Saltar al contenido principal

quant

Since v2.19.0 · Used inside a declaration

Grammar​

# Flow-body block. The attribute header is OPTIONAL and goes in PARENS;
# the braces hold real flow steps (let / for / yield), like `par`.
quant(encoding: amplitude,        # `amplitude` (default) | `angle`
      observable: <ObservableName>, # the Hermitian operator to measure
      qubits: <n>, depth: <d>,    # all optional
      bandwidth: <γ>, reupload: <L>, # `reupload: L≥2` interleaves the
      backend: quant_sim) { #   data encoding L times (v2.23.0)
    let surrogate = <continuous-carrier>   # bind the carrier (a Tensor)
    yield surrogate                        # collapse → the ⟨observable⟩ expectation
}

# Bare form (every attribute defaulted: encoding=amplitude, backend=quant_sim):
quant { let s = carrier  yield s }

quant is AXON's bridge between sub-symbolic embeddings and the algebra of quantum-kernel methods. Inside a flow, it lifts a continuous carrier tensor into a finite-dimensional Hilbert space, optionally evolves it under a variational circuit, and collapses it back to classical silicon by measuring a declared observable — yielding a single real expectation a downstream step can consume.

It is a cognitive primitive: the quantum machinery is a means to a geometry (the projected-kernel route to a provable convex advantage), not an end. The honest claim is convexity — a valid quantum kernel Gram is PSD, so the downstream classical SVM dual has a global optimum — not "tunneling through barriers."

Charter split (free syntax / paid scale)​

The keyword, the static rules, and a usable CPU reference simulator ship in OSS axon-lang. That simulator is hard-capped at n ≤ 10 qubits (axon-E0783 past that). The efficient execution substrate — the Q32.32 bit-exact arithmetic, the QuIDD decision-diagram compression for n ≫ 10, per-tenant VRAM control, and locked hardware / QPU-native backends — is Axon Enterprise only. The standard is unified; the scale is the paid privilege.

Surface​

quant is a flow-body block (nested, like transact or forge). The optional attribute header is in parentheses; the braces hold real flow steps. This exact program passes axon check (0 errors):

observable Energy {
    qubits: 1
    term: 1.0 * "Z"
}

flow Classify(embedding: Tensor) -> String {
    quant(encoding: amplitude, observable: Energy, qubits: 1) {
        let surrogate = embedding   // bind the continuous carrier
        yield surrogate             // collapse → ⟨ψ(embedding)| Energy |ψ⟩
    }
    return "done"
}

Nota de gramática (la forma que compila): los atributos van en quant( … ) (paréntesis), la clave es encoding (no encode), y las llaves { } contienen pasos de flow (let / for / yield). El yield toma una referencia (un let o un parámetro) — NO usa los brackets unicode ⟨⟩ (no lexean). El carrier debe ser un tipo continuo (Tensor), si no salta axon-E0782.

Anatomy​

encoding: — the lift (header attribute, in ( ))​

  • amplitude (default) — the carrier becomes the state's amplitude vector (must be unit-norm; the runtime asserts ‖x‖₂ = 1). n = ⌈log₂ d⌉ qubits for a length-d carrier.
  • angle — each carrier component drives a rotation angle (one qubit per component). Resists the amplitude form's normalization constraint.

The other header attributes (all optional, order-free, in the parens): observable:, qubits:, depth:, bandwidth:, reupload:, backend:.

reupload: — data re-uploading layers (header attribute, v2.23.0)​

reupload: L interleaves the carrier encoding L times through the circuit instead of once. L = 1 (the default when omitted) is plain single-shot encoding; L ≥ 2 is the only provable escape from the amplitude+Pauli quadratic kernel bound — each re-upload lets the learned feature map reach frequencies a single encoding cannot. L < 1 is axon-E0784. This is the lever an adopter pulls when the single-encoding feature map is too low-frequency to separate their data:

flow Reupload(embedding: Tensor) -> String {
    quant(encoding: angle, reupload: 3, qubits: 2) {
        let surrogate = embedding
        yield surrogate
    }
    return "done"
}

observable: — the measurement (header attribute)​

Resolves (closed-catalogue, axon-E0784) to a declared observable Pauli-sum. Its width fixes the qubit count n.

yield <reference> — the collapse (a step in the body)​

yield <reference> is a step inside the quant braces, only legal there (axon-E0787 otherwise). The reference is a let-bound name or a flow parameter (the carrier being measured) — it reuses the let-value grammar, so there are no ⟨⟩ brackets. It emits the measured expectation ⟨ψ|M|ψ⟩ back into the classical flow.

Una expectativa = un Float (feature map). Cada bloque quant produce UNA expectativa escalar de UN observable. Para un projected / seed kernel se ensambla clásico: declarás k observables, hacés yield de cada uno → φ(x) = [⟨M₁⟩, …, ⟨Mₖ⟩], y k(x,y) = sim(φ(x), φ(y)). La navegación estructural (p.ej. signed-EPR) no se toca; quant solo puntúa el seed.

Runtime behaviour​

quant lowers to a QuantBlock IR node. At execution:

  1. Encode the carrier into a state vector under the chosen scheme.
  2. Evolve (optional variational circuit; gate application).
  3. Measure the observable: ⟨ψ| M |ψ⟩, real by Hermiticity.
  4. Yield the expectation as a classical value.

In the enterprise backend the whole path is bit-reproducible on the Q32.32 substrate (Pauli measurement is exact), audited (quant:started → measured → completed, the raw carrier never written to the chain — only its SHA-256 digest), RBAC-gated (quant:execute), shielded, and VRAM-quota'd per tenant.

Static guarantees​

  • axon-E0782 — Continuous Type Invariant: the carrier must be a continuous tensor (a discrete value is rejected).
  • axon-E0783 — capacity: n > 10 on the OSS simulator is a compile error (enterprise lifts it).
  • axon-E0784 — header validity: the observable: must resolve to a declared observable, and qubits/depth/bandwidth/reupload must be in range (reupload >= 1, v2.23.0).
  • axon-E0787 — yield outside a quant block is rejected.
  • axon-W005 — a circuit-depth advisory (soft barren-plateau note).

What this primitive is NOT​

  • Not a tool. A tool binds an external capability; quant is an in-language transform over a Hilbert space.
  • Not a compute. Compute selects an LLM backend + effort; quant performs a quantum-kernel measurement.
  • Not infinite-precision. The substrate is deterministic (bit-reproducible) but quantized at Δ = 2⁻³² — determinism is not exactness.
  • Not a tunneling optimizer. The advantage is the convexity of the kernel-SVM dual, validated by the geometric-difference witness — a potential-advantage signal, necessary but not sufficient.

See also​

  • axon://primitives/observable — the Hermitian operator a quant block measures.
  • axon://primitives/flow — the parent of every quant block.
  • axon://primitives/compute — backend selection (a different axis).