--- title: Reed–Muller Distillation Factory url: https://www.emergentmind.com/topics/reed-muller-distillation-factory type: topic --- # Reed–Muller Distillation Factory A Reed–Muller distillation factory is a magic-state distillation construction whose coding-theoretic core is supplied by Reed–Muller codes, quantum Reed–Muller codes, or punctured/shortened Reed–Muller-derived CSS and triorthogonal codes. In the literature, the term encompasses at least three closely related objects: Reed–Muller code families that admit transversal non-Clifford action and therefore serve as candidate distillation ingredients; explicit distillation modules such as \(15\ket T\to\ket T\), \(64\ket T\to\ket{CCZ}\), and \(127\ket T\to\ket T\) routines; and architectural realizations in which the relevant stabilizer measurements are made geometrically local in 2D or 3D. The subject therefore sits at the intersection of algebraic coding theory, transversal-gate constructions, and factory-level resource optimization [2305.06423] [2605.06284] [2510.10852] [1205.3104]. ## 1. Conceptual scope and factory model In the code-theoretic sense, Reed–Muller constructions are relevant because transversal \(T\)-type operations act nontrivially on suitable Reed–Muller codes, so the codes can be used as distillation or checking gadgets for noisy \(T\) states, or for producing \(CCZ\) states [2605.06284]. In the stabilizer-code sense, a CSS-T code is a CSS code \(CSS(C_1,C_2)\) built from binary linear codes \(C_2\subseteq C_1\) such that the code space is preserved by transversal physical \(T\) and \(T^\dagger\), making it a natural ingredient for implementing logical non-Clifford resources [2305.06423]. In the protocol sense, a factory is a recursively concatenated family of triorthogonal-code modules. For a triorthogonal code with parameters \([[n,k,d]]_p\), the routine consumes \(n\) noisy input magic states and outputs \(k\) improved ones, with leading-order suppression \[ \epsilon_{\rm out}=O(A_d \epsilon_{\rm in}^d), \] where \(A_d\) is the number of weight-\(d\) logical \(Z\)-type operators [2510.10852]. After \(z\) concatenation levels, \[ \epsilon_{\rm out}=O\!\left(A_d^{\frac{d^z-1}{d-1}}\epsilon_{\rm in}^{d^z}\right) \] from \[ \frac{n^z}{k^{z-1}} \] noisy magic states, and the asymptotic overhead scales as \[ O\!\left(\log^\gamma\!\frac{1}{\epsilon_{\rm out}}\right), \qquad \gamma=\frac{\log(n/k)}{\log d}. \] This gives a precise operational meaning to “factory”: an iterated family of distillation blocks whose performance is governed by \(n\), \(k\), \(d\), \(A_d\), and postselection [2510.10852]. A common misconception is to identify every Reed–Muller transversal-\(T\) code with a complete factory protocol. That identification is too strong. Some works provide explicit distillation maps and leading-order error formulas, whereas others provide only the code-family input that such a factory would require. In particular, the CSS-T Reed–Muller analysis of [2305.06423] is explicitly coding-theoretic rather than protocol-theoretic. ## 2. Reed–Muller algebraic foundations The binary Reed–Muller code used in the CSS-T setting is \[ \RM(r,m)=\{ev(f)\mid f\in \mathbb F_2[x_1,\dots,x_m],\ \deg(f)\le r\}, \] with parameters \[ \RM(r,m)\text{ is }[2^m,\ \sum_{i=0}^r \binom{m}{i},\ 2^{m-r}], \] nesting \[ \RM(r_2,m)\subseteq \RM(r_1,m)\qquad\text{if }r_2\le r_1, \] and duality \[ \RM(r,m)^\perp=\RM(m-r-1,m) \] [2305.06423]. The local-unfolding literature uses a quantum Reed–Muller notation \[ QRM_m(q,r), \] defined on \(n=2^m\) physical qubits, with \(X\)-stabilizer group \(RM_m(q)\) and \(Z\)-stabilizer group \(RM_m(m-r-1)\). Equivalently, the \(X\) stabilizers are generated by subcubes of dimension \(m-q\), and the \(Z\) stabilizers by subcubes of dimension \(r+1\). A subset of vertices is represented by \[ V_P=\{v\in \mathbb{F}_2^m\mid