research35 min studyLesson 556 in Physics

Dynamical mean-field self-consistency

Map a lattice model to a dynamical impurity bath and state precisely which limit makes a local self-energy exact.

On this page

What you’ll understand

  • Specify the large-coordination hopping scaling.
  • Close the impurity and local lattice Green functions self-consistently.
  • Distinguish exact limiting DMFT from a finite-dimensional approximation and its solver errors.

Dynamical mean-field theory keeps the time-dependent quantum fluctuations of one site while replacing the rest of a lattice by a self-consistent bath. Its name can suggest a static average field, but the bath is an entire function of energy or time. The construction becomes exact in a specified infinite-coordination limit with scaled hopping. In ordinary finite dimensions, locality of self-energy is an approximation whose usefulness depends on the observable and the missing spatial correlations.

Take a nontrivial large-coordination limit

Let zc denote the number of nearest neighbors and scale a bond hopping as a fixed energy t star divided by the square root of zc. The sum of squared hopping amplitudes then remains finite as coordination grows. Keeping hopping fixed instead would make the kinetic-energy scale grow and would define a different singular limit. The interaction U stays local and unscaled in the usual single-band Hubbard construction. This limit is a property of a family of models, not a claim that a cubic material literally has infinitely many neighbors.

For appropriate lattices and local interactions, irreducible self-energy becomes local in that limit while retaining its full energy dependence. Locality does not make every correlation function local: propagation through the lattice still produces momentum-dependent Green functions and response structure. Single-site DMFT retains nontrivial local charge and spin dynamics. A purely static Hartree shift is therefore a much stronger approximation and cannot reproduce the general dynamical bath or the separated coherent and Hubbard-band spectral features.

tij=t∗zc,∑j neighbors∣tij∣2=t∗2,Σij(z)⟶δijΣ(z)t_{ij}=\frac{t_*}{\sqrt{z_c}},\qquad \sum_{j\ {\rm neighbors}}|t_{ij}|^2=t_*^2,\qquad \Sigma_{ij}(z)\longrightarrow\delta_{ij}\Sigma(z)
zc is coordination, not complex energy; t* is an energy and z is a complex energy in the Green functions. Locality is exact in the specified large-coordination limit for the usual local-interaction setting.↔ Scroll wide equations horizontally.

Close the impurity and lattice equations

Assume a local self-energy and use it in the lattice momentum sum, or equivalently in an integral over the noninteracting density of states. This gives the local lattice Green function. Independently define an interacting impurity with the same on-site interaction and an adjustable hybridization function describing exchange with a noninteracting bath. Solving the impurity yields its Green function and self-energy. Self-consistency requires the impurity Green function to equal the local lattice one.

The bath is often written through a noninteracting impurity Green function, called the Weiss function. Its inverse contains the hybridization, and subtracting the interacting impurity inverse gives the self-energy. A practical iteration updates these functions, solves an interacting impurity problem and repeats until suitable residuals converge. A converged numerical loop alone does not certify physical uniqueness, thermodynamic stability or spectral accuracy. Coexisting solutions and solver discretization, sampling, continuation and truncation errors need separate checks.

Gloc(z)=∫dϵ ρ0(ϵ)z+μ−ϵ−Σ(z),G0−1=z+μ−Δ(z),Σ=G0−1−Gimp−1,Gimp=GlocG_{\rm loc}(z)=\int d\epsilon\,\frac{\rho_0(\epsilon)}{z+\mu-\epsilon-\Sigma(z)},\quad \mathcal G_0^{-1}=z+\mu-\Delta(z),\quad \Sigma=\mathcal G_0^{-1}-G_{\rm imp}^{-1},\quad G_{\rm imp}=G_{\rm loc}
ρ0 is a normalized noninteracting density of states with inverse-energy units. G has inverse-energy units, while Σ and bath hybridization Δ have energy units. The impurity interaction is solved dynamically, not replaced by this algebra alone.↔ Scroll wide equations horizontally.

Use a lattice-specific closure and state its limits

For an infinite-coordination Bethe lattice with the semicircular noninteracting density of states of half-bandwidth twice t star, self-consistency simplifies: hybridization equals squared t star times the local Green function. That convenient relation is lattice specific. A different density of states generally requires the full integral, and inserting the same shortcut for a square lattice changes the problem. The noninteracting limit provides a clean branch and units check before using an interacting impurity solver.

In finite-dimensional correlated metals, a single-site approximation omits momentum-dependent self-energy effects, short-range singlets and some collective fluctuations. Cluster or diagrammatic extensions can recover selected spatial structure, at further cost and with their own checks. A Mott boundary must specify the density of states, temperature, magnetic constraints and solver. The optimization lab only illustrates convergence of one elementary iteration; its controls are not U, filling, bath discretization or an impurity solver and cannot validate a DMFT phase diagram.

