08.10.13 · stat-mech / qft

Parisi-Sourlas dimensional reduction and random-field supersymmetry

shipped3 tiersLean: nonepending prereqs

Anchor (Master): Parisi & Sourlas, *Phys. Rev. Lett.* 43, 744 (1979) (the founding paper); Parisi & Sourlas, *Phys. Rev. Lett.* 44, 979 (1980) (branched polymers and lattice animals); Cardy, *Phys. Lett. B* 159, 415 (1985) (superfield formulation and nonperturbative breakdown); Imbrie, *Phys. Rev. Lett.* 53, 1747 (1984) (RFIM order at d=3, contradicting naive reduction); Tarjus & Tissier, *Phys. Rev. B* 78, 024203 (2008) (functional RG and the dimension where reduction fails)

Intuition Beginner

Take a magnet and pour static into it: at every point of space, fix a tiny frozen random push that nudges the local spin one way or the other. The pushes do not move; they are baked in, like impurities frozen into a crystal. This is a system in a random environment, and the random-field Ising model is its cleanest example. The question is how such a system orders and where its critical point sits, once you average over all the ways the frozen static could have been arranged.

Parisi and Sourlas found something startling in 1979. When you average over the frozen randomness in the right way, the resulting theory secretly carries an extra symmetry. The symmetry rotates the ordinary directions of space into two new "ghostly" directions that behave like dimensions but anticommute: swapping their order flips a sign. They are bookkeeping directions, not places you can walk to, yet the symmetry treats them on the same footing as real space.

The payoff is a clean slogan. Because the two ghostly directions act like extra dimensions with the opposite sign, they cancel two real ones. So a random-environment system in $d$ dimensions behaves, at its critical point, like a clean system in $d - 2$ dimensions. The static you poured in costs you exactly two dimensions of critical behaviour.

This is a remarkable trick when it works. It also fails in an instructive way at low dimensions, where the hidden symmetry breaks and the slogan stops holding. Knowing both halves — the trick and its breakdown — is the whole story of this unit.

Visual Beginner

A schematic with three panels. On the left, a square lattice of spins with a frozen random arrow at each site pointing up or down, representing the quenched random field. In the middle, the same space drawn with two extra short axes labelled $θ$ and $\overset{ˉ}{θ}$ , shaded to mark them as anticommuting "ghost" directions rather than real space. On the right, a smaller lattice in two fewer dimensions, with a curved arrow from the middle panel indicating that averaging over the random field plus the two ghost directions lands you on the clean system in $d - 2$ dimensions.

The picture captures the essence: real space plus two sign-flipping ghost directions, glued together by a hidden symmetry, behaves like real space with two dimensions removed.

Worked example Beginner

Watch the two ghost directions cancel two real ones in the simplest possible case: a single Gaussian integral.

Step 1. Start with an ordinary Gaussian integral over $d$ real coordinates $x = (x_{1}, \dots, x_{d})$ with a rotation-invariant weight that depends only on the radius $r^{2} = x_{1}^{2} + \dots + x_{d}^{2}$ . Call the integrand $f (r^{2})$ . Its integral over all of space gives some number that depends on $d$ through the surface area of spheres.

Step 2. Now attach two ghost coordinates $θ$ and $\overset{ˉ}{θ}$ . They are not ordinary numbers; they anticommute, and the rule for integrating them is special: the integral of $1$ is zero, and the integral of the product $θ \overset{ˉ}{θ}$ is one. This Berezin rule is the whole of ghost-direction calculus.

Step 3. Build the super-radius $R^{2} = r^{2} + 2 θ \overset{ˉ}{θ}$ , treating the ghost piece on the same footing as the real piece. Expand $f (R^{2}) = f (r^{2}) + 2 θ \overset{ˉ}{θ} f^{'} (r^{2})$ , because $(θ \overset{ˉ}{θ})^{2} = 0$ kills every higher term.

Step 4. Integrate over the two ghost directions first. The constant part $f (r^{2})$ integrates to zero by the Berezin rule; only the $2 θ \overset{ˉ}{θ} f^{'} (r^{2})$ term survives, leaving $2 f^{'} (r^{2})$ .

Step 5. Now integrate $2 f^{'} (r^{2})$ over the $d$ real coordinates. A short computation with sphere areas shows the answer equals the integral of $f$ over $d - 2$ real coordinates, up to a fixed factor of $- 2 π$ . The two ghost directions have removed two real ones.

What this tells us: the cancellation is an exact identity about super-rotation-invariant integrands, not a vague analogy. Parisi and Sourlas showed that the disorder-averaged random-field theory is built entirely out of such super-invariant integrands, so the same cancellation propagates to every correlation function and hence to the critical exponents.

Check your understanding Beginner

Exercise (easy, numeric).

Naive Parisi-Sourlas reduction maps the random-field Ising model to pure Ising two dimensions lower. The pure Ising model has its lower critical dimension at $d = 1$ (no order in one dimension). What lower critical dimension for the random-field Ising model does naive reduction therefore predict?

Hint

If the random-field system in $d$ acts like the pure system in $d - 2$ , then it loses order exactly when the pure system two dimensions down does.

Answer

$d = 3$ . Pure Ising loses order at $d = 1$ , so reduction predicts the random-field Ising model loses order at $d = 1 + 2 = 3$ . This prediction is famously wrong: Imbrie proved the three-dimensional random-field Ising model does order, so the true lower critical dimension is $d = 2$ , and the discrepancy is the signal that the supersymmetry breaks down.

