Formal in Silico

Domain · goal G1 · status snapshot 8 October 2026

mini-DFT

A one-dimensional periodic reduced Hartree–Fock model with plane waves — the project’s first domain, where the trust ledger, the certificate checkers and trace validation were built. Goal G1 is feature parity with DFTK.jl 0.8.0 on the same discrete problem, with every output covered by a Lean theorem.

1,169claims in the trust ledger
76%machine-proved: 589 at T3, 299 at T4
10 / 46DFTK parity items done, LDA and collinear spin among them; 2 more agree but lack full certificates
≤ 5×10⁻¹⁵energy difference from DFTK on the same discrete problem

01

Claims in the trust ledger

Of the ledger’s 1,598 claims, 1,169 concern mini-DFT; they are grouped below by topic. The tiers are defined on the programme page.

T4 end to end · 299 T3 machine proof · 589 T2 model-checked · 6 T1 paper proof · 16 T0 tested · 258 not yet reached · 1
DFT foundations
discrete model · algorithm · parallelism · floating point
18
Ground-state energy certificate
8
Finite-temperature free energy
14
Gaussian and cold smearing
5
SCF-solution certificates
Methfessel–Paxton of any order · Marzari–Vanderbilt
14
k-point sampling and supercells
7
Discretisation error
truncation K → ∞ · true Gaussian potential · continuum limit: E∞ equals the continuum ground energy
30
Problem binding: the external potential
named potentials in Lean and their Fourier series · certificate problem equals the named one at Γ and with k-points · Gaussian deviation bound · mandatory for every certificate · perturbation estimate: Gaussian SCF solutions, spectra, bands and entropy hold for the named potential · user coefficient files and Gaussian sites · Gaussian LSDA
21
Density and eigenvalues
10
Band structure and density of states
zero temperature · finite-temperature SCF solutions, rHF and LDA · k-points: primitive-cell bands from the supercell solution · per-spin LSDA bands and DOS
11
LDA exchange-correlation (Slater, PW92)
exchange-correlation on a real-space grid · nonlinear SCF-solution certificates · energy · PW92 interval bounds · Fermi–Dirac and Gaussian occupations · k-point supercell · zero temperature under a gap
91
Collinear spin (LSDA)
bounds for two spin channels · block-diagonal SCF solution with magnetisation · k-points · zero temperature, also with k-points · band error split · energy and free-energy checker · spin-interpolated PW92 with local box bounds: SCF solution, zero temperature, energy, bands and DOS, Gaussian potential · checkers
32
SCF-solution uniqueness, input checks and echo
every SCF solution is a zero of the auxiliary map · header and input format checked before checking · k-points, weights and DOS width echoed · zero-temperature gap printing rule (2026-10-07 audit) · one exit-code rule, spin-down k-point gate, LSDA band echo (2026-10-07 second review)
14
Forces (Hellmann–Feynman)
first-order expansion with a second-order remainder · derivative of the ground (free) energy · zero total force · force-interval checker · certified site forces FORCE_OK, FORCE_SUM at zero and finite temperature
8
Exact exchange (Hartree–Fock)
exchange and Fock matrices for any real kernel · Frobenius bounds · energy expansion · self-interaction cancellation · q = 0 kernel shift · SCF solution exists and is unique in a ball · checker · driver, DFTK fixed-point check, known answers
11
3D plane waves: Cohen–Bergstresser model and band intervals
d-dimensional plane-wave model · eigenvalue intervals by inertia · band edges and gaps · Schur-complement truncation · continuous Bloch operator · certify-pw · gap over the whole Brillouin zone · DFTK, empty lattice, positive control
864
3D parallel protocols (TLA+)
k-point pools × band groups · bit-reproducible reductions · refinement of the 1D specifications · slab and pencil FFT transposes · any number of processes (paper proof) · MPI driver (S3): trace validation, bitwise identity in the tested layouts, planted bugs, certificates
11

The T4 claims are the soundness theorems of the certificate checkers (which run in exact rational arithmetic, so no floating point is involved) and the Frama-C/WP proofs about the trusted C kernel: no run-time errors (DFT-F03) and, since 6 October, the floating-point correctness of its interval arithmetic on the C code under an IEEE 754 model — the charge interval, the residual bound and the inertia count (DFT-F01, F02; the step from the exact elimination to the eigenvalue count is the Lean theorem exactElim_inertia). Three claims rest on paper proofs in domains/dft/notes/ (T1; no published result covers the 1D periodic Hartree kernel directly): continuum well-posedness of 1D rHF, the a priori plane-wave estimate and the local convergence of linear mixing for α < 2/λmax of the dielectric operator, all at their target; since 8 October a fourth extends the 3D parallel protocols, model-checked by TLC on finite instances (DFT-R08–R11, T2), to any number of processes, k-points, bands and grid points (DFT-R12, T1, at its target). One 3D claim is registered but not yet reached: PW_CBAND, band intervals for the continuous Bloch operator (DFT-W15, target T4). The equality of the truncation limit with the continuum energy (DFT-C02) is machine-proved since 8 October (T3): the continuum model is written in Fourier space, and its equivalence with the operator model on H¹ is assumption A-DFT-08. The Lean–Rocq interface claim DFT-I01 was withdrawn on 6 October, when decision D1 chose Frama-C rather than Rocq.

02

Proof dependencies

Which claims are built on which, read from the Lean proof terms themselves (not from the ledger’s hand-written dependency column). The graph starts with this domain only; click the legend to add the others, or switch to “All Lean theorems” for the theorem level. All domains together are on the programme page.

