High-Temperature Expansion of Lattice Models

You cannot solve the Ising model in three dimensions, and even in two dimensions Onsager's exact solution is a tour de force. But there is a humble question you can answer to any order you like: what happens just below infinite temperature? At $T=\infty$ the spins are independent coins and the partition function is trivially $2^N$ . Cool a little, and neighbours begin to correlate — bond by bond, in powers of the dimensionless coupling $\beta J$ , or better, in powers of $v=\tanh(\beta J)$ , which vanishes as $T\to\infty$ . This is the high-temperature expansion, and it turns thermodynamics into pure combinatorics: counting graphs on a lattice. The technical heart is one exact identity. For Ising spins $s_i=\pm1$ the bond weight $e^{\beta J s_is_j}$ takes only two values, so it can be written exactly as $\cosh(\beta J)(1+v\,s_is_j)$ — linear in $s_is_j$ , no truncation. Multiply this over every bond, sum over spins, and each term is labelled by a subgraph of the lattice with a factor $v$ per selected bond. The spin trace then performs a ruthless selection: any site touched by an odd number of bonds traces to zero. Only closed graphs — every vertex of even degree — survive. The partition function has become an integer-coefficient generating function counting closed lattice loops. That is the whole engine, and it pays off twice. Taking $\ln Z$ keeps only connected closed graphs (the linked-cluster theorem again), giving an extensive free energy and clean series for energy, specific heat, and — with a field — the susceptibility. And because a phase transition is a singularity of the generating function, the radius of convergence in $v$ pins the critical temperature. For the 2D square lattice, Kramers–Wannier self-duality fixes it exactly at $v_c=\sqrt2-1$ ; in three dimensions, long series plus ratio and Padé analysis still give some of the sharpest critical exponents we have.

Key takeaways

The high-temperature expansion develops the partition function of a lattice spin model in powers of $v=\tanh(\beta J)$ around the trivial $T\to\infty$ limit, where the spins are independent.
The exact Ising bond identity $e^{\beta J s_is_j}=\cosh(\beta J)\,(1+v\,s_is_j)$ converts each Boltzmann weight into a polynomial, turning $Z$ into a sum over subgraphs with one factor $v$ per bond.
The spin trace kills every graph containing an odd-degree vertex, so only closed (even-degree) subgraphs contribute — $Z$ becomes an integer-coefficient generating function for closed lattice loops.
Taking $\ln Z$ retains only connected closed graphs (the linked-cluster theorem), giving an extensive free energy and controlled series for $U$ , $C_V$ , and $\chi$ .
The nearest singularity in $v$ is the critical point: for the 2D square Ising model the series diverges at $v_c=\sqrt2-1$ , the exact Onsager value, fixed by Kramers–Wannier duality.
Ratio and Padé analysis of long series (Domb–Sykes) extract $v_c$ and the critical exponents, competitive with Monte Carlo in three dimensions.
Key idea: rewrite each Boltzmann bond as $\cosh(\beta J)(1+v\,s_is_j)$ with $v=\tanh\beta J$ ; the spin trace then selects only closed even-degree subgraphs, so $Z$ and $\ln Z$ become integer-coefficient power series in $v$ counting closed and connected-closed lattice loops.

The core insight

Rewrite each Boltzmann bond factor as $\cosh(\beta J)(1+v\,s_is_j)$ with $v=\tanh(\beta J)$ ; the spin trace then selects only closed (even-degree) subgraphs, so $Z$ and $\ln Z$ become integer-coefficient power series in $v$ counting closed and connected-closed lattice graphs respectively.

