How Can Dynamical Lie Algebras Be Computed for Fixed-Gate Quantum Ansätze?

Fixed gates are not infinitesimal directions

A dynamical Lie algebra (DLA) collects the infinitesimal directions a parameterized quantum circuit can move in as its trainable parameters vary. For a circuit built from continuously parameterized gates of the form e^{-iθH}, the DLA is the Lie closure of the skew-Hermitian generators iH: the smallest real vector space containing them and closed under commutators. Expressivity, controllability, trainability and the geometry of the loss landscape all depend on which directions this algebra contains.

Hardware-efficient ansätze rarely match that template. They typically interleave local trainable rotations with fixed entangling gates such as controlled-X (CX) or controlled-Z (CZ). A fixed gate can always be written as U = e^{-iK}, but that notation does not make iK an available direction: the architecture never provides the continuous family e^{-itK} with a tunable t. Inserting iK into the generator set would describe a different control system from the one actually being run.

The paper also records why this is not a pedantic distinction. Earlier work treated fixed CZ gates as independent generators; a counterexample showed that this can produce a DLA that misrepresents the circuit's dynamics, and the original derivation was subsequently corrected to apply to arbitrary-angle controlled-Z rotations rather than fixed CZ gates. The opposite shortcut fails too: dropping the fixed gates loses their effect entirely, because each fixed block transforms the trainable generators that pass through it.

Conjugation turns fixed blocks into effective generators

The paper's central move is a rewriting of the circuit. Because applying a fixed unitary U and then a trainable gate is equivalent to conjugating that gate by U, all fixed blocks can be pushed to one side of the matrix product, at the cost of transforming the trainable generators they cross. For a single layer nothing changes and the available generators are the original set. For two layers, the second layer's generators become Ad_U(H) = U†HU; for three layers the third becomes Ad²_U(H), and so on.

This yields the paper's definition. Given an initial generator set G₀ for one trainable block and a fixed block U repeated at every layer, the trainable directions available up to depth L are the union of Ad^j_U(G₀) for j from 0 to L−1, and the depth-L DLA is the Lie closure of that union. The construction assumes the fixed block is Clifford, is identical in every layer, and that each layer reuses the same generator set with independently chosen parameter values.

A worked check supports the definition. The RZZ(θ) gate has generator Z⊗Z. Decomposed into CX gates surrounding an RZ rotation, the parameterized gate is conjugated by the surrounding CX, and the effective generator comes out again as Z⊗Z, so the decomposition preserves the DLA. The definition recovers what the original gate would give rather than inventing a new algebra.

Clifford blocks make the generator sequence repeat

The depth-dependent algebras form a filtration: each additional layer can only add directions, so the dimension never decreases, and because the ambient algebra is finite-dimensional the sequence must eventually stop growing. The paper calls the smallest depth at which it stops the Lie saturation depth. This is explicitly not the overparameterization threshold; more layers are generally needed to explore every direction, and a necessary condition for overparameterization is that the total parameter count KL be at least the DLA dimension.

For Clifford fixed blocks and Pauli-string generators, the sequence is guaranteed to repeat. Clifford conjugation maps Pauli operators to Pauli operators up to a phase, and the Pauli strings modulo phase form a finite set, so repeated conjugation by the same block must eventually return to the starting class. Beyond that depth no new effective directions appear. The period for a particular generator set may be shorter than the global period of the block.

The practical value is that the period can be computed rather than simulated. Encoding Pauli classes as binary vectors over F₂ turns Clifford conjugation into multiplication by an invertible binary matrix M_U, and the global conjugation period T_U is the order of that matrix. The minimal common period of the generator set divides T_U. The paper derives simplified matrix forms for blocks built only from CX gates, where the period reduces to the order of one n×n block, and for blocks built only from CZ gates, where any nontrivial block has order two.

A ring of CX gates yields explicit periods

The paper applies the framework to an ansatz of RY rotations on n qubits with a fixed ring of CX gates. Because the block is built from CX gates, its binary matrix is block-diagonal and the conjugation period equals the order of an n×n matrix B_n describing the action on the X-component. The paper derives B_n explicitly and shows its characteristic polynomial is χ_n = (x+1)^n + x over F₂.

The order then follows from factoring that polynomial. The paper proves the minimal and characteristic polynomials coincide, so the period is the least common multiple of the multiplicative orders associated with the distinct irreducible factors. For n = 2 the polynomial x² + x + 1 is irreducible and the period is 3; for n = 3, x³ + x² + 1 is irreducible and the period is 7. For n = 5 the polynomial factors as (x² + x + 1)(x³ + x + 1), giving a period of lcm(3,7) = 21.

The table of results from n = 2 to n = 8 separates two quantities that are easy to conflate. The conjugation period T_U grows quickly, reaching 127 at n = 7 and 63 at n = 8, but the Lie saturation depth is far smaller: the paper reports it as 4 when n is a multiple of 3 and 2 otherwise, and notes this is checked rather than proved. The paper also reports that the DLA is so(2^n) with dimension 2^{n−1}(2^n − 1) for every n checked up to 8, and considers it very likely true for all n.

Saturation depth is not the depth you need

The distinction matters because the two depths answer different questions. The Lie saturation depth says when the algebra stops growing; the overparameterization condition says how many parameters are needed to fill it. With K = n parameters per layer, the necessary condition KL ≥ dim g becomes a lower bound on depth of 2^{n−1}(2^n − 1)/n. That corresponds to at least 3 layers for n = 2, 10 for n = 3, 30 for n = 4 and 4080 for n = 8.

The paper also connects the algebra's size to trainability. Under the assumptions cited from prior work, the variance of the cost function is inversely proportional to the DLA dimension. Since the dimension here grows as O(4^n), the ansatz is susceptible to barren plateaus in the sufficiently deep regime. The same exponential dimension that makes the algebra expressive also makes the required depth grow rapidly with qubit count.

The method covers Clifford blocks and Pauli generators

The construction is stated for homogeneous fixed-block ansätze: the same fixed unitary is repeated at every layer, the trainable block reuses one generator set, and the fixed block is Clifford. The period results rely on the generators being Pauli strings, which keeps conjugated generators inside the finite Pauli group modulo phase. Extending the approach to non-Clifford fixed gates is listed as future work, not as a result.

The paper also proposes a comparison it did not run: benchmarking two circuits with the same parameterized part but different fixed Clifford blocks, one built from CX gates and one from CZ gates, to see how the choice of entangling block changes the effective generators and the resulting algebra. That comparison is a stated perspective, so the paper does not establish which block is preferable.

Within its scope, the paper's priority is a change of procedure rather than a new circuit. For an ansatz with trainable Pauli rotations and a repeated fixed Clifford block, the DLA should be built from the finite set of conjugated generators, with the period obtained from the order of a binary symplectic matrix, instead of from the fixed gates themselves. The CX-ring example shows why both numbers are needed: a small saturation depth of 2 or 4 coexists with a depth requirement that reaches 4080 layers at n = 8.

Sources used in this answer