BFSS matrix models: threshold states and positive-energy eigenstates

Source and verification

This article explains family 270 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

🏷️ Matrix quantum mechanics and normalizable states

A quantum Hamiltonian can have continuous spectrum beginning at zero and still possess square-integrable eigenvectors, either at zero or at positive energies. Family 270 addresses both possibilities for the relative BFSS matrix model. One manuscript claims a unique zero-energy state for every fixed matrix size $N\ge 2$. Its companion claims infinitely many positive-energy eigenstates for $N=2$. These statements concern different parts of the same spectral picture. (OpenAI, 2026) (OpenAI, 2026a)

The prerequisites are matrix-valued differential operators, quadratic forms, the spectral theorem, harmonic oscillators, and Sobolev compactness. We state the required symmetry-sector facts explicitly and concentrate on their analytic consequences.

The bosonic configuration consists of nine traceless Hermitian $N\times N$ matrices $X_1,\dots ,X_9$. Choosing an orthonormal matrix basis identifies this space with ${\mathbb{R}}^{9(N^2-1)}$. The wave function takes values in a fixed finite-dimensional fermionic space ${\mathcal{F}}_N$. Its internal matrices encode the fermionic degrees of freedom.

The physical Hilbert space is the subspace of

$$L^2({\mathbb{R}}^{9(N^2-1)};{\mathcal{F}}_N)$$

invariant under simultaneous unitary conjugation of the matrices, with the corresponding action on the internal space. This is the gauge constraint. Tracelessness removes the free center-of-mass coordinates; the papers study this relative system, not a product with free translational motion.

There are sixteen symmetric first-order operators $Q_{\alpha }$, called supercharges. Each is a constant-matrix first-order differential operator plus a matrix-valued quadratic expression in the coordinates. The normalization and operator realization are fixed by closing the form

$${\mathfrak{q}}_N(\Psi )=\frac{1}{16}\sum _{\alpha =1}^{16}\Vert Q_{\alpha }\Psi {\Vert }_2^2$$

on smooth compactly supported gauge-invariant wave functions. The resulting nonnegative self-adjoint operator is $H_N$. Thus a zero-energy vector must be annihilated by all supercharges in the appropriate closed domains. A formal solution at infinity is not enough: it must be globally square integrable and belong to this realization.

The two claimed conclusions

The first manuscript asserts $\dim \ker H_N=1$ for every finite $N\ge 2$. Its zero state is invariant under spatial rotations and even in the specified fermion grading. Constants may depend on $N$. The companion asserts an orthonormal sequence ${\Psi }_j\in \mathrm{Dom}(H_2)$ with $H_2{\Psi }_j=E_j{\Psi }_j$, where $E_j>0$ and $E_j\to \infty$. (OpenAI, 2026b) (OpenAI, 2026a)

Neither result asserts a positive gap above zero for the whole relative Hamiltonian. The zero state is at the threshold of a continuum. Positive eigenstates can coexist with that continuum: even the direct sum of a multiplication operator on $L^2(0,\infty )$ and a diagonal operator with positive eigenvalues has this property. The companion’s claim concerns the exact finite-$N=2$ operator; it is not a refutation of the separate large-$N$ Matrix Theory conjecture.

🏷️ Flat valleys and transverse oscillators

For $N=2$, write the coefficients of the nine matrices as three vectors $x^1,x^2,x^3\in {\mathbb{R}}^9$. In the source normalization, expanding the supercharge form gives

$$\begin{array}{rl}{\mathfrak{q}}_2(\Psi )=& \frac{1}{2}\Vert \nabla \Psi {\Vert }_2^2\\ & +\int V(x)\Vert \Psi (x){\Vert }^{2}dx\\ & +\int ⟨\Psi ,B(x)\Psi ⟩dx,\end{array}$$

where

$$\begin{array}{c}V(x)=\sum _{a<b}|x^a\wedge x^b{|}^2,\\ B(x)=B(x{)}^{\ast },\\ \Vert B(x)\Vert \le C|x|.\end{array}$$

Here $|u\wedge v{|}^2=|u{|}^2|v{|}^2-(u\cdot v{)}^2$. The scalar potential vanishes when the three vectors are collinear. These unbounded flat valleys explain why confinement is not apparent from the classical potential alone. The internal term $B$ can have either sign, although the full form is nonnegative by its supercharge definition.

Fix one color vector $u=x^a\ne 0$. Decompose each other vector as $x^b=s_b\hat{u}+z_b$, with $z_b\perp u$. There are sixteen transverse coordinates, and

$$V(x)\ge |u{|}^2\sum _{b\ne a}|z_b{|}^2.$$

Reserving half of the kinetic energy for later compactness, the transverse operator is

$$h_r=-\frac{1}{4}{\Delta }_z+r^2|z{|}^2,r=|u|.$$

The change $\xi =\sqrt{2r}z$ gives $h_r=\frac{r}{2}(-{\Delta }_{\xi }+|\xi {|}^2)$. In sixteen dimensions its levels are $r(n+8)$ for integers $n\ge 0$. Exciting transverse modes therefore creates an energy cost growing linearly along the valley.

🏷️ Symmetry exclusion of low oscillator levels

A rotation sector is the closed sum of all copies of one irreducible spatial rotation type. The companion labels such types by four ordered weights ${\lambda }_1\ge {\lambda }_2\ge {\lambda }_3\ge {\lambda }_4\ge 0$. Only one consequence of the labeling is needed: after fixing the direction $u$, every symmetry type on the transverse slice has first weight at least ${\lambda }_2$.

On the other hand, oscillator level $n$ consists of a Gaussian times polynomials of degree at most $n$. The first weight contributed by such polynomials is at most $n$. The fermionic space is finite-dimensional, so its contribution has a fixed upper bound $M$. Any type occurring both in the slice and in level $n$ must therefore satisfy

$${\lambda }_2\le {\nu }_1\le n+M.$$

If ${\lambda }_2>m+M$, all levels $n\le m$ are absent. The least possible transverse energy becomes $(m+9)|u|$. The second weight is essential; increasing only ${\lambda }_1$ would not give this exclusion. The source proves the required restriction inequality, including half-integral weights. (OpenAI, 2026a)

Two domain points keep this argument valid. First, fixing a slice is done on smooth vectors in a finite-dimensional span of their rotations, so it does not evaluate arbitrary $L^2$ equivalence classes pointwise. Second, the closed form is rotation invariant. Its resolvent commutes with each sector projection, making each sector a reducing subspace of the actual self-adjoint Hamiltonian.

For each choice of color, integrate the transverse oscillator estimate by Fubini. Only derivatives in the transverse variables are used; no derivative of the frame with respect to $u$ is introduced. Averaging the three color choices and using ${\sum }_a|x^a|\ge |x|$ yields

$$\begin{array}{rl}{\mathfrak{q}}_2(\Psi )& \ge \frac{1}{4}\Vert \nabla \Psi {\Vert }_2^2\\ & +c_m\int |x|\Vert \Psi (x){\Vert }^{2}dx,\\ c_m& =\frac{m+9}{3}-C.\end{array}$$

Choose $m$ so that $c_m>0$. This positive linear term dominates the potentially negative fermionic term. The inequality extends from the smooth sector core to its full form domain by closedness of the gradient and weighted multiplication operators.

🏷️ Compactness and positive eigenvalues

The displayed form bound gives a uniform tail estimate:

$${\int }_{|x|>R}\Vert \Psi (x){\Vert }^{2}dx\le \frac{{\mathfrak{q}}_2(\Psi )}{c_{m}R}.$$

A bounded set in the form norm is also bounded in $H^1$. Rellich compactness on each ball, applied componentwise in the finite internal space, gives convergent subsequences locally in $L^2$. The uniform tail estimate upgrades these to convergence on the full space. Thus the sector Hamiltonian has compact resolvent. Its kernel is zero because a zero form value would force the positive weighted integral to vanish.

One still needs an infinite-dimensional physical sector satisfying the large-weight condition. The source constructs gauge-invariant polynomial wave functions whose second weight grows without bound. Multiplying one fixed such polynomial and a suitable internal vector by smooth radial cutoffs with disjoint supports produces infinitely many linearly independent vectors in the same sector. Gauge invariance and the rotation type are preserved by radial multiplication.

Choose one of these sectors beyond the threshold. A nonnegative operator with compact resolvent on an infinite-dimensional space has eigenvalues tending to infinity; the trivial kernel makes them positive. Since the sector reduces $H_2$, its eigenvectors lie in the full operator domain and remain exact eigenvectors there. This proves the companion’s conclusion from the symmetry and form estimates, without approximating eigenstates by resonances or numerical levels. (OpenAI, 2026a)

🏷️ A massive index and the missing compactness step

The zero-energy problem uses a different mechanism. Introduce an auxiliary charge $D(h,m)$, where $h>0$ multiplies the commutator interaction and $m\ge 0$ is a confining mass parameter. At zero mass and unit coupling, its gauge-invariant $L^2$ kernel agrees with $\ker H_N$. The manuscript fixes a particular charge and a commuting rotation-invariant subspace for the massive comparison.

