Skip to content
Merged
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
57 changes: 57 additions & 0 deletions .github/workflows/verify.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,57 @@
name: Verify science and hosted mathematics

on:
pull_request:
workflow_dispatch:

permissions:
contents: read

jobs:
verify:
runs-on: ubuntu-24.04
timeout-minutes: 20
env:
SOURCE_REVISION: ${{ github.event.pull_request.head.sha || github.sha }}
OPENBLAS_NUM_THREADS: 1
OMP_NUM_THREADS: 1
steps:
- uses: actions/checkout@v4
with:
ref: ${{ github.event.pull_request.head.sha || github.sha }}
fetch-depth: 0
persist-credentials: false
- uses: actions/setup-python@v5
with:
python-version: '3.12'
- uses: actions/setup-node@v4
with:
node-version: '20'
- name: Reproduce finite scientific checks
run: |
python -m pip install -r verification/requirements.txt
python verification/check_reference_identification.py > science-results.json
- name: Install browser and source compiler
run: |
npm install --prefix "$RUNNER_TEMP/qri-hosted-deps" --no-audit --no-fund playwright@1.56.1 mathjax-full@3.2.1
"$RUNNER_TEMP/qri-hosted-deps/node_modules/.bin/playwright" install --with-deps chromium
- name: Capture actual pinned GitHub pages at two widths
env:
NODE_PATH: ${{ runner.temp }}/qri-hosted-deps/node_modules
run: |
node verification/hosted/capture_hosted_math.cjs \
--root "$GITHUB_WORKSPACE" \
--repository "$GITHUB_REPOSITORY" \
--revision "$SOURCE_REVISION" \
--contract verification/hosted/math-contract.json \
--output "$GITHUB_WORKSPACE/hosted-evidence"
- name: Retain verification evidence
if: always()
uses: actions/upload-artifact@v4
with:
name: qri-verification-${{ env.SOURCE_REVISION }}
path: |
science-results.json
hosted-evidence/
if-no-files-found: error
retention-days: 14
2 changes: 2 additions & 0 deletions .gitignore
Original file line number Diff line number Diff line change
@@ -1,3 +1,5 @@
__pycache__/
*.py[cod]
verification/results.new.json
/science-results.json
/hosted-evidence/
29 changes: 13 additions & 16 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,29 +2,26 @@

How much of a quantum reference library survives measurement before a future query is known? This project studies identification from supplied copies: retain the references as quantum states, or measure the entire library first and keep an unrestricted classical record.

A library contains $M$ independent Haar-distributed pure states in dimension $d$. The input consists of $m$ labeled copies of every reference and $t$ copies of a uniformly selected library member. The output is its label. No classical reference descriptions, preparation or inverse oracles, or QRAM are supplied.
A library contains $`M`$ independent Haar-distributed pure states in dimension $`d`$. The input consists of $`m`$ labeled copies of every reference and $`t`$ copies of a uniformly selected library member. The output is its label. No classical reference descriptions, preparation or inverse oracles, or QRAM are supplied.

Write

$$
```math
D_s=\binom{d+s-1}{s},\qquad C=\frac{D_mD_t}{D_{m+t}}.
$$

For $M\ge 2$, the optimal average error obeys
```
For $`M\ge 2`$, the optimal average error obeys

$$
```math
\frac{M-1}{8C+4(M-2)}\le e_{\mathrm{opt}}\le\frac{M-1}{C+M-1}.
$$

Sequential symmetric-subspace tests give an explicit decoder with error at most $2(M-1)/C$. Reliable identification therefore requires and permits $C$ of order $M/\epsilon$ for small target error $\epsilon$, up to universal constants.
```
Sequential symmetric-subspace tests give an explicit decoder with error at most $`2(M-1)/C`$. Reliable identification therefore requires and permits $`C`$ of order $`M/\epsilon`$ for small target error $`\epsilon`$, up to universal constants.

