---
title: Ultra-High-Rate Quantum Codes
url: https://www.emergentmind.com/topics/ultra-high-rate-quantum-codes
type: topic
---

# Ultra-High-Rate Quantum Codes

Searching arXiv for recent and foundational papers on ultra-high-rate quantum codes and related high-rate qLDPC constructions.
Ultra-high-rate quantum codes are quantum error-correcting codes designed to maximize the encoding rate $R=k/n$ while retaining useful error suppression, often with $R$ comparable to or exceeding $1/2$, so that roughly two physical qubits encode one logical qubit [2604.16209]. In adjacent parts of the literature, closely related goals appear under the labels *high-rate* or *constant-rate*, including families with explicit asymptotic bounds such as $R \ge 13/72$ or $R=0.2$ [2001.03568][2402.09648]. Across these variants, the central problem is the same: reducing qubit overhead without sacrificing decodability, circuit-level performance, or hardware realizability.

## 1. Rate, distance, and the meaning of “ultra-high-rate”

The basic metric is the encoding rate $R=k/n$, with $[[n,k,d]]$ denoting a stabilizer code of block length $n$, dimension $k$, and distance $d$. For finite-size qLDPC codes, the pressure toward high rate is immediately constrained by the quantum Singleton bound,
$$
n-k \ge 2(d-1),
$$
so pushing $R$ above $1/2$ forces modest distances at practical block lengths [2604.16209]. This is why most practical finite-size qLDPC codes studied before recent ultra-high-rate constructions operated around or below $1/10$ when tuned for useful logical performance at $n \lesssim$ a few thousand [2604.16209].

The literature does not use the term uniformly. In the reconfigurable-neutral-atom setting, *ultra-high-rate* denotes $R=k/n \ge 1/2$ and specifically targets finite-size codes such as $[[1152,580,\leq 12]]$ and $[[2304,1156,\leq 14]]$ with overheads $n/k \approx 2$ [2604.16209]. In geometric qLDPC work, *ultra-high-rate* may instead mean non-vanishing constant rate, as in hyperbolic-manifold codes with $R \ge 13/72$ [2001.03568]. Other constructions push nominal rates much higher by altering the resource model: less-noisy or entanglement-assisted LDPC schemes reach rates around $0.86$ at moderate length, while QSBC–QURC turbo constructions realize $R_Q=1-2/n$ and hence approach unity as the outer block length grows [1309.5587][1903.10927].

This suggests that the topic is best understood as a design regime rather than a single asymptotic class. Some families prioritize $R>1/2$ at finite $n$, some prioritize constant asymptotic rate with polynomial distance, and others trade standard stabilizer constraints for auxiliary resources.

## 2. Algebraic and geometric construction paradigms

A major branch of ultra-high-rate research remains within CSS qLDPC design. Hyperbicycle codes interpolate between hypergraph-product and generalized bicycle codes through a quasi-cyclic ansatz, preserve LDPC sparsity, and admit lower and upper bounds with $D=\Theta(\sqrt{N})$, but the concrete examples in the original study remain in the moderate-rate regime, roughly $0.05$–$0.28$, rather than the $R \gtrsim 1/2$ regime [1212.6703]. They are historically important because they clarify how finite rate, bounded stabilizer weight, and cyclic structure can coexist, even when the resulting rates are not yet ultra-high in the narrowest modern sense.

More recent algebraic qLDPC work reaches substantially higher rates. A two-branch finite-field base over $\mathbb{F}_{19}$ combined with a circulant-permutation-matrix lift of degree $P=101$ yields a $(3,18)$-regular CSS code with parameters $[[34542,23032,d\le 310]]$ and rate $R \approx 0.6667824677$; the Tanner graphs of $H_X$ and $H_Z$ separately have girth $8$, and decoder experiments reported no failures in $10^8$ trials at $p=0.01$, with a finite-length transition near $p=0.029$ [2606.27130]. In a different HGP direction, TGRE-hypergraph-product codes were constructed with exact asymptotic rate $R=0.2$, blocklength $N=(5/4)n^2$, logical count $K=(1/4)n^2$, distance scaling $D=O(\log n)$, and an observed code-capacity threshold around $0.096$ under depolarizing noise with FDBP decoding [2402.09648].

