---
title: Phenomenological Renormalization Group (PRG)
url: https://www.emergentmind.com/topics/phenomenological-renormalization-group-prg
type: topic
---

# Phenomenological Renormalization Group (PRG)

Searching arXiv for recent and foundational papers on phenomenological renormalization group.
Phenomenological renormalization group (PRG) denotes a class of renormalization procedures in which the RG step is defined operationally from finite-size observables or empirical correlation structure rather than from a fully specified microscopic Hamiltonian. In the lattice-model setting, PRG refers to Nightingale-style matching of finite-strip correlation lengths or mass gaps, possibly combined with cluster decimation, to construct recursion relations in coupling space and extract critical data [1311.0469]. In the neural-data setting, PRG is a model-agnostic, data-driven coarse-graining of activity variables based on correlation-preserving pair merges in direct space or low-variance mode elimination in “momentum” space, with the central question being whether the resulting distributions flow toward a non-Gaussian critical fixed point [2001.04353, 2506.14053].

## 1. Conceptual basis

The RG logic underlying PRG is the standard one: a block transformation at scale factor $b$ maps microscopic variables $\sigma$ into coarse variables $\sigma'=\mathcal{R}[\sigma;b]$, and correspondingly maps couplings $K$ into renormalized couplings $K'=R_b(K)$. Critical points satisfy the fixed-point condition $K^*=R_b(K^*)$, and the linearization around $K^*$ determines relevant and irrelevant directions. PRG retains this fixed-point viewpoint while replacing an exact microscopic derivation of $R_b$ with phenomenological matching conditions imposed on observables that can be computed from finite systems or directly from data [2506.14053].

In the finite-size formulation, the relevant observables are strip correlation lengths, second-moment lengths, transfer-matrix eigenvalues, and free-energy-related quantities. In the neuronal formulation, the observables are covariance matrices, cluster variances, silence probabilities, covariance spectra, autocorrelations, and the full distribution of PCA-based coarse variables. The common principle is that scale invariance is inferred from the flow of observables under repeated coarse-graining rather than from explicit knowledge of all microscopic interactions. PRG is therefore phenomenological in the precise sense that it defines RG steps by matching finite-resolution descriptions of the same system [1311.0469].

A related point is methodological rather than definitional: PRG provides an alternative to explicit avalanche constructions. In the avalanche paradigm one measures $P(S)\sim S^{-\tau}f(S/S_c)$ and $P(T)\sim T^{-\alpha}g(T/T_c)$, with hyperscaling relations such as $\alpha=(\tau-1)/D+1$; PRG instead probes scale invariance by asking how observables transform under coarse-graining, without necessarily constructing avalanches at all [2506.14053].

## 2. Finite-size matching in lattice and gauge models

Borisenko et al. formulate a modified phenomenological RG by combining Nightingale’s PRG with the cluster-decimation approximation (CDA). The starting point is a general $Z(N)$ spin model on a hypercubic lattice $\Lambda_0$, with partition function
$$
Z(\Lambda_0;\{t_k\})=\prod_{x\in\Lambda_0}\frac{1}{N}\sum_{s(x)=0}^{N-1}\prod_{\langle x,y\rangle}Q[\{t_k\};s(x)-s(y)],
$$
where
$$
Q[\{t_k\};s]=\sum_{k=0}^{N-1}t_k\,e^{(2\pi i/N)ks},\qquad t_0=1,\quad t_k=t_{-k}=t_{k+N}.
$$
The two-point correlator in representation $r$ is
$$
\Gamma_r(\Lambda_0;R)=\Big\langle e^{(2\pi i/N)r[s(x)-s(x+R)]}\Big\rangle.
$$
Embedding the model on an $M\times L$ strip with $L\to\infty$ and transfer matrix $T$, one denotes by $\lambda_0(M)$ the largest eigenvalue and by $\lambda_r(M)$ the leading eigenvalue in sector $r$, so that
$$
F(M)=\lim_{L\to\infty}\frac{1}{L}\ln Z(M\times L)=\ln\lambda_0(M),
$$
$$
\Gamma_r(M;R)=D_r(M;R)\,[B_r(M)]^R,\qquad B_r(M)=\frac{\lambda_r(M)}{\lambda_0(M)},
$$
and the mass gap is
$$
m_r(M)=-\ln B_r(M).
$$
These strip quantities furnish exact input for the CDA-based RG step [1311.0469].