Where this lives today

High-temperature series are far from a historical curiosity. Modern "linked-cluster" and "numerical linked-cluster" expansions (NLCE) compute properties of quantum lattice models — Hubbard, Heisenberg, frustrated magnets — by summing over finite clusters, a direct descendant of the closed-graph counting shown here, and they are a standard tool for interpreting cold-atom and quantum-magnet experiments where quantum Monte Carlo suffers from the sign problem. For classical models, series of thirty or more terms analysed with differential approximants still give three-dimensional Ising exponents $\gamma\approx1.237$ , $\nu\approx0.630$ , $\eta\approx0.036$ at a precision that rivals the best Monte Carlo and conformal-bootstrap results. The deepest legacy is conceptual. The bond factorization that turns spins into a $\mathbb{Z}_2$ loop gas is the classical seed of lattice gauge theory: the strong-coupling (high-temperature) expansion of a gauge model, expanded in character sums instead of $\tanh\beta J$ , produces the area law for the Wilson loop and Wilson's 1974 demonstration of quark confinement at strong coupling. That a Taylor series around infinite temperature can both pin a critical point it cannot reach and, generalized, explain confinement, is a lasting testament to the power of counting graphs.

The physical picture

At infinite temperature the spins are random coins and the partition function is just the number of configurations — $2^N$ for $N$ Ising spins. Cooling a little correlates neighbours weakly, and the expansion asks, in order: what is the leading correction, then the next? Each correction is a picture of correlated spins — a subgraph of the lattice — and its weight is one factor of $v=\tanh(\beta J)$ per selected bond. Because $v\to0$ as $T\to\infty$ , larger graphs are more strongly suppressed, and truncating after a few orders is accurate at high temperature. The reason only closed graphs survive is a conservation law for "spin parity". Tracing a single spin gives $\sum_{s=\pm1}s^{\,d}=2$ if $d$ is even and $0$ if $d$ is odd. A graph term carries $s_i$ raised to the degree $d_i$ of site $i$ , so any site touched by an odd number of selected bonds annihilates the whole term. The survivors are loops and unions of loops — closed circuits drawn on the lattice. The first figure shows this selection rule at work: the smallest square plaquette (four bonds, every corner of degree two) survives, while an open path or a "dangling bond" has odd-degree endpoints and traces to zero.

Figure

Only closed even-degree subgraphs survive the spin trace: the square plaquette (all vertices degree 2) contributes at order v⁴, while open paths and dangling bonds have odd-degree ends and vanish

Think of $v=\tanh\beta J$ as a "fugacity per edge": the partition function is a grand sum over edge subsets with this fugacity, restricted to even-degree configurations — precisely a $\mathbb{Z}_2$ loop gas. This is why duality and topology enter so naturally. The transition is not something you build in; it emerges as the radius of convergence. The integer graph coefficients grow like $\mu^{\,n}$ , and $v_c=1/\mu$ is where loops proliferate uncontrollably — the onset of long-range order. The second figure is the payoff: a Domb–Sykes ratio plot, where the successive coefficient ratios $a_n/a_{n-1}$ of the susceptibility series are plotted against $1/n$ and extrapolated to a straight line whose intercept is $1/v_c$ . For the 2D square lattice that intercept lands exactly on $1+\sqrt2\approx2.414$ .

Figure

Domb–Sykes ratio plot: the susceptibility-series ratios aₙ/aₙ₋₁ plotted against 1/n extrapolate linearly to the intercept 1/v_c = 1+√2 ≈ 2.414, locating the 2D Ising critical point

The core equations

We start from the Ising model with a coupling $J$ and, later, a field $H$ ; the whole method rests on rewriting the nearest-neighbour Boltzmann weight and then tracing over spins.

Z=\sum_{\{s\}}\exp\!\Big(\beta J\sum_{\langle ij\rangle} s_i s_j+\beta H\sum_i s_i\Big) \qquad (1)

where: $s_i=\pm1$ is the Ising spin at site $i$ ; $\langle ij\rangle$ runs over nearest-neighbour bonds; $J>0$ is the ferromagnetic exchange; $\beta=1/k_BT$ ; and $H$ is the external field (set $H=0$ except when computing the susceptibility). At $\beta J\to0$ the sum collapses to $Z=2^N$ : the reference point of the whole expansion. The single identity that powers the method exploits the fact that $s_is_j$ takes only the values $\pm1$ , so the exponential is exactly linear in it — this is not a Taylor truncation but an equality.

