Engineer James Orlando published exact states for two vertices of a quantum polytope
Engineer James Orlando published exact states for two vertices of a quantum polytope
On 21 July, James Orlando released a preprint, code, and data addressing a problem involving the generalized Pauli constraints. His program provides explicit quantum states for two vertices that a 2008 paper had verified numerically. The earlier calculation came very close to the answer, but no exact formulas for the states were available at the time.
Electrons are fermions, and the Pauli exclusion principle limits the number of particles that can occupy each quantum state. The generalized Pauli constraints impose additional conditions on the entire distribution of electrons among those states. Mathematicians represent the allowed distributions as points in a multidimensional polytope. Its vertices define the limiting combinations permitted by quantum mechanics.
In a 2008 paper by Murat Altunbulak and Alexander Klyachko, two such vertices for a system of four fermions in nine orbitals remained numerical results. Orlando published a preprint containing explicit formulas and a repository in which they can be checked. One vertex can be represented by a combination of seven electronic configurations with integer weights. The other requires quantum phases. Its amplitudes reinforce or cancel one another, and this interference produces the required distribution.
The method used to obtain the result is also informative. Claude suggested possible approaches, while the program enumerated vertices using rational arithmetic, constructed candidate states, and verified them with solver certificates. The final state was then checked again without rounding. In his account of the work, Orlando writes that Claude was “often confidently wrong”: the model misread a digit in a PDF. Testing the acceptance criterion uncovered a more important error. The search had been restricted to real amplitudes without making that limitation explicit. Once complex amplitudes were included, the state corresponding to the second vertex was found. The verification code and audit detected both errors.
The model reduces the cost of generating hypotheses, but confidence comes from a procedure that can reject them. One week earlier, a Lean project verified a new lower bound proposed by GPT-5.6 Sol: the program’s kernel independently derived every logical step. Here, rational arithmetic and solver certificates serve the same role. The open data allow another person to run the same sequence of checks and verify the specific formula, instead of judging whether the model’s answer merely sounds convincing. Feedback from a mathematician in the relevant field became part of the same process. Orlando revised the wording in response to the comments and checked the reasoning again.