High- and Low-Temperature Series Expansions

Long before Onsager solved the two-dimensional Ising model and decades before the renormalization group, physicists already had quantitative estimates of critical temperatures and critical exponents. Their tool was the humblest one imaginable: a power series, computed term by term and then cleverly extrapolated. The idea is to convert the intractable Ising partition function into a rigorous expansion in a small parameter — and there are two natural small parameters, one at each temperature extreme. At high temperature the spins are nearly independent, and the good variable is $v = \tanh\beta J$ , which vanishes as $T \to \infty$ . Each term of the expansion is a closed graph (an Eulerian loop) drawn on the lattice, weighted by $v$ raised to the number of bonds it uses. At low temperature the system sits near its perfectly aligned ground state, and the good variable is $u = e^{-2\beta J}$ , the Boltzmann cost of breaking a single bond, which vanishes as $T \to 0$ . Now each term is a flipped droplet, weighted by $u$ raised to its boundary length. Thermodynamics has become an exercise in graph enumeration — one you can do by hand at low orders and by computer to fifty orders and beyond. The two expansions are not independent. A high-temperature loop on the lattice is a low-temperature domain wall on the dual lattice, and in two dimensions Kramers–Wannier duality makes this correspondence exact, fixing the critical temperature at $\sinh 2\beta_c J = 1$ without summing a single series. And because the critical point is a genuine singularity of the free energy, the pattern of the series coefficients encodes the critical exponents: ratio tests and Padé approximants reconstruct the singularity from a finite Taylor germ. Even today, high-order series remain among the most accurate determinations of three-dimensional Ising exponents, standing shoulder to shoulder with Monte Carlo and the conformal bootstrap.

Key takeaways

High-temperature series expand the Ising partition function in $v = \tanh(\beta J)$ , generating a sum over closed loops on the lattice; each loop contributes $v^{(\text{bonds})}$ times a combinatorial multiplicity.
Low-temperature series expand in $u = e^{-2\beta J}$ about the ordered ground state; each term is a flipped droplet weighted by $u^{(\text{boundary length})}$ .
The two are dual: a high- $T$ loop on the lattice is a low- $T$ domain wall on the dual, and in 2D self-duality fixes $\sinh 2\beta_c J = 1$ , $k_B T_c/J = 2/\ln(1+\sqrt2) \approx 2.269$ .
On the 2D square lattice the leading low- $T$ correction is a single flip breaking $2d = 4$ bonds, giving $\ln Z/N = 2\beta J + u^4 + 2u^6 + \cdots$ .
Series of 20–50 terms, analyzed by ratio, Padé, or Dlog-Padé methods, yield critical temperatures and exponents (3D Ising $\gamma \approx 1.237$ , $\nu \approx 0.630$ ) — historically the first quantitative critical data.
A finite series cannot cross its own radius of convergence (set by $T_c$ ) by direct summation; reaching the singularity requires analytic continuation.
Key idea: reorganize the partition function as a sum over graphs — high- $T$ counts closed loops weighted by $\tanh\beta J$ , low- $T$ counts domain-wall boundaries weighted by $e^{-2\beta J}$ — and extract $T_c$ and exponents from the coefficient pattern.

The core insight

Reorganize the partition function as a sum over graphs: high- $T$ counts closed loops weighted by $\tanh\beta J$ , low- $T$ counts domain-wall boundaries weighted by $e^{-2\beta J}$ . Order-by-order rigor plus clever extrapolation pins down $T_c$ and the critical exponents.

Where this lives today

Series expansions remain a front-line quantitative tool precisely where exact solutions do not exist. For the three-dimensional Ising model, and for lattices (face-centered cubic, body-centered cubic) and models (antiferromagnets, Potts) with no closed-form solution, high-temperature susceptibility series of 20–50 terms analyzed by Dlog-Padé pin the critical temperature to several digits — the 3D simple-cubic estimate $k_B T_c/J \approx 4.5115$ matches Monte Carlo — and deliver the exponents $\gamma \approx 1.237$ , $\nu \approx 0.630$ , $\beta \approx 0.326$ , $\alpha \approx 0.110$ in agreement with the renormalization group, the conformal bootstrap, and large-scale simulation. That four-way agreement is one of the most convincing confirmations of universality we have. The same graph technology extends far beyond magnets. Potts and clock models, Heisenberg and XY spins (where the bond factor becomes a spherical-harmonic expansion), and lattice gauge theories all admit strong-coupling (= high-temperature) expansions in plaquettes. The strong-coupling expansion of lattice quantum chromodynamics — the direct field-theory descendant of the Ising high-temperature series — is how physicists first argued that quarks are confined, by showing the confining "loop" series has a finite radius of convergence. A 1941 bookkeeping trick for magnets became, seventy years on, a probe of the strong nuclear force.

