Bethe-Peierls Approximation

Weiss mean-field theory replaces every neighbour of a spin by the same average value, discarding all correlation. It is astonishingly useful, but it commits one blatant crime: it predicts a finite ordering temperature even for a one-dimensional chain, where the exact transfer-matrix solution says there is no order at any positive temperature. The Bethe–Peierls approximation is the next honest step up. Instead of averaging a spin's whole environment, it keeps a small cluster — a central spin $\sigma_0$ together with its $q$ nearest neighbours — and treats the interactions inside that cluster exactly. The rest of the lattice is represented by an effective cavity field $H'$ acting on the boundary spins. Its value is determined self-consistently so that the boundary reproduces the average environment of the infinite lattice. That one field $H'$ is not free. It is pinned by a fairness demand: in a uniform crystal a central spin and a neighbour spin are physically the same kind of site, so the cluster must assign them the same average magnetization. This self-consistency condition, $\langle\sigma_0\rangle=\langle\sigma_j\rangle$ , closes the theory and determines both $H'$ and the magnetization. The payoff is a substantial improvement over Weiss mean-field theory while remaining analytically tractable. Because the local bonds are kept exact, the method feels the coordination number $q$ . A chain ( $q=2$ ) is too weakly connected to hold order, and Bethe–Peierls correctly returns $T_c=0$ there — the precise 1D result that Weiss theory butchers. For $q\ge3$ it gives a finite critical temperature that sits below the Weiss estimate and much closer to the truth: on the square lattice it overshoots the exact Onsager value by about a quarter, where Weiss overshoots by three-quarters. The construction becomes exact on the Bethe lattice — the loop-free interior of a Cayley tree — because there the effective-field replacement discards no correlation at all. What it still cannot fix is the universality class: the static uniform boundary field cannot represent critical fluctuations, so the exponents stay classical.

Key takeaways

Bethe–Peierls keeps a central spin and its $q$ nearest neighbours exact, replacing the rest of the lattice by an effective field $H'$ acting only on the neighbours.
$H'$ is fixed by self-consistency: the central spin and a neighbour must have equal magnetization, $\langle\sigma_0\rangle=\langle\sigma_j\rangle$ , so all sites are statistically equivalent.
Linearizing gives criticality at $\tanh(\beta_c J)=1/(q-1)$ , i.e. $k_BT_c=2J/\ln[q/(q-2)]$ — finite only for $q>2$ .
It crucially gives $T_c=0$ for the chain ( $q=2$ ), the EXACT result, where Weiss wrongly predicts $k_BT_c=2J$ ; on the square lattice it gives $2J/\ln2\approx2.885\,J$ versus Onsager's $2.269\,J$ .
It is exact on the Bethe lattice (loop-free Cayley tree) and is the ancestor of the modern cavity method and belief propagation.
Key idea: Treat one spin plus its $q$ neighbours exactly, replace the rest by a boundary field $H'$ , and fix $H'$ by demanding the central spin and a neighbour share the same magnetization — this alone yields $k_BT_c=2J/\ln[q/(q-2)]$ and repairs mean field's fatal 1D error.

The core insight

Treat one spin plus its $q$ nearest neighbours exactly; replace the remainder of the lattice by a self-consistent cavity field $H'$ acting on the neighbours, and determine $H'$ by requiring that boundary and interior sites have the same average magnetization in the infinite lattice. The result is $k_BT_c=2J/\ln[q/(q-2)]$ , which is finite only for $q>2$ and vanishes for the one-dimensional chain.

Where this lives today

The loop-free heart of Bethe's 1935 construction has become one of the most reused ideas in computational science. On a tree — or, more generally, on a locally tree-like sparse graph — the Bethe–Peierls approximation becomes exact. The self-consistent cavity fields become the messages of belief propagation, each summarizing the influence of an entire branch of the graph. This is the engine behind error-correcting codes (LDPC decoding), inference in graphical models, and the statistical-physics theory of random constraint-satisfaction problems, where the cavity method of Mézard, Parisi, and Zecchina predicts the satisfiability thresholds of random $k$ -SAT. In condensed matter the same equations describe spin glasses on Bethe lattices and dynamical mean-field theory's local-cluster impurity solvers. The through-line from a 1935 alloy paper to modern machine-learning inference is the single move that defines the method: treat local interactions exactly while representing the rest of the system through self-consistent cavity fields.