Formal definition Intermediate+

Fix $d$ -dimensional Euclidean space and a real scalar field $ϕ : R^{d} \to R$ with bare action $S_{0} [ϕ] = \int d^{d} x [\frac{1}{2} (\nabla ϕ)^{2} + V (ϕ)]$ , where $V (ϕ) = \frac{1}{2} m^{2} ϕ^{2} + \frac{λ}{4 !} ϕ^{4}$ is the pure $ϕ^{4}$ potential. The random-field problem couples $ϕ$ linearly to a quenched Gaussian random field $h : R^{d} \to R$ : $$ S_h[\phi] = S_0[\phi] - \int d^dx, h(x),\phi(x), $$ with $h$ drawn from the centred Gaussian ensemble $\overline{h (x)} = 0$ , $\overline{h (x) h (y)} = Δ δ^{d} (x - y)$ 08.06.01. The overbar denotes the quenched (disorder) average. The word quenched means observables are computed at fixed $h$ and only then averaged over $h$ — in contrast to the annealed average, which would average $Z$ rather than $ln Z$ or correlation functions.

The stationary-point (mean-field / leading-order) equation for a configuration is the field equation $$ -\nabla^2\phi(x) + V'(\phi(x)) = h(x). $$ Because the source $h$ on the right is Gaussian, this is structurally a Langevin/Nicolai map 08.10.02, 08.10.11: it sends the field $ϕ$ to a Gaussian variable $h$ , with Jacobian $det (δ h / δ ϕ) = det (- \nabla^{2} + V^{''} (ϕ))$ . Disorder averaging of an observable $O [ϕ_{h}]$ over the Gaussian $h$ produces $$ \overline{O} = \int \mathcal{D}h; e^{-\frac{1}{2\Delta}\int h^2}, O[\phi_h] = \int \mathcal{D}\phi; \big|\det(-\nabla^2 + V'')\big|; e^{-\frac{1}{2\Delta}\int(-\nabla^2\phi + V')^2}, O[\phi]. $$

The supersymmetric (BRST) representation exponentiates the Jacobian determinant by a pair of anticommuting (Grassmann) ghost fields $ψ, \overset{ˉ}{ψ}$ and the Gaussian weight by an auxiliary field $ω$ . Collecting $ϕ, ω, ψ, \overset{ˉ}{ψ}$ into a single superfield on superspace $(x, θ, \overset{ˉ}{θ})$ , $$ \Phi(x, \theta, \bar\theta) = \phi(x) + \bar\theta,\psi(x) + \bar\psi(x),\theta + \theta\bar\theta,\omega(x), $$ where $θ, \overset{ˉ}{θ}$ are anticommuting coordinates, recasts the disorder-averaged generating functional as a single superspace action $$ \mathcal{S}[\Phi] = \int d^dx, d\theta, d\bar\theta;\Big[\tfrac12,\Phi,(-\Delta_{\mathrm{super}}),\Phi + W(\Phi)\Big], \qquad \Delta_{\mathrm{super}} = \nabla^2 + 2,\Delta^{-1}\partial_\theta\partial_{\bar\theta}. $$ The defining structural fact is that $S$ is invariant under the orthosymplectic group $OSp (d ∣2)$ of super-rotations mixing the $d$ commuting coordinates $x^{μ}$ with the two anticommuting coordinates $θ, \overset{ˉ}{θ}$ , with the super-metric assigning the odd directions the opposite signature.

Counterexamples to common slips

Quenched is not annealed. Averaging $Z [h]$ over $h$ (annealed) gives a different, easier, and physically wrong answer for a system with frozen disorder; the quenched average of correlation functions is what carries the supersymmetry. Conflating the two is the most common error in disorder problems.
The supersymmetry is a symmetry of the disorder-averaged theory, not of any single disorder realisation. At fixed $h$ there is no $OSp (d ∣2)$ ; it emerges only after the Gaussian integral over $h$ generates the ghost-superpartner structure.
Dimensional reduction is a statement about leading critical behaviour, not an exact equality of full partition functions. It holds order by order in perturbation theory and breaks nonperturbatively, so reading it as an identity of complete theories is the error that predicts the wrong lower critical dimension.

Key theorem with proof Intermediate+

Theorem (Parisi-Sourlas dimensional reduction; Parisi-Sourlas 1979 ^{[Parisi-Sourlas-1979]}, Zinn-Justin Ch. 27 ^{[Zinn-Justin]}). Let $G (x_{1}, \dots, x_{n})$ be a connected correlation function of the disorder-averaged random-field theory $S_{h}$ in $d$ dimensions, computed in perturbation theory in $λ$ . If the supersymmetry $OSp (d ∣2)$ is unbroken, then $G$ restricted to a common super-rotation-invariant configuration equals the corresponding correlation function $G^{pure}$ of the pure theory $S_{0}$ in $d - 2$ dimensions, to all orders in $λ$ : $$ G_{d}^{\mathrm{RF}}(x_1, \ldots, x_n) = G_{d-2}^{\mathrm{pure}}(x_1, \ldots, x_n). $$ In particular the critical exponents of the random-field theory in $d$ dimensions coincide with those of the pure theory in $d - 2$ .