Geometric constant-rate qLDPC families provide another axis of comparison. Hyperbolic $4$-manifold codes based on the regular $\{5,3,3,5\}$ tessellation achieve the explicit lower bound
$$
R = k/n \ge 13/72,
$$
with polynomially growing distance and single-shot decoding behavior consistent with a phenomenological threshold around $4\%$ under syndrome noise [2001.03568]. The earlier “Golden codes,” built from the regular $\{4,3,3,5\}$ tessellation, give an explicit lower bound
$$
R \ge 17/360-o(1),
$$
together with $d=\Omega(n^{0.1})$ and a local decoder that exploits the regular hypercube structure of the tessellation [1712.08578].

| Family | Representative rate or parameters | Distinguishing feature |
|---|---:|---|
| APM CSS on atom arrays | $[[2304,1156,\le 14]]$, $R \approx 0.502$ | finite-size $R>1/2$ with hardware co-design |
| Two-branch finite-field + CPM lift | $[[34542,23032,d\le 310]]$, $R \approx 2/3$ | same-type girth $8$ |
| TGRE-HGP | $R=0.2$ | constant-rate HGP variant with threshold around $0.096$ |
| Hyperbolic $\{5,3,3,5\}$ family | $R \ge 13/72$ | single-shot decodable constant-rate qLDPC |
| Golden codes | $R \ge 17/360$ | regular $\{4,3,3,5\}$ tessellation and local decoder |

A persistent misconception is that high rate and asymptotic goodness are interchangeable. The constructions above show otherwise: some achieve $R>1/2$ at finite size with modest distance, while others keep constant rate asymptotically but at smaller constants and different geometric or decoder trade-offs.

## 3. Hardware co-design on neutral-atom platforms

The most explicit recent finite-size ultra-high-rate program is the co-design of CSS qLDPC codes for reconfigurable neutral-atom arrays [2604.16209]. In that family, both $H_X$ and $H_Z$ are assembled from $P \times P$ affine permutation matrix blocks arranged in a $3 \times 12$ block-circulant pattern, with three active block rows and twelve block columns. Each affine permutation matrix implements $x \mapsto ax+b \pmod P$ with $\gcd(a,P)=1$, data-qubit degree is approximately $6$, ancillas participate in $12$ two-qubit gates per syndrome round, and the search enforces girth $\ge 6$ rather than girth $8$ in order to preserve viable finite-size instances [2604.16209].

The distinctive contribution is not only algebraic but architectural. The construction imposes orbit-commutation relative to a reference APM with large orbits, Chinese-Remainder-Theorem separability for $P=m\ell$ with $\gcd(m,\ell)=1$, and a maximal abelian column subgroup that linearizes the column action. For $P=96$, the relevant subgroup is cyclic $\mathbb{Z}_{32}$; for $P=192$, it is $\mathbb{Z}_{32}\times \mathbb{Z}_2$ [2604.16209]. Under these constraints, each inter-block transition compiles into only $2$–$3$ product moves rather than $O(\log P)$ generic moves, and each step requires at most one horizontal cyclic column shift and one vertical row swap or shift [2604.16209].

On the hardware side, atoms are placed on a $2$D grid with data atoms spaced $12\,\mu\text{m}$ horizontally and vertically, ancillas $2\,\mu\text{m}$ from associated data atoms, and Rydberg CZ gates assumed at $1\,\mu\text{s}$ per parallel layer [2604.16209]. The syndrome-extraction schedule has CNOT depth $12$ per basis under baseline coloration, total depth $24$ for sequential $X/Z$, while Bell-pair-assisted extraction reduces this to about $7$ [2604.16209]. The measured movement times per round are $8.3$ ms and $9.9$ ms with four AOD pairs for the $P=96$ and $P=192$ codes respectively, and $13.3$ ms and $16.9$ ms with two AOD pairs [2604.16209].

These implementation constraints do not prevent strong logical performance. Under a circuit-level model with $p=0.1\%$, the $[[2304,1156,\leq 14]]$ code achieves a per-logical-per-round error rate of $1.3_{-0.9}^{+3.0}\times 10^{-13}$ and a block error per round of $1.5_{-1.0}^{+3.5}\times 10^{-10}$, corresponding to $N_{\text{ops}} \approx 7.7\times 10^{12}$ and entry into the teraquop regime. The smaller $[[1152,580,\leq 12]]$ code reaches $2.9_{-1.5}^{+3.1}\times 10^{-11}$ per logical per round and $N_{\text{ops}} \approx 3.4\times 10^{10}$ [2604.16209].

A related neutral-atom strategy uses static long-range connectivity rather than dynamic rearrangement. La-cross hypergraph-product codes on static $2$D arrays implement weight-$6$ checks with depth-$10$ syndrome rounds, and circuit-level simulations show that these codes outperform surface codes when the nearest-neighbor two-qubit gate error probability is below $\sim 0.1\%$ under a range-dependent neutral-atom model; under range-independent errors, the most non-local instance improves below $p \lesssim 0.5\%$ [2404.13010]. The contrast between the two neutral-atom programs is informative: one turns non-locality into shallow motion, the other into bounded long-range blockade gates.