The physical picture

Stand on the central spin and ask what holds it in place. Its $q$ neighbours tug on it through real exchange bonds $J$ . But those neighbours are themselves anchored to the rest of the crystal. Rather than track every distant spin, Bethe–Peierls summarizes their collective grip on each neighbour by one number, the effective field $H'$ — a stand-in for the absent lattice. The central spin, whose ties to the world are the explicit bonds inside the cluster, feels only the true field $H$ . The crucial check is fairness. In a uniform crystal the central spin and any neighbour are identical sites, so whatever magnetization the cluster assigns the centre it must assign a neighbour. That single demand determines $H'$ . If instead you carelessly set $H'=qJm$ you would double-count the centre's own influence and slide right back down to ordinary mean-field theory.

Figure

The Bethe–Peierls cluster — one spin plus its q neighbours treated exactly, the rest replaced by the field H′

Why does loop-freeness matter so much? On a tree, removing the central spin cleanly disconnects its branches; each branch then acts on its neighbour independently, so its influence really is a single field and no information is lost. On a real lattice the neighbours remain connected through multiple closed loops outside the chosen cluster. Replacing the influence of those external correlations by a self-consistent cavity field is the essential approximation of the Bethe–Peierls method. The method nonetheless "knows" about coordination number: a spin with only two neighbours is too weakly held to sustain order at any finite $T$ , while a spin with many neighbours orders easily. This dependence on $q$ — invisible to the leading single-site picture — is exactly why Bethe–Peierls gets the dimensional trend right.

Figure

Critical temperature vs coordination number — Bethe–Peierls above the exact, below Weiss; both meet only as q→∞

The core equations

The cluster Hamiltonian keeps the central-to-neighbour bonds exact and gives the neighbours an extra field $H'$ that replaces their bonds to the outside. This is the whole content of the approximation, so we write it first.

H_{\text{cluster}}=-J\sigma_0\sum_{j=1}^{q}\sigma_j-H\sigma_0-(H+H')\sum_{j=1}^{q}\sigma_j \qquad (1)

where: $\sigma_0=\pm1$ is the central spin; $\sigma_j=\pm1$ are the $q$ nearest neighbours; $J>0$ is the ferromagnetic exchange; $H$ the true external field (in energy units); $H'$ the effective boundary field standing in for the neighbours' outside bonds; $q$ the coordination number. Physically: the centre feels only $H$ because its lattice ties are the explicit $J\sigma_0\sigma_j$ bonds, whereas each neighbour, having $q-1$ absent bonds, feels $H+H'$ . Fixing the central spin decouples the neighbour sums, so the partition function factorizes into a product of single-spin sums — one power of a hyperbolic cosine per neighbour.

Z=e^{\beta H}\big[2\cosh\beta(J+H+H')\big]^{q}+e^{-\beta H}\big[2\cosh\beta(-J+H+H')\big]^{q} \qquad (2)

where: $\beta=1/k_BT$ ; the first term is $\sigma_0=+1$ , the second $\sigma_0=-1$ . Writing $C_\pm=2\cosh\beta(\pm J+H+H')$ , this is $Z=e^{\beta H}C_+^{\,q}+e^{-\beta H}C_-^{\,q}$ . Each neighbour sees an effective field $\pm J+H+H'$ set by the sign of the central spin. The central-spin magnetization is a log-derivative of $Z$ with respect to the field on the centre; it is the ratio of the two weighted branches.

\langle\sigma_0\rangle=\frac{e^{\beta H}C_+^{\,q}-e^{-\beta H}C_-^{\,q}}{e^{\beta H}C_+^{\,q}+e^{-\beta H}C_-^{\,q}} \qquad (3)

where: $C_\pm=2\cosh\beta(\pm J+H+H')$ . Physically: at $H=0$ this is a competition between the $+$ and $-$ branches, tipped only by a spontaneous $H'\ne0$ . A single neighbour, one of whose bonds is explicit and $q-1$ of which are folded into $H'$ , has a magnetization with one hyperbolic sine replacing one cosine.

\langle\sigma_j\rangle=\frac{e^{\beta H}C_+^{\,q-1}S_+ + e^{-\beta H}C_-^{\,q-1}S_-}{e^{\beta H}C_+^{\,q}+e^{-\beta H}C_-^{\,q}} \qquad (4)

where: $S_\pm=2\sinh\beta(\pm J+H+H')$ ; it is the same for every neighbour by symmetry. Physically: the $S_\pm$ arise from differentiating with respect to the neighbour field. The self-consistency condition that closes the theory demands the two magnetizations be equal, because in a uniform lattice the centre and a neighbour are the same kind of site.

\langle\sigma_0\rangle=\langle\sigma_j\rangle \qquad (5)

where: the equality is enforced by translational invariance; it is one equation for the one unknown $H'$ . Physically: below the critical temperature, the self-consistency equation admits a nonzero cavity field $H'$ even when the external field $H=0$ . This spontaneous solution corresponds to spontaneous magnetization of the infinite lattice. Reducing that equality at $H=0$ collapses everything to a compact transcendental equation for $H'$ , whose bifurcation gives the critical temperature.

e^{2\beta H'}=\left[\frac{\cosh\beta(J+H')}{\cosh\beta(J-H')}\right]^{q-1},\qquad k_BT_c=\frac{2J}{\ln\!\big(q/(q-2)\big)} \qquad (6)

where: the left equation determines $H'$ ; the trivial root $H'=0$ loses stability at $T_c$ , where $\tanh(\beta_cJ)=1/(q-1)$ . Physically: $T_c$ is finite only for $q>2$ , and $q=2$ (the chain) sends $\ln[q/(q-2)]\to\infty$ so $T_c\to0$ .

Derivation AFactorized cluster partition function and the two magnetizations

The goal is to show that once the central spin is held fixed, the neighbour spins decouple, so the $(q{+}1)$ -spin cluster partition function collapses to a sum of two products of single-spin sums, and to read off the central and neighbour magnetizations from it. The strategy is to condition on $\sigma_0=\pm1$ : with the centre frozen, no two neighbours interact directly (the cluster is a star graph, a tiny tree), so each neighbour sum is an independent two-state sum. This same conditioning, differentiated with respect to the two distinct applied fields, delivers both magnetizations that the self-consistency condition will later equate.

Step 1. Write $Z=\sum_{\sigma_0=\pm1}e^{\beta H\sigma_0}\prod_{j=1}^{q}\sum_{\sigma_j=\pm1}e^{\beta(J\sigma_0+H+H')\sigma_j}$ . With $\sigma_0$ fixed, the exponent is linear in each $\sigma_j$ separately. Why: the only interaction each neighbour has inside the cluster is with the centre, which is now a fixed number, so the neighbour sums are independent.

Step 2. Each neighbour sum is $\sum_{\sigma_j=\pm1}e^{\beta(J\sigma_0+H+H')\sigma_j}=2\cosh\beta(J\sigma_0+H+H')$ . Substituting $\sigma_0=\pm1$ gives $C_\pm=2\cosh\beta(\pm J+H+H')$ and hence $Z=e^{\beta H}C_+^{\,q}+e^{-\beta H}C_-^{\,q}$ . Why: the standard identity $\sum_{s=\pm1}e^{as}=2\cosh a$ applies to every neighbour identically.

Step 3. The central magnetization is $\langle\sigma_0\rangle=\tfrac1\beta\,\partial\ln Z/\partial H_{\text{centre}}$ , differentiating only the explicit $e^{\pm\beta H}$ prefactors: this returns the ratio $(e^{\beta H}C_+^{\,q}-e^{-\beta H}C_-^{\,q})/Z$ . Why: the centre couples to the field solely through $e^{\beta H\sigma_0}$ , so its magnetization is the population difference of the two branches.

Step 4. The neighbour magnetization differentiates with respect to the neighbour field $H'$ (or $H$ on the neighbours), turning one $\cosh$ into a $\sinh$ : $\langle\sigma_j\rangle=(e^{\beta H}C_+^{\,q-1}S_+ + e^{-\beta H}C_-^{\,q-1}S_-)/Z$ with $S_\pm=2\sinh\beta(\pm J+H+H')$ . Why: each of the $q$ identical neighbours contributes, and $\partial\cosh/\partial(\text{field})\propto\sinh$ .