Δ(z)=t∗2Gloc(z)(semicircular Bethe closure),U=0, μ=0: G=1z−t∗2G\Delta(z)=t_*^2G_{\rm loc}(z)\quad\text{(semicircular Bethe closure)},\qquad U=0,\ \mu=0:\ G=\frac1{z-t_*^2G}
The Bethe closure uses a semicircular band of half-width 2t*. For an upper-half-plane energy, choose the Green-function branch with negative imaginary part and asymptotic behavior G(z) ∼ 1/z.↔ Scroll wide equations horizontally.

A noninteracting causal branch check

For the noninteracting Bethe closure, take t∗=1 eVt_*=1\,\mathrm{eV}, μ=0\mu=0 and z=i eVz=i\,\mathrm{eV}. Seek G=−iy eV−1G=-iy\,\mathrm{eV^{-1}} with positive y. Find the physical solution and the hybridization.

  1. Insert the proposed imaginary form into G=1/(z−t∗2G)G=1/(z-t_*^2G). The denominator is i(1+y) eVi(1+y)\,\mathrm{eV}, so the dimensionless equation is y=1/(1+y)y=1/(1+y).

  2. Rearrange to y2+y−1=0y^2+y-1=0. The positive solution is y=(5−1)/2≈0.618y=(\sqrt5-1)/2\approx0.618. The other root would give a Green function with positive imaginary part and is not the retarded branch in the upper half-plane.

  3. The hybridization is Δ=t∗2G=−0.618i eV\Delta=t_*^2G=-0.618i\,\mathrm{eV}. Substituting back gives inverse denominator −i/(1.618 eV)=−0.618i eV−1-i/(1.618\,\mathrm{eV})=-0.618i\,\mathrm{eV^{-1}}, confirming the closure without solving any interacting impurity.

The result

The physical local Green function is −0.618i eV⁻¹ and its bath hybridization is −0.618i eV in this noninteracting check.

Explore the model

Interactive lab

Interactive lab · optimization

Iteration convergence as a separate numerical issue

Reset and set BOTH Step size η = 0.50 and Initial position x₀ = 2; predict convergence of xₙ = (1 − η)ⁿx₀. Reset and set η = 2.20 and x₀ = 2; predict alternating growth and eventual out-of-range reporting. This is gradient descent on x²/2, not a DMFT iteration, impurity solver or proof of thermodynamic stability.

Iteration convergence as a separate numerical issueGradient descent on f(x) = x²/2: xₙ = (1 − η)ⁿx₀. Playback advances two discrete iterations per second. Values outside |x| ≤ 4 are explicitly reported. Current readings: Position x0: 2; Objective x²/2: 2; Iteration behavior: Stable step. All parameters are adjustable with the labeled controls below.−4−202402.174.356.528.7Parameter xf(x) = x²/2n = 0
ObjectiveDiscrete iterate
Position x0
2
Objective x²/2
2
Iteration behavior
Stable step
n = 0 · t = 0 s

Gradient descent on f(x) = x²/2: xₙ = (1 − η)ⁿx₀. Playback advances two discrete iterations per second. Values outside |x| ≤ 4 are explicitly reported. Parameter changes pause playback and recompute the model at the current elapsed time. Reset restores both parameters and returns time to zero. Playback starts paused. Adjust either parameter, then choose Play or Step.

Look forAn iteration needs convergence checks, while the physical approximation needs separate validation.

Check your understanding

What does single-site DMFT retain despite its local self-energy assumption?

Try it yourself

At one bath-update step for the Bethe closure, let t∗=0.50 eVt_*=0.50\,\mathrm{eV} and Gloc=(0.20−0.40i) eV−1G_{\rm loc}=(0.20-0.40i)\,\mathrm{eV^{-1}}. Find Δ and the next Weiss inverse at z=0.50i eVz=0.50i\,\mathrm{eV} with μ = 0. Does this single update establish convergence?

Take a moment to make a prediction or work it out on paper. Open a hint when you need one.

Give me a hint

Multiply G by the squared hopping scale, then subtract Δ from z + μ. The supplied Green function need not already be a fixed point.

Compare with the solution

The squared hopping is 0.25 eV20.25\,\mathrm{eV^2}, giving Δ=(0.050−0.100i) eV\Delta=(0.050-0.100i)\,\mathrm{eV}. The Weiss inverse is z−Δ=(−0.050+0.600i) eVz-\Delta=(-0.050+0.600i)\,\mathrm{eV}. These are one-step algebraic updates, not a converged solution: one must solve the interacting impurity for this bath and test agreement with the next local lattice Green function. Causal signs and units are useful diagnostics, but neither proves uniqueness, stability or an accurate many-body spectrum.

Take it with you

  • The infinite-coordination limit requires explicit hopping scaling.
  • DMFT is a dynamical self-consistent impurity construction, not a static average potential.
  • The Bethe shortcut and finite-dimensional accuracy require separate model-specific assumptions.

Read further

Original sources and deeper explanations to follow at your own pace.

Optional progress, saved in this browser when available. Revisit any lesson at any time.

Lesson not marked complete.