Quantum Hamming Bound

• Quantum Hamming Bound is a finite-length sphere-packing limit that equates the code dimension with the number of correctable error patterns for exact quantum error correction.
• It differentiates between nondegenerate (pure) and degenerate codes by addressing overlaps in error sectors, which challenge the naive disjoint-sphere interpretation.
• Recent proofs extend the bound to all quantum subspace codes using linear-programming and Fourier-analytic techniques, solidifying its role in quantum coding theory.

The quantum Hamming bound is the finite-length sphere-packing count for exact quantum error correction: the code dimension times the number of correctable local error patterns must fit inside the ambient Hilbert space. For an exact $q$-ary quantum subspace code $((n,K,d))$, with $t=\lfloor(d-1)/2\rfloor$, it can be written as

$$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$

In the binary stabilizer setting $[[n,k,d]]$, this becomes

$\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$

For nondegenerate or pure codes, the bound follows from a direct counting argument on correctable errors and syndrome space. The central difficulty is degeneracy: distinct physical errors may coincide on the code subspace, so the naive disjoint-sphere picture is no longer literally valid. Whether that could invalidate the finite-length bound was a central question in quantum coding theory until recent exact results closed it in full generality (Zhang et al., 21 Jun 2026).

1. Standard formulation and sphere-packing interpretation

A quantum error-correcting code with parameters $((n,K,d))_q$ has length $n$, codespace dimension $K$, and distance $d$. The distance is the largest integer such that the code can detect any error acting nontrivially on at most $((n,K,d))$0 qudits, and the correction radius is

$((n,K,d))$1

In the $((n,K,d))$2-ary stabilizer formalism, there are $((n,K,d))$3 nontrivial local error patterns of weight $((n,K,d))$4. For pure or nondegenerate codes, distinct correctable errors must map the codespace to linearly independent subspaces, which yields the standard quantum Hamming bound

$((n,K,d))$5

For binary stabilizer codes with $((n,K,d))$6, this specializes to

$((n,K,d))$7

or equivalently

$((n,K,d))$8

(0711.4603).

The underlying interpretation is the quantum analogue of the classical sphere-packing bound. For nondegenerate codes, each logical state together with all correctable errors generates a family of linearly independent states. In the binary case, the total number

$((n,K,d))$9

cannot exceed the full Hilbert-space dimension $t=\lfloor(d-1)/2\rfloor$0, and division by $t=\lfloor(d-1)/2\rfloor$1 yields the stated inequality. In stabilizer language, the same argument says that the number of correctable Pauli errors of weight at most $t=\lfloor(d-1)/2\rfloor$2 must fit inside the syndrome space of size $t=\lfloor(d-1)/2\rfloor$3 (Dallas et al., 2022).

The bound is a necessary condition for exact correction, not an existence theorem. It limits admissible triples $t=\lfloor(d-1)/2\rfloor$4 or $t=\lfloor(d-1)/2\rfloor$5 by comparing code dimension against the volume of a radius-$t=\lfloor(d-1)/2\rfloor$6 error ball. In later work, this finite-length count is reformulated using $t=\lfloor(d-1)/2\rfloor$7-ary Hamming-scheme balls $t=\lfloor(d-1)/2\rfloor$8, which is the natural language for the exact arbitrary-dimension results (Zhang et al., 21 Jun 2026).

2. Degeneracy, purity, and the failure of naive disjointness

The conceptual obstruction to extending the standard counting proof is degeneracy. In the Knill–Laflamme condition

$t=\lfloor(d-1)/2\rfloor$9

nondegenerate means that $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$0 has full rank, while degenerate means that $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$1 is rank-deficient. Equivalently, different correctable errors need not generate linearly independent states on the code space. This is precisely the point at which the standard sphere-packing proof can fail: several physical errors may occupy the same effective error sector, so the naive count can overestimate the true amount of Hilbert-space used (Dallas et al., 2022).

A related distinction is pure versus impure. Purity concerns orthogonality of error subspaces: a code is pure with respect to a set of errors if different errors map the codespace to mutually orthogonal subspaces. A pure code must be nondegenerate, but the converse need not hold. This distinction matters because the textbook counting argument is immediate for pure or nondegenerate codes, whereas the long-standing open question asked whether impure or degenerate codes might exploit overlaps to evade the Hamming count (Dallas et al., 2022).

For stabilizer codes, degeneracy has a particularly simple structural characterization. An $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$2 stabilizer code is degenerate if and only if its stabilizer group contains a non-identity element of weight less than $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$3. Equivalently, some low-weight Pauli errors share syndromes because a nontrivial stabilizer of weight $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$4 identifies them on the code space. In this setting, the paper on degenerate stabilizer bounds also notes that “pure/impure” and “degenerate/nondegenerate” line up in the way needed for its arguments (Nemec et al., 2023).

