A phase transition with no order parameter, no broken symmetry and no ordered phase, in which the defects turn out to be the physics. Page 1 of the critical-lines series.
Every phase transition you were taught works the same way: a symmetry breaks, an order parameter turns on, and something diverges as a power law. This one does none of that. There is no order parameter, no symmetry is broken, and a theorem forbids the ordered phase from existing at all. Yet there is a sharp, experimentally confirmed transition, with a parameter-free prediction that was measured and found correct.[3] It is also the most treacherous critical line there is for an instrument that reads the renormalization-group flow, which is why it opens this series.
The Compass’s Detector gives one of four verdicts: three describe what the renormalization-group flow does, the fourth (REFUSE) says the data cannot tell, and this page is about the one place where they are hardest to tell apart.
The panels above follow a point as it is coarse-grained: a picture of the renormalization-group flowCoarse-grain a system (average over blocks, rescale) and track how its effective couplings change. The couplings trace a path: the flow. A fixed point is a set of couplings that coarse-graining leaves unchanged.. The strip under each panel is the speed of the flow. The Criticality CompassAn instrument under design with two parts: a Detector that classifies a system’s critical geometry from observations, and a Navigator that finds a critical line and moves along it. is an instrument under design; its Detector reads observations of a system and returns one of four verdicts:[18]
The Detector follows the flow over two kinds of relation: SCALE (a system against its own coarse-grained copy: level n against n + 1, or size L/2 against L) and CONTROL (runs at neighbouring values of a knob such as the temperature). A breaker is a test case chosen to fool the instrument: a good failure on it is REFUSE, a bad one is a confident wrong verdict. Each reading of a flow velocity is sorted by a five-valued sign (SIG+, SIG−, ZERO!, SLOW, ?), taught on the Techniques page, §3; applied to BKT in §12 below.
Each section has three layers. The picture and the prose are open. Working equations (a closed box) holds the equations a working physicist uses. Full derivation, nested inside it, holds the step-by-step argument. Every section ends with a cyan box saying what it means for the Criticality CompassAn instrument under design with two parts: a Detector that classifies a system's critical geometry from observations (POINT / LINE / WALKING / REFUSE), and a Navigator that finds a critical line and moves along it. and a short Check yourself quiz. The box under the introduction, “The Compass in one minute”, collects every Compass term in one place. The final test at the bottom draws different questions on each attempt; your best and latest scores stay in this browser, and an Export button gives a short text summary you can paste anywhere.
BKT is on the syllabus because it is the LINE that ends treacherously: near its end the distance to the end shrinks only as 1/ℓ (the per-level velocity as 1/ℓ²), so a finite system can look like WALKING. It is the Detector’s deliberate breaker, and the good answer there is REFUSE. ◆ design
In two dimensions, at any temperature above zero, a continuous symmetry cannot break: long-wavelength fluctuations always win.
Take the two-dimensional XY modelUnit spins confined to a plane, one angle θ per site. Its symmetry, rotating every angle by the same amount, is continuous.: planar spins on a square lattice, each described by one angle, with H = −J ∑⟨ij⟩ cos(θi − θj). Step through the trace. The mean-square wobble of one spin’s angle grows like the logarithm of the sample’s size, so however cold the system is, a big enough sample loses its alignment. The magnetisation dies too, as a small power of 1/L. That is slow: at kBT = 0.5J the sample must be tens of millions of lattice spacings across before half the order is gone. But it is without bound.
A continuous symmetry cannot be spontaneously broken at any temperature T > 0 in d ≤ 2 dimensions with short-range interactions.[1]
So the model should be boring: disordered everywhere, no transition. Except that thin superfluid films and two-dimensional magnets show something sharp happening at a definite temperature. Both facts are true. Resolving that is the whole subject.
The Detector never sees “order” here, because there is none to see. It sees flow: how the effective couplings move under coarse-graining. That is the only reason BKT is within its reach at all.
At low temperature, expand the cosine. The Gaussian (spin-wave) model
H ≈ (J/2) ∫ d²r |∇θ|²has ⟨|θk|²⟩ = kBT/(Jk²) for each Fourier mode, and so
⟨θ²⟩ ≈ (kBT/2πJ) ln(L/a), m = ⟨cos θ⟩ = e−⟨θ²⟩/2 = (a/L)kBT/4πJ → 0.The theorem itself is an inequality (Bogoliubov’s) and does not rely on the Gaussian approximation; the approximation shows the mechanism.[1] This page uses one convention throughout; the 1973 paper prints some coefficients in another.[20]
With θ(r) = L−1 ∑k θk eik·r, the energy is (J/2) ∑k k² |θk|², a sum of independent Gaussian modes, so equipartition gives ⟨|θk|²⟩ = kBT/(Jk²). Then
⟨θ(r)²⟩ = (kBT/J) ∫ d²k/(2π)² · 1/k² = (kBT/2πJ) ∫π/Lπ/a dk/k = (kBT/2πJ) ln(L/a).The integrand d²k/k² = 2π dk/k is exactly scale-free in two dimensions: every octave of wavelength contributes the same amount. In three dimensions it is d³k/k² ∝ k dk, convergent at small k, and order survives. For a Gaussian variable ⟨cos θ⟩ = e−⟨θ²⟩/2, which gives the power law for m.