  • a claim or theorem, coloured by domain
  • uses — points at what the proof relies on
  • bigger — used by more results
  • what the selection uses
  • what uses the selection

Drag to pan, scroll or pinch to zoom, drag a node to move it; click a node to see what it uses and what uses it, and “Only its chain” to keep just those, directly or not. An arrow from claim A to claim B means that a proof behind A uses a theorem behind B (through lemmas that belong to no claim); two claims can use each other. Claims without a Lean theorem (tests, T0–T2) are not drawn. Generated by scripts/depgraph.py; a reading aid, not part of the trust chain.

03

Goal G1 — mini-DFT reaches DFTK parity

Reference: DFTK.jl 0.8.0 on the same discrete problem. Order: complete the 1D feature set (M0–M3), generalise the Lean model to d dimensions for 2D/3D (M4–M5), then infrastructure (M6).

Done — implemented, agrees, certified Agrees — certificates incomplete Implemented — not yet compared Not started
M0
Existing core
5 / 9
M1
1D solvers, occupation, k-points
2 / 7
M2
1D physics terms
2 / 7
M3
1D post-processing and response
0 / 9
M4
d dimensions
1 / 7
M5
Full 3D
0 / 4
M6
Infrastructure and interfaces
0 / 3

M0 · existing core

FeatureStatusWhat is certified
Kinetic, external (Fourier) and Hartree termsDoneGround-state energy, density, eigenvalues and each energy component; since 4 October the checker also verifies the external potential against its Lean definition (DFT-I02, I03, V09)
Fermi–Dirac smearing and entropyDoneFree energy, entropy and energy; existence of the finite-temperature ground state and uniqueness of its density are proved
Γ-centred grids with an odd number of k-pointsDonePer-cell energy through the equivalent supercell, plus its truncation limit
Even grids, Monkhorst–PackDoneSupercell with half-integer frequencies; grids that contain the zone boundary are excluded
Dense diagonalisationDoneEach eigenpair enclosed by a residual bound; Sylvester inertia confirms they are the lowest ones
Linear, Kerker and Anderson mixingImplementedDoes not affect the result, which is checked by certificate
Non-self-consistent bands, Gaussian-broadened DOSAgreesCertified for zero-temperature Γ-point runs and, since 4 October, for finite-temperature Γ-point SCF-solution certificates — rHF with MV, MP, Fermi–Dirac or Gaussian smearing, and LDA (DFT-S04, S05). Since 6 October also for Monkhorst–Pack k-point runs — rHF with any of these smearings, LDA, and rHF at zero temperature — through the primitive-cell Bloch Hamiltonian of the block-averaged supercell solution (DFT-S06–S08). DFTK’s bands and DOS fall inside the Lean intervals. Still missing: zero-temperature LDA with k-points and the named-Gaussian (GAUSS_BAND) version with k-points
SupercellsImplementedSupercell equivalence is proved
Parallelism over k-pointsImplementedThreads (1D) and MPI (3D point-group LDA, showcase S3); the protocols are model-checked (T2); results are bitwise identical across thread counts and in the tested MPI layouts, and the logs pass trace validation (T0)

M1 · 1D solvers, occupation and k-points

FeatureStatusWhat is certified
Gaussian, Methfessel–Paxton, Marzari–Vanderbilt smearingDoneGaussian: free energy, entropy and energy. MP and MV occupations can exceed 1, so there is no variational principle; instead the checker proves that the discrete SCF equations have exactly one solution near the computed one, bounds its chemical potential, and encloses its energy and free energy (free-energy interval about 5×10−14 wide). Since 3 October this covers Methfessel–Paxton of any order: its occupation, derivatives and entropy are Hermite polynomials times e−x², differentiated term by term once for every order, and the checker computes the order’s derivative bounds from the Hermite coefficients (DFT-T30, T31). DFTK’s Fermi level, energy and free energy all fall inside the Lean intervals, for second-order MP as well.
Monkhorst–Pack grids, shifts, time-reversal reductionDoneEvery grid without the zone boundary is certified; DFTK’s energies lie inside the Lean intervals
Bands and DOS aligned with DFTKAgreesAs in M0, at the Γ point and with k-points. The zone-boundary point X of a band path is certified but not compared: there the certificate keeps the plane waves |k| ≤ K, while DFTK’s spherical cutoff takes 2K + 2 of them, a different discretisation
Grid-based Hamiltonian action, LOBPCG, preconditioningNot startedAlias-free grid action is already proved
Adaptive band count and diagonalisation toleranceNot started
LDOS, dielectric and χ0 mixing; potential mixingNot started
Newton’s method, direct minimisationNot started