P(v)=1\}, \] with associated Pauli operators \[ Z(V)=\prod_{v\in \mathbb{F}_2^m}(Z_v)^{v\in V},\qquad X(V)=\prod_{v\in \mathbb{F}_2^m}(X_v)^{v\in V} \] [2605.06284]. For prime dimension \(p\), the classical Reed–Muller code generalizes to \[ RM_p(r,m)=\{ f : \deg(f)\le r,\; f\in \mathbb F_p[x_1,\dots,x_m]/\langle \dots, x_i^p-x_i,\dots\rangle\}, \] whose codewords are evaluations \[ {\rm ev}(f)=(f(\vec v): \vec v\in \mathbb F_p^m)\in \mathbb F_p^{p^m}. \] Its dual is \[ RM_p(r,m)^\perp = RM_p(\tilde r,m), \qquad \tilde r = m(p-1)-r-1 \] [2510.10852]. A distinct odd-prime construction uses shortened first-order \(d\)-ary Reed–Muller codes to define a one-logical-qudit CSS code \[ \mathcal{QRM}_d(m) \] on \[ n=d^m-1 \] physical qudits, with \[ \mathcal{L}_X=\mathcal{RM}_d^*(1,m), \qquad \mathcal{L}_Z=[\mathrm{span}(\mathcal{L}_X,\mathbf{1})]^\perp \] [1205.3104]. This family underlies explicit qutrit and ququint distillation protocols rather than only asymptotic code families. ## 3. Transversal non-Clifford structure For binary CSS-T codes, the defining conditions are that \(C_2\) is an even code and that for each codeword \(x\in C_2\), there exists a self-dual code in \(C_1^\perp\) supported on \(x\) [2305.06423]. A structural criterion used in the analysis is: > An \([n,k]\) binary linear code \(C\) contains a self-dual code if and only if \(n\) is even and \(C^\perp\) is self-orthogonal, meaning that \(C^\perp\subseteq C\). Combined with puncturing/shortening duality, \[ short(C,I)^\perp=punct(C^\perp,I), \] this yields the equivalent support condition \[ punct(C_1,\mathbb F_2^m\setminus \operatorname{supp}(x)) \subseteq short(C_1^\perp,\mathbb F_2^m\setminus \operatorname{supp}(x)) \qquad\forall x\in C_2 \] [2305.06423]. The central Reed–Muller CSS-T characterization takes \[ C_1=\RM\!\left(\frac{m-1}{2}-t,\ m\right),\qquad C_2=\RM(r_2,m), \] and proves that \(CSS(C_1,C_2)\) is a CSS-T code if and only if \[ r_2\le 2t+1 \quad\text{for \(m\) even,} \] or \[ r_2\le 2t \quad\text{for \(m\) odd.} \] The proof proceeds by translating the support condition into a puncturing/shortening inclusion and then using degree bounds for products of polynomials [2305.06423]. In the prime-dimensional triorthogonal setting, the exact algebraic criterion is \[ RM_p(r,m)\ \text{is a classical triorthogonal space if}\ 3r < m(p-1). \] Given a triorthogonal generator matrix \(G\), one punctures \(k\) coordinates to obtain \(G'\), shortens on the same coordinates to obtain \(G_0\), and forms \[ CSS(G_0\to X,\; G'{}^\perp\to Z). \] If the surviving submatrix has full rank, then the resulting CSS code has length \(n_0-k\), dimension \(k\), and in the nondegenerate case considered, \[ d=\min(d(G_0^\perp), d(G'))=d(G_0^\perp) \] [2510.10852]. For odd prime dimension \(d\), the transversal structure is expressed in terms of a diagonal non-Clifford gate \(M\in\mathcal{M}_d^m\). An \(\mathcal{M}_d^m\)-distillation code satisfies \[ M^{\otimes n}\Pi (M^{\otimes n})^\dagger = M_L^\dagger \Pi M_L, \] with logical operators \[ X_L = X[\mathbf{1}],\qquad Z_L = Z[(d-1)\mathbf{1}] \] [1205.3104]. This transversality is the mechanism that turns a Reed–Muller code into a distillation primitive. The 3D local “rubik’s code” \(QRM_6(1,2)\) makes this logical action particularly explicit. It encodes 15 logical qubits indexed by 2-element subsets of \(\{1,\dots,6\}\), and a transversal \(T\) or \(\widetilde T\) on all 64 qubits acts logically as the product of all \(CCZ\) gates on triples of logical qubits that partition \(\{1,\dots,6\}\): \[ \widetilde T_A \equiv \prod_{\{I,J,K\}\text{ partition }\{1,\dots,6\}} CCZ_{I,J,K} \] [2605.06284]. ## 4. Distillation modules and representative constructions The Reed–Muller factory literature contains both one-output and multi-output modules. In odd prime dimension, one round of the protocol