Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
264 changes: 264 additions & 0 deletions research/clifford-bell-fibres/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,264 @@
# A Gaussian-dyadic left inverse at the sharp four-qubit width

Let `R = Z[i,1/2]`, `d = 16`, and fix one Gaussian-dyadic unitary `U` in
each four-qubit Clifford coset modulo scalar powers of `i`. Let `A_f` have
columns `(vec(U)/4)^(tensor f)`. Vectorizing the tensor-power matrices in
one step instead merely permutes the coordinates.

**Theorem.** There is an explicit matrix over `R` with `B_f A_f = I` at
every integer `f >= 13`. Together with the reported complex rank failure
below width 13 in [Ali Romman's #7766](https://github.com/aliromman/integer-mult-bounds/blob/d7c2dd1868c70f0addce2c0a4d3220149452f14b/research/clifford-four-qubit-threshold/PROOF.md),
this determines the least Gaussian-dyadic width as 13. The construction
of the upper bound is independent of that spectral calculation.

## 1. Six Bell pairs and one matrix

Write `V = F_2^4 ⊕ F_2^4` and use unphased real Weyl matrices
`W_t = X^x Z^z` for `t=(x,z)`. The commutation sign is the symplectic
pairing `<t,t'> = x·z' + z·x'`, and `W_t^T = (-1)^q(t) W_t` with
`q(t)=x·z`. The 256 vectors `vec(W_t)/4` form an orthonormal basis of
the two-register space. Conjugation by a Clifford gives

```text
U W_t U* = gamma_U(t) W_(St), S in G = Sp(V).
```

All phases here and below are units of `R`. A scalar of modulus one in
`R` is a power of `i`, by reducing a sum of two squares equal to a power
of four modulo four and dividing through by four repeatedly. The usual
Clifford action surjects onto `G`. A Hermitian Pauli lift `P` of each
nonzero `p` gives a dyadic unitary `(1+i)(I+iP)/2` inducing the
symplectic transvection at `p`, and these generate the symplectic group.
Its projective kernel consists of the Weyl matrices. Indeed, a Clifford
fixing every label has conjugation signs in `{+1,-1}` by squaring
`W_t`. These signs form a character of `V`, hence have the form
`(-1)^<p,t>`. After removing `W_p` the unitary commutes with every Weyl
matrix, which forces it to be scalar. Thus each fibre of the map to
`S` consists of 256 representatives proportional to `W_p U_S`.

On a Bell pair the action of `U ⊗ U` reads

```text
U W_t U^T = theta_U(t) W_(L_S t), L_S t = St + b_S .
```

In fact `U bar(U)^(-1) = U U^T` is a Pauli up to phase, since entrywise
conjugation preserves the symplectic action on the real Weyls. Left
multiplication of `U` by any Pauli preserves all labels in this formula.
The affine maps compose, as follows directly by applying the two-register
action twice. Transposition also gives `q(L_S t)=q(t)`.

For an ordered six-tuple `t=(t_1,...,t_6)` put `w=sum_i t_i` over the
binary field. If its output is `t'=L_S t` componentwise, then
`w'=sum_i t'_i = Sw`, since six copies cancel the affine offset. Under
the Bell basis change on six pairs, the block of `U^(tensor 13)` between
these tuples is `beta_U(t) U`, where `beta_U(t)=product_i theta_U(t_i)`.
Every other output block for this input is zero. Left multiply this
observed block by `W_(w')` and right multiply by `W_w^(-1)`. The block
becomes

```text
sigma_S(t) U, sigma_S(t) = beta_U(t) / gamma_U(w).
```

This phase depends only on `S`. Replacing `U` by `W_p U` multiplies each
Bell phase by `(-1)^<p,t'_i>` since `W_p^T=W_p^(-1)`, hence multiplies the product by
`(-1)^<p,w'>`. The same sign multiplies `gamma_U(w)`. A scalar lift
`zU` contributes `z^12=1` to the product and leaves conjugation unchanged.
Composition in the two matrix actions gives
`sigma_(S_1 S_2)(t) = sigma_(S_1)(L_(S_2)t) sigma_(S_2)(t)`.