## 4. Concatenated, subsystem, assisted, and specialized high-rate families

Ultra-high-rate design is not confined to LDPC constructions. Many-hypercube codes concatenate the $[[6,4,2]]$ error-detecting code to produce
$$
[[6^L,4^L,2^L]], \qquad R=(2/3)^L,
$$
with explicit instances $[[216,64,8]]$ at $L=3$ and $[[1296,256,16]]$ at $L=4$ [2403.16054]. They are not LDPC asymptotically, because check weights grow with level, but at practical sizes they deliver rates around $30\%$ and $20\%$ together with a level-by-level minimum-distance decoder whose bit-flip threshold is about $5.6\%$, and a logical CNOT threshold around $0.9\%$ in the stated circuit-level model [2403.16054]. A subsystem variant based on concatenated $[[4,2,2]]$ blocks replaces exponentially growing syndrome weight by constant measurement weight $4$, with subsystem parameters
$$
[[4^r,2^r,4^r+2^r-2\cdot 3^r,2^r]],
$$
rate $R=2^{-r}$, and an estimated bit-flip threshold around $2\%$ using neural-network decoding [2510.04526]. A later optimization showed that mixed-level families such as $D_{6,4,4}$ and $D_{6,6,4,4}$ can outperform the lower-rate all-$4$ versions at the same level, while level-$3$ encoder overhead can be reduced by about $60\%$ relative to the original design [2512.00561].

Concatenated quantum Hamming architectures give another high-rate route. For $L$ levels of the $[[15,7,3]]$ code,
$$
n_L=15^L,\quad k_L=7^L,\quad R_L=(7/15)^L,\quad d_L=3^L,
$$
and a bidirectional hard-decision decoder raises the code-capacity threshold under independent bit-flip noise from approximately $1.56\%$ to $4.35\%$ while empirically preserving the full $3^L$ distance scaling for at least $L\le 3$ [2601.09131].

Assisted formalisms push the nominal rate far higher by relaxing standard stabilizer constraints. In the less-noisy model, any binary $[n,k,d]$ code can be converted into a quantum code with parameters $[[2n-k,k,d;2(n-k)]]$, and an affine-geometry example gives $[[1242,1080,4;162]]$ with rate $1080/1242 \approx 0.8696$ [1309.5587]. In the entanglement-assisted model, using one ebit yields the example $[[1161,997,6;1]]$ with rate $\approx 0.8587$ [1309.5587]. These constructions are ultra-high-rate in the most literal sense, but the auxiliary-resource assumptions place them in a distinct operational category.

Other specialized families target particular noise or coding goals. QSBC–QURC turbo codes realize
$$
R_Q=1-2/n
$$
for outer block lengths $n=4,6,8$, giving supported rates $1/2$, $2/3$, and $3/4$, and at $R_Q=1/2$ operate at distance $D=0.029$ from the quantum hashing bound at QBER $10^{-3}$ [1903.10927]. Constant-excitation stabilizer codes immune to collective coherent $Z$ rotations were constructed with parameters
$$
[[2^{r+1},2^r-(r+1),4]],
$$
so that the rate tends to $1/2$ while retaining exact immunity to collective $Z$ rotations on the code space [2503.05249].

## 5. Decoding, single-shot operation, and fault-tolerant computation

In high-rate quantum coding, decoder design is often as decisive as the code family itself. The neutral-atom APM codes use a hierarchical three-tier decoder: T1 belief propagation on a sliding-window Tanner graph, T2 relay BP, and T3 exact integer-programming maximum-likelihood decoding on the rare tail [2604.16209]. For the $[[1152,580,\leq 12]]$ and $[[2304,1156,\leq 14]]$ codes, the escalation rates are $q_2=0.8\%, q_3=0.003\%$ and $q_2=1.3\%, q_3=0.07\%$ respectively, and FPGA/GPU acceleration reduces the average per-round decoding work to about $120$ ns and $1.0\,\mu\text{s}$, well below the $8$–$10$ ms syndrome-extraction cycle [2604.16209].

Single-shot decoding provides a complementary paradigm for constant-rate qLDPC families. In the hyperbolic-manifold codes derived from the $\{5,3,3,5\}$ tessellation, parity-check redundancy supports both a cellular-automaton decoder and BP under noisy measurements, and the data are consistent with a threshold around $4\%$ in the phenomenological model with syndrome noise, or above $5\%$ for BP with perfect measurements [2001.03568]. The existence of such decoders is significant because constant rate alone is not enough; operationally useful high-rate codes must also avoid repeated measurement overhead and decoder bottlenecks.