takes \(n\) copies of a noisy single-qudit state, twirls each under the Clifford \(C_M\), measures the \(Z\)-type stabilizers, applies a Clifford correction conditioned on those outcomes, measures the \(X\)-type stabilizers and postselects on the trivial syndrome, then decodes to one qudit [1205.3104]. In the punctured-triorthogonal setting, the design recipe is: choose a triorthogonal \(RM_p(r,m)\), puncture a set \(S\subset \mathbb F_p^m\), shorten on the same coordinates, form the CSS code, and use its transversal \(T\) to distill \(k\) magic states from \(n\) noisy ones [2510.10852]. | Module | Map or parameters | Reported leading performance | |---|---|---| | Punctured \(QRM_4(1,1)\) | \([[15,1,3]]\)-type Reed–Muller \(T\)-state factory | \(35p^3\) | | \(QRM_6(1,2)\) | \(64\ket T \to 15\ket{CCZ}\) | \(10416\,p^4\) | | Big unfolded code | \(64\ket T \to \ket{CCZ}\) | \(8256\,p^4\) | | Punctured \(QRM_7(2,2)\) | \(127\ket T \to \ket T\) | \(11811\,p^7\) | | \(\mathcal{QRM}_3(2)\) | \(n=3^2-1=8\) qutrit code | \(\epsilon' \le K\epsilon^2\) | | \(\mathcal{QRM}_5(1)\) | \(n=5^1-1=4\) ququint code | \(\epsilon' \sim \frac{3}{2}\epsilon^2+\frac{7}{2}\epsilon^3+O(\epsilon^4)\) | For the local binary factories, the leading-order rule is \[ p_{\text{out}} = m\, p_{\text{in}}^{d_Z}, \] where \(d_Z\) is the \(Z\)-distance and \(m\) is the number of minimum-weight nontrivial \(Z\) logical operators [2605.06284]. Specific values reported are: punctured \(QRM_4(1,1)\), \(d_Z=3\), \(m=35\); \(QRM_6(1,2)\), \(d_Z=4\), \(m=10416\); the big unfolded code, \(d_Z=4\), \(m=8256\); and punctured \(QRM_7(2,2)\), \(d_Z=7\), \(m=11811\) [2605.06284]. For odd-prime protocols, the qutrit code \(\mathcal{QRM}_3(2)\) has threshold \[ \epsilon^*=0.20015 \] for all noise types and \[ \epsilon_{\mathrm{dep}}^*=0.211001 \] for depolarizing noise, while the ququint code \(\mathcal{QRM}_5(1)\) has \[ \epsilon^*=0.31195 \] for generic noise and \[ \epsilon_{\mathrm{dep}}^*=0.3631226 \] for depolarizing noise [1205.3104]. The ququint update under depolarizing noise is given exactly by \[ \epsilon' = \frac{\epsilon^2(96-160\epsilon+75\epsilon^2)} {64-256\epsilon+480\epsilon^2-400\epsilon^3+125\epsilon^4}. \] The prime-\(p\) punctured Reed–Muller program emphasizes multi-output modules and asymptotic overhead rather than only one-output routines. Its most practically striking searched example is the ququint code \[ [[519,106,5]]_5 \] with \[ \gamma=0.99. \] The same table gives \[ A_d = 2180, \] and with \(\delta_{\rm in}=10^{-3}\), one round gives \[ \delta_{\rm out}\approx 8\times 10^{-18}, \] with distillation cost \[ C=\frac{n}{\bar n_T}\approx 7.4 \] [2510.10852]. This suggests that search-optimized puncturing can materially outperform analytically convenient puncturing rules at finite size. ## 5. Geometric locality and unfolding A major recent development is the conversion of Reed–Muller distillation constructions into geometrically local factories by “unfolding” the hypercube description of the code into 2D or 3D layouts in which a basis of the \(Z\)-stabilizer group becomes local [2605.06284]. The underlying algebraic mechanism is a product decomposition of polynomial subspaces associated with subcube types, rather than an ad hoc layout search. For \(QRM_4(1,1)\), two cube coordinates are grouped into one planar axis and the other two into the second planar axis. For \(QRM_6(1,1)\), the coordinates are split as \(\{1,2,3\}\) and \(\{4,5,6\}\), producing a planar \(7\times 7\) structure; 49 local square checks are obtained from the product grid, and 8 more checks are appended to span the full 57-dimensional \(Z\)-stabilizer group [2605.06284]. The Gray-code basis used for \((\mathbb{F}_2[X,Y,Z])_{\le 2}\) is \[ \big((Y+1)(Z+1),\; X(Z+1),\; Y(Z+1),\; (X+1)Y,\; (Y+1)Z,\; XZ,\; YZ\big). \] For the 64-qubit “rubik’s code” \(QRM_6(1,2)\), the