All these operations are fixed `R`-linear maps on the observations.
Concretely, first multiply by `4^13` to undo the column normalization,
then unvectorize, regroup six pairs, and take the stated input and output
Bell matrix elements. This reasoning by columns therefore also applies
to an arbitrary mixture with coefficients `c_U`.

## 2. Which sums the orbit supplies

Write `Z_j, X_j` for the corresponding binary labels, and let
`L = span(Z_1,...,Z_4)`. Use the ordered seed

```text
T = (0, Z_1, Z_2, Z_3, Z_4, X_4).
```

A symplectic map stabilizing this tuple in the Bell action has `b_S=0`,
fixes `L` pointwise and fixes `X_4`. Fixing `L` pointwise forces a shear
`(x,z) -> (x,z+Bx)` with `B` symmetric. Fixing `X_4` means its fourth
row and column vanish. Preserving `q`, now with zero offset, forces
zero diagonal. Conversely the three ordinary real controlled-Z gates
on qubit pairs 12, 13 and 23 generate all eight such shears. Their lifts
fix each seed Weyl by ordinary conjugation and have `UU^T=I`.
They also fix the product of the seed Weyls, hence `gamma_U(w)=1`.
Thus the stabilizer `C` has order eight and `sigma_c(T)=1` throughout.

Identify the orbit of `T` with `G/C`. For each tuple `i` there, fix any
transporter `g_i` from the seed, and put `j_i = sigma_(g_i)(T)`. The
cocycle and its trivial value on `C` make this phase independent of the
transporter. For `S i=j` we have `sigma_S(i)=j_j/j_i`. Define the
unknown fibre matrix `M_S=sum_(U over S) c_U U`. Multiplying each
corrected observation block by `j_i/j_j` therefore exposes the sum

```text
sum_(S in g_j C g_i^(-1)) M_S .
```

An explicit transporter can always be chosen by finite enumeration.
For any subgroup `H` of power-of-two order write `e_H = sum_(h in H) h/|H|`
in the group algebra `R[G]`. The coefficient vectors of the available
coset sums are precisely the coefficient vectors of the elements
`8 g_j e_C g_i^(-1)`. Arbitrary left or right translates yield the
same set, since group elements can be replaced by coset representatives.
Consequently we can recover every `M_S` linearly over `R` once

```text
R[G] e_C R[G] = R[G].
```

## 3. A dyadic identity for the stabilizer average

Take the full abelian shear group `A = Sym_4(F_2)` fixing `L` pointwise,
including its diagonal entries. Regard a shear parameter `B` as a
symmetric matrix on a chosen basis of `L`. For any vectors `u,v` of
`L`, write `B_uv = uv^T + vu^T`. Each character of `A` is determined
by a binary quadratic form `Q` on `L` as follows:

```text
chi_Q(uu^T) = (-1)^Q(u), chi_Q(B_uv) = (-1)^lambda(u,v),
lambda(u,v) = Q(u+v)+Q(u)+Q(v).
```

Indeed, the four diagonal coefficients and six off-diagonal
coefficients of a linear functional on symmetric matrices freely
specify the four linear and six cross coefficients of `Q`. For each
of these 1,024 characters set
`e_Q = 2^-10 sum_(B in A) chi_Q(B) B`, using group elements in this
sum. Character orthogonality gives pairwise orthogonal idempotents with
`sum_Q e_Q=1`. A basis change `g` of `L`, acting contragrediently on
the complementary dual, sends `B` to `g B g^T`, so acts on the
characters by the natural change of variables.

For any isotropic three-plane `W <= V` the conjugate `C_W` of `C`
consists of the eight products of its three pair shears. In basis-free
terms the generators for basis pairs `u,v` send
`s -> s + <s,u>v + <s,v>u`. This is the seed description above for
`W=span(Z_1,Z_2,Z_3)`, transported by a symplectic change of basis.
The stated map for each pair and all their products depend only on
`W`. Such basis changes exist by extension of an isotropic basis to
a symplectic basis. In particular, each `e_(C_W)` belongs to the
two-sided ideal generated by `e_C`.

If the alternating polar form `lambda` of `Q` is degenerate, choose a
three-plane `W <= L` isotropic for `lambda`. Its existence follows
from the rank being zero or two. In rank two, extend the two-dimensional
radical by any vector outside it. Then `C_W <= A` and `chi_Q` is one
on its three generators. Hence `e_Q e_(C_W) e_Q=e_Q`.