Go deeper · Derivation B · 1 minCritical temperature from linearizing the self-consistency equation

Derivation BCritical temperature from linearizing the self-consistency equation

The goal is to derive the Bethe–Peierls critical temperature $k_BT_c=2J/\ln[q/(q-2)]$ by finding the temperature at which the disordered solution $H'=0$ first becomes unstable to a small spontaneous field. The strategy is to set $H=0$ , reduce the equality $\langle\sigma_0\rangle=\langle\sigma_j\rangle$ to a single equation for $H'$ , take its logarithm, and expand to first order in $H'$ ; a nonzero solution appears exactly when the linear-response coefficient reaches unity, i.e. exactly where the disordered solution $H'=0$ becomes unstable. This is the standard bifurcation criterion that determines the critical temperature $T_c$ . The inversion uses the identity $\tanh^{-1}x=\tfrac12\ln[(1+x)/(1-x)]$ .

Step 1. At $H=0$ the equality $C_+^{\,q}-C_-^{\,q}=C_+^{\,q-1}S_+ + C_-^{\,q-1}S_-$ rearranges to $C_+^{\,q-1}(C_+-S_+)=C_-^{\,q-1}(C_-+S_-)$ . Using $\cosh\theta-\sinh\theta=e^{-\theta}$ and $\cosh\theta+\sinh\theta=e^{\theta}$ , the $e^{\mp\beta H'}$ cancel to give $e^{2\beta H'}=[\cosh\beta(J+H')/\cosh\beta(J-H')]^{q-1}$ . Why: the hyperbolic identities turn the boundary factors into pure exponentials, isolating $H'$ .

Step 2. Take logs: $2\beta H'=(q-1)[\ln\cosh\beta(J+H')-\ln\cosh\beta(J-H')]$ , using that $\cosh$ is even. This is exact; the transition is where a nontrivial root branches off $H'=0$ . Why: the log linearizes the product structure and exposes the antisymmetric combination in $H'$ .

Step 3. Expand the right side to first order in $H'$ . Since $\tfrac{d}{dx}\ln\cosh\beta(J+x)\big|_{x=0}=\beta\tanh(\beta J)$ , the antisymmetric combination doubles it: $\text{RHS}\approx(q-1)\,2\beta H'\tanh(\beta J)$ . Why: the leading response of $\ln\cosh$ to a small field shift is $\beta\tanh(\beta J)$ , the single-spin susceptibility.

Step 4. The linearized equation $2\beta H'=(q-1)2\beta H'\tanh(\beta J)$ has a nonzero solution only when $1=(q-1)\tanh(\beta_cJ)$ , i.e. $\tanh(\beta_cJ)=1/(q-1)$ . Inverting, $\beta_cJ=\tfrac12\ln\frac{q}{q-2}$ , so $k_BT_c=2J/\ln[q/(q-2)]$ . Why: below the critical slope only $H'=0$ is stable (disordered); above it a spontaneous branch appears (ordered). *As a consistency check: the linearized slope is $g'(0)=2\beta(q-1)\tanh(\beta J)$ ; for $q=4$ the criticality condition $\tanh(\beta_cJ)=1/(q-1)$ gives $\beta_cJ=\tfrac12\ln2$ , and the general inversion $\tanh^{-1}(1/(q-1))=\tfrac12\ln[q/(q-2)]$ confirms the closed-form result.*

Go deeper · Derivation C · 1 minLimits: the exact 1D chain and the Weiss recovery, and the classical exponent β=1/2

Derivation CLimits: the exact 1D chain and the Weiss recovery, and the classical exponent β=1/2

The goal is to extract the qualitative triumph and the residual limitation of the method by pushing $k_BT_c=2J/\ln[q/(q-2)]$ to its extremes and expanding the self-consistency equation one order beyond linear. The strategy is to evaluate $q=2$ (the chain), $q\to\infty$ (the Weiss/upper-critical-dimension limit), and then the cubic balance that governs how the magnetization vanishes at $T_c$ . The first limit tests whether the method fixes mean field's worst error; the second tests whether it degrades gracefully to mean field where mean field is exact; the third tests the universality class.

Step 1. Chain $q=2$ : $\ln[q/(q-2)]=\ln(2/0)\to\infty$ , so $k_BT_c\to0$ . Equivalently $\tanh(\beta_cJ)=1/(q-1)=1$ forces $\beta_c\to\infty$ . There is no finite- $T$ order on the chain — exactly the transfer-matrix result. Why: with only two neighbours a single domain wall costs finite energy $2J$ but gains $\sim\ln N$ entropy, so disorder always wins at $T>0$ .

