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
70 changes: 70 additions & 0 deletions research/clifford-accelerated-dyadic-decoder/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,70 @@
# Two finite schedules at tensor width 13

Let $R=\mathbb Z[i,1/2]$. For one Gaussian-dyadic unitary $U_a$ in each projective four-qubit Clifford class, let $A$ have columns $(\operatorname{vec}(U_a)/4)^{\otimes13}$. There are $N=12,128,668,876,800$ columns. Write $K=A^*A$, $J=K^{-1}A^*$, and use the Euclidean operator norm. The width-13 spectrum gives **52 distinct positive rational levels** $x_0<\cdots<x_{51}$. In particular,
$$
x_0=470624175/2^{32},\qquad x_1=359889075/2^{31},\qquad x_{51}=7841980575/2^{30}.
$$
The starting matrix $B$ is any left inverse with $\|B-J\|<2^{56}$. The dyadic Bell-fibre construction satisfies this hypothesis. We count one application of $K$ to the current output vector per step, with $B y$ and the reusable $A^*y$ supplied. [Exact spectrum, PR 7772](https://github.com/aliromman/integer-mult-bounds/blob/1398fbdca3536389d8ac3c8b3aaa408f03e8f6c2/research/clifford-threshold-conditioning/PROOF.md) · [Bell inverse and starting error, PR 7799 §5](https://github.com/apoursh/integer-mult-bounds/blob/7f08a74037788311cc6b7da27297b25f1b188e45/research/clifford-bell-fibres/PROOF.md)

For a real step $h$, update $D\gets D+h(A^*-KD)$, starting at $D=B$. For an observation $y$ this means $d\gets d+h(A^*y-Kd)$. Each update preserves $DA=I$. The error after any prefix is $D-J=p(K)(B-J)$, where $p(X)=\prod_j(1-h_jX)$. As $K$ is Hermitian, $\|p(K)\|=\max_i|p(x_i)|$. Every exact left inverse also satisfies
$$
DD^*=K^{-1}+(D-J)(D-J)^*.
$$
Indeed $(D-J)A=0$ kills both cross terms. The least attainable squared norm is therefore $1/x_0$.

## 1. A dyadic schedule with relative residual below $2^{-129}$

For each level $x$ define
$$
S_x=x\prod_{y\ne x}|1-x/y|,\qquad
T_x=\max_{y\ne x}64y/|1-y/x|,\qquad
b_x=\max(0,\lceil\log_2(2^{128}S_x)\rceil,\lceil\log_2 T_x\rceil).
$$
Use $h_x=2^{-b_x}\operatorname{round}(2^{b_x}/x)$, nearest with ties to even, in **decreasing order of $x$**. These 52 real-dyadic steps are specified without large matrix operations.

For the residual at $x$, its own multiplier has absolute value at most $x2^{-b_x-1}$. The other multipliers obey
$$
|1-h_y x|\le (1+1/128)|1-x/y|
$$
by the choice of $T_y$. Thus the 52-factor product is bounded by $2^{-b_x-1}S_x(1+1/128)^{51}<2^{-128}$. The last inequality follows from $(1+1/128)^{51}\le1/(1-51/128)<2$. Exact evaluation at the 52 rational nodes gives the stronger certificate
$$
2^{-130}<\max_i|p(x_i)|<2^{-129}.
$$
Consequently $\|D-J\|<2^{-73}$ and $\|D\|^2-1/x_0<2^{-146}$. All operations stay in $R$ for $B\in R^{N\times 256^{13}}$.

Both schedules in this note have ordered positive steps $h_1\le\cdots\le h_m$. For a fixed $x>0$, the factors $|1-h_jx|$ are at most one until $h_jx>2$, and at least one thereafter. Prefix products have their maximum at an endpoint. The final product at each node is at most one, so **every prefix has norm at most one** as an operator on the initial error. The order matters numerically even in this finite example. Applying the same 52 factors in reverse gives a prefix norm strictly between $2^{138}$ and $2^{139}$ at step 48.