M2 · 1D physics terms

FeatureStatusWhat is certified
LDA exchange-correlation (Slater exchange, PW92 correlation)DoneEvaluated on the same real-space grid as DFTK (M = 4K + 1 points); six systems — zero temperature, Fermi–Dirac, Marzari–Vanderbilt, Gaussian with Monkhorst–Pack k-points, bands and DOS — agree with DFTK in energy, every component and every eigenvalue. The energy is no longer convex, so there is no variational certificate; for Slater exchange, with or without PW92 correlation, at the Γ point with any finite-temperature smearing (Fermi–Dirac, Gaussian, MV, or MP of any order), the checker proves that the nonlinear SCF equations have exactly one solution within about 10⁻¹² of the computed one and encloses its chemical potential, energy (exchange-correlation included) and free energy (DFT-X01–X15). For PW92 the correlation is written in s = √rs, where every derivative of its logarithm becomes a rational function regular at s = 0; the checker encloses it by interval arithmetic and bounds |v′| and |v″| for all densities above the grid minimum. The Fermi–Dirac occupation 1/(1 + ex) has its own proved derivative bounds, and its entropy is proved equal to the binary entropy used by the variational certificates (DFT-X16, X17). DFTK’s Fermi level, energy and free energy fall inside the Lean intervals. With Monkhorst–Pack k-points the certificate describes the equivalent supercell on a grid of nk·M points: shifting by one cell is then a grid shift, so the exchange-correlation potential is equivariant, the unique solution is cell-periodic, and the per-cell energy and free energy — exchange-correlation integrated on the cell grid, as DFTK does — are enclosed (DFT-X18–X20); the supercell grid must satisfy Ms ≥ 4Ks + 1, so these runs use M = 4K + 3, and DFTK on the same grid lands inside the intervals. Since 6 October also at zero temperature (certificate fis-lda0): the aufbau occupation is a step, so the checker replaces it by a smooth step that equals it outside (−1, 1) and checks that the Hartree–exchange-correlation Hamiltonian has no eigenvalue within the gap half-width T of μ₀ anywhere in the ball; there the smooth problem and the zero-temperature problem coincide, so the unique solution in the ball is the spectral projector below the gap, with N electrons, and its energy is enclosed (DFT-X21–X25), at the Γ point and with Monkhorst–Pack k-points.
Collinear spin (LSDA), magnetic termDoneSince 6 October, a first version: --spin collinear --xc lda_x --nelec N --bfield B, at the Γ point and finite temperature, with spin-polarised Slater exchange, one chemical potential for both spins, a free magnetisation and an optional uniform Zeeman field B. Each spin has its own density matrix γs = f((H[ρ↑ + ρ↓] + Vxc,s − μ ∓ B)/T), with tr γ↑ + tr γ↓ = N. The two spins form a block-diagonal problem, and the Newton–Kantorovich argument of the LDA certificate carries over block by block: the checker proves that the LSDA SCF equations have exactly one solution within 3×10⁻¹⁴–1×10⁻¹² of the computed one and encloses its magnetisation ν = tr γ↑ − tr γ↓ and chemical potential (DFT-X26–X28). DFTK’s magnetisation and Fermi level (collinear, Xc([:lda_x]), the Zeeman field as a spin-dependent external potential) fall inside the Lean intervals on four systems. DFTK’s Magnetic term is a vector potential, not a Zeeman field, and is not done. Since 7 October: Monkhorst–Pack k-points (--nk, an equivalent-supercell certificate; per-cell magnetisation intervals of width 2×10⁻¹³–5×10⁻¹³, DFT-X35–X37), zero temperature at the Γ point (fis-lsda0: exact integer magnetisation and a certified common gap of both spins, DFT-X40–X42; DFTK’s Smearing.None gives the same ν), and machine-proved bounds on the LSDA energy and free energy at the SCF solution (DFT-X45, X46; the driver’s floating-point E, F agree with DFTK to 7×10⁻¹⁵). Also since 7 October: zero temperature with k-points (per-cell magnetisation as an exact rational, e.g. 4/3 for free electrons with nk = 3 and B = 0.2, and one gap for all k-points and both spins, DFT-X43, X44); per-spin bands, DOS and total DOS of finite-temperature LSDA solutions at Γ and with k-points (DFT-S09–S11, X38; band half-widths 2×10⁻¹¹–3×10⁻¹⁰, all 192 of DFTK’s band values and 252 DOS values inside); certified intervals for the energy and free energy of the LSDA solution, per cell with k-points (DFT-X47; widths 1.5×10⁻¹⁰–1×10⁻⁸ for E and 4×10⁻¹⁵–7×10⁻¹⁴ for F, DFTK inside on six systems); and the named Gaussian potential at the Γ point (GAUSS_LSDA*, DFT-V15–V17); and spin-interpolated PW92 correlation (--xc lda, DFTK Xc([:lda_x, :lda_c_pw])) at finite temperature, at Γ and with k-points: the cross derivatives between the spins make the uniform bounds about 100× too loose, so the checker bounds the exchange-correlation Jacobian on a small box around each grid point instead (DFT-X29–X34; magnetisation width 1.1×10⁻¹³, 10–22 s at Γ, 61 s with nk = 3; DFTK’s ν and Fermi level inside on four systems). Since 8 October PW92 has every result that Slater exchange has: zero temperature at Γ and with k-points (DFT-X48–X50; radius 1.1×10⁻¹⁵–8.4×10⁻¹⁵ at Γ in 7–12 s, ≤ 1.1×10⁻¹³ with k-points in about 60 s), energy and free energy (DFT-X51–X53; widths 1.3×10⁻⁸–5.9×10⁻⁷ for E, the uniform PW92 constants make them one to two orders wider than for Slater exchange, and 3.6×10⁻¹⁴–1.6×10⁻⁹ for F), per-spin bands and DOS (DFT-X54, X55; half-widths ≤ 1.4×10⁻¹⁰ and ≤ 1.9×10⁻⁸, all 192 band and 252 DOS values of DFTK inside) and the named Gaussian potential at Γ (DFT-X56–X58), with DFTK inside every interval. Missing: DFTK’s Magnetic vector potential, bands and DOS at zero temperature, Gaussian LSDA with k-points and Gaussian LSDA bands; physical uniqueness (DFT-T34, T37) is proved for Slater exchange only
Exact exchange (Hartree–Fock, hybrids)Partial: closed-shell HF at finite temperature, Γ pointSince 8 October --xc hf adds exact exchange to rHF, the 1D analogue of DFTK’s ExactExchange: Fock matrix F[γ] = H[γ] − K[γ] with the kernel 4π/q² for 1 ≤ |m| ≤ 2K and v0 = πL²/6 at q = 0 (in 1D the Coulomb kernel is fixed only up to a constant, which changes the exchange energy, so the choice is part of the model and echoed as PROBLEM HF_KERNEL). The checker proves that the HF SCF equations have exactly one solution in a ball around the certificate and bounds its chemical potential (HF_OK, checkHF_sound, DFT-H01–H08); seven systems give radii 4.5×10⁻¹⁵–9.3×10⁻¹⁴ in 0.4–3.1 s. DFTK’s own HF SCF does not converge on these 1D systems, so the cross-check is a fixed-point check: at the certificate’s centre DFTK’s Hamiltonian equals the Fock matrix to 1.7×10⁻¹⁴ and the energy terms agree to 9×10⁻¹⁴ (DFT-H11). Missing: the HF energy interval (DFT-H09), zero temperature, k-points, hybrids and the Gaussian potential

