Random Peaked Circuits Overview

• Random peaked circuits are quantum circuits that produce one or few designated outcomes with significantly higher probabilities than others, diverging from uniform distributions.
• They integrate random gate layers with optimized peaking layers and obfuscation techniques to selectively amplify target outputs while retaining a pseudo-random structure.
• They are analyzed using classical simulation methods and tailored diagnostics to benchmark quantum systems and to explore complexity and verification challenges.

Random peaked circuits are quantum circuits whose output distribution over computational-basis bitstrings contains an anomalously large probability mass on one or a few designated outcomes while the rest of the circuit retains random or pseudo-random structure. In the most common contemporary usage, a circuit with unitary $U$ acting on $|0^n\rangle$ is peaked when there exists a hidden bitstring $s^\*$ such that $P(s^\*) \gg P(x)$ for most $x \neq s^\*$, where $P(x)=|\langle x|U|0^n\rangle|^2$; closely related literature also studies random low-depth circuits whose sorted output probabilities exhibit a high-and-narrow peak with a long Porter–Thomas tail, a distinct but historically important notion of “peakedness” (Aaronson et al., 2024, Gharibyan et al., 29 Oct 2025, Chen et al., 2019).

1. Definitions and statistical characterizations

For engineered peaked circuits, the central quantity is the maximum output probability

$p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$

A circuit is often called $\delta$-peaked when

$$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$

and, in the shallow-circuit literature, “peaked” can also mean

$$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$

These formulations all encode the same basic property: one basis state is polynomially or macroscopically heavier than the generic $|0^n\rangle$0 scale (Gharibyan et al., 29 Oct 2025, Bravyi et al., 2023).

A second usage arises from random circuit sampling. For an $|0^n\rangle$1-qubit random circuit, output probabilities $|0^n\rangle$2 are modeled by the Porter–Thomas law

$|0^n\rangle$3

and, when sorted in ascending order,

$|0^n\rangle$4

This sorted population exhibits a high, narrow peak at small $|0^n\rangle$5 and a long tail. In that setting, “peaked” refers not to a single planted bitstring but to the shape of the ensemble statistics (Chen et al., 2019).

Several auxiliary diagnostics recur across the literature. HQAP work uses peak weight, peak-to-average ratio $|0^n\rangle$6, and min-entropy $|0^n\rangle$7 (Gharibyan et al., 29 Oct 2025). Other work emphasizes collision probability

$|0^n\rangle$8

support-overlap scores such as

$|0^n\rangle$9

or biased expectation ratios

$s^\*$0

when the purpose of peaking is to enhance a target observable rather than merely identify a hidden string (Aaronson et al., 2024, Kim et al., 15 Apr 2025).

2. Construction paradigms

One explicit model augments $s^\*$1 layers of uniformly random two-qubit gates by $s^\*$2 optimized “peaking” layers in a 1D brick-wall or all-to-all architecture. The optimization target is

$s^\*$3

This model isolates the basic tradeoff between randomness and peak formation: if $s^\*$4, one can trivially invert the random layers and achieve $s^\*$5, whereas the interesting regime is $s^\*$6 (Aaronson et al., 2024).

The HQAP construction adds a more elaborate obfuscation layer. It first trains a shallow circuit to amplify a designated bitstring, then inserts a large mirrored identity block $s^\*$7, and finally applies tensor patch optimization, angle sweeping, masking, and swap transformations to frustrate direct cancellation. In the notation used there, the final unitary has the structure

$s^\*$8

The largest reported instance used 56 qubits, all-to-all connectivity, and 2044 two-qubit gates, with peak weight $s^\*$9 (Gharibyan et al., 29 Oct 2025).

A distinct quasi-random construction uses a random brick-wall first half

$P(s^\*) \gg P(x)$0

followed by an approximate inverse. Hidden $P(s^\*) \gg P(x)$1 gates are inserted according to a target string and commuted through the circuit using identities such as

$P(s^\*) \gg P(x)$2

and

$P(s^\*) \gg P(x)$3

The mirror half is then modified so that each block remains close in matrix action while differing significantly in parameters. Peakedness is controlled by an average deviation parameter $P(s^\*) \gg P(x)$4, with the ideal case $P(s^\*) \gg P(x)$5 giving $P(s^\*) \gg P(x)$6, and the paper adopts the criterion

$P(s^\*) \gg P(x)$7

to classify a circuit as peaked (Udalov, 10 Aug 2025).

Multi-peaked generalizations also appear. By inserting a small entangling block, for example an $P(s^\*) \gg P(x)$8 followed by $P(s^\*) \gg P(x)$9, one can construct two-peak or multi-peak output distributions rather than a single designated maximum (Udalov, 10 Aug 2025).

