An obfuscator rewrites a program so that its code reveals nothing beyond what the program computes. The only workable version of this idea, indistinguishability obfuscation (iO), makes a narrower promise. If two programs compute exactly the same function, nobody can tell which one was obfuscated. That promise depends on a question cryptographers rarely ask out loud: for the class of programs in hand, can we even decide whether two programs compute the same function? For one class this question has no answer yet. These are continuous-variable (CV) quantum programs, which act on bosonic modes such as light in a fiber or microwave photons in a cavity, where each mode has an infinite-dimensional state space. Equivalence checkers exist for qubit circuits. Quantum obfuscation has been defined for qubit programs. Bosonic programs are covered by neither.
This note proposes to settle the equivalence half of the question exactly for a natural hierarchy of bosonic circuit families, and to measure what that answer buys for obfuscation. The approach tracks how a circuit transforms the position and momentum operators of each mode. It does not simulate states, which would require an arbitrary photon-number cutoff. For Gaussian gates plus a bounded number of cubic-phase gates, this transformation is a finite polynomial with a unique normal form. Equivalence is therefore decidable, and the normal form is a canonical description, which by The Theoretical Boundaries of Program Obfuscation is all that perfect indistinguishability obfuscation of the circuit description requires.
The quantum part of a program, the state it carries with it, is different. Hiding it requires a randomizing ensemble, and the proof that standard CV 2-designs do not exist (Continuous-Variable Quantum State Designs: Theory and Applications) implies that any physical, finite-energy randomizer must leak. We propose to measure how that leakage falls as the energy budget grows.
Core Claim
The decidability frontier for exact equivalence of bosonic circuits falls in a definite place. Below it sit Gaussian circuits, and Gaussian circuits with a bounded number of cubic-phase layers (a cubic-phase gate is the non-Gaussian operation $e^{i\gamma \hat x^3}$). Above it sit circuits containing the Kerr gate $e^{i\kappa \hat n^2}$, where $\hat n$ is the photon-number operator. Kerr gates map the mode's lowering operator to a non-polynomial function of $\hat n$.
The reason the lower fragment is decidable is as follows. A unitary $U$ on $n$ modes is determined, up to a global phase, by how it conjugates the quadrature operators:
$$U^\dagger \hat x_j U = P_j(\hat x,\hat p), \qquad U^\dagger \hat p_j U = Q_j(\hat x,\hat p), \quad j=1,\dots,n,$$
where $\hat x_j,\hat p_j$ are the position and momentum quadratures of mode $j$, and $P_j, Q_j$ are polynomials in the non-commuting quadratures, written in a fixed (Weyl-symmetric) operator ordering. The "determined" part follows from the irreducibility of the canonical commutation relations. Gaussian gates make $P_j,Q_j$ affine. A cubic-phase gate sends $\hat p \mapsto \hat p + 3\gamma \hat x^2$, so substituting degree-$d$ polynomials yields degree at most $2d$. After $k$ non-Gaussian layers the degree is therefore at most $2^k$. Two circuits are equivalent exactly when their normal-ordered polynomials coincide. With symbolic gate parameters, the coefficients become polynomials in the parameters, and equivalence for all settings becomes a polynomial identity.
This yields a decision procedure whose cost is polynomial in the number of modes for fixed non-Gaussian depth and exponential in that depth. That is the same shape Scalable Equivalence Checking and Verification of Shallow Quantum Circuits found for qubit circuits, which is linear in qubit count and exponential in depth.
The second claim concerns the state a program carries. Every perfect hiding argument for quantum advice states uses a twirl, meaning an average over random unitaries that wipes out all information about which state was used. The qubit construction in How to Use Quantum Indistinguishability Obfuscation uses a Clifford one-time pad for this. In CV systems, the natural twirls are averages over displacements or symplectic transformations, and these are non-normalizable. Their physical, energy-capped versions are the regularized rigged designs of Continuous-Variable Quantum State Designs: Theory and Applications and the random GKP lattice ensembles of Continuous-variable designs and design-based shadow tomography from random lattices. We claim that the best achievable distinguishing advantage $\varepsilon(E)$ between two twirled advice states is strictly positive at any finite mean photon number $E$. We will measure its decay with $E$ and test whether random-lattice ensembles achieve it.
Relation to Prior Work
Two qubit equivalence checkers supply the method template and constrain what counts as progress.
Equivalence Checking of Parameterised Quantum Circuits decides equivalence of parameterized qubit circuits for all parameter values. It does this with decision diagrams whose edge weights are trigonometric polynomials, and it cannot handle non-trigonometric dependencies or infinite-dimensional modes. Our Weyl-polynomial normal form plays the role its trigonometric decision diagrams play, but for bosonic operators. Because we work on operators in the Heisenberg picture, no state-space representation, and hence no photon cutoff, is ever needed.
