Cat Breeding Protocol in Quantum Optics

• Cat breeding protocol is a heralded optical procedure that combines smaller non-Gaussian cat-like states to produce larger-amplitude Schrödinger cat or multi-peak grid states.
• It utilizes Gaussian interference with homodyne or photon-number heralding to iteratively amplify states, achieving scalability while managing fidelity and event rates.
• Recent advances employ adaptive, cluster-state, and gate-based embedded architectures to mitigate loss, improve resource synchronization, and enhance fault-tolerant grid state generation.

Cat breeding protocol denotes a family of conditional continuous-variable optical procedures in which smaller Schrödinger cat states or cat-like non-Gaussian resources are combined to produce either a larger-amplitude cat state or a multi-peak grid/GKP-like state. The ideal targets are the familiar even and odd two-component cats,

$$|\mathrm{Cat}_{\pm}\rangle \propto |\alpha\rangle \pm |-\alpha\rangle,$$

but many experimentally relevant realizations operate with squeezed cat states or approximate cat-like states rather than perfect unsqueezed cats. In the narrow optical sense, breeding is a heralded interference-and-measurement primitive for iterative cat growth; in the broader contemporary sense, it also includes cat-to-grid breeding, adaptive variants without post-selection, and architectures that embed breeding inside cluster-state or gate-based processing (Etesse et al., 2014, Weigand et al., 2017, Endo et al., 22 Jun 2026).

1. Definition and conceptual scope

In the optical literature, “cat breeding” most directly means a heralded operation that takes two smaller cat-like states and combines them into a larger Schrödinger cat state. The core motivation is scalability: direct heralded generation of large free-propagating optical cat states becomes rapidly inefficient as the target cat size increases, whereas a breeding step can in principle be iterated, allowing one to grow cat amplitude stage by stage (Etesse et al., 2014).

The relevant state manifold is parity-structured. Even cats contain only even photon numbers and odd cats only odd photon numbers. This parity structure is central both to the original breeding logic and to later grid-state protocols. In the small-amplitude regime, an odd cat reduces to a single-photon resource,

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$

which explains why single-photon Fock states can act as minimal odd-cat seeds in the first experimental breeding step (Etesse et al., 2014).

A standard amplitude-growth law appears repeatedly: one successful breeding round enlarges the effective cat amplitude approximately as

$\alpha \rightarrow \sqrt{2}\,\alpha.$

This scaling underlies both iterative cat enlargement and the construction of many-peak lattices for GKP-like states. At the same time, later work makes clear that the resources actually available in the laboratory are often squeezed cat states or generalized photon-subtraction outputs rather than ideal cats. In particular, the picosecond GPS experiment explicitly states that its measured states are squeezed cat states or cat-like states and that the work does not experimentally implement a full cat-breeding enlargement loop as the central result, even though the generated states are intended as resources for adaptive breeding toward logical-qubit generation (Endo et al., 22 Jun 2026).

The term therefore has two technically distinct but historically connected uses. One is cat enlargement: two cat-like inputs, a Gaussian interaction, and a conditional measurement yield a larger cat. The other is cat-to-grid breeding: repeated combination of two-peak states builds a multi-peak comb approaching an approximate grid or GKP state. The second use has become increasingly prominent because GKP-state preparation is the principal fault-tolerance application of cat breeding (Weigand et al., 2017).

2. Canonical optical breeding operations

The first experimental implementation of a breeding operation used two heralded single-photon Fock states as the smallest odd-cat-like resources. The inputs

$|1\rangle \otimes |1\rangle$

were mixed on a symmetric $50{:}50$ beamsplitter, one output arm was measured by homodyne detection of the $x$-quadrature, and the other arm was retained conditionally. For the Fock input $|1\rangle|1\rangle$, Hong–Ou–Mandel interference gives

$$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$

and ideal conditioning at $x'=0$ projects the remaining arm onto

$$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$

That state has $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$0 fidelity with an even squeezed cat target, while the experimentally reconstructed conditional output had fidelity $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$1 with an even squeezed Schrödinger cat state of amplitude $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$2 and squeezing factor $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$3 along the $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$4 quadrature (Etesse et al., 2014).

The same experiment implemented homodyne heralding with a finite acceptance window

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$5

explicitly realizing the usual trade-off between success probability and fidelity. The corrected Wigner negativity was reported as

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$6

while the uncorrected value was

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$7

The practical rates were also characteristic of early free-propagating breeding: single-photon heralding occurred at about $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$8 per source, double coincidences at about $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$9, and with the homodyne conditioning window the final conditioned event rate was about $\alpha \rightarrow \sqrt{2}\,\alpha.$0 (Etesse et al., 2014).