3. Classical simulation methods

The classical simulation landscape is heterogeneous because different peaked-circuit families expose different exploitable structures. For constant-depth peaked circuits, a central result is that an $x \neq s^\*$0-qubit peaked shallow circuit can be approximately sampled in quasipolynomial time $x \neq s^\*$1, while nearest-neighbor circuits on a 2D grid admit polynomial-time sampling $x \neq s^\*$2 and fixed $x \neq s^\*$3 grids admit $x \neq s^\*$4 algorithms. The key mechanism is Hamming-ball concentration implied by bounded light cones and a polynomially heavy output (Bravyi et al., 2023).

Low-depth random circuits with Porter–Thomas-peaked sorted distributions admit a different route. The teleportation-inspired algorithm rewires the circuit into a logical circuit whose number of qubits scales with depth rather than physical width. For a 1D chain, an $x \neq s^\*$5-qubit, $x \neq s^\*$6-depth circuit maps to an $x \neq s^\*$7-qubit, $x \neq s^\*$8-depth circuit; for an $x \neq s^\*$9 grid with depth $P(x)=|\langle x|U|0^n\rangle|^2$0, the logical-qubit count is $P(x)=|\langle x|U|0^n\rangle|^2$1; for Bristlecone it is $P(x)=|\langle x|U|0^n\rangle|^2$2. Memory then scales as $P(x)=|\langle x|U|0^n\rangle|^2$3 rather than $P(x)=|\langle x|U|0^n\rangle|^2$4. Demonstrated examples include a 1000-qubit depth-42 chain mapped to 42 logical qubits, a $P(x)=|\langle x|U|0^n\rangle|^2$5 depth-42 grid mapped to 40 logical qubits, and a 72-qubit depth-32 Bristlecone circuit mapped to 44 logical qubits. The same work combines this with threshold-rejection sampling; at

$P(x)=|\langle x|U|0^n\rangle|^2$6

the reported sample fidelity is approximately $P(x)=|\langle x|U|0^n\rangle|^2$7 and efficiency approximately $P(x)=|\langle x|U|0^n\rangle|^2$8 (Chen et al., 2019).

The HQAP family later received a near-exact classical attack based on full tensor-network contraction with a mirrored MPO and “unswapping.” The contraction exploits the approximate $P(x)=|\langle x|U|0^n\rangle|^2$9 core, factors hidden permutations as

$p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$0

and greedily reduces MPO bond dimension by testing left, right, or bilateral SWAP insertions. On the largest studied instance—56 qubits and 1,917 two-qubit RZZ gates—the method fully contracted the circuit on a single Nvidia A100 GPU in 4,059 seconds and recovered the peak with frequency $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$1, consistent with the designed $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$2 (Kremer et al., 23 Apr 2026).

A complementary approximate strategy is sparse truncated state-vector simulation. Here the state vector is stored as pairs $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$3, truncated either by top-$p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$4 magnitude or by retained probability mass, and updated blockwise using grouped contributions and segmented sums. On the 44-qubit “sharp peak” ring circuit with 580 instructions, the correct $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$5 was recovered using fewer than $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$6 retained amplitudes; the same study reports roughly one order-of-magnitude GPU speedup at larger $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$7, but also observes that $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$8 mass retention can already force more than $p_{\mathrm{peak}} = \max_x P(x)=P(s^\*).$9 terms in an intermediate step (Ferreira, 8 Jul 2026).

4. Hardness results and structural complexity

The theoretical picture is intentionally mixed: some peaked families are classically tractable because their peak is generated by shallow locality or mirror structure, whereas other formulations are designed to retain average-case hardness. In the explicit random-plus-peaking model, well-spread ensembles obey

$\delta$0

which implies

$\delta$1

for the probability that a random instance is $\delta$2-peaked. The same work proves that obtaining $\delta$3 peakedness from a random circuit of depth $\delta$4 requires

$\delta$5

optimized peaking layers with overwhelming probability (Aaronson et al., 2024).

A later formalization of random peaked circuits via postselection on row overlap takes a random circuit $\delta$6 from a $\delta$7-design, samples a second circuit $\delta$8 from the same architecture, and accepts $\delta$9 when a designated output becomes heavy. Conditioned on exact alignment, the orthogonal block is distributed as

$$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$0

with $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$1 Haar-random on $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$2 in the exact Haar case, and moment-wise equivalent to a $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$3-design in the design case. This yields non-compressibility and circuit-complexity lower bounds of $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$4 with high probability, and it underpins an average-case $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$5-hardness statement for estimating the peak to $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$6 additive error (Zhang, 30 Sep 2025).

At inverse-polynomial additive accuracy, the worst-case decision version changes complexity class. The same work notes that the corresponding worst-case problem is PromiseBQP-complete, while the HQAP paper proves QCMA-completeness for the promise problem PCBS, which asks whether there exists a basis state $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$7 such that $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$8 exceeds a threshold when neither the peaked input nor the output is known (Zhang, 30 Sep 2025, Gharibyan et al., 29 Oct 2025).

This separation between planted structure and average-case randomness is central. A plausible implication is that “random peaked circuits” do not denote a single hardness class but a design space in which simulability depends sensitively on whether the heavy output is produced by shallow geometry, mirrored identities, approximate inverses, or postselected design structure.

5. Verification, benchmarking, and applications

The practical appeal of peaked circuits is that verification can be reduced to detecting one designated outcome rather than reconstructing a full output distribution. In HQAP, verification is “peak match”: one samples the circuit and checks equality with the publicly known peak. The shot complexity to see the peak at least once is $$\max_{s\in\{0,1\}^N} |\langle s|C|0^N\rangle|^2 \ge \delta,$$9, and the largest 56-qubit instance was reported to produce the target peaked bitstring in under 2 hours on Quantinuum H2 (Gharibyan et al., 29 Oct 2025).

This verification simplicity motivated system-level benchmarking. Peaked Random Circuits were run as a matrix over

$$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$0

and the benchmark defined

$$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$1

Across multiple instances,

$$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$2

This protocol was applied to IQM superconducting and AQT trapped-ion systems, and was presented as comparable to Quantum Volume while exhibiting greater sensitivity to interference effects (Brieger et al., 25 May 2026).

Peaked-state preparation has also been used as a measurement primitive rather than a hardness primitive. For infinite-temperature correlation functions, engineered peaked states prepared by Grover-style amplitude amplification or shallow structured circuits increase support on operator-relevant subspaces. On 12 qubits with 8192 shots, the reported metrics were: Haar $$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$3, $$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$4, $$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$5; Grover with $$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$6, $$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$7, $$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$8, $$\max_{x\in\{0,1\}^n} |\langle x|U|0^n\rangle|^2 \ge n^{-a}.$$9; shallow circuits, $|0^n\rangle$00, $|0^n\rangle$01, $|0^n\rangle$02 (Kim et al., 15 Apr 2025).

These applications sharpen an important distinction from standard random circuit sampling. RCS emphasizes anti-concentration and verification by XEB or related global metrics; peaked circuits trade that for a single heavy-string witness, often at the cost of introducing exploitable structure (Movassagh, 2019, Aaronson et al., 2024).

6. Limitations, controversies, and open problems

The main controversy concerns whether peaked circuits can simultaneously satisfy near-term feasibility, strong classical hardness, and efficient verification. HQAP was introduced precisely as a candidate satisfying all three, with extrapolations suggesting years of classical runtime for the largest instances (Gharibyan et al., 29 Oct 2025). That claim was then directly challenged by a near-exact classical simulation based on mirrored MPO contraction and unswapping, which extracted the peak of the largest tested circuit in about one hour on a single GPU and argued that mirror-driven peaked circuits are generically vulnerable when obfuscation is permutation-dominated (Kremer et al., 23 Apr 2026).

Simulation limitations are family-dependent. The teleportation-inspired logical-qubit method is effective only for low-depth circuits because the logical-qubit count grows with depth; the sparse-truncated simulator has no formal guarantee of preserving $|0^n\rangle$03 and can approach dense $|0^n\rangle$04 behavior when tail mass is broadly dispersed; and MPS methods, while effective for shallow quasi-random peaked circuits, can require bond dimension $|0^n\rangle$05 in deeper instances, eliminating practical advantage (Chen et al., 2019, Ferreira, 8 Jul 2026, Udalov, 10 Aug 2025).

Several open problems recur across the literature. One is efficient generation: current constructions often rely on gradient-based peaking or expensive postselection, and the existence of a polynomial-time method that produces strongly peaked yet classically hard circuits remains unresolved (Aaronson et al., 2024, Zhang, 30 Sep 2025). Another is classical distinguishability: it is not known in general whether postselected or engineered peaked circuits can be distinguished from fully random circuits in classical polynomial time (Aaronson et al., 2024). Additional unresolved issues include analytical control of $|0^n\rangle$06 in approximate-inverse constructions, quantitative noise robustness, tighter hardness results at inverse-polynomial additive error, and peaking mechanisms based on problem-theoretic hardness rather than structural mirror identities (Udalov, 10 Aug 2025, Kremer et al., 23 Apr 2026).

In that sense, random peaked circuits form a research program rather than a settled model. They sit at the intersection of random circuit sampling, variational state engineering, tensor-network simulability, and complexity theory, and the central technical question remains whether one can plant a verifiable heavy output without simultaneously planting a classical shortcut.