Proof. The engine is a single integral identity. Let $f : R \to R$ be smooth with rapid decay and consider the super-integral of $f$ evaluated on the super-radius $R^{2} = x^{2} + 2 θ \overset{ˉ}{θ}$ over superspace $R^{d} \times {θ, \overset{ˉ}{θ}}$ : $$ I_d = \int d^dx, d\theta, d\bar\theta; f(x^2 + 2\theta\bar\theta). $$ Expand in the nilpotent variable: $f (x^{2} + 2 θ \overset{ˉ}{θ}) = f (x^{2}) + 2 θ \overset{ˉ}{θ} f^{'} (x^{2})$ , since $(θ \overset{ˉ}{θ})^{2} = 0$ . The Berezin rules $\int d θ d \overset{ˉ}{θ} (1) = 0$ and $\int d θ d \overset{ˉ}{θ} (θ \overset{ˉ}{θ}) = 1$ kill the first term and retain the second: $$ I_d = \int d^dx; 2 f'(x^2). $$ Pass to radial coordinates with sphere area $Ω_{d - 1} r^{d - 1}$ and integrate the derivative by parts in $r$ : $$ I_d = \Omega_{d-1}\int_0^\infty 2 f'(r^2), r^{d-1}, dr = \Omega_{d-1}\int_0^\infty \frac{d}{dr}!\big[f(r^2)\big], r^{d-2}, dr = -(d-2),\Omega_{d-1}\int_0^\infty f(r^2), r^{d-3}, dr. $$ Comparing $Ω_{d - 1} = 2 π^{d /2} /Γ (d /2)$ with $Ω_{d - 3} = 2 π^{(d - 2) /2} /Γ ((d - 2) /2)$ and using $Γ (d /2) = \frac{d - 2}{2} Γ ((d - 2) /2)$ gives $(d - 2) Ω_{d - 1} = 2 π Ω_{d - 3}$ , so $$ I_d = -2\pi \int d^{,d-2}x; f(x^2). $$ This is the Parisi-Sourlas reduction identity: an $OSp (d ∣2)$ -invariant integrand integrated over $d$ commuting and $2$ anticommuting coordinates equals $- 2 π$ times the same integrand over $d - 2$ commuting coordinates.

Now apply this diagram by diagram. Every Feynman amplitude of the super-symmetric theory is, by $OSp (d ∣2)$ invariance, a function of super-distances $(x_{i} - x_{j})^{2} + 2 (θ_{i} - θ_{j}) (\overset{ˉ}{θ}_{i} - \overset{ˉ}{θ}_{j})$ only. The super-propagator $⟨ ΦΦ ⟩$ inverts $- Δ_{super}$ , and at coincident Grassmann arguments reduces, by the identity above applied to each loop integration, to the ordinary propagator of the pure theory in $d - 2$ dimensions. Vertices carry no extra Grassmann structure beyond the superspace measure. Performing the Berezin integrations loop by loop converts each $d$ -dimensional super-loop into a $(d - 2)$ -dimensional ordinary loop with the same combinatorial weight. So every amplitude of the random-field theory in $d$ dimensions equals the corresponding amplitude of the pure theory in $d - 2$ dimensions, order by order in $λ$ . The critical exponents, being determined by these amplitudes at the Wilson-Fisher fixed point, coincide. $□$

Bridge. The reduction identity builds toward the entire modern theory of disordered critical phenomena, and it appears again in the branched-polymer correspondence proved below, where the same two-Grassmann cancellation maps lattice animals in $d$ to the Lee-Yang edge in $d - 2$ . The foundational reason the trick works is exactly the Berezin rule $\int d θ d \overset{ˉ}{θ} = \partial_{θ} \partial_{\overset{ˉ}{θ}}$ , which turns the two ghost directions into a second derivative that lowers the effective dimension by two; this is exactly the same superpotential/Nicolai structure of 08.10.11, now read on a field theory rather than on a single quantum-mechanical degree of freedom. Putting these together, the random field $h$ plays the role of the Langevin noise of 08.10.02, the field equation $- \nabla^{2} ϕ + V^{'} = h$ is its Nicolai map, and the Jacobian determinant exponentiated by ghosts is the central insight that creates the supersymmetry out of nothing more than a Gaussian source. The bridge is that dimensional reduction is not a coincidence of exponents but a representation-theoretic consequence of $OSp (d ∣2)$ , and the same representation theory generalises to every observable built from super-invariant integrands — which is why it must be checked, and is found to fail, nonperturbatively at low dimensions.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Verify the Parisi-Sourlas reduction identity $\int d^{d} x d θ d \overset{ˉ}{θ} f (x^{2} + 2 θ \overset{ˉ}{θ}) = - 2 π \int d^{d - 2} x f (x^{2})$ for the Gaussian choice $f (u) = e^{- u}$ , by computing both sides directly.

Hint

Left side: expand the exponential to first order in $θ \overset{ˉ}{θ}$ and use the Berezin rule, then do the $d$ -dimensional Gaussian. Right side: do the $(d - 2)$ -dimensional Gaussian.

Answer

Left side. Expand $e^{- (x^{2} + 2 θ \overset{ˉ}{θ})} = e^{- x^{2}} (1 - 2 θ \overset{ˉ}{θ})$ , since $(θ \overset{ˉ}{θ})^{2} = 0$ . The Berezin integral retains the $- 2 θ \overset{ˉ}{θ}$ term: $\int d θ d \overset{ˉ}{θ} e^{- x^{2}} (1 - 2 θ \overset{ˉ}{θ}) = - 2 e^{- x^{2}}$ . Then $$ \int d^dx,(-2),e^{-x^2} = -2,\pi^{d/2}. $$ Right side. The $(d - 2)$ -dimensional Gaussian is $\int d^{d - 2} x e^{- x^{2}} = π^{(d - 2) /2}$ , so $- 2 π \cdot π^{(d - 2) /2} = - 2 π^{d /2}$ . The two sides agree: $- 2 π^{d /2} = - 2 π^{d /2}$ . The factor of $- 2 π$ and the drop of two dimensions are exactly reproduced by the simplest Gaussian integrand. Rubric: full credit for both sides evaluated to $- 2 π^{d /2}$ .

Exercise 4 (medium, short-answer).

Explain why the disorder average must be performed on correlation functions at fixed $h$ (quenched) rather than on the partition function $Z [h]$ (annealed), and state which average produces the Parisi-Sourlas supersymmetry.

Hint

The random field is frozen on the timescale of thermal fluctuations. The free energy, not the partition function, is the self-averaging observable.