A distinct linear-optical amplification variant replaces homodyne heralding by photon-number heralding at the final breeding stage. In that protocol, two odd kitten states generated by one-photon subtraction from squeezed vacuum are interfered on a third beamsplitter and heralded by a photon-number detector so that the output has the same parity as the inputs. For the main case

$\alpha \rightarrow \sqrt{2}\,\alpha.$1

the output parity is odd. The reported simulations give two odd kittens of $\alpha \rightarrow \sqrt{2}\,\alpha.$2 and $\alpha \rightarrow \sqrt{2}\,\alpha.$3, generated from one-photon-subtracted squeezed vacuum states of $\alpha \rightarrow \sqrt{2}\,\alpha.$4 dB, amplified to an odd cat state of $\alpha \rightarrow \sqrt{2}\,\alpha.$5 with fidelity $\alpha \rightarrow \sqrt{2}\,\alpha.$6; the main text also describes an amplified state with $\alpha \rightarrow \sqrt{2}\,\alpha.$7 and $\alpha \rightarrow \sqrt{2}\,\alpha.$8. With input squeezed vacuum increased to $\alpha \rightarrow \sqrt{2}\,\alpha.$9 dB, the predicted amplified odd cat is $|1\rangle \otimes |1\rangle$0 with $|1\rangle \otimes |1\rangle$1. Under a modeled realistic regime with $|1\rangle \otimes |1\rangle$2 and $|1\rangle \otimes |1\rangle$3, the amplified state remains at $|1\rangle \otimes |1\rangle$4 but the fidelity drops to $|1\rangle \otimes |1\rangle$5 (Song et al., 2021).

These two lines of work establish the canonical optical picture. A breeding step uses Gaussian interference, conditions on a homodyne or photon-number event, and produces a larger parity-structured non-Gaussian state. The experimentally difficult part is not the Gaussian network but the synchronized supply of sufficiently pure input resources.

3. Cat-to-grid breeding and the elimination of post-selection

A major extension of cat breeding is the generation of grid or GKP-like states. In this setting, the goal is not merely a larger two-component cat but a many-peak comb approaching an approximate simultaneous eigenstate of commuting displacement operators. The 2017 no-post-selection analysis formulates a squeezed cat as

$|1\rangle \otimes |1\rangle$6

and shows that a single breeding round consists of a $|1\rangle \otimes |1\rangle$7 beamsplitter followed by $|1\rangle \otimes |1\rangle$8-quadrature homodyne measurement on one output arm. After the measurement outcome $|1\rangle \otimes |1\rangle$9, the operator strings are updated by

$50{:}50$0

$50{:}50$1

The efficient version uses $50{:}50$2 initial cats and $50{:}50$3 rounds; for final spacing $50{:}50$4, the initial cat amplitude is

$50{:}50$5

The central result is that post-selection on $50{:}50$6 is unnecessary: all runs can be kept, and the full homodyne record determines the shifted grid state that was obtained, after which one either applies a correcting displacement or tracks the offset in a displacement frame (Weigand et al., 2017).

This reformulation matters because the original post-selected breeding logic becomes exponentially impractical with depth. The no-post-selection protocol accepts all measurement outcomes and uses classical post-processing instead of discarding runs. Numerical simulations with $50{:}50$7 and $50{:}50$8 showed that the no-post-selection protocol still improves the effective squeezing parameter $50{:}50$9 strongly with each round, although it remains somewhat worse than the idealized post-selected curve (Weigand et al., 2017).

Loss fundamentally alters this optimistic asymptotic picture. A 2025 analysis models optical loss on the input squeezed cats before breeding and evaluates the resulting output in terms of effective squeezing relative to the qunaught threshold of $x$0 dB symmetric effective squeezing. The conclusion is sharp: fault-tolerant GKP preparation becomes impossible once the input loss exceeds about $x$1, and the protocol fails to produce a fault-tolerant grid state when the input states experience $x$2 loss or more. For $x$3 cats postselected on $x$4, the reported values are

$x$5

at $x$6, but only

$x$7

at $x$8 (Solodovnikova et al., 8 Aug 2025).

That loss analysis also resolves a common misconception. Nonzero homodyne outcomes do not merely induce an overall displacement; they induce relative complex phases on different Gaussian peaks. A feedforward displacement may align the lattice, but it cannot in general undo measurement-induced incoherent phase structure among the peaks. The paper therefore argues that the protocol is deterministic only in the narrow sense that every run yields a conditional output state; it is not deterministic in the sense that every output is fault-tolerant or even useful by a chosen quality criterion (Solodovnikova et al., 8 Aug 2025).

