gpt 6 astra ultra attempt on Phys-4

Extended States in the Anderson Model

Definitions and mathematical statement

For independent identically distributed real random variables $V_x$, the standard lattice Anderson operator on $\ell^2({\mathbb{Z}}^d)$ has the form \[ (H_\lambda\psi)(x)=-\sum_{|y-x|_1=1}\psi(y)+\lambda V_x\psi(x). \] For a vector $\psi$, its spectral measure is $\mu_\psi(B)=\langle\psi,1_B(H_\lambda)\psi\rangle$; an absolutely continuous spectral component has density with respect to Lebesgue measure. The central target is a nonzero absolutely continuous spectral subspace in a weak-disorder regime in $d\ge3$. The catalogue does not specify a probability law or exact energy interval. Those choices must be fixed in a full model-specific theorem. Its additional language about transport and a mobility edge is not, by definition alone, equivalent to existence of absolutely continuous spectrum.

Formulation and concept references: [S1].

Short English statement

Can weak random disorder in a lattice of dimension at least three be proved to permit an extended-state spectral regime?

Research attempt

Status and a fixed model.

This is an unresolved attempt, dated 27 September 2026. Fix an integer $d\geq3$ and let the $V_x$ be independent uniform random variables on $[-1,1]$. This is a nondegenerate law with variance $1/3$. Use exactly the nearest-neighbor hopping and sign in the statement, and write $H_{\lambda,\omega}=H_0+\lambda V_\omega$. The target considered here is a nonzero absolutely continuous spectral subspace almost surely for every fixed $0<\lambda<\lambda_0(d)$. A result only at $\lambda=0$ would not establish this target.

The attempt proves a free spectral calculation, a uniform positive-mass observation, and a sufficient resolvent criterion. The latter criterion is not proved for the random model. It isolates a quantitative missing estimate and supplies countertests to two tempting replacements: finite-time perturbation theory and absolute continuity only after averaging disorder.

Kirsch [S1] explains the spectral formulation and the distinction between the lattice problem and trees. The recent manuscript [S3], inspected in its online full text [E1], treats independent Gaussian potentials and proves diffusive estimates on specified weak-coupling scales. Its Theorems 1.1--1.2 do not assert the uniform boundary estimate below for a fixed positive $\lambda$. No transfer from a Bethe lattice, Gaussian disorder, or a joint scaling limit to the present uniform-law lattice model is assumed.

1. Free Fourier analysis and a regular energy interval.

Use normalized Haar measure on $\mathbb T^d=[-\pi,\pi]^d$. The Fourier transform maps $H_0$ to multiplication by \[ h(k)=-2\sum_{j=1}^d\cos k_j. \tag{31.1} \] Thus the free spectral measure of $\psi$ is the pushforward of $|\widehat\psi(k)|^2\,dk/(2\pi)^d$ under $h$. Its critical points have each $k_j$ equal to $0$ or $\pi$, so their energies belong to \[ \{-2d+4j:0\leq j\leq d\}. \] Fix the concrete open interval \[ E_*=-2d+2,\qquad I=(E_*-1/2,E_*+1/2). \tag{31.2} \] Its closure lies inside the free band and contains no critical energy. For $\psi=\delta_0$, coarea gives on $I$ the density \[ \rho_0(E)=\frac{1}{(2\pi)^d} \int_{h^{-1}(E)}\frac{d\sigma(k)}{|\nabla h(k)|}. \tag{31.3} \] It is finite, continuous, and positive there. Finiteness and continuity follow by covering the compact inverse image of $\overline I$ with finitely many implicit-function charts, since $|\nabla h|$ has a positive minimum on that set. Positivity follows from a nonempty regular level and its local $(d-1)$-dimensional surface patch. Every energy in $I$ is attained, for example by varying the first coordinate while holding the others at zero.

The same coarea argument with the integrable weight $|\widehat\psi|^2$ shows absolute continuity away from the finitely many critical values for any $\psi\in\ell^2$. A level set at any one of those values has Haar measure zero: fix the other coordinates and solve for $\cos k_1$, which gives at most two values of $k_1$. Fubini therefore shows that there is no residual spectral mass on the critical values. This proves the free operator has purely absolutely continuous spectrum. It gives a reference density for the attack, not stability under nonzero random perturbations.

