Scalable Postselection in Quantum Systems

• Scalable postselection is the conditioning of quantum operations on designated measurement outcomes while renormalizing probabilities to maintain operational significance.
• It addresses nonlinear challenges by mitigating the amplification of small errors and ensuring modular error analysis across varying system sizes.
• The concept underpins advancements in error correction, bounded-space complexity, and metrology by guiding designs for controlled abort rates and constant overhead.

Postselection is the operation of conditioning a computation or measurement protocol on a designated outcome and discarding all runs in which that outcome does not occur. In quantum information, the conditioned map is operationally meaningful but generally nonlinear, because outputs are renormalized by an input-dependent success probability. “Scalable postselection” therefore does not denote a single technique. Across current work, it refers to several distinct problems: whether postselected computations admit a modular error theory; whether postselection retains a clean complexity-theoretic meaning under severe time-space restrictions; whether abort-and-retry rules can improve quantum error correction without driving acceptance to zero; whether fault-tolerant architectures can use local or hierarchical postselection with only constant or controlled overhead; and whether metrological, cryptographic, and variational protocols can preserve useful information or trainability after conditioning (Gavorová, 2020, English et al., 2024, Birchall et al., 5 Mar 2026, Staples et al., 9 Mar 2026).

1. Domain-specific meanings of scalability

The literature uses “scalable postselection” in several non-equivalent senses. In all cases, the common issue is whether the benefits of conditioning survive increasing system size, code distance, runtime, or estimator precision, rather than being nullified by vanishing success probability, abort-rate blowup, or loss of composability (Dhillon, 2011, English et al., 2024, Birchall et al., 5 Mar 2026, Zhong et al., 2024).

Setting	Scalability criterion	Representative papers
Postselected computations	Preserve enough distance/contractivity structure for modular error analysis	(Gavorová, 2020)
Space-bounded complexity	Characterize postselection under log-space or simultaneous time-space bounds	(Gall et al., 2021, Tani, 2022, Yakaryilmaz et al., 2011)
QEC thresholds	Conditional logical failure and abort probability both vanish with code size	(English et al., 2024, Chen et al., 6 Oct 2025)
Fault-tolerant architectures	Postselect constant-size windows or fixed fractions with controlled overhead	(Birchall et al., 5 Mar 2026, Staples et al., 9 Mar 2026, Lee et al., 7 Oct 2025)
Metrology	Recover or redirect useful Fisher information after conditioning	(Alves et al., 2016, Zhong et al., 2024, Ho et al., 2018, Qin et al., 2023)
Variational photonics	Avoid exponential gradient concentration induced by postselection geometry	(Xie et al., 12 May 2026)

A recurring distinction is between the underlying linear completely positive map and the conditioned map. In the channel-theoretic setting, one begins with a linear completely positive trace-nonincreasing map $\Phi$, but the actual postselected computation is

$$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$

which is nonlinear in $\rho$ (Gavorová, 2020). In QEC and metrology, the same structural issue reappears as conditioning on syndromes or measurement branches. In complexity theory, by contrast, postselection is usually formalized as allowing non-postselecting halts to be discarded, with acceptance probabilities renormalized over the remaining branches (Gall et al., 2021, Yakaryilmaz et al., 2011).

2. Nonlinearity, distinguishability, and modular error analysis

A foundational obstacle to scalable postselection is that renormalization breaks the linear structure on which ordinary channel theory relies. For postselection superoperators $\Psi,\Phi$, the postselected trace distance and postselected diamond distance compare normalized outputs,

$$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$

with stabilization defining $\widehat d_\diamondsuit$ (Gavorová, 2020). These are only pseudometrics: two different underlying CP maps can induce the same normalized map. That already signals a scalability problem, because exact postselected behavior does not by itself determine success probabilities or linear structure.

The paper on distinguishability of postselected computations shows that several standard tools of scalable error analysis fail outright in this setting. Convexity can fail so severely that mixed inputs are much worse than pure inputs. Ordinary contractivity under postprocessing can fail, and naive subadditivity under composition can fail as well. The mechanism is always the same: small differences in success probabilities are magnified by renormalization. This directly obstructs fault-tolerance-style reasoning, where local errors are normally chained using convexity, contractivity, and subadditivity (Gavorová, 2020).

The same work also identifies the restricted structure that survives. If the ideal comparison map is trace-preserving, then closeness of normalized outputs forces the success probability of the postselected approximation to be nearly input-independent. This is the role of the conversion lemma. It permits replacement of the nonlinear conditioned map by a suitably rescaled linear CP map and thereby imports weaker forms of ordinary inequalities. In particular, weak contractivity and weak subadditivity hold when the ideal later-stage map is trace-preserving, and a postselected computation close in $\widehat d_{tr}$ to an isometry is also controlled in stabilized distance and in Stinespring form (Gavorová, 2020).