Correlations decay as a power law whose exponent changes smoothly with temperature; a power law looks the same after coarse-graining, so each temperature is its own RG fixed point, and together they form a line.
Compute the correlation between two spins anyway. It is a power law, not an exponential: correlations do not survive to infinity, but they do not die off at any finite length either. And the exponent η(T) varies continuously with temperature. Drag T in the trace: the straight line on the log–log plot tilts, and it stays straight.
In renormalization-group language: coarse-graining by a factor b rescales r to r/b and multiplies a power law by the constant b−η, which a rescaling of the spins absorbs. So the correlations at temperature T are unchanged by the flow: T is a fixed point, with its own η(T). One fixed point per temperature is a line of fixed points, and it is what makes BKT structurally unlike anything in the standard catalogue. Nothing has to be tuned; criticality is the generic condition of the whole phase.[6] But the spin-wave calculation assumed that θ is a smooth, single-valued field. It is not.
This is the LINE: codimension zero, criticality without tuning. A Detector reading exponents sees them move with the CONTROL (temperature) but not with SCALE: at a fixed T, coarse-graining leaves η alone.
With vortices included the bare J is replaced by the renormalised stiffness. Nelson and Kosterlitz write it in superfluid variables as η(T) = m²kBT / 2πℏ²ρs(T), i.e. η = 1/2πKR with KR = ℏ²ρs/m²kBT. This is the XY model’s KR = JR/kBT in superfluid units: the stiffness ℏ²ρs/m² plays the role of the renormalised coupling JR. (Here m is the helium atom’s mass, not the magnetisation of §1; Kosterlitz and Thouless write m* for its effective value in a film, §4.)[3] At the transition KR = 2/π, so η(TKT) = 1/4 exactly (§7).
For a Gaussian field, ⟨ei(θ0 − θr)⟩ = e−⟨(θ0 − θr)²⟩/2 (Wick). The difference has variance (kBT/J)∫d²k/(2π)² · 2(1 − cos k·r)/k² ≈ (kBT/πJ) ln(r/a) for r » a, the modes with k < 1/r cancelling between the two points. Half of that is (kBT/2πJ) ln(r/a), so the correlation is (a/r)kBT/2πJ. The 1973 paper prints this exponent as kBT/4πJ (its eq. 53), a convention factor of two.[20]
The angle lives on a circle, so going once round a loop it can come back having turned by a whole number of turns, the winding number; a vortex is a point that a loop encloses with winding +1, and that integer cannot change smoothly.
This is a live XY Monte Carlo lattice; colour is the spin angle. It starts cold, with one vortexA point round which the angle winds once, by +2π. All colours meet there. and one antivortex planted. Choose Draw a loop and click two corners: the page walks the loop, adds up the phase differences between neighbours (each wrapped into (−π, π]) and reports the total in turns. Round the vortex you get +1; round the antivortex −1; round both, 0. Then heat it (the T slider, and Run or Step) and watch vortex–antivortex pairs appear everywhere. The integer is π1(S1) = ℤ, the same invariant that classifies superfluid vortices.[2]
The toy has periodic edges and a seeded random number generator; it is for looking, and nothing on this page reports a number measured with it. Note one thing it shows honestly: the count of winding plaquettes rises smoothly with T. A 2024 lattice study states it flatly: “The density of vortices in itself is not a good indicator of the KT transition.”[13] What changes at TKT is not whether vortices exist but whether they are bound.
A winding number is an integer: it cannot drift, so it needs little data. In the kernel’s language the raw phase differences along the bonds are directed differences (A over the declared bond relation, with each bond’s orientation declared); wrapping each into (−π, π] is a further declared transform, since A is only x[s] − x[t]. Their sum round a loop is a line sum, not the kernel’s R. The kernel’s declaration rules exclude one shortcut: a closed loop is a ring, and rings are not admissible as declared relations, so the loop sum must be declared as a projection (it keeps the integer winding and discards the bond values), not smuggled in as an edge. In plain terms: counting a winding is allowed, but it has to be written down as its own declared step, not assembled from the operators. ◆ design
Round any closed lattice contour ∑ Δφ(r) = 2πq, q = 0, ±1, ±2, …, the total vortex strength enclosed (Kosterlitz & Thouless eq. 45). Split φ = ψ + φ̃ into a smooth spin-wave part (∑Δψ = 0) and a vortex part (∑Δφ̃ = 2πq); the energy separates, H − E0 = J∑(Δψ)² + J∑(Δφ̃)², with no cross term (their eq. 46). One vortex has energy πJ ln(R/a) (their eq. 44).[2]
Kosterlitz and Thouless also explain why the isotropic Heisenberg model (spins on a sphere) has no such transition: a 2π twist in the plane can be undone through the third direction, and the one remaining invariant, the skyrmion number, costs only a finite, size-independent energy.[2]
For one vortex, θ = the polar angle about its core, so |∇θ| = 1/r and
E = (J/2)∫aR (1/r)² 2πr dr = πJ ln(R/a).The integral diverges at large R: a single free vortex costs an energy that grows with the system. A vortex–antivortex pair at separation r has fields that cancel beyond r, so its energy is finite, about 2πJ ln(r/a). That is why pairs appear at any T > 0, and why the question is only whether they come apart.
