Density of States

Definition

The density of states is the number of available quantum states per unit energy:

g(E) = dN/dE

Here N(E) is the cumulative count of single-particle states with energy below E, and g(E) is its derivative — how fast new states appear as you climb the energy axis. In a solid we usually quote it per unit volume, so the units are states per joule per cubic meter (or, more conveniently, states per eV per atom).

The crucial point: g(E) counts seats, not occupants. A state existing at energy E says nothing about whether an electron sits there. To get the actual electron density you multiply by the occupation probability — the Fermi-Dirac function f(E) — so the number of electrons in a window dE is g(E)·f(E)·dE. Almost every electronic property of a material is some weighted integral of g(E) against f(E).

How it works — counting states in k-space

For a free electron in a box, the allowed wavevectors k form a uniform grid in momentum space, each cell occupying a tiny fixed volume (2π/L)^d in d dimensions. The energy depends only on the magnitude of k:

E = ℏ²k² / (2m)   ⟹   k = √(2mE)/ℏ ∝ √E

So a fixed energy E corresponds to a "shell" of constant |k| in k-space — a sphere in 3D, a circle in 2D, two points in 1D. To count states up to energy E, you measure how much k-space lies inside that shell, then divide by the cell volume. Differentiating with respect to E hands you g(E). The dimensionality of the shell is what makes each case behave so differently:

• 3D: states fill a sphere, volume ∝ k³ ∝ E^(3/2). So N(E) ∝ E^(3/2) and g(E) ∝ E^(1/2) — the famous √E rise.
• 2D: states fill a disk, area ∝ k² ∝ E. So N(E) ∝ E and g(E) = constant — energy-independent.
• 1D: states fill a line segment, length ∝ k ∝ E^(1/2). So N(E) ∝ E^(1/2) and g(E) ∝ E^(-1/2) — a divergence at the band bottom.

The pattern is clean: each lost dimension knocks one power of √E off g(E). The same factor of two for spin multiplies all three, but does not change the energy dependence.

Dimensional comparison

Dimension	k-shell	N(E) ∝	g(E) ∝	Shape vs. E	Realized in
3D bulk	Sphere	E^(3/2)	E^(1/2)	Rising √E curve from 0	Ordinary metals, doped semiconductors
2D sheet	Circle	E	E⁰ (constant)	Flat staircase, one step per subband	Quantum wells, 2DEG, graphene channels
1D wire	Two points	E^(1/2)	E^(-1/2)	Diverging spike at each band bottom	Nanotubes, quantum wires, polymers
0D dot	Single point	step	Sum of delta functions	Discrete spikes, "artificial atom"	Quantum dots, single molecules
Band edge (any d)	Flat ∇ₖE = 0	kink/jump	van Hove singularity	Sharp peak or kink	Real crystals, ARPES/STM spectra
Saddle point (2D)	Flat ∇ₖE = 0	—	log divergence	Logarithmic spike	Graphene, cuprate Fermi surfaces

Worked example — sodium's g(E_F)

Take sodium, a near-perfect free-electron metal with one valence electron per atom. Its conduction electrons number n ≈ 2.65 × 10²⁸ per m³, giving a Fermi energy E_F ≈ 3.24 eV. For a 3D free-electron gas the density of states at the Fermi level is

g(E_F) = (3/2) · n / E_F

This compact form follows because N ∝ E^(3/2) means g(E) = (3/2)·N(E)/E. Plugging in:

g(E_F) = 1.5 × (2.65×10²⁸ m⁻³) / (3.24 eV)
       ≈ 1.23 × 10²⁸ states · eV⁻¹ · m⁻³
       ≈ 7.6 × 10⁴⁶ states · J⁻¹ · m⁻³

Now feed that into the electronic heat capacity. At room temperature k_B·T ≈ 0.025 eV, far smaller than E_F = 3.24 eV, so only a sliver of electrons — about g(E_F)·k_B·T per unit volume — can absorb heat. The Sommerfeld result is

C_el = (π²/3) · g(E_F) · k_B² · T   →   γ = C_el/T = (π²/3) g(E_F) k_B²

For sodium this predicts γ ≈ 1.1 mJ·mol⁻¹·K⁻², within a factor of order one of the measured 1.38 mJ·mol⁻¹·K⁻². The leftover discrepancy is precisely the "effective mass" correction — the lattice and electron interactions renormalize m, and hence g(E_F), away from the bare free-electron value. So a single thermodynamic measurement reads off the density of states at the Fermi level.

