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
55 changes: 55 additions & 0 deletions research/clifford-compact-adjoint-fibres/CANONICAL.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,55 @@
# Five rational weights for a repeated rank-one stabilizer site

Let $a,b$ be normalized four-qubit stabilizer states in absolute Gaussian-dyadic gauges, with diagonalizing Pauli Lagrangians $L_a,L_b$. Put $B=|a\rangle\langle b|$ and $y=(\mathrm{vec}B)^{\otimes13}$. For the column labelled $(S,p)$ let $V=W_pU_S$, $\tau=\langle a|V|b\rangle$, and
$$
j=d(L_a,SL_b)=4-\dim(L_a\cap SL_b).
$$
With the dictionary normalization in [the adjoint proof](PROOF.md), the full Moore–Penrose solution is
$$
(A^\dagger y)_{S,p}=h_j\,\overline\tau.
$$

| $j$ | $h_j$ |
|---:|---:|
| 0 | $279183413/16081633325721600$ |
| 1 | $52701673/32163266651443200$ |
| 2 | $36859/253254068121600$ |
| 3 | $-777583/64326533302886400$ |
| 4 | $-14688133/1029224532846182400$ |

The input to the executable specifies $a=U_{T_a}|r\rangle$ and $b=U_{T_b}|c\rangle$; changing either by a Gaussian unit gives the thirteenth power of the corresponding phase on $y$. All 36,720 stabilizer lines (2,295 Lagrangians, sixteen eigencharacters each) occur on each side. When $T_a=T_b=I$, $B$ is an ordinary computational matrix unit. For a query, evaluate one four-qubit matrix element and the rank of the displayed intersection. The class computation below is finite preprocessing shared by all inputs and queries. [Executable](canonical.py)

## From the adjoint to five distances

The two stabilizer projectors for $a$ and $Vb$ are normalized signed sums of their sixteen Pauli stabilizers. Their trace pairing vanishes when the two characters disagree on the intersection; otherwise it equals $2^{-j}$ because the intersection contains $2^{4-j}$ labels. Hence $|\tau|^2$ is zero or $2^{-j}$. In the nonzero case $\tau(1+i)^j$ is a Gaussian dyadic of norm one. Factoring its Gaussian-integer numerator over the unique prime above two leaves only a Gaussian unit, so $\overline\tau^{12}=(-1/64)^j$. This gives
$$
(A^*y)_{S,p}=(\overline\tau/4)^{13}=u_j\overline\tau,\qquad
u_j=4^{-13}(-1/64)^j
$$
also when $\tau=0$. Keeping the complex phase here is essential even for real-looking rank-one inputs.

In each fibre the 256 matrices $V_{S,p}$ are a trace-orthogonal basis of squared norm 16. Under synthesis $M_c(S)=\sum_pc_{S,p}V_{S,p}$, the adjoint data therefore become $16u_j B$. Coefficient extraction is $c_{S,p}=\operatorname{Tr}(V_{S,p}^*M_c(S))/16$. By #7816, $K=A^*A$ acts entrywise on $M$ by the real central convolution with $F(g)=16^{-12}(-64)^{\dim\operatorname{Fix}(g)}$ on $G=\mathrm{Sp}(8,2)$. This Gram is invertible. On functions $f(S)=w(d(L_a,SL_b))$, left and right transport identify its action with that at a single diagonal reference Lagrangian $Z$ on both sides. Namely, for $L_a=A_0 Z$ and $L_b=B_0 Z$ substitute $A_0^{-1}SB_0$ in the function argument; centrality makes both translations commute with the convolution. The stabilizer $P$ of $Z$ acts transitively on Lagrangians $X$ at each fixed distance from $Z$ by extending a pair of intersection-adapted symplectic bases. Thus the space of such radial functions is five-dimensional and retained by both $K$ and $K^{-1}$. [Central convolution and source conventions](https://github.com/apoursh/integer-mult-bounds/blob/7e7e62c62fcccd11f5c8e22677b4523863a48f4b/research/clifford-transported-fibre-gram/PROOF.md)