The decimation step relates the original lattice $\Lambda_0$ to a lattice $\Lambda_1$ of spacing $2$ through new couplings $\{t_k^{(1)}\}$:
$$
Z(\Lambda_0;\{t\})=A(\{t\})\,Z(\Lambda_1;\{t^{(1)}\}),
$$
$$
\Gamma_r(\Lambda_0;\{t\};R)=G_r(\{t\},R)\,\Gamma_r(\Lambda_1;\{t^{(1)}\};R/2).
$$
Within CDA, the prefactors are taken from exact strip data, in particular
$$
A(\{t\})=\left[\frac{\lambda_0(M;\{t\})}{\lambda_0(M/2;\{t^{(1)}\})}\right]^{L^2/M},
$$
with analogous expressions for $G_r$. The phenomenological matching condition is the preservation of the mass gap, up to the scale factor $b=2$:
$$
b\,m_r(M;\{t\})=m_r(M/2;\{t^{(1)}\})
\quad\Leftrightarrow\quad
B_r(M;\{t\})^2=B_r(M/2;\{t^{(1)}\}),
$$
for each representation $r=1,\dots,N-1$. An equivalent formulation preserves the second-moment correlation length $\xi_2$ through
$$
b\,\xi_2(M;\{t\})=\xi_2(M/2;\{t^{(1)}\}).
$$
The resulting nonlinear system defines the renormalized couplings and hence an RG flow in coupling-constant space [1311.0469].

Fixed points $\{t^*\}$ satisfy
$$
B_r(M;\{t^*\})=B_r(M/2;\{t^*\}),\qquad \forall r.
$$
Linearization gives a Jacobian $J_{ij}=\partial t_i^{(1)}/\partial t_j|_{t^*}$, whose leading eigenvalue $\Lambda_{\max}$ yields the correlation-length exponent
$$
\nu=\frac{\ln b}{\ln \Lambda_{\max}}.
$$
Because the free-energy prefactor $A(\{t\})$ is retained along the flow, the method also permits computation of the bulk free energy and specific heat [1311.0469].

## 3. Data-driven coarse-graining in neuronal systems

In neuronal applications, PRG is defined directly on activity variables. The raw variables are denoted $\sigma_i^{(1)}$, $i=1,\dots,N$, with equal-time covariance matrix
$$
C_{ij}=\langle \sigma_i^{(k)}\sigma_j^{(k)}\rangle-\langle \sigma_i^{(k)}\rangle\langle \sigma_j^{(k)}\rangle
$$
and normalized correlation matrix
$$
c_{ij}=\frac{C_{ij}}{\sqrt{C_{ii}C_{jj}}}.
$$
The direct-space construction iteratively forms disjoint pairs by selecting the maximal off-diagonal correlation $c_{i,j_{\max}}$, removing that pair, and repeating until $N_k/2$ pairs have been formed. For each selected pair, the coarse variable is defined by simple summation,
$$
\sigma_i^{(k+1)}=\sigma_i^{(k)}+\sigma_{j_{\max}}^{(k)},\qquad i=1,\dots,N_{k+1}=N_k/2.
$$
After $k$ steps, each $\sigma_i^{(k)}$ is the sum of $K=2^{k-1}$ original variables, and the covariance is recomputed after each blocking step. Because all $\sigma_i\ge 0$, coarse blocks can vanish, which makes the silence probability a natural observable [2001.04353].

The complementary momentum-space construction starts from the eigendecomposition of the original covariance matrix,
$$
\sum_j C_{ij}^{(1)}u_{jr}=\lambda_r u_{ir},\qquad \lambda_1\ge \cdots \ge \lambda_N.
$$
One forms the rank-$\hat K$ projector
$$
P_{ij}(\hat K)=\sum_{r=1}^{\hat K}u_{ir}u_{jr},
$$
and defines coarse variables
$$
\phi_i(\hat K)=z_i(\hat K)\sum_j P_{ij}(\hat K)\bigl[\sigma_j^{(1)}-\langle \sigma_j^{(1)}\rangle\bigr],
$$
where $z_i$ enforces $\mathrm{Var}[\phi_i]=1$. As $\hat K$ decreases, low-variance high-rank modes are discarded, mimicking a momentum-space RG cutoff. The same PCA-based construction is used in the later neuronal study, there written in terms of binary variables $\phi_i(t)$ and coarse variables $\psi_i(N_{\mathrm{cutoff}})$ with a normalization constant $Z$ chosen so that $\mathrm{Var}(\psi)=1$ [2001.04353, 2506.14053].

A basic interpretive premise of this formulation is that no assumption of Gaussianity is built into the procedure. For weakly correlated variables, a Gaussian, central-limit fixed point is expected. Persistent non-Gaussian tails under repeated coarse-graining are taken as evidence for a nontrivial interacting fixed point rather than a trivial Gaussian one [2001.04353].