Now let `lambda` be nondegenerate. There exists a nonzero `r in L`
with `Q(r)=0`: on a two-dimensional `lambda`-isotropic plane `Q`
is linear, so has a nonzero zero. Set `H=r^perp` for this polar form,
take a basis `z_1=r,z_2,z_3` of `H` and extend it by `z_4` to `L`.
Then `lambda(z_4,r)=1`. Put `P=span(r,z_2)`. In the complementary
dual Lagrangian take `x_4` pairing one with `z_4` and zero with `H`
in the original symplectic form on `V`. Choose the mixed isotropic plane

```text
W = span(r,z_2,x_4).
```

Its group `C_W` consists of the eight products of `B_(r,z_2)` and the
four basis changes `T_p` for `p in P` sending `z_4 -> z_4+p` and fixing
`H` pointwise, with contragredient action on the dual. This follows
from the three pair-shear formulas. Every element normalizes `A` and
`chi_Q(B_(r,z_2))=1`.

For each nonzero `p in P`, the character moves under `T_p`. If `p=r`,
then `Q(z_4+r)=Q(z_4)+1`. If `p` is outside the span of `r`, then
`p` is outside the radical of `lambda|H`, which is the line through
`r`. Pick `y in H` with `lambda(p,y)=1`. The transformation changes
`lambda(z_4,y)`, hence changes `Q`. Thus only the two elements of
`C_W` with `p=0` stabilize the character. For each of those,
`e_Q c e_Q=e_Q`; for every other `c`, orthogonality of `e_Q` and
`c e_Q c^(-1)` gives zero. Consequently

```text
e_Q e_(C_W) e_Q = (2/8) e_Q = e_Q/4 .
```

This constructs the identity in the ideal. For each `Q` choose a plane
and a transporter `g_Q` as above. With coefficient `d_Q=1` in the
degenerate case and `d_Q=4` in the nondegenerate case,

```text
1 = sum_Q d_Q e_Q g_Q e_C g_Q^(-1) e_Q .
```

Expanding the character idempotents expresses this formula as a
dyadic linear combination of the available translates. Multiplying
by each group element on the left gives all single-fibre indicators.
There is no division by the order of the symplectic group.

## 4. Decode the Pauli fibre and pad

For each fixed projective representative `U` over `S`, return
`Tr(U* M_S)/16`. Any other representative over `S` differs by a
nonidentity Pauli on the left up to scalar, and has zero trace pairing
with `U`. The target pairing is 16, so this recovers exactly `c_U`.
It uses coefficients in `R`, even for arbitrary fixed scalar lifts.
Together the steps specify a linear left inverse on the entire
observation space. At larger widths, tensor each row with `f-13`
copies of its target's single-copy functional `(vec(U)/4)^*`. This
functional is one on the target column. All other columns are already
annihilated by the first thirteen factors. Applying any ring homomorphism
from R to a residue field preserves the identity B_f A_f = I. In
particular, the columns remain independent modulo every odd Gaussian
prime for f >= 13, including characteristics dividing the group order.

## 5. Stability on the normalized observations

The single functional which extracts one corrected coset block and traces
it against `U*/16` has Euclidean row norm `4^13/4 = 4^12`. The Bell
input and output vectors have unit norm, the Weyl corrections are unitary,
and `||U/16||_F=1/4`. Each `e_Q` has coefficient absolute sum one.
The expansion in Section 3 therefore costs at most `sum_Q d_Q/8 <= 512`
in absolute coefficients of coset sums. This bounds the norm of each row
of `B_13` by `2^33`. Tensoring with the unit-norm padding preserves that
bound for every `f>=13`. With `N=12,128,668,876,800` rows we obtain
`||B_f||_2 <= sqrt(N) 2^33 < 2^55`.

