↵ open · Esc close
Home/ Papers/ Barren Plateaus & DLA Theory
2024–2025 Theory

Barren Plateaus via Dynamical Lie Algebras

Ragone, Bakalov, Sauvage, Larocca, Cerezo et al.
Nature Communications (2024) · arXiv:2309.09342
Review: Larocca, Thanasilp, Wang, Cerezo et al. — arXiv:2405.00781 (2024, updated 2025)

Contents

  1. What are Barren Plateaus?
  2. Formal Definition
  3. Dynamical Lie Algebras
  4. Main Theorem: DLA & Variance Bound
  5. Architecture Diagram
  6. Types of Barren Plateaus
  7. Mitigation Strategies
  8. Expressibility-Trainability Tradeoff
  9. Open Questions

1. What are Barren Plateaus?

Barren plateaus (BP) are regions of the parameter landscape of parameterised quantum circuits (PQCs) where the gradient of the cost function vanishes exponentially with system size $n$. Concretely, the gradient norm is exponentially small almost everywhere, making gradient-based optimisation of PQCs intractable for large $n$.

Practical consequence: On an $n$-qubit circuit, the gradient magnitude is $O(2^{-n})$, meaning you need $O(2^{2n})$ measurements just to resolve the sign of the gradient — exponentially more than what any classical or quantum device can provide efficiently.

Originally discovered by McClean et al. (2018) for random circuits, barren plateaus have been shown to afflict: global cost functions, deep circuits, hardware-efficient ansätze, quantum kernel methods, and even some problem-inspired ansätze. The DLA framework provides a unified theory of why they occur.

2. Formal Definition

Consider a PQC $U(\boldsymbol{\theta}) = \prod_k e^{-i\theta_k H_k}$ and cost function:

$$\mathcal{L}(\boldsymbol{\theta}) = \langle \psi_0 | U^\dagger(\boldsymbol{\theta})\, O\, U(\boldsymbol{\theta}) | \psi_0 \rangle$$
Eq. (1) — Variational cost function for observable $O$

The gradient with respect to parameter $\theta_\mu$ is:

$$\partial_\mu \mathcal{L} = -i\langle \psi_0 | U^\dagger [H_\mu,\, U_\mu^\dagger O\, U_\mu] U | \psi_0 \rangle$$
Eq. (2) — Gradient via commutator (parameter-shift derivation)

Barren plateau definition: A PQC exhibits a barren plateau if:

$$\text{Var}_{\boldsymbol{\theta}}\!\left[\partial_\mu \mathcal{L}(\boldsymbol{\theta})\right] \in O\!\left(\frac{1}{2^n}\right)$$
Eq. (3) — Variance decays exponentially in the number of qubits $n$

By Chebyshev's inequality, this means $|\partial_\mu \mathcal{L}|$ is exponentially small with high probability:

$$\Pr\!\left[|\partial_\mu \mathcal{L}| \geq \epsilon\right] \leq \frac{\text{Var}[\partial_\mu \mathcal{L}]}{\epsilon^2} \in O\!\left(\frac{1}{\epsilon^2 \cdot 2^n}\right)$$
Eq. (4) — Probability of large gradient vanishes exponentially

3. Dynamical Lie Algebras

The Dynamical Lie Algebra (DLA) $\mathfrak{g}$ of a PQC is the Lie algebra generated by the circuit's generators $\{iH_k\}$ under the commutator:

$$\mathfrak{g} = \text{Lie}\!\left(\{iH_k\}_{k=1}^{K}\right) = \text{span}\!\left\{[iH_j, iH_k],\, [[iH_j, iH_k], iH_l],\, \ldots\right\}$$
Eq. (5) — DLA as iterated commutator closure of generators

The DLA determines the reachable set of the circuit:

$$\mathcal{M} = \{e^A \mid A \in \mathfrak{g}\} \subseteq SU(2^n)$$
Eq. (6) — Reachable manifold (Lie group corresponding to DLA)

DLA Dimension as the Key Quantity

The dimension $|\mathfrak{g}| = \dim(\mathfrak{g})$ is the central quantity:

Circuit type $\dim(\mathfrak{g})$ Barren plateau?
Generic (hardware-efficient) random PQC $\Theta(4^n)$ — full $\mathfrak{su}(2^n)$ ✓ Yes — exponential BP
Quantum Alternating Operator Ansatz (QAOA) Polynomial in $n$ (problem-dependent) Partial — shallow QAOA OK
ADAPT-VQE (chemistry-inspired) $O(n^2)$ to $O(n^4)$ Mitigated — polynomial DLA
Matchgate / free-fermion circuits $O(n^2)$ (quadratic) ✓ No BP — classically simulable

4. Main Theorem: DLA & Variance Bound

Theorem (Ragone-Larocca 2024):
Let $U(\boldsymbol{\theta})$ be a PQC with DLA $\mathfrak{g}$. For a traceless observable $O$ and random parameter initialisation: $$\text{Var}_{\boldsymbol{\theta}}\!\left[\partial_\mu \mathcal{L}\right] \leq \frac{\|O\|_F^2 \cdot \|H_\mu\|_F^2}{|\mathfrak{g}|}$$ where $|\mathfrak{g}| = \dim(\mathfrak{g})$ and $\|\cdot\|_F$ is the Frobenius norm.
$$\boxed{\text{Var}_{\boldsymbol{\theta}}\!\left[\partial_\mu \mathcal{L}\right] \leq \frac{C}{|\mathfrak{g}|}}$$
Eq. (7) — DLA variance upper bound (constant $C$ depends on $O$ and $H_\mu$ norms)

Since generic $n$-qubit circuits have $|\mathfrak{g}| = \dim(\mathfrak{su}(2^n)) = 4^n - 1$:

$$\text{Var}_{\boldsymbol{\theta}}\!\left[\partial_\mu \mathcal{L}\right] \leq \frac{C}{4^n - 1} \in O(4^{-n})$$
Eq. (8) — Exponential gradient suppression for generic PQCs
Corollary: A PQC avoids barren plateaus if and only if $|\mathfrak{g}| \in O(\text{poly}(n))$. This provides an algebraic diagnostic: compute the DLA dimension to check trainability.

Tightness

The bound is tight for 2-designs (unitaries drawn from circuits that form approximate unitary 2-designs), where equality holds up to constants. For structured circuits (QAOA, ADAPT-VQE), the bound is not always tight.

$$\text{Var}_{\boldsymbol{\theta}}\!\left[\partial_\mu \mathcal{L}\right] = \frac{\|[H_\mu, O]\|_F^2}{|\mathfrak{g}|(|\mathfrak{g}|+1)} \quad \text{(exact, for 2-designs)}$$
Eq. (9) — Exact variance for 2-design circuits (Cerezo et al. 2025)

5. Landscape Architecture Diagram

PARAMETER SPACE  θ ∈ ℝᵏ
──────────────────────────────────────────────────────────

Large DLA (generic PQC):         Small DLA (structured PQC):

  ℒ(θ)                            ℒ(θ)
   │                               │
   │   "flat desert"               │    has structure!
   │ ~~~~~~~~~~~~~~~~              │  /\
   │ ~~~~~~~~~~~~~~~~              │ /  \    /\
   │ ~~~~~~~~~~~~~~~~              │/    \  /  \
   │ ~~~~~~~~~~~~~~~~              │      \/    \___
   └──────────────── θ             └────────────────── θ

  |𝔤| = O(4ⁿ)                     |𝔤| = O(poly(n))
  Var[∂ₘℒ] ~ 1/4ⁿ ≈ 0             Var[∂ₘℒ] ~ 1/poly(n)

  Training: IMