Short remarks on shallow unitary circuits

(i) We point out that every local unitary circuit of depth smaller than the linear system size is easily distinguished from a global Haar random unitary if there is a conserved quantity that is a sum of local operators. This is always the case with a continuous onsite symmetry or with a local energy conservation law. (ii) We explain a simple algorithm for a formulation of the shallow unitary circuit learning problem and relate it to an open question on strictly locality-preserving unitaries (quantum cellular automata). (iii) We show that any translation-invariant quantum cellular automaton in $D$-dimensional lattice of volume $V$ can be implemented using only $O(V)$ local gates in a staircase fashion using invertible subalgebra pumping.

💡 Research Summary

The paper investigates three interrelated aspects of shallow quantum circuits that respect a conserved quantity. First, it establishes a simple yet rigorous protocol showing that any local unitary circuit whose depth is smaller than the linear size of the system can be efficiently distinguished from a global Haar‑random unitary whenever there exists a conserved charge Q that is a sum of local operators (e.g. a continuous on‑site U(1) symmetry or a local energy conservation law). By preparing an initial state with a non‑zero expectation value of Q on a region A, applying a symmetric Haar‑random unitary on A∪B, and then applying the unknown circuit W of depth ≤c n, one measures Q on A. Because a shallow symmetric circuit factorises as W = W_BC W_AB, the randomisation on AB leaves the expectation of Q_A at ≈q/2, whereas a fully symmetric Haar unitary spreads the charge over the whole system and yields ≈q/4. The difference is O(q) and can be resolved with O(log 1/δ) repetitions, implying that any ensemble of symmetric shallow circuits that forms an approximate symmetric k‑design with k≥1 must have depth at least linear in the system size.

Second, the authors present an explicit learning algorithm for shallow circuits. They exploit the identity U⊗U† = (U⊗1) S (U†⊗1) S, where S swaps the system with an auxiliary copy. Since S decomposes into commuting two‑site swaps S_i,i′, learning U reduces to learning the conjugated swaps (U⊗1) S_i,i′ (U†⊗1). By preparing local density matrices ρ supported on a small region, applying U, and performing full tomography on U ρ U†, one can reconstruct U P U† for any Pauli‑type operator P. Because the map P↦U P U† is an algebra homomorphism, it suffices to learn the images of two generators (X and Z), after which the entire Pauli algebra is known. The tomography cost depends only on the circuit depth (the spread of U) and not on the total system size; parallelisation across well‑separated sites reduces the number of queries from O(n) to O(n/v), where v is the number of sites within the spread radius. Consequently, the algorithm outputs a shallow circuit (of comparable depth) that implements the original unitary (or its tensor product with its inverse) without requiring exponential resources.

Third, the paper addresses the gate‑complexity of translation‑invariant quantum cellular automata (QCA). A QCA is defined as an automorphism α of the local operator algebra with finite spread r: a single‑site operator X is mapped to an operator supported within distance r. Prior work showed that any such QCA on n qudits can be realized with O(n) gates. The authors give an alternative proof based on “invertible subalgebra pumping.” By inductively enlarging and then compressing a subalgebra that is invariant under α, they construct a unitary U consisting of O(V) local gates (V is the lattice volume) that implements α via conjugation, i.e. α(O)=U O U† for all operators O. In one dimension, the shift QCA—a non‑trivial example with spread = 1—can be implemented by a staircase of swap gates around a ring; the depth scales with the system size, yet the total gate count remains linear. The result generalises to D‑dimensional lattices, showing that any translation‑invariant QCA can be compiled into a circuit with gate count linear in the number of sites, regardless of its depth.

Overall, the work ties together symmetry‑protected distinguishability, efficient learning of shallow circuits, and the structural simplicity of translation‑invariant QCA. It demonstrates that conserved quantities fundamentally limit the randomness achievable by shallow circuits, provides a practical method to reconstruct such circuits using only local tomography, and clarifies that despite potentially large depth, QCA can always be compiled with a linear number of local gates. These insights have implications for quantum complexity theory, quantum simulation of many‑body dynamics, and the design of fault‑tolerant quantum architectures.