## Exact radial normal equations

Write $\Omega=G/P$ for the 2,295 Lagrangians, with shell sizes $(n_0,\ldots,n_4)=(1,30,280,960,1024)$. These also follow by selecting the intersection in $Z$ and then a symmetric graph on its quotient, $n_j={4\brack j}_2\,2^{j(j+1)/2}$. The executable generates all of $\Omega$ by closing $Z$ under specified symplectic transvections; the count, shell sizes and Lagrangian invariant certify coverage. For an ordinary conjugacy representative $g_i$ let $D_{ij}=\#\{X\in\Omega:d(X,g_iX)=j\}$. The operator $\sum_g F(g)g$ on functions on $\Omega$ has common entry $b_j$ on every ordered pair at distance $j$. Counting all such ordered pairs gives
$$
b_j=\frac{\sum_i |C_i| F(g_i)D_{ij}}{|\Omega|n_j}.
$$
The pushforward from the group has exactly this normalization: the entry from $X$ to $Y$ sums over every group element carrying $X$ to $Y$, including the entire stabilizer multiplicity. Equivalently one may use $g^{-1}$ since $F(g^{-1})=F(g)$. We compute the displacement counts on every retained ordinary representative and check equality on the fused rational pair before using its combined class size. The class sizes, fixed dimensions and all 81 representatives were computed natively in the referenced #7816 GAP source. [Frozen class data](source/CLASS-ALGEBRA.json) · [Native regeneration](https://github.com/apoursh/integer-mult-bounds/blob/7e7e62c62fcccd11f5c8e22677b4523863a48f4b/research/clifford-transported-fibre-gram/CLASS-ALGEBRA.md)

Choose one $Y_i$ at distance $i$ from $Z$ and form the $5\times5$ matrix
$$
K_{ik}^{\rm rad}=\sum_{X:d(Z,X)=k} b_{d(Y_i,X)}.
$$
Exact elimination gives the displayed solution of $K^{\rm rad}h=u$. Multiplying each $16h_jB$ by $K$ gives $16u_jB$ on *all* group arguments. Since the full Gram is invertible, synthesis and extraction prove the claimed Moore–Penrose formula. Tests check the five normal equations, $n_iK_{ik}=n_kK_{ki}$, all distance branches under independent dense lift multiplication, and a separate spherical-eigenvalue reconstruction. The displacement construction uses $81\cdot2295$ class-representative/Lagrangian pairs and five distance rows. It never forms the $12,128,668,876,800$-column dictionary. [Direct checks](verify_canonical.py) · [Independent spectral reconstruction](spherical.py)

## Error and support on the complete dictionary

Trace orthogonality gives $\sum_p|\tau|^2=16$, so each fibre has exactly $16\cdot2^j$ nonzero $\tau$. All five $h_j$ are nonzero. Each shell of distance $j$ contains $|P|n_j$ group elements, and $|P|=47,377,612,800/2295$. Thus $|P|\sum_j n_j16\cdot2^j=8,338,459,852,800$ canonical coefficients are nonzero, the same support as the adjoint on this family. The orthogonal projector satisfies
$$
\|AA^\dagger y\|^2=\langle A^*y,A^\dagger y\rangle
=16|P|\sum_j n_j u_j h_j
=\frac{478566778595}{5856573836677349376}.
$$
The input has unit norm. Subtracting this fraction from one gives the minimum squared least-squares residual in the full Clifford span. The sparse known-fibre interface of #7816 by itself would require nearly every fibre; here the rank-one stabilizer symmetry gives a compact inverse despite that density. The argument applies to a repeated separable stabilizer site. Site-varying computational points have the adjoint census in [COORDINATES.md](COORDINATES.md); their inverse-Gram reduction remains open.
57 changes: 57 additions & 0 deletions research/clifford-compact-adjoint-fibres/COORDINATES.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,57 @@
# Exactly how dense is the adjoint of one computational coordinate?