Variants and regimes

• Free-electron (Sommerfeld) gas. The textbook √E in 3D. Excellent for the alkali metals, aluminum, and lightly doped semiconductors near a band edge.
• Effective-mass parabolic bands. Same √E shape but with m replaced by an effective mass m*. A heavy band (large m*) means a large g(E_F): flat bands pack in states. This is why "heavy-fermion" compounds have enormous specific heats.
• Tight-binding / real crystals. The dispersion is no longer parabolic, so g(E) develops band edges, gaps, and van Hove peaks. The total DOS is the sum over all bands; gaps are regions where g(E) = 0.
• Reduced dimensions. Confine electrons to a sheet and you get the flat 2D staircase; confine to a wire and you get the 1D E^(-1/2) spikes; confine to a dot and the spectrum collapses to discrete delta-function levels. Nanostructures let you engineer g(E) by geometry.
• Linear (Dirac) dispersion. In graphene E ∝ |k|, so in 2D g(E) ∝ |E| — it vanishes linearly at the Dirac point rather than being constant. Dimensionality alone is not the whole story; the band shape matters too.
• Phonon DOS. The same machinery applies to lattice vibrations. In the Debye model g(ω) ∝ ω² in 3D, the acoustic analogue of √E, and it governs the lattice heat capacity that dominates at higher temperature.

JavaScript — computing g(E) by counting in k-space

// Free-electron density of states by binning a k-space grid.
// Works for d = 1, 2, 3. Energy units: E = |k|^2 (set hbar^2/2m = 1).
function densityOfStates(dim, kMax, nGrid, nBins) {
  const counts = new Array(nBins).fill(0);
  const eMax = kMax * kMax;
  const dE = eMax / nBins;
  const dk = (2 * kMax) / nGrid;          // grid spacing per axis

  // Recursively sweep a dim-dimensional grid of k-points.
  function sweep(coords) {
    if (coords.length === dim) {
      const k2 = coords.reduce((s, k) => s + k * k, 0);
      const E = k2;
      if (E > 0 && E < eMax) {
        const bin = Math.min(nBins - 1, Math.floor(E / dE));
        counts[bin] += 2;                  // factor 2 for spin
      }
      return;
    }
    for (let i = 0; i < nGrid; i++) {
      sweep([...coords, -kMax + i * dk]);
    }
  }
  sweep([]);

  // g(E) = (states in bin) / (bin width * total k-space measured)
  const cellVolume = Math.pow(dk, dim);
  return counts.map((c, i) => ({
    E: (i + 0.5) * dE,
    g: c / (dE / cellVolume),             // states per unit energy (unnormalized)
  }));
}

// 3D should rise like sqrt(E); 2D should be flat; 1D should fall like 1/sqrt(E).
const dos3D = densityOfStates(3, 4, 60, 24);
console.log('3D ratio g(4E0)/g(E0):', (dos3D[19].g / dos3D[4].g).toFixed(2)); // ~2 (sqrt of 4)

const dos2D = densityOfStates(2, 4, 400, 24);
console.log('2D is flat? early/late:', (dos2D[3].g / dos2D[20].g).toFixed(2)); // ~1

const dos1D = densityOfStates(1, 4, 20000, 24);
console.log('1D diverges at low E:', (dos1D[1].g / dos1D[20].g).toFixed(2));   // >1, grows as E->0

The grid count is crude but it makes the scaling visible: quadruple the energy in 3D and the binned g roughly doubles (√4 = 2); the 2D result stays flat; the 1D result piles up near the band bottom.

