Discontinuity of the q-State Potts Phase Transition on Z^d for d >= 3 and q >= 3

Discontinuity of the q-State Potts Phase Transition on Z^d for d >= 3 and q >= 3

Definitions and mathematical statement

In the nearest-neighbor ferromagnetic Potts model, spins $\sigma_x\in\{1,\ldots,q\}$ on ${\mathbb{Z}}^d$ have finite-volume weights proportional to \[ \exp\!\left(\beta\sum_{\langle x,y\rangle} \mathbf1_{\{\sigma_x=\sigma_y\}}\right), \] with fixed boundary spins used to select an ordered state. Write $m_q(\beta)$ for its excess probability of spin $1$ at the origin above $1/q$, in the infinite-volume limit, and let $\beta_c$ be the threshold for positive order. The order-parameter discontinuity formulation asks $m_q(\beta_c)>0$ for every $d\ge3$, $q\ge3$. Equivalently in the random-cluster convention, the wired critical measure has positive probability of an infinite cluster, with $p=1-e^{-\beta}$. The catalogue calls the transition first-order but does not specify a latent-heat normalization; the displayed statement specifies the order/disorder discontinuity convention rather than adding a separate quantitative energy-jump demand.

Formulation and concept references: [S1].

Short English statement

Does every nearest-neighbor Potts model with at least three spin states in dimension at least three acquire long-range order discontinuously at its transition?

Research attempt

The exact discontinuity under discussion. The target is positive order at the critical point in the wired state. It is not a numerical lower bound for latent heat. The higher-dimensional discussion and Conjecture 2 in [S1] distinguish the full nearest-neighbor assertion from the established large-$q$ and sufficiently-high-dimensional regimes. Those regimes do not cover every pair $(d,q)$ in the statement. This attempt derives the cluster formulation, proves elementary noncritical estimates, and tests a tree recursion which displays the expected discontinuous mechanism but does not transfer to the lattice.

Exact finite-volume spin--cluster correspondence. Let $G=(V,E)$ be finite, with its boundary vertices wired together and assigned spin $1$. Put $p=1-e^{-\beta}$. For each edge, \[ e^{\beta\mathbf1_{\{\sigma_x=\sigma_y\}}} =e^\beta\left((1-p)+p\mathbf1_{\{\sigma_x=\sigma_y\}}\right). \] Expanding the product over edges and then summing spins gives the random-cluster weight \[ p^{|\omega|}(1-p)^{|E|-|\omega|}q^{k_0(\omega)}, \] apart from the constant $e^{\beta|E|}$. Here $k_0$ counts open components not connected to the wired boundary. Given $\omega$, assign spin $1$ to the boundary cluster and independent uniform spins to the other clusters. If the origin is not boundary-connected its spin is uniform; otherwise it is $1$. Consequently \[ \mathbb P(\sigma_0=1)-\frac1q =\left(1-\frac1q\right)\phi^w_{G,p,q}(0\leftrightarrow\partial G). \] The same identity holds in the wired infinite-volume limit. In particular, the normalization in this entry is $m_q(\beta)=(1-1/q)\theta^w(p)$, not $\theta^w(p)$ itself.

Conditional edge probabilities and elementary bounds. Fix all edges except $e$. If its endpoints are already connected, opening $e$ does not change $k_0$ and its conditional open probability is $p$. If they are not connected, opening merges two components, at least one of which is not the wired component, and reduces $k_0$ by one. The probability is then \[ r=\frac{p}{p+q(1-p)}. \] Thus every single-edge conditional probability lies in $[r,p]$. Revealing edges in any fixed order and using independent uniform random variables couples the configuration between Bernoulli configurations of parameters $r$ and $p$: each conditional probability given the revealed prefix is an average of the full conditional probabilities and remains in that interval. This proves the finite-volume stochastic bounds directly. Limits preserve inequalities for increasing cylinder events, hence give the corresponding infinite-volume comparisons.

For a self-avoiding path with $n$ edges, its all-open probability is at most $p^n$. There are at most $2d(2d-1)^{n-1}$ such paths starting at the origin. An infinite open cluster contains paths of every length, so \[ \theta^w(p)\leq 2d(2d-1)^{n-1}p^n, \qquad p<\frac1{2d-1}\ \Longrightarrow\ \theta^w(p)=0. \] The argument works for every $q\geq1$ and makes no critical-point claim.

For completeness, a low-temperature positivity bound can be proved by restricting the dominated Bernoulli configuration to a coordinate copy of $\mathbb Z^2$. If the origin's Bernoulli cluster is finite, a closed-edge dual circuit surrounds it. The number of length-$n$ candidates is bounded by $4n3^n$: choose an intersection with a fixed positive coordinate ray and then a nonbacktracking walk. The crude bound suffices. Therefore \[ \mathbb P_r(0\not\leftrightarrow\infty) \leq\sum_{n\geq4}4n\,[3(1-r)]^n. \] For $r$ sufficiently close to one this sum is below one, proving $\theta^w(p)>0$. Since $r\uparrow1$ as $p\uparrow1$ for fixed finite $q$, these estimates bracket a nontrivial threshold. They leave a large interval between the explicit high- and low-temperature conditions, and do not give a positive lower bound at its actual critical endpoint.

