Advanced Statistical & Quantum-Specific Features

Welcome to Lesson 38 of the SNAP ADS Learning Hub! Lesson 37 introduced ADACL’s 394 physics-informed features at the family level — statistical (~120), quantum-specific (~140), frequency-domain (~70), and physics-constraint (~64). This lesson goes deeper into the two largest families: statistical and quantum-specific. By the end, you should know what concrete measurements those 260 features capture, why the simple ones are surprisingly powerful, and where the genuinely novel quantum-informed signal sits.

This is the most technically dense lesson in Module 7. We will reference specific moments of distributions, specific quantum quantities like T₁/T₂ coherence times and von Neumann entropy, and specific cross-talk metrics. If any of this is unfamiliar, the prior modules cover the foundations — Module 1 for quantum states and density matrices, Module 2 for Lindblad dynamics and coherence-time decay, Lesson 37 for the four-family overview.

A useful framing: a doctor’s “vital signs” panel is heart rate, blood pressure, temperature, oxygen saturation, respiration rate. Five simple numbers, and the entire field of medicine compresses an enormous amount of patient state into them. ADACL’s statistical features are the equivalent for a quantum system. The quantum-specific features are the extra panel of measurements only a quantum-domain specialist would think to take — like ordering a tropopause measurement for a lung patient because you happen to know they’re a high-altitude pilot.

The Statistical Family (~120 Features)

The statistical features are computed over rolling windows of each raw sensor stream. They are mostly classical summary statistics — but the specific combination and granularity is what matters.

Moment-based features (~40)

For each sensor stream and each window size:

• Mean — the baseline level. Useful for slow drifts in absolute value.
• Variance — the spread. A widening variance is one of the earliest signals of noise injection.
• Skewness — third moment. Asymmetric distributions signal non-Gaussian noise sources (two-level systems, cosmic-ray events on superconducting chips).
• Kurtosis — fourth moment. Heavy tails signal rare large excursions. Particularly diagnostic for cryogenic-system glitches that produce brief but extreme readings.
• Range — max − min over the window. Cheap to compute, surprisingly informative.
• IQR — interquartile range. A robust scale measure that resists single-point outliers.

The features are computed at multiple window sizes (e.g., 1-minute, 5-minute, 30-minute) so the network sees both fast and slow timescales of the same statistic. This multi-scale view is where deep CNNs can extract temporal patterns that single-scale features miss.

Autocorrelation-based features (~25)

Autocorrelation measures how correlated a signal is with a delayed copy of itself. Quantum systems have characteristic autocorrelation signatures — 1/f noise, narrow-band oscillations from poorly-shielded electronics, periodic drift from environmental cycles.

ADACL includes autocorrelation at multiple lags (1ms, 10ms, 100ms, 1s, 10s) for the most-informative sensor streams. A change in the lag-1ms autocorrelation often signals fast electronic noise; a change in the lag-1s autocorrelation often signals slower thermal drift. The two are uncorrelated; both belong in the feature set.

Quantile-based features (~25)

Quantiles capture the shape of the distribution beyond what moments reveal. ADACL tracks the 5th, 25th, 50th, 75th, and 95th percentiles per window. Why bother when we already have mean and variance? Because heavy-tailed distributions can have the same mean and variance as Gaussian distributions but vastly different 95th percentiles. The quantile features are the bridge to the kurtosis story.

Cross-stream features (~30)

Pairwise statistics between streams: Pearson correlation, mutual information, lagged cross-correlation. These catch failure modes that are invisible in any single stream. A common example: a control-pulse parameter and a measurement outcome that are normally uncorrelated suddenly correlating signals a calibration drift.

The cross-stream features are computed on the most operationally important pairs (e.g., neighbouring qubit measurements, control pulse vs. measurement) rather than all O(N²) pairs, both for compute cost and to avoid feature-explosion in the network.

The Quantum-Specific Family (~140 Features)

These are the features that distinguish ADACL from a generic anomaly detector. They only make sense in a quantum context, and they’re derived from the Lindblad master-equation framework introduced in Module 2.

Coherence-time features (~30)

T₁ and T₂ are the bread-and-butter of qubit characterisation. T₁ is the energy-relaxation time; T₂ is the dephasing time. Both are estimated per-qubit from short Ramsey or echo sequences interleaved with the main workload.

ADACL tracks not just point estimates but distributional features:

• T₂ rolling mean and std over windows of varying length.
• T₂ ratio to baseline — current estimate / baseline calibration value. Drops below 0.8 typically signal cause for concern.
• T₂ skewness across recent estimates — asymmetric distributions point at two-level-system coupling.
• T₁ – T₂ relationship — a healthy qubit has T₂ ≤ 2T₁. Deviations from this relationship are diagnostic of specific noise channels.