e^{\beta J s_i s_j}=\cosh(\beta J)\,\big(1+ v\, s_i s_j\big),\qquad v\equiv\tanh(\beta J) \qquad (2)

where: $v=\tanh\beta J\in[0,1)$ for $0<T<\infty$ is the graph-expansion variable. Reading it physically: $\cosh\beta J$ is a spin-independent prefactor, and $v\,s_is_j$ is the only correlation a single bond can carry — one factor of $v$ per bond, with the sign set by whether the two spins agree. Multiplying the identity over all $N_b$ bonds and tracing over spins turns $Z$ into a purely combinatorial object: a sum over subgraphs, one factor $v$ per bond, with only closed graphs surviving.

Z=(\cosh\beta J)^{N_b}\,2^{N}\!\!\sum_{g\;\text{closed}}\! v^{\,b(g)} \qquad (3)

where: $N_b=\tfrac12 qN$ is the number of bonds ( $q$ the coordination number), $g$ ranges over subgraphs in which every vertex has even degree, and $b(g)$ counts the bonds in $g$ . The empty graph ( $b=0$ ) gives the leading $2^N(\cosh\beta J)^{N_b}$ ; the smallest loop is the first correction. Taking the logarithm keeps only the connected closed graphs — the linked-cluster theorem — so the free energy per site is extensive and finite in the thermodynamic limit.

-\beta f=\ln 2+\tfrac{1}{2}\,q\,\ln\cosh\beta J+\frac1N\!\!\sum_{g\;\text{conn., closed}}\!\! v^{\,b(g)} \qquad (4)

where: $f$ is the Helmholtz free energy per spin and $q$ the coordination number. The first two terms are the independent-spin reference; the graph sum is the interaction correction, and dividing by $N$ makes it intensive. The same connected-graph structure governs the Mayer cluster expansion for gases (recall aq-sm-mayer-cluster-expansion) and $\ln Z$ in field theory. With a small field one gets the correlation function and susceptibility from open two-point graphs — paths from site $0$ to site $r$ .

\chi=\beta\sum_{r}\langle s_0 s_r\rangle_c,\qquad \langle s_0 s_r\rangle=\frac{\sum_{g:\,0\to r} v^{\,b(g)}}{\sum_{g\;\text{closed}} v^{\,b(g)}} \qquad (5)

where: the numerator sums graphs with odd degree only at the two marked sites $0$ and $r$ (a path connecting them), and $\langle\,\rangle_c$ is the connected correlation. To leading order $\langle s_0 s_r\rangle\approx v^{\,\ell(0,r)}$ , with $\ell$ the shortest path length, so the reduced susceptibility becomes a path-counting series $\bar\chi=\sum_r\langle s_0 s_r\rangle=1+qv+\cdots$ . Finally, the payoff for two dimensions: the critical point, fixed exactly by Kramers–Wannier self-duality rather than by resumming the series.

v_c=\tanh(\beta_c J)=\sqrt2-1,\qquad \sinh(2\beta_c J)=1,\qquad \frac{k_BT_c}{J}=\frac{2}{\ln(1+\sqrt2)}\approx2.269 \qquad (6)

where: $v_c$ is the radius of convergence of the loop series and the exact Onsager critical value. This article owns the series route to this number and the estimate it provides; the full exact 2D solution is Onsager's, handed forward by name (aq-sm-kramers-wannier-duality). A "which variable when" note: never confuse $v=\tanh\beta J$ (graph variable, integer coefficients) with $\beta J$ (cumulant variable) — they agree only at first order.

