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
49 changes: 49 additions & 0 deletions research/clifford-two-site-canonical/ADDRESS-CENSUS.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,49 @@
# Exact support distribution over two freely varying coordinates

Fix a repeated matrix unit in common input and output stabilizer bases. Among the 65,536 ordered pairs of matrix units on the two remaining sites, there are seven possible numbers of nonzero coefficients of the Hilbert adjoint. Every pair appears once, including both choices equal to the repeated coordinate.

| Nonzero Clifford labels | Ordered pairs |
|---:|---:|
| 5,929,571,450,880 | 420 |
| 5,978,984,546,304 | 1,350 |
| 6,028,397,641,728 | 18,900 |
| 6,078,693,113,856 | 44,100 |
| 6,670,767,882,240 | 90 |
| 6,781,947,346,944 | 675 |
| 8,338,459,852,800 | 1 |

At width 13 with **all** sites free, there are exactly 30 support sizes among the $2^{104}$ coordinates. With the first address fixed, the number attaining the minimum of 5,411,658,792,960 is 74,959,609,988,606,102,773,113,260,160 out of $2^{96}$ (94.6123%). The maximum is attained only when every site agrees. With zero through six free sites beyond one common address, there are respectively 1, 3, 7, 13, 22, 27 and 30 sizes. The full histogram stabilizes in its set of sizes by six free sites; the multiplicities continue to change. The census concerns the raw adjoint $A^*$.

## Counting quotient isomorphisms

Work in computational bases by conjugating by the two fixed Clifford basis changes. Write $X,Y$ for the row and column spaces $\mathbb F_2^n$. Subtract the repeated or first coordinate from all sites, and let $D\le X\oplus Y$ be the span of the resulting displacement vectors. Put $a=\dim\mathrm{pr}_X D$, $b=\dim\mathrm{pr}_YD$, and $c=a+b-\dim D$. The pure row and pure column subspaces of $D$ have dimensions $a-c,b-c$. Their two projection quotients are identified by $D$ and have common dimension $c$.

At Hadamard rank $h$, the displacement support of a Clifford lift is a linear relation $R$ given by two $h$-planes $U\le X$, $V\le Y$ and an isomorphism $X/U\to Y/V$. This follows from the column-support derivation in [the coordinate census of #7821](https://github.com/apoursh/integer-mult-bounds/blob/eead545a4ba6c50010fcac0919a9fdc6b14bd673/research/clifford-compact-adjoint-fibres/COORDINATES.md). Every column is one coset of $U$, and the surjection from columns to $X/U$ has kernel $V$. Each such relation occurs for
$$
m_h=2^{h(h+1)/2}|P_Z| / \prod_{i=0}^{n-h-1}(2^n-2^i)
$$
symplectic fibres, where $|P_Z|=2^{n(n+1)/2}|\mathrm{GL}(n,2)|$. The coefficient is nonzero when $D\subseteq R$ and the Pauli translate includes the first address, giving $2^{n+h}$ nonzeros per occupied fibre. The absolute square of each coefficient is $2^{-w(n+h)}$ at width $w$.

For $D\subseteq R$, its pure subspaces must lie in $U,V$. The intersections $U\cap\mathrm{pr}_XD$ and $V\cap\mathrm{pr}_YD$ must induce the **same** subspace of the common quotient. Otherwise a vector equivalent to zero on one side maps to a nonzero class on the other. If this induced subspace has dimension $e$, there are $\genfrac{[}{]}{0pt}{}{c}{e}_2$ choices. Set $k=c-e$. The intersections now have prescribed dimensions $a-k,b-k$. The number of $h$-planes having one prescribed intersection of dimension $s$ with an $a$-plane is
$$
L(a,s)=\genfrac{[}{]}{0pt}{}{n-a}{h-s}_2\,2^{(a-s)(h-s)}.
$$
Indeed choose the projection of the $h$-plane to the ambient quotient of the $a$-plane, then its graph into the $a$-plane modulo the fixed intersection. Infeasible dimensions have count zero. Finally the isomorphism on the two $(n-h)$-dimensional quotients must extend a fixed isomorphism between $k$-planes. There are $2^{k(n-h-k)}|\mathrm{GL}(n-h-k,2)|$ extensions. Choose the isomorphism on the complementary quotient and an arbitrary component in the target $k$-plane. Thus the exact number of relations is
$$
\sum_{e=0}^c \genfrac{[}{]}{0pt}{}{c}{e}_2 L(a,a-k) L(b,b-k)
2^{k(n-h-k)}|\mathrm{GL}(n-h-k,2)|,\qquad k=c-e,
$$
omitting $k>n-h$. Multiplying by $m_h$ gives occupied fibres. The formula is symmetric in $a,b$ and uses only three ranks. It yields both the support size and the squared adjoint norm by shell summation.

