Kramers-Wannier Duality

In 1941 two physicists at Bell Labs computed the exact temperature at which a two-dimensional magnet loses its magnetism — and they did it without ever solving the model. Three years before Lars Onsager's celebrated exact solution, Hendrik Kramers and Gregory Wannier found the critical point of the square-lattice Ising model using nothing but a symmetry. The trick is one of the most beautiful in all of statistical mechanics: the same partition function, expanded around high temperature and around low temperature, tells the same combinatorial story twice. Here is the story. Expand a hot Ising magnet and the corrections are closed loops of activated bonds, each loop weighted by $\tanh K$ (small when hot). Expand a cold Ising magnet and the corrections are closed domain walls — the boundaries of flipped regions — each wall weighted by $e^{-2K}$ (small when cold). Kramers and Wannier noticed these two ledgers count the same closed loops, one on the lattice, one on its dual. Equating them yields a map $K \leftrightarrow K^{*}$ that trades strong coupling for weak, order for disorder. Since a single phase transition must sit at the fixed point of that map, solving $\sinh 2K = 1$ hands you $k_B T_c/J = 2/\ln(1+\sqrt2) \approx 2.269$ for free. Duality turned out to be far more than a shortcut to one number. It is the prototype of every order–disorder duality in physics: the disorder operator conjugate to the spin, Wegner's $\mathbb{Z}_2$ lattice gauge duality, and the strong–weak dualities that organize modern quantum field theory and string theory all descend from this 1941 insight. Mastering it here, in the cleanest possible setting, is how you learn to read a partition function from two ends at once.

Key takeaways

Kramers–Wannier duality is an exact map sending the 2D Ising partition function at coupling $K$ to the same model at a dual coupling $K^{*}$ , trading high temperature for low, weak coupling for strong.
The map is the involution $\sinh 2K\,\sinh 2K^{*} = 1$ , equivalently $\tanh K^{*} = e^{-2K}$ : apply it twice and you return to $K$ .
Geometrically, high- $T$ closed bond-graphs on the lattice and low- $T$ closed domain walls on the dual lattice are the same loops; the square lattice is its own dual (self-dual).
If exactly one transition exists it must lie on the mirror line $K = K^{*}$ , forcing $\sinh 2K_c = 1$ , i.e. $K_c = \tfrac12\ln(1+\sqrt2) \approx 0.4407$ and $k_B T_c/J \approx 2.269$ — got in 1941, three years before Onsager.
Duality alone does not prove a transition exists; it only locates one if there is a single one. That single-transition assumption is the one physical input.
The construction generalizes: triangular and honeycomb lattices are duals of each other, and the same order–disorder logic seeds gauge dualities and field-theory strong–weak dualities.
Key idea: one closed-loop combinatorics counts hot bond-graphs on a lattice and cold domain walls on its dual; equating the two generating functions gives an exact strong–weak map whose fixed point pins the critical temperature.

The core insight

The same closed-loop combinatorics counts high-temperature bond graphs on a lattice and low-temperature domain walls on its dual; equating the two generating functions gives an exact strong–weak coupling map whose fixed point pins the critical temperature.

Where this lives today

Kramers–Wannier duality has aged into a foundational idea rather than a historical curiosity. In condensed-matter and high-energy physics it is the ancestor of the entire web of dualities: electric–magnetic duality, the strong–weak $S$ -duality of gauge theories, and the order–disorder dualities that classify topological phases of matter. The disorder operator $\mu$ that Kadanoff and Ceva made precise is now understood as a defect operator, and the same self-duality that fixes the 2D Ising $T_c$ reappears in the self-dual point of the transverse-field Ising chain, a canonical model of quantum criticality and a standard benchmark for quantum simulators built from trapped ions and superconducting qubits. The practical descendant is just as alive. Because high- and low-temperature series are duality images, terms of one determine terms of the other, giving a stringent consistency check on the series-expansion estimates of critical temperatures and exponents for the many lattice models — three-dimensional Ising, Potts, lattice gauge theories — that have no exact solution. Every time a researcher cross-checks a strong-coupling lattice-QCD expansion against a weak-coupling one, they are using the great-grandchild of a 1941 Bell Labs calculation.

