The Ising Model: Definition and Hamiltonian

The Ising model is the fruit fly of statistical mechanics — the simplest system that still does something profound. Place a classical two-state variable $\sigma_i=\pm1$ on every site of a lattice, let neighbouring sites prefer to agree with an energy $-J$ per aligned bond, and add an optional field $h$ that tilts every spin. That is the entire model. Wilhelm Lenz proposed it in 1920 and handed it to his student Ernst Ising, who solved the one-dimensional chain in his 1924 Hamburg thesis (published 1925) and found — disappointingly, he thought — no phase transition at any positive temperature. He concluded the model was a failure and left physics research. He was wrong about the model, if not about $d=1$ . In two and higher dimensions the same Hamiltonian has a genuine thermodynamic phase transition: below a critical temperature $T_c$ a macroscopic fraction of spins lock together even with no external field, spontaneously breaking the $\mathbb{Z}_2$ symmetry between "all up" and "all down". Lars Onsager's 1944 exact solution of the two-dimensional square lattice was the first rigorous proof that a many-body partition function can become non-analytic in the thermodynamic limit — the mathematical signature of a phase transition. The model then became the proving ground for essentially every idea in the modern theory of criticality: order parameters, spontaneous symmetry breaking, duality, series expansions, scaling, and the renormalization group. At bottom the physics is a contest between energy and entropy. The Boltzmann weight $e^{-\beta H}$ rewards aligned, low-energy configurations, while entropy rewards disorder; temperature sets the exchange rate. What decides the winner is dimensionality — because disordering an ordered state means inserting domain walls, and the cost of a wall scales differently in $d=1$ (cheap) than in $d\ge2$ (expensive). And the model is not just about magnets: the identical Hamiltonian describes lattice gases (occupied/empty sites), binary alloys (A/B atoms), adsorption, and simplified neural memories. That a single toy model governs the critical behaviour of such different systems is the concrete face of universality, and it is why the Ising model is the most-studied model in all of statistical physics.

Key takeaways

The Ising model puts spins $\sigma_i=\pm1$ on a lattice with nearest-neighbour coupling $J$ (ferromagnetic $J>0$ favours alignment) and a field $h$ that biases each spin.
Its Hamiltonian $H=-J\sum_{\langle ij\rangle}\sigma_i\sigma_j-h\sum_i\sigma_i$ is the simplest with a true thermodynamic phase transition and spontaneous symmetry breaking in $d\ge2$ .
At $h=0$ it has a global $\mathbb{Z}_2$ symmetry $\sigma_i\to-\sigma_i$ ; below $T_c$ this breaks spontaneously and a nonzero magnetization appears.
It is exactly solvable in $d=1$ (no transition at $T>0$ , Ising 1925) and $d=2$ at $h=0$ (Onsager 1944, $k_BT_c/J=2/\ln(1+\sqrt2)\approx2.269$ ); $d\ge3$ needs numerics and the renormalization group.
Domain-wall bookkeeping explains why dimension decides: in $d=1$ a wall costs a fixed $2J$ but gains $\sim k_BT\ln N$ entropy, so order always dies; in $d\ge2$ the wall energy grows with the system, so order survives.
Key idea: a binary variable per site plus a short-range alignment energy is the minimal recipe for a genuine thermodynamic phase transition with spontaneous symmetry breaking — and every concept of critical phenomena can be illustrated inside it.

The core insight

A binary variable per site plus a short-range alignment energy is the minimal ingredient set that produces a true thermodynamic phase transition with spontaneous symmetry breaking; whether that transition happens is decided by dimensionality, and essentially every concept in critical phenomena can be illustrated within this one model.

Where this lives today

The Ising model remains the benchmark against which every new idea in statistical physics is first tested — and it keeps delivering. Because $d=2$ is exactly solved and $d=3$ is exhaustively studied numerically, it is the standard proving ground for Monte Carlo algorithms, transfer-matrix and tensor-network methods, series expansions, the renormalization group, and the conformal bootstrap, which has pinned the 3D Ising exponents ( $\nu\approx0.6300$ , $\eta\approx0.0363$ ) to a precision rivaling experiment. Its reach far exceeds magnetism: the same Hamiltonian, through the lattice-gas mapping, governs the liquid–gas critical point of real fluids (whose measured $\beta\approx0.326$ matches the 3D Ising value), the order–disorder transitions of binary alloys such as $\beta$ -brass, adsorbed monolayers, and — with modified couplings — spin glasses, neural-network memories (the Hopfield model), and inference problems in machine learning. When a physicist wants to understand a cooperative phenomenon from scratch, the first move is almost always to write down an Ising-like Hamiltonian.