The physical picture

At infinite temperature the spins are perfectly independent, the free energy is analytic, and the partition function is dominated by the empty graph — nothing is happening. Cool the magnet slightly and correlations switch on; the leading corrections are the smallest closed loops the lattice allows. Each loop is a coherent fluctuation of a ring of spins, and as $T$ falls the loops grow more numerous and longer, until at $T_c$ loops of every size proliferate and the series diverges. The proliferation of loops is the phase transition, seen from the hot side. From the cold side the same transition looks different but is the same event. Near zero temperature the ground state is frozen and the only excitations are rare flipped droplets, each bounded by a domain wall. The cheapest excitation flips a single spin, breaking its $2d$ neighboring bonds, so the leading correction to every quantity scales as $u^{2d} = e^{-4d\beta J}$ — for the 2D square lattice, $u^4$ . The transition arrives when droplets grow without bound and the entropy of their wiggling boundaries overwhelms the energy cost. These are two descriptions of one physics, stitched together by duality: a hot loop on the lattice is a cold wall on the dual.

Figure

Ratio-method estimates of k_B T_c/J climb toward the exact value as the series order grows

The figure is the whole method in one glance. Each blue dot is the critical-temperature estimate you get from truncating the susceptibility series at order $n$ and applying the ratio test; the red dashed line is the exact answer $k_B T_c/J = 2.2692$ . A finite series cannot reach the singularity by direct summation — you can never cross the radius of convergence — but the trend of the estimates converges to it. The power of the technique is that it works even where no exact answer exists: for the three-dimensional Ising model, this same climb pins $k_B T_c/J$ and the exponents to several digits.

Figure

The two expansions: high-T counts closed loops (weight v), low-T counts flipped droplets (weight u)

A finite series is a Taylor germ, and the critical behavior is its analytic continuation. Ratios $a_n/a_{n-1}$ tend to $1/v_c$ with a slope set by the exponent; Dlog-Padé approximants, which fit the logarithmic derivative $\frac{d}{dv}\ln\chi$ , place a pole exactly at $v_c$ whose residue is $-\gamma$ . The series does not merely approximate the singularity — it encodes it, and the analyst's job is to decode it. And by the duality drawn above, the high- $T$ plaquette loop $v^4$ and the low- $T$ single flip $u^4$ are literally the same four-bond object seen on the two lattices — the reason the two series are mirror images.

The core equations

Everything begins with the same algebraic identity that powers duality: because $s_i s_j = \pm 1$ , a bond's Boltzmann weight is not really an exponential but a two-term polynomial.

e^{\beta J s_i s_j}=\cosh\beta J\,\bigl(1+s_i s_j\,\tanh\beta J\bigr) \qquad (1)

where: $s_i = \pm1$ is the spin at site $i$ ; $\beta = 1/k_B T$ ; $J > 0$ the ferromagnetic coupling; $v = \tanh\beta J$ the high-temperature expansion variable. Read physically: the "1" is an inactive bond, the " $s_i s_j \tanh\beta J$ " an activated one carrying a factor $v$ and a spin pair. The identity is exact — no approximation is hidden in it. Multiplying (1) over all bonds and summing over spins keeps only graphs where every site has even degree, because the sum over a single spin, $\sum_{s=\pm1} s^n$ , is $2$ for even $n$ and $0$ for odd. The survivors are closed loops:

Z=(\cosh\beta J)^{N_b}\,2^{N}\sum_{\text{closed graphs}\,g} v^{\,\ell(g)} \qquad (2)

where: $N_b$ is the number of bonds; $N$ the number of sites; $\ell(g)$ the number of bonds in graph $g$ ; a "closed graph" has every vertex of even degree (Eulerian). The smallest closed graph on the square lattice is the four-bond plaquette, contributing $v^4$ with multiplicity $N$ — the first correction to the disordered free energy $N\ln 2$ . The low-temperature expansion measures energy from the aligned ground state; flipping clusters of spins breaks the bonds on their boundaries, and the excitations are closed domain walls:

