Anisotropic Tanner-Graph Structure

• Anisotropic Tanner-graph structure is defined by using two interlinked Cayley-type graphs with swapped inner codes that create non-symmetric, direction-dependent constraints.
• It underpins both classical and quantum error-correcting codes by leveraging distinct local constraints to achieve strong spectral expansion and improved minimum distances.
• The construction utilizes Ramanujan graph properties and explicit parity-check systems, delivering practical enhancements in code performance and local testability.

An anisotropic Tanner-graph structure is an explicit framework underpinning the construction of certain quantum and classical error-correcting codes—particularly Quantum Tanner codes—where distinct roles are assigned to multiple directions within the underlying graph, leading to fundamentally non-symmetric (anisotropic) properties. In these constructions, “anisotropy” arises from the deliberate swapping of inner codes along two intertwined yet structurally different graphs defined over the same set of “bit positions” (squares), leading to advantageous code performance, including high minimum distance and efficient local testability (Leverrier et al., 2022).

1. Underlying Graph Structures

The core of an anisotropic Tanner-graph structure is the use of two Cayley-type graphs, denoted $\Gamma_L$ and $\Gamma_R$, defined over a finite non-Abelian group $G$ and two symmetric generating subsets $A = A^{-1}$ and $B = B^{-1}$ (with $|A| = |B| = A$). The construction forms a “square complex” in which each square

$q_{g,a,b} = \{ (g,0), (ag,1), (gb,1), (agb,0) \}$

is indexed by $(g, a, b) \in G \times A \times B$. The set $Q$ of all such squares serves as the global coordinate (bit) set for the codes.

Two disjoint copies of $G$ are defined:

$V_0 = G \times \{0\}, \qquad V_1 = G \times \{1\},$

so the total vertex set is $V = V_0 \sqcup V_1$. The left graph $\Gamma_L$ is defined on $V_0$ and the right graph $\Gamma_R$ on $V_1$, both with the same edge set $Q$. For $v = (g,0) \in V_0$, $\Gamma_L(v)$ comprises all $A^2$ incident squares; similarly for $\Gamma_R(w)$, $w \in V_1$.

If $Cay(G, A)$ and $Cay(G, B)$ are Ramanujan graphs, then both $\Gamma_L$ and $\Gamma_R$ exhibit strong spectral expansion, with second eigenvalue bounded by $O(\sqrt{A}) \ll A$.

2. Classical Tanner Codes on the Graphs

Each vertex in $\Gamma_L$ is assigned an inner code $C^0 = C_A \otimes C_B \subset F_2^{A \times B}$, while each vertex in $\Gamma_R$ uses the swapped interleaved code $C^1 = C_B \otimes C_A \subset F_2^{A \times B}$. Here, $C_A \subset F_2^A$ and $C_B \subset F_2^B$ are classical codes with parameters $[A, k_A, d_A]$ and $[A, k_B, d_B]$ respectively.

The local parity-check matrices are formed as block matrices:

$$H^0 = \begin{pmatrix} H_A \otimes I_{|B|} \ I_{|A|} \otimes H_B \end{pmatrix}, \qquad H^1 = \begin{pmatrix} H_B \otimes I_{|A|} \ I_{|B|} \otimes H_A \end{pmatrix},$$

where $H_A$ and $H_B$ are the parity-check matrices of $C_A$ and $C_B$.

The global parity-check matrices $H_L$ and $H_R$ are constructed as direct sums over all vertices of $V_0$ and $V_1$, respectively, each summand projecting onto the $A^2$ coordinates associated with its vertex. The resulting classical codes are $C_L = \ker H_L$ and $C_R = \ker H_R$, both subspaces of $F_2^Q$, and both are LDPC with constant-weight rows and columns.

3. Manifestation of Anisotropy

Anisotropy arises from the manner in which the row and column codes are swapped between the left and right graphs. At each $v \in V_0$, the local code enforces $C_A$ row constraints and $C_B$ column constraints; at $w \in V_1$, this is reversed. Explicitly:

• Along directions where $a \in A$ varies ($b$ fixed): on $\Gamma_L$, the local constraint is $C_A$, on $\Gamma_R$ it is $C_B$.
• Along directions where $b \in B$ varies ($a$ fixed): on $\Gamma_L$, the constraint is $C_B$, on $\Gamma_R$ it is $C_A$.

The relative minimum distances $\delta_A = d_A / A$ and $\delta_B = d_B / A$ may differ, producing “strong” and “weak” directions as a function of which code (and thus which minimum distance) is enforced along which set of edges.

4. Implications for Code Distance and Local Testability

This inherent anisotropy directly impacts global code properties:

• A nonzero $x \in C_R$ of weight below approximately $n/A^{3/2-\varepsilon}$ can be eliminated via local adjustment at $V_0$ vertices, with iterative application ensuring a linear minimum distance $d_{\min} = \Omega(n/A^{3/2+\varepsilon})$.
• The local tester for the code—a two-stage process when $C_A = C_B$ (as in Dinur–like LTCs)—benefits from anisotropy: the number of rejections satisfies $(\# \text{rejects}) \geq K \cdot d(x, C_L)$ for $K = \Omega(\delta_A + \delta_B - \lambda)$, reflecting distinct “strengths” in the two directions and explicitly capturing anisotropic influences.

5. Synthesis with Quantum Coding Theory

The anisotropic Tanner-graph structure supports the CSS (Calderbank-Shor-Steane) quantum code construction: Because all inner-products $H_L H_R^T = 0$ from the graph and code properties, the pair $(C_L, C_R)$ defines a quantum code manifesting both high minimum distance and desirable LDPC characteristics. This structure simplifies prior quantum LDPC code constructions, yielding improved minimum distance scaling and explicitly connecting the design with local testability properties central in classical LTC literature (Leverrier et al., 2022).

6. Summary Table: Structure and Anisotropy

Graph	Vertex Set	Inner Code Assignment
$\Gamma_L$	$V_0$	$C_A \otimes C_B$
$\Gamma_R$	$V_1$	$C_B \otimes C_A$

A key feature is that both graphs act on the same “edges” $Q$, but impose fundamentally different (anisotropic) constraints according to the direction—an arrangement essential for the proofs of global minimum distance and local testability. All construction details (incidence matrices $P_v$, Kronecker products, expansion parameters, and parity-check systems) are explicit and central to the anisotropic Tanner-graph framework (Leverrier et al., 2022).