The physical picture

Picture a chessboard of tiny compass needles, each forced to point strictly up or down. Every needle "wants" to agree with its neighbours — each aligned bond lowers the energy by $J$ — but thermal kicks randomly flip them, and temperature is the strength of those kicks. There are two knobs: temperature (entropy versus energy) and field (an external thumb on the scale). At zero field the up and down states are perfectly symmetric, so if the board orders it must spontaneously choose a direction — it breaks a symmetry it was under no obligation to break. The deep question is why dimensionality decides whether ordering happens at all. Disordering an ordered state requires inserting a domain wall — a boundary between an up region and a down region. In one dimension a wall is a single broken bond: its energy cost is a fixed $2J$ , but it can sit at any of roughly $N$ positions, so it carries configurational entropy $\sim k_B\ln N$ . The free-energy balance $2J-k_BT\ln N$ always turns negative as $N\to\infty$ for any $T>0$ : walls proliferate and order is destroyed. In two dimensions a wall is an extended line whose energy grows with the system size, so at low temperature the energy cost outruns the available entropy and an ordered phase survives up to a finite $T_c$ . That contrast is the heart of the Peierls argument and the reason Ising's 1D result was misleadingly pessimistic.

Figure

The Ising board: aligned bonds cost −J, a domain wall costs +2J per broken bond

The Ising model is "maximally simple but not trivial". Strip away the lattice entirely and let every spin couple to every other (mean field) and the model becomes exactly solvable — but you lose the fluctuations that make criticality interesting. Keep the lattice in $d=2$ and you keep the fluctuations, and Onsager's exact solution rewards you with the celebrated non-classical exponents $\beta=1/8$ , $\gamma=7/4$ , $\nu=1$ . The punchline figure below makes the dimensional story vivid: the order parameter $m(T)$ is pinned at zero for all $T>0$ in $d=1$ , but rises below $T_c$ in $d=2$ with the famous $1/8$ -power shape.

Figure

Order parameter m(T): d=1 is pinned at zero for all T＞0, while d=2 orders below T_c with β=1/8

The core equations

The whole model is one Hamiltonian: a sum over nearest-neighbour bonds that rewards alignment, plus a coupling of each spin to an external field.

H = -J\sum_{\langle ij\rangle}\sigma_i\sigma_j - h\sum_{i}\sigma_i \qquad (1)

where: $\sigma_i=\pm1$ is the spin at site $i$ ; $J$ the nearest-neighbour coupling ( $J>0$ ferromagnetic, $J<0$ antiferromagnetic); $\langle ij\rangle$ runs over distinct nearest-neighbour pairs (each bond counted once); $h$ the external field (absorbing the magnetic moment); $N$ the number of sites. Physically: aligned bonds ( $\sigma_i\sigma_j=+1$ ) lower the energy by $J$ , so at low $T$ the system wants uniform alignment. All thermodynamics flow from the partition function — the sum of Boltzmann weights over all $2^N$ configurations.

Z = \sum_{\{\sigma\}} e^{-\beta H} = \sum_{\sigma_1=\pm1}\!\cdots\!\sum_{\sigma_N=\pm1} \exp\!\Big[\beta J\!\sum_{\langle ij\rangle}\!\sigma_i\sigma_j + \beta h\sum_i\sigma_i\Big] \qquad (2)

where: $\beta=1/(k_BT)$ ; $\{\sigma\}$ is one configuration (an assignment of $\pm1$ to all $N$ sites). For finite $N$ , $Z$ is a finite sum of exponentials and hence perfectly analytic — a phase transition, i.e. a singular free energy $f=-\lim_{N\to\infty}k_BT\,(\ln Z)/N$ , can only emerge in the thermodynamic limit (the Yang–Lee mechanism). The order parameter is the magnetization per spin, obtained as a field derivative of the free energy.