Answer

The random field $h (x)$ is frozen: it does not fluctuate on the timescale over which the spins equilibrate. Physical observables are therefore thermal averages computed at fixed $h$ , and only then averaged over the disorder ensemble — the quenched average $\overline{⟨ O ⟩_{h}}$ . The annealed average $\overline{Z [h]}$ would let $h$ fluctuate together with the spins, describing a different physical system (mobile rather than frozen disorder) and generically giving wrong critical behaviour. The supersymmetry arises specifically in the quenched calculation: averaging a quenched observable over the Gaussian $h$ produces the Jacobian determinant $det (- \nabla^{2} + V^{''})$ , whose exponentiation by anticommuting ghost fields supplies the superpartners of $ϕ$ and hence the $OSp (d ∣2)$ structure. The annealed average has no such determinant and no supersymmetry. Rubric: full credit for the frozen-disorder distinction plus the identification of the quenched average as the source of the SUSY.

Exercise 5 (medium, symbolic).

Write the superfield $Φ (x, θ, \overset{ˉ}{θ}) = ϕ + \overset{ˉ}{θ} ψ + \overset{ˉ}{ψ} θ + θ \overset{ˉ}{θ} ω$ and compute $\int d θ d \overset{ˉ}{θ} \frac{1}{2} Φ^{2}$ , expressing the result in terms of the component fields.

Hint

Square the superfield, keep only the $θ \overset{ˉ}{θ}$ component (the rest Berezin-integrates to zero), and remember the ghosts anticommute, so $ψ^{2} = 0$ and $\overset{ˉ}{ψ} ψ = - ψ \overset{ˉ}{ψ}$ .

Answer

Square the superfield and collect the $θ \overset{ˉ}{θ}$ coefficient. Cross terms: $ϕ$ with $θ \overset{ˉ}{θ} ω$ gives $ϕ ω θ \overset{ˉ}{θ}$ twice (from $Φ^{2}$ ), and the two ghost terms $\overset{ˉ}{θ} ψ$ and $\overset{ˉ}{ψ} θ$ multiply to give $\overset{ˉ}{θ} ψ \overset{ˉ}{ψ} θ = - ψ \overset{ˉ}{ψ} θ \overset{ˉ}{θ}$ (one sign from anticommuting $θ$ past $\overset{ˉ}{ψ}$ ), counted twice. So $$ \tfrac12\Phi^2\big|_{\theta\bar\theta} = \phi,\omega - \psi,\bar\psi. $$ The Berezin integral picks out exactly this component: $$ \int d\theta, d\bar\theta;\tfrac12\Phi^2 = \phi,\omega - \psi,\bar\psi. $$ The auxiliary field $ω$ pairs with $ϕ$ and the two ghosts pair with each other — the component structure that, with the kinetic super-Laplacian, gives the disorder-averaged action its bosonic ( $ϕ, ω$ ) and fermionic ( $ψ, \overset{ˉ}{ψ}$ ) content. Rubric: full credit for $ϕ ω - ψ \overset{ˉ}{ψ}$ with the sign from anticommutation justified.

Exercise 6 (hard, short-answer).

State the Imry-Ma argument for the lower critical dimension of the random-field Ising model and explain why its answer $d_{l c} = 2$ contradicts the naive Parisi-Sourlas prediction.

Hint

Balance the surface energy of a domain wall enclosing a region of linear size $L$ against the bulk energy the random field gains by flipping that region. Compare the $L$ -scaling of the two.

Answer

Imry-Ma 1975 ^[Imry-Ma]. Consider flipping a domain of linear size $L$ in $d$ dimensions. The cost is the domain-wall surface energy, scaling as $σ L^{d - 1}$ with $σ$ the wall tension. The gain is the alignment of the spins in the domain with the local random field; by the central limit theorem the net random-field energy over a region of volume $L^{d}$ scales as $- h_{rms} Δ L^{d /2}$ . Order is unstable when the gain dominates the cost for large $L$ , i.e. when $d /2 > d - 1$ , that is $d < 2$ . So discrete-symmetry random-field systems order for $d > 2$ and the lower critical dimension is $d_{l c} = 2$ .

Contradiction with reduction. Naive Parisi-Sourlas maps the random-field Ising model to pure Ising in $d - 2$ , which loses order at $d - 2 = 1$ , predicting $d_{l c} = 3$ . Imry-Ma (a rigorous energy-entropy argument, later confirmed by Imbrie's proof ^[Imbrie] of order at $d = 3$ ) gives $d_{l c} = 2$ . The two dimensions of disagreement are exactly the signal that the supersymmetry is spontaneously broken below some dimension and dimensional reduction fails there. Rubric: full credit for the $L^{d - 1}$ vs. $L^{d /2}$ balance giving $d_{l c} = 2$ and the explicit clash with the reduced prediction $d_{l c} = 3$ .

Advanced results Master

Theorem (the OSp $(d ∣2)$ superrotation symmetry; Parisi-Sourlas 1979 ^{[Parisi-Sourlas-1979]}). The disorder-averaged random-field action, written in superspace, is invariant under the orthosymplectic group $OSp (d ∣2)$ acting on the supercoordinate $(x^{μ}, θ, \overset{ˉ}{θ})$ by linear transformations preserving the super-quadratic form $x^{2} + 2 θ \overset{ˉ}{θ}$ . The $d$ ordinary rotations $SO (d)$ , the $Sp (2)$ rotations of the Grassmann pair, and the $2 d$ supercharges mixing $x^{μ}$ with $θ, \overset{ˉ}{θ}$ generate the full superalgebra.

The supercharges $Q^{μ}$ that rotate a commuting direction into an anticommuting one are the field-theoretic descendants of the Nicolai/SUSY-QM supercharges of 08.10.11: there the superpotential $W$ generated partner Hamiltonians, here the action gradient $- \nabla^{2} ϕ + V^{'}$ generates the partner structure that pairs $ϕ$ with the ghosts $ψ, \overset{ˉ}{ψ}$ . The invariance of the action under $Q^{μ}$ is the precise statement that the two ghost directions are interchangeable with two real ones, and it is this invariance that the reduction identity exploits.