### A coefficient bound for writing out the decoder

The nominal $b_x$ range from 69 to 266 with sum 5,676. Reducing each $h_x$ gives a total denominator exponent of 5,627. This counts dyadic scalar coefficients. To bound coefficients in the full decoder, the stabilizer-state result $4U_a\in\mathbb Z[i]^{16\times16}$ gives $2^{52}A$ integral. Also $K(a,b)=(\operatorname{tr}(U_a^*U_b)/16)^{13}$. The relative Clifford has finite order, and its Gaussian-rational trace, a sum of roots of unity, is a Gaussian integer. Thus $2^{52}K$ is integral. [Stabilizer entry bound, PR 7802 §4](https://github.com/apoursh/integer-mult-bounds/blob/ebe2c019e398d73444e4f33f2d71bd5775795a01/research/clifford-dyadic-noise-floor/PROOF.md)

For the explicit initial Bell inverse of PR 7799, $2^{27}B$ is integral. Indeed its normalization multiplier $4^{13}$ and the six Bell bras and six Bell kets with entries of denominator four leave the factor $4^{13}/4^{12}=4$. The Weyl permutations and phase corrections are integral. Isolating a fibre costs at most $2^{23}$ in the denominator, through two shear-character averages of order $2^{10}$ and a coset average of order eight. The terminal coefficient comes from $U_a^*/16$, with denominator dividing $2^6$. The scalar ledger gives the exponent $23+6-2=27$. Adding the exponents under the 52 updates therefore places every coefficient of every prefix on the $2^{-8358}\mathbb Z[i]$ grid, since $27+52\cdot52+5627=8358$. This is a sufficient common grid. Movement, accumulator headroom, and a gate count for $A^*$ and $K$ require a concrete implementation. [Bell maps and averages, PR 7799 §§1–4](https://github.com/apoursh/integer-mult-bounds/blob/7f08a74037788311cc6b7da27297b25f1b188e45/research/clifford-bell-fibres/PROOF.md)

## 2. Exact optimum with 50 dyadic steps and one odd reciprocal

Construct a real polynomial $r$ of degree at most 50 by prescribing at the 51 nonbottom nodes
$$
r(x_j)=(-1)^{j-1}/(1-x_j/x_0),\qquad 1\le j\le51.
$$
Its values alternate strictly, so it has one simple real root in each open interval $(x_j,x_{j+1})$, $1\le j\le50$, and degree exactly 50. For each root $z$ use the scalar $2^{-192}\lfloor 2^{192}/z\rfloor$, in order of decreasing root. Complete the sequence by the exact factor $h=1/x_0=2^{32}/470624175$. These steps approximate the normalized polynomial $(1-X/x_0)r(X)/r(0)$, chosen to vanish on the bottom level and equioscillate on the rest. The package rounds reciprocals using only rational interpolation and integer sign comparisons. In particular, its generator never approximates a root in floating point. All 51 factors are listed in SCHEDULE.json.

Exact substitution of the stored shifts gives $p(x_0)=0$ and $2^{-59}<\max_i|p(x_i)|<2^{-58}$. Set $C=D-J$. The range of $C$ is orthogonal to the bottom eigenspace and $\|C\|<1/4$. On that eigenspace $DD^*=K^{-1}+CC^*$ is exactly $1/x_0$. On its orthogonal complement the largest eigenvalue is at most $1/x_1+1/16<1/x_0$, since the spectral gap $1/x_0-1/x_1$ exceeds three. Hence the exact corrected decoder **attains** $\|D\|^2=1/x_0$. If $B$ is dyadic, all odd denominators enter through the single last scalar. The sharp necessity in PR 7802 then gives the least odd denominator exactly $470,624,175$. This realizes the optimum among left inverses, including noncanonical ones. [Sharp denominator theorem, PR 7802](https://github.com/apoursh/integer-mult-bounds/blob/ebe2c019e398d73444e4f33f2d71bd5775795a01/research/clifford-dyadic-noise-floor/PROOF.md) · [Separate canonical-denominator result, PR 7805](https://github.com/apoursh/integer-mult-bounds/blob/d3df403a13bee082a8cef62f190024aaf0cd08dd/research/clifford-canonical-denominator/PROOF.md)