Z=e^{\beta J N_b}\sum_{\text{domain walls}\,\Gamma} u^{\,|\Gamma|},\qquad u=e^{-2\beta J} \qquad (3)

where: $|\Gamma|$ is the total number of broken bonds (wall length); $e^{\beta J N_b}$ the ground-state weight; $u = e^{-2\beta J}$ the cost of one broken bond. The leading term is a single flipped spin breaking $2d$ bonds, contributing $N u^{2d}$ — for $d = 2$ that is $u^4$ . Both expansions have a radius of convergence set by the same physical singularity, the critical point, which in 2D is fixed exactly by self-duality:

\sinh 2\beta_c J = 1,\qquad k_B T_c/J = \dfrac{2}{\ln(1+\sqrt2)}\approx 2.269 \qquad (4)

where: $\beta_c$ is the critical inverse temperature and $v_c = \tanh\beta_c J = \sqrt2 - 1 \approx 0.414$ is the critical value of the high-temperature variable — the radius of convergence of the high- $T$ series on the positive axis. The most useful quantity to expand is the susceptibility, because its singularity gives the exponent $\gamma$ directly:

\chi(T)=\beta\sum_{n=0}^{\infty} a_n\,v^{n}\sim A\,(v_c-v)^{-\gamma} \qquad (5)

where: $\chi$ is the zero-field susceptibility per spin; $a_n$ are graph-counting coefficients (numbers of loop configurations); $\gamma$ the susceptibility exponent; $A$ a non-universal amplitude. A power-law singularity at $v_c$ forces a specific coefficient asymptotics, and that is what the ratio test reads off:

\frac{a_n}{a_{n-1}}\to \frac{1}{v_c}\Bigl(1+\frac{\gamma-1}{n}+\cdots\Bigr) \qquad (6)

where: the intercept of $a_n/a_{n-1}$ plotted against $1/n$ gives $1/v_c$ , and the slope of $n(v_c\,a_n/a_{n-1}-1)$ gives $\gamma - 1$ . Physically, longer loops proliferate at a rate set by $v_c$ , with a sub-leading correction whose size measures the exponent. The more robust extractor is the Dlog-Padé approximant, which handles the singularity structure directly:

\frac{\mathrm{d}}{\mathrm{d}v}\ln\chi(v)=\frac{P_L(v)}{Q_M(v)}\approx \frac{-\gamma}{v-v_c} \qquad (7)

where: $P_L, Q_M$ are polynomials fit to the series of $\frac{d}{dv}\ln\chi$ ; $v_c$ is the nearest real zero of $Q_M$ (the physical pole); $-\gamma$ is its residue. Because $\frac{d}{dv}\ln(v_c-v)^{-\gamma} = \gamma/(v-v_c)$ has a simple pole with residue $-\gamma$ , fitting a rational function to the log-derivative reads off both $v_c$ and $\gamma$ from the same object. Finally, the low-temperature magnetization series records how droplets eat away at the order:

m(T)=1-2u^{2d}-2d\,u^{4d-2}+\cdots,\qquad u=e^{-2\beta J} \qquad (8)

where: $m$ is the spontaneous magnetization per spin; $u^{2d}$ the single-flip term ( $u^4$ in 2D, $u^6$ in 3D). Each flipped droplet lowers the magnetization by twice its volume, weighted by its boundary cost — so the leading correction is exponentially small in $\beta J$ .

Derivation AThe 1D Ising chain as an exact high-temperature resummation

Our goal is to compute the high-temperature loop expansion for a periodic Ising ring and show it reproduces the exact transfer-matrix free energy — a clean, controlled check that the graph rules are right. The strategy exploits a special feature of a ring: with coordination number two at every site, the only closed (even-degree) subgraphs are the empty graph and the single loop that winds all the way around. Summing just those two contributions gives the entire partition function in closed form. The subtlety to watch is that the winding loop's weight $v^N$ survives at finite $N$ but vanishes in the thermodynamic limit, which is exactly why a 1D chain has no finite-temperature transition.