Assume the published uniform spectral enclosure of Ali Romman's
[#7772, exact source](https://github.com/aliromman/integer-mult-bounds/blob/1398fbdca3536389d8ac3c8b3aaa408f03e8f6c2/research/clifford-threshold-conditioning/PROOF.md).
For `A=A_f` and `K=A* A` it gives eigenvalues in
`[470624175/2^32, 7841980575/2^30]`. Start at `D_0=B_f` and use the
exact dyadic iteration from [#7798, Section 4](https://github.com/apoursh/integer-mult-bounds/blob/0ae10f215780296ca50142fd36fdf904b8c1b146/research/clifford-width17/PROOF.md#4-exact-dyadic-stabilization):

```text
D_(j+1) = (I-K/4) D_j + A*/4 .
```

Induction gives `D_j A=I`, and subtracting the canonical fixed point
`A^+=K^(-1) A*` multiplies its error on the left by `I-K/4`. The
enclosure gives `||I-K/4||_2 < 63/64` and `||A^+||_2 < 3.02095`
(the latter is checked by squaring against the lower endpoint).
The initial error is less than `2^56`. Since `(63/64)^64<1/2`,
at `j=8192` the operator-norm error is less than `2^-72`, giving
`||D_8192||_2 < 3.021`. This yields bounded noise amplification for
an exact Gaussian-dyadic decoder at the first possible width as well
as all larger widths. The width and splitting proof in Sections 1–4
uses no spectral bound.

The construction concerns exact algebraic recovery. It supplies no
priced physical realization, new supplier or improvement in κ. The
rank failure at width 12 is an input from #7766. An inverse at any
smaller width could be padded to width 12 by Section 4, contradicting
that rank failure. The full
Clifford alphabet and production multiplier are not enumerated here.
The earlier width-17 construction and its general stabilizer result
retain their distinct scope. [Width-17 handoff #7798](https://github.com/CrocSwap/integer-mult-bounds/pull/7798).

Clifford actions on Weyl operators, Bell coordinates, symplectic
transvections and character idempotents are classical. The projective
alphabet and complex threshold follow Ali Romman and the retained work
of alephnan and Nick Petalas. Prepared by **apoursh** with substantial
OpenAI Codex and separate agent assistance. Apache-2.0.
34 changes: 34 additions & 0 deletions research/clifford-bell-fibres/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,34 @@
# Exact Gaussian-dyadic Clifford recovery at width 13

For the full four-qubit projective Clifford alphabet, **13 is the minimum tensor width for exact coefficient recovery over the Gaussian dyadics**. At width 13 an explicit left inverse of the normalized tensor map uses only coefficients in `Z[i,1/2]`. Tensoring each output row with its target's unit-norm single-copy functional extends it to every larger width. Ali Romman's complex rank obstruction at width 12 supplies the lower bound. The construction resolves all four widths left open by #7798. [Proof](PROOF.md) · [Complex threshold #7766](https://github.com/CrocSwap/integer-mult-bounds/pull/7766) · [Width-17 handoff #7798](https://github.com/CrocSwap/integer-mult-bounds/pull/7798)

The observation is an arbitrary linear combination of `(vec(U)/4)^(tensor f)`, with one fixed Gaussian-dyadic representative per Clifford scalar coset. The proof constructs a linear decoder on the whole observation space and proves exact recovery on the image. It also implies full column rank after reduction modulo every odd Gaussian prime. The established width-17 threshold for all eight-qubit stabilizer rays concerns a larger alphabet.

## Construction

1. Use six Bell pairs to read which symplectic transformations carry one ordered tuple of labels to another. The thirteenth copy retains a matrix. A known Weyl correction removes phases depending on the chosen Pauli lift, exposing the sum of the unknown matrices over eight symplectic transformations.
2. Translate these sums to isolate one symplectic transformation at a time. The certificate checks all 1,024 characters of the binary symmetric shear group. For 576 characters, the subgroup average acts as the identity on the character component. For the other 448, a change of symplectic coordinates gives one quarter of that component. Multiplication by four completes the same reconstruction over the dyadics.
3. Within each recovered matrix, a normalized trace against each of its 256 possible Clifford representatives returns that representative's coefficient.