This yields a restricted notion of scalable analysis. It supports modular chaining when postselected gadgets are compared to trace-preserving ideals, often unitaries or isometries, and when the relevant success probabilities are forced to be almost constant by the comparison. It does not solve the physical problem that the total success probability of many postselection steps may still become exponentially small. Nor does it restore a full analogue of ordinary channel theory. The best-supported conclusion is therefore partial: postselected computations admit a usable but restricted distinguishability theory, not a complete scalable substitute for CPTP-channel analysis (Gavorová, 2020).

3. Complexity-theoretic scalability under time and space bounds

In complexity theory, postselection remains meaningful under stringent resource constraints, but its meaning depends sharply on the model. For logarithmic-space quantum computation, the central result is

$\mathsf{PostBQL}=\mathsf{PL},$

where postselection is formalized by conditioning on halting in accept or reject rather than in a non-postselecting halt state (Gall et al., 2021). The same work shows

$$\mathsf{PL}=\mathsf{PQL}=\mathsf{PostBQL}=\mathsf{BQL(exp)}=\mathsf{BQL(\infty)},$$

so postselection closes the gap between bounded-error log-space quantum computation and a standard unbounded-error classical log-space class. This is a complexity-theoretic notion of scalability: postselection remains exactly characterizable even with only $O(\log n)$ workspace.

A related structural result shows that in space-bounded bounded-error quantum computation, intermediate measurements and intermediate postselections can be eliminated without changing computational power. For every space-constructible $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$0,

$$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$1

and in particular $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$2 (Tani, 2022). Here the significance is compositional: repeated postselection during an $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$3-space computation can be compiled away into a unitary $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$4-space process with only final postselection. This is scalability in the complexity-theoretic sense, not in the hardware sense.

The machine-theoretic study of postselection under simultaneous time-space bounds proves several stricter power increases. Postselection is equivalent to restart for the relevant resource-bounded PTMs and QTMs, and this equivalence lets one prove that postselection strictly augments the power of classical machines using $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$5 space and polynomial expected time, and of real-time constant-space quantum machines as well (Yakaryilmaz et al., 2011). The same paper gives a clean low-space quantum-over-classical separation for postselected real-time machines, and shows that

$$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$6

For $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$7, this implies that $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$8 would equal $$\rho \mapsto \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]},$$9 if randomized machines had postselection (Yakaryilmaz et al., 2011).

These results are mathematically strong but physically idealized. They allow postselected events of exponentially small probability and therefore do not constitute a hardware scalability theory. A short note arguing that adaptive measurements can be simulated by postselection is relevant only conceptually and complexity-theoretically: it does not provide an explicit efficient transformation, branch-probability analysis, or a rigorous efficient simulation theorem (Dhillon, 2011).

4. Statistical-mechanical thresholds and syndrome-selective QEC

In QEC, the most precise definition of scalable postselection is that both the conditional logical failure probability and the abort probability vanish as code size grows. The statistical-mechanics analysis of post-selected QEC formulates postselection as a partition of the error set into $\rho$0 and $\rho$1, then studies logical decoding on the constrained disorder ensemble (English et al., 2024). For surface and toric codes, this leads to two distinct asymptotic thresholds: a conditional logical threshold and an abort threshold. Their interplay yields four thermodynamic phases. Only the phase below both thresholds is fully scalable: accepted runs become reliable, and accepted runs still occur with asymptotically unit probability.

That framework also provides a simple heuristic for surface codes that avoids decoding entirely. Using the non-equilibrium magnetization

$\rho$2

one aborts when $\rho$3 falls below the cutoff $\rho$4, equivalently when

$\rho$5

For the toric code under bit-flip noise, the full-postselection limit $\rho$6 reduces the disorder model to a clean Ising model and gives the exact conditional threshold

$\rho$7

For depolarizing noise, the corresponding clean isotropic eight-vertex model yields $\rho$8 (English et al., 2024). These are conditional logical thresholds, not unconditional operating thresholds: scalability still requires remaining below the abort threshold.

A complementary line of work explains why postselection can improve logical accuracy at all. In the toric code with perfect syndrome measurements, the logical failure rate can be written in terms of the free-energy difference $\rho$9 between competing homology sectors. Within the coding phase, dangerous low-confidence syndromes are exponentially unlikely, and the distribution obeys a large deviation principle. As a result, postselecting away those rare syndromes suppresses the logical error from $\Psi,\Phi$0 to $\Psi,\Phi$1, with $\Psi,\Phi$2 in general; for the toric code with perfect syndrome measurements, the numerical estimate is

$\Psi,\Phi$3

The improvement is scalable precisely because the discarded syndromes are themselves exponentially rare (Chen et al., 6 Oct 2025).

