Scaling Theory of Localization

The scaling theory of localization asks how the dimensionless conductance of a phase-coherent disordered sample changes when the observation length is enlarged. Its central hypothesis is that, after microscopic details and geometry have been fixed, long-distance transport can be organized by a flow

$$\beta(g) \equiv \frac{d\ln g}{d\ln L}.$$

The sign of $\beta$ is immediately physical. If $\beta(g)>0$, conductance grows as the sample is enlarged and the flow is metallic. If $\beta(g)<0$, conductance decreases and the flow is toward an insulator. A zero of $\beta$ is a scale-invariant fixed point and can represent a continuous metal–insulator transition.

This compact statement is powerful, but it is conditional. The original one-parameter theory concerns coherent, noninteracting, disordered single-particle systems with a fixed symmetry class. It does not say that every disordered film is insulating, that every resistance crossing is an Anderson transition, or that interactions, topology, absorption, gain, and non-Hermiticity can always be hidden inside one number.

The aim is to connect the interference correction with a flow in system size, then use a fixed point to derive critical scaling. The original argument is due to Abrahams et al. (1979).

Required background. Disorder in Quantum Matter supplies the microscopic ensemble and elastic transport scales. Anderson Localization supplies localized-state and transmission diagnostics. Weak Localization supplies the perturbative interference correction from which the large-conductance scaling regime is approached.

Size-dependent conductance and localization scaling

Use $G_0=e^2/h$, counting spin and valley multiplicities in the channels. Compare samples from one stationary disorder ensemble at fixed energy and aspect ratio, with consistent leads and contacts. The basic flow is a zero-temperature, phase-coherent construction; finite-temperature interpretation requires the additional discussion below.

The conductance of a mesoscopic sample is random. A notation such as $g(L)$ therefore means a declared statistic of a fixed ensemble and measurement protocol. In the metallic regime, the mean may be adequate. Deep in the localized regime, the typical conductance

$$g_{\mathrm{typ}} \equiv \exp\!\left( \overline{\ln g} \right)$$

is usually more representative than $\overline g$. At criticality, the whole distribution can remain broad and scale invariant.

Dimensionless Conductance

From terminal conductance to a scale variable

For a $d$-dimensional hypercube of side $L$ with bulk conductivity $\sigma$, the geometry-matched dimensionless conductance is

$$g_\sigma(L) = \frac{ \sigma(L)L^{d-2} }{ e^2/h }.$$

This expression is dimensionless in every spatial dimension. If $\sigma$ did not renormalize, then

$$g_\sigma(bL) = b^{d-2}g_\sigma(L).$$

The corresponding classical beta function would be

$$\beta_{\mathrm{cl}} = d-2.$$

That elementary geometry already distinguishes the dimensions. A fixed bulk conductivity gives decreasing conductance in one dimension, size-independent conductance in two dimensions, and increasing conductance in three dimensions. Quantum interference changes the flow, especially where the classical term vanishes.

For an open coherent conductor, the Landauer formula gives

$$g_G = \frac{G}{G_0} = \sum_n T_n.$$

This terminal $g_G$ includes the chosen leads and contacts. A scaling sequence must preserve lead geometry and aspect ratio, or consistently remove contact contributions, before comparing different $L$. Otherwise a contact crossover can be mistaken for bulk renormalization.

Thouless energy and level spacing

The same scale variable has a spectral interpretation. For diffusion constant $D$,

$$E_{\mathrm{Th}} = \frac{\hbar D}{L^2}$$

is the Thouless energy: the energy associated with diffusing across the sample. If $\nu(E_{\mathrm F})$ is the density of states per volume, including the declared internal degeneracy, then the mean single-particle level spacing in the box is

$$\Delta_L = \frac{1}{ \nu(E_{\mathrm F})L^d }.$$

Using the Einstein relation

$$\sigma = e^2\nu(E_{\mathrm F})D,$$

one finds

$$\begin{aligned} g_\sigma &= \frac{\sigma L^{d-2}}{e^2/h}, \\ &= h\nu(E_{\mathrm F})D L^{d-2}, \\ &= 2\pi \frac{E_{\mathrm{Th}}}{\Delta_L}. \end{aligned}$$

Thus conductance measures, up to a convention-dependent factor, how many levels are mixed during one traversal time. Many authors define the Thouless number as $E_{\mathrm{Th}}/\Delta_L$; others absorb $2\pi$ or define the traversal energy with $h$ instead of $\hbar$. Numerical coefficients should never be transferred between conventions without matching these definitions.