Step 1. Enumerate the closed graphs. On a ring every site touches exactly two bonds, so an even-degree subgraph must use either zero bonds (the empty graph) or all $N$ bonds (the full winding loop) — no other even-degree configuration exists.

\#\{\text{closed graphs}\} = 2 : \quad g_{\rm empty}\ (\ell=0), \quad g_{\rm ring}\ (\ell=N)

Why: a partial loop would leave two sites with odd degree, killed by the spin sum.

Step 2. Assemble the partition function. With $N_b = N$ bonds on the ring, equation (2) gives

Z=(\cosh\beta J)^{N}\,2^{N}\,(v^0 + v^{N})=(2\cosh\beta J)^{N}\big(1+\tanh^N\beta J\big)

Why: the empty graph contributes $v^0 = 1$ , the winding loop $v^N = \tanh^N\beta J$ ; the prefactor $(\cosh\beta J)^N 2^N = (2\cosh\beta J)^N$ .

Step 3. Compare with the transfer matrix. Its eigenvalues are $\lambda_\pm = e^{\beta J} \pm e^{-\beta J}$ , so $Z = \lambda_+^N + \lambda_-^N = (2\cosh\beta J)^N + (2\sinh\beta J)^N$ . Factoring out $(2\cosh\beta J)^N$ leaves $1 + \tanh^N\beta J$ — identical to Step 2.

Z_{\rm TM} = (2\cosh\beta J)^N\big(1 + \tanh^N\beta J\big) = Z_{\rm series}

Why: the two methods must agree, and they do term for term; the series is exact for the ring.

Step 4. Take the thermodynamic limit. For any $T > 0$ , $\tanh\beta J < 1$ , so $\tanh^N\beta J \to 0$ as $N \to \infty$ , leaving the bulk free energy $f = -k_B T\ln(2\cosh\beta J)$ , analytic for all $T > 0$ . The would-be transition sits at $v_c = 1$ , i.e. $T_c = 0$ .

f = -k_B T \ln(2\cosh\beta J), \qquad v_c = 1 \Leftrightarrow T_c = 0

Why: the singularity is pushed to zero temperature — no finite- $T$ transition in one dimension, exactly as the domain-wall argument predicts. *Verified: `verify_aq-sm-high-low-temperature-series-expansions.py` checks $Z_{\rm TM} = Z_{\rm series}$ numerically for several $(\beta J, N)$ and confirms $\tanh^N\beta J \to 0$ .*

Go deeper · Derivation B · 2 minThe low-temperature series of the 2D square lattice by droplet counting

Derivation BThe low-temperature series of the 2D square lattice by droplet counting

Our goal is to derive the leading terms of $\ln Z/N = 2\beta J + u^4 + 2u^6 + \cdots$ for the square lattice directly from the domain-wall expansion, illustrating how the boundary length of each flipped droplet sets the power of $u$ . The strategy is to enumerate the cheapest excitations — one flipped spin, then a nearest-neighbor pair — count how many bonds each breaks, and then take the logarithm, which automatically discards disconnected (reducible) droplet clusters and leaves only connected ones. The care point is the logarithm: it is what converts the raw partition-function sum into a proper per-site free energy with the right combinatorial factors.

Step 1. Single flip. Flipping one spin breaks its four nearest-neighbor bonds, costing $4 \times 2J = 8J$ , i.e. weight $u^4$ . There are $N$ sites, so the leading term of $Z/e^{2\beta J N}$ is $1 + N u^4 + \cdots$ .

Z/e^{2\beta J N} = 1 + N u^4 + \cdots, \qquad u^4 = e^{-8\beta J}

Why: $2d = 4$ broken bonds for a single flip in 2D; the leading correction is $u^{2d} = u^4$ .

Step 2. Nearest-neighbor pair. Two adjacent flipped spins share one bond, so together they break $6$ bonds (not $8$ ), giving $u^6$ . Each of the $2N$ lattice bonds hosts one such pair, contributing $2N u^6$ before the logarithm removes overcounting.

\text{NN pair: } 6 \text{ broken bonds} \Rightarrow u^6, \quad \text{multiplicity } 2N

Why: the shared bond is not broken, so a pair costs $6$ rather than $8$ — boundary length, not volume, sets the power.