The physical picture

Think of the Ising partition function as a bookkeeping of loops. Cool a hot magnet a little and the leading corrections are the smallest closed circuits of "activated" bonds, each carrying a weight $\tanh K$ that is small at high temperature. Warm a cold magnet a little and the leading corrections are the smallest closed domain walls — the perimeters of flipped droplets — each carrying a weight $e^{-2K}$ per unit length that is small at low temperature. Duality is the statement that these two ledgers are the same book, read from opposite ends. The geometric heart of it is the dual lattice. Put a site at the center of every square plaquette and connect neighbors across each bond; the dual of a square lattice is again a square lattice. So the cold-magnet domain walls, which live on the dual, sit on an identical copy of the original lattice. The model is its own dual — and that self-duality is exactly why the critical point lands on a single self-consistent equation rather than a relation between two different lattices.

Figure

The duality map K* versus K — the self-dual fixed point sits where the curve crosses the diagonal

Read the figure as a mirror. The curve $K^{*}(K)$ is decreasing — a hot, weakly coupled model ( $K$ small) maps to a cold, strongly coupled one ( $K^{*}$ large), and vice versa. Any singularity of the free energy must therefore come in dual pairs $(K, K^{*})$ , one on each side of the mirror. The only way a singularity can be unpaired is to sit exactly on the mirror line $K = K^{*}$ , where the blue curve crosses the dashed diagonal. A model with a single transition has no choice: its critical point is forced onto that intersection.

Figure

Hot bond-loops on the lattice map to cold domain walls on the dual lattice — the same closed curves

The deeper picture is a reciprocity between order and disorder. The spin operator, whose average is the magnetization, is ordered in the cold phase and disordered in the hot one. Its dual partner — Kadanoff and Ceva's disorder operator $\mu$ , nonlocal in the spins — is ordered in the hot phase and disordered in the cold. They are conjugate variables, like position and momentum: you cannot make both sharp at once, and the temperature at which their roles exchange is the critical point. This order–disorder reciprocity is the seed of every strong–weak duality in modern physics.

The core equations

The whole construction rests on rewriting one partition function two ways. We start from the square-lattice Ising model in zero field, then expand it once for hot and once for cold. The model is

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

where: $s_i = \pm1$ is the Ising spin at site $i$ ; $\langle ij\rangle$ runs over nearest-neighbor bonds; $K = \beta J$ is the dimensionless coupling (large when cold, small when hot); $Z$ is the canonical partition function. Everything below is a way of counting the terms in this sum by their loop content. Because each product $s_i s_j$ is only ever $\pm1$ , the Boltzmann weight of a single bond collapses from an exponential into a two-term polynomial — the algebraic key that turns thermodynamics into graph-counting:

e^{K s_i s_j} = \cosh K\,\big(1 + s_i s_j \tanh K\big) \qquad (2)

where: $\tanh K$ is the high-temperature expansion variable, small at high $T$ ; $\cosh K$ is a spin-independent prefactor carried by every bond. Reading it physically: the "1" is the bond doing nothing, and the " $s_i s_j \tanh K$ " is the bond being activated, contributing a factor $\tanh K$ and a spin pair that the sum over spins will police. Multiplying (2) over all bonds and summing over spins keeps only graphs in which every site is touched an even number of times, because $\sum_{s=\pm1} s^n$ vanishes for odd $n$ . The survivors are closed loops:

Z = (\cosh K)^{N_b}\, 2^{N} \sum_{\Gamma} (\tanh K)^{|\Gamma|} \qquad (3)

where: $N_b = 2N$ is the number of bonds on the square lattice; $N$ the number of sites; $\Gamma$ a closed graph (every vertex of even degree); $|\Gamma|$ its bond count. On the square lattice the smallest closed graph is the four-bond plaquette, contributing $v^4$ with $v = \tanh K$ — the first correction to the disordered free energy. Now expand the cold magnet instead. Measure energy from the aligned ground state; flipping a cluster of spins breaks exactly the bonds on its boundary, each costing $2J$ , so the excitations are closed domain walls on the dual lattice:

Z = e^{K N_b}\, \sum_{\Gamma^{*}} e^{-2K|\Gamma^{*}|} \qquad (4)

where: $\Gamma^{*}$ is a closed domain-wall configuration on the dual lattice; $|\Gamma^{*}|$ its total length; $e^{K N_b}$ the ground-state weight (all bonds satisfied). Physically: at low temperature the only thing happening is the occasional flipped droplet, and its Boltzmann cost is $e^{-2K}$ per unit of boundary — the low-temperature counterpart of $\tanh K$ . Comparing (3) and (4), both are sums over closed loops weighted by a per-length fugacity. Matching the fugacity of one expansion to the other defines the dual coupling:

\tanh K^{*} = e^{-2K} \quad\Longleftrightarrow\quad \sinh 2K\,\sinh 2K^{*} = 1 \qquad (5)

where: $K^{*}$ is the dual coupling. The left form says "the hot loops of the dual equal the cold walls of the original"; the right, symmetric form makes the involution manifest — swapping $K \leftrightarrow K^{*}$ leaves it unchanged, so the map is its own inverse. A cold, ordered model is dual to a hot, disordered one. The fixed point of the involution is the only candidate for a self-dual critical point, and solving it is a one-line quadratic:

\sinh 2K_c = 1 \;\Rightarrow\; K_c = \tfrac{1}{2}\ln(1+\sqrt{2}) \approx 0.4406868, \quad \frac{k_B T_c}{J}=\frac{2}{\ln(1+\sqrt2)}\approx 2.2692 \qquad (6)

where: $K_c$ is the critical coupling and $T_c$ the critical temperature. Equivalently the high-temperature variable takes the value $v_c = \tanh K_c = \sqrt2 - 1 \approx 0.4142$ at the transition. Finally, equating the two full expansions relates the free energy per site at dual couplings — a functional equation whose non-analytic part must be symmetric under the map:

f(K) = \tfrac12\ln\!\big(\tfrac12\sinh 2K\big) + f(K^{*}) \qquad (7)

where: $f = \lim_{N\to\infty} N^{-1}\ln Z$ is the dimensionless free energy per site. The first term is an explicit, analytic prefactor; the singular part obeys $f_{\rm sing}(K) = f_{\rm sing}(K^{*})$ , which is what pins any lone singularity to $K = K^{*}$ .

Derivation ALocating the critical point from self-duality

Our goal is to turn the pure symmetry $K \leftrightarrow K^{*}$ into a single equation for the critical temperature, assuming only that the model has one phase transition. The strategy is to equate the hot and cold expansions of $Z$ , extract the smooth analytic prefactor, note that the leftover singular part must be shared between $K$ and its dual, and then argue that a lone singularity has nowhere to sit except the fixed line $K = K^{*}$ . The one thing to watch is that this reasoning locates a transition but does not by itself prove one exists — that is the extra physical input, justified after the fact by Onsager.

Step 1. Write $Z$ two ways. The high-temperature form is the left side, the low-temperature form the right:

(\cosh K)^{N_b}2^N\!\sum_\Gamma (\tanh K)^{|\Gamma|} = e^{KN_b}\!\sum_{\Gamma^{*}} e^{-2K|\Gamma^{*}|}

Why: on the self-dual square lattice the loop sums $\sum_\Gamma$ and $\sum_{\Gamma^{*}}$ are combinatorially identical — the same closed curves counted on two overlaid copies of the lattice.

Step 2. Identify the per-length fugacities to make the two loop sums literally the same generating function:

\tanh K^{*} = e^{-2K}

Why: matching the dual's high- $T$ weight $\tanh K^{*}$ to the original's low- $T$ wall weight $e^{-2K}$ is what defines the dual temperature.

Step 3. Separate the smooth prefactors from the singular content:

f_{\rm sing}(K) = f_{\rm sing}(K^{*})

Why: the prefactors $(\cosh K)^{N_b}$ , $2^N$ , $e^{KN_b}$ are analytic in $K$ , so all non-analytic behavior of the free energy is shared between dual couplings.