An equivalent closed-system diagnostic is sensitivity to a boundary twist. If changing a periodic boundary phase shifts a typical level by $\delta E$, then

$$g_{\mathrm T} \sim \frac{\delta E}{\Delta_L}.$$

Extended states communicate with the boundary and have appreciable curvature. Exponentially localized states far from the boundary are much less sensitive. This spectral definition avoids physical leads but introduces its own boundary and averaging conventions.

The one-parameter hypothesis

Suppose a block of size $L$ is replaced by an effective block of size $bL$. One-parameter scaling asserts that the new conductance depends only on the old conductance,

$$g(bL) = F_b\!\left(g(L)\right),$$

once symmetry class, energy, disorder ensemble, boundary protocol, and shape have been fixed. Successive coarse-graining must compose:

$$F_{b_1b_2} = F_{b_1}\circ F_{b_2}.$$

For an infinitesimal change of scale this semigroup property produces an autonomous differential flow, the beta function. “One parameter” does not mean that conductance has no fluctuations. It means that the long-distance family is parameterized by one running variable. Near a critical point, a single relevant direction can produce a one-variable scaling form after irrelevant corrections decay. That local property alone does not prove a one-parameter flow across the whole phase diagram; finite-size drift can reveal the additional variables.

The hypothesis must be tested rather than assumed when:

• interactions generate independent amplitudes or dephasing scales;
• magnetic field or spin relaxation changes with $L$;
• a topological angle distinguishes sectors at the same longitudinal conductance;
• disorder has long-range correlations or several active length scales;
• gain, loss, non-Hermiticity, or driving changes probability conservation;
• the sample crosses between ballistic, diffusive, and hydrodynamic regimes.

Scheme dependence and invariant conclusions

The detailed beta-function curve depends on how conductance is defined. Suppose a second valid convention uses a smooth, strictly increasing positive coordinate

$$\widetilde g = f(g).$$

Its beta function is

$$\begin{aligned} \widetilde\beta(\widetilde g)&=J(g)\beta(g),\\ J(g)&=\frac{d\ln f}{d\ln g}. \end{aligned}$$

Multiplying $g$ by a constant, subtracting a contact term before forming $g$, or replacing a terminal conductance by a Thouless number can therefore shift the plotted curve and the numerical value of $g_c$. Under this nonsingular increasing reparameterization, zeros map to zeros and increasing conductance still maps to increasing conductance. A decreasing coordinate, such as resistance, reverses the numerical flow direction; the physical fixed-point stability remains the same.

The critical exponent is invariant under either nonsingular choice. At a fixed point,

$$\left. \frac{d\widetilde\beta}{d\ln\widetilde g} \right|_{\widetilde g_c} = \left. \frac{d\beta}{d\ln g} \right|_{g_c},$$

because the term differentiating the coordinate Jacobian is multiplied by $\beta(g_c)=0$. This is why a critical exponent can be universal while the critical conductance is not. A singular coordinate change or an inconsistent contact subtraction does not enjoy this protection.

Beta Function

Reading a flow

With

$$\beta(g) = \frac{d\ln g}{d\ln L},$$

the flow between two sizes obeys

$$\ln \left( \frac{L_2}{L_1} \right) = \int_{g(L_1)}^{g(L_2)} \frac{d\ln g}{\beta(g)}.$$

A fixed point $g_\ast$ satisfies

$$\beta(g_\ast)=0.$$

It is stable if nearby flows return to it as $L$ increases and unstable if they move away. A conventional three-dimensional Anderson transition is governed by an unstable fixed point: samples on one side flow toward a metal and those on the other flow toward an insulator.

Metallic and localized asymptotes

At large conductance, ordinary geometric scaling gives

$$\beta(g) \longrightarrow d-2.$$

For the time-reversal-invariant orthogonal class, coherent return paths add a negative weak-localization correction. Its large-$g$ structure is

$$\beta(g) = d-2 -\frac{a_d}{g} +O\!\left(g^{-2}\right),$$

where $a_d>0$ depends on the conductance, degeneracy, and cutoff convention. The sign is robust; the bare coefficient is not convention-free.