Theorem (dimensional reduction for branched polymers; Parisi-Sourlas 1980/1981 ^{[Parisi-Sourlas-1981]}, Brydges-Imbrie 2003 ^{[Brydges-Imbrie]}). The configurational statistics of branched polymers (lattice animals) in $d$ dimensions are governed by the same critical exponents as the Lee-Yang edge singularity — equivalently a pure $i ϕ^{3}$ field theory — in $d - 2$ dimensions. In particular the branched-polymer radius-of-gyration exponent in $d$ equals the Lee-Yang correlation-length exponent in $d - 2$ .

This is a second, independent instance of Parisi-Sourlas reduction, and the one that was eventually placed on rigorous footing. Brydges and Imbrie 2003 ^{[Brydges-Imbrie]} proved, by a probabilistic forest-root expansion rather than supersymmetry heuristics, that the pressure of the branched-polymer gas in dimension $d + 2$ equals that of the repulsive hard-core gas in dimension $d$ . Their proof confirms the field-theoretic prediction exactly and shows that for this observable the reduction does not suffer the breakdown that plagues the random-field Ising case — the difference being that the branched-polymer supersymmetry is not spontaneously broken.

Theorem (nonperturbative breakdown of dimensional reduction; Cardy 1985 ^[Cardy], Tarjus-Tissier 2008 ^{[Tarjus-Tissier]}). For the random-field Ising model the perturbative dimensional reduction $d \to d - 2$ fails below a critical dimension $d_{DR} \approx 5.1$ . The failure is driven by nonperturbative configurations (instantons / multiple stationary points of the field equation, equivalently a cusp in the functional-RG fixed-point effective action) that spontaneously break the $OSp (d ∣2)$ supersymmetry. The true lower critical dimension is $d_{l c} = 2$ , not the reduced value $3$ .

Cardy 1985 ^[Cardy] identified the mechanism: the field equation $- \nabla^{2} ϕ + V^{'} (ϕ) = h$ can have several solutions for a given $h$ , and summing over them with the correct signed Jacobian — rather than a single dominant saddle — generates contributions invisible to perturbation theory that violate the supersymmetric Ward identities. Tarjus and Tissier's nonperturbative functional RG ^{[Tarjus-Tissier]} makes this quantitative: a cusp develops in the disorder-renormalised effective potential below $d_{DR} \approx 5.1$ , the field-theoretic analogue of replica-symmetry breaking, and dimensional reduction holds only above that dimension.

Synthesis. Parisi-Sourlas dimensional reduction puts together the foundational reason a disordered critical system can be controlled at all: a quenched Gaussian random field generates, through its Jacobian determinant, a pair of anticommuting superpartners for the order-parameter field, and the resulting $OSp (d ∣2)$ supersymmetry is exactly the symmetry whose two odd directions cancel two real ones in every super-invariant integral. The central insight is that this is the same Nicolai/superpotential structure of 08.10.11 and the same noise-as-Langevin structure of 08.10.02, now lifted from a single degree of freedom to a field theory, so that the random source $h$ is literally the Langevin noise and the field equation is its Nicolai map. The bridge is that dimensional reduction is not numerology about exponents but a representation-theoretic consequence of super-rotation invariance, which is exactly why it generalises cleanly to the branched-polymer / Lee-Yang correspondence — rigorously proved by Brydges-Imbrie.

Putting these together, the failure of the reduction for the random-field Ising model must come from somewhere outside perturbation theory. This is exactly what happens: the supersymmetry is spontaneously broken below $d \approx 5.1$ by multiple solutions of the field equation, a cusp in the functional-RG effective action that is dual to replica-symmetry breaking in the disorder average. The foundational reason the naive prediction $d_{l c} = 3$ is wrong, and the true value is the Imry-Ma value $d_{l c} = 2$ , is this nonperturbative breaking — and tracking exactly where supersymmetry survives (branched polymers, high $d$ ) and where it shatters (random-field Ising, low $d$ ) is the whole content of the modern theory of disordered critical phenomena.

Full proof set Master

Proposition (Berezin reduction identity). For smooth rapidly-decaying $f$ and integer $d \geq 3$ , $$ \int d^dx, d\theta, d\bar\theta; f(x^2 + 2\theta\bar\theta) = -2\pi \int d^{,d-2}x; f(x^2). $$

Proof. Expand $f (x^{2} + 2 θ \overset{ˉ}{θ}) = f (x^{2}) + 2 θ \overset{ˉ}{θ} f^{'} (x^{2})$ using $(θ \overset{ˉ}{θ})^{2} = 0$ . The Berezin integral $\int d θ d \overset{ˉ}{θ}$ annihilates the constant term and returns the coefficient of $θ \overset{ˉ}{θ}$ , giving $\int d^{d} x 2 f^{'} (x^{2})$ . In radial coordinates with $Ω_{d - 1} = 2 π^{d /2} /Γ (d /2)$ the surface area of the unit $(d - 1)$ -sphere, $$ \int d^dx, 2 f'(x^2) = \Omega_{d-1}\int_0^\infty 2 f'(r^2), r^{d-1}, dr. $$ Since $\frac{d}{d r} f (r^{2}) = 2 r f^{'} (r^{2})$ , the integrand is $\frac{d}{d r} [f (r^{2})] r^{d - 2}$ . Integrate by parts; the boundary terms vanish by decay of $f$ and $d \geq 3$ : $$ \Omega_{d-1}\int_0^\infty \frac{d}{dr}[f(r^2)], r^{d-2}, dr = -(d-2),\Omega_{d-1}\int_0^\infty f(r^2), r^{d-3}, dr. $$ The recursion $Γ (d /2) = \frac{d - 2}{2} Γ (\frac{d - 2}{2})$ gives $(d - 2) Ω_{d - 1} = (d - 2) \frac{2 π ^{d /2}}{Γ ( d /2 )} = \frac{2 π ^{d /2}}{Γ (( d - 2 ) /2 ) /2} \cdot \frac{1}{1} = 2 π \cdot \frac{2 π ^{(d - 2) /2}}{Γ (( d - 2 ) /2 )} = 2 π Ω_{d - 3}$ . Therefore $$ \int d^dx, 2 f'(x^2) = -2\pi,\Omega_{d-3}\int_0^\infty f(r^2), r^{d-3}, dr = -2\pi\int d^{,d-2}x; f(x^2). \qquad \square $$

