents $\alpha =1/2$, $\beta =1/3$ by a dynamic renormalisation-group argument resting on Galilean invariance ^{[Kardar-Parisi-Zhang 1986]}. The nonlinear term was not invented for the occasion: the noisy Burgers equation it reduces to had been studied by Forster, Nelson, and Stephen in 1977 as a model of a randomly stirred fluid, and the dynamic-RG technology was already in place ^{[Forster-Nelson-Stephen 1977]}. What KPZ supplied was the recognition that one equation unifies a sprawling experimental zoo — ballistic deposition, the Eden colony model, the asymmetric simple exclusion process, paper-burning fronts, fluid imbibition, and bacterial growth — into a single universality class, the out-of-equilibrium analogue of the Ising class for static criticality. The comprehensive 1995 review of Halpin-Healy and Zhang consolidated the field ^{[Halpin-Healy-Zhang 1995]}.

The philosophical turn came two decades later, when the exponent $\beta =1/3$ was sharpened into a full distributional law: the height fluctuations follow the Tracy-Widom distribution of random-matrix theory, and the conjectured scaling limit, the KPZ fixed point, was constructed by the integrable-probability program ^{[Corwin 2012]}. The lesson is that universality is layered — first the exponents, then the entire limiting distribution coincide across systems sharing only their symmetries and conservation laws, so that a burning sheet of paper and the edge of a random matrix spectrum obey one law.

Bibliography Master

@article{KardarParisiZhang1986,
  author  = {Kardar, Mehran and Parisi, Giorgio and Zhang, Yi-Cheng},
  title   = {Dynamic Scaling of Growing Interfaces},
  journal = {Physical Review Letters},
  volume  = {56},
  year    = {1986},
  pages   = {889--892}
}

@article{HalpinHealyZhang1995,
  author  = {Halpin-Healy, Timothy and Zhang, Yi-Cheng},
  title   = {Kinetic Roughening Phenomena, Stochastic Growth, Directed Polymers and All That},
  journal = {Physics Reports},
  volume  = {254},
  year    = {1995},
  pages   = {215--414}
}

@article{Corwin2012,
  author  = {Corwin, Ivan},
  title   = {The Kardar-Parisi-Zhang Equation and Universality Class},
  journal = {Random Matrices: Theory and Applications},
  volume  = {1},
  year    = {2012},
  pages   = {1130001}
}

@article{ForsterNelsonStephen1977,
  author  = {Forster, Dieter and Nelson, David R. and Stephen, Michael J.},
  title   = {Large-distance and long-time properties of a randomly stirred fluid},
  journal = {Physical Review A},
  volume  = {16},
  year    = {1977},
  pages   = {732--749}
}

@book{BarabasiStanley1995,
  author    = {Barab{\'a}si, Albert-L{\'a}szl{\'o} and Stanley, H. Eugene},
  title     = {Fractal Concepts in Surface Growth},
  publisher = {Cambridge University Press},
  year      = {1995}
}