## Counting address lists

With $t$ varying sites and a fixed base address, the number of ordered lists of displacement type $(a,b,c)$ is
$$
\genfrac{[}{]}{0pt}{}{n}{a}_2\genfrac{[}{]}{0pt}{}{n}{b}_2\genfrac{[}{]}{0pt}{}{a}{c}_2\genfrac{[}{]}{0pt}{}{b}{c}_2 |\mathrm{GL}(c,2)|
\prod_{i=0}^{a+b-c-1}(2^t-2^i).
$$
The first factors choose the projected subspaces, their pure kernels and the isomorphism between the $c$-dimensional quotients. These data recover $D$ uniquely as a relation. The final product counts surjections $\mathbb F_2^t\to D$, since they are exactly the ordered $t$-tuples spanning $D$. A varying base contributes an extra factor $2^{2n}$. The formulas have no dependency on the other sites' order or on a chosen coordinate gauge.

The [calculator](address_census.py) produces the table and complete histograms. The [independent checker](check_address_census.py) constructs every quotient plane and surjection directly for $n\le4$. In four qubits this visits all 65,536 distinct support relations and checks all 55 ordered displacement types, a different enumeration from summing subspace intersection counts. It also counts all two-site address pairs, checks normalization, and independently verifies that the sum of squared adjoint norms over all addresses equals the number of unit dictionary columns for widths 1, 2, 6 and 13. The calculation extends apoursh's support criterion, individual-address counts and extrema in #7821.

The same closed formula also checks every support multiplicity in the [two-site normal operator](TWO-SITE-SPECTRUM.md), including its off-diagonal entries. Form $D$ from the two directly matched addresses, the two crossed addresses, or their union. The three shell counts reproduce the independently enumerated direct, crossed and simultaneous support counts on every row and column type.
12 changes: 12 additions & 0 deletions research/clifford-two-site-canonical/CROSS-CHECK.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,12 @@
# Independent check across the two kinds of nonzero mark

At $X=L$ take a nonzero interior mark $u\in L$ and an exterior mark $w\notin L$. There are two parabolic orbits of such ordered pairs, classified by the symplectic pairing $s=\langle w,u\rangle$. In coordinates $L=Z$, the external $X$ part of $w$ is a nonzero functional on $L$. A change of basis fixing $u$ is transitive on the functionals of each specified value at $u$. Symmetric shears add an arbitrary $Z$ component to the vector with a given nonzero $X$ part (prescribe the corresponding column of a symmetric matrix). For each interior $u$ the numbers of allowed exterior $w$ are $n_0=112,n_1=128$.

Write $k_C$ for the central inverse value on conjugacy class $C$, computed directly from the full character table by the formula in #7823. Set $f_j=4^{-13}(-1/32)^j$. For each class representative $g$, sum over the 2,295 Lagrangians and form the weighted displacement count
$$
T_s(g)=\sum_L f_{d(L,gL)}\,
\#\{w\in (g^{-1}L)\setminus L:\langle w,gw\rangle=s\}.
$$
Indeed in the raw right hand side after shifting $(L,w)$ by $g$, the only surviving interior mark is $u=gw$. For fixed $s$ the inverse result is constant on all $2295\cdot15\cdot n_s$ triples $(L,u,w)$ by transitivity. Its value is therefore $\sum_C |C| k_C T_s(g_C)/(2295\cdot 15\cdot n_s)$. Reversing the two marks gives the same value: substitute $g^{-1}$ for $g$, using $k(g)=k(g^{-1})$ and $d(L,gL)=d(L,g^{-1}L)$. The checker compares all members of each fused rational class packet before applying its multiplicity.

This calculation makes no use of the marked orbit traversal, separating operator, or polynomial interpolation. Its exact answers are $224457234251/202337455041975052800000$ for the perpendicular pair and $-235362773012219/27517893885708607180800000$ for the paired case. Both agree with both orientations in the decoder. [Class-sum checker](check_cross.py)
25 changes: 25 additions & 0 deletions research/clifford-two-site-canonical/DENOMINATOR.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,25 @@
# Exact common odd denominator

The least positive odd integer clearing every canonical two-site coefficient into $\mathbf Z[i,1/2]$ for all Gaussian-dyadic joint inputs is
$$D=21{,}556{,}663{,}020{,}569{,}210{,}160{,}884{,}269{,}703{,}315{,}625.$$
It has 115 bits. The analogous one-site divisor from #7823 has 73 bits. This statement prices the fixed coordinate interface. Circuit costs depend on the producer, precision and schedule.