In two dimensions, the classical term vanishes. To leading logarithmic accuracy,

$$\frac{dg}{d\ln L} \simeq -a_2,$$

and therefore

$$g(L) \simeq g(\ell) -a_2 \ln \left( \frac{L}{\ell} \right).$$

This is the scaling form of the weak-localization correction. Extrapolating until $g$ becomes order unity gives only an estimate,

$$\xi \sim \ell \exp\!\left[ \frac{ g(\ell)-g_{\mathrm{cross}} }{ a_2 } \right].$$

Here $g_{\mathrm{cross}}\sim1$ marks the limit of the weak-coupling estimate. The exponential explains how a two-dimensional sample can look metallic over every accessible length even though the asymptotic orthogonal-class flow is insulating.

Deep in a localized regime,

$$g_{\mathrm{typ}}(L) \sim A(L) \exp\!\left( -\frac{L}{\xi} \right),$$

Here $\xi$ is defined from conductance decay, and $A(L)$ varies at most algebraically. If the amplitude length is used instead, the exponential is $e^{-2L/\xi_{\mathrm{amp}}}$; this factor does not change the following asymptote. Hence

$$\beta(g) \sim -\frac{L}{\xi} \sim \ln g$$

as $g\to0$. The flow becomes increasingly negative because adding length multiplies exponentially small transmission factors.

Schematic conductance flow. In the conventional orthogonal class, one and two dimensions flow toward localization, while a three-dimensional curve can cross an unstable fixed point $g_c$. Arrows on either side of $g_c$ show the direction of increasing sample size. Curve shapes away from controlled asymptotes are qualitative.

Linearized fixed-point flow

It is convenient to write

$$x \equiv \ln g.$$

Near an isolated fixed point $x_c=\ln g_c$, write $\delta x=x-x_c$. Then

$$\frac{d\delta x}{d\ln L} =y_g\delta x+O(\delta x^2).$$

with

$$y_g = \left. \frac{d\beta}{d\ln g} \right|_{g_c}.$$

The solution is

$$x(L)-x_c \propto L^{y_g}.$$

An unstable transition has $y_g>0$. Matching the scale at which a microscopic detuning becomes order unity gives

$$\nu = \frac{1}{y_g} = \left[ \left. \frac{d\beta}{d\ln g} \right|_{g_c} \right]^{-1}.$$

This relation assumes one relevant direction. If two independent relevant couplings are present, one scalar slope cannot determine the full critical behavior.

One, Two, and Three Dimensions

Dimension and class	Expected flow
1D, generic random disorder	Toward localization; long multichannel wires can have a diffusive interval
2D, orthogonal	Toward localization, sometimes logarithmically slowly
2D, unitary without topology	Higher-order corrections still localize
2D, symplectic	Metallic and insulating regimes can occur
2D, quantum Hall	Additional topological scaling data are needed
3D, conventional classes	An unstable fixed point can separate metal and insulator

One dimension

For a generic one-dimensional random potential, repeated coherent backscattering drives transmission downward at every scale. A single spinless propagating channel has $g\le1$, so a large-$g$ diffusive starting point is not available in every one-dimensional system. In a quasi-one-dimensional wire with many channels, the localization length can be parametrically larger than the transport mean free path,

$$\xi_{\mathrm{q1D}} \sim N\ell,$$

up to symmetry and boundary factors. The wire can therefore support a long diffusive crossover before exponential localization appears.

“All states localize in one dimension” is not a theorem about every Hamiltonian called one-dimensional. Correlated disorder, special chiral points, quasiperiodicity, protected boundary channels, long-range hopping, and non-Hermitian transport can invalidate the generic assumptions. The Hamiltonian and symmetry class must precede the slogan.

Two dimensions

Two dimensions are marginal at the classical level because $d-2=0$. In the orthogonal class, the negative interference correction makes the beta function negative at large $g$, and the conventional flow has no metallic fixed point. The approach can be logarithmically slow before $L$ reaches the very large localization length.

Breaking time reversal removes the elementary Cooperon contribution, but that fact alone does not guarantee an ordinary two-dimensional metal. In the conventional unitary class, higher-order localization corrections remain. A strong magnetic field can also create Landau levels and a topological term, producing the distinct integer quantum Hall flow with critical extended energies.

Strong spin–orbit coupling places a time-reversal-invariant system in the symplectic class. The weak correction can then be antilocalizing, and a two-dimensional metallic phase separated from an insulator by a critical point is possible. This is why the dimension alone never determines the beta function; dimension and symmetry class must both be specified.

Three dimensions

For $d=3$,

$$\beta(g) \longrightarrow +1 \qquad (g\to\infty),$$

while the localized asymptote is negative. If the flow is continuous, it must cross zero at some $g_c$. The crossing is unstable:

$$\begin{array}{lll} g>g_c &\Longrightarrow& g(L)\ \text{grows},\\[3pt] g<g_c &\Longrightarrow& g(L)\ \text{shrinks}. \end{array}$$

The numerical value of $g_c$ is not universal because it depends on boundaries and conductance normalization. The critical exponent is universal within a specified class. For the standard three-dimensional orthogonal Anderson model, numerical estimates are near $\nu\approx1.57$. For example, Slevin and Ohtsuki (2014) report a combined estimate $1.571$ with an interval $[1.563,1.579]$. Differences between finite-size estimates reflect uncertainty and corrections to scaling, not a permitted dependence of the exact exponent on conductance convention. This exponent does not apply automatically to other symmetry classes or interacting transitions.

At the fixed point, it is $g(L)$ rather than the bulk conductivity that is scale independent. Rearranging the definition of $g$ gives

$$\sigma_c(L) = \frac{e^2}{h} g_cL^{2-d}.$$

Thus in three dimensions $\sigma_c(L)\propto L^{-1}$ and vanishes in the thermodynamic zero-frequency limit. The metallic side instead reaches a nonzero bulk $\sigma(0)$ after the flow leaves the critical region. Confusing these statements leads to the false expectation that an Anderson critical point should look like an ordinary diffusive metal.

Metal–Insulator Transitions

Relevant and irrelevant variables

Let $u$ measure distance from criticality and choose its sign so that $u>0$ is the metallic side. The diverging correlation or localization length is

$$\xi(u) = \xi_0 \lvert u\rvert^{-\nu}.$$

Measure lengths in a fixed microscopic unit. A finite-size ansatz allowing a leading irrelevant variable $v$ is:

$$g(L,u,v) = \mathcal F \left( uL^{1/\nu}, vL^y \right).$$

Here $y<0$ makes the correction decay with size. At $u=0$, the leading term is scale invariant,

$$g(L,0,0)=g_c,$$

but the irrelevant term produces crossing drift at finite $L$. A single clean crossing is suggestive, not a complete exponent analysis.

Near criticality one may expand

$$\ln g(L,u) \simeq \ln g_c +A uL^{1/\nu} +B L^y +\cdots.$$

Fits should vary the minimum size, polynomial order, irrelevant structure, disorder sample count, and covariance treatment. Stability under those choices is part of the result.

Conductivity exponent

On the metallic side in $d>2$, match the critical block at $L\sim\xi$ to ordinary transport at larger scales:

$$g_c \sim \frac{ \sigma(0)\xi^{d-2} }{ e^2/h }.$$

Therefore

$$\sigma(0) \propto \xi^{2-d} \propto u^\mu,$$

with the Wegner scaling relation

$$\mu = (d-2)\nu.$$

This relation belongs to the noninteracting one-parameter framework. Interactions, dangerous irrelevant variables, or additional dynamical scaling can change the relation between a correlation-length exponent and a measured conductivity exponent.

Transfer-matrix finite-size scaling

A common numerical route uses a very long bar with transverse width $M$. From the smallest positive Lyapunov exponent one obtains a quasi-one-dimensional decay length $\lambda_M$. The normalized quantity

$$\Lambda_M \equiv \frac{\lambda_M}{M}$$

obeys a scaling form

$$\Lambda_M = \mathcal G \left( uM^{1/\nu}, vM^y \right).$$

On the localized side, $\Lambda_M$ decreases with $M$; on the metallic side it increases; at criticality it approaches a class-dependent constant. The MacKinnon–Kramer recursive transfer-matrix method made this crossing logic quantitatively practical. Reliable extraction still requires long-bar convergence, uncertainty in Lyapunov exponents, irrelevant fields, energy-window control, and independent fit variants.

A discrete beta-function estimate can be formed from matched sizes:

$$\beta_{\mathrm{eff}} \left( g;L,b \right) = \frac{ \ln g(bL)-\ln g(L) }{ \ln b }.$$

This estimate is meaningful only if both points lie on the same branch of a one-parameter flow. Averaging data from opposite sides of a transition can manufacture a smooth but fictitious beta function.

Broad distributions and critical states

Near localization, conductance and wavefunction intensities are broadly distributed. One-parameter scaling is more faithfully expressed as a flow of the probability distribution,

$$P_L(g) \longrightarrow P_{bL}(g),$$

whose long-distance form is controlled by the same relevant coordinate. At a critical point, the distribution can approach an $L$-independent non-Gaussian shape while wavefunctions remain multifractal. Equality of mean conductance alone does not establish equality of distributions.