The charge reverses a decomposition into even and odd fermion states. Its index is the difference between the dimensions of the even and odd kernels. At fixed positive mass, confinement makes the problem Fredholm. The source computes

$$\mathrm{ind}D(h,1)=p(N),$$

where $p(N)$ is the number of unordered integer partitions of $N$. This is initially a signed count on the specified invariant subspace, not a statement that the undeformed Hamiltonian has $p(N)$ or one ground state.

To approach zero mass, send $h\to \infty$ at mass one. Rescaling coordinates by $y=h^{1/3}x$ converts the charge into $h^{1/3}D(1,h^{-2/3})$. The small-mass limit is therefore encoded in a strong-coupling limit. A family of normalized massive states can lose all its norm on bounded sets during this limit. Index constancy alone does not prevent that loss.

The source also proves, before determining its dimension, that the whole undeformed kernel is rotation invariant and even. A rotation generator is expressed through supercharge anticommutators. Radial cutoffs move these identities onto the closed domains; their error is bounded by escaping-annulus $L^2$ tails. No global moment bound or prior finite-dimensionality of the kernel is assumed. This parity information later converts a signed count into a dimension count. (OpenAI, 2026b)

🏷️ Cluster profiles and corrected test states

A proper partition $C=(n_1,\dots ,n_k)$ with $k\ge 2$ describes separated matrix blocks. Their centers $c_{\alpha }\in {\mathbb{R}}^9$ satisfy ${\sum }_{\alpha }{n}_{\alpha }{c}_{\alpha }=0$. A profile is a wave function of these relative centers, with values in the tensor product of the internal block kernels and center fermions. Equal-sized blocks have the prescribed permutation symmetry. The internal kernels need not be assumed finite-dimensional when defining these spaces.

Near separated centers, the charge has the structure

$$\begin{array}{rl}D(h,m)=& \sqrt{h}E\\ & +D^{\mathrm{slow}}(h,m)+R.\end{array}$$

The transverse operator $E$ has an even one-dimensional ground line and a gap on its orthogonal complement. Estimates in the charge graph norm force concentration on this line. But state concentration alone does not identify the limiting charge equation: the charge applied to the leading product can leave an order-one perpendicular component.

The source corrects the product by applying the reduced inverse of $E$ to that component, with a factor $h^{-1/2}$. For a fixed test profile $f$, the corrected lift $I_{h}f$ and leading identification ${\mathcal{J}}_{h}f$ satisfy

$$\begin{array}{rl}\Vert I_{h}f-{\mathcal{J}}_{h}f\Vert & \to 0,\\ \Vert D(h,m)I_{h}f-{\mathcal{J}}_h{D}^C(m)f\Vert & \to 0.\end{array}$$

Here $D^C(m)$ is the center charge. The second convergence allows the charge equation to pass to a profile by duality. The construction works simultaneously on fixed finite families and preserves their limiting inner products.

Internal block wave functions shrink on scale $h^{-1/3}$, while relative centers remain on scale one. Fixed physical cutoffs have derivatives supported away from an internal origin, so their commutator errors vanish using only the $L^2$ tails after rescaling. No unproved moment bound on an internal zero state is inserted.

🏷️ Excluding intermediate scales

Profiles at the internal scale and at separated centers could miss norm between them. The crucial additional estimate is an exterior inequality

$$\Vert rD(h,0)F\Vert \ge c_N\Vert F\Vert ,$$

where $r=|x|$, for compactly supported gauge-invariant graph-domain vectors whose support satisfies $h^{1/3}r>R_N$. It does not impose the auxiliary rotation restriction used for the massive index.

Its quantitative consequence is particularly transparent. Suppose $\Vert F_h\Vert \le B$ and $\Vert D(h,0)F_h\Vert \le K$. A cutoff supported on a logarithmically enlarged annulus, with $|r{\chi }^{\prime }(r)|\le C/L$, gives

$$\begin{array}{rl}& \Vert 1_{\{M{h}^{-1/3}<r<\epsilon \}}{F}_h\Vert \\ & \le c_N^{-1}(e^L\epsilon K+CB/L),\end{array}$$

provided $M>e^L{R}_N$. Indeed, apply the exterior estimate to $\chi F_h$ and use $D(\chi F_h)=\chi D{F}_h+[D,\chi ]F_h$. The principal symbol gives $\Vert [D,\chi ](x)\Vert =|\nabla \chi (x)|/\sqrt{2}$.