Where the density of states shows up

• Electronic heat capacity. The linear-in-T term C_el = (π²/3)g(E_F)k_B²T is a direct readout of g(E_F). Heavy-fermion metals have γ values hundreds of times larger than copper because their bands are nearly flat.
• Electrical conductivity. Only electrons near E_F carry current, and the conductivity in Boltzmann transport is proportional to g(E_F) times the squared velocity and scattering time. A material with zero g(E_F) is an insulator.
• Optical absorption and color. Absorption strength tracks the joint density of states between occupied and empty bands; van Hove peaks produce the characteristic absorption edges that give materials their color.
• Superconductivity. The BCS transition temperature depends exponentially on g(E_F): T_c ∝ exp(−1/[g(E_F)V]). Pushing a van Hove singularity to the Fermi level is a recipe for raising T_c.
• Magnetism. The Stoner criterion for spontaneous ferromagnetism is g(E_F)·I > 1 — a large density of states at the Fermi level destabilizes the paramagnet. This is why iron, with its peaked d-band DOS, is magnetic.
• Semiconductor devices. Carrier concentrations, threshold voltages, and the 2D staircase DOS in quantum-well lasers and HEMTs all follow directly from g(E) for the relevant dimensionality.
• Spectroscopy. Scanning tunneling spectroscopy's dI/dV is proportional to the local DOS; ARPES reconstructs the bands and hence g(E). The spikes you see in the data are van Hove singularities.

Common pitfalls and misconceptions

• Confusing g(E) with occupied states. g(E) counts available seats, not electrons. The number of electrons is g(E)·f(E); a huge g(E) above E_F is irrelevant at low temperature because f(E) ≈ 0 there.
• Forgetting spin. Each k-state holds two electrons (spin up and down). Drop the factor of two and your g(E) — and every property derived from it — is off by half.
• Assuming dimensionality alone fixes the power law. The √E, constant, and E^(-1/2) rules assume parabolic free-electron bands. Graphene's linear dispersion gives g(E) ∝ |E| in 2D, not a constant. The band shape matters as much as the dimension.
• Treating van Hove singularities as physical infinities. The 1D divergence and 2D log spike are integrable — the total number of states stays finite. They are sharp peaks, not unphysical infinities, and real broadening (temperature, disorder, finite lifetime) rounds them off.
• Thinking a band gap means g(E) is small. Inside a gap g(E) is exactly zero, not merely small. That is the defining difference between a metal (finite g(E_F)) and an insulator (zero g(E_F)).
• Mixing up g(E) and the dispersion E(k). The dispersion is the input; g(E) is the derived count. Flat regions of E(k) — where the group velocity vanishes — are exactly where g(E) spikes.

Performance and derivation analysis

The clean way to get every result is to write the count as an integral over the constant-energy surface, weighting by how slowly the energy changes:

g(E) = (1/(2π)^d) ∮_{E(k)=E} dS / |∇ₖ E(k)|

This single formula contains everything. The numerator dS is the size of the constant-energy shell (sphere, circle, or point pair), and the denominator is the gradient of the dispersion — the group velocity. Two consequences fall out immediately:

• The √E family. For E ∝ k², the gradient |∇ₖE| ∝ k ∝ √E and the shell measure dS scales as k^(d−1) ∝ E^((d−1)/2). Dividing gives g(E) ∝ E^((d−1)/2 − 1/2) = E^((d−2)/2): that is E^(1/2) in 3D, E⁰ in 2D, and E^(-1/2) in 1D. The three rules are one formula evaluated at d = 3, 2, 1.
• van Hove singularities. Wherever ∇ₖE → 0 the denominator vanishes and g(E) spikes. These flat spots occur at band bottoms, tops, and saddle points. In 1D the band edge gives a hard E^(-1/2) divergence; in 2D a band edge gives a step and a saddle gives a logarithmic divergence; in 3D the slope of g(E) jumps. Léon van Hove proved in 1953 that any periodic crystal must have at least these critical points, so every real band structure carries these features.

Computationally, evaluating g(E) for a real material means sampling E(k) on a fine k-mesh and histogramming, then improving the bins with the linear-tetrahedron method to capture the van Hove features without artificial smearing. The cost scales with the number of k-points times bands; for a typical density-functional calculation a 30×30×30 mesh over a dozen bands is enough to resolve the peaks that drive magnetism and superconductivity. The payoff is large: from one g(E) curve you can read off the carrier density (integrate to E_F), the heat capacity coefficient (the value at E_F), the optical edges (joint DOS), and the likelihood of an instability (peaks near E_F) — which is why g(E) is the first thing a condensed-matter physicist computes.