Proposition (uniqueness of the surviving Grassmann monomial). In any polynomial in the anticommuting pair $θ, \overset{ˉ}{θ}$ over a commutative base, only the coefficient of $θ \overset{ˉ}{θ}$ survives Berezin integration, and the map $P \mapsto \int d θ d \overset{ˉ}{θ} P$ equals $\partial_{θ} \partial_{\overset{ˉ}{θ}} P$ .

Proof. A polynomial in $θ, \overset{ˉ}{θ}$ has the general form $P = a + b θ + c \overset{ˉ}{θ} + e θ \overset{ˉ}{θ}$ with commuting coefficients, because $θ^{2} = \overset{ˉ}{θ}^{2} = 0$ truncates the expansion. The Berezin rules $\int d θ d \overset{ˉ}{θ} (1) = \int d θ d \overset{ˉ}{θ} (θ) = \int d θ d \overset{ˉ}{θ} (\overset{ˉ}{θ}) = 0$ and $\int d θ d \overset{ˉ}{θ} (θ \overset{ˉ}{θ}) = 1$ give $\int d θ d \overset{ˉ}{θ} P = e$ . Differentiation: $\partial_{\overset{ˉ}{θ}} (e θ \overset{ˉ}{θ}) = - e θ$ followed by $\partial_{θ} (- e θ) = - e$ , while the convention $\int d θ d \overset{ˉ}{θ} = \partial_{θ} \partial_{\overset{ˉ}{θ}}$ with $\partial_{θ} \partial_{\overset{ˉ}{θ}} (θ \overset{ˉ}{θ}) = \partial_{θ} (θ) = 1$ (ordering the derivatives to match the measure) gives $e$ . Either ordering convention isolates the top monomial; the coefficient $e$ is the integral. $□$

Proposition (reduction identity fails for non-super-invariant integrands). If the integrand depends on $θ, \overset{ˉ}{θ}$ other than through the combination $x^{2} + 2 θ \overset{ˉ}{θ}$ , the reduction identity need not hold. Concretely, $\int d^{d} x d θ d \overset{ˉ}{θ} g (x^{2}) θ \overset{ˉ}{θ}$ equals $\int d^{d} x g (x^{2})$ , which is not $- 2 π \int d^{d - 2} x g$ in general.

Proof. By the previous proposition $\int d θ d \overset{ˉ}{θ} [g (x^{2}) θ \overset{ˉ}{θ}] = g (x^{2})$ , so the super-integral equals $\int d^{d} x g (x^{2})$ , the full $d$ -dimensional integral with no dimensional drop. The reduction identity required the $θ \overset{ˉ}{θ}$ dependence to be locked to the $x^{2}$ dependence through super-rotation invariance, so that the surviving term was a derivative $f^{'}$ rather than an independent function $g$ . When that link is absent, the integration-by-parts step that lowered the dimension has no derivative to act on, and the dimensional reduction does not occur. This is the algebraic shadow of the physical statement that dimensional reduction requires unbroken $OSp (d ∣2)$ invariance: a super-noninvariant operator, such as one generated by a second stationary point of the field equation, evades the reduction. $□$

Proposition (random-field upper critical dimension). The random-field $ϕ^{4}$ theory has upper critical dimension $d_{u} = 6$ , consistent with dimensional reduction from the pure $ϕ^{4}$ value $d_{u}^{pure} = 4$ .

Proof. Power-count the disorder-averaged theory. The random-field coupling $\overline{hh} = Δ δ^{d}$ makes the effective propagator at small momentum behave as $Δ/ k^{4}$ rather than $1/ k^{2}$ , because the random-field insertion contributes two extra inverse powers of momentum (a $δ$ -function disorder correlator integrated against two propagators). The quartic coupling $λ$ then has engineering dimension $[λ] = 6 - d$ in this modified power counting: marginal at $d = 6$ , irrelevant above, relevant below. So mean-field exponents hold for $d > 6$ and the Wilson-Fisher fixed point governs $d < 6$ , giving $d_{u} = 6$ . Dimensional reduction maps this to the pure $ϕ^{4}$ value $d_{u}^{pure} = 4$ via $6 = 4 + 2$ , and this prediction — unlike the lower-critical-dimension one — is correct, because the upper critical dimension is controlled by the same perturbative regime in which reduction holds. $□$