Scalable Equivalence Checking and Verification of Shallow Quantum Circuits decides equivalence of constant-depth qubit circuits using light-cone local constraints. We borrow its framing: the hard parameter is depth, not width. Here, though, the relevant depth is the number of non-Gaussian layers rather than total depth.
The broader decidability literature explains why a frontier should exist at all. On Decidable and Undecidable Extensions of Simply Typed Lambda Calculus shows that adding single features such as references or exceptions flips contextual equivalence to undecidable. Our Kerr boundary is the bosonic analogue of such a feature. We do not claim Kerr circuits are undecidable; we claim our representation stops being finite there, and we will map where it stops.
The obfuscation side supplies the reduction and the motivating gaps. The Theoretical Boundaries of Program Obfuscation shows that efficient canonical forms give indistinguishability obfuscation. On Low-End Obfuscation and Learning studies obfuscation for restricted circuit classes, where general iO's overhead can be avoided. Our fragment is a concrete restricted class of this kind. The open question is whether its canonical form is efficient, and the doubling degree bound suggests it is efficient only for logarithmically many non-Gaussian layers.
How to Use Quantum Indistinguishability Obfuscation defines quantum obfuscation for qubit programs. It assumes the equivalence promise without supplying a way to check it, and it leaves continuous systems open. We do not attempt a standard-model quantum obfuscator, which remains open even for qubits. Instead, we supply the missing equivalence check for bosonic programs and quantify the energy price of the twirling step.
Two backdoor papers motivate why the frontier matters for machine learning. Backdoor defense, learnability and obfuscation shows that iO makes general Boolean circuits undefendable against backdoors, while restricted classes admit defenses. Injecting Undetectable Backdoors in Obfuscated Neural Networks and Language Models obtains undetectability only by digitizing continuous models into Boolean circuits. A fragment with a cheap canonical form is one where this kind of implementation-level hiding is unavailable, because the canonical form exposes the function directly. Our hierarchy marks where that guarantee holds for bosonic models.
Research Program
The first component is theoretical. We will prove that the Heisenberg-picture normal form is complete and sound for equivalence up to global phase on the Gaussian-plus-cubic fragment. We will prove the degree and term-count bounds as functions of mode number and non-Gaussian depth, and give a matching lower-bound family showing the exponential blow-up in depth is real.
We will then try to push the frontier upward by adding gates whose Heisenberg action stays polynomial, such as higher-order phase gates $e^{i\gamma \hat x^m}$. For Kerr gates, we will test whether working in a photon-number-diagonal representation recovers decidability when the input energy is bounded. If it does, the decidability claim becomes energy-relative, which would directly mirror the physicality conjecture of Continuous-variable designs and design-based shadow tomography from random lattices.
The second component is a tool. It will be a symbolic engine that represents gate parameters as symbols, composes Heisenberg maps gate by gate, normal-orders the result, and compares coefficients. Coefficient comparison has two interchangeable back ends. One is exact, using canonical polynomial normal forms over parameter symbols. The other is randomized, using Schwartz–Zippel evaluation in the style of Global value numbering using random interpretation. The second component also includes a canonicalizer that outputs the normal form as the obfuscated description, with measured size against circuit size.
The third component is the hiding tradeoff. For one and two modes, where exact numerics remain trustworthy, we will compute the optimal distinguishing advantage between pairs of functionally equivalent advice states. Each pair will be passed through three energy-capped twirling ensembles: energy-capped displacement averages, regularized rigged 2-designs from Continuous-Variable Quantum State Designs: Theory and Applications, and finite GKP random-lattice ensembles from Continuous-variable designs and design-based shadow tomography from random lattices. We will fit the decay of $\varepsilon(E)$ against mean energy $E$. We do not presuppose an exponent; it is a measured quantity. Analytically, we target a lower bound on $\varepsilon(E)$ derived from the design non-existence argument.
Baselines
The naive reference point is Fock truncation with numerical comparison. Each circuit is simulated as a matrix on photon numbers up to a cutoff $D$ per mode, and the two truncated unitaries are compared up to phase. It is run on the same circuits and hardware, with the cutoff swept upward until memory runs out. To beat it, our tool must return correct verdicts at mode counts where truncation exhausts memory. It must also certify equivalence in cases where truncation reports spurious differences, which arise because cubic-phase gates leak amplitude across the cutoff.
The strongest existing method has no direct CV counterpart. No reported result was found on Lacuna for exact equivalence checking of continuous-variable circuits. The nearest exact competitor is the parameterized-circuit checker of Equivalence Checking of Parameterised Quantum Circuits. It reports verifying a 16-qubit RealAmplitudes benchmark in roughly 10.7 seconds against 44.3 seconds for the ZX-calculus tool QCEC, and handles circuits up to 65 qubits. For the shallow-circuit regime, Scalable Equivalence Checking and Verification of Shallow Quantum Circuits reports 4.46 seconds for weak equivalence of 100-qubit depth-3 circuits.
