OctoInformBeta

OpenAI preprint, September 2026 · §4. The rank-one comparison · §4

Lemma 4.2: On positive semidefinite Hermitian matrices, with a divided-difference identity for the mixed resolvent integral on nonnegative reals

Resolvent formula and tangent bound for the regularized primitive.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Take n×nn \times n positive semidefinite Hermitian complex matrices X,Y∈ℂn×nX, Y \in \mathbb{C}^{n \times n}, a spectral parameter σ∈(0,1)\sigma \in (0, 1), and a regularization parameter δ>0\delta > 0. For a regularized scalar primitive applied via Hermitian functional calculus, the question is how it is represented as an integral of matrix resolvents and what tangent upper bound compares its values at XX and YY in the Loewner order. On nonnegative real numbers, a second question is how the normalized mixed resolvent integral matches the primitive divided difference.

Definitions

Given σ,δ,y∈ℝ\sigma, \delta, y \in \mathbb{R}, the regularized scalar primitive fσ,δ(y)∈ℝf_{\sigma,\delta}(y) \in \mathbb{R} is the real number fσ,δ(y)=∫0y(t++δ)σ−3/2dt,f_{\sigma,\delta}(y) = \int_0^y (t_+ + \delta)^{\sigma - 3/2}\,dt,(1)✦ where t+=max⁡(t,0)t_+ = \max(t, 0). For a positive semidefinite Hermitian matrix A∈ℂn×nA \in \mathbb{C}^{n \times n}, the matrix fσ,δ(A)∈ℂn×nf_{\sigma,\delta}(A) \in \mathbb{C}^{n \times n} is the value of fσ,δf_{\sigma,\delta} on AA under Hermitian functional calculus.

Associated with any β∈ℝ\beta \in \mathbb{R}, the resolvent normalization constant cβ∈ℝc_\beta \in \mathbb{R} is the real number cβ=(∫0∞vβ(1+v)2dv)−1.c_\beta = \left(\int_0^\infty \frac{v^\beta}{(1 + v)^2}\,dv\right)^{-1}.(2)✦

For β,δ,v∈ℝ\beta, \delta, v \in \mathbb{R} and A∈ℂn×nA \in \mathbb{C}^{n \times n}, the resolvent matrix integrand Rβ,δ(A,v)∈ℂn×nR_{\beta,\delta}(A, v) \in \mathbb{C}^{n \times n} is the n×nn \times n complex matrix Rβ,δ(A,v)=vβ((δ+v)−1I−(A+(δ+v)I)−1),R_{\beta,\delta}(A, v) = v^\beta \bigl((\delta + v)^{-1} I - (A + (\delta + v) I)^{-1}\bigr),(3)✦ where I∈ℂn×nI \in \mathbb{C}^{n \times n} is the identity matrix.

With β,δ,v∈ℝ\beta, \delta, v \in \mathbb{R}, matrices Y,Z∈ℂn×nY, Z \in \mathbb{C}^{n \times n}, and identity matrix I∈ℂn×nI \in \mathbb{C}^{n \times n}, the resolvent tangent kernel Kβ,δ(Y,Z,v)∈ℂn×nK_{\beta,\delta}(Y, Z, v) \in \mathbb{C}^{n \times n} is the n×nn \times n complex matrix Kβ,δ(Y,Z,v)=vβ(Y+(δ+v)I)−1Z(Y+(δ+v)I)−1.K_{\beta,\delta}(Y, Z, v) = v^\beta (Y + (\delta + v) I)^{-1} Z (Y + (\delta + v) I)^{-1}.(4)✦

The primitive divided difference Dσ,δ(x,y)∈ℝD_{\sigma,\delta}(x, y) \in \mathbb{R} at σ,δ,x,y∈ℝ\sigma, \delta, x, y \in \mathbb{R} is the real number Dσ,δ(x,y)=∫01(τx+(1−τ)y+δ)σ−3/2dτ.D_{\sigma,\delta}(x, y) = \int_0^1 (\tau x + (1 - \tau) y + \delta)^{\sigma - 3/2}\,d\tau.(5)✦

The Claim

Theorem 1. Let σ∈(0,1)\sigma \in (0, 1) and δ>0\delta > 0. For every positive semidefinite Hermitian matrix A∈ℂn×nA \in \mathbb{C}^{n \times n}, fσ,δ(A)=cσ−1/2∫0∞Rσ−1/2,δ(A,v)dv.f_{\sigma,\delta}(A) = c_{\sigma - 1/2} \int_0^\infty R_{\sigma - 1/2, \delta}(A, v)\,dv.(6)✦ Given two positive semidefinite Hermitian matrices X,Y∈ℂn×nX, Y \in \mathbb{C}^{n \times n}, the Loewner inequality fσ,δ(X)≤fσ,δ(Y)+cσ−1/2∫0∞Kσ−1/2,δ(Y,X−Y,v)dvf_{\sigma,\delta}(X) \le f_{\sigma,\delta}(Y) + c_{\sigma - 1/2} \int_0^\infty K_{\sigma - 1/2, \delta}(Y, X - Y, v)\,dv(7)✦ holds. Whenever x,y∈ℝx, y \in \mathbb{R} are nonnegative, cσ−1/2∫0∞vσ−1/2(x+δ+v)(y+δ+v)dv=Dσ,δ(x,y).c_{\sigma - 1/2} \int_0^\infty \frac{v^{\sigma - 1/2}}{(x + \delta + v)(y + \delta + v)}\,dv = D_{\sigma,\delta}(x, y).(8)✦

In the paper, the result reads: for y≥0y \ge 0, the function ff of (3.1) has the representation f(y)=cβ∫0∞vβ((δ+v)−1−(y+δ+v)−1)dvf(y) = c_\beta \int_0^\infty v^\beta \bigl((\delta + v)^{-1} - (y + \delta + v)^{-1}\bigr)\,dv with cβ−1=∫0∞tβ(1+t)−2dtc_\beta^{-1} = \int_0^\infty t^\beta (1 + t)^{-2}\,dt, and if X,Y≥0X, Y \ge 0 are Hermitian matrices, then f(X)≤f(Y)+TY(X−Y)f(X) \le f(Y) + T_Y(X - Y), where in an eigenbasis of YY with eigenvalues yiy_i, (TYZ)ij=f[yi,yj]Zij(T_Y Z)_{ij} = f[y_i, y_j] Z_{ij} and f[x,y]=∫01f′(τx+(1−τ)y)dτf[x, y] = \int_0^1 f'(\tau x + (1 - \tau) y)\,d\tau, with fσ,δf_{\sigma,\delta} and Dσ,δD_{\sigma,\delta} standing for ff and f[⋅,⋅]f[\cdot,\cdot], and σ−1/2\sigma - 1/2 standing for β\beta. Unlike the paper, (6) applies directly to matrices rather than scalars, while (7) and (8) replace TYT_Y and the integral of f′f' by Kσ−1/2,δK_{\sigma - 1/2, \delta} and Dσ,δD_{\sigma,\delta}.

What Is Checked

The claims (6), (7), and (8) are stated for parameters σ∈(0,1)\sigma \in (0, 1) and δ>0\delta > 0, positive semidefinite Hermitian matrices A,X,Y∈ℂn×nA, X, Y \in \mathbb{C}^{n \times n} with the Loewner order, and real numbers x,y≥0x, y \ge 0, in terms of the regularized scalar primitive fσ,δf_{\sigma,\delta} via Hermitian functional calculus, the resolvent normalization constant cσ−1/2c_{\sigma - 1/2}, the resolvent matrix integrand Rσ−1/2,δR_{\sigma - 1/2, \delta}, the resolvent tangent kernel Kσ−1/2,δK_{\sigma - 1/2, \delta}, and the primitive divided difference Dσ,δD_{\sigma,\delta}.