Connections Master

Supersymmetric quantum mechanics 08.10.11. The superpotential/supercharge structure of Hamiltonian SUSY-QM is the single-degree-of-freedom prototype of the random-field supersymmetry. The partner pairing generated there by a superpotential $W (x)$ is, in the field theory, the pairing of the order-parameter field $ϕ$ with the ghost superpartners $ψ, \overset{ˉ}{ψ}$ produced when the Gaussian random field is integrated out; the field equation $- \nabla^{2} ϕ + V^{'} = h$ is the field-theoretic Nicolai map whose Jacobian carries the supersymmetry.
Gaussian field theory and the quenched random source 08.06.01. The random field $h (x)$ with $\overline{h (x) h (y)} = Δ δ^{d} (x - y)$ is the Gaussian field of 08.06.01 used as frozen disorder rather than as a fluctuating order parameter. Its delta-correlated Gaussian statistics are exactly what make the disorder average produce a determinant rather than an intractable functional, and hence what make the supersymmetric representation available.
Fokker-Planck equation and equilibrium distribution 08.10.02. The structural identity between the random-field equation and a Langevin/Nicolai map is the bridge to stochastic quantisation: the random source $h$ plays the role of the Langevin noise, and the deterministic field equation relating $ϕ$ to $h$ is the same map that 08.10.02 uses to send a stochastic field to its Gaussian driving noise. Disorder averaging and noise averaging are the same Gaussian integral, which is why both generate the ghost-determinant supersymmetry.
Renormalization group and critical exponents (lateral). Dimensional reduction is ultimately a statement about Wilson-Fisher fixed points: the random-field exponents in $d$ equal the pure exponents in $d - 2$ at the corresponding fixed point. Its nonperturbative breakdown is detected by the functional renormalization group through a cusp in the disorder-renormalised effective action, the field-theoretic analogue of replica-symmetry breaking in disordered-system RG.
Replica method and disorder averaging (lateral). The supersymmetric (BRST) average and the replica average are two routes to the same quenched free energy. The supersymmetric route makes the $OSp (d ∣2)$ symmetry manifest and dimensional reduction transparent; the replica route makes replica-symmetry breaking manifest. The two descriptions meet at the breakdown of dimensional reduction, where supersymmetry breaking in one language is replica-symmetry breaking in the other.

Historical & philosophical context Master

The random-field problem was sharpened by Yoseph Imry and Shang-keng Ma in 1975 ^[Imry-Ma] (Phys. Rev. Lett. 35, 1399), whose domain-wall energy-balance argument fixed the lower critical dimension of random-field systems at $d = 2$ for discrete symmetry and $d = 4$ for continuous symmetry. The argument is elementary and robust, and it became the benchmark against which all later field-theoretic predictions were measured. Giorgio Parisi and Nicolas Sourlas's 1979 paper Random magnetic fields, supersymmetry, and negative dimensions ^{[Parisi-Sourlas-1979]} (Phys. Rev. Lett. 43, 744) made the startling discovery that the disorder-averaged theory carries a hidden supersymmetry mixing space with two anticommuting coordinates, and that this supersymmetry forces the leading critical behaviour in $d$ dimensions to equal that of the pure system in $d - 2$ — a result they phrased provocatively in terms of integration in "negative dimensions." Their 1980-1981 work ^{[Parisi-Sourlas-1981]} (Phys. Rev. Lett. 46, 871) extended the same mechanism to branched polymers and the Lee-Yang edge singularity, a second and ultimately cleaner instance of the reduction.

The trouble was immediate and famous. The reduction predicted that the random-field Ising model should lose order at $d = 3$ (mapping to pure Ising at $d = 1$ ), yet John Imbrie's 1984 rigorous proof ^[Imbrie] (Phys. Rev. Lett. 53, 1747) established that the three-dimensional random-field Ising model does order, in agreement with Imry-Ma's $d_{l c} = 2$ and in flat contradiction with dimensional reduction. John Cardy's 1985 superfield analysis ^[Cardy] (Phys. Lett. B 159, 415) located the resolution: the field equation admits multiple stationary points, and the nonperturbative sum over them spontaneously breaks the supersymmetry, invalidating the reduction below some dimension. Gilles Tarjus and Matthieu Tissier's nonperturbative functional-RG program ^{[Tarjus-Tissier]} (Phys. Rev. B 78, 024203, 2008) made this quantitative, placing the breakdown at $d \approx 5.1$ via a cusp in the renormalised effective action — the field-theoretic face of replica-symmetry breaking. The branched-polymer reduction, by contrast, was vindicated: David Brydges and John Imbrie's 2003 Annals of Mathematics proof ^{[Brydges-Imbrie]} established it rigorously without any supersymmetry heuristic, by a probabilistic forest-expansion identity. The episode is a model case in mathematical physics of a beautiful symmetry argument that is exactly right in its perturbative domain, spectacularly wrong outside it, and instructive precisely because of the gap.

Bibliography Master

@article{ParisiSourlas1979,
  author  = {Parisi, Giorgio and Sourlas, Nicolas},
  title   = {Random magnetic fields, supersymmetry, and negative dimensions},
  journal = {Physical Review Letters},
  volume  = {43},
  year    = {1979},
  pages   = {744--745}
}

@article{ParisiSourlas1981,
  author  = {Parisi, Giorgio and Sourlas, Nicolas},
  title   = {Critical behavior of branched polymers and the {L}ee-{Y}ang edge singularity},
  journal = {Physical Review Letters},
  volume  = {46},
  year    = {1981},
  pages   = {871--874}
}

@article{Cardy1985,
  author  = {Cardy, John L.},
  title   = {Nonperturbative effects in a scalar supersymmetric theory},
  journal = {Physics Letters B},
  volume  = {159},
  year    = {1985},
  pages   = {415--418}
}

@article{ImryMa1975,
  author  = {Imry, Yoseph and Ma, Shang-keng},
  title   = {Random-field instability of the ordered state of continuous symmetry},
  journal = {Physical Review Letters},
  volume  = {35},
  year    = {1975},
  pages   = {1399--1401}
}

@article{Imbrie1984,
  author  = {Imbrie, John Z.},
  title   = {Lower critical dimension of the random-field {I}sing model},
  journal = {Physical Review Letters},
  volume  = {53},
  year    = {1984},
  pages   = {1747--1750}
}

@article{TarjusTissier2008,
  author  = {Tarjus, Gilles and Tissier, Matthieu},
  title   = {Nonperturbative functional renormalization group for random-field models and beyond},
  journal = {Physical Review B},
  volume  = {78},
  year    = {2008},
  pages   = {024203}
}

@article{BrydgesImbrie2003,
  author  = {Brydges, David C. and Imbrie, John Z.},
  title   = {Branched polymers and dimensional reduction},
  journal = {Annals of Mathematics},
  volume  = {158},
  year    = {2003},
  pages   = {1019--1039}
}