Step 3. Take the logarithm to get the free energy per site. The logarithm strips disconnected clusters (the reducible $(N u^4)^2/2$ piece and its kin), leaving only connected droplets:

\frac{\ln Z}{N} = 2\beta J + u^4 + 2u^6 + \tfrac{9}{2}u^8 + \cdots

Why: $\ln$ is the connected-cluster generating function; the extensive coefficients ( $N u^4 \to u^4$ per site) come out finite and lattice-specific.

Step 4. Read the internal energy. Differentiating, $E/N = -\,d(\ln Z/N)/d\beta$ , and with $du/d\beta = -2Ju$ the leading correction to the ground-state energy $-2J$ is $\propto u^4 = e^{-8\beta J}$ — the activation gap of a single flip.

\frac{E}{N} = -2J + O(u^4), \qquad u^4 = e^{-8\beta J}

Why: the $8J$ gap of one flip governs the exponentially small low- $T$ excitation, and $u^4$ is its Boltzmann factor. *Verified: `verify_aq-sm-high-low-temperature-series-expansions.py` brute-force enumerates all $2^{16}$ spin configurations on a $4\times4$ torus, confirming $2$ ground states, single-flip excitations at $4$ broken bonds, NN pairs at $6$ , and — after taking the log — the coefficients $u^4 \to N$ and $u^6 \to 2N$ , i.e. $\ln Z/N = 2\beta J + u^4 + 2u^6 + \cdots$ .*

Limiting cases

The behavior of the two expansion variables and the series across temperature regimes:

Case	Result	Meaning
$T\to\infty$	$v=\tanh\beta J\to 0,\;\ln Z/N \to \ln 2$	Only the empty graph survives; $\chi \to \beta$ (Curie law).
$T\to 0$	$u=e^{-2\beta J}\to 0,\; m\to 1$	Series about the ordered ground state; leading excitation $u^{2d}$ , $m = 1 - 2u^{2d} + \cdots$ .
$T\to T_c$	$\chi\sim\vert v_c-v\vert^{-\gamma},\; C\sim\vert T-T_c\vert^{-\alpha}$	Radius of convergence is $v_c$ ; coefficients grow as $a_n\sim n^{\gamma-1}v_c^{-n}$ , encoding the exponent.
$d=1$	$Z=(2\cosh\beta J)^N(1+\tanh^N\!\beta J),\; v_c=1\;(T_c=0)$	Only the winding loop contributes; series converges for all $T > 0$ , no transition.
$d=2$ square	$\sinh 2\beta_c J=1,\; v_c=\sqrt2-1$	Self-dual point; the loop series' radius of convergence is exactly the Kramers–Wannier/Onsager $T_c$ .

Worked examples

1 · The 2D square-lattice critical point from the series' radius of convergence (confidence builder)

The high-temperature loop series converges up to the critical value $v_c = \tanh\beta_c J$ , fixed by self-duality at $\sinh 2\beta_c J = 1$ . Compute $v_c$ , $\beta_c J$ , and $k_B T_c/J$ .

Step 1. Self-dual condition: $\sinh 2\beta_c J = 1$ . Then $\cosh 2\beta_c J = \sqrt{1 + \sinh^2 2\beta_c J} = \sqrt2$ .

Step 2. Critical variable via the half-angle identity $v_c = \tanh\beta_c J = \dfrac{\sinh 2\beta_c J}{1 + \cosh 2\beta_c J} = \dfrac{1}{1+\sqrt2} = \sqrt2 - 1 = \boxed{0.41421}$ .

Step 3. Critical coupling: $\beta_c J = \tfrac12\ln(1+\sqrt2) = \tfrac12(0.88137) = \boxed{0.44069}$ .

Step 4. Temperature: $k_B T_c/J = 1/(\beta_c J) = 2/\ln(1+\sqrt2) = \boxed{2.26919}$ . So the high-temperature susceptibility series, a Taylor series in $v$ , has radius of convergence $v_c = 0.4142$ — and that number is the critical temperature. Truncate at any finite order and you approach it; you never quite reach it by summation.

2 · Reading the exponent γ off the coefficient ratios (exam level)

A power-law singularity $\chi \sim A(v_c - v)^{-\gamma}$ forces the coefficient asymptotics $a_n \sim C\,n^{\gamma-1} v_c^{-n}$ . Using this idealized form with the 2D Ising values $v_c = \sqrt2 - 1$ and $\gamma = 7/4$ , show how the ratio method recovers both $v_c$ and $\gamma$ .