The modern finite-length resolution does not restore literal disjoint spheres. Instead, it replaces the geometric count by a linear-programming or Fourier-analytic inequality that remains valid even when correctable error sectors overlap. This suggests that the right invariant is not disjointness itself, but a more refined control of ball intersections and Lloyd-type responses.

3. Partial results and strengthened variants before the full resolution

Before the general theorem was proved, several results established the quantum Hamming bound in important special cases. A notable early milestone showed that every $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$5 stabilizer code with $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$6 satisfies

$$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$7

which is exactly the distance-$$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$8 quantum Hamming bound, hence a proof for all double error-correcting stabilizer codes, including degenerate nonbinary ones. The proof used a Delsarte-style linear-programming argument with $$K \le \frac{q^n}{V_t^{(Q)}(n)},\qquad V_t^{(Q)}(n)=\sum_{j=0}^{t}(Q-1)^j\binom{n}{j},\qquad Q=q^2.$$9-ary Krawtchouk polynomials. The same note also derived explicit maximal-length constraints for quantum MDS stabilizer codes, including $[[n,k,d]]$0 for single error-correcting additive quantum MDS codes (0711.4603).

A different line of work strengthened the usual Hamming estimate rather than merely extending its validity. The “quantum Hamming-Singleton bound” interpolates between the Hamming bound and the quantum Singleton bound, and a further Lloyd-based refinement gives a strict improvement whenever the relevant Lloyd polynomial has a non-integer zero. In the notation of that paper, the strengthened form is

$[[n,k,d]]$1

whenever at least one Lloyd root is non-integer. For $[[n,k,d]]$2 or $[[n,k,d]]$3, the strengthened bound was proved to hold even for impure codes; for stabilizer codes, the paper also identified an infinite family of lengths

$[[n,k,d]]$4

for which the strengthened bound improves the ordinary binary stabilizer qHB by at least one logical qudit (Yu et al., 2010).

Additional progress came from combining two earlier bounds. Using Rains’ analytical bound $[[n,k,d]]$5 together with the Li–Xing threshold $[[n,k,d]]$6 beyond which the qHB must hold, it was shown that no $[[n,k,d]]$7 code with $[[n,k,d]]$8 can violate the quantum Hamming bound. This result included degenerate codes and reduced the search space for any possible counterexample to distances $[[n,k,d]]$9 (Dallas et al., 2022).

For degenerate stabilizer codes, a sharper statement was later obtained. For every fixed $\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$0, there exists a positive integer $\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$1 such that every degenerate $\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$2 stabilizer code with $\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$3 satisfies

$\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$4

This is stricter than the usual quantum Hamming bound. In the single-error-correcting case $\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$5, the inequality becomes

$\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$6

and the paper states that it holds for all degenerate $\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$7 stabilizer codes. A further consequence is that, outside explicitly listed exceptional lengths, any optimal distance-$\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$8 stabilizer code must be nondegenerate (Nemec et al., 2023).

4. Finite-length resolution for exact codes

The binary case was resolved by proving that every exact binary quantum subspace code $\sum_{j=0}^{t}3^j\binom{n}{j}\le 2^{n-k}.$9 with $((n,K,d))_q$0 obeys

$((n,K,d))_q$1

without assuming either nondegeneracy or additivity. The proof uses the Li–Xing linear-programming theorem with the choice $((n,K,d))_q$2, where

$((n,K,d))_q$3

is the Lloyd polynomial built from quaternary Krawtchouk polynomials. A central identity identifies the polynomial evaluation with an exact Hamming-ball intersection count,

$((n,K,d))_q$4

so the LP condition reduces to proving

$((n,K,d))_q$5

Monotonicity in block length reduces the problem to the shortest admissible length $((n,K,d))_q$6, after which a local node–edge charging inequality completes the proof. The conclusion is explicit: degeneracy can merge correctable error sectors, but cannot enlarge the finite-length binary Hamming bound (Zhang et al., 14 Jun 2026).

The nonbinary case was then proved for all exact $((n,K,d))_q$7-ary quantum subspace codes with $((n,K,d))_q$8. Writing $((n,K,d))_q$9, the bound takes the form

$n$0

As in the binary proof, the key quantity is the normalized two-center intersection ratio

$n$1

where $n$2 is the $n$3-ary Lloyd sum. For $n$4, the argument reduces to the critical length $n$5 and proves the stronger half-gap estimate

$n$6

The qutrit case $n$7, i.e. $n$8, requires a separate short-range, long-range, and bridge-window analysis; the bridge regime is handled by a filtered Lloyd square

$n$9

together with exact coefficient certificates and a Stein-tangent positivity argument. Combined with the independent binary endpoint theorem, this yields the quantum Hamming bound in arbitrary local dimension for exact subspace codes (Zhang et al., 21 Jun 2026).