m = \langle\sigma_i\rangle = \frac{1}{N}\sum_i\langle\sigma_i\rangle = -\frac{1}{N}\frac{\partial F}{\partial h},\qquad F=-k_BT\ln Z \qquad (3)

where: $m\in[-1,1]$ is the magnetization per spin and $\langle\cdot\rangle$ the canonical (Boltzmann) average. Physically $m$ measures how far the system has ordered: $m=0$ in the disordered (paramagnetic) phase, $|m|>0$ once symmetry is broken. At zero field the Hamiltonian is invariant under flipping every spin — the global $\mathbb{Z}_2$ symmetry that the ordered phase must break.

H \longmapsto H \quad\text{under}\quad \sigma_i\to-\sigma_i\ \ \text{(only when } h=0) \qquad (4)

where: $\mathbb{Z}_2$ is the two-element group of the global flip. The bond term survives because $(-\sigma_i)(-\sigma_j)=\sigma_i\sigma_j$ , but the field term $-h\sum_i\sigma_i$ changes sign unless $h=0$ — so the symmetry is exact only at $h=0$ , and below $T_c$ the system spontaneously picks one of the two symmetry-related ordered states. The response of the magnetization to the field is fixed by the equilibrium fluctuations of the total magnetization — a fluctuation–response relation.

\chi = \frac{\partial m}{\partial h}\Big|_{h=0} = \frac{\beta}{N}\big(\langle M^2\rangle-\langle M\rangle^2\big),\qquad M=\sum_i\sigma_i \qquad (5)

where: $\chi$ is the isothermal susceptibility and $M$ the extensive total magnetization. Physically: a system that fluctuates strongly (near $T_c$ ) responds strongly to a field, so $\chi$ diverges at criticality. A remarkable feature is that the same spin Hamiltonian is a lattice gas after a linear relabelling of variables.

\sigma_i = 2n_i-1\in\{-1,+1\},\quad n_i\in\{0,1\};\qquad \sigma_i\sigma_j = 4n_in_j - 2n_i - 2n_j + 1 \qquad (6)

where: $n_i=1$ means "site occupied", $n_i=0$ "empty". Substituting turns the ferromagnetic bond into a lattice-gas attraction with $J=\varepsilon/4$ plus chemical-potential and constant terms — so the liquid–gas critical point shares the Ising universality class. The two-dimensional square lattice at $h=0$ has an exactly known critical temperature — Onsager's landmark result.

\left.\frac{k_B T_c}{J}\right|_{\text{2D, sq}} = \frac{2}{\ln(1+\sqrt2)} \approx 2.2692 \qquad (7)

where: $T_c$ is the critical temperature of the square-lattice Ising model. Since $\ln(1+\sqrt2)\approx0.8814$ , the number is $\approx2.2692$ ; below it, spontaneous magnetization appears with $\beta=1/8$ .

Derivation AThe smallest Ising systems: exact partition function of a short chain

The goal is to solve the Ising chain exactly for a handful of spins, extracting the partition function, mean energy, and heat capacity as concrete numbers, and to see explicitly that a finite system has no spontaneous magnetization — order is strictly a thermodynamic-limit phenomenon. The strategy is direct enumeration: for an open chain of $N$ spins at $h=0$ there are $2^N$ configurations, each with an energy set by the number of aligned versus broken bonds, and summing the Boltzmann weights gives a closed form we can evaluate. This tiny calculation is the honest confidence-builder: everything larger is the same sum organized more cleverly (transfer matrices), and it already shows why finite systems stay analytic.

Step 1. For an open chain the energy is $E=-J\sum_{i=1}^{N-1}\sigma_i\sigma_{i+1}$ , and using $e^{\beta J\sigma_i\sigma_j}=\cosh\beta J\,(1+\sigma_i\sigma_j\tanh\beta J)$ on each of the $N-1$ independent bonds, the spin sums factorize to give $Z_N=2^N(\cosh\beta J)^{N-1}$ . Why: each bond variable $\sigma_i\sigma_{i+1}$ is $\pm1$ , so the exponential truncates exactly; summing $\pm1$ over each of the $N$ sites gives $2^N$ and leaves only the $\cosh$ factors.