We will run that checker on GKP-encoded qubit versions of our test circuits, which are the subset with a faithful qubit encoding, using the same machine. The goal is not to beat qubit numbers on qubit problems. The goal is to show that our tool, run on the native bosonic description, reaches a comparable seconds scale at comparable widths, while the qubit route covers only the encodable subset.
A third baseline is numerical instantiation testing. It samples random parameter settings and input states, given an equal wall-clock budget. Beating it means detecting inequivalences whose difference is confined to a measure-zero or low-probability set of parameter values. Such cases will be constructed deliberately.
The key ablation swaps our exact coefficient back end for the randomized one while keeping the Heisenberg representation fixed. This isolates whether the gain comes from the representation or from exactness.
For the hiding experiments, the reference is the unrandomized advantage, meaning no twirl at all. The strongest competitor is the energy-capped displacement average, which is the obvious CV analogue of a one-time pad.
Why not just…?
Why not just truncate to a finite photon number and use qubit tools? Truncation works well for one or two modes at low energy. We expect it to tie or win there, and we will report that honestly. Its failure is structural. Memory grows as $D^n$, and the qubit simulators cited in Scalable Equivalence Checking and Verification of Shallow Quantum Circuits hit a wall near 30 to 34 qubits. More importantly, a truncated cubic-phase gate is not unitary, so truncation can never certify equivalence, only fail to find a difference. The mode-scaling experiment below settles where the crossover lies.
Why not just test at random parameters? Random instantiation is cheap and usually catches gross bugs, which is exactly how Equivalence Checking of Parameterised Quantum Circuits frames its need for symbolic methods. Passing one parameter setting proves nothing about the others. For obfuscation the gap is fatal. Backdoor defense, learnability and obfuscation shows that sampling catches random triggers but not adversarially placed ones, and an equivalence promise broken on a thin parameter set is precisely an adversarial trigger. Our planted-inequivalence suite measures this directly.
Why not just use randomized polynomial identity testing? This is the most serious alternative, and we adopt rather than fight it. Once circuits are written as Heisenberg polynomials, Schwartz–Zippel testing (Global value numbering using random interpretation) is a valid back end with tunable, exponentially small error. What it cannot do alone is supply the representation: applied to circuits it would treat bosonic gates as uninterpreted symbols and miss physical identities. It also does not output a canonical form, so it cannot serve as the obfuscator. The ablation will show whether exactness costs anything in runtime. If it costs a lot, the randomized back end becomes the default verifier and the exact one is kept only for canonicalization.
Proposed Experiments
No bosonic equivalence benchmark exists in the file, so we will build one. Its structure mirrors the protocol of Equivalence Checking of Parameterised Quantum Circuits: ansatz templates, rewritten equivalents, and injected errors. The ansatz templates are layered beam-splitter, squeezing, and displacement networks interleaved with cubic-phase layers. Rewritten equivalents are produced by known Gaussian identities, such as Bloch–Messiah decompositions and commuting displacements through symplectic maps, and by cubic-gate commutation rules. Injected errors include parameter perturbations and, separately, errors confined to special parameter values. The suite will sweep from 1 to 100 modes and from 0 to 6 non-Gaussian layers, with symbolic parameters throughout. The variational families in VeriQBench: A Benchmark for Multiple Types of Quantum Circuits will serve as the qubit-side templates for the GKP-encoded comparison.
The metrics are runtime, peak memory, verdict accuracy against ground truth known by construction, and normal-form size against gate count.
The first decisive check tests the claimed scaling. For fixed non-Gaussian depth, runtime should grow polynomially in mode number. For fixed mode number, it should grow exponentially in non-Gaussian depth. If depth-3 instances at 50 modes do not finish within minutes, the claim that this fragment is practically decidable is wrong.
The second decisive check is the crossover against truncation, and whether truncation's false "inequivalent" verdicts appear where predicted.
The third decisive check concerns hiding. Here $\varepsilon(E)$ must remain bounded away from zero at every finite energy for all three twirling ensembles. If some finite-energy ensemble achieved zero advantage, it would contradict the design no-go theorem as we are applying it, and that would itself be the finding. Whether finite GKP random-lattice ensembles match the displacement average at equal energy also bears on the open question in Continuous-variable designs and design-based shadow tomography from random lattices of whether finite lattice subsets can act as approximate designs.
References
- On Best-Possible Obfuscation
- Continuous-Variable Quantum State Designs: Theory and Applications
- Scalable Equivalence Checking and Verification of Shallow Quantum Circuits
- How to Use Quantum Indistinguishability Obfuscation
- Continuous-Variable Designs and Design-Based Shadow Tomography from Random Lattices
- Equivalence Checking of Parameterised Quantum Circuits
- On Decidable and Undecidable Extensions of Simply Typed Lambda Calculus
- On Low-End Obfuscation and Learning
- Backdoor Defense, Learnability and Obfuscation
- Injecting Undetectable Backdoors in Obfuscated Neural Networks and Language Models
- Global value numbering using random interpretation
- VeriQBench: A Benchmark for Multiple Types of Quantum Circuits