## 3. Why the pass counts are sharp in the stated model

A normalized residual polynomial has $p(0)=1$ and may have complex coefficients. For any $d+1$ distinct positive nodes $X$ and any $p$ of degree at most $d$, Lagrange interpolation at zero gives
$$
1\le\max_{x\in X}|p(x)|W(X),\qquad W(X)=\sum_{x\in X}\prod_{y\in X,\, y\ne x}|y/(y-x)|.
$$
The verifier computes the following strict integer bounds. The independent replay uses Newton divided differences on the alternating data $(-1)^i$ in increasing node order, since these are the signs of the Lagrange basis values at zero.

| Nodes, indexed from zero in increasing order | Certified weight sum | Consequence for the full spectrum |
|---|---:|---|
| All 52 levels | $2^{60}<W<2^{61}$ | Every degree at most 51 has residual norm $>2^{-61}$. |
| All levels except $x_{21}$ | $2^{52}<W<2^{53}$ | Every degree at most 50 has residual norm $>2^{-53}$. |

Thus 52 is the least degree for contraction by $2^{-128}$. On the full set, the exact degree-51 minimax residual equals $1/W$, attained by interpolation of the alternating signs divided by $W$. This also explains the abrupt gain in the last degree.

For the second claim, fix a polynomial of degree at most 50 in advance. There is a level $x$ with $|p(x)|>2^{-53}$. Let $u$ be a unit $K$-eigenvector at that level and $v$ a unit ambient vector perpendicular to the columns of $A$. Give the method the initial inverse $B'=J+2^{55}uv^*$. It has $B'A=I$ and admissible error $2^{55}<2^{56}$. Its final error norm exceeds four, so the identity for $DD^*$ forces $\|D\|>4>1/\sqrt{x_0}$. The same strict obstruction holds for some dyadic starting inverse. Indeed choose a dyadic split $L$, put $P=I-AL$, and observe that $B'=L+(B'-L)P$. Approximating $B'-L$ entrywise by a dyadic $H$ gives the exactly feasible $L+HP$ arbitrarily close to $B'$. Continuity preserves both strict bounds. Therefore 51 is the smallest degree of a **fixed normalized residual polynomial guaranteeing the optimal norm for every such initial dyadic left inverse**, when the final odd reciprocal is allowed.

These are uniform polynomial lower bounds given only a spectral residual or initial norm requirement. The Bell inverse may have additional structure in its error, and more general algorithms may use projections or side information. Neither lower bound prices a physical multiplier. Larger tensor widths require a new spectral check for these node-specific schedules.
31 changes: 31 additions & 0 deletions research/clifford-accelerated-dyadic-decoder/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
# Sharp polynomial pass counts for the four-qubit Clifford decoder

At tensor width 13, **52 applications of the Gram matrix suffice for a Gaussian-dyadic left inverse within $2^{-73}$ of the canonical decoder** in operator norm. The previous explicit construction used 8,192 applications of a fixed step and also covered larger widths. The 52-step schedule uses the exact discrete width-13 spectrum. Every prefix recovers noiseless inputs exactly and amplifies the initial error relative to the canonical decoder by at most one. [Derivation, lower certificates and coefficient bounds](PROOF.md) · [Complete scalar schedules](SCHEDULE.json) · [Previous uniform construction, PR 7799 §5](https://github.com/apoursh/integer-mult-bounds/blob/7f08a74037788311cc6b7da27297b25f1b188e45/research/clifford-bell-fibres/PROOF.md)