Step 2. For $N=2$ this is $Z_2=4\cosh\beta J=2e^{\beta J}+2e^{-\beta J}$ , exactly the sum over the four states $\{\uparrow\uparrow,\downarrow\downarrow\}$ (energy $-J$ ) and $\{\uparrow\downarrow,\downarrow\uparrow\}$ (energy $+J$ ). Why: two aligned states weighted $e^{\beta J}$ and two anti-aligned weighted $e^{-\beta J}$ — a direct check of the closed form.

Step 3. The mean energy is $\langle E\rangle=-\partial\ln Z_N/\partial\beta=-(N-1)J\tanh\beta J$ : each of the $N-1$ bonds contributes $-J\tanh\beta J$ independently. For $N=4$ at $\beta J=1$ this gives $\langle E\rangle=-3J\tanh(1)\approx-2.28478\,J$ . Why: taking $\ln Z=N\ln2+(N-1)\ln\cosh\beta J$ and differentiating isolates the per-bond energy $-J\tanh\beta J$ .

Step 4. The heat capacity follows from the energy variance, $C/k_B=\beta^2(\langle E^2\rangle-\langle E\rangle^2)$ ; for the $N=4$ open chain at $\beta J=1$ direct enumeration gives $Z_4\approx58.788$ and $C/k_B\approx1.25992$ . Crucially, at $h=0$ the magnetization $m=0$ for every finite $N$ : the $\mathbb{Z}_2$ symmetry is exact and the two ordered states cancel. Why: a finite sum is analytic and $\mathbb{Z}_2$ -symmetric, so it cannot single out an ordered direction — spontaneous magnetization needs $N\to\infty$ . *Verified: verify_aq-sm-ising-model-definition-hamiltonian.py enumerates the $2^N$ configurations for $N=2,3,4$ , confirms $Z_N=2^N\cosh^{N-1}\beta J$ (open) and $Z_N=(2\cosh\beta J)^N+(2\sinh\beta J)^N$ (periodic), and reproduces $Z_4\approx58.788$ , $\langle E\rangle\approx-2.28478\,J$ , $C/k_B\approx1.25992$ at $\beta J=1$ .*

Go deeper · Derivation B · 2 minWhy one dimension has no transition: the single-domain-wall free energy

Derivation BWhy one dimension has no transition: the single-domain-wall free energy

The goal is to show, by a free-energy estimate of one domain wall, that the one-dimensional Ising chain cannot sustain long-range order at any positive temperature — the result that so disappointed Ising in 1925. The strategy is to start from the perfectly aligned ground state, insert a single domain wall (one broken bond), and compare the energy it costs to the entropy it gains from the many places it can sit. Because the entropy grows with system size while the energy does not, the free-energy change becomes negative in the thermodynamic limit, so walls always proliferate. The same bookkeeping run in $d\ge2$ reverses the conclusion, which is exactly why dimension is the decisive variable.

Step 1. Insert one wall into an aligned open chain of $N$ spins: exactly one bond flips from aligned ( $-J$ ) to broken ( $+J$ ), so the energy cost is $\Delta E=2J$ , independent of $N$ . Why: only the single bond straddling the up/down boundary changes; every other bond is untouched.

Step 2. The wall can sit on any of the $N-1$ bonds, giving a configurational entropy $\Delta S=k_B\ln(N-1)$ . Why: the number of distinct one-wall configurations is $N-1$ , and entropy is $k_B$ times the log of that multiplicity.

Step 3. The free-energy change is $\Delta F=\Delta E-T\Delta S=2J-k_BT\ln(N-1)$ . For any fixed $T>0$ , $\Delta F\to-\infty$ as $N\to\infty$ : creating a wall lowers the free energy, so walls proliferate and destroy long-range order. Why: the entropy term grows without bound with system size while the energy cost stays fixed, so entropy always wins eventually.