M3 · 1D post-processing and response

FeatureStatusWhat is certified
ForcesPartial: Gaussian sites, rHF, Γ pointSince 8 October --pot sites --forces gives certified forces −∂E/∂xj on every Gaussian site, at zero temperature and with Fermi–Dirac or Gaussian smearing (FORCE_OK, and FORCE_SUM, whose interval contains 0 by theorem; DFT-N04–N08). With L = 4, K = 16 the half-widths are 1.8×10⁻⁴ at zero temperature and 4.5–4.7×10⁻⁵ at T = 0.05, about the square root of the energy width; finite differences of certified energies and DFTK’s compute_forces (ElementGaussian with AtomicLocal, six forces, ≤ 2.7×10⁻¹³ from the driver) fall inside. Missing: k-points (UNSUPPORTED FORCE), SCF-solution certificates (MV, MP), LDA and LSDA (DFT-N09), and the wells of --pot gauss (floating-point forces only)

M4 · d dimensions

FeatureStatusWhat is certified
Lean model generalised to d-dimensional latticesDoneReduced Hartree–Fock in plane waves on any Bravais lattice in d dimensions: a ground state exists, its density is unique and the energy is monotone in the cutoff; the 1D model is the case d = 1, and without the Hartree term the Hamiltonian is the Cohen–Bergstresser one (DFT-M01–M06). For LDA minimisers exist but the density need not be unique (M16). The 1D certificates do not carry over by themselves; the 3D checkers are new (next rows).
2D/3D plane waves, FFTWPartial: 3D done; 2D and FFTW notcertify-pw3d certifies SCF solutions of reduced Hartree–Fock and Kohn–Sham LDA (Slater exchange, PW92 correlation) at the Γ point and on Monkhorst–Pack meshes, at finite and zero temperature, with band intervals at every k point. DFTK’s solutions of the same discrete problems lie inside the uniqueness balls, and every DFTK eigenvalue compared lies in its interval, 8192 of 8192 on an 8×8×8 mesh (DFT-M07–M79, K20, Q10). For real Si at the settings of showcase S1 the mesh gap is certified in [0.5307362, 0.5307368] eV (DFT-K29). The drivers use numpy’s FFT, not FFTW.
Lattices, atomic structures, symmetryPartial: fcc, time reversal, point groupsThe checkers bind fcc cells with the diamond structure. Time reversal (DFT-K08–K10) and point-group reduction (K11–K18, K21–K28) are proved: the checker verifies the space-group operations itself, computes only the irreducible k points and draws its conclusion about the full mesh. Arbitrary lattices and automatic symmetry detection are not done.
HGH and UPF pseudopotentialsPartial: GTH, s and p channelsSeparable GTH/HGH pseudopotentials written as DFTK implements them: local part, two s projectors and the p channel, with file hash and provenance recorded (DFT-Q01–Q12). The d channel and UPF files are not done.
DFTK examples: silicon, graphene, GaAs surface, Cohen–BergstresserPartial: silicon, Cohen–BergstresserSilicon with the GTH-LDA q4 pseudopotential is certified at the settings of the pre-registered S1 protocol (DFT-K29, M78); Cohen–Bergstresser band intervals for Si, Ge and GaAs contain DFTK’s values (DFT-W13, W14). Graphene and the GaAs surface are not done.
Planned: M2–M6 (21 items not started)

M2 · 1D physics terms

  • Local nonlinearity (Gross–Pitaevskii)
  • Atoms and pseudopotentials: local and nonlocal parts, Ewald, correction terms
  • Pairwise interatomic potentials
  • DFT+U

M3 · 1D post-processing and response

  • Geometry optimisation
  • Stresses
  • Response: χ0, Hessian, DFPT
  • Polarisability
  • Phonons
  • Elastic constants
  • A posteriori error estimates and refinement
  • Currents

M4 · d dimensions

  • 3D Ewald and pseudopotential correction
  • Anyons (2D)

M5 · full 3D

  • GGA and meta-GGA
  • 3D forces, stresses, phonons, elastic constants
  • Exact exchange (ACE)
  • MPI k-point parallelism, bitwise identical on 1–64 processes (partly done: S3, 25 tested layouts with between 1 and 32 processes)

M6 · infrastructure (last; may be trimmed)

  • GPU
  • Arbitrary floating-point types, automatic differentiation
  • Wannier90, VTK, JSON and plotting interfaces