Derivation AFrom Ising spins to a sum over closed graphs

The goal is to convert the Ising partition function, an intractable $2^N$ -term spin sum, into a generating function that counts even-degree subgraphs of the lattice — replacing statistical mechanics with graph combinatorics. The strategy is to apply the exact bond identity $(2)$ to every nearest-neighbour pair, expand the resulting product over bonds into a sum over subgraphs (each bond either "in" with weight $v\,s_is_j$ or "out" with weight $1$ ), and then perform the spin trace, which acts as a parity filter. The crucial point to watch is the trace rule $\sum_{s=\pm1}s^{\,d}=2\,[d\text{ even}]$ : it is what annihilates every graph with an odd-degree vertex, leaving only the closed loops. No approximation is made anywhere; the equality is exact for any $\beta$ .

Step 1. Factor each bond. Using $(2)$ on all $N_b$ bonds, $Z=\sum_{\{s\}}\prod_{\langle ij\rangle}\cosh\beta J\,(1+v\,s_is_j)$ . Why: the $\cosh\beta J$ factors are spin-independent and pull out; the identity is exact because $s_is_j=\pm1$ has only two possible values, both matched by the linear form.

Step 2. Pull out the prefactor. $Z=(\cosh\beta J)^{N_b}\sum_{\{s\}}\prod_{\langle ij\rangle}(1+v\,s_is_j)$ . Why: there are exactly $N_b$ bonds, so the product of $\cosh\beta J$ over bonds is $(\cosh\beta J)^{N_b}$ , independent of the spin configuration.

Step 3. Expand the product into graphs. $\prod_{\langle ij\rangle}(1+v\,s_is_j)=\sum_{g\subseteq E} v^{\,b(g)}\prod_{\langle ij\rangle\in g}s_is_j$ , where $E$ is the edge set. Why: expanding the product, each bond independently contributes either $1$ (omitted) or $v\,s_is_j$ (selected); a choice of which bonds to select is exactly a subgraph $g$ , carrying $v^{\,b(g)}$ times a product of spins.

Step 4. Trace over spins. For each site, $\sum_{s_i=\pm1}s_i^{\,d_i(g)}=2$ if the degree $d_i(g)$ is even and $0$ if it is odd. Why: $s_i$ appears in the graph term raised to the number of selected bonds meeting site $i$ ; summing $s_i=\pm1$ gives $2$ for even powers and $0$ for odd — an odd-degree site kills the whole term.

Step 5. Keep closed graphs only. The survivors have every $d_i$ even — closed subgraphs (loops and unions of loops) — each contributing $2^N$ from the $N$ site-traces, so $Z=(\cosh\beta J)^{N_b}2^N\sum_{g\,\text{closed}}v^{\,b(g)}$ . Why: only even-degree graphs pass the parity filter; the factor $2$ per site multiplies to $2^N$ , and the remaining $v$ -weighted sum runs over closed graphs alone. *Verified: setting $v=0$ gives $Z=2^N(\cosh\beta J)^{N_b}\to 2^N$ as $T\to\infty$ ; the leading nontrivial closed graph on the square lattice is the elementary plaquette ( $b=4$ ). The exact bond identity for $s_is_j=\pm1$ and the even-degree trace rule $\sum_{s}s^{\,k}=2[k\text{ even}]$ are checked symbolically and numerically in verify_high-temperature-expansion.py.*

Go deeper · Derivation B · 2 minLocating the 2D critical point by Kramers–Wannier duality

Derivation BLocating the 2D critical point by Kramers–Wannier duality

Here the goal is to fix the exact critical coupling of the square-lattice Ising model without solving the model, by recognizing that its high-temperature loop series and its low-temperature domain-wall series share identical combinatorics. The strategy is to write both series as sums over closed loops — the high- $T$ series in the variable $v=\tanh\beta J$ , the low- $T$ series in $e^{-2\beta^*J}$ — match them term by term, deduce the duality relation between the two couplings, and then impose that the (assumed unique) transition sits at the single self-dual fixed point. What to watch is the assumption of a single transition: duality alone gives a candidate point, and it is the physical input that there is exactly one that promotes it to the critical point.