The energy of one free vortex and the entropy of where to put it both grow like ln L, so the system size cancels and a pure number decides which wins.
This is the heart of the subject and it fits on a napkin: Berezinskii in 1971–72, independently and first, and Kosterlitz and Thouless in 1973.[2] Put one vortex in a system of size L. Its energy grows like ln(L/a): a single vortex looks forbidden. But its core can sit at any of about (L/a)² places, so its entropy grows like ln(L/a) too, with the same functional form. Step through the trace and then drag T. The free-energy line pivots through zero, and at that temperature free vortices stop costing anything.
Below kBT = πJ/2 a free vortex costs an infinite free energy and vortices survive only in bound pairs. Above it, entropy wins and free vortices proliferate.
The right-hand panel of the trace shows why this is only an estimate. The napkin number is kBT = πJ/2 ≈ 1.571J. The measured transition of the square-lattice XY model is at 0.8929J.[10][11] Kosterlitz and Thouless said so themselves: other pairs “will relax in the field of the first pair thereby renormalizing the rigidity modulus … downwards and consequently reducing the critical temperature.”[2]
The whole argument is about how things change with SCALE (ln L), which is exactly the axis the Detector reads. The lesson in the gap between 1.571 and 0.893: a bare estimate is not the fixed point, and the flow (§6) is what moves one to the other.
As Kosterlitz and Thouless print it, for a dislocation in a crystal (Burgers vector b, 2D rigidity modulus v, Poisson ratio τ, area A):[2]
E = [vb²(1+τ)/4π] ln(A/A0), S = kB ln(A/A0) + O(1), kBTc = vb²(1+τ)/4π(their eqs 1–3), with the statement that these are upper bounds. “Identical considerations for the xy model give” their eq. 4, printed as kBTc = πJ. In the convention used on this page (H = −J∑cos, the same paper’s own vortex energy πJ ln(R/a), and S = 2kB ln(R/a)) the argument gives πJ/2; the printed coefficient differs by a convention factor.[20]
For the neutral superfluid they give kBTc = πℏ²ρ/2m (eq. 5), and for a film kBTc = πℏ²ρs/2m* (eq. 82), with ρs the superfluid density per unit area.
Energy: πJ ln(L/a) + Ecore (§3’s L1 with R = L). Entropy: the number of places for the core is about (L/a)², so S = kB ln(L/a)² = 2kB ln(L/a). Then F = E − TS = (πJ − 2kBT) ln(L/a) + Ecore. As L → ∞ only the sign of the bracket matters, and it vanishes at kBT = πJ/2. Setting η = kBT/2πJ there gives 1/4, which is not a coincidence: both numbers are the point where the vortex fugacity becomes marginal (§6).
Winding number behaves exactly like electric charge in two dimensions, and the transition is a dielectric turning into a conductor.
Two vortices with winding numbers ni, nj at distance r interact as Eij = −2πJ ninj ln(rij/a), which is the two-dimensional Coulomb law with winding number as charge.[2] So the XY model maps onto a neutral two-dimensional Coulomb gas, and BKT becomes a dielectric-to-conductor transition. Below it, charges are bound into dipoles that only polarise; above it, free charges screen. In the trace, raise T and watch the mean pair size run away when its denominator reaches zero.
A second language for the same flow: nothing new for the Detector, but a vocabulary to recognise (dielectric constant = inverse stiffness, pair = dipole, screening = free vortices).
Kosterlitz and Thouless’s model system: charges ±q with U(r) = −2qiqj ln(r/r0) + 2μ for r > r0 (their eq. 6), 2μ the energy of a pair at the cutoff r0. Ignoring the other pairs,
⟨r²⟩ = r0² (βq² − 1)/(βq² − 2)(their eq. 7), which diverges at βq² = 2: “a phase transition to a conducting state will take place” at kBTc ≈ q²/2 (eq. 10). For the XY vortex gas a ±1 pair costs 2πJ ln(r/a), which matches the model’s 2q² ln(r/r0) for q² = πJ, and gives πJ/2 again (the paper’s own bookkeeping differs by the convention factor).[2][20]
The paper’s real step is self-consistent screening. A pair of size r is screened by all smaller pairs, giving a scale-dependent dielectric constant ε(r). Write ℓ = ln(r/r0) for the scale and u(ℓ) = 2βq²/ε(ℓ) − 4 for the running stiffness (the paper calls them x and y; on this page y is reserved for the fugacity of §6). They obtain (eq. 19)
du/dℓ = −π² e−2βμ (u + 4)² e−ℓu,which rescales to a parameter-free equation, in the paper’s own letters dỹ/dx̃ = −exp(−x̃ỹ) (eq. 21). Its solutions fall into two classes: u(∞) ≥ 0, a finite macroscopic dielectric constant (T < Tc, one end-point per temperature: a line), or u → −∞ (T > Tc, free charges). This is the ancestor of the recursion relations of §6, written as a flow in ln r before the name RG was attached to it. The paper derives an essential singularity from it (eq. 27) and says frankly that “since we have used a mean field theory, this form of the singularity is unlikely to be the correct one.”[2]
Two couplings, stiffness and fugacity, flow into a whole line of fixed points, and the line ends exactly where the transition is.