For independent short-range disorder under the assumptions of the Chayes et al. (1986) bound, the finite-size correlation-length exponent obeys

$$\nu_{\mathrm{FS}} \ge \frac{2}{d}$$

The usual one-parameter Anderson analysis identifies $\nu_{\mathrm{FS}}$ with the exponent $\nu$ above. A fitted small-system exponent or a model with long-range correlated disorder must be examined before applying that identification.

What an experiment actually scales

An electronic experiment cannot increase coherent size indefinitely. Sample size, dephasing and escape limit the paths that contribute. Thermal energy averaging can blur an energy-dependent transition, but there is no universal rule that replaces every length in this theory by $\min(L,L_\phi,L_T)$. For example, the systematic weak-localization correction and sample-specific conductance fluctuations respond differently to thermal averaging.

Temperature can also change density, screening, interactions and the disorder landscape. A collapse versus temperature is therefore not automatically the zero-temperature flow of a noninteracting Anderson model. Compare matched sizes and disorder statistics, include crossing drift, and test both the localized and metallic branches with an independent spatial or spectral observable.

Relation to RG and Quantum Field Theory

The original scaling argument is renormalization-group reasoning expressed directly in an observable. Increasing $L$ removes sensitivity to short-distance details and moves the system along a trajectory in conductance space. Fixed points, relevant perturbations, irrelevant corrections, universality classes, and critical exponents have their usual renormalization-group meaning.

A microscopic field-theory derivation begins with disorder-averaged products of retarded and advanced Green functions. Replica, supersymmetry, or Keldysh methods can perform the disorder average while preserving access to correlation functions. The long-wavelength diffuson and Cooperon modes are encoded in a constrained matrix field

$$Q^2=1.$$

For example, after rescaling position to a dimensionless coordinate $\mathbf x$, a gradient term has the schematic form

$$S_{\mathrm{grad}}[Q] =\frac{1}{16t}\int d^dx\, \operatorname{Tr}(\nabla_{\!x}Q)^2.$$

The dimensionless coupling $t$ is proportional to $1/g$ at the running scale. This specification matters: in unrescaled physical coordinates the gradient coefficient carries a power of length away from two dimensions. The target space of $Q$ depends on symmetry. Frequency terms regulate long-time propagation, and topological terms can distinguish flows with the same longitudinal conductance. See Evers and Mirlin (2008) for the full constructions and normalization choices.

Integrating a momentum shell and rescaling changes $t$. Since $t\propto1/g$,

$$\frac{d\ln t}{d\ln L} = -\beta(g).$$

Perturbation theory about small $t$ reproduces weak localization and its symmetry-dependent sign. Expansion near $d=2+\epsilon$ organizes the Anderson fixed point when controlled. Nonperturbative configurations and topological terms are essential in several two-dimensional classes.

The field theory therefore explains why symmetry, topology and interactions cannot always be compressed into one conductance: they determine the allowed fields and additional couplings. Rare events and multifractality also require information beyond a saddle-point conductivity. The renormalization-group introduction develops relevant and irrelevant directions more generally.

Exercises

1. Exponentially large two-dimensional crossover

Use

$$\frac{dg}{d\ln L} = -a_2$$

with $a_2=2/\pi$, $g(\ell)=12$, and $g_{\mathrm{cross}}=1$. Estimate $\xi/\ell$.

Solution

Integration gives

$$g(L) = 12-\frac{2}{\pi} \ln \left( \frac{L}{\ell} \right).$$

Set $g(\xi)=1$:

$$\ln \left( \frac{\xi}{\ell} \right) = \frac{11}{2/\pi} = \frac{11\pi}{2} \simeq 17.28.$$

Therefore

$$\frac{\xi}{\ell} \simeq e^{17.28} \simeq 3.2\times10^7.$$

The precise number is not controlled because the weak-coupling equation was extrapolated to $g\sim1$. The robust lesson is the exponential scale separation.

2. Localized asymptote

Suppose

$$g(L) = \left( \frac{L}{\ell} \right)^p \exp \left( -\frac{L}{\xi} \right).$$

Compute $\beta(g)$ as a function of $L$ and show that $\beta\sim\ln g$ for $L\gg\xi$.