A recurrent criticism of high-rate qLDPC is that they are better suited to memory than to computation. Several recent works directly address that objection. Batched protocols for arbitrary CSS qLDPC codes introduce batched single-shot syndrome extraction, batched code switching, and batched addressable Cliffords with constant per-block space-time overhead, assuming fast classical computation [2510.06159]. In a near-term self-dual bivariate-bicycle implementation, the reported space-time cost per logical qubit per layer is about $1.8\times 10^5$, compared with about $6.7\times 10^5$ for a surface-code protocol and $5.7\times 10^5$ for a low-rate BB protocol [2510.06159].

Addressable logic has also been demonstrated in a small 3D-local high-rate qLDPC family. Lift-connected surface codes scale as $[[n,\mathcal{O}(n^{1/3}),\mathcal{O}(n^{1/3})]]$, and the $[[15,3,3]]$ instance provides deterministic fault-tolerant circuits for the logical Clifford set together with magic-state preparation [2511.10191]. Although the asymptotic rate vanishes, the small-instance ratio $k/n=0.2$ is high relative to comparable surface-code instances, and the reported circuit-level pseudothresholds for algorithmic FT lie in the range $4.8\cdot 10^{-3}$ to $1.2\cdot 10^{-2}$ [2511.10191].

## 6. Performance frontiers, trade-offs, and open problems

The strongest present finite-size evidence for ultra-high-rate practicality comes from teraquop-scale memory estimates. The reconfigurable-neutral-atom APM family reaches $N_{\text{ops}}\approx 7.7\times 10^{12}$ for the $[[2304,1156,\le 14]]$ code at $p=0.1\%$ [2604.16209]. A different route concatenates algebraic quantum Reed–Solomon outer codes over the gross bivariate-bicycle inner code by treating each inner block as a single Galois qudit of dimension $q=2^{11}=2048$; at uniform $10^{-3}$ physical noise, the concatenated gross code enters the teraquop regime and crosses below the two-gross code in space overhead at $4.10\times 10^{-15}$ per logical qubit-round [2605.21898]. In that architecture, list decoding, time-like RS protection against measurement errors, and spacetime error mixing compensate for the correlated logical failures inside each high-rate inner block [2605.21898].

The trade-offs are correspondingly sharp. Finite-size $R>1/2$ families generally have modest distances, often in the low teens, because of the Singleton constraint [2604.16209]. Constant-rate qLDPC families with stronger asymptotic guarantees usually operate at lower rates, larger blocklengths, or both [2001.03568][1712.08578]. Some constructions keep bounded check weight but accept moderate rates, as in hyperbicycle and many bivariate-bicycle implementations; others keep high rate but pay with heavier checks, such as TGRE-HGP, where the maximum stabilizer weight scales as $O(\sqrt{N})$ [1212.6703][2402.09648]. Assisted codes achieve the most aggressive nominal rates, but only by assuming phase-only ancillas, ebits, or analogous auxiliary structure [1309.5587].

Several open problems recur across the literature. One is how to raise distance at similar rate and size; this is explicit in the neutral-atom APM work, where $d\le 14$ is already close to the finite-size ceiling imposed by high rate [2604.16209]. A second is how to model realistic hardware noise beyond idealized depolarizing channels, including idling, atom loss, correlated faults, and movement-induced errors [2604.16209]. A third is certification: the rate-$2/3$ CPM-lifted code reports an explicit upper bound $d\le 310$, but no nontrivial lower bound is claimed, and symmetry-reduced logical-operator searches remain open [2606.27130]. A fourth is computational overhead: many constant-overhead logical-gate results assume classical decoding latency far below the code cycle time [2510.06159].

Ultra-high-rate quantum codes therefore do not define a single architecture but a spectrum of strategies for compressing the physical-to-logical overhead of fault-tolerant quantum information. The current landscape ranges from finite-size $R>1/2$ qLDPC codes co-designed with reconfigurable atom arrays, through constant-rate hyperbolic and product-code families, to concatenated, assisted, and specialized constructions that exploit nonlocality, auxiliary qubits, or tailored noise structure. What unifies them is the attempt to replace the conventional high-overhead paradigm with codes whose logical density is large enough to matter at experimentally relevant scales, while still admitting concrete syndrome extraction, decoding, and fault-tolerant control.

Source: https://www.emergentmind.com/topics/ultra-high-rate-quantum-codes