These results change the status of the subject. The finite-length exact qHB is no longer conditional on purity, nondegeneracy, or stabilizer structure. Degeneracy remains a structural feature of codes, but not a loophole in the exact Hamming count.

5. Relation to existence bounds, approximate correction, and analogues

The quantum Hamming bound is a necessary upper bound. This should be distinguished from Gilbert–Varshamov-type results, which are sufficient existence guarantees. In the $K$0-ary stabilizer setting based on symplectic self-orthogonal codes, the standard quantum GV condition is

$K$1

whereas a recent probabilistic improvement proves existence whenever

$K$2

This is an $K$3 multiplicative improvement over the standard quantum GV bound. It does not challenge the qHB; rather, it sharpens the known existence region below the packing frontier (Yuan et al., 26 Jan 2026).

Approximate quantum error correction alters the meaning of “attaining” the Hamming limit. For a Haar random $K$4-dimensional code $K$5 and any unitary error set of size $K$6, it was shown that approximate correction is possible with disturbance essentially

$K$7

and hence whenever $K$8. This is the sense in which Haar random codes approximately attain the quantum Hamming bound: they realize a soft version of the exact nondegenerate counting condition $K$9, but only in an approximate decoding framework (Ma et al., 8 Oct 2025).

There are also structurally analogous but genuinely different bounds. For bosonic codes against displacement noise, a continuous-variable analogue replaces discrete Pauli-weight balls by phase-space balls and replaces the discrete MacWilliams transform by a Bessel/Fourier transform. The resulting “Quantum Cohn–Elkies bound” and “Quantum Levenshtein bound” constrain $d$0-QEDCs through inequalities such as

$d$1

and

$d$2

but the paper explicitly notes that this is not the standard finite-dimensional Hamming bound; the analogy is structural rather than literal (Burchards, 13 Feb 2025).

A further source of confusion is the “CSS Hamming bound” benchmark used in concatenated GKP-QLDPC decoding. In that setting, the benchmark is presented as

$d$3

and finite-rate LP-QLDPC-GKP schemes with analog information and sequential min-sum decoding are reported to surpass that CSS bound for the studied families. This concerns a different decoding model, with inner GKP analog information and an outer iterative decoder, rather than a counterexample to the exact finite-length quantum Hamming bound for arbitrary qubit errors (Raveendran et al., 2021).

6. Quantum Hamming codes and distance-$d$4 structure

The canonical distance-$d$5 family associated with the bound is the family of quantum Hamming codes

$d$6

These are described as “optimal distance-3 codes,” and are treated as the quantum analogues of classical Hamming codes, which are “perfect codes” that “achieve the theoretical Hamming bound.” In this sense, quantum Hamming codes occupy the most direct operational position relative to the qHB in the single-error-correcting regime (Shi et al., 15 Jan 2026).

Recent work on fault-tolerant syndrome extraction for this family studies not the derivation of the bound itself, but the operational consequences of the structure behind bound-saturating distance-$d$7 codes. For $d$8, the paper constructs a fault-tolerant measurement sequence for $d$9 of length exactly

$((n,K,d))$00

which is “only one additional measurement beyond the original non-fault-tolerant sequence.” The construction uses a cyclic transformation matrix $((n,K,d))$01 generated by

$((n,K,d))$02

modulo $((n,K,d))$03, preserves a self-dual CSS-like symmetry, and enables reuse of the second half of the measurement circuits by adding boundary Hadamard gates to the first half (Shi et al., 15 Jan 2026).

The distance-$((n,K,d))$04 regime is also where degeneracy is now understood most concretely. For degenerate $((n,K,d))$05 stabilizer codes, the strengthened Hamming-like inequality

$((n,K,d))$06

holds for all lengths, and the classification consequences imply that a degenerate stabilizer code can achieve the optimal number of logical qubits only for explicitly listed exceptional lengths. Outside those exceptional families, any optimal distance-$((n,K,d))$07 stabilizer code must be nondegenerate (Nemec et al., 2023).

Taken together, these results show that the quantum Hamming bound is not merely a counting lemma. It organizes the taxonomy of exact quantum codes, separates exact from approximate correction regimes, determines when degeneracy matters and when it does not, and remains the finite-length reference point against which both optimal code families and stronger LP-based refinements are measured.