Step 4. Impose a single transition. If exactly one coupling $K$ is singular, it must be its own image, or it would have an unwanted twin at $K^{*} \ne K$ :

K = K^{*}

Why: singularities come in dual pairs; only a self-dual point is unpaired, so a unique transition is forced onto the mirror line.

Step 5. Solve the fixed-point condition. Setting $K = K^{*}$ in the symmetric relation $\sinh 2K\,\sinh 2K^{*} = 1$ gives $\sinh^2 2K_c = 1$ , hence

\sinh 2K_c = 1

taking the positive root (the coupling is positive).

Step 6. Invert to a number. Writing $\sinh 2K_c = \tfrac12(e^{2K_c} - e^{-2K_c}) = 1$ is a quadratic in $x = e^{2K_c}$ , namely $x^2 - 2x - 1 = 0$ , whose positive root is $x = 1 + \sqrt2$ :

K_c = \tfrac12\ln(1+\sqrt2) \approx 0.44069, \qquad \frac{k_BT_c}{J} = \frac{1}{K_c} = \frac{2}{\ln(1+\sqrt2)} \approx 2.26919

Why: this is a real, closed-form critical temperature obtained from symmetry alone — no transfer-matrix spectrum, no exact free energy. *Verified: `verify_aq-sm-kramers-wannier-duality.py` confirms $\sinh 2K_c = 1$ symbolically, $K_c = 0.4406868$ , $k_BT_c/J = 2.269185$ , and $v_c = \tanh K_c = \sqrt2 - 1$ , all matching Onsager's exact solution to full precision.*

Go deeper · Derivation B · 1 minThe symmetric involution from the loop-matching condition

Derivation BThe symmetric involution from the loop-matching condition

Our goal is to convert the asymmetric-looking matching condition $\tanh K^{*} = e^{-2K}$ into the manifestly symmetric form $\sinh 2K\,\sinh 2K^{*} = 1$ , thereby proving the duality map is a true involution — its own inverse. The route is a short exercise with hyperbolic double-angle identities: express $\sinh 2K^{*}$ in terms of $\tanh K^{*}$ , substitute the matching condition, and simplify until the original coupling $K$ appears only through $\sinh 2K$ . Watch that the algebra stays exact; no small-parameter approximation is used anywhere.

Step 1. Start from the loop-matching condition:

\tanh K^{*} = e^{-2K}

Why: this is the identification of the dual high-temperature weight with the original low-temperature wall weight, from the previous derivation.

Step 2. Express the dual $\sinh 2K^{*}$ through $\tanh K^{*}$ using $\sinh 2\theta = 2\tanh\theta/(1-\tanh^2\theta)$ :

\sinh 2K^{*} = \frac{2\tanh K^{*}}{1-\tanh^2 K^{*}} = \frac{2e^{-2K}}{1-e^{-4K}}

Why: this rewrites everything on the dual side in terms of the original coupling $K$ .

Step 3. Simplify by multiplying numerator and denominator by $e^{2K}$ :

\frac{2e^{-2K}}{1-e^{-4K}} = \frac{2}{e^{2K}-e^{-2K}} = \frac{1}{\sinh 2K}

Why: $e^{2K} - e^{-2K} = 2\sinh 2K$ , so the whole expression collapses to the reciprocal of $\sinh 2K$ .

Step 4. Read off the symmetric relation:

\sinh 2K^{*} = \frac{1}{\sinh 2K} \quad\Longrightarrow\quad \sinh 2K\,\sinh 2K^{*} = 1

Why: the product is symmetric under $K \leftrightarrow K^{*}$ , so applying the map twice returns the original coupling — an involution with fixed point $\sinh 2K_c = 1$ . *Verified: `verify_aq-sm-kramers-wannier-duality.py` checks $\tanh K^{*} = e^{-2K}$ is solved by $K^{*} = \operatorname{atanh}(e^{-2K})$ , confirms $\sinh 2K\,\sinh 2K^{*} = 1$ symbolically, and verifies $(K^{*})^{*} = K$ (the involution) numerically over a grid of couplings.*

Limiting cases

The single map $\sinh 2K\,\sinh 2K^{*} = 1$ governs every regime, and reading off its limits is the fastest sanity check on the whole construction.