4. Resource-state generation and the temporal-mode bottleneck

Recent work has shifted from breeding itself to the supply of breeding-ready resource states. The most explicit example is the picosecond generalized photon-subtraction experiment, which demonstrates multi-photon generalized photon subtraction in picosecond optical wave packets and explicitly frames the output as a resource for adaptive breeding rather than a completed breeding loop. The architecture interferes two independently phase-controlled ultrashort squeezed vacua on a beam splitter and heralds with photon-number-resolving detection. Experimentally, the two squeezed modes were generated by broadband type-0 PPLN waveguide OPAs pumped by $x$9-nm pulses at $|1\rangle|1\rangle$0 MHz with duration about $|1\rangle|1\rangle$1 ps. One output arm was spectrally filtered and sent to a transition-edge sensor operated as a PNRD, while the other was measured by pulsed balanced homodyne detection matched to a $|1\rangle|1\rangle$2-ps temporal mode (Endo et al., 22 Jun 2026).

Several hardware parameters are directly protocol-relevant. The local-oscillator autocorrelation gave a $|1\rangle|1\rangle$3 ps full width at half maximum, temporal and spatial overlaps exceeded $|1\rangle|1\rangle$4 and $|1\rangle|1\rangle$5, the TES had calibrated system detection efficiency $|1\rangle|1\rangle$6 at $|1\rangle|1\rangle$7 and decay time constant $|1\rangle|1\rangle$8 ns, the modeled heralding-arm efficiency was $|1\rangle|1\rangle$9, the homodyne efficiency was $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$0, the two OPA gains were $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$1 dB and $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$2 dB, the beam-splitter reflectivity was $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$3, and the interference phase was fixed at $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$4 (Endo et al., 22 Jun 2026).

The observed outputs were reconstructed without loss correction and showed clear non-Gaussianity. The Wigner functions exhibited up to four distinct negative regions for four-photon heralding. After inverse squeezing to vacuum-width peaks, the reported effective cat amplitudes were

$$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$5

for $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$6, with partially corrected reference amplitudes

$$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$7

The distinction is explicit: $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$8 is extracted from the experimentally reconstructed, largely uncorrected state and reflects the practically available resource, whereas $$|1\rangle|1\rangle \longrightarrow \frac{1}{\sqrt 2}\big(|2\rangle|0\rangle - |0\rangle|2\rangle\big),$$9 is only a reference estimate before homodyne loss (Endo et al., 22 Jun 2026).

The same work derives the cat-to-grid geometry relevant to breeding. For a cat of amplitude $x'=0$0, the initial quadrature peak spacing is

$x'=0$1

and after $x'=0$2 breeding steps the neighboring-peak spacing becomes

$x'=0$3

For two-step breeding to the symmetric qunaught lattice constant, the required initial amplitude is

$x'=0$4

The reported $x'=0$5 for $x'=0$6 is therefore within about $x'=0$7 of that target while preserving clear Wigner negativity (Endo et al., 22 Jun 2026).

The supplementary tree-protocol simulation makes the breeding step explicit: each step interferes two copies of the state on a balanced beamsplitter and heralds on a homodyne measurement of the conjugate quadrature at outcome zero. In the lossless counterpart of the $x'=0$8 state, Wigner negativity survives over two breeding steps with minima

$x'=0$9

For the experimentally reconstructed uncorrected state, the comb structure still forms in position, but the negativity decays as

$$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$0

so after two steps the negativity is statistically indistinguishable from zero. The practical implication is precise: amplitude and peak geometry are near the required regime, but purity and loss remain the limiting factors for repeated breeding (Endo et al., 22 Jun 2026).

5. Architectural variants and embedded breeding

Cat breeding is no longer confined to the original free-space beamsplitter picture. One measurement-based variant embeds known breeding protocols inside a continuous-variable cluster state. In that framework, PhANTM generates embedded cat states by photon-counting-assisted node teleportation, and cat breeding is implemented by teleporting one cat through another cat-state ancilla. Operationally, the ancilla’s $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$1-wavefunction multiplies the input’s $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$2-wavefunction,

$$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$3

With one quadrature orientation, this sharpens the two-peak structure and, after Gaussian rescaling, yields a larger cat; with another orientation, it produces a three-peak state and repeated application yields a grid/GKP-like state. The same work shows that previous beamsplitter breeding protocols map directly to cluster-state $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$4-based processing up to Gaussian corrections. Quantitatively, after $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$5 PhANTM steps the average fidelity with the nearest cat exceeds $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$6 in all examined cases, and for $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$7 dB cluster squeezing, $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$8 steps generate $$|\psi_{\rm out}\rangle= \frac{1}{\sqrt3}|0\rangle+\sqrt{\frac23}|2\rangle.$$9 with $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$00 success and $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$01 with $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$02 success; one subsequent breeding round gives approximate GKP qunaughts with average fidelity above $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$03, and above $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$04 for all states with $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$05 (Eaton et al., 2021).