04

What a certified run reports today

Rigorous intervals for

  • the discrete ground-state energy, and the truncation limit with a bound on the discretisation error; since 8 October the truncation limit is proved equal to the ground energy of the continuum model, so the LIMIT_OK interval also contains the continuum energy (DFT-C02, C09)
  • the energy under the true Gaussian potential, Fourier tail included
  • the ground-state density (Coulomb norm and pointwise) and the Hartree potential
  • every eigenvalue of the ground-state Hamiltonian, and the gap
  • kinetic, external and Hartree energy components
  • band energies along a path and density-of-states values, at zero temperature and for finite-temperature SCF solutions (rHF and LDA), at the Γ point and, since 6 October, for k-point runs, where the primitive-cell Bloch Hamiltonian is built from the block average of the supercell solution
  • free energy, entropy and energy at finite temperature
  • for MP (any order) and MV smearing: a locally unique SCF solution, its chemical potential, energy and free energy
  • for LDA (Slater exchange, or Slater exchange with PW92 correlation) at finite temperature, with any of the four smearings, at the Γ point or with k-points: a locally unique solution of the nonlinear SCF equations, its chemical potential, energy including the exchange-correlation energy, and free energy
  • for LDA at zero temperature, when the gap stays open over the ball: a locally unique solution that is the spectral projector below the gap, the gap itself and the energy (since 6 October)
  • for collinear-spin LSDA (spin-polarised Slater exchange, or with spin-interpolated PW92 correlation, which since 8 October has all the same results; optional uniform Zeeman field) at the Γ point and finite temperature: a locally unique LSDA SCF solution, its magnetisation and its chemical potential (LSDA_MAG, LSDA_MU; since 6 October), with k-points and at zero temperature with an exact integer magnetisation and a common gap (LSDA0_*), and since 7 October its energy and free energy (LSDA_E, LSDA_F), per-spin bands and DOS (LSDA_BAND, LSDA_DOS, LSDA_DOS_TOT) and, for a Gaussian potential at Γ, the same for the named potential (GAUSS_LSDA*)
  • forces on Gaussian sites (--pot sites --forces, rHF at the Γ point, zero temperature or Fermi–Dirac and Gaussian smearing): an interval for each force FORCE_OK and for their sum FORCE_SUM (since 8 October)
  • for closed-shell Hartree–Fock (--xc hf, finite temperature, Γ point): a locally unique HF SCF solution and its chemical potential (HF_OK, since 8 October)
  • per-cell values for k-point runs, through the equivalent supercell

Every certificate written since 4 October also names its external potential, and the checker prints the problem it certified before any result (next section). Since 7 October “locally unique” means the SCF solution itself: every SCF solution is a zero of the auxiliary map the certificate checks, so the ball contains exactly one SCF solution (DFT-T32–T37).

05

Problem binding: which potential was certified

Until 3 October a certificate carried the external potential only as Fourier coefficients written by the untrusted driver, and the checker certified whatever problem those coefficients described (assumption A-DFT-05): a driver that computed the wrong potential produced a self-consistent certificate that the checker accepted. Since 4 October every certificate the driver writes carries a pot line naming the potential — free electrons, a cosine V₀ cos(2πRx/L) or the periodised Gaussian double well, with the exact doubles used for V₀ and σ and the number of copies R per cell — and the checker compares the coefficients with Lean’s own definition of that potential. Lean also proves that these coefficients are the Fourier series of the real-space potential (DFT-V07), so they are not one more convention to trust.

LineWhat it means
PROBLEM …The problem being certified, in exact decimals: the potential and its parameters, L, K, electron pairs, number of k-points and, by certificate kind, temperature, smearing, exchange-correlation and grid size, and for LSDA the electron count, spin and Zeeman field. Echoed, not recomputed; make verify-all compares it with the command line.
POT_OK exactFree electrons and cosine: every coefficient equals Lean’s, so the certified problem is the named problem — at the Γ point (DFT-I02) and, for k-point supercells, the per-cell problem (DFT-I03). Every result of the certificate is then a statement about the named potential.
POT_OK gauss dGaussian double well: its coefficients involve √(2π) and exp, so the checker encloses Lean’s exact coefficients in rational intervals and proves an ℓ¹ deviation of at most d (about 10⁻¹⁵). Variational certificates — zero or finite temperature, Γ point or k-points — also print GAUSS_OK intervals about the named potential, widened by 2N·d (DFT-V09).
POT_OK file d · POT_OK sites dSince 7 October, your own potential: --pot file --potfile F (exact rational Fourier coefficients) or --pot sites --sites F|t:V0:σ,… (periodised Gaussian sites). The certificate carries the data as exact fractions and the checker echoes every value (PROBLEM POTC, PROBLEM SITE); given the same --potfile or --sites it re-reads the input itself and rejects a mismatch. Dyadic coefficients bind exactly (POT_OK exact); non-dyadic ones and sites give an exactly computed deviation d and then the same GAUSS_… lines as the Gaussian well (DFT-V18–V24). Not yet: the cutoff limit for sites, LSDA and zero-temperature LDA.
POT_FAILThe coefficients are not those of the named potential, or the certificate has no pot line at all (since 5 October the binding is mandatory, DFT-I04); nothing else is printed.

Since 5 October the pot line is mandatory (DFT-I04): the checker runs the binding first for every kind of certificate and rejects one without a pot line, so a driver cannot escape the binding by leaving it out. The archived showcase and research certificates, written before the binding, now carry the pot line too (regenerated, or added after checking that their problem data match the current driver line by line). Since 6 October a proved perturbation estimate (DFT-V10) moves the other Gaussian results to the named potential: changing the potential by d in ℓ¹ shifts every Hamiltonian by a Toeplitz matrix of norm at most d, so the checker re-runs the Newton–Kantorovich test with the widened residual and prints GAUSS_SCF/GAUSS_LDA (a unique SCF solution of the named problem, rHF or LDA, Γ point or k-points), GAUSS_EIG, GAUSS_RHO, GAUSS_BAND, GAUSS_DOS and GAUSS_ENTROPY/GAUSS_ENERGY (DFT-V11–V14). The price is a wider interval: for a double well with Marzari–Vanderbilt smearing (d ≈ 1.3×10⁻¹⁵) the radius of uniqueness grows from 1.3×10⁻¹² to 2.6×10⁻¹¹ and the energy interval from 3.7×10⁻¹⁰ to 7.6×10⁻⁹. Caveats, as in the ledger: zero-temperature LDA (fis-lda0) and the zero-temperature energy components HAR_OK, EXT_OK, KIN_OK still speak about the certificate’s coefficients (within d), and the k-point truncation limit has no bound for the Gaussian tail. What remains trusted is that Lean’s definition of the named potentials matches the one the specification describes.