The classical comparison allows **any collective measurement of all references**, unlimited computation and classical storage, but no quantum information from training may survive until the query arrives. Even revealing the exact later query state does not remove the fixed-reference ceiling. For each fixed positive integer $m$, as $M\to\infty$ and $d/\log M\to\infty$,
The classical comparison allows **any collective measurement of all references**, unlimited computation and classical storage, but no quantum information from training may survive until the query arrives. Even revealing the exact later query state does not remove the fixed-reference ceiling. For each fixed positive integer $`m`$, as $`M\to\infty`$ and $`d/\log M\to\infty`$,

$$
```math
P_{\mathrm{classical}}^{\mathrm{revealed}}
=(1+o(1))\frac{(\log M)^m}{m!M}.
$$

```
The logarithm is natural. The ceiling is attained to leading order by independent covariant reference measurements; its converse includes collective training. These statements quantify loss across a delayed-query memory boundary, with a fixed supplied reference budget.

| Question | Authoritative account |
Expand All @@ -38,7 +35,7 @@ The logarithm is natural. The ceiling is attained to leading order by independen
| Which ingredients and results are inherited? | [Related work](references/RELATED_WORK.md) |
| How can the finite identities be reproduced? | [Verification](verification/README.md) |

Start with the [connected science route](science/README.md). The [LLM reading guide](llms.txt) maps questions directly to proofs.
Start with the [connected science route](science/README.md), or follow the [three-qubit walkthrough](learning/WALKTHROUGH.md). The [learning route](learning/README.md) connects the selected Watrous anchor to a [30-milestone map](learning/MILESTONES.md) and the full proofs. The [LLM reading guide](llms.txt) maps questions directly to authoritative sources.