**Reduction to five full Walsh sweeps.** For each distance $j$, use the symplectic map exchanging $Z_s$ and $X_s$ on the first $j$ qubits and fixing the others. Every relative $S'$ at this distance has the form $A S B^{-1}$ for this representative $S$ and $A,B$ preserving $Z$. Indeed pairs of Lagrangians of the same distance lie in a single symplectic orbit, by extending a common intersection basis to adapted symplectic bases. Because the scalar Gram is central, its transformed flag weight is invariant under simultaneously moving the two marked flags. Therefore the weight at $S'$ in the coefficient formula equals the weight at $S$ with the output mark transformed by $A^{-1}$ and the input mark by $B^{-1}$. A linear symplectic map preserving $Z$ acts on a fixed coset $(z,x)$ by an invertible linear change of $x$ and, within the coset, an affine change of $z$. Such changes permute Walsh frequencies and introduce only signs. Thus the set of absolute scalar multipliers is identical throughout each shell, also when the physical bases of the repeated site are rotated.

Write $x=r\oplus i$, $y=c\oplus k$, $\alpha=r\oplus m$ and $\beta=c\oplus n$. The four independent four-bit strings parametrize all $65,536$ Walsh values per shell, including the extraction factor $1/16$. The remaining factor is a conjugated matrix entry. When nonzero this is a Gaussian-dyadic unit: its squared modulus is $2^{-j}$, so its inverse is also Gaussian-dyadic. At least one Pauli translate has a nonzero entry at any requested matrix address since translations of any nonempty column support reach all rows in the left stabilizer basis. All four strings are realizable taking $r=c=0$. These observations prove that the lcm of the odd denominators of the Walsh values is exactly the desired joint divisor. Linear superposition with Gaussian-dyadic coefficients preserves the bound.

| Distance $j$ | Distinct absolute multipliers | Zero multipliers out of 65,536 | Distinct odd divisors | Shell lcm |
|---:|---:|---:|---:|:---|
| 0 | 10 | 64,140 | 10 | $D/3$ |
| 1 | 25 | 62,720 | 22 | $D$ |
| 2 | 42 | 57,216 | 35 | $D$ |
| 3 | 44 | 38,912 | 36 | $D$ |
| 4 | 27 | 0 | 24 | $D$ |

No individual matrix-unit entry realizes all the prime powers of $D$. A pair already forces it. For $B=E_{00}$, take $p=0$ in both queries. At the identity, the defects $E_{11},E_{11}$ have canonical odd denominator $562788258996425089753125$. At the one-qubit exchange representative, $E_{22},E_{44}$ have canonical odd denominator $6293916210385170849893217431625$. Their lcm is $D$. The verifier recovers these divisors from actual complex output coordinates. [Complete sweep and witnesses](denominators.py)

A single query attains the complete divisor if the joint observation is allowed to be a superposition. With $B=E_{00}$, choose the four-qubit exchange representative ($j=4$), $p=0$, and
$$
J=E_{00}\otimes E_{11}+E_{11}\otimes E_{22}+2E_{11}\otimes E_{42}.
$$
Its canonical coefficient is exactly $35434033216309953801405149/(2^{16}D)$ in lowest terms. Thus one real output of an integer input can require the entire 115-bit odd divisor.
38 changes: 38 additions & 0 deletions research/clifford-two-site-canonical/DOUBLE-READOUT.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@
# Canonical readout with two arbitrary sites

Use the Gaussian-dyadic four-qubit Clifford lifts and the width-13 dictionary of #7816 and #7823. In fixed left and right stabilizer bases let the repeated site be the matrix unit $B=E_{rc}$. At the other two sites take an arbitrary joint tensor $J$ with entries $J_{ikmn}$ in the basis $E_{ik}\otimes E_{mn}$. This includes entangled two-site inputs. The observation is $(\mathrm{vec}\,B)^{\otimes 11}\otimes J$. All thirteen factors of a dictionary column are identical, so the choice of the two site positions is immaterial.

Write $V$ for the specified lifted Pauli translate expressed in those two bases, $Z$ for the computational Lagrangian and $S$ for the relative binary action. Thus $V=U_L^*W_pU_TU_R$ and $S=L^{-1}TR$. Write $j=4-\dim(Z\cap SZ)$ and $\tau=V_{rc}$. The width-13 adjoint entry for a basis pair $E_{ik}\otimes E_{mn}$ is $4^{-13}\overline\tau^{11}\overline{V_{ik}}\overline{V_{mn}}$. A nonzero $\tau$ has squared modulus $2^{-j}$ and $\overline\tau^{12}=(-1/64)^j$ by the Gaussian-dyadic phase identity of #7821. Consequently, including $\tau=0$,
$$
4^{-13}\overline\tau^{11}=h_j\tau,
\qquad h_j=4^{-13}(-1/32)^j.
$$