Step 1. High-temperature loop sum. $Z_{\text{HT}}\propto\sum_{g\,\text{closed}}v^{\,b(g)}$ with $v=\tanh\beta J$ ; on the square lattice these closed loops are precisely the boundaries of spin-down domains drawn on the dual lattice. Why: from the graph representation $(3)$ , the interacting part of $Z$ is a sum over closed loops, and a closed loop on the lattice is a domain-wall configuration on the dual.

Step 2. Low-temperature domain-wall sum. $Z_{\text{LT}}\propto\sum_{\text{domain walls}}e^{-2\beta^* J\,\ell}$ , each wall of length $\ell$ costing $e^{-2\beta^*J\,\ell}$ ; on the square lattice these are also closed loops. Why: at low $T$ the excitations above the ordered state are flipped-spin droplets whose energy is proportional to boundary length, and each boundary is a closed loop of exactly the same kind counted at high $T$ .

Step 3. Match the two loop sums. Identical loop combinatorics force $v=\tanh\beta J\;\longleftrightarrow\;e^{-2\beta^*J}$ , the duality relation between the coupling $\beta$ and its dual $\beta^*$ . Why: two power series with the same integer loop coefficients are equal as functions once their per-bond weights are identified, which pins $\tanh\beta J=e^{-2\beta^*J}$ .

Step 4. Impose the self-dual condition. A single transition must sit at the fixed point $\beta=\beta^*$ , i.e. $\tanh\beta_cJ=e^{-2\beta_cJ}$ , which rearranges to $\sinh(2\beta_cJ)=1$ . Why: if the map $\beta\leftrightarrow\beta^*$ has one fixed point and there is one transition, the transition can only be at that self-dual coupling.

Step 5. Solve. $\sinh(2\beta_cJ)=1\Rightarrow 2\beta_cJ=\ln(1+\sqrt2)\Rightarrow v_c=\tanh(\beta_cJ)=\sqrt2-1$ and $k_BT_c/J=2/\ln(1+\sqrt2)\approx2.269$ . Why: $\operatorname{arcsinh}(1)=\ln(1+\sqrt2)$ , and back-substituting gives $v_c=\sqrt2-1\approx0.4142$ , the exact Onsager value — obtained purely from series duality. *Verified: $\sinh(2\beta_cJ)=1$ solves to $2\beta_cJ=\ln(1+\sqrt2)$ , $\beta_cJ\approx0.4407$ , $v_c=\tanh(0.4407)=0.41421=\sqrt2-1$ , and $k_BT_c/J=2.26919$ — all checked numerically in verify_high-temperature-expansion.py.*

Limiting cases

The behaviour of the expansion variable across the characteristic temperature regimes:

Case	Result	Meaning
Infinite temperature	$v=\tanh\beta J\to 0$ , $Z\to 2^N$	Only the empty graph survives; spins independent, $\ln Z=N\ln 2$ .
Small coupling	$v=\beta J-\tfrac13(\beta J)^3+\cdots$	Graph and cumulant variables coincide only at lowest order; conversions matter beyond first order.
Near criticality (2D)	$v\to v_c=\sqrt2-1\approx0.4142$	Series radius of convergence; $\chi\sim(v_c-v)^{-\gamma}$ with $\gamma=7/4$ exactly.
Zero temperature	$v=\tanh\beta J\to 1$	Far outside the convergence disk; the dual low- $T$ (domain-wall) series in $e^{-2\beta J}$ is the natural description.

Common trapThe high-temperature series does not converge past the critical point

A power series in $v$ has a finite radius of convergence set by the nearest singularity, and that singularity is the critical point at $v_c$ . So the series literally cannot be summed for $v>v_c$ (i.e. $T<T_c$ ) — the ordered phase is not reachable by term-by-term high- $T$ expansion. What the series can do is locate $v_c$ from the outside (as the radius of convergence, via ratio or Padé analysis) and encode the exponent $\gamma$ in the nature of the singularity. Two further traps: the two expansion variables $\beta J$ (cumulant form) and $v=\tanh\beta J$ (graph form) agree only to first order, $v\approx\beta J-\tfrac13(\beta J)^3$ , so mixing them corrupts coefficients; and for Ising spins only even-degree graphs survive — odd-degree partial graphs vanish identically under the trace, they do not merely cancel against others.