## Reproduce

Expand All @@ -47,11 +44,11 @@ python3 -m pip install -r verification/requirements.txt
OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 python3 verification/check_reference_identification.py
```

The verification guide states the tested instances, numerical tolerances and limits. Dense finite checks supplement the analytical proofs; they do not establish an unrestricted optimal measurement numerically.
The verification guide states the tested instances, numerical tolerances and limits. The [hosted-rendering check](verification/hosted/README.md) captures commit-pinned mathematical pages at wide and narrow widths. Dense finite checks supplement the analytical proofs; they do not establish an unrestricted optimal measurement numerically.

## Scope

The results are Bayesian for an independent Haar pure-state library, a uniform label, equal reference-copy counts, and one lookup. They do not establish a worst-case arbitrary-library converse, mixed-state robustness, repeatable database use, or a total-runtime speedup. The decoder scans references, consumes real quantum storage, and charges all $Mm+t$ supplied specimens. Classical shadows already provide a logarithmic reference-copy comparator for a later known query; the fixed-$m$ success separation is not an exponential reference-sample advantage.
The results are Bayesian for an independent Haar pure-state library, a uniform label, equal reference-copy counts, and one lookup. They do not establish a worst-case arbitrary-library converse, mixed-state robustness, repeatable database use, or a total-runtime speedup. The decoder scans references, consumes real quantum storage, and charges all $`Mm+t`$ supplied specimens. Classical shadows already provide a logarithmic reference-copy comparator for a later known query; the fixed $`m`$ success separation is not an exponential reference-sample advantage.

## Purpose and contact

Expand Down
88 changes: 88 additions & 0 deletions learning/ANCHOR_AND_BRIDGE.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,88 @@
# Educational anchor and bridge

The sole educational anchor is John Watrous, *The Theory of Quantum Information* (Cambridge University Press, 2018). Use the [author’s book page](https://cs.uwaterloo.ca/~watrous/TQI/) and its linked manuscript; the section and equation numbers below refer to the author’s standard-layout 2018 PDF. The book supplies the mathematical language. The linked project proofs supply the identification model and its evaluated bounds. Reading the whole book is unnecessary.

## A targeted route through the anchor

| Project prerequisite | Verified Watrous pointer | Continue here |
| --- | --- | --- |
| Density operators, mixtures, pure and flat states | [Chapter 2](https://cs.uwaterloo.ca/~watrous/TQI/TQI.2.pdf), §2.1.2; Eqs. (2.12), (2.14) | [Model](../science/model.md) |
| Partial trace and joint registers | Chapter 2, §2.1.3 | [Collective-training posterior](../science/classical_preprocessing.md#a-collective-training-measurement-still-has-positive-one-label-posteriors) |
| POVMs and outcome probabilities | Chapter 2, §2.3.1 | [Identification objective](../science/quantum_identification.md) |
| Projective and nondestructive measurements | Chapter 2, §2.3.2 | [Physical symmetry instrument](../science/resources.md) |
| Binary discrimination and the trace norm | [Chapter 3](https://cs.uwaterloo.ca/~watrous/TQI/TQI.3.pdf), §3.1.1, Theorem 3.4 | [Worked qubit example](WALKTHROUGH.md) |
| Ensemble discrimination and the PGM | Chapter 3, §3.1.2; Eqs. (3.49)–(3.52) | [PGM proof](../science/quantum_identification.md#pretty-good-measurement) |
| Permutation projector and symmetric dimension | [Chapter 7](https://cs.uwaterloo.ca/~watrous/TQI/TQI.7.pdf), §7.1.1; Proposition 7.1, Corollary 7.3 | [Averaged hypotheses](../science/quantum_identification.md#haar-averaging-gives-flat-projector-hypotheses) |
| Uniform pure-state measure and its tensor moments | Chapter 7, §7.2.1; Lemma 7.24, Eq. (7.159) | [Haar averaging](../science/model.md#symmetric-spaces) and [posterior moments](../science/classical_preprocessing.md#finite-dimensional-revealed-query-converse) |

The pure-copy integral is in **§7.2.1**, before §7.2.2. Section 7.3 concentration machinery is not required for the project’s fixed $`m`$ asymptotic proof.

## Translating notation

| Watrous convention | This project |
| --- | --- |
| Unit vector $`u`$; associated pure density operator $`uu^*`$ | $`\lvert\psi\rangle`$; often the same symbol $`\psi`$ also denotes $`\lvert\psi\rangle\langle\psi\rvert`$ |
| Adjoint $`A^*`$ and identity $`1_{\mathcal X}`$ | $`A^\dagger`$ and $`I_{\mathcal X}`$ |
| POVM value $`\mu(j)`$ | Measurement effect $`E_j`$ |
| Subnormalized ensemble value $`\eta(j)`$ | $`\rho_j/M`$, because labels are uniform |
| Symmetric subspace on $`s`$ tensor factors | $`\mathcal S_s=\mathrm{Sym}^s(\mathbb C^d)`$, with dimension $`D_s`$ |

The book’s average ensemble state is $`\sum_j\eta(j)=S/(Mr)`$ here, where $`S=\sum_jP_j`$ and $`\rho_j=P_j/r`$. Substituting this into its PGM formula cancels the scalar factors and gives $`E_j=S^{-1/2}P_jS^{-1/2}`$. All inverse square roots are restricted to their support; assigning zero-probability kernel directions completes the measurement.

## Bridge 1: averaging does not provide descriptions

The physical input is an unknown reference library and an unknown matching query. The algorithm cannot insert the unknown $`\psi_j`$ into a measurement formula. Nevertheless, averaging success over the specified prior is linear in the input density operator. Under label $`j`$, the query and that reference are copies of the same state, so their average is $`\Pi_{m+t}/D_{m+t}`$. Every other reference independently averages to $`\tau_m=I_{\mathcal S_m}/D_m`$.

Their product is $`\rho_j=P_j/r`$ with $`r=D_{m+t}D_m^{M-1}`$. Thus the unknown-library problem becomes discrimination of a known **averaged** ensemble. This is a way to evaluate a universal measurement, not an oracle that reveals the sampled library. The ratio