Density-matrix-derived features (~40)

For systems where state tomography (or DeCoN-PINN inference) gives an estimate of the density matrix ρ, ADACL derives:

• Trace distance to baseline ρ — ½‖ρ_now − ρ_baseline‖₁. The canonical quantum-state distance metric.
• Fidelity to baseline — (Tr √(√ρ_b · ρ · √ρ_b))². Complements trace distance with a smoother metric near the baseline.
• Von Neumann entropy — −Tr(ρ log ρ). Sensitive to changes in state purity; rising entropy often precedes decoherence events.
• Purity — Tr(ρ²). Inverse-related to entropy; convenient computationally and direct interpretation as “how mixed is this state.”

These four features are mathematically related, but each emphasises a different aspect of the state change; redundancy here is by design.

Gate-error features (~30)

For systems that run periodic gate-tomography or randomised-benchmarking sequences:

• Per-gate error rate — typical magnitudes 1e-4 to 1e-3 for modern superconducting hardware.
• Cross-talk per-pair — how much does activating gate A on qubit i affect qubit j.
• Leakage out of the computational subspace — a particularly destructive failure mode that’s invisible to most metrics.
• Drift in calibrated gate parameters — pulse amplitude, duration, phase.

These features are also the ones operators most often want to see in the explanation when ADACL raises an alert. The explanation layer (covered later) surfaces the top contributing features; gate-error features dominate that surfacing for hardware-fault alerts.

Readout-discrimination features (~20)

How well can the system tell the |0⟩ and |1⟩ states apart at measurement time?

• Readout contrast — fraction of correctly-discriminated states.
• Confusion-matrix off-diagonals — explicit |0⟩→|1⟩ and |1⟩→|0⟩ misclassification rates.
• IQ-plane separability — for systems exposing the raw IQ readout, the cluster separation in IQ space.

Readout failures degrade everything downstream; these features catch them upstream.

Multi-qubit features (~20)

Multi-qubit quantities that don’t reduce to per-qubit measures: Bell-state fidelity for adjacent pairs, multi-qubit entropy, residual classical correlation. These are the features that catch correlated-noise failure modes — the hardest to diagnose and the most damaging when they appear.

Why So Many?

A natural question: surely a few of these features dominate, and the rest are redundant? Empirically, no. The deviation function (Lesson 33) routinely uses 30-50 features as primary contributors across different anomaly types. Coherence-time anomalies surface through different features than cross-talk anomalies than readout-degradation anomalies. Pruning the feature set to “the important ones” works only if you know in advance which anomaly you’re looking for, which defeats the purpose of generic anomaly detection.

The 394-feature design is deliberate over-completeness: include every measurement that has a clear physics motivation, let the CNN learn which to weight for which kind of anomaly, and use the explanation layer to surface the top contributors at runtime.

Feature Computation Cost

A practical concern: 394 features × 10 streams × multiple window sizes is non-trivial compute. ADACL’s feature pipeline is structured so that:

• The cheap features (moments, quantiles) are computed every window.
• The expensive features (state tomography, randomised benchmarking) are computed on a slower schedule and interpolated.
• A small subset (the “tier-1” features) are computed continuously and trigger expensive re-computation only when they themselves exceed a coarse threshold.

The compute cost is well under 1 % of system overhead even on heavily-loaded production quantum hardware.

Key Takeaways

• Understanding the fundamental concepts: ADACL’s 260 statistical + quantum-specific features span moment-based summaries at multiple time scales, autocorrelations at multiple lags, quantile distributions, cross-stream correlations, coherence-time descriptors, density-matrix-derived quantities, gate-error metrics, readout discrimination, and multi-qubit entanglement.
• Practical applications in quantum computing: Each family catches a different class of failure mode. Coherence-time features catch decoherence drift; cross-talk features catch correlated noise; readout features catch measurement degradation. Pruning the feature set to “just the important ones” breaks generic detection.
• Connection to the broader SNAP ADS framework: These features are the substrate the CNN regressor (Lesson 35) reasons over and the explanation layer surfaces to operators. Every alert ADACL raises ultimately points back at one or more of these features.

What’s Next?

Lesson 39 covers ADACL’s anomaly detection and scoring module — how the trained CNN’s outputs are turned into operator-facing alerts, how the scoring relates to adaptive baselines, and how the calibration is maintained over time. Lesson 40 then covers production optimisation: making the entire pipeline fast, robust, and resource-efficient at deployment scale.

---

Ready to continue? Use the navigation buttons below to move to the next lesson or return to the module overview.