06

Cross-checks and supporting evidence

Benchmarks do not raise a tier, but they catch misunderstandings: a Lean interval that excludes an independent reference value would reveal a wrong model.

Silicon bands along L–Γ–X–W–K–Γ, every point certified

S1 model (GTH-LDA, a = 5.416 Å, 8×8×8 mesh, the SCF solution of DFT-O04). At each of the 73 listed points the Lean checker encloses bands 1–8 in intervals 0.163–0.170 meV wide (DFT-W282, T4); the lines join the interval midpoints, nothing is claimed between points. Energies relative to the valence-band top at Γ. Vertical lines on Γ–X: the window t ∈ (0.8242, 0.8613) outside which band 5 is proved to lie higher than at t = 0.843, so the minimum along the path lies inside it (DFT-W283, T4); DFTK’s minimum (t = 0.845) lies inside it.

  • DFT benchmark suite, including cross-checks with DFTK.jl (run of 4 October)167 / 167 pass
  • Energies (free energies) against DFTK.jl on the same discrete problem≤ 5×10⁻¹⁵
  • All eigenvalues and energy components against DFTK.jl≤ 10⁻¹²
  • DFTK’s Gaussian free energy, and its MP (first and second order) and MV Fermi level, energy and free energyinside the Lean intervals
  • LDA against DFTK.jl with the same grid: energies, components including exchange-correlation, eigenvalues (six systems); for certified LDA runs DFTK’s Fermi level, energy and free energyagree; inside the Lean intervals
  • LDA SCF-solution certificate, K = 8 (Marzari–Vanderbilt): radius of uniqueness / energy interval / free-energy interval / checking time4×10⁻¹² / 1×10⁻⁸ / 2×10⁻¹⁵ / 6 s
  • LDA SCF-solution certificates with Fermi–Dirac and Gaussian smearing, K = 8: radius of uniqueness / energy interval / free-energy interval / checking time1–8×10⁻¹² / 4×10⁻⁹–2×10⁻⁸ / 4×10⁻¹³–1×10⁻¹² / 8–14 s
  • k-point LDA certificates (equivalent supercell, nk = 2–4, K = 4, M = 19): radius of uniqueness / per-cell energy interval / free-energy interval / checking time2–9×10⁻¹² / 1×10⁻⁹–2×10⁻⁸ / 3×10⁻¹⁵–8×10⁻¹⁴ / 9–53 s
  • Band and DOS certificates for finite-temperature SCF solutions (double well, L = 10, K = 8, T = 0.01): half-width of the band intervals for rHF with MV / LDA (PW92, Fermi–Dirac); DFTK’s non-self-consistent bands and DOS≈ 10⁻¹¹ / ≈ 10⁻⁹; inside the Lean intervals
  • Band and DOS certificates for k-point runs (double well, nk = 3): half-width of the band intervals for rHF with MV (L = 10, K = 4) / LDA (PW92, Fermi–Dirac, L = 4, K = 4) / rHF at zero temperature (L = 4, K = 6); DFTK’s compute_bands (zone boundary X excluded) and compute_dos≤ 5.6×10⁻¹² / ≤ 3.9×10⁻¹¹ / ≤ 3.5×10⁻⁴; inside the Lean intervals
  • Zero-temperature LDA certificates (double well, L = 4, K = 8, M = 33): radius of uniqueness / energy interval / checking time, Slater exchange and PW922.5–5×10⁻¹³ / 9×10⁻¹⁶–3×10⁻¹⁵ / 4–6 s
  • Collinear LSDA certificates (cosine potential and free electrons, N = 2 or 3, B = 0, 0.05 or 0.1, K = 6–8, Fermi–Dirac): radius of uniqueness / checking time; DFTK’s magnetisation and Fermi level, four systems3.3×10⁻¹⁴–9.9×10⁻¹³ / 2.7–6.1 s; inside the Lean intervals
  • Zero-temperature LSDA with k-points (nk = 2–3, K = 4–6): radius of uniqueness / checking time; DFTK’s per-cell magnetisation (the same rational) and Fermi level, four systems5.5×10⁻¹⁵–6.4×10⁻¹⁴ / 6–19 s; inside the Lean gap
  • LSDA bands and DOS (Γ point and nk = 2, 3): half-width of the band intervals / of the DOS intervals; DFTK’s compute_bands and per-spin compute_dos1.6×10⁻¹¹–3.0×10⁻¹⁰ / 1×10⁻⁹–4×10⁻⁸; 192 + 252 values inside
  • LSDA with spin-interpolated PW92 correlation (cosine potential and free electrons, Fermi–Dirac): radius of uniqueness / checking time; DFTK’s magnetisation and Fermi level, four systems3.5×10⁻¹⁴–4.3×10⁻¹³ / 10–22 s (61 s with nk = 3); inside, ≤ 4.5×10⁻¹⁴ from the driver
  • LSDA energy and free energy (cosine, free electrons, nk = 3): width of E / of F / checking time; Gaussian double well (V₀ = 2, T = 0.1) for the named potential1.5×10⁻¹⁰–1×10⁻⁸ / 4×10⁻¹⁵–7×10⁻¹⁴ / 8–55 s; Gaussian E 7.4×10⁻⁹; DFTK inside
  • Forged certificates (adversarial reviews of 6 and 7 October): forged names, huge exponents, partial k-point sets, scaled weights, swapped files and headers, exit codes, spin-down k-point lines, phase tables, in every domain307 / 307 rejected or unchanged
  • PW92 LSDA completion (zero temperature, energy, bands, Gaussian): radius at zero temperature / width of E / band half-width; DFTK Xc([:lda_x, :lda_c_pw])1.1×10⁻¹⁵–1.1×10⁻¹³ / 1.3×10⁻⁸–5.9×10⁻⁷ / ≤ 1.4×10⁻¹⁰; inside everywhere
  • Certified site forces (L = 4, K = 16): half-width at zero temperature / at T = 0.05; DFTK’s compute_forces, six forces1.8×10⁻⁴ / 4.5–4.7×10⁻⁵; inside, ≤ 2.7×10⁻¹³ from the driver
  • Hartree–Fock certificates (seven systems, K = 4–6): radius of uniqueness / checking time; DFTK fixed-point check of the Fock matrix and energy terms; free electrons (mpmath)4.5×10⁻¹⁵–9.3×10⁻¹⁴ / 0.4–3.1 s; ≤ 1.7×10⁻¹⁴ and ≤ 9×10⁻¹⁴; μ inside
  • SCF-solution certificates for Methfessel–Paxton of order 2, 3, 4, K = 8: radius of uniqueness / free-energy interval1–6×10⁻¹² / 4×10⁻¹⁴–5×10⁻¹³
  • Problem binding: a driver that solves another potential while the pot line names the requested one (free electrons at Γ and with k-points, LDA at Γ and with k-points, Gaussian at zero temperature and with MV); a Gaussian V₀ of 2.001 written as 2POT_FAIL
  • Problem binding: one coefficient changed in its last bitPOT_FAIL
  • Problem binding: every DFT test certificate with its pot line removed (and each of the three entropy certificates on its own)POT_FAIL
  • Known answers computed from the command line, not the certificate: free electrons at zero temperature (E = π²/2, π²/54 per cell with k-points), at finite temperature (F, S, E), with Γ-point and k-point LDA (μ, E, F) and with collinear LSDA (ν, μ, E, F, bands, and at zero temperature N↑, N↓ and the gap, also with k-points), computed independently with mpmath; at B = 0 the LSDA magnetisation interval contains 0 and its μ interval meets the LDA certificate’sinside the Lean intervals
  • A driver run with L off by 0.001: the certificate is self-consistent and the checker accepts itcaught by the known answer and the PROBLEM echo
  • Reports with 1, 2, 4 and 7 threadsbitwise identical
  • Trace validation: runs ending converged, gap closed, max iterations, certificate failedall accepted
  • Trace validation: tampered traces (claimed convergence without a certificate, a dropped record, wrong limit)all rejected
  • Mutation tests on the SCF controller spec (remove the certificate check or the iteration cap)caught by TLC
  • Trace validation of the k-point threads: every claim, join and reduction step replayed against the protocol, 1, 4 and 7 threadsall accepted
  • Tampered thread logs (double claim, claim out of counter order, reduction out of order, join before exit, missing chemical-potential step)all rejected
  • Report validation: every Lean line in a report is the checker’s own output; a rejecting, crashing or missing checker makes the driver exit with an errorpass
  • Fuzzing of the driver under AddressSanitizer and UBSan (about 4 000 mutated inputs), after hardening its input parsing; on 4 October it found that a huge k-point count made the supercell certificate ask for about 10¹⁴ bytes, and such certificates are now refused as too largeno memory errors, undefined behaviour or crashes