This statistical-mechanical picture also clarifies a frequent misconception. Postselection does not generally rescue above-threshold devices into scalable fault tolerance. The threshold paper argues that if a system is above the standard logical threshold, postselection does not bring it below the conditional logical threshold without typically pushing it above the abort threshold (English et al., 2024). The large-deviation paper makes a related point in a different language: the method improves the accepted-run exponent, but for very long computations the cumulative abort constraint can erase the benefit (Chen et al., 6 Oct 2025). The strongest use case is therefore finite-size improvement in already subthreshold devices.

5. Fault-tolerant constructions, resource states, and general QLDPC postselection

Architectural work on fault tolerance uses “scalable postselection” in a more constructive sense: postselection is made local, hierarchical, or confidence-filtered so that acceptance does not collapse with distance. The clearest example is Macromux, or macroscale multiplexing, which postselects only on constant-size space-time windows of a fault-tolerant protocol. A protocol is diced into constant-size bricks, $\Psi,\Phi$4 copies of each brick are prepared, each copy is scored using syndrome or erasure information, and equal ranks are recursively paired so that the best bricks are fused with the best. Because the windows are constant size, the acceptance rate per brick stays bounded away from zero and the additional overhead remains a constant multiplicative factor set by $\Psi,\Phi$5. In surface-code fusion networks, this yields threshold gains as large as $\Psi,\Phi$6 for Pauli errors, erasure thresholds from $\Psi,\Phi$7 to about $\Psi,\Phi$8 in the asymptotic limit, and in some photonic settings doubled loss thresholds with only $\Psi,\Phi$9 overhead (Birchall et al., 5 Mar 2026).

A different route keeps the postselected object extensive in code distance, but rejects only a fixed fraction of low-confidence instances. In teleportation-based logical gates through surface-code cluster states, the relevant pre-consumption information is incomplete because boundary syndrome bits are hidden until the resource is used. The paper on scalable postselection of quantum resources introduces the partial gap,

$$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$0

which estimates, from visible bulk syndrome alone, the logical gap expected after consumption. Postselecting on this decoder soft information produces scalable improvements in logical error rate, with an effective distance enhancement of about $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$1 in the bulk-limited regime and a reported $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$2 reduction in spacetime overhead per logical gate at the same logical error probability (Staples et al., 9 Mar 2026).

For general quantum LDPC codes, earlier logical-gap postselection was limited by an overhead exponential in the number of logical qubits, because it required decoding over all $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$3 logical classes. The QLDPC post-selection work replaces this with heuristic cluster-based inverse-confidence metrics extracted from a single clustering-decoder run. The two main scores are normalized cluster-size and cluster-LLR $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$4-norm fractions, with acceptance rule $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$5. This makes confidence-based postselection compatible with arbitrary QLDPC codes and with sliding-window real-time decoding. The strongest example is the $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$6 bivariate bicycle code, where the method yields approximately three orders of magnitude reduction in logical error rate with abort rate only $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$7 at physical error rate $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$8, and $$\widehat d_{tr}(\Psi,\Phi) := \sup_{\rho\in\mathcal D(H)} \left\| \frac{\Psi(\rho)}{\mathrm{tr}[\Psi(\rho)]} - \frac{\Phi(\rho)}{\mathrm{tr}[\Phi(\rho)]} \right\|_{tr},$$9 at $\widehat d_\diamondsuit$0. In sliding-window mode it supports early mid-circuit abort decisions and favorable scaling with the number of rounds (Lee et al., 7 Oct 2025).

Taken together, these constructions show that scalable fault-tolerant postselection need not mean the same thing everywhere. One approach fixes the postselected object size and keeps overhead constant; another keeps rejection fraction fixed for large resource states; a third performs decoder-native confidence filtering for finite-rate QLDPC blocks. What they share is refusal of “perfect-syndrome or reject everything” logic.

6. Metrological uses: information splitting, information transfer, and conditional precision

Metrological work treats scalability as a question of whether postselection can improve precision without throwing away so much data that the gain evaporates. The most robust conclusion is that the relevant information is generally split between the postselected meter and the postselection statistics themselves. In the optical experiment comparing weak-value amplification with a non-amplifying post-selection strategy, the total Fisher information is

$\widehat d_\diamondsuit$1

where $\widehat d_\diamondsuit$2 comes from the surviving meter and $\widehat d_\diamondsuit$3 comes from success/failure counts (Alves et al., 2016). Both strategies can approach the quantum limit for sufficiently small $\widehat d_\diamondsuit$4, but only if the post-selection statistics are included. The high-amplification weak-value regime suffers from low success probability and non-Gaussian meter distortion, while the non-amplifying strategy with $\widehat d_\diamondsuit$5 keeps $\widehat d_\diamondsuit$6 and appears markedly more scalable.