Step 4. Hence $T_c=0$ in $d=1$ : the chain orders only at exactly $T=0$ . In $d\ge2$ a domain wall is an extended interface whose energy grows with its length ( $\propto L$ in $2$ D), so $\Delta E$ outscales the $\sim\ln(\text{positions})$ entropy at low $T$ and an ordered phase becomes stable up to a finite $T_c$ — the essence of the Peierls argument. Why: the geometry of a wall changes from a point (fixed cost) in $1$ D to a line (growing cost) in $2$ D, flipping the energy–entropy balance. *Verified: verify_aq-sm-ising-model-definition-hamiltonian.py checks $\Delta F=2J-k_BT\ln(N-1)$ and confirms $\Delta F\to-\infty$ as $N\to\infty$ for any $T>0$ , giving $T_c=0$ ; it also confirms the transfer-matrix free energy of the 1D chain is analytic for all $T>0$ .*

Limiting cases

The Ising model's behaviour depends sharply on dimension, field, and the sign of $J$ :

Case	Result	Meaning
$d=1$ , $h=0$	$T_c=0$	No ordered phase at any $T>0$ ; free energy analytic for all $T>0$ . Solved exactly by the transfer matrix.
$d=2$ , $h=0$ , square lattice	$k_BT_c/J = 2/\ln(1+\sqrt2)\approx2.269$	Onsager's exact result; continuous transition with $\beta=1/8$ , $\nu=1$ , $\gamma=7/4$ , logarithmic specific heat ( $\alpha=0$ ).
$d=3$ , $h=0$ , simple cubic	$k_BT_c/J\approx4.5115$	No exact solution; exponents $\beta\approx0.326$ , $\nu\approx0.630$ from RG / Monte Carlo / conformal bootstrap (3D Ising class).
Infinite-range / mean field	$k_BT_c/J = z$ (coordination number)	Exact above the upper critical dimension $d_c=4$ ; classical exponents $\beta=1/2$ , $\gamma=1$ , $\delta=3$ .
High temperature, any $d$	$\langle\sigma_0\sigma_r\rangle\sim e^{-r/\xi}$ , $\xi\to0$	Spins decorrelate; paramagnetic phase, zero spontaneous magnetization.

Worked examples

Example 1 — solving a four-spin chain exactly. Compute the partition function, mean energy, and heat capacity of an open four-spin Ising chain at $h=0$ and $\beta J=1$ , working entirely from direct enumeration.

Step 1. With $N=4$ open bonds ( $N-1=3$ ), the closed form is $Z_4=2^4(\cosh\beta J)^3=16\cosh^3(\beta J)$ . At $\beta J=1$ , $\cosh(1)\approx1.54308$ , so $Z_4=16(1.54308)^3\approx\mathbf{58.788}$ . (Cross-check: direct sum over all $2^4=16$ configurations gives the same $58.788$ .)

Step 2. The mean energy is $\langle E\rangle=-(N-1)J\tanh\beta J=-3J\tanh(1)$ . With $\tanh(1)\approx0.76159$ , $\langle E\rangle\approx\mathbf{-2.28478\,J}$ — each of the three bonds sits, on average, at $-0.76159\,J$ , well aligned but not frozen.

Step 3. The heat capacity comes from the energy variance, $C/k_B=\beta^2(\langle E^2\rangle-\langle E\rangle^2)$ . Enumerating the sixteen states gives $C/k_B\approx\mathbf{1.25992}$ at $\beta J=1$ . And at $h=0$ , $m=0$ identically for this finite chain: no spontaneous magnetization at finite $N$ — order genuinely requires the thermodynamic limit.

Example 2 — locating the 2D critical temperature and comparing to mean field. For the square lattice ( $z=4$ ), compare Onsager's exact $T_c$ with the Weiss mean-field prediction $k_BT_c^{\rm MF}/J=z$ .

Step 1. Onsager: $k_BT_c/J=2/\ln(1+\sqrt2)$ . Since $1+\sqrt2\approx2.41421$ and $\ln(2.41421)\approx0.88137$ , this is $2/0.88137\approx\mathbf{2.2692}$ .

Step 2. Mean field: each spin has $z=4$ neighbours, so $k_BT_c^{\rm MF}/J=z=\mathbf{4}$ . Mean field overestimates $T_c$ by a factor $4/2.269\approx1.76$ , because it ignores the fluctuations that make ordering harder in low dimensions.