Step 2. Large $q$ : expand $\ln[q/(q-2)]=\ln[1/(1-2/q)]=2/q+2/q^2+\cdots$ , so $k_BT_c=2J/(2/q+\cdots)\to qJ$ . This is precisely the Weiss result $k_BT_c^{\text{MF}}=qJ$ ; mean field becomes exact as $q\to\infty$ . Why: with many neighbours the fluctuation of the local field is suppressed by the law of large numbers, so the uniform-field replacement is exact.

Step 3. Cubic balance near $T_c$ : expanding the self-consistency equation to $O(H'^3)$ gives $[1-(q-1)\tanh(\beta J)]H'\propto-H'^3$ . The bracket vanishes linearly in $(T_c-T)$ , so $H'^2\propto(T_c-T)$ and $H'\propto(T_c-T)^{1/2}$ . Why: this is the same cubic-equation balance as Weiss/Landau theory — the confining nonlinearity is quartic in the free energy.

Step 4. Since $m=\langle\sigma_0\rangle\propto H'$ to leading order, $m\propto(T_c-T)^{1/2}$ , so $\beta=1/2$ for every $q\ge3$ . Although the Bethe–Peierls approximation predicts a much better critical temperature than Weiss theory, it still neglects long-wavelength critical fluctuations. Consequently the critical exponents remain those of mean-field theory, $\beta=\tfrac12$ . Why: the static uniform boundary field cannot represent the diverging correlation length, so the closure stays in the mean-field universality class. *As a consistency check: $q\to2^+\Rightarrow T_c\to0$ ; the series $\ln[q/(q-2)]\approx2/q+2/q^2$ gives $k_BT_c\to qJ$ as $q\to\infty$ ; and the cubic balance yields $H'\sim(T_c-T)^{1/2}$ , confirming $\beta=1/2$ .*

Limiting cases

Behaviour of the Bethe–Peierls critical temperature $k_BT_c=2J/\ln[q/(q-2)]$ as the coordination number $q$ varies (all in units of $J$ , with the Weiss value $qJ$ and Onsager's exact square-lattice value for comparison):

Case	Result	Meaning
One dimension ( $q=2$ )	$\ln(2/0)\to\infty\Rightarrow k_BT_c=0$	No finite- $T$ order on the chain — the EXACT result. Weiss fails here, predicting $k_BT_c=2J$ .
Square lattice ( $q=4$ )	$k_BT_c=2J/\ln2\approx2.885\,J$	Onsager exact is $2J/\ln(1+\sqrt2)\approx2.269\,J$ ; Weiss gives $4J$ . Bethe overshoots ${\sim}27\%$ , Weiss ${\sim}76\%$ .
Triangular / simple cubic ( $q=6$ )	$k_BT_c=2J/\ln(3/2)\approx4.933\,J$	Improves on Weiss $6J$ ; the more neighbours, the closer both approximations get.
Large coordination ( $q\to\infty$ )	$k_BT_c\to qJ$	Approaches the Weiss result from below; mean field becomes exact, the upper-critical-dimension limit.
Critical exponents (all $q$ )	$\beta=\tfrac12,\ \gamma=1,\ \delta=3,\ \alpha=0$	Classical (mean-field) critical exponents for all finite coordination numbers because the static self-consistent boundary field neglects long-wavelength critical fluctuations.

Worked examples

Example 1 — the Bethe–Peierls $T_c$ across lattices, against the exact benchmarks. Compute $k_BT_c/J$ for the chain, square, and simple-cubic lattices and compare to Weiss and Onsager.

Step 1. Chain, $q=2$ : $\ln[2/(2-2)]=\ln\infty=\infty$ , so $k_BT_c=2J/\infty=\mathbf{0}$ . Weiss gives $qJ=2J$ ; the exact answer is $0$ , so Bethe–Peierls is exact and Weiss is qualitatively wrong.

Step 2. Square lattice, $q=4$ : $k_BT_c=2J/\ln2=2J/0.6931=\mathbf{2.885\,J}$ . Onsager's exact value is $2J/\ln(1+\sqrt2)=2J/0.8814=2.269\,J$ ; Weiss gives $4J$ . The Bethe overshoot is $(2.885-2.269)/2.269=\mathbf{27.2\%}$ , versus the Weiss overshoot $(4-2.269)/2.269=\mathbf{76.3\%}$ .