For a chosen tolerance, first take $L$ large, then $M$ large enough, and finally $\epsilon$ small. The bound excludes norm that escapes in rescaled internal coordinates while remaining at a shrinking physical radius. This order of choices is essential.

The exterior estimate is proved simultaneously with the kernel theorem by induction on $N$. On an annulus away from the origin, only proper clusters occur. Smaller-block results give a decomposition that accounts for all norm there. Each center space has dimension $d_C=9(k-1)\ge 9$. Hardy’s inequality and the free center-charge normalization give

$$\Vert D^C(0)(|c|f)\Vert \ge \frac{d_C-2}{2\sqrt{2}}\Vert f\Vert .$$

Passing this bound through the profile limit and paying for the commutator with $r$ gives a strictly positive annular estimate. Dilation and a squared partition of unity in $\log r$ extend it to the exterior region; the localization error is $O(L^{-2})$ for a sufficiently wide partition. This proves the current-size exterior estimate before assuming any current-size ground state exists. (OpenAI, 2026b)

🏷️ Preserving dimensions through the massless limit

After the exterior estimate, a whole-block profile $F_0\in \ker H_N$ captures the concentration at the origin. A separate massive confinement estimate prevents escape to infinity. The full decomposition then preserves both norm and inner products:

$$\underset{j}{\lim }\Vert {\Psi }_j{\Vert }^2=\Vert F_0{\Vert }^2+\sum _{C\mathrm{proper}}\Vert F_C{\Vert }^2.$$

For states whose charge norm tends to zero, each proper profile has the unique even massive center-oscillator factor. By induction, each internal block contributes one even kernel line. Thus there is one line for each of the $p(N)-1$ proper partitions. The whole-block profile is also even, without yet knowing its dimension.

For this massive comparison, we work on the chosen gauge- and auxiliary-rotation-invariant space. An odd normalized sequence with charge norm tending to zero would have every profile vanish, contradicting preservation of its norm. This proves a uniform massive gap on its odd subspace at large $h$. On each positive eigenspace of the squared charge, the charge divided by the square root of that eigenvalue is an isometric bijection between the even and odd parts. The same gap therefore holds on the even complement of the massive kernel. The index now implies that this restricted massive kernel has exactly $p(N)$ dimensions.

For existence, extract profiles from an orthonormal basis of that massive kernel. The limit preserves all $p(N)$ inner products, whereas proper profiles supply only $p(N)-1$ dimensions. Therefore $\ker H_N$ cannot be zero.

For uniqueness, take any finite orthonormal list of $M$ undeformed zero states. The source constructs recovery vectors for them and for all proper profile lines. They are asymptotically orthonormal and their charge norms tend to zero. Projecting onto the massive kernel changes them by a norm tending to zero because of its uniform gap. Their Gram matrix remains positive definite for large $h$, so

$$\begin{array}{c}p(N)-1+M\le p(N),\\ M\le 1.\end{array}$$

This argument applies to every finite list and does not presuppose a finite-dimensional undeformed kernel. At $N=2$ all proper internal blocks are scalar, initiating the induction. The graph-domain identification finally identifies the charge kernel with that of the full-space form Hamiltonian. (OpenAI, 2026b)

🏷️ Verification boundary and connections

The positive-eigenvalue companion’s proof text was read throughout, including its form normalization, symmetry-sector reduction, oscillator bound, compactness argument and construction of nonzero sectors. For the longer threshold manuscript, the review covers its model, main statements and selected domain, parity, massive-index, profile, exterior and dimension-comparison passages. The full massive localization, fast-mode estimates, recovery constructions and all nested-cluster compatibility estimates have not been independently verified. No numerical spectrum or finite certificate is substituted for those analytic obligations.

• Spin-one Haldane gaps concerns uniform spectral separation; the BFSS threshold problem instead allows a continuum beginning at zero.
• Laughlin gaps and disorder stability also distinguishes finite or auxiliary information from the uniform analytic estimates needed to pass to a larger system or a limiting regime.