| Goal at width 13 | Exact scalar arithmetic | Gram applications | Lower certificate |
|---|---|---:|---|
| Contract the initial error uniformly by at least $2^{-128}$ | All 52 steps dyadic, actual residual below $2^{-129}$ | **52** | Every normalized polynomial of degree at most 51 has residual above $2^{-61}$. |
| Attain the optimal squared noise norm $2^{32}/470624175$ | 50 dyadic steps on a 192-bit grid, then one multiplication by $2^{32}/470624175$ | **51** | Every fixed normalized polynomial of degree at most 50 fails for some exact dyadic initial inverse within the stated error bound. |

For observation $y$, initialize $d=By$ with the Bell inverse and precompute $t=A^*y$. Each listed scalar $h$ prescribes the update $d\gets d+h(t-Kd)$ with $K=A^*A$. Only the second schedule uses odd division, and its denominator realizes the previously proved sharp minimum among all optimal-norm decoders. The scalar precisions of the first schedule range from 69 to 266 fractional bits. The proof gives a sufficient grid for writing out every coefficient of that decoder. [Sharp arithmetic necessity, PR 7802](https://github.com/apoursh/integer-mult-bounds/blob/ebe2c019e398d73444e4f33f2d71bd5775795a01/research/clifford-dyadic-noise-floor/PROOF.md)

Both lower bounds concern fixed residual polynomials $p$ with $p(0)=1$. The second uses only the hypothesis $\|B-J\|<2^{56}$ about the initial error and allows other exact dyadic starting inverses besides the Bell inverse. The 52-level program counts exact Gram applications. Producing the input, pricing an individual application and realizing a multiplication supplier remain open. No change to the selected multiplication exponent is claimed.

## Reproduce

From the repository root, use CPython 3.12 and the standard library.

~~~sh
python3.12 -B scripts/research_package.py --package research/clifford-accelerated-dyadic-decoder
~~~

The primary verifier regenerates the shifts from Ali Romman's spectrum using rational interpolation, integer root comparisons and explicit rounding. It checks both schedules and both degree lower bounds. The second verifier recovers the levels from Nick Petalas's fixed-space character projections. It checks integer prefix products, computes the lower certificates using divided differences and applies the actual updates to a rectangular dyadic test with a nondiagonal Gram matrix. Reversing the 52 shifts gives an intermediate amplification above $2^{138}$, which the prescribed order avoids. The generator supports writing a fresh copy of SCHEDULE.json via its `--emit-schedule` option. [Source pins](SOURCE-PINS.json) · [Registered checks](VERIFY.json)

## Credit

Ali Romman supplied the four-qubit spectrum and conditioning. Nick Petalas supplied the projected quotient character data from the ATLAS and Thomas Breuer's GAP Character Table Library. The four-qubit Clifford conventions also build on alephnan's Pauli-sector work. The starting dyadic splitting, the sharp optimal-noise denominator, and the separate larger arithmetic cost of the canonical inverse are earlier contributions of apoursh. The present results use classical Lagrange interpolation, alternation and spectral polynomial iteration. Prepared by **apoursh** with substantial OpenAI Codex assistance. Apache-2.0. [Source files and predecessor credits](SOURCE-PINS.json) · [Canonical denominator, PR 7805](https://github.com/apoursh/integer-mult-bounds/blob/d3df403a13bee082a8cef62f190024aaf0cd08dd/research/clifford-canonical-denominator/PROOF.md) · [Investigation claim](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19755740)

## Open next questions

1. Can one Gram application be evaluated on a compact acquisition interface? Start with one specified observation layout and count data movement, arithmetic, and intermediate precision for $K d$. The pass counts leave those costs open.
2. Does the particular Bell inverse already have zero error on any of the 52 levels? Project its error onto the smallest block first. Extra exact zeros could reduce the required polynomial degree for this specific starting matrix.
Loading
Loading