Worked examples

1 · The 2D square-lattice susceptibility series (confidence builder)

We build the first two nontrivial coefficients of the reduced susceptibility $\bar\chi=\sum_r\langle s_0 s_r\rangle$ for the square lattice by counting weighted paths, recovering the standard series $\bar\chi=1+4v+12v^2+\cdots$ .

Step 1. The reduced susceptibility is $\bar\chi=\sum_r\langle s_0 s_r\rangle$ with $\langle s_0 s_r\rangle\approx v^{\,\ell(0,r)}$ to leading order, $\ell$ the shortest path length. The $r=0$ term contributes the constant $\langle s_0^2\rangle=1$ , so $\bar\chi=1+\cdots$ .

Step 2. The order- $v$ term counts the sites reachable in one step: on the square lattice each site has $q=2d=4$ nearest neighbours, each contributing $v^1$ . So the coefficient of $v$ is $q=\mathbf{4}$ .

Step 3. The order- $v^2$ term counts length-2 weighted walks from the origin to any site $r\neq0$ . There are $4\times4=16$ ordered two-step walks; of these, $4$ return to the origin (step then reverse) and are excluded from $r\neq0$ , leaving $16-4=\mathbf{12}$ . Hence the coefficient of $v^2$ is $12$ .

Step 4. Collecting terms, $\bar\chi=\mathbf{1+4v+12v^2+\cdots}$ , exactly the leading Domb high-temperature series for the square-lattice Ising model. The first ratio $a_1/a_0=4$ is already the coordination number, and higher ratios climb toward $1/v_c=1+\sqrt2\approx2.414$ — the seed of the Domb–Sykes extrapolation.

2 · The exact 2D critical temperature from self-duality (exam level)

We extract the exact Onsager critical temperature of the square-lattice Ising model from the self-dual condition $\sinh(2\beta_c J)=1$ , and cross-check the loop-variable value $v_c=\sqrt2-1$ .

Step 1. Kramers–Wannier duality maps the high- $T$ loop series onto the low- $T$ domain-wall series, forcing the self-dual condition $\sinh(2\beta_c J)=1$ . Solve for the coupling: $2\beta_c J=\operatorname{arcsinh}(1)=\ln\!\big(1+\sqrt{1^2+1}\big)=\ln(1+\sqrt2)$ .

Step 2. Numerically $\ln(1+\sqrt2)=\ln(2.41421)=0.88137$ , so $\beta_c J=0.44069$ .

Step 3. The critical temperature is $k_BT_c/J=1/(\beta_c J)=1/0.44069=\mathbf{2.2692}$ , i.e. $k_BT_c/J=2/\ln(1+\sqrt2)\approx2.269$ — the exact Onsager value.

Step 4. Cross-check the loop variable: $v_c=\tanh(\beta_c J)=\tanh(0.44069)=\mathbf{0.41421}=\sqrt2-1$ , and its inverse $1/v_c=1+\sqrt2\approx2.414$ is exactly where the Domb–Sykes ratio plot extrapolates. A phase transition the high- $T$ series can never actually reach has been pinned exactly — from the series' own combinatorics.

Origins · 1 minSeries expansion and how this was discovered

Series expansion

method of expanding a partition function in a small parameter and analysing the series to locate a transition (Opechowski 1937; Domb and Sykes, 1950s–60s) · graph counting from infinite temperature down to criticality