Step 3. The discrepancy shrinks as $d$ (and $z$ ) grow: mean field becomes exact above the upper critical dimension $d_c=4$ . The gap between $4$ and $2.269$ is precisely the fingerprint of the fluctuations that Onsager's exact solution keeps and mean field throws away — and it is why the exact 2D exponents ( $\beta=1/8$ ) differ so sharply from the classical $\beta=1/2$ .

Origins · 1 minernst ising and how this was discovered

ernst ising

1900–1998 · Ernst Ising was a German-born physicist who gave his name to the model after solving its one-dimensional case in his 1924 doctoral thesis at Hamburg (published 1925), supervised by Wilhelm Lenz, who had proposed the model in 1920. Ising found that the linear chain has no phase transition at any positive temperature and, generalizing incorrectly, concluded the model failed to describe ferromagnetism; discouraged, he left research. In fact the model's triumph came in higher dimensions, through Peierls, Kramers, Wannier and above all Onsager's 1944 exact 2D solution. Ising's later life was shaped by history: as a Jew in Nazi Germany he lost his teaching post, survived the war, and emigrated to the United States in 1947, where he taught physics for decades at Bradley University, largely unaware for years of how central "his" model had become. He lived to 98, long enough to see the Ising model become the single most-studied model in statistical physics.

History

1920

— Wilhelm Lenz proposes the model of two-state spins on a lattice

1925

— Ernst Ising solves the 1D chain and finds no phase transition at T ＞ 0

1936

— Rudolf Peierls gives the contour argument for long-range order in 2D

1941

— Kramers and Wannier locate the 2D critical point via duality and the transfer matrix

1944

— Lars Onsager solves the 2D square lattice exactly, proving a nonanalytic free energy

1952

— Yang derives the spontaneous magnetization, giving the exponent β = 1/8

1971

— Wilson's renormalization group finally cracks the 3D Ising exponents

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 12–13 · Kardar, Statistical Physics of Fields, Ch. 1–3 · Reif, Fundamentals of Statistical and Thermal Physics, Ch. 10 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 8 · Landau & Lifshitz, Statistical Physics Part 1, §§145–151 · E. Ising, Z. Phys. 31, 253 (1925) · L. Onsager, Phys. Rev. 65, 117 (1944).

Provenance. The Hamiltonian, partition function, and $\mathbb{Z}_2$ symmetry are standard (Pathria & Beale Ch. 12; Kardar SPF Ch. 1). The two-spin and short-chain partition functions are textbook enumerations (Kardar SPF Ch. 3) — recomputed here by direct summation: open chain $Z_N=2^N\cosh^{N-1}\beta J$ , and the worked $N=4$ values $Z_4\approx58.788$ , $\langle E\rangle\approx-2.28478\,J$ , $C/k_B\approx1.25992$ at $\beta J=1$ . The Peierls single-wall free energy $2J-k_BT\ln(N-1)$ follows Pathria/Sethna and Peierls (1936). Onsager $k_BT_c/J=2/\ln(1+\sqrt2)\approx2.2692$ from Onsager (1944); $\ln(1+\sqrt2)\approx0.88137$ is checked in the verify script. The 2D exact exponents $\beta=1/8,\gamma=7/4,\nu=1,\eta=1/4,\alpha=0$ are Onsager/Yang; the 3D simple-cubic $k_BT_c/J\approx4.5115$ and $\beta\approx0.326$ , $\nu\approx0.630$ , $\eta\approx0.036$ are standard conformal-bootstrap / Monte-Carlo values (Pelissetto–Vicari review; Sethna Ch. 12). The lattice-gas relabel $\sigma_i=2n_i-1$ giving $J=\varepsilon/4$ is verified symbolically.

Builds on: Canonical Ensemble and Partition Function · Order Parameter and Spontaneous Magnetization. Leads to: One-Dimensional Ising Model and Transfer Matrix Method · Mean-Field (Weiss) Theory of the Ising Model · Onsager Solution of the Two-Dimensional Ising Model · Lattice Gas Model and Ising Mapping · Peierls Argument for Long-Range Order.

Why is the Ising model called the "fruit fly" or "drosophila" of statistical mechanics?