07

Research result: several exact SCF solutions

The rHF energy is convex, so its SCF solution is unique. LDA exchange is concave, and uniqueness can fail — but a floating-point SCF that lands on different points from different starting densities does not prove it: they could be approximations of one solution, or of none. SCF-solution certificates can. Each proves that exactly one solution lies in a ball around its centre, so accepted certificates for the same problem whose balls are pairwise disjoint prove as many distinct exact solutions.

For a 1D LDA electron gas (Slater exchange with PW92 correlation) with four electrons per period L = 0.6, a cosine potential of period L/4 (V₀ = 20), K = 6, M = 25 grid points and Fermi–Dirac smearing at T = 0.03, five certificates with byte-identical problem data are accepted by the Lean checker. Their radii are at most 1.6×10⁻¹² and their centres 0.048–0.096 apart, so the discrete SCF equations have at least five distinct exact solutions, with pairwise different densities. Disjoint free-energy intervals also order them: the symmetric solution is not the lowest.

SolutionStarting densityRadius rFree energy F, Lean interval
0charge-density wave, φ = π/21.59×10⁻¹²[102.565322714574288, 102.565322714624328]
1charge-density wave, φ = 3π/21.51×10⁻¹²[102.565322714574313, 102.565322714617116]
2charge-density wave, φ = π1.37×10⁻¹²[102.578707142620807, 102.578707142638557]
3charge-density wave, φ = 01.37×10⁻¹²[102.578707142620873, 102.578707142625617]
4uniform (symmetric solution)2.13×10⁻¹³[102.604009214520563, 102.604009214563429]

Mechanism: the half-filled shell k = ±2π/L forms a charge-density wave under exchange, and the potential, which has only frequency-4 components, pins its phase at a few discrete values; φ = 0 and π are related only by a shift the 25-point grid does not have, so they are genuinely different solutions. Limits: the result is about this discrete problem (K = 6, M = 25) — floating-point runs find five solutions at K = 8, 12 and 16 too, without certificates; it is a lower bound on the number of solutions, not a count, and which are stable is not proved; it is an open-shell effect of a small cell at the Γ point, not a bulk instability. The certificates predate the problem binding and have no pot line, which does not affect the conclusion since all five share the same problem data. They are archived in research/kohn-sham-multiplicity/ and re-checked by make verify-all.

08

Open gaps

An output without a Lean theorem is unfinished. These are the ones we know about, listed rather than hidden.