The series-expansion approach to critical phenomena treats the partition function of a lattice model as a generating function in a small parameter — here $v=\tanh\beta J$ — whose coefficients are integers counting lattice graphs, and whose singularity structure encodes the phase transition. Its power lies in being systematic and exact order by order: unlike mean-field theory, no correlations are neglected, so the true critical behaviour is recovered if enough terms are computed. The method was pioneered by Opechowski in 1937 and turned into a precision instrument by Cyril Domb, Martin Sykes, and the King's College London series-expansion school through the 1950s and 60s, who computed high-temperature series to many orders and developed the ratio and Padé-approximant analyses (the Domb–Sykes plot) that extract critical points and exponents. For the two-dimensional Ising model the high-temperature loop series is mapped onto its own low-temperature form by Kramers–Wannier duality, fixing the critical point exactly at $v_c=\sqrt2-1$ ; in three dimensions, where no exact solution exists, long series remain among the most accurate sources of the critical exponents $\gamma$ , $\nu$ , and $\eta$ , competitive with Monte Carlo and the renormalization group. The same bond-factorization generalizes to $O(n)$ , Potts, and lattice gauge models, where the strong-coupling (high-temperature) expansion demonstrates confinement.

History

1937

— Opechowski introduces the high-temperature expansion for the Ising model.

1941

— Kramers and Wannier use self-duality to locate the exact 2D square-lattice critical point.

1944

— Onsager's exact solution confirms

k_BT_c/J=2/\ln(1+\sqrt2)

and

\gamma=7/4

1957

— Domb and Sykes develop systematic ratio and Padé analysis of long series.

1974

— Domb and Green begin the multivolume "Phase Transitions and Critical Phenomena", codifying series methods.

Going further

Texts. Pathria & Beale, Statistical Mechanics, 3rd ed., §13.4 (Ising model, high-temperature expansion). Kardar, Statistical Physics of Fields, Ch. 7 (low- and high-temperature series, duality). Reif, Fundamentals of Statistical and Thermal Physics, Ch. 10. Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 8. Domb & Green (eds.), Phase Transitions and Critical Phenomena, Vol. 3 (series-expansion methods). Landau & Lifshitz, Statistical Physics Part 1, §151.

Provenance. Base article: AllQFT track `sm`, "High-Temperature Expansion of Lattice Models" (STAT.INTER.HTEXP.1261); source Pathria & Beale §13.4 and Kardar SPoF Ch. 7. All symbolic and numeric claims re-verified here in verify_high-temperature-expansion.py: the exact bond identity $e^{\beta Js_is_j}=\cosh\beta J(1+v\,s_is_j)$ , the even-degree trace rule, the reduced-susceptibility series $\bar\chi=1+4v+12v^2+\cdots$ , the self-dual critical point $v_c=\sqrt2-1$ with $k_BT_c/J=2/\ln(1+\sqrt2)\approx2.269$ , and the leading energy $U/N=-\tfrac12qJv$ . No base physics error found. The exact 2D solution and duality are handed forward by name to aq-sm-kramers-wannier-duality (Onsager/Kramers–Wannier own the exact solution; this article owns the series route and the estimate).

Prerequisites. Configuration integral · Mayer cluster expansion · Linked-cluster theorem · The Mayer f-function.

Leads to. Kramers–Wannier duality · Critical Exponents and Series Analysis · Cluster integrals.

How can a Taylor series around infinite temperature find a phase transition it can never actually reach?

Because a phase transition is a singularity of the free energy as a function of $v=\tanh\beta J$ , and a power series "knows" about the nearest singularity through its radius of convergence. The coefficients $a_n$ of the susceptibility series grow asymptotically like $\mu^{\,n}$ , and $v_c=1/\mu$ is exactly the critical value — even though the series diverges there and cannot be summed beyond it. Analysing how the coefficients grow (their ratios and the subleading corrections) reveals both the location $v_c$ and the exponent $\gamma$ of the divergence. So you never step into the ordered phase; you infer the wall from the way the series accelerates toward it.

What is the difference between expanding in βJ and in v = tanh βJ?