Step 1. Form the ratio: $\dfrac{a_n}{a_{n-1}} = \dfrac{n^{\gamma-1}v_c^{-n}}{(n-1)^{\gamma-1}v_c^{-(n-1)}} = \dfrac{1}{v_c}\Big(\dfrac{n}{n-1}\Big)^{\gamma-1}$ .

Step 2. Expand $\big(1 + \tfrac{1}{n-1}\big)^{\gamma-1} \approx 1 + \dfrac{\gamma-1}{n} + O(1/n^2)$ , so $\dfrac{a_n}{a_{n-1}} \to \dfrac{1}{v_c}\Big(1 + \dfrac{\gamma-1}{n}\Big)$ .

Step 3. Intercept gives $v_c$ : plot $a_n/a_{n-1}$ against $1/n$ ; the $n \to \infty$ intercept is $1/v_c = 1/0.41421 = 2.41421$ , so $v_c = \boxed{0.41421}$ .

Step 4. Slope gives $\gamma$ : form $n\big(v_c\,a_n/a_{n-1} - 1\big) \to \gamma - 1$ . At $n = 320$ this evaluates to $0.75$ (checked numerically), so $\gamma - 1 = 3/4$ and $\gamma = \boxed{7/4}$ .

Step 5. Interpretation. The intercept is where loops proliferate (the radius of convergence, $= v_c$ ); the sub-leading correction measures how fast the susceptibility diverges (the exponent $\gamma$ ). For the real 2D Ising model, high-order series analyzed this way reproduce $\gamma = 7/4$ exactly — a triumph of pure counting.

Origins · 1 minduality and how this was discovered

duality

order ↔ disorder · the bridge between the two series

Duality is the concept that binds the two temperature expansions into one. A closed loop of activated bonds in the high-temperature series, drawn on the lattice, is the same object as a closed domain wall in the low-temperature series, drawn on the dual lattice. Kramers and Wannier recognized in 1941 that these two generating functions are identical on the self-dual square lattice, so the high- and low-temperature series are mirror images: knowing the coefficients of one determines the other. This is why the series can locate the exact critical temperature, $\sinh 2\beta_c J = 1$ , without ever being summed — the singularity is pinned by the symmetry of the coefficients, not by their numerical value. Duality also polices the enumeration: a mistake in one series shows up as a mismatch with its dual, catching graph-counting errors before they corrupt an exponent estimate.

History

1912

— Onsager and others develop cluster and virial expansions, the ancestors of graphical series methods in statistical mechanics.

1936

— Peierls uses low-temperature droplet counting to prove the 2D Ising model orders — the first rigorous domain-wall argument.

1941

— Kramers and Wannier equate the high- and low-temperature loop series to fix the exact 2D

T_c

from self-duality.

1944

— Onsager's exact solution confirms

\sinh 2\beta_c J = 1

and the exponents the series had been probing.

1963

— Domb, Sykes, Fisher and collaborators push high-temperature series to high order, extracting the first accurate 3D Ising exponents by ratio and Padé analysis.

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 13 (Ising model, series expansions) · Kardar, Statistical Physics of Fields, Ch. 7 (series expansions and 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 Expansions).

Provenance. All structural results (bond identity, graph expansion, duality, self-dual $T_c$ ) are standard and derived above; a verifier should reproduce: (i) the 1D resummation $Z = (2\cosh\beta J)^N(1+\tanh^N\beta J)$ against the transfer matrix; (ii) $\sinh 2\beta_c J = 1$ , $v_c = \sqrt2 - 1$ , $k_B T_c/J = 2/\ln(1+\sqrt2) \approx 2.269$ ; (iii) the low- $T$ coefficients $\ln Z/N = 2\beta J + u^4 + 2u^6 + \cdots$ from single-flip and nearest-neighbor-pair droplets (confirmed here by brute-force enumeration on a $4\times4$ torus). The 3D exponent values ( $\gamma \approx 1.237$ , $\nu \approx 0.630$ , $\beta \approx 0.326$ , $\alpha \approx 0.110$ ) and $k_B T_c/J \approx 4.5115$ for the simple cubic lattice are standard literature numbers (Domb & Green Vol. 3; modern series/MC reviews). All numbers are reproduced in `verify_aq-sm-high-low-temperature-series-expansions.py`.