The same redistribution principle underlies the later coherent-state proposal based on Fisher information transfer. There the total postselected information decomposes as

$\widehat d_\diamondsuit$7

with information distributable among successful photons, failed photons, and success/failure probabilities (Zhong et al., 2024). By tuning the relative phase $\widehat d_\diamondsuit$8, the protocol transfers quantum-scale $\widehat d_\diamondsuit$9-scaling Fisher information from the postselection statistics into the successful output meter state, and a power-recycling cavity further reshapes the branch probabilities in favor of the higher-information branch. The claim is not that postselection creates information, but that it can redirect information to experimentally more usable observables.

A stronger claim appears in the quadratic nonlinear metrology protocol with pre- and post-selection. For a coherent-state probe with mean photon number $\widehat d_{tr}$0, the conventional quadratic nonlinear scheme has $\widehat d_{tr}$1, corresponding to $\widehat d_{tr}$2, while the paper argues numerically that the success-probability-weighted Fisher information of the postselected scheme scales as $\widehat d_{tr}$3, yielding $\widehat d_{tr}$4 (Qin et al., 2023). Because the success probability is included, this is presented as an unconditional scaling claim within the idealized model. The paper itself, however, treats it as numerically supported rather than as a general asymptotic theorem.

In multiparameter estimation, postselection is used differently. The formalism averages over all orthonormal postselection branches, with pCFIM and pQFIM weighted by branch probabilities. The resulting hierarchy

$\widehat d_{tr}$5

shows that postselection cannot exceed the sensor’s total QFIM, but in the worked phase-and-fluctuation example the MA after postselection can in principle recover the whole sensor information, $\widehat d_{tr}$6, even when the sensor is mixed (Ho et al., 2018). The unresolved issue is constructive attainability: for mixed states, the paper does not provide a general optimal POVM.

7. Trainability, communication, and speculative extensions

Beyond error correction and sensing, several recent works ask whether postselection itself remains operationally manageable in broader architectures. In passive photonic variational circuits, the decisive quantity is not the full Fock-space dimension but how postselection reshapes the effective observable $\widehat d_{tr}$7. Exact simulations show that allow-bunching evolution and collision-free filtering have gradient variance consistent with polynomial rather than exponential decay over the tested sizes, whereas dual-rail postselection induces exponential concentration beyond moderate sizes, robust across three initialization ensembles (Xie et al., 12 May 2026). This indicates that postselection is not intrinsically incompatible with trainability; the obstruction is specific postselection geometry.

In continuous-variable QKD with measurement-based noiseless linear amplification, the main issue is finite-size sample discard. Gaussian post-selection can improve secure range asymptotically, but in realistic finite blocks the best operating point is typically non-Gaussian because larger cutoffs discard too many samples. The finite-size key rate depends on $\widehat d_{tr}$8, so post-selection reduces both the prefactor and the effective sample size used in finite-size penalties (Hosseinidehaj et al., 2019). This is a different but closely related notion of scalability: one must preserve enough accepted data for composable security estimates to remain nontrivial.

A more implementation-oriented proposal replaces explicit measure-and-restart postselection by tracing out repeated fresh ancilla blocks. In the ancilla-thermalization scheme, post-selection on $\widehat d_{tr}$9 ancilla qubits is replaced with tracing out

$\mathsf{PostBQL}=\mathsf{PL},$0

ancillae blocks to attain the same target accuracy, where $\mathsf{PostBQL}=\mathsf{PL},$1 is the one-shot success probability (Wright et al., 2020). The output is approximate rather than exact, but the method removes mid-circuit measurement and was motivated especially by superconducting hardware. It is best viewed as a measurement-free replacement for some repeat-until-success gadgets, not as a universal asymptotic improvement.

Finally, some algorithmic proposals remain speculative. The two-way quantum computing paper argues that “postparation,” conceived as a time-reversed analogue of state preparation, could mitigate the poor scaling of ordinary postselection by enforcing final states physically rather than by repeated discard. Its strongest claims are qualitative: it does not provide a rigorous complexity-theoretic proof of polynomial savings or an algorithm-specific efficient simulation theorem (Linden et al., 2024). Likewise, the short note on adaptive measurements and postselection is best understood as a conceptual bridge rather than a scalable construction (Dhillon, 2011).

Across these diverse settings, two conclusions recur. First, postselection does not by itself create information, threshold, or trainability advantages; it redistributes confidence, Fisher information, or error weight across branches. Second, scalability is domain-specific. In some contexts it means constant-overhead local filtering; in others, vanishing abort and logical-failure probabilities; in others, retention of enough accepted samples or gradient variance to avoid collapse. The modern literature therefore treats scalable postselection not as a blanket property of conditioning, but as a constrained design principle for making conditioning compatible with asymptotic control.