Step 3. Simple cubic, $q=6$ : $k_BT_c=2J/\ln(3/2)=2J/0.4055=\mathbf{4.933\,J}$ , versus Weiss $6J$ and the numerically exact $\approx4.511\,J$ . Bethe–Peierls tracks the exact trend and improves monotonically with $q$ , exactly the behaviour expected of a controlled local-cluster closure.

Example 2 — why the chain has no order: the self-consistency slope. Show directly from the linearized condition that $q=2$ forbids a finite-temperature transition.

Step 1. The criticality condition is $\tanh(\beta_cJ)=1/(q-1)$ . For $q=2$ this reads $\tanh(\beta_cJ)=1$ .

Step 2. But $\tanh(x)<1$ strictly for every finite $x$ , and $\tanh(x)\to1$ only as $x\to\infty$ . So the equation $\tanh(\beta_cJ)=1$ has no finite solution: $\beta_c\to\infty$ , i.e. $T_c\to0$ .

Step 3. For $q=3$ , by contrast, $\tanh(\beta_cJ)=1/2$ gives $\beta_cJ=\tanh^{-1}(1/2)=\tfrac12\ln3=0.549$ , so $k_BT_c=J/0.549=\mathbf{1.820\,J}$ — a genuine finite transition. The chain sits exactly at the boundary $q=2$ where the required slope $1/(q-1)=1$ can never be met at any positive temperature.

Origins · 1 minHans Bethe and how this was discovered

Hans Bethe

1906–2005 · Hans Bethe introduced the cluster approximation that bears his name in 1935, in a paper on the statistical theory of superlattices (order–disorder transitions in alloys), where the Ising spin encodes which atomic species occupies a site. His construction — treat a central atom and its coordination shell exactly, embed them in a self-consistent effective field — was the first systematic improvement on the Bragg–Williams (mean-field) theory and correctly captured short-range order persisting above the ordering temperature. Rudolf Peierls extended and clarified the method shortly afterwards, which is why it is usually called the Bethe–Peierls approximation. Bethe went on to a vast career: the theory of the deuteron and nuclear forces in the 1930s, the calculation of stellar energy generation through the carbon–nitrogen cycle and the proton–proton chain (for which he received the 1967 Nobel Prize in Physics), leadership of the theoretical division at Los Alamos, and the first estimate of the Lamb shift. The loop-free version of his 1935 idea — exact on the Bethe lattice — is now the foundation of the cavity method and belief propagation used across statistical physics, computer science, and information theory.

History

1907

— Weiss introduces the molecular-field (mean-field) theory of ferromagnetism

1925

— Ising solves the one-dimensional model and finds no transition

1935

— Bethe formulates the cluster approximation for order–disorder in alloys

1936

— Peierls extends and clarifies the method (Bethe–Peierls approximation)

1944

— Onsager's exact square-lattice solution fixes the true 2D critical temperature

1951

— Kikuchi generalizes to the cluster-variation method (larger exact clusters)

2001

— Mézard–Parisi cast the cavity method as the modern descendant of Bethe–Peierls

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., §12.6 (Bethe approximation) · Kardar, Statistical Physics of Fields, Ch. 3 (mean field and beyond) · Reif, Fundamentals of Statistical and Thermal Physics, Ch. 10 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity · Landau & Lifshitz, Statistical Physics Part 1, §149 · H. A. Bethe, Proc. R. Soc. Lond. A 150, 552 (1935) · R. Peierls, Proc. Camb. Phil. Soc. 32, 477 (1936). Key results: the factorization by conditioning on $\sigma_0$ ; the linearization giving $\tanh(\beta_cJ)=1/(q-1)$ and its inversion $\beta_cJ=\tfrac12\ln[q/(q-2)]$ ; $q=2\Rightarrow T_c=0$ ; $q=4\Rightarrow2J/\ln2\approx2.885\,J$ against Onsager $2J/\ln(1+\sqrt2)\approx2.269\,J$ and Weiss $4J$ ; $q\to\infty\Rightarrow k_BT_c\to qJ$ ; and $\beta=1/2$ from the cubic balance. Overshoot fractions ( $27\%$ Bethe, $76\%$ Weiss on the square lattice) follow from these values.