Drag the white handle. The arrows are the flow field; the thick curve is the trajectory from your start; the dashed curve is the hyperbola E = y² − x² through it, which the flow can never leave. Here x = πK − 2 measures how far the running stiffness K (the renormalised J/kBT) is above the critical value 2/π, and y, the vortex fugacity, measures how cheap vortices are. (Hasenbusch, Marcu and Pinn write x = πβ − 2, with β the coupling of the dual Gaussian model; on this page β always means 1/kBT.) The scale variable is t = ln b, the logarithm of the coarse-graining factor; Nelson and Kosterlitz call it ℓ.[9] The separatrices are the lines y = x and y = −x; they cut the upper half-plane into three regions. Region I is below y = x on the right (x > 0, y < x): the fugacity dies and the trajectory lands on the red fixed line (y = 0) at x∞ = √(x² − y²), a different point for every start. Region II is above both separatrices (y > |x|): y grows without bound and vortices take over. Region III is below y = −x on the left (x < 0, y < −x): y also grows without bound. The branch x = y is the critical trajectory, and it runs into the end of the line, the point (0, 0).
The second trace takes the two equations apart: the fugacity equation alone, the stiffness equation alone, the combined field, the conserved hyperbolas, and finally the approach along the critical trajectory, y(t) = 1/(1/y0 + t). That last curve is the treachery. An irrelevant coupling near an ordinary fixed point dies exponentially in the scale time t = ln b. On the critical trajectory the fugacity dies only as 1/t: after a factor of 1040 in scale it has still not gone.
This is what “the flow stops on a line” looks like, and what happens at the line’s end. In region I (x > 0, y < x) the velocity drops to zero exponentially: ZERO! within any declared margin, a LINE. On the critical trajectory the velocity falls as 1/t², and the five-valued sign of §12 reads it SIG, then SLOW, then ?. That sequence, SIG → SLOW → ?, is a named cell of the Detector’s table, and its verdict is REFUSE(“slow”). ◆ design
Hasenbusch, Marcu and Pinn write the flow (their eqs 22–24), with t = ln B the scale:
dy/dt = −xy, dx/dt = −y², x = πK − 2, y = cz (c a non-universal constant), E = y² − x² conserved.On the critical branch x = y, y(t) = 1/(1/y0 + t) (their eq. 60). For E > 0, with ε = √E, x(t) = ε tan(arctan(x0/ε) − εt) (eq. 69).[9] The same flow in the couplings (yK, yV) of Ueda and Oshikawa is dyK/dl = −yV², dyV/dl = −yVyK, with the BKT line yK = yV; they extracted (yK, yV) from transfer-matrix spectra and watched the trajectories follow the hyperbolas, with the T = 0.893 points lying on the BKT line.[11]
In the original variables, the Kosterlitz recursion relations as Nelson and Kosterlitz print them (their eq. 8), with K = J/kBT and y the vortex fugacity:[3]
dK−1/dℓ = 4π³y², dy/dℓ = (2 − πK) yThe fugacity changes sign of growth at πK = 2: that is where x = 0.
A resemblance, not a claim. The E > 0 branch, dx/dt = −(ε² + x²), has exactly the form of the walking beta function that Gorbenko, Rychkov and Zan write, dλ/dt = −y − λ², whose passage takes a scale time of order π/√y.[19] The functional forms match; nothing here says the two small parameters are the same thing.
Hasenbusch (2005) derives the flow from the sine-Gordon action (here β is the coupling of this dual model, the running stiffness K of the text, not 1/kBT) S = (1/2β)∑(φx − φx+μ)² − z∑cos(2πφx), the dual of the XY model: for β > 2/π the cosine is irrelevant. Near (x, z) = (0, 0): dz/dt = −xz, dx/dt = −const·z² (eqs 13–15).[10] On the critical trajectory x = const1/2z, so dx/dt = −x² and
x = 1/(ln l + C)(eq. 18). The next order, dx/dt = −x² − x³/2, adds corrections of order ln|ln L|/(ln L)².
Conservation: dE/dt = 2y(−xy) − 2x(−y²) = 0. On x = y the first equation becomes dy/dt = −y², which integrates to 1/y = 1/y0 + t. For E < 0 and x > 0 the trajectory reaches y = 0 at x∞ = √(−E): the approach there is exponential in t, with rate x∞, which vanishes as the start approaches the separatrix.