Because it is the simplest system that still exhibits the full richness of a phase transition, just as the fruit fly is the simplest organism that displays the essential mechanisms of genetics. Every major concept — order parameters, spontaneous symmetry breaking, critical exponents, universality, duality, the renormalization group, the conformal bootstrap — can be introduced, tested, and understood inside the Ising model before being applied to more complicated systems. It is exactly solvable in one and two dimensions and exhaustively studied in three, so it serves as the universal benchmark: any new theoretical or computational method in statistical mechanics is validated against it first.

What exactly is the Z2 symmetry, and how can it be "spontaneously broken"?

$\mathbb{Z}_2$ is the two-element group whose nontrivial operation flips every spin, $\sigma_i\to-\sigma_i$ . At $h=0$ the Hamiltonian is invariant under this global flip, because each bond $\sigma_i\sigma_j$ is unchanged. Spontaneous breaking means that although the Hamiltonian keeps the symmetry, a single equilibrium state below $T_c$ does not: the system settles into either the mostly-up or the mostly-down phase, not a symmetric mixture. Formally, one takes $N\to\infty$ first (so the free energy can become nonanalytic) and then $h\to0^+$ , which selects the up phase; $h\to0^-$ selects the down phase. The two states are related by the symmetry but are separated by an infinite free-energy barrier in the thermodynamic limit.

Why does the one-dimensional chain refuse to order at any positive temperature?

Because a domain wall is too cheap. To disorder the aligned chain you insert a wall (one broken bond), which costs a fixed energy $2J$ regardless of system size. But that wall can sit at any of $\sim N$ positions, giving it configurational entropy $k_B\ln N$ . The free-energy change is $\Delta F=2J-k_BT\ln N$ , which becomes negative for any $T>0$ once $N$ is large enough — so walls always proliferate and wash out order. In two dimensions a wall is an extended line whose energy grows with the system size, so it cannot be created for free, and an ordered phase survives up to $T_c$ . This energy-versus-entropy contrast is the core of the Peierls argument and the reason dimensionality is decisive.

How does the antiferromagnetic model (

J<0

) differ from the ferromagnetic one?

On a bipartite lattice (square, cubic, honeycomb) the antiferromagnet is secretly the ferromagnet in disguise. Flip all the spins on one of the two sublattices, $\sigma_i\to-\sigma_i$ for $i$ in sublattice B; this sends $J<0$ to $J>0$ , so at $h=0$ the antiferromagnet has exactly the same $|T_c|$ and thermodynamics as the ferromagnet, with the staggered (sublattice) magnetization playing the role of order parameter. The correspondence breaks down on non-bipartite lattices such as the triangular lattice, where antiferromagnetic bonds cannot all be satisfied simultaneously — this is geometric frustration, which suppresses ordering and can leave residual entropy at $T=0$ .

What does the lattice-gas mapping mean physically?

Relabel each spin by an occupation number, $n_i=(1+\sigma_i)/2$ , so $\sigma_i=+1$ becomes "site occupied" and $\sigma_i=-1$ becomes "empty". Substituting into the Ising Hamiltonian turns the ferromagnetic bond into a nearest-neighbour attraction between particles (with $J=\varepsilon/4$ ), the magnetic field into a chemical potential, and the magnetization into the density. The magnetic phase transition then is the liquid–gas transition: the ordered (magnetized) phase corresponds to a dense liquid, the disordered phase to a dilute gas, and the magnetic critical point to the liquid–gas critical point. This is why simple fluids and uniaxial magnets share critical exponents — a striking, experimentally confirmed instance of universality.

Why is Onsager's 2D solution such a landmark?

Onsager's 1944 exact solution of the two-dimensional square-lattice Ising model at $h=0$ was the first rigorous demonstration that the partition function of an interacting many-body system can become genuinely non-analytic in the thermodynamic limit — the mathematical signature of a phase transition. Before it, one could argue heuristically (Peierls) that a transition should occur, but no one had computed the full free energy of a nontrivial interacting model exactly. Onsager found $k_BT_c/J=2/\ln(1+\sqrt2)\approx2.269$ , a logarithmically divergent specific heat, and (with Yang) the magnetization exponent $\beta=1/8$ — values sharply different from mean-field predictions, proving that fluctuations qualitatively change critical behaviour in low dimensions. It set the standard for exact results in statistical mechanics.