## 4. Scaling observables and fixed-point diagnostics

The neuronal PRG literature emphasizes several scaling relations expected at a nontrivial critical fixed point. The first is variance scaling:
$$
M_2(K)=\frac{1}{N_k}\sum_{i=1}^{N_k}\Big[\langle (\sigma_i^{(k)})^2\rangle-\langle \sigma_i^{(k)}\rangle^2\Big]\propto K^{\tilde\alpha}.
$$
The second is the silence probability,
$$
P_{\mathrm{silence}}(K)\equiv \mathrm{Prob}[\sigma_i^{(k)}=0],\qquad
F(K)\equiv \log P_{\mathrm{silence}}(K)\sim -K^{\tilde\beta}.
$$
The third is covariance-spectrum scaling: if at criticality in $d$ dimensions $G(r)\sim r^{-(d-2+\eta)}$, then the ordered covariance eigenvalues obey
$$
\lambda_r\sim (K/r)^\mu,\qquad \mu=\frac{2-\eta}{d}.
$$
The fourth is dynamical scaling of the mean autocorrelation
$$
C^{(k)}(t)=\frac{1}{N_k}\sum_i
\frac{\langle \sigma_i^{(k)}(t_0)\sigma_i^{(k)}(t_0+t)\rangle-\langle \sigma_i^{(k)}\rangle^2}
{\mathrm{Var}[\sigma_i^{(k)}]},
$$
from which a characteristic time $\tau_c(K)$ is extracted, with
$$
\tau_c\propto K^{\tilde z}.
$$
These observables operationalize the fixed-point question in direct space and dynamics [2001.04353].

The momentum-space literature adds a distributional diagnostic. One monitors the full distribution of the PCA coarse variables and, in particular, the kurtosis
$$
\kappa=\frac{\langle \psi^4\rangle}{\langle \psi^2\rangle^2}.
$$
At a true non-Gaussian fixed point, $P^*(\psi)$ remains distinct from a Gaussian under many coarse-graining steps, so $\kappa$ remains distinct from the Gaussian value $\kappa=3$. This criterion is presented as especially sensitive in neuronal models: momentum-space PRG asks whether non-Gaussian structure survives removal of a large fraction of modes, whereas a trivial or weakly correlated regime flows rapidly toward Gaussianity [2506.14053].

Taken together, these diagnostics motivate a multi-observable reading of PRG outputs. Variance growth, silence scaling, eigenvalue spectra, autocorrelation times, and the flow of coarse-grained distributions probe different aspects of the same candidate fixed point. A central practical claim in the literature is that no single observable suffices on its own [2001.04353].

## 5. Benchmark systems and quantitative behavior

The systematic benchmark in the contact process considers $N$ binary sites $\sigma_i\in\{0,1\}$ on either a $2$D lattice with $N=40^2$ or a small-world network. Each occupied site empties at rate $1$, and each empty site with $n_i$ occupied neighbors becomes occupied at rate $\lambda n_i/k_i$. The model has an absorbing state with all $\sigma_i=0$, order parameter $\rho=\langle \sigma\rangle$, and control parameter $\lambda$, with transitions at $\lambda_c\approx 1.6488$ in $2$D and $\lambda_c\approx 1.7961$ on the small-world network. At $\lambda=\lambda_c$, the variance exponent is reported as $\tilde\alpha\approx 1.15$, the silence exponent as $\tilde\beta\approx 0.65$, the covariance-spectrum exponent as $\mu=0.63\pm 0.02$ consistent with $\mu=(2-\eta)/d\approx 0.6$ in directed percolation, and the dynamical exponent as $\tilde z\approx 0.50\pm 0.06$. For $\lambda\gg\lambda_c$, $\tilde\alpha$ relaxes toward $1$, the silence exponent is smaller and decays faster, the spectrum flattens, and the autocorrelation becomes exponential rather than scaling. In momentum space, $P_{\hat K}(\phi)$ flows toward Gaussian for $\lambda>\lambda_c$, while at $\lambda_c$ it retains clear non-Gaussian tails even when keeping only $1\%$ of modes [2001.04353].

The same study highlights near-critical ambiguity. At $\lambda\simeq 1.1\,\lambda_c$, variance and $F(K)$ still show apparent power laws, with exponents intermediate between critical and deep super-critical values, but the covariance spectrum and $\tau_c(K)$ do not exhibit clean scaling, and the momentum-space distribution rapidly becomes Gaussian. This is presented as evidence that some PRG observables can remain deceptive at distances of $10\%$ from the critical point, whereas spectrum and momentum-space diagnostics are more discriminating [2001.04353].