The correlation length diverges faster than any power, the stiffness jumps by a universal amount, and η = 1/4 exactly.
An essential singularity. Because the flow near the end of the line is marginal, the correlation length above TKT does not diverge as a power of the reduced distance from the transition, κ = |βc − β|/βc (on this page t is kept for the scale time of §6). It goes as ξ ∼ A exp(C/√κ).[9] The trace plots the square-lattice XY model’s own form, with the constants A = 0.223 and C = 1.78 that Hasenbusch, Marcu and Pinn extracted. Flip the axes: against κ−1/2 the points fall on a straight line; against ln κ they curve, and a power-law fit returns a “ν” that grows the closer you get.
Quoting a correlation-length exponent for this transition is a category error. The divergence is faster than any power law, and the transition is of infinite order: every derivative of the free energy is continuous through it. This is also the easiest experimental signature to check: plot ln ξ against κ−1/2 and look for a straight line.
The universal jump. At the transition the renormalised stiffness takes the value KR = 2/π exactly and then drops discontinuously to zero. For a superfluid film this is a statement about the superfluid density with nothing adjustable in it:[3]
ρs(TKT−) / TKT = 2m²kB / πℏ² (= 3.49 × 10−9 g cm−2 K−1 for ⁴He; Nelson and Kosterlitz print 3.52 × 10−9)The film’s thickness, substrate and disorder do not enter. Nelson and Kosterlitz framed it as a single number that could decide between competing theories; Bishop and Reppy measured it in 1978 with helium films on a torsional oscillator.[4] Twenty-eight years later Hadzibabic and co-workers saw the same jump in the coherence of a trapped two-dimensional atomic gas, coinciding with the onset of free vortices.[7]
“There is no ν”: a Detector that fits a power law here manufactures an exponent, and the exponent drifts with the fitting window. Drift with scale is WALKING’s signature, so an unwary Detector reads this LINE’s end as WALKING. That is the bad failure.
Correlation length (Hasenbusch, Marcu & Pinn eq. 25): ξ ∼ A exp(C|(βc − β)/βc|−1/2), A, C, βc non-universal; for the XY cosine model A = 0.223(13), C = 1.78(2).[9]
Universal jump (Nelson & Kosterlitz eq. 10): limT→Tc− m²kBT / ℏ²ρs(T) = π/2, i.e. KR → 2/π; and η(Tc) = 1/4. Below Tc they find a square-root cusp, ρs(T)/T ≈ [ρs(Tc)/Tc][1 + 0.50(1 − T/Tc)1/2], and stress that its amplitude is non-universal. The same 2/π holds for the Frank constants of a 2D nematic.[3]
At TKT the susceptibility carries a logarithmic correction, χ ∝ L2−η(ln L + C)−2r with η = 1/4 and r = −1/16 (Hasenbusch 2005, eqs 42–44). Fits without the C term had returned r ≈ −0.023 to −0.027; with it, r = −0.056(7) at Lmin = 256.[10]
The jump is the fixed-point condition. In dy/dℓ = (2 − πK)y the fugacity stops shrinking where πK = 2. Below Tc the flow ends on the fixed line with some KR ≥ 2/π; the critical trajectory ends exactly at its end, KR = 2/π; above Tc y grows and K is driven to zero. So KR(Tc−) = 2/π is where the separatrix meets the line, and no material parameter can enter it.[3]
The essential singularity. Just above Tc, E = ε² > 0 is small; if E is a smooth function of T that vanishes at Tc, then ε² ∝ κ. By eq. 69 the trajectory needs a scale time of order π/ε to pass the bottleneck near the origin and escape. The correlation length is the scale at which it escapes: ln ξ ∼ π/ε ∝ κ−1/2.[9]
At TKT the helicity modulus Υ (the free-energy cost of a twist across the sample: the stiffness a finite sample shows) and ξ2nd/L (the second-moment correlation length divided by the lattice size) creep toward their limits as 1/(ln L + C), so even at L = 2048, the largest lattice of Hasenbusch’s 2005 study, Υ is still 5.6 % away.
Hasenbusch derived the exact limits and the leading corrections of two quantities at TKT on an L × L torus, and simulated lattices up to L = 2048 (one lattice took about 76 days).[10] The dots in the trace are his Table 1; the curves are his fits. Slide L. At L = 2048 the helicity modulusΥ: the free-energy cost of a small twist of the boundary conditions, i.e. the spin stiffness a finite sample shows. Its infinite-volume value at TKT is the universal 2/π. Υ is still 5.6 % above its limit. The second quantity is ξ2nd/L, the second-moment correlation lengthξ2nd: a correlation length computed from the susceptibility and the lowest nonzero Fourier mode of the spin correlations. Divided by L it is dimensionless and tends to a finite value at a critical point. divided by the lattice size. To get within 1 % you would need L of order 1020. The second readout is the trap: fit a power law L−ω to the approach and you get a small exponent that keeps shrinking as you add larger lattices. It is not an exponent at all. It is 1/(ln L + C).