```math
C=\frac{\dim\mathcal H}{r}
=\frac{D_mD_t}{D_{m+t}}
```
measures how concentrated each averaged hypothesis is relative to the common uniform state on $`\mathcal H`$.

## Bridge 2: a collective record still has a positive posterior

The earlier training POVM may couple every reference block. To analyze one selected label after outcome $`y`$, integrate all other references out. Their priors are scalar identities on their symmetric supports. Partial tracing the positive training effect therefore gives a positive one-block effect $`A_{y,j}`$; normalization gives a density operator $`a_{y,j}`$.

The label remains uniform because it was chosen independently of training. Its later-query likelihood is

```math
q_{y,j}(\psi)
=D_m\langle\psi^{\otimes m}|a_{y,j}|\psi^{\otimes m}\rangle.
```
Conditioned on $`y`$, the query itself need not have the Haar distribution. Its density relative to Haar is $`\sum_jq_{y,j}/M`$, while its label posterior is $`q_{y,j}/\sum_iq_{y,i}`$. Multiplying the winning posterior by that query density cancels the denominator. The conditional success is exactly $`M^{-1}\int\max_jq_{y,j}(\psi)\,d\psi`$.

Now take a positive integer moment $`s`$. Haar integration places $`ms`$ copies inside one symmetric projector:

```math
\int q_{y,j}^s\,d\psi
=\frac{D_m^s}{D_{ms}}
\mathrm{Tr}(a_{y,j}^{\otimes s}\Pi_{ms})
\leq\frac{D_m^s}{D_{ms}}.
```
The last step uses only positivity, trace one and $`0\leq\Pi_{ms}\leq I`$. Consequently it covers collective training, not merely independent estimation. Bounding the maximum by an $`s`$-norm, then choosing $`s`$ of order $`\log M/m`$, produces the sharp fixed $`m`$ ceiling. The [classical proof](../science/classical_preprocessing.md) supplies both that limiting calculation and the covariant strategy attaining it.

## Bridge 3: why the quantum converse needs three moments

For the retained references, exact pair overlaps determine a variance-like quantity $`\beta=(M-1)/C`$. Pair overlaps alone could all come from one shared ambiguous subspace, so they do not force the desired accumulation of error.

Concatenate bases of the hypothesis subspaces and form the Gram matrix $`G`$. Its dimension is $`n=Mr`$; include all its zero eigenvalues. The first three exact moments give

```math
\mathbb E\lambda=1,\qquad
\mathbb E(\lambda-1)^2=\beta,\qquad
\mathbb E[\lambda(\lambda-1)^2]=\beta+\gamma,
```
where $`\gamma=(M-1)(M-2)/C^2`$. The inherited discrimination trace bound turns success into $`\mathbb E\sqrt\lambda`$, so error is at least $`\frac12\mathbb E(\sqrt\lambda-1)^2`$. Cauchy–Schwarz converts the two controlled variance expressions into

```math
e_{\mathrm{opt}}\geq\frac{\beta^2}{4(2\beta+\gamma)}.
```
The [full proof](../science/quantum_identification.md#error-converse-from-three-moments) gives the polar-decomposition step and constants. Watrous provides the operator and discrimination foundations; this evaluated argument is supplied locally.

## Research ingredients beyond the anchor

The [related-work account](../references/RELATED_WORK.md) credits Tyson’s generalized Holevo–Curlander bound, Gao’s sequential union bound and the inherited measure-and-prepare decomposition. These are primary research dependencies, not additional educational anchors. The [resource account](../science/resources.md) supplies the explicit symmetry instrument and its charged finite precision. Neither a textbook discrimination formula nor a POVM existence theorem supplies those gate and disturbance guarantees automatically.

Use the [worked example](WALKTHROUGH.md) for a complete small instance, then return to the [scientific route](../science/README.md). The common milestone map organizes understanding; these pages record no individual reader’s assessment or progress.
Loading
Loading