Case	Result	Meaning
High-temperature limit	$K\to 0 \;\Rightarrow\; K^{*}\to\infty$	A nearly disordered hot model is dual to a deeply frozen cold one — weak coupling $\leftrightarrow$ strong coupling.
Low-temperature limit	$K\to\infty \;\Rightarrow\; K^{}\to 0$ , with $e^{-2K}=\tanh K^{}$	Each domain-wall bond of weight $e^{-2K}$ maps to a hot bond of weight $\tanh K^{*}$ .
Self-dual point	$K=K^{*} \;\Rightarrow\; \sinh 2K_c=1,\; K_c=\tfrac12\ln(1+\sqrt2)\approx 0.4407$	Fixed point of the involution; the unique candidate for criticality.
Triangular / honeycomb	dual lattices of each other, not self-dual: $\sinh 2K_c=1/\sqrt3$ (triangular), $\cosh 2K_c=2$ (honeycomb)	Duality plus the star–triangle relation still locates $K_c$ even without self-duality.

Worked examples

1 · The critical temperature from the fixed point (confidence builder)

Compute the exact square-lattice Ising critical coupling $K_c$ , the temperature $k_B T_c/J$ , and the high-temperature variable $v_c = \tanh K_c$ from the self-dual condition $\sinh 2K_c = 1$ .

Step 1. Write the condition explicitly: $\sinh 2K_c = \tfrac12(e^{2K_c} - e^{-2K_c}) = 1$ . Multiply through by $2e^{2K_c}$ to clear the negative exponential: $e^{4K_c} - 2e^{2K_c} - 1 = 0$ .

Step 2. This is a quadratic in $x = e^{2K_c}$ : $x^2 - 2x - 1 = 0$ , so $x = \dfrac{2 + \sqrt{4+4}}{2} = 1 + \sqrt2$ (taking the positive root).

Step 3. Solve for the coupling: $2K_c = \ln(1+\sqrt2)$ , hence $K_c = \tfrac12\ln(1+\sqrt2) = \tfrac12\ln(2.41421) = \tfrac12(0.881374) = \boxed{0.440687}$ .

Step 4. Temperature: $k_B T_c/J = 1/K_c = 2/\ln(1+\sqrt2) = 2/0.881374 = \boxed{2.26919}$ .

Step 5. High-temperature variable: $v_c = \tanh K_c$ . Since $\cosh 2K_c = \sqrt{1 + \sinh^2 2K_c} = \sqrt2$ and $\tanh K_c = \sinh 2K_c/(1 + \cosh 2K_c) = 1/(1+\sqrt2) = \sqrt2 - 1 = \boxed{0.41421}$ . That single number, $k_B T_c/J \approx 2.269$ , is what Kramers and Wannier extracted from symmetry alone in 1941 — and it agrees with Onsager's 1944 exact solution to every digit.

2 · A non-self-dual lattice: the triangular Ising critical point (exam level)

The triangular lattice is not self-dual — its dual is the honeycomb lattice. Using the star–triangle relation, the triangular critical coupling satisfies $\sinh 2K_c = 1/\sqrt3$ . Find $K_c$ and $k_B T_c/J$ , and explain why $T_c$ is higher than on the square lattice.

Step 1. Solve $\sinh 2K_c = 1/\sqrt3 = 0.57735$ . Then $2K_c = \operatorname{arcsinh}(0.57735) = \ln(0.57735 + \sqrt{1 + 1/3}) = \ln(0.57735 + 1.15470) = \ln(1.73205)$ .

Step 2. Note $\ln(1.73205) = \ln\sqrt3 = \tfrac12\ln 3 = 0.549306$ , so $K_c = \tfrac12(0.549306) = \boxed{0.274653}$ .