Even a width-13 chart supported at one raw address produces trillions of nonzero adjoint coefficients. Their exact count depends on the correlations between the row and column nibbles of its 13 sites. The counts below concern the standard point observation with amplitude one. They do not assume a Clifford decomposition of that observation.

| Attained extremum over all raw addresses | Occupied symplectic fibres | Nonzero Pauli-labelled coefficients |
|---|---:|---:|
| Minimum | 21,139,292,160 | 5,411,658,792,960 |
| Maximum | 47,377,612,800 | 8,338,459,852,800 |

## Count for any coordinate

Write $V=\mathbb F_2^n$, $q=2^n$, and denote the raw site row and column nibbles by $(r_i,c_i)\in V^2$. Use $n=4,f=13$ for the counts above. Let $Z$ denote the diagonal Pauli Lagrangian. For a fixed symplectic element $S$, write $H$ for the projection of $SZ$ onto the output $X$ bits and put $h=\dim H$. Let $J$ be the $X$-to-$X$ block of $S$. The support of the zero column of its unitary lift $U_S$ is an affine translate of $H$. Indeed it is stabilized by $SZ$ with specified signs. Diagonal stabilizers give $n-h$ independent support equations; the remaining stabilizer translations generate $H$ and preserve absolute value. Column $c$ is obtained by applying the conjugated input $X^c$, so its support is shifted by $Jc$. Each nonzero matrix entry has absolute value $2^{-h/2}$ by normalization.

For the label $X^xZ^z U_S$ the adjoint entry on this coordinate is a product of $f$ conjugated unitary matrix entries divided by $q^{f/2}$. It is nonzero exactly when the constraints on $x$ at the individual sites agree. With differences from site one, this requires
$$
\Delta r_i+J\Delta c_i\in H \quad\text{for every }i .
$$
When compatible, $x$ runs through one coset of $H$ and $z$ is arbitrary. Exactly $2^{n+h}$ coefficients survive in the fibre, each of absolute value $2^{-(n+h)f/2}$.

For each candidate subspace $H\le V$, quotient the output by $H$ and write $m=n-h$. The map $\bar J:V\to V/H$ is surjective since the full symplectic map has surjective output $X$ block and the images of the input $Z$ directions generate $H$. The site differences prescribe $\bar J(\Delta c_i)=\overline{\Delta r_i}$ on a subspace of dimension $s=\operatorname{rank}(\Delta c_i)$. If the prescribed pairs do not form the graph of a linear map on that subspace, this $H$ contributes zero. Otherwise let $b=\operatorname{rank}(\overline{\Delta r_i})$ and set $E(a,d)=\prod_{j=0}^{d-1}(2^a-2^j)$ for $d\le a$, zero for $d>a$. The number of surjective extensions is
$$
2^{b(n-s)}E(n-s,m-b).
$$
To see this, choose a complement to the prescribed domain. Its map to the codomain modulo the prescribed image must be onto, and each such map has $2^{b(n-s)}$ lifts.

The multiplicity of each surjective $\bar J$ at fixed $H$ is
$$
2^{h(h+1)/2}\,\frac{|P_Z|}{E(n,m)},\qquad
|P_Z|=2^{n(n+1)/2}|\operatorname{GL}(n,2)|.
$$
Indeed any Lagrangian $L$ projecting onto $H$ satisfies $L\cap Z=H^\perp$. Modulo $H^\perp$, its lifts of $H$ are the graphs of maps $H\to H^*$, with isotropy equivalent to symmetry. This gives $2^{h(h+1)/2}$ choices. The symplectic group acts transitively on Lagrangians by extending a basis to a symplectic basis. For each $L$ there are $|P_Z|$ maps with $SZ=L$. The stabilizer $P_Z$ of $Z$ consists of a general linear basis change and a symmetric shear, giving its displayed order. Right composition with general linear basis changes preserving $Z$ permutes the surjections $V\to V/H$ transitively, so every surjection has the same multiplicity. Summing the product of the extension count and this multiplicity over $H$ counts exactly the occupied fibres.