A later analysis examines two neuronal models with known critical points: an excitable cellular automaton in the mean-field directed-percolation class on $N=10^4$ sites with branching ratio $\sigma$ and $\sigma_c=1$, and a stochastic excitatory–inhibitory integrate-and-fire network with inhibition strength $g$ and $g_c=1.5$. In both cases, $T=5\times 10^3\,\mathrm{ms}$ time series are recorded from $256$ randomly subsampled neurons, binned adaptively via $\Delta t=f\langle \mathrm{ISI}\rangle$, and analyzed with PRG. The momentum-space kurtosis remains near the Gaussian baseline $\kappa\approx 3$ away from criticality but exhibits a sharp peak at $\sigma\approx \sigma_c$ or $g\approx g_c$; surrogate data obtained by shuffling ISIs stays near $\kappa\approx 3$ throughout. Real-space coarse-graining yields $M_2\sim C_{\mathrm{size}}^\alpha$, with $\alpha\approx 1$ away from criticality, $\alpha\to 2$ at extreme saturation, and nontrivial $1<\alpha<2$ only within a narrow window around the true critical point. The reported conclusion is that PRG detects scaling only in a very narrow range around criticality under proper preprocessing [2506.14053].

In the lattice-model tradition, Borisenko et al. report quantitatively accurate flows for several systems. For the $2$d $Z(4)$ spin model, four fixed lines are found, including the standard Potts line $t_1=t_2$, where the critical coupling and exponent converge toward the exact values $t_s^*=1/3$ and $\nu_s=2/3$, with deviations $\lesssim 1\%$ for the sequence of steps $M=8\to 6\to 4$. Using $\xi_2$ matching on $L\times L$ clusters, the predicted $\beta_c$ and $\nu$ for the $2$d Ising model approach the exact values $\beta_c=0.440687$ and $\nu=1$ as $L$ grows to $64$, with similarly good accuracy for $N=3,5,13$. For $3$d $Z(N)$ lattice gauge models, the method predicts $\beta_c$ within a few percent of large-scale Monte Carlo and yields $\nu\approx 0.62$–$0.70$, consistent with known universality classes [1311.0469].

## 6. Limitations, artifacts, and interpretive cautions

A recurrent theme in the modern PRG literature is that scaling signatures are necessary but not sufficient conditions for criticality. In the contact-process benchmark, no single PRG observable suffices: real-space variance and silence probability can show apparent power laws even when the system is not exactly critical, whereas covariance spectra, autocorrelation scaling, and momentum-space non-Gaussianity provide stricter cross-checks. The same study reports that the qualitative RG flow does not depend on the presence of long-range interactions, indicating that PRG is agnostic to network topology at that level, but this topological robustness should not be conflated with proof of criticality [2001.04353].

A distinct source of ambiguity is preprocessing. In neuronal spike trains, binarization over time windows $\Delta t$ induces an average firing density $\rho_{\mathrm{bin}}=\langle \phi\rangle$ that depends strongly on $\Delta t$. If $\Delta t$ is too small, most bins are empty and spurious correlations arise among all zeros, producing artificially high kurtosis; if $\Delta t$ is too large, almost every bin is nonempty, again producing large kurtosis. The proposed remedy is adaptive binning,
$$
\Delta t=f\cdot \langle \mathrm{ISI}\rangle,
$$
with $f\sim O(1)$, which keeps $\rho_{\mathrm{bin}}$ roughly constant, for example around $\sim 0.15$, across subcritical and supercritical regimes. Under fixed rather than adaptive $\Delta t$, spurious large kurtoses are found deep in subcritical or supercritical regimes [2506.14053].

The literature also records more subtle failure modes. Large system size $N$ and a wide hierarchy of block sizes $K$ are required to resolve power laws and to separate Gaussian from non-Gaussian flows. For asynchronous updates, uncorrelated time points should be subsampled to avoid artificial temporal correlations. Moreover, simple extrinsic-noise models can produce non-Gaussian momentum-space tails without true interactions; in the formulation summarized for the contact-process study, direct-space pairing is presented as a way to catch such cases. These caveats place PRG in the category of a stringent but indirect inferential framework: it can rule out triviality more effectively when multiple observables agree, but it does not by itself replace a full dynamical or microscopic theory [2001.04353].

A plausible implication is that the strongest use of PRG lies in comparative diagnosis rather than in isolated exponent fitting. When direct-space, momentum-space, spectral, and dynamical observables all point to the same non-Gaussian scale-invariant regime, the case for a critical fixed point is materially stronger than when only one observable displays an approximate power law. That interpretive standard is common to both the finite-size lattice formulation and the neural-data formulation, despite their different operational definitions of the RG step [1311.0469, 2506.14053].

Source: https://www.emergentmind.com/topics/phenomenological-renormalization-group-prg