A useful limit-order check. Let $\Lambda_N$ be increasing boxes and $\theta_N(p)=\phi^w_{\Lambda_N,p,q}(0\leftrightarrow\partial\Lambda_N)$. For fixed $N$ this is a ratio of finite polynomials with positive denominator on $0<p<1$, hence continuous there. The spatial Markov property and the maximality of wired boundary conditions imply that $\theta_N$ decreases with $N$. Its limit is $\theta^w(p)$. To check the latter point rather than assume an exchange of limits, fix a smaller box $\Lambda_R$. For $N\geq R$, \[ \theta_N(p)\leq\phi^w_{\Lambda_N,p,q}(0\leftrightarrow\partial\Lambda_R). \] Taking $N\to\infty$ first gives the infinite-volume probability on the right; then $R\to\infty$ gives $\theta^w(p)$. The reverse inequality follows from the wired-domain comparison. The standard monotone construction of the wired measure underlies these comparisons.

Thus $\theta^w(p)=\inf_N\theta_N(p)$ is upper semicontinuous. It is also nondecreasing in $p$, by the random-cluster monotonicity for $q\geq1$. These two facts imply right continuity at interior parameters: \[ \theta^w(p)=\lim_{s\downarrow p}\theta^w(s). \] They do not imply left continuity. In fact the conjecture asks for a jump from the zero left-hand regime to a positive value at $p_c$. Finite-volume analyticity does not prohibit such a jump after the volume limit. Equally, right continuity does not prove a jump: the function $\max\{p-p_c,0\}$ has all these properties and equals zero at $p_c$. This elementary model is a countercheck on the attempted inference, not a candidate Potts order parameter.

An exactly solvable recursion showing a discontinuous branch. On a rooted tree with $b$ children per vertex, let the incoming branch partition functions favor color $1$ by a ratio $z$ over every other color. Writing $E=e^\beta$, summation over each child's spin yields the recursion \[ z\longmapsto\left(\frac{Ez+q-1}{z+E+q-2}\right)^b. \] The symmetric solution $z=1$ is always fixed. The derivative of this map there equals $b(E-1)/(E+q-1)$, so its linear stability changes at $E=(b+q-1)/(b-1)$.

For the concrete choice $b=2$, $q=3$, put $y=\sqrt z>0$. The fixed-point equation becomes \[ y^3-Ey^2+(E+1)y-2 =(y-1)\left(y^2+(1-E)y+2\right)=0. \] The nonsymmetric positive roots first appear at $E_*=1+2\sqrt2$, when $y=\sqrt2$ and $z=2$. They therefore appear a definite distance from $z=1$, while the latter's linear derivative is still less than one because $E_*<4$. This gives an explicit coexistence-of-fixed-points mechanism rather than a continuous bifurcation from the symmetric branch. At $E=4$ the two extra roots are $y=1,2$, agreeing with the computed loss of linear stability.

These equations describe a tree message recursion, not the nearest-neighbor lattice model. Even on the tree, selecting a thermodynamic phase requires specifying boundary conditions and the relevant free energy; the mere birth of two fixed points is not a proof that this value is the transition point. On $\mathbb Z^d$, loops correlate incoming branches, so multiplying $b$ identical independent contributions is not an exact calculation. Replacing $b$ by $2d-1$ cannot repair this loss of independence.

Where a contour attack must gain a uniform estimate. The cluster bounds suggest trying to separate an ordered ensemble from a disordered one and show that interfaces between them are costly. An estimate that works only when $p$ is very close to one proves low-temperature order, which is already supplied above. To prove the target it must remain quantitatively valid at the actual coexistence parameter, without assuming that $q$ or $d$ is large. This requires enough interface suppression to keep the wired infinite cluster density bounded away from zero as $p\downarrow p_c$. Neither the single-edge comparison nor the tree recursion establishes such suppression for $(d,q)=(3,3)$, the smallest requested case. Numerical free-energy crossings also do not bound a thermodynamic-limit order parameter without uniform errors and a controlled volume limit.

Verification and outcome. The diagnostic enumerates all spin and bond configurations of a small wired graph at rational $e^\beta$, checking the expansion and the exact factor $1-1/q$. It checks the tree polynomial factorization with exact integer coefficients and evaluates the crude Peierls sum at one explicit high-density parameter. These are finite identities and a numerical bound, not evidence for a uniform critical jump. The first missing step remains a strictly positive critical-density estimate valid for every requested $d$ and $q$.

Sources

• S1 H. Duminil-Copin, 'Lectures on the Ising and Potts models on the hypercubic lattice', arXiv:1707.00520 $2017$, in Random Graphs, Phase Transitions and the Gaussian Free Field, Springer PIMS 123 $2020$. https://ar5iv.labs.arxiv.org/html/1707.00520. Formulation source.