🏷️ Key steps and proof outlines

The reduced inverse cancels a specific order-one error

The threshold proof needs convergence in the charge graph norm, which measures both a vector and its charge output. In a fixed separated-cluster chart, let $\chi (b)$ be the normalized fast oscillator ground state, $P(b)$ its projection, and $f_h$ a rescaled internal-and-center profile. The leading lift has the expansion

$$D(h,m)(\chi f_h)=\chi (D^C(m)f{)}_h+A(b)f_h+o_{L^2}(1).$$

The unwanted map $A$ contains the derivative of $\chi$ in slow coordinates, together with the order-one geometric and density-conjugation terms. Thus $\Vert \chi f_h\Vert$ being correct does not imply that its charge output is correct.

The source’s diagonal-compression calculation gives $P(c)A(c)=0$ on the limiting cluster configuration. Concentration of the internal variables and smoothness then make $P(b)A(b)f_h$ tend to zero. Define

$$V=(1-P)A,W=(E{|}_{\mathrm{Ran}(1-P)}{)}^{-1}V,$$

and replace the lift by

$$I_{h}f=\chi f_h-h^{-1/2}W{f}_h.$$

Since the leading fast charge is $\sqrt{h}E$, it sends the correction to $-V{f}_h$, exactly cancelling the remaining order-one error.

Two additional checks are necessary. First, the slow and remainder outputs of the correction must tend to zero; the source proves bounds involving $\Vert p_b{f}_h/\sqrt{h}\Vert$ and $\Vert \sqrt{h}[b_i,b_j]f_h\Vert$, both controlled by its internal rescaling estimates. Second, parity must survive: $V$ and the reduced inverse of $E$ are odd, so their product $W$ is even. This yields a recovery map that preserves the grading used by the index argument, as well as the limiting charge equation. No assertion that arbitrary internal Clifford matrices preserve the internal kernel is needed (OpenAI, 2026b).

A concrete polynomial creates the confining sectors

For the positive-energy companion, the oscillator estimate is useful only if physical sectors with large second weight actually exist. The construction can be made explicit. Set

$$\begin{array}{c}{w}_1=e_1+i{e}_2,w_2=e_3+i{e}_4,\\ {\zeta }_a=w_1\cdot x^a,{\eta }_a=w_2\cdot x^a,\end{array}$$

using complex bilinear dot products. Define

$$\begin{array}{rl}P(x)& =\sum _{a<b}({\zeta }_a{\eta }_b-{\zeta }_b{\eta }_a{)}^2\\ & =(\zeta \cdot \zeta )(\eta \cdot \eta )-(\zeta \cdot \eta {)}^2.\end{array}$$

Gauge transformations rotate the three color coordinates orthogonally, so the second expression proves gauge invariance directly. At $x=t(e_1,e_3,0)$, $P(x)=t^4$; hence it does not vanish identically on any radial shell.

The source’s rotation calculation says that multiplying by $P^k$ raises both of the first two weights by $2k$. Choose a nonzero gauge-invariant highest-weight internal vector $v$ of one fixed rotation type $\mu$, and a smooth radial cutoff $\chi$ supported away from zero. Then

$${\Psi }_{\chi }(x)=\chi (|x|)P(x{)}^{k}v$$

is a smooth physical vector of type

$$(2k+{\mu }_1,2k+{\mu }_2,{\mu }_3,{\mu }_4).$$

The existence of a gauge-invariant internal vector is established by constructing a fermionic vacuum line; it is preserved by the gauge action and that action on the line is trivial.

Fixing $k$ sufficiently large puts the second weight above the transverse-exclusion threshold. Varying $\chi$ over disjoint radial shells then produces infinitely many nonzero orthogonal vectors in this same sector. This supplies the infinite-dimensional reducing space required by the compact-resolvent argument. The polynomial construction is therefore the bridge from a conditional confinement estimate for selected symmetry types to actual positive eigenstates of the physical Hamiltonian (OpenAI, 2026a).

🏷️ Manuscripts and verification scope

• The unique threshold bound state of the SU(N) BFSS model — pinned manuscript and companion source.
• Positive eigenvalues of the relative SU(2) BFSS Hamiltonian — pinned manuscript and companion source.

No family scope page was present at lean/docs/270.md in the pinned snapshot. This does not establish that all related formalization code is absent. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. Positive eigenvalues of the relative SU(2) BFSS Hamiltonian.

🐻  OpenAI 2026b. The unique threshold bound state of the SU(N) BFSS model.