  • ForcesSince 8 October certify prints certified force intervals for Gaussian sites (--pot sites) on rHF Γ-point certificates at zero temperature and with Fermi–Dirac or Gaussian smearing (FORCE_OK, FORCE_SUM, DFT-N05–N08), resting on the Hellmann–Feynman theorem proved on 7 October (DFT-N01–N03). The half-width is set by the density error and is about the square root of the energy width (1.8×10⁻⁴ at K = 16). Missing: k-point certificates (UNSUPPORTED FORCE), SCF-solution certificates (MV, MP, --scf-cert), LDA and LSDA (the derivative along a locally unique SCF branch needs the implicit function theorem, DFT-N09), and the well centres of --pot gauss (floating-point forces only; the same potential written as two --pot sites sites is certified).
  • k-point sums versus the infinite crystalEach k-point run is certified for its own discrete problem, but the error of the k-point sum relative to the infinite crystal has no bound yet; k-point energies are not variational.
  • Grids containing the Brillouin-zone boundaryNo certificate; DFTK’s treatment there is also a different discrete problem, so neither side is compared.
  • Scaling of SCF-solution certificatesChecking time grows roughly with the fourth power of the supercell size; the derivative bounds used for higher-order Methfessel–Paxton are several times the true maxima and loosen as the order rises; for gapped systems with the chemical potential mid-gap the SCF equations are ill-conditioned and the certificate honestly fails; uniqueness is only local, and for LDA several solutions do exist (above).
  • LDA limitsZero temperature is certified only when the gap stays open (a metal at zero temperature has no certificate); k-point LDA certificates need M ≥ 4K + 3 (the default 4K + 1 only agrees with DFTK); near-insulating systems (∂N/∂μ ≈ 0) fail the SCF-solution certificate; the LDA energy interval is a first-order estimate.
  • Collinear spin (LSDA)Certified with spin-polarised Slater exchange and with spin-interpolated PW92 correlation, each with an optional uniform Zeeman field: finite temperature at Γ and with k-points (SCF solution, magnetisation, μ, energy, free energy, per-spin bands and DOS), zero temperature at Γ and with k-points under a common gap, and the named Gaussian potential at Γ. The PW92 energy interval uses uniform constants on [lo, ∞)², which are about two orders of magnitude above the true maxima, so it is one to two orders wider than for Slater exchange; physical uniqueness (DFT-T34, T37) is proved for Slater exchange only. Missing: user sites and non-dyadic coefficient files (rejected), zero-temperature bands and DOS, the gapless or fully polarised zero-temperature case, Gaussian LSDA with k-points (rejected) and Gaussian LSDA bands (the band lines of a Gaussian certificate are about its rounded coefficients). For the Gaussian potential the Newton–Kantorovich condition can fail for the named potential although it holds for the certificate (double well V₀ = 2 at T = 0.05: GAUSS_FAIL). DFTK’s Magnetic term, a vector potential, is not implemented.
  • Exact exchange (Hartree–Fock)Certified: the HF SCF solution and its chemical potential at finite temperature and the Γ point, for exact potentials (DFT-H08). Missing: the energy and free-energy interval (DFT-H09), zero temperature, k-points (the Bloch Fock operator needs pair densities at q ≠ 0), the named Gaussian potential and hybrids. Checking time grows like n⁵ (the unknowns are the n² coordinates of the mean-field matrix), so the driver allows K ≤ 12; gapped systems at small T fail honestly (cosine potential at T = 0.05), and the benchmarks use T = 0.3–0.4. DFTK’s own HF SCF does not converge on these 1D systems, so the DFTK comparison is a fixed-point check, not an SCF-to-SCF comparison.
  • Discretisation error at finite temperatureCertified at zero temperature only.
  • Truncation limit equals the continuum energyMachine-proved since 8 October (DFT-C02, T3, with C03–C09): at the Γ point and zero temperature, for absolutely summable potential coefficients and Npair ≤ 2K + 1, E∞ = infK′ ≥ K EK′ equals the ground energy of the continuum model written in Fourier space. Missing: the equivalence of that model with density-matrix operators on H¹ (assumption A-DFT-08, paper only), the summability link for the named Gaussian and site potentials, finite temperature and k-points; DFT-C01 and D06 stay paper proofs.
  • Density, eigenvalues, bands and DOS beyond zero temperatureAll certified for zero-temperature Γ-point runs; bands and DOS also for finite-temperature SCF-solution certificates and, since 6 October, for k-point runs (rHF with any smearing, LDA, zero-temperature rHF; DFT-S06–S08). Missing: bands and DOS of zero-temperature LDA with k-points, of the named Gaussian potential with k-points, and of finite-temperature variational certificates (Fermi–Dirac or Gaussian without --scf-cert); the comparison with DFTK at the zone-boundary point X, which needs the same discretisation there; and the density, eigenvalues and energy components at finite temperature.
  • The rest of the problem written in the certificateSince 4 October the external potential is checked against its Lean definition (A-DFT-05, above). The other problem data — cell length, cutoff, electron count, temperature, smearing, exchange-correlation, grid size, number of k-points — are the problem itself: the checker echoes their exact values but cannot recompute them, and only make verify-all compares them with the command line (T0); since 7 October the k-points, weights and DOS width of band and DOS runs are echoed too (DFT-I05), and an unknown header or key, a repeated scalar key or a non-ASCII byte is rejected before checking (DFT-I06, I07). Since 5 October a certificate without a pot line is rejected; for a Gaussian potential zero-temperature LDA and the energy components are about the certificate’s coefficients (within about 10⁻¹⁵ of the named potential). Since 6 October make verify-all compares every echo, smearing, functional and grid size included, and the checker parses names and numbers defensively (adversarial review, progress page); its density of states is printed only for a complete k-point set whose weights sum to one.
  • ACSL versus Lean statementsSince 6 October the interval arithmetic is proved on the C code; that its ACSL contracts state the same as the Lean theorems is checked by hand (A-DFT-07), and the proof assumes a contract for nextafter and the value of INFINITY (A-DFT-06).