They are two reorganizations of the same expansion. The $\beta J$ form is the cumulant (moment) expansion: you write $Z=Z_0\langle e^{\beta J\sum s_is_j}\rangle_0$ and expand in powers of $\beta J$ , keeping connected averages in the infinite-temperature ensemble. The $v$ form uses the exact bond identity and gives integer graph coefficients directly. They agree at first order, $v\approx\beta J$ , but differ beyond it because $v=\beta J-\tfrac13(\beta J)^3+\tfrac{2}{15}(\beta J)^5-\cdots$ . The $v$ variable is preferred for graph counting because its coefficients are clean integers counting lattice loops; the $\beta J$ variable is closer to the cumulant/linked-cluster language of aq-sm-cumulant-expansion.

Why does taking ln Z keep only connected closed graphs?

This is the linked-cluster theorem, the same one behind the Mayer cluster expansion for gases. The full closed-graph sum factorizes into independent placements of connected pieces: a general closed graph is a disjoint union of connected loops, and summing over how many copies of each connected loop appear reproduces the exponential of the connected-loop sum. Taking the logarithm therefore strips away the disconnected products and leaves exactly the connected closed graphs. Because a single connected loop lives at one place on the lattice (up to translation), the connected sum is proportional to $N$ — which is why $\ln Z$ , and hence the free energy, is extensive.

Why is v = tanh(βJ) rather than just βJ the natural small parameter?

Two reasons. First, it comes from an exact identity — the bond weight $e^{\beta J s_is_j}$ is genuinely $\cosh\beta J(1+v\,s_is_j)$ , so no information is lost, whereas an expansion in $\beta J$ truncates the exponential. Second, $v$ stays bounded, $v\in[0,1)$ , whereas $\beta J\to\infty$ as $T\to0$ ; the whole physical temperature range maps into a finite interval, which is exactly what a convergent power series needs. The critical point then sits at a finite $v_c<1$ , and duality relations (which swap high and low temperature) are simplest in $v$ .

Does the high-temperature expansion work for models other than Ising?

Yes, and this is one of its great strengths. The bond-factorization generalizes: for continuous $O(n)$ spins one uses a "character expansion" of $e^{\beta\,\mathbf s_i\cdot\mathbf s_j}$ into orthogonal functions, for $q$ -state Potts models a similar decomposition, and for lattice gauge theories a character expansion of the plaquette weight. In every case the trace imposes a parity/closure rule and $Z$ becomes a graph sum in a small coupling. The gauge case is historically decisive: the strong-coupling expansion produces the area law for the Wilson loop, which is Wilson's demonstration of confinement. The Ising version shown here is the simplest member of a very large family.

What exactly is a Domb–Sykes plot and how does it give v_c and γ?

It is a plot of the successive coefficient ratios $r_n=a_n/a_{n-1}$ of the susceptibility series against $1/n$ . For a power-law singularity $\chi\sim(v_c-v)^{-\gamma}$ , the ratios behave as $r_n\to\frac{1}{v_c}\big(1+\frac{\gamma-1}{n}+\cdots\big)$ , so they fall on an approximately straight line whose intercept (at $1/n\to0$ ) is $1/v_c$ and whose slope fixes $\gamma$ . For the 2D square Ising model the intercept extrapolates to $1/v_c=1+\sqrt2\approx2.414$ and the slope is consistent with the exact $\gamma=7/4$ . In three dimensions, where no exact answer exists, the same plot (with refinements like Neville tables and differential approximants) yields $\gamma\approx1.237$ .

Why do closed graphs correspond to domain walls in the duality argument?

On the square lattice, a closed loop drawn on the bonds of the original lattice is exactly the boundary of a region on the dual lattice — it separates an "inside" from an "outside". In the low-temperature picture, the excitations above the fully ordered ground state are droplets of flipped spins, and each droplet's energy cost is proportional to the length of its boundary, which is itself a closed loop. So the same closed loops are counted twice: once as high-temperature graph contributions weighted by $v^{\,\ell}$ , once as low-temperature domain walls weighted by $e^{-2\beta^*J\,\ell}$ . Matching the weights, $v=e^{-2\beta^*J}$ , is the Kramers–Wannier duality, and its fixed point is the critical coupling.