This is the breaker. A GOOD failure on a BKT file is REFUSE(“slow”): “the flow is still moving at every scale I can reach, and too slowly to classify.” A BAD failure is a confident WALKING (fooled by the drifting exponent) or a confident POINT (fooled by a power-law fit). The test is pre-registered on this page: it tests the Detector’s honesty, not its skill. ◆ design
(Hasenbusch 2005, eqs 31 and 36; the amplitudes are exact, C is an integration constant fitted to data). With Lmin = 256 the fits give, for Υ, amplitude 0.3033(29) and C = 0.847(67); for ξ2nd/L, amplitude 0.2112(97) and C = 1.69(36). The fitted C drifts upward as Lmin rises and χ²/dof stays above 1 to Lmin = 512: “neglected corrections die out only logarithmically”.[10]
The same analysis decided a disagreement about TKT itself: at β = 1.1199 (T = 0.8929) the fitted amplitudes approach the exact ones; at β = 1.12091 (T = 0.8921) they overshoot and keep rising, so that value is excluded. Tensor-network level spectroscopy later gave Tc = 0.892943(2), removing the logarithms by a symmetry argument rather than fighting them.[11]
At z = 0 (no vortices) both quantities are exact Gaussian-model numbers; their O(z) corrections vanish, so to leading order only the running stiffness matters: Υ = 0.63650817819 + 1.001852182(K − 2/π) + … (eq. 30, with K the Gaussian stiffness, Hasenbusch’s β). On the critical trajectory the stiffness runs as x = πK − 2 = 1/(ln L + C), i.e. K − 2/π = 1/π(ln L + C), and 1.001852182/π = 0.318899454 gives eq. 31. The limit 0.63650817819 differs from 2/π = 0.63662 because on a finite torus the Gaussian field can also wind round the periodic directions; that winding-sector sum is the difference.[10]
Why the power-law fit misleads: if f(L) = a/(ln L + C), then −d ln f/d ln L = 1/(ln L + C). Any window of L returns that value as its “exponent”, and it falls toward zero as the window moves out.
A two-dimensional crystal has two kinds of defect, so it melts in two BKT-type steps with a hexatic phase between them.
Kosterlitz and Thouless already applied their argument to a 2D crystal, where the defect is a dislocation and melting is the unbinding of dislocation pairs.[2] Halperin and Nelson worked out the dislocation-unbinding step (Young did so independently) and showed that it is only the first step; the hexatic phase and the second, disclination-unbinding step are Halperin and Nelson’s.[5] Unbinding dislocations destroys positional order but leaves the hexatic phase, which still remembers bond directions (quasi-long-range). Unbinding disclinationsOrientational defects: round one, the local bond direction turns by 60° (for a triangular lattice). (Standard geometry; not checked here.) destroys that too, and leaves an isotropic liquid. Slide T in the trace across the three phases. The hexatic has been reported in colloidal monolayers; that experimental literature is not checked here.
A LINE can end in two steps, one after another. Neither step is a POINT, and a Detector that finds “two transitions” should expect two log-slow ends, not two power laws.
Kosterlitz and Thouless’s dislocation pair energy (their eq. 72): Upair(r) ≈ (Kb²/4π)(ln|r/a| − ½cos 2θ) + 2μ, with K = 2v(1 + τ) (KT’s elastic constant, not the stiffness K) and θ the angle between b and r. The angular term only multiplies their flow equation by a Bessel-function factor that is constant for small y (their eq. 81), so “we can immediately apply all the results” of the Coulomb-gas section: melting is the unbinding of dislocation pairs, with the rigidity modulus renormalised downward.[2] The hexatic phase and its second transition are the later theory, cited here by reference only (not held locally).[5]
BKT needs three conditions, two dimensions, a continuous symmetry and a logarithmic defect energy, and its numbers also assume thermal equilibrium: drop a condition and it is a different transition.
| System | Defect | Status |
|---|---|---|
| 2D superfluid ⁴He film | vortex, π₁(S¹) = ℤ | BKT: universal jump measured[4] |
| Trapped 2D Bose gas | vortex | BKT crossover; exponent jump and free vortices coincide[7] |
| 2D nematic film | ±½ disclination | same 2/π for the Frank constants (prediction)[3] |
| 2D colloidal crystal | dislocation, disclination | KTHNY; hexatic reported (not checked here)[5] |
| Thin-film superconductor | flux line | not this transition in the 1973 argument: a finite penetration depth screens the currents and makes the vortex energy finite (which is why §1 names superfluid films and magnets, not superconductors)[2] |
| Driven-dissipative polariton condensate | vortex | BKT-like, but the exponent at the transition exceeds the equilibrium bound 1/4 (in one of the two parameter sets studied)[12] |
| 3D superfluid (λ-point) | vortex loops | different: an ordinary 3D XY critical point (textbook; not checked here) |
| 2D Ising | domain wall | discrete symmetry: an ordinary critical point (Onsager; not held here) |
| Cardiac spirals, active nematics | π₁ defects | driven, not equilibrium: an analogy under test |
Quasi-two-dimensional systems with π₁ defects are everywhere in biology: spiral waves in cardiac tissue, pinwheels in visual cortex, ±½ defects in active nematic films. The defect-unbinding picture is suggestive. But those systems are driven and dissipative, and the equilibrium machinery on this page (the universal jump, the Kosterlitz recursion, the free-energy argument itself) assumes a Boltzmann distribution that a driven system does not have. The polariton study shows the picture transferring while the numbers do not: bound pairs, then free vortices, but an exponent between 0.25 and 0.35 at the transition, in one of their two parameter sets. Why “bound”: in equilibrium η = 1/2πKR grows with T through the low-temperature phase, and KR never falls below 2/π there, so η rises to exactly 1/4 at TKT (§7) and no higher. The exact value of §7 is the top of the equilibrium range.[12] The discriminating measurement is cheap: an essential ln ξ ∝ κ−1/2 means BKT; a power law means something else. Measure before assuming.