A different gate-based variant replaces beamsplitter interference by an iterative QND entangling gate. In that protocol, a small cat ancilla and a target mode interact through

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$06

after which the ancilla momentum is measured by homodyne detection. For a target squeezed coherent state, the ideal heralding condition is

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$07

The single-step output has size $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$08 and inverse quadrature squeezing

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$09

The amplified output is then rotated by $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$10 and reused as the next ancilla, yielding after $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$11 successful rounds

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$12

This is breeding-like amplification in a strict operational sense—small cat resources are turned into larger squeezed cats step by step—but it is not the standard two-cat beamsplitter fusion picture. The protocol also provides exact finite-window success-probability and mixed-state-fidelity formulas, and it reports a strong parity asymmetry: even ancillas are more favorable because the probability density peaks at the ideal outcome, while odd ancillas are more sensitive to finite acceptance windows (Goncharov et al., 1 Jun 2026).

Cat breeding also appears as a preprocessing stage for hybrid entanglement. A linear-optical 2026 scheme starts from small-amplitude odd cats, uses breeding to increase the non-Gaussianity of a mode-1 resource $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$13, and then feeds that bred state into a three-mode circuit that outputs a hybrid entangled state between an approximate GKP-like logical mode and a photon-number qubit. In the baseline form, the target is

$$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$14

and the exact-state fidelity analysis shows a nontrivial optimum at $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$15 with $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$16. For an average fidelity of $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$17, the success probability is about $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$18; for success probability $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$19, the average fidelity remains about $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$20; and for a relaxed homodyne window $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$21, the average fidelity is about $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$22 while the success probability rises to about $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$23 (Kiryu et al., 20 Mar 2026).

6. Limitations, misconceptions, and present research directions

The central limitation of cat breeding is loss. In free-propagating optical settings, losses degrade cat coherence, suppress Wigner negativity, broaden peaks, and distort the measurement-conditioned phase relations that breeding relies on. The most explicit threshold analysis concludes that fault-tolerant qunaught generation is still possible at about $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$24 loss but fails at $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$25 loss or more under the paper’s assumptions, even before including additional nonidealities such as homodyne inefficiency or propagation loss beyond the modeled input-state loss (Solodovnikova et al., 8 Aug 2025).

A second limitation is event rate. The original homodyne-heralded breeding experiment already emphasized low coincidence rates and the need for synchronized resource supply, which is why quantum memories were identified as an enabling technology for iterative implementations (Etesse et al., 2014). The ultrafast GPS resource-state experiment changes the rate regime by moving from nanosecond modes to $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$26-ps wave packets at $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$27 MHz, but it also makes clear that the current rate is limited mainly by the timing jitter of the TES ($$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$28 ns) and the bandwidth of the homodyne receiver ($$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$29 MHz), and that $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$30 events lacked sufficient statistics for reliable tomography (Endo et al., 22 Jun 2026).

A third limitation is conceptual. Not every cat-state protocol is a breeding protocol. Direct preparation of entangled cat states in Kerr-nonlinear resonators using cavity-mediated cat-code Mølmer–Sørensen gates and composite two-photon drives is explicitly not cat breeding in the narrow optical sense: it does not combine two smaller cats into a larger-amplitude single-mode cat, and it does not rely on heralded interference. It is instead a direct state-preparation and control protocol for entangled cat states in a cat-qubit architecture (Gu et al., 2024).

The main contemporary misconception is therefore twofold. First, “cat breeding” does not always mean simple amplitude doubling of a free-propagating optical cat; in much of the recent literature it means cat-to-grid conversion, adaptive breeding, or embedded breeding inside larger architectures. Second, feedforward is not a universal cure for poor heralding outcomes: in lossy regimes, the outcome-dependent relative phases among peaks can destroy the target structure in ways that a single corrective displacement cannot repair (Solodovnikova et al., 8 Aug 2025).

The present research trajectory is correspondingly stratified. One direction develops better front-end resource sources, exemplified by picosecond generalized photon subtraction near the $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$31 two-step qunaught threshold (Endo et al., 22 Jun 2026). A second direction refines breeding laws and architectures, including no-post-selection grid breeding, cluster-state embeddings, and iterative $$|\alpha\rangle-|-\alpha\rangle \propto 2\alpha |1\rangle + O(\alpha^3),$$32-gate amplification (Weigand et al., 2017, Eaton et al., 2021, Goncharov et al., 1 Jun 2026). A third direction quantifies the narrow operating margins imposed by loss, detector efficiency, and heralding statistics. Taken together, these results define cat breeding not as a single protocol but as a technical program: the controlled conversion of small non-Gaussian bosonic resources into larger cats or grid states by Gaussian interactions, conditional measurements, and, increasingly, architecture-specific feedforward or embedding.