Step 3. Temperature: $k_B T_c/J = 1/K_c = 1/0.274653 = \boxed{3.6410}$ .

Step 4. Compare to the square lattice ( $k_B T_c/J = 2.269$ ). The triangular lattice has coordination number $z = 6$ versus $z = 4$ for the square lattice. Each spin has more aligned neighbors reinforcing it, so a higher temperature is needed to disorder it — exactly the trend $K_c^{\rm tri} = 0.275 < K_c^{\rm sq} = 0.441$ , i.e. a smaller coupling suffices at the same physical $T_c$ , so the transition temperature rises.

Step 5. Sanity check via the honeycomb dual, which obeys $\cosh 2K_c = 2$ : $2K_c = \operatorname{arccosh}(2) = \ln(2 + \sqrt3) = 1.31696$ , giving $K_c^{\rm hex} = 0.65848$ and $k_B T_c/J = 1.5185$ . The honeycomb ( $z = 3$ ) has the lowest $T_c$ of the three, completing the coordination-number trend $z = 3 < 4 < 6 \Rightarrow T_c$ increasing.

Origins · 1 minGregory Wannier and how this was discovered

Gregory Wannier

1911–1983 · Basel, then Bell Labs and Oregon

Gregory Wannier was a Swiss-American physicist who left an outsized mark on solid-state theory. As a young researcher he introduced what are now called Wannier functions — spatially localized counterparts to Bloch waves that remain a workhorse of modern electronic-structure theory — and analyzed the "Wannier–Stark ladder" of electrons in a crystal under an electric field. His 1941 paper with Hendrik Kramers on the Ising model is a masterpiece of getting a hard exact result from a soft symmetry argument: without solving the model, they located its critical point by recognizing that its high- and low-temperature expansions are mirror images. The duality they discovered became the template for order–disorder dualities across statistical mechanics and field theory. Wannier spent his career at Bell Labs and the University of Oregon, and his textbook Statistical Physics is prized for its clarity.

History

1936

— Rudolf Peierls argues, via domain-wall counting, that the 2D Ising model must order at low temperature — establishing that a transition exists.

1941

— Kramers and Wannier publish the duality and extract the exact critical temperature

k_B T_c/J = 2/\ln(1+\sqrt2)

from the self-dual point alone.

1944

— Onsager's exact transfer-matrix solution confirms the Kramers–Wannier

T_c

and delivers the full thermodynamics.

1971

— Wegner generalizes duality to

\mathbb{Z}_2

lattice gauge theory in three dimensions, introducing gauge-invariant order parameters.

1971

— Kadanoff and Ceva formalize the disorder operator dual to the spin, making the order–disorder reciprocity precise.

Going further

References. Kramers, H. A. & Wannier, G. H., Statistics of the Two-Dimensional Ferromagnet, Part I, Phys. Rev. 60, 252 (1941) · Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 13 · Kardar, Statistical Physics of Fields, Ch. 7 · Savit, R., Duality in field theory and statistical systems, Rev. Mod. Phys. 52, 453 (1980) · Baxter, Exactly Solved Models in Statistical Mechanics, Ch. 6.

Provenance. The critical coupling $K_c = \tfrac12\ln(1+\sqrt2) \approx 0.4406868$ and $k_B T_c/J \approx 2.2692$ are standard (Kramers–Wannier 1941; Onsager 1944); a verifier should recompute both from $\sinh 2K_c = 1$ , along with $v_c = \tanh K_c = \sqrt2 - 1$ . The duality relation $\sinh 2K\,\sinh 2K^{*} = 1$ and the two loop expansions follow Kardar Ch. 7 and Pathria & Beale Ch. 13. The triangular ( $\sinh 2K_c = 1/\sqrt3$ , $k_B T_c/J = 3.641$ ) and honeycomb ( $\cosh 2K_c = 2$ , $k_B T_c/J = 1.519$ ) critical points are standard (Baxter Ch. 6). All numbers are reproduced in `verify_aq-sm-kramers-wannier-duality.py`.