Solution

Taking the logarithm,

$$\ln g = p\ln \left( \frac{L}{\ell} \right) -\frac{L}{\xi}.$$

Therefore

$$\beta = \frac{d\ln g}{d\ln L} = p-\frac{L}{\xi}.$$

For $L/\xi\gg1$, the linear term dominates both expressions:

$$\ln g \simeq -\frac{L}{\xi}, \qquad \beta \simeq -\frac{L}{\xi}.$$

Hence $\beta\sim\ln g$ up to subleading logarithmic and constant terms.

3. Fixed-point slope and correlation length

Near an unstable fixed point, let

$$\beta = y_g \left( \ln g-\ln g_c \right), \qquad y_g>0.$$

Show that $\nu=1/y_g$.

Solution

Let

$$\delta x = \ln g-\ln g_c.$$

The flow equation becomes

$$\frac{d\delta x}{d\ln L} = y_g\delta x,$$

so

$$\delta x(L) = \delta x(\ell) \left( \frac{L}{\ell} \right)^{y_g}.$$

If the microscopic detuning is proportional to $u$, the crossover occurs when

$$\lvert u\rvert \xi^{y_g} \sim1.$$

Thus

$$\xi \propto \lvert u\rvert^{-1/y_g}.$$

Comparison with $\xi\propto\lvert u\rvert^{-\nu}$ gives

$$\nu = \frac{1}{y_g}.$$

4. A discrete beta estimate

Matched samples give $g(16a)=3.40$ and $g(32a)=3.15$. Estimate $\beta_{\mathrm{eff}}$ between the two sizes and interpret its sign.

Solution

Here $b=2$, so

$$\begin{aligned} \beta_{\mathrm{eff}} &= \frac{\ln(3.15)-\ln(3.40)}{\ln2}, \\ &\simeq -0.110. \end{aligned}$$

The negative value means the chosen conductance statistic decreases over this scale interval. By itself, one interval does not prove asymptotic localization; larger sizes and a check of one-parameter collapse are required.

Continue along the two branches

• Anderson Insulators follows localized orbitals into finite-temperature hopping transport.
• Mobility Edges locates the transition in energy and disorder strength.
• Finite-Size Scaling in Numerics develops fits with irrelevant corrections, correlated uncertainties and stability checks.

References

• E. Abrahams, P. W. Anderson, D. C. Licciardello, and T. V. Ramakrishnan, “Scaling Theory of Localization: Absence of Quantum Diffusion in Two Dimensions,” Physical Review Letters 42, 673–676 (1979), doi:10.1103/PhysRevLett.42.673. Introduces the one-parameter conductance beta function and dimensional flow argument.
• J. T. Chayes, L. Chayes, D. S. Fisher, and T. Spencer, “Finite-Size Scaling and Correlation Lengths for Disordered Systems,” Physical Review Letters 57, 2999–3002 (1986), doi:10.1103/PhysRevLett.57.2999. Proves a disorder-sensitive bound on correlation-length scaling.
• F. Evers and A. D. Mirlin, “Anderson Transitions,” Reviews of Modern Physics 80, 1355–1417 (2008), doi:10.1103/RevModPhys.80.1355. Reviews symmetry classes, field theory, multifractality, and criticality.
• S. Hikami, “Anderson Localization in a Nonlinear-Sigma-Model Representation,” Physical Review B 24, 2671–2679 (1981), doi:10.1103/PhysRevB.24.2671. Develops perturbative localization flow in the nonlinear-sigma-model language.
• A. MacKinnon and B. Kramer, “One-Parameter Scaling of Localization Length and Conductance in Disordered Systems,” Physical Review Letters 47, 1546–1549 (1981), doi:10.1103/PhysRevLett.47.1546. Establishes recursive transfer-matrix finite-size scaling.
• K. Slevin and T. Ohtsuki, “Corrections to Scaling at the Anderson Transition,” Physical Review Letters 82, 382–385 (1999), doi:10.1103/PhysRevLett.82.382. Demonstrates the importance of irrelevant scaling fields in precision exponent extraction.
• K. Slevin and T. Ohtsuki, “Critical Exponent for the Anderson Transition in the Three-Dimensional Orthogonal Universality Class,” New Journal of Physics 16, 015012 (2014), doi:10.1088/1367-2630/16/1/015012. Refines transfer-matrix estimates and systematic-error controls.
• F. J. Wegner, “Electrons in Disordered Systems. Scaling near the Mobility Edge,” Zeitschrift für Physik B 25, 327–337 (1976), doi:10.1007/BF01315248. Establishes critical scaling relations near a mobility edge.