the Stieltjes-transform / resolvent method
/ Stieltjes = STEEL-tyes /
The Stieltjes-transform method is the analytic counterpart to the combinatorial moment method, and it is the engine behind almost all modern random matrix results — local laws, universality, edge behaviour. The idea is to study a measure not through its moments but through a single complex-analytic function that encodes it, the Stieltjes transform, and to derive a self-consistent equation for that function's deterministic limit. Where the moment method counts walks, this method tracks a resolvent.
For a probability measure mu on R, its Stieltjes transform is the function m(z) = integral of 1/(x - z) dmu(x), defined for z in the upper half plane (where the integral is finite and m is analytic). The transform determines mu uniquely, and the measure is recovered by the inversion formula: the density is the boundary value rho(x) = (1/pi) Im m(x + i0). For a Hermitian matrix H_N the relevant object is the resolvent (also called the Green's function) G(z) = (H_N - z I)^(-1); its normalised trace is exactly the Stieltjes transform of the empirical spectral distribution: m_N(z) = (1/N) trace G(z) = (1/N) sum over i of 1/(lambda_i - z). The strategy is now: show m_N(z) concentrates around a deterministic m(z), and find an equation m must satisfy. For a Wigner matrix the celebrated self-consistent equation is m(z) = 1/(-z - m(z)), a quadratic whose solution m(z) = (-z + sqrt(z^2 - 4))/2 is the Stieltjes transform of the semicircle. Inverting recovers the semicircle density. The self-consistent equation comes from a Schur-complement / leave-one-out argument: deleting one row and column changes the resolvent in a controllable way, and averaging gives the fixed-point relation.
Why this method dominates: the resolvent localises in a way moments cannot. By taking z = E + i eta with the imaginary part eta as small as 1/N (up to logarithms), one obtains local laws that control the eigenvalue density on scales of a few eigenvalue spacings, not just the macroscopic bulk. These local laws are the modern route to universality (the four-moment theorem and the Erdos-Schlein-Yau approach both rely on resolvent control), and the same machinery handles sample covariance matrices (the Marchenko-Pastur equation z m^2 + (z + 1 - y) m + 1 = 0 in one convention) and the edge. A caveat: the method requires care about the imaginary part eta. The deterministic equation holds with high probability only for eta above the local scale; pushing eta below the typical eigenvalue spacing is exactly where the analysis is hardest and where individual eigenvalue fluctuations live. The Stieltjes transform is also exactly the Cauchy transform whose inverse defines the R-transform of free probability — the two viewpoints are deeply linked.
Derive the semicircle by fixed point. The Schur-complement argument shows the diagonal resolvent entries each satisfy G_{ii}(z) approximately 1/(-z - m_N(z)), so averaging over i gives m_N(z) approximately 1/(-z - m_N(z)). Passing to the limit yields the exact equation m(z) = 1/(-z - m(z)). Solving the quadratic m^2 + z m + 1 = 0 and taking the root with positive imaginary part gives m(z) = (-z + sqrt(z^2 - 4))/2; its boundary imaginary part is (1/2) sqrt(4 - x^2)/pi up to constants — the semicircle density.
A self-consistent equation for the resolvent's trace pins down the limiting density.
The self-consistent (deterministic) equation holds with high probability only for spectral parameter z = E + i eta with eta above the local eigenvalue scale; below it the resolvent fluctuates and the analysis is genuinely delicate — that regime is exactly where edge and spacing statistics live.