coordinates are grouped as \[ x\leftrightarrow \{1,2\},\qquad y\leftrightarrow \{3,4\},\qquad z\leftrightarrow \{5,6\}. \] The 27 bulk cubes correspond to the little cubes of a Rubik’s cube, and 15 more local cubes are placed on the boundary faces, giving 42 \(Z\)-stabilizer generators in total, matching \[ \binom66+\binom65+\binom64+\binom63=42. \] According to the introduction, the 3D layout uses **64 data qubits plus 42 additional qubits** [2605.06284]. The “big unfolded code” interpolates between \(QRM_6(1,2)\) and \(QRM_6(1,1)\): it has the same \(X\)-stabilizer group as \(QRM_6(1,2)\), but its \(Z\)-stabilizer generators are chosen as a subset of the square checks from \(QRM_6(1,1)\). Three omitted \(Z\)-stabilizer generators become three logical \(Z\)’s, corresponding to logical qubits of types \(\{1,2\},\{3,4\},\{5,6\}\), and the resulting logical action is \[ \widetilde T_{\langle \{1,2,3,4,5,6\}\rangle} \equiv CCZ_{\{\{1,2\},\{3,4\},\{5,6\}\}}. \] This yields a \(64\ket T\to\ket{CCZ}\) distillation factory [2605.06284]. For \(QRM_7(2,2)\), the coordinates are grouped as \[ x\leftrightarrow \{1,2,3\},\qquad y\leftrightarrow \{4,5,6\},\qquad z\leftrightarrow \{7\}, \] and the construction yields 99 generators total, matching \[ \binom77+\binom76+\binom75+\binom74+\binom73=99. \] The 3D local implementation uses **127 data qubits plus 152 additional qubits** [2605.06284]. A common misunderstanding is that unfolding makes the entire stabilizer structure local in a symmetric way. The paper is explicit that locality is achieved for a **basis of the \(Z\)-stabilizer group**, while nonlocal \(X\)-stabilizers may remain acceptable, especially for biased-noise hardware such as cat qubits [2605.06284]. ## 6. Asymptotic behavior, overhead tradeoffs, and limitations The asymptotic behavior of Reed–Muller factory families depends strongly on which Reed–Muller formalism is used. For the binary CSS-T construction, the exact CSS dimension is \[ k= \sum_{i=0}^{\frac{m-1}{2}-t}\binom{m}{i} - \sum_{i=0}^{r_2}\binom{m}{i}, \] the block length is \[ n=2^m, \] and the CSS lower bound on distance simplifies to \[ d\ge 2^{r_2+1} \] [2305.06423]. The asymptotic Reed–Muller rate satisfies \[ \lim_{m\to\infty} R(\RM(r(m),m)) = \Phi\!\left( \lim_{m\to\infty}\frac{2r(m)-m}{\sqrt m} \right), \] and for \[ r_1(m)=\frac{m-1}{2}-t(m) \] the CSS-T family rate tends to \[ \Phi(-2c), \qquad c=\lim_{m\to\infty}\frac{t(m)}{\sqrt m}, \] with \[ \Phi(-2c)\le \frac12. \] The maximum rate of a CSS-T code defined by Reed–Muller codes is therefore \[ \frac12 \] [2305.06423]. This same work proves an important limitation: a family of classical Reed–Muller codes cannot have both nonvanishing rate and nonvanishing relative distance, and CSS codes built only from nested Reed–Muller codes therefore have vanishing quantum relative distance [2305.06423]. For distillation-factory interpretation, that limitation does not eliminate usefulness, because the minimum distance can still diverge. If \(t(m)=\Theta(\sqrt m)\), then the asymptotic rate remains nonzero and \(r_2(m)\) can scale like \(O(\sqrt m)\), yielding a diverging distance lower bound while \(k/n\) stays bounded away from zero [2305.06423]. A plausible implication is that these families are best viewed as candidates for constant-overhead magic-state distillation rather than asymptotically good quantum memories. For punctured prime-dimensional Reed–Muller factories, the central yield parameter is \[ \gamma = \frac{\log(n/k)}{\log d}, \] and the paper derives sublogarithmic magic-state cost \[ O\!\left(\log^\gamma(\epsilon^{-1})\right) \] for all prime dimensions \(p\) [2510.10852]. In the analytically tractable Manhattan-weight puncturing family, \[ n=\binom{m}{>w}_p,\qquad k=\binom{m}{\le w}_p,\qquad d=\Delta_p(m,\tilde r,w), \] with \[ \tilde r = m(p-1)-r-1,\qquad 3r