Prerequisites. The Ising Model: Definition and Hamiltonian · High- and Low-Temperature Series Expansions · Domain Walls and Interfacial Excitations · Transfer Matrix Formalism for Lattice Models. Leads to. Onsager Solution of the Two-Dimensional Ising Model · Peierls Argument for Long-Range Order · Order Parameter and Spontaneous Magnetization.

If duality is exact, why did anyone need Onsager's solution at all?

Duality is exact but incomplete: it tells you where the singularity of the free energy sits (the self-dual point), assuming there is exactly one. It says nothing about the nature of that singularity — the exponents, the shape of the specific-heat divergence, the magnetization. Onsager's 1944 transfer-matrix solution supplied all of that: the logarithmic specific heat, $\beta = 1/8$ , and the full free energy. Duality gives you one number cheaply; the exact solution gives you the whole thermodynamics.

Why does the single-transition assumption matter so much?

Because singularities of the free energy come in dual pairs $(K, K^{*})$ . If the model had two transitions, generically they would sit at a pair of dual couplings, and duality would relate them but not pin either one. Only a self-dual singularity, at $K = K^{*}$ , is unpaired and therefore forced to a definite location. The assumption of a single transition is what makes that self-dual point the unique candidate. It is justified after the fact: Onsager found exactly one transition, exactly there.

What exactly is the dual lattice, and why is the square lattice special?

The dual lattice places a site at the center of each plaquette of the original and connects two dual sites whenever their plaquettes share a bond. Domain walls (boundaries of flipped regions) live on this dual lattice. For the square lattice the dual is again a square lattice, so the low-temperature walls sit on an identical copy — the model is self-dual. Triangular and honeycomb lattices are duals of each other, not themselves, so duality relates their two different critical points rather than fixing one self-consistently.

How can a hot phase and a cold phase be "the same"?

They are not the same physically — one is disordered, one ordered. Duality says their partition functions are related by an exact analytic map, because the same closed-loop combinatorics counts hot bond-graphs and cold domain walls. What is genuinely exchanged is order and disorder: the spin, ordered when cold, is dual to the disorder operator $\mu$ , which is ordered when hot. So duality is a statement about the mathematical structure of $Z$ and about which operator carries the order, not a claim that hot equals cold.

Is there a factor-of-two subtlety I should watch in the dual coupling?

Yes. The dual coupling is $K^{*} = \operatorname{atanh}(e^{-2K})$ , which equals $-\tfrac12\ln\tanh K$ — note the one-half. A common slip writes $K^{*} = -\ln\tanh K$ (dropping the $\tfrac12$ ), which is off by a factor of two and corrupts any quantity built from $K - K^{*}$ , such as the interfacial surface tension. Always define $K^{*}$ through $\tanh K^{*} = e^{-2K}$ or, symmetrically, $\sinh 2K\,\sinh 2K^{*} = 1$ , and you will not go wrong.

Does duality work in three dimensions?

Not in the same clean form. The 3D Ising model is dual to a $\mathbb{Z}_2$ lattice gauge theory, not to another Ising model — this is Wegner's 1971 result. So there is no self-dual point pinning the 3D Ising $T_c$ , and its critical temperature must be found by series, Monte Carlo, or the conformal bootstrap. The order–disorder logic survives, but the object dual to the 3D Ising model is a gauge theory with a different order parameter (the Wilson loop), which is why 3D critical data cannot be gotten "for free."

Why is the critical high-temperature variable v_c = √2 − 1 rather than a rounder number?

Because it is the value of $\tanh K$ at the self-dual point. Solving $\sinh 2K_c = 1$ gives $e^{2K_c} = 1 + \sqrt2$ , and then $\tanh K_c = (e^{2K_c}-1)/(e^{2K_c}+1) = \sqrt2/(2+\sqrt2) = \sqrt2 - 1 \approx 0.4142$ . The irrational value is a direct fingerprint of the quadratic $x^2 - 2x - 1 = 0$ that the self-dual condition produces. It is also the radius of convergence of the high-temperature susceptibility series, which is why series analysts watch for it.