## One Pauli conjugation

Let $P_{(z,x)}=i^{|x\cap z|}X_xZ_z$ be a trace-orthogonal Hermitian Pauli basis, with $z$ stored in the low four bits. Expand $|c\rangle\langle k|=16^{-1}\sum_v\langle k|P_v|c\rangle P_v$. Then
$$
\tau\overline{V_{ik}}=\langle r|V|c\rangle\langle k|V^*|i\rangle
=\frac1{16}\sum_{u,v}\langle r|P_u|i\rangle\langle k|P_v|c\rangle\,
\mathbf1_{\{u=Sv\}}\epsilon_v(V),
$$
where $VP_vV^*=\epsilon_v(V)P_{Sv}$. The signs depend on the Pauli translate. When $u=Sv$, multiplication $V\mapsto P_u VP_v$ acts on this particular basis vector by that sign.

For a fixed $S$, synthesize the adjoint coefficients by $M(S)=\sum_p d(S,p)V$. The trace basis identity is $\sum_p\overline{V_{mn}}V=16E_{mn}$. Multiplication of each coefficient by $\epsilon_v(V)$ therefore replaces $16E_{mn}$ by $16P_uE_{mn}P_v$. With $t_{uv}=\langle r|P_u|i\rangle\langle k|P_v|c\rangle$ this gives
$$
M_{A^*y}(S)=\sum_{u,v}t_{uv}\,h_j\mathbf1_{\{u=Sv\}}P_u E_{mn}P_v.
$$
The Gram of #7816 acts on each scalar entry of $M$ by the central $F(g)=16^{-12}(-64)^{\dim\operatorname{Fix}(g)}$ on $\mathrm{Sp}(8,2)$. Constant left and right basis changes preserve this action, as does replacing $T$ by its relative coordinate $S$ since $F$ is central. Each scalar input $S\mapsto h_j\mathbf1_{\{u=Sv\}}$ is the pullback of $(X,w)\mapsto h_{d(Z,X)}\mathbf1_{\{u=w\}}$ along $S\mapsto (SZ,Sv)$. Its inverse $K(Z,u;X,w)$ is computed on that permutation module. This uses one arbitrary marked vector, including vectors outside either Lagrangian. A single incident-flag table covers the smaller one-site problem in #7823.

## Coefficient formula

Extraction of a coefficient of $M$ is $\mathrm{Tr}(V^*M)/16$. For a nonzero $t_{uv}$ we must have $u=(z,r\oplus i)$ and $v=(t,c\oplus k)$. Multiplying $t_{uv}$ with $P_uE_{mn}P_v$ cancels both Hermitian phase choices and leaves the single matrix unit $E_{r\oplus i\oplus m,c\oplus k\oplus n}$ with sign $(-1)^{z\cdot(r\oplus m)+t\cdot(c\oplus n)}$. Indeed the two factors of $i^{|x\cap z|}$ on the left multiply to $(-1)^{x\cdot z}$, and the signs from the two input indices there are $(-1)^{z\cdot(i\oplus m)}$. On the right the signs from the two input indices are $(-1)^{t\cdot(c\oplus n\oplus c\oplus k)}$, which combines with $(-1)^{t\cdot(c\oplus k)}$ to give $(-1)^{t\cdot(c\oplus n)}$. Thus the basis-pair answer is
$$
\frac{\overline{V_{r\oplus i\oplus m,c\oplus k\oplus n}}}{16}
\sum_{z,t\in\mathbf F_2^4}
(-1)^{z\cdot(r\oplus m)+t\cdot(c\oplus n)}
K\bigl(Z,(z,r\oplus i);SZ,S(t,c\oplus k)\bigr).
$$
Multiply each answer by $J_{ikmn}$ and sum. The implementation uses one length-256 Walsh transform for each occupied pair of displacements $(r\oplus i,c\oplus k)$ and combines all terms reading the same output unit. There are at most 256 such transforms and 256 output units. All requested Pauli translates of a given target reuse the same transformed weights. The implementation starts with explicit rational real and imaginary input entries and the supplied symplectic labels and bases. A physical producer and tape cost for general two-site inputs remain open.

Independent controls compare the untransformed formula with direct degree-13 entries across all five Lagrangian distances, including rotated bases. To exercise the normal equation without using the full-group inverse implementation, a second test takes the six embedded one-qubit symplectic elements and all 256 four-qubit Pauli translates. The principal scalar Gram on those six elements is inverted directly. Every one of the 1,536 resulting entries satisfies the actual Choi normal equation, whose factors are computed independently from $[\operatorname{Tr}(V^*V')/16]^{13}$. Full-group controls interchange genuinely distinct singleton matrix units and reduce either chosen singleton to the repeated site, recovering the independently computed #7823 weights.
Loading
Loading