Builds on: The Ising Model: Definition and Hamiltonian · Mean-Field (Weiss) Theory of the Ising Model · Order Parameter and Spontaneous Magnetization · One-Dimensional Ising Model and Transfer Matrix Method. Leads to: Onsager Solution of the Two-Dimensional Ising Model · Binary Alloy Order-Disorder Transition · Kramers-Wannier Duality · High- and Low-Temperature Series Expansions.

How is the effective field H′ different from the Weiss molecular field?

The Weiss field $qJm$ replaces all of a spin's neighbours by their average and acts on the spin itself, throwing away every correlation. The Bethe–Peierls field $H'$ acts only on the boundary (neighbour) spins, standing in for their bonds to the outside, while the central-to-neighbour bonds are kept exact. And $H'$ is not proportional to $m$ by fiat — it is fixed by the site-equivalence condition $\langle\sigma_0\rangle=\langle\sigma_j\rangle$ . This is what lets the method feel the coordination number $q$ and correctly kill the transition on the chain, which Weiss cannot do.

Why does the central spin feel H but each neighbour feels H + H′?

The central spin's coupling to the lattice is fully represented by its explicit bonds $J\sigma_0\sigma_j$ inside the cluster, so it needs no surrogate — only the real field $H$ . Each neighbour, however, has exactly one explicit bond (to the centre) but $q-1$ bonds to spins outside the cluster that are not modelled. Those omitted bonds are replaced by the effective field $H'$ , so a neighbour feels $H+H'$ in total. Adding $H'$ to the centre too would double-count its own influence and destroy the improvement.

Why is the approximation exact on the Bethe lattice but only approximate on a square lattice?

On a Bethe lattice (a Cayley tree) there are no closed loops, so removing the central spin genuinely disconnects its $q$ branches into independent pieces. Each branch then influences its neighbour through a single well-defined effective field, and summarizing it that way discards no information — the construction is exact. On a real lattice the neighbours are also bonded to each other through loops, and replacing those correlated bonds by one uniform field is a genuine approximation. This is why the method's error shrinks as the lattice becomes more tree-like (larger $q$ , higher dimension).

If it gets T_c much better than Weiss, why are the exponents still wrong?

Getting $T_c$ right is a question of where the transition sits; getting the exponents right is a question of how fluctuations behave at the transition. Bethe–Peierls captures short-range correlations, which shifts $T_c$ toward the true value, but its effective field is still static and spatially uniform. It cannot represent the correlation length diverging or the long-wavelength critical modes, so the order parameter still vanishes as $(T_c-T)^{1/2}$ and all exponents stay classical. Fixing the exponents requires the renormalization group, which treats fluctuations on every length scale.

Where does the number 1/(q−1) in the criticality condition come from?

It is the ratio of a neighbour's explicit "known" bond to its total connectivity. Each neighbour has $q$ bonds; one is the exact bond to the centre, and $q-1$ are folded into $H'$ . When you linearize the self-consistency equation, the response of the $q-1$ replaced bonds appears as the factor $(q-1)$ multiplying $\tanh(\beta J)$ , and criticality is $1=(q-1)\tanh(\beta_cJ)$ . So $\tanh(\beta_cJ)=1/(q-1)$ : the more neighbours a spin has, the smaller the coupling needed to order, and the higher $T_c$ . At $q=2$ the required $\tanh=1$ is unreachable at finite $T$ , which is why the chain never orders.

Is Bethe–Peierls related to anything used outside physics?

Very much so. On tree-like graphs the self-consistency equations become belief propagation, a message-passing algorithm that is exact on trees and a powerful approximation on sparse graphs. It underlies the decoding of low-density parity-check (LDPC) error-correcting codes, inference in probabilistic graphical models, and the cavity-method analysis of random satisfiability problems in computer science. The single idea — keep local structure exact, replace the far environment self-consistently — is one of the most transplanted concepts from statistical physics into information science.

Does the sign of J change the story?

For ferromagnetic $J>0$ the effective field aligns neighbours and the cluster orders into a uniform magnetized state below $T_c$ . For antiferromagnetic $J<0$ on a bipartite lattice the same machinery applies to the staggered (sublattice) magnetization, and one finds a Néel temperature of the same form. The subtlety is frustration: on non-bipartite lattices (like the triangular one) antiferromagnetic bonds cannot all be satisfied, and the simple single-cluster Bethe–Peierls closure must be extended. For the standard ferromagnetic and bipartite-antiferromagnetic cases treated here, only the interpretation of the order parameter changes, not the algebra.