Driven analogues often turn out to be absorbing-state or other nonequilibrium transitions with ordinary power laws. The Compass classifies the geometry it measures; it must not import BKT’s equilibrium numbers into a system that has not earned them.
The three conditions, and what breaks. Two dimensions make a point defect’s energy logarithmic; in three dimensions vortices are loops and the transition is ordinary 3D XY (textbook; not checked here). A continuous symmetry gives Goldstone modes and winding defects; with a discrete symmetry order survives and the transition is an ordinary critical point. The logarithm is what makes energy and entropy scale identically; a superconductor’s screening currents cut it off (Kosterlitz & Thouless eqs 83–84), and an isotropic Heisenberg magnet has no stable winding at all.[2]
Comaron et al. find, in a driven-dissipative condensate, an algebraic exponent 0.25 < αs < 0.35 at the transition in one parameter set (in the other it does not exceed 1/4) and a vortex number ⟨Nv⟩ ∝ ξ−1/2.[12]
BKT is a transition driven by a topological defect, but its low-temperature phase is gapless and critical, which is the opposite of what “topological order” means today.
The word is used for two different things, and the confusion starts in 1973. The Kosterlitz–Thouless abstract begins: “A new definition of order called topological order is proposed for two-dimensional systems in which no long-range order of the conventional type exists”. Their topological order is defined by the system’s response (rigidity, the stability of persistent currents).[2] About two decades later Wen gave the phrase its modern meaning: an order with no broken symmetry and no local order parameter, characterised by a ground-state degeneracy that depends on the topology of space and is robust against arbitrary perturbations, as in the fractional quantum Hall liquids (the ν = 1/q Laughlin state has qg ground states on a genus-g surface).[14] Fill in the ledger above: the two columns share almost nothing.
The reasoning, in three lines. (1) The BKT low-temperature phase is a line of fixed points with power-law correlations: gapless and critical (§2, §6). (2) Wen’s definition requires a topology-dependent ground-state degeneracy, which is protected by a gap. (3) So the BKT phase is not topologically ordered in Wen’s sense; the 2016 Nobel citation’s “topological phase transitions” is the correct description of it.[8] This conclusion is this page’s synthesis of the two premises, not a quotation.
The bridge to the Compass is the WINDING, not topological order. A winding number is a topological invariant computed from directed phase differences through one declared projection (§3). The Compass classifies the geometry of the flow, not phases of matter, so it needs no theory of topological order to handle BKT.
Wen (1995). The FQH liquid has “hidden long range orders”: long-range order despite no long-range correlation of any local operator. The degeneracy “is not a consequence of symmetry of the Hamiltonian” and survives impurities that break every symmetry. The edges are gapless chiral Luttinger liquids.[14]
McGreevy (2023). In the language of generalised symmetries, a superfluid (a broken U(1) in two space dimensions) has an emergent higher-form symmetry whose charge counts the winding of the phase, Q[C] = ∮C dφ; it “is generally not an exact symmetry of a microscopic Hamiltonian for a superfluid; it is explicitly broken by the presence of vortex configurations.” In higher dimensions the analogue of BKT is a Coulomb → Zk topological-order transition, predicted to show a universal jump in the stiffness; “many of these ideas were anticipated by Cardy”; Monte Carlo finds that transition weakly first order.[15]
Ji & Wen (2020). At a symmetry-breaking critical point the symmetry is doubled: at the Ising point the Z2 and its Kramers–Wannier dual appear together, written Z2 ∨ Z̃2, and the two cannot both be spontaneously broken. “Symmetry breaking transitions and topological transitions are treated at equal footing.”[16]
Chatterjee & Wen (2023). “Finite symmetry (with or without anomaly) is the shadow of topological order in one higher dimension.”[17]
The categorical-symmetry sketch for the Ising chain. Kramers–Wannier duality maps spins to domain walls; each representation makes one Z2 manifest, and the model has both. A patch operator of one Z2 creates a charge of the other at each end, UZ₂(i,j)UZ̃₂(i′,j′) = −UZ̃₂(i′,j′)UZ₂(i,j) for interleaved patches, which is why they cannot both be broken.[16] In the holographic picture the anomaly-free Z2 is the boundary shadow of the 2D Z2 topological order {1, e, m, f} (the toric code), and an anomalous Z2 that of the double semion order.[17]
BKT is a known-answer test for the LINE verdict, and a treacherous one: near the line’s end the honest reading is REFUSE.