Conditional on #7772's uniform spectral enclosure, the exact dyadic iteration from #7798 yields an inverse with Euclidean operator norm **below 3.021 at every width f ≥ 13**. This controls arbitrary additive observation error. The canonical least-squares map has odd denominators at width 13. The two decoders agree on the noiseless image and may differ on arbitrary observations. The physical cost of implementing either the initial decoder or the 8,192 iterations used for the norm bound remains open. This contribution makes no change to κ. [Stability proof](PROOF.md#5-stability-on-the-normalized-observations) · [Spectral bounds #7772](https://github.com/CrocSwap/integer-mult-bounds/pull/7772) · [Canonical obstruction #7792](https://github.com/CrocSwap/integer-mult-bounds/pull/7792)

## Reproduce

Run from the repository root with CPython 3.12 and the standard library.

```sh
python3.12 -B scripts/research_package.py --package research/clifford-bell-fibres
```

- `verify.py` constructs binary symplectic matrices, checks the eight-element stabilizer among all 64 candidate shears, and verifies every character identity using literal matrix conjugation. It also checks the Fourier resolution, failure controls and the numerical margins for the stability corollary.
- `independent_check.py` uses separate computational-basis four-qubit matrices. Three calibration frames cover 297 operator instances with different global phases and 6,144 within-fibre coefficient checks. A second experiment passes an actual mixture with 80 nonzero coefficients through the Bell measurements. It recovers all 2,048 coefficients on the restricted alphabet, including 1,968 zeros and deliberate cancellations.
- Both checkers run normally and under optimized Python. The proof supplies the full-alphabet statement. No full tensor matrix, production multiplier, external spectrum replay, human proof certification or formal-kernel check is included. [Replay commands](VERIFY.json) · [Pinned research inputs and context](SOURCES.json)

## Credit and coordination

Prepared by **apoursh** with substantial OpenAI Codex and separate agent assistance, under Apache-2.0. Ali Romman's work supplies the original complex threshold, spectral enclosure, Bell-sector analysis and canonical obstruction. The Pauli and character-theoretic lineage includes alephnan and Nick Petalas. The starting question and the dyadic stabilization iteration are from #7798. Projective Clifford actions, finite Fourier analysis and group-algebra idempotents are classical. The exact width-13 upper bound in Sections 1–4 has its own proof. [Bell-sector predecessor #7789](https://github.com/CrocSwap/integer-mult-bounds/pull/7789) · [Research claim](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19753161) · [Task scope and overlap](TASK.json)

## Open next questions

1. What does one transported Bell block cost at a named caller? Include coordinate access, precision, movement, workspace and cleanup before summing the cost of the full decoder.
2. Can a sparse, exact width-13 implementation reach a useful noise bound with fewer Gram applications? Start with the character sum for one output coefficient and quantify cancellations at a fixed input layout.
50 changes: 50 additions & 0 deletions research/clifford-bell-fibres/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,50 @@
{
"schema_version": 1,
"base_commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"files": [
{
"path": "certificates/selected-result.json",
"usage": "provenance",
"git_blob": "75d77b328ea02ded535e7f23aee5b1e55d0bb1a8",
"sha256": "154c3d30455d778fd3343b6b4851f0fa4665f0f7625b0adb3675e2bafbbf6bf1"
},
{
"path": "docs/research/agent-task-catalog.json",
"usage": "provenance",
"git_blob": "3404fbdd5e7529a0674eef26114b5071e4586cec",
"sha256": "ead6cdc9d0af6d58e8a49744d8ec8c7c15e13d2ba93c276c434bcf9b51b93a81"
}
],
"cited_research": [
{
"pr": 7798,
"head": "0ae10f215780296ca50142fd36fdf904b8c1b146",
"author": "apoursh",
"use": "Width-17 split, broader stabilizer lower bound, stability iteration and handoff."
},
{
"pr": 7766,
"head": "d7c2dd1868c70f0addce2c0a4d3220149452f14b",
"author": "Ali Romman",
"use": "Complex rank failure at width 12 for sharpness."
},
{
"pr": 7772,
"head": "1398fbdca3536389d8ac3c8b3aaa408f03e8f6c2",
"author": "Ali Romman",
"use": "Uniform spectral bounds used only in the stability corollary."
},
{
"pr": 7789,
"head": "a5aa2ae21ae68338390f584b8fca58ecba0df409",
"author": "Ali Romman",
"use": "Bell-line representation and spectral precedent."
},
{
"pr": 7792,
"head": "9321c67e29c3c41a188154bc5c2add7d7802b694",
"author": "Ali Romman",
"use": "Canonical odd denominator comparison."
}
]
}
Loading
Loading