Prerequisites. The Ising Model: Definition and Hamiltonian · One-Dimensional Ising Model and Transfer Matrix Method · Peierls Argument for Long-Range Order · Domain Walls and Interfacial Excitations. Leads to. Kramers-Wannier Duality · Onsager Solution of the Two-Dimensional Ising Model · Order Parameter and Spontaneous Magnetization · Transfer Matrix Formalism for Lattice Models.

Why is the high-temperature variable tanh βJ rather than βJ?

Because the exact bond identity is $e^{\beta J s_i s_j} = \cosh\beta J\,(1 + s_i s_j\tanh\beta J)$ — linear in $s_i s_j$ with coefficient $\tanh\beta J$ . That linearization is precisely what lets the spin sum select closed graphs. If you tried to expand in $\beta J$ instead, the bond weight would not factor into "inactive + activated" and the graph rules would collapse. There is also a convergence bonus: $\tanh\beta J \le 1$ even when $\beta J$ is large, so the natural variable stays bounded all the way to low temperature.

Why can't I just compute a hundred terms and sum right up to T_c?

Because a power series diverges at its radius of convergence, and for these series that radius is the critical point ( $v_c$ or $u_c$ ). No matter how many terms you keep, direct summation cannot cross $T_c$ ; the partial sums simply blow up as you approach it. To get into the critical region you must analytically continue — Padé or Dlog-Padé approximants reconstruct the singularity from the coefficient pattern, effectively summing the divergent tail in a controlled way.

Where does the leading u⁴ in the 2D low-temperature series come from?

From the cheapest possible excitation: flipping a single spin. On the square lattice that spin has four neighbors, so flipping it breaks four bonds, each costing $2J$ , for a total of $8J$ . The Boltzmann factor is $e^{-8\beta J} = (e^{-2\beta J})^4 = u^4$ . In $d$ dimensions a single flip breaks $2d$ bonds, so the leading low-temperature correction is always $u^{2d}$ — $u^4$ in 2D, $u^6$ in 3D. Boundary length, not droplet volume, sets the power.

What is the difference between the ratio method and Dlog-Padé?

The ratio method uses successive coefficient ratios $a_n/a_{n-1}$ : their $n\to\infty$ intercept gives $1/v_c$ and the slope gives $\gamma - 1$ . It is simple but sensitive to sub-leading "confluent" corrections and to spurious singularities on the negative real axis. Dlog-Padé fits a rational function to the log-derivative $\frac{d}{dv}\ln\chi$ , whose physical pole sits at $v_c$ with residue $-\gamma$ . Because it models the singularity structure directly, Dlog-Padé is usually more accurate and is the standard modern tool.

How reliable are series estimates compared with Monte Carlo?

Very. For the 3D Ising model, high-order series analyzed by Dlog-Padé give $\gamma \approx 1.237$ , $\nu \approx 0.630$ , $\beta \approx 0.326$ — in agreement with large-scale Monte Carlo, the renormalization group, and the conformal bootstrap to several digits. Series and Monte Carlo have different systematic errors (truncation and extrapolation versus finite-size and statistical), so their agreement is a genuine cross-check rather than a coincidence. That four-way concordance is one of the strongest confirmations of universality.

Do these series methods work for models other than the Ising model?

Yes — the same graph technology extends broadly. Potts and clock models, Heisenberg and XY spins (where the bond factor becomes a spherical-harmonic expansion), and lattice gauge theories all admit high-temperature (strong-coupling) expansions in loops or plaquettes. The strong-coupling expansion of lattice quantum chromodynamics is the direct field-theory descendant, and it was historically how physicists first argued for quark confinement. Wherever you can write the partition function as a sum over graphs, you can build a series.

Why does the 1D chain give a series that converges for all T > 0?

Because on a ring the only closed graphs are the empty graph and the single loop winding all the way around, contributing $1 + \tanh^N\beta J$ . There is no singularity at any finite temperature — the would-be critical point sits at $v_c = 1$ , i.e. $T_c = 0$ . So the high-temperature series has infinite radius of convergence in $T$ (it only "diverges" at $T = 0$ ), reflecting the exact statement that the 1D Ising model does not order at any positive temperature.