The trace applies the Detector’s five-valued sign (taught on the Techniques page, §3) to the flow velocity at successive coarse-graining levels. At each level, v is the estimate of the flow velocity and w is the half-width of its band, so the band is v ± w (it widens with depth, because a coarser level has fewer blocks). The band is read against a margin ±ε that is declared before any data are seen, like the equivalence bounds of a two-one-sided-tests procedure.[18] Four yes/no tests: a = the whole band lies above +ε (v − w > ε); b = the whole band lies below −ε (v + w < −ε); c = the whole band lies inside ±ε (|v| + w < ε); r = the band excludes zero (v − w > 0 or v + w < 0). The cell is looked up from the key abcr: a with r is SIG+ (1001); b with r is SIG− (0101); c without r is ZERO! (0010); c with r (key 0011) is SLOW by the overlap rule; r without a, b, c (0001) is SLOW; every other key is ?. The overlap rule says that where the band is inside ±ε yet excludes zero, the cell is SLOW: a zero is never claimed while a nonzero is resolved, and LINE is never asserted on an ambiguous reading. The whole table is printed under “Working equations”, read from the object the trace uses. Move ε and the band and watch the BKT velocity profile, 1/(n + n0)², read out.
Two typed verdicts appear below. REFUSE(“slow”) is REFUSE with its reason attached: the flow was resolved as moving, then as slow, then lost in the noise, with no turn that would mark WALKING. LINE(ε, nlevels) is a LINE verdict that carries its own terms: stopped within the declared margin ε, over the nlevels coarse-graining levels read, and nothing more. The scoring side is where exact answers (solutions, known phase diagrams) live; they grade a verdict after it is given and never feed the instrument.
| Technique | How it sees BKT | Honest failure |
|---|---|---|
| MCRG level matching | levels n and n+1 agree better and better, but only as 1/n | REFUSE(“slow”) |
| Drifting exponents | an effective exponent that shrinks as 1/(ln L + C), not a plateau | REFUSE, never a fitted ν |
| Kernel signs | SIG → SLOW → ? with no turn back up | REFUSE(“slow”) |
| Transfer matrix | central charge 1 and a marginal operator; exact on the scoring side | truth for scoring, not an instrument input |
The techniques page of this series takes each row apart.
Pre-registered: on BKT the honest Detector answers REFUSE. The breaker tests honesty, not skill. And the verdict depends on ε, which is exactly why ε is declared first: chosen afterwards, it could make any answer come out.
The table has one cell for each of the 16 keys abcr of the tests above. Keys such as 1100 (the band both above +ε and below −ε) cannot occur for w, ε ≥ 0; they read ?, so the table has no gaps. There is no chain of conditions: the verdict is a lookup, and “?” is a value. With ε = 0 it collapses to the three-valued sign (SIG+, SIG−, ?).[18] The full table, read from the same object the trace uses:
The BKT row: velocity vn ∝ dx/dt = −x² with x = 1/(t + C) along the critical trajectory, so |vn| ∝ 1/(n + n0)² (n0 a constant, 1.5 in the trace). Once v − w no longer exceeds ε the key is 0001 or 0011 and the cell is SLOW; once the band covers zero it reads ZERO! if it also lies inside ±ε (0010), and ? otherwise. The named cell SIG → SLOW → ? with no later turn maps to REFUSE(“slow”). The verdict on a LINE is typed LINE(ε, nlevels): equivalent to stopped within the declared margin over the levels read, and nothing more.
Ten lines to keep, then a twenty-question test that draws a different set each time.
Twenty items drawn from the page’s bank, never from the section quizzes; a retake draws a mostly different set. Task items ask you to set a graphic above and then submit.
Related. The Renormalization Group for the Working Mathematician: fixed points, relevant directions and universality, of which the Kosterlitz flow is the most unusual example. Critical lines: the models (Ashkin–Teller, XXZ, Potts walking, Blume–Capel and more) and the techniques (MCRG, drifting exponents, the kernel lens). Confined Helium KT: the transition measured in ⁴He nanochannels of 10, 15 and 20 nm, with rotons, the dynamic theory and a live fit to the published data.
Criticality at codimension zero. Almost every route to critical behaviour requires setting a parameter to a precise value, which raises an awkward question whenever criticality is claimed in a natural system: who tuned it? BKT dissolves the question. A system anywhere in the low-temperature phase is critical without anything having been arranged, and the theorem that forbids the ordered phase is the one that creates the critical one.
Kosterlitz–Thouless for the Working Mathematician — B. Stephenson, 2026.