## Sharp extrema

For $h=n$ the compatibility conditions disappear at every address. This yields the lower bound; take $c_i=0$ and use rows $0,e_1,\ldots,e_n$ among the $f$ sites to force $H=V$. Taking every site equal makes all fibres compatible, attaining the upper bound. These two witnesses exist for $f\ge n+1$. Each occupied rank-$h$ fibre always contributes the same number of entries and the same positive squared norm, so the witnesses also minimize and maximize the squared norm. At width 13 the extrema of $\|A^*\delta\|_2^2$ are $315/2^{70}$ and $87222165730875/2^{70}$, respectively.

For the all-equal address the numbers of compatible fibres by $h=0,1,2,3,4$ are 20,643,840, 619,315,200, 5,780,275,200, 19,818,086,400 and 21,139,292,160. These follow from the same multiplicity or by counting the $h$-dimensional $H$. The computation enumerates only the 67 binary subspaces in dimension four. The independent small-qubit checker generates the 6 and 720 symplectic classes of lifts modulo Paulis in dimensions two and four via actual Hadamard, phase and controlled-NOT matrices. It tests all pairs of site addresses at each size and 240 further multisite lists, for 512 literal comparisons. [Census implementation](coordinate_support.py) · [Small-qubit verification](check_coordinates.py)

## Independent check of the five canonical weights

For a repeated site the canonical calculation in [the main proof](PROOF.md) reduces to a radial Gram on the graph of the 2,295 Lagrangians. The following check uses the earlier quotient character spectrum and the nearest-neighbour graph alone to compute this action, independently of the conjugacy-class displacement computation. Write $i=4-\dim(L\cap L_0)$ and connect vertices sharing a hyperplane. The number of neighbours at distances $i-1$ and $i+1$ is
$$
c_i=2^i-1,\qquad b_i=2(16-2^i).
$$
Indeed each hyperplane of $L$ extends to two other Lagrangians, since its perpendicular quotient has dimension two. Exactly $2^i-1$ hyperplanes contain $L\cap L_0$. For each of these, one of the other two extensions adds the line cut out by $L_0$ in that quotient, reducing the distance; the other keeps it fixed. If the hyperplane misses part of $L\cap L_0$, the original $L$ already uses the unique line from $L_0$, so both alternatives increase the distance. This gives shell sizes $k=(1,30,280,960,1024)$.

The stabilizer of $L_0$ acts transitively on each shell by extending bases adapted to two Lagrangians to symplectic bases. Consequently the permutation representation has a five-dimensional commutant, one dimension per orbit on ordered pairs. The adjacency matrix has five distinct eigenvalues: with $u(0)=1$, the tridiagonal radial eigenvalue equation
$$
\theta u(i)=c_i u(i-1)+(30-c_i-b_i)u(i)+b_i u(i+1)
$$
gives and checks $\theta=30,13,3,-5,-15$ at both boundaries. The adjacency algebra thus exhausts the commutant and every eigenspace is an irreducible constituent of multiplicity one. Its projection kernel based at a vertex is a multiple of $u$. Transitivity, the trace and idempotence give its multiplicity $2295/\sum_i k_i u(i)^2$. The respective dimensions are 1, 135, 1190, 918 and 51, each unique among the 81 quotient characters in the frozen #7772 table. Their normalized Gram eigenvalues are, in this order,
$$
\frac{470624175}{2^{32}},\quad \frac{1132984125}{2^{32}},\quad
\frac{5790052125}{2^{33}},\quad \frac{15074084025}{2^{33}},\quad
\frac{22846598775}{2^{32}} .
$$
Project the radial adjoint values onto each $u$, divide by the corresponding Gram eigenvalue, and sum. All five canonical weights and the squared projection norm $478566778595/5856573836677349376$ agree exactly with the displacement-based solve. The norm is strictly between zero and one and all weights are nonzero. [Independent computation](spherical.py) · [Source spectrum](source/width13-spectrum.json)
Loading
Loading