@book{ZinnJustin2002,
  author    = {Zinn-Justin, Jean},
  title     = {Quantum Field Theory and Critical Phenomena},
  publisher = {Oxford University Press},
  edition   = {4th},
  year      = {2002}
}

@book{Cardy1996,
  author    = {Cardy, John L.},
  title     = {Scaling and Renormalization in Statistical Physics},
  publisher = {Cambridge University Press},
  year      = {1996}
}

@book{DeDominicisGiardina2006,
  author    = {De Dominicis, Cirano and Giardina, Irene},
  title     = {Random Fields and Spin Glasses: A Field Theory Approach},
  publisher = {Cambridge University Press},
  year      = {2006}
}

Prerequisites

08.10.11 pending
08.06.01 pending
08.10.02 pending

Tier anchors

beginner: Parisi, *Statistical Field Theory* (Addison-Wesley, 1988), Ch. 5 (disorder and the random-field picture); Nattermann, *Theory of the Random Field Ising Model* (in Young, ed., *Spin Glasses and Random Fields*, World Scientific, 1998), §1 (the Imry-Ma domain-wall argument as physical motivation); Cardy, *Scaling and Renormalization in Statistical Physics* (Cambridge, 1996), Ch. 8 (intuition for dimensional reduction)
intermediate: Cardy, *Scaling and Renormalization in Statistical Physics* (Cambridge, 1996), Ch. 8 (supersymmetric formulation, OSp superrotations, the reduction theorem); Zinn-Justin, *Quantum Field Theory and Critical Phenomena* (Oxford, 4th ed. 2002), Ch. 27 (random-field problem, BRST/SUSY method, dimensional reduction); De Dominicis & Giardina, *Random Fields and Spin Glasses* (Cambridge, 2006), Chs. 4-5 (replica vs. supersymmetric averaging)
master: Parisi & Sourlas, *Phys. Rev. Lett.* 43, 744 (1979) (the founding paper); Parisi & Sourlas, *Phys. Rev. Lett.* 44, 979 (1980) (branched polymers and lattice animals); Cardy, *Phys. Lett. B* 159, 415 (1985) (superfield formulation and nonperturbative breakdown); Imbrie, *Phys. Rev. Lett.* 53, 1747 (1984) (RFIM order at d=3, contradicting naive reduction); Tarjus & Tissier, *Phys. Rev. B* 78, 024203 (2008) (functional RG and the dimension where reduction fails)

References

Parisi, G. & Sourlas, N. — Random magnetic fields, supersymmetry, and negative dimensions · Phys. Rev. Lett. 43, 744 (1979) — the founding paper: a system in a quenched Gaussian random field in d dimensions, averaged over disorder, acquires a supersymmetry mixing the d coordinates with two anticommuting coordinates, so that its leading critical behaviour equals that of the pure system in d-2 dimensions (dimensional reduction), formalised via integration in 'negative dimensions'
Parisi, G. & Sourlas, N. — Critical behavior of branched polymers and the Lee-Yang edge singularity · Phys. Rev. Lett. 46, 871 (1981) — the branched-polymer / lattice-animal application: the configurational statistics of branched polymers in d dimensions reduce to the Lee-Yang edge singularity (a pure i*phi^3 field theory) in d-2 dimensions, a second instance of Parisi-Sourlas dimensional reduction proven rigorously much later by Brydges-Imbrie
Cardy, J. L. — Nonperturbative effects in a scalar supersymmetric theory · Phys. Lett. B 159, 415 (1985) — the superfield formulation of the Parisi-Sourlas theory and the identification of the nonperturbative (instanton) effects that spontaneously break the supersymmetry and so invalidate naive dimensional reduction at small d, explaining why the random-field Ising model orders at d=3 rather than d=5
Imry, Y. & Ma, S.-K. — Random-field instability of the ordered state of continuous symmetry · Phys. Rev. Lett. 35, 1399 (1975) — the Imry-Ma domain-wall energy-balance argument establishing the lower critical dimension of random-field systems (d=2 for discrete, d=4 for continuous symmetry), the physical benchmark against which the Parisi-Sourlas reduced prediction is tested and found to disagree
Imbrie, J. Z. — Lower critical dimension of the random-field Ising model · Phys. Rev. Lett. 53, 1747 (1984) — rigorous proof that the three-dimensional random-field Ising model has a phase transition (long-range order at low temperature and weak disorder), directly contradicting the naive dimensional-reduction prediction d_lc = 3 and establishing that the Parisi-Sourlas supersymmetry must break down
Tarjus, G. & Tissier, M. — Nonperturbative functional renormalization group for random-field models · Phys. Rev. B 78, 024203 (2008) — a nonperturbative functional-RG analysis locating the dimension d ~ 5.1 below which dimensional reduction and the underlying supersymmetry fail, driven by a cuspy fixed-point effective action (the analogue of replica-symmetry breaking in the field-theoretic language)
Zinn-Justin, J. — Quantum Field Theory and Critical Phenomena · Oxford University Press (4th ed., 2002), Ch. 27 — the random-field problem treated by the BRST/supersymmetric method: disorder averaging by Grassmann ghosts, the OSp(d|2) superrotation symmetry, and the perturbative proof of dimensional reduction order by order in the loop expansion
Brydges, D. C. & Imbrie, J. Z. — Branched polymers and dimensional reduction · Annals of Mathematics 158, 1019 (2003) — a rigorous probabilistic proof of Parisi-Sourlas dimensional reduction for branched polymers: the pressure of the branched-polymer gas in dimension d+2 equals that of the hard-core gas in dimension d, recovering the field-theoretic prediction without supersymmetry heuristics

Estimated time

beginner: 20m
intermediate: 50m
master: 85m