2. Small norm perturbations preserve mass, but the basic estimates fail.

On the probability-one event that all $|V_x|\leq1$, $\|H_{\lambda,\omega}-H_0\|\leq\lambda$. Duhamel's formula and the resolvent identity consequently give, uniformly in $\omega$, \[ \|e^{-itH_{\lambda,\omega}}-e^{-itH_0}\|\leq\lambda|t|, \qquad \|(H_{\lambda,\omega}-z)^{-1}-(H_0-z)^{-1}\| \leq\lambda(\operatorname{Im}z)^{-2}. \tag{31.4} \] In the first estimate the integrand is the product of two unitaries and $\lambda V_\omega$; in the second each self-adjoint resolvent has norm at most $(\operatorname{Im}z)^{-1}$. Neither estimate is uniform in the limit needed for spectral type: the errors grow as $|t|\to\infty$ or as $\operatorname{Im}z\downarrow0$ at fixed positive $\lambda$.

One useful conclusion nevertheless survives. Choose a continuous function $0\leq f\leq1$ supported in $I$ and positive on a smaller interval. Then \[ c_0=\langle\delta_0,f(H_0)\delta_0\rangle =\int_I f(E)\rho_0(E)\,dE>0. \] For sufficiently small $\lambda$, uniformly in $\omega$, $\|f(H_{\lambda,\omega})-f(H_0)\|<c_0/2$. To justify this without a spectral-type assumption, put all spectra for $\lambda\leq1$ in $[-M,M]$, where $M=2d+1$, and approximate $f$ uniformly there by a polynomial $P$. For $A,B$ with norms at most $M$, \[ \|A^j-B^j\|\leq jM^{j-1}\|A-B\| \] by telescoping. First make $2\|P-f\|_\infty$ small, then make $\lambda\sum_{j\geq1}|P_j|jM^{j-1}$ small. The asserted continuity follows. Consequently, with $\mu_{0,\lambda,\omega}$ denoting the spectral measure of $\delta_0$, \[ \mu_{0,\lambda,\omega}(I) \geq\langle\delta_0,f(H_{\lambda,\omega})\delta_0\rangle >c_0/2. \tag{31.5} \] This is positive spectral mass for all sufficiently small perturbations; it does not say whether that mass has a density.

3. A sufficient estimate with an explicit almost-sure conclusion.

For fixed $\lambda>0$, define the nonnegative Poisson transform \[ F_{\eta,\omega}(E)=\frac1\pi \operatorname{Im}\langle\delta_0, (H_{\lambda,\omega}-E-i\eta)^{-1}\delta_0\rangle =\int\frac{\eta/\pi}{(t-E)^2+\eta^2} \,d\mu_{0,\lambda,\omega}(t). \tag{31.6} \] Here the sign of the imaginary part agrees with the chosen resolvent. The following condition, for some $1<p<\infty$, would suffice: \[ \sup_{0<\eta\leq1}\mathbb E \int_I F_{\eta,\omega}(E)^p\,dE \leq C_{\lambda,d,p,I}<\infty. \tag{31.7} \] The constant need not be uniform as $\lambda\downarrow0$. What matters is uniformity as $\eta\downarrow0$ for each fixed positive disorder strength.

Here is a proof of sufficiency that does not exchange an uncountable collection of probability-one events. Take the deterministic sequence $\eta_n=2^{-n}$. Fatou's lemma applied to the nonnegative random variables $\|F_{\eta_n,\omega}\|_{L^p(I)}^p$ gives \[ \mathbb E\liminf_n\|F_{\eta_n,\omega}\|_{L^p(I)}^p\leq C_{\lambda,d,p,I}. \] Hence for almost every fixed $\omega$ there is a subsequence with bounded $L^p(I)$ norm. Reflexivity gives a further subsequence converging weakly to some $g_\omega\in L^p(I)$. For each $\varphi\in C_c(I)$, extended by zero to $\mathbb R$, the approximate-identity property of the Poisson kernel gives \[ \lim_{\eta\downarrow0}\int_I\varphi(E)F_{\eta,\omega}(E)\,dE =\int\varphi(t)\,d\mu_{0,\lambda,\omega}(t). \] Weak $L^p$ convergence identifies the same limit with $\int_I\varphi g_\omega$. Uniqueness of measures tested against $C_c(I)$ proves \[ \mu_{0,\lambda,\omega}|_I=g_\omega(E)\,dE \quad\text{almost surely}. \tag{31.8} \] For $\lambda$ small enough for $31.5$, this density has positive integral. The vector $1_I(H_{\lambda,\omega})\delta_0$ is therefore nonzero and belongs to the absolutely continuous subspace. No claim that $\delta_0$ is cyclic is needed. Thus $31.7$, if established, really would imply the catalogue's central spectral target for the fixed model and interval.

4. Why averaging a measure is too weak.

Consider a separate, exactly solvable countermodel with the same iid uniform variables but no hopping: $D_\omega\delta_x=V_x(\omega)\delta_x$. Every realization has a complete orthonormal basis of eigenvectors, so the operator is pure point. Yet \[ \mu^{D}_{\delta_0,\omega}=\delta_{V_0(\omega)},\qquad \mathbb E\mu^{D}_{\delta_0,\omega}(dE) =\tfrac12 1_{[-1,1]}(E)\,dE. \tag{31.9} \] The averaged measure is absolutely continuous although every sample measure is atomic. This operator is not the weak-disorder hopping model; it refutes only the proposed inference from an averaged density to samplewise absolute continuity.

The distinction is detected by the power in $31.7$. Set $J=(-1/2,1/2)$ for this countermodel. If $|V_0|\leq1/4$ and $\eta<1/4$, then $[V_0-\eta,V_0+\eta]\subset J$, and its Poisson kernel obeys \[ \int_J\left(\frac{\eta/\pi}{(E-V_0)^2+\eta^2}\right)^p dE \geq 2^{1-p}\pi^{-p}\eta^{1-p}. \tag{31.10} \] The conditioning event has probability $1/4$, so the expected $L^p$ norm diverges for every $p>1$. In contrast the first moment of the Poisson transform is bounded by $1/2$, being the convolution of a probability kernel with the bounded averaged density. At $p=1$, integration over the whole line is always one for any probability spectral measure. A first moment bound therefore cannot replace the stronger condition.

Verification, scaling limits, and the remaining gap.

The companion checks verify the critical-energy list and that $31.2$ avoids it for dimensions three through twelve. They also check the telescoping identity on explicit noncommuting rational matrices, and the exact scaling in $31.10$. These are checks of the reductions, not simulations establishing an infinite-volume spectral type. A finite lattice matrix always has atomic spectral measures, even when its empirical measures approach a continuous limit. For example equally weighted atoms on a regular grid in $[0,1]$ converge weakly to Lebesgue measure by Riemann sums. Finite-volume eigenvalue plots alone cannot settle the required almost-sure assertion.

The precise gap is $31.7$ at a fixed $0<\lambda<\lambda_0$ all the way to $\eta=0$. Bounds down to a scale $\eta=\lambda^{2+\kappa}$ still stop at a positive number for that fixed $\lambda$; taking $\lambda\to0$ along the scale does not fill the missing limit. The elementary resolvent estimate in $31.4$ becomes worse there, and $31.9$ shows why replacing the missing moment control by a smooth averaged measure is invalid. No such uniform $L^p$ estimate is proved here, nor is transport or a mobility edge inferred from the partial results. The Anderson extended-state problem remains UNRESOLVED.

Sources

• S1 W. Kirsch, "An Invitation to Random Schrodinger Operators", in Random Schrodinger Operators, Panoramas et Syntheses 25, Societe Mathematique de France, 2008; arXiv:0709.3707. https://arxiv.org/abs/0709.3707. Formulation source.
• S2 A Note on Extended States on the Bethe Lattice. https://arxiv.org/html/2609.06987v1. Further reference; not a proof endorsement.
• S3 Self-consistent equations and quantum diffusion for the Anderson model. https://arxiv.org/abs/2506.06468. Further reference; not a proof endorsement.

• E1 Adam Black, Reuben Drogin, and Felipe Hernández, Self-consistent equations and quantum diffusion for the Anderson model, arXiv:2506.06468v2 (6 November 2025), Introduction and Theorems 1.1--1.2. https://arxiv.org/abs/2506.06468v2. https://arxiv.org/html/2506.06468v2. Version-specific identification of [S3]. The inspected Gaussian-law diffusion estimates are not used as a proof of the fixed-disorder, zero-height resolvent condition for uniform disorder proposed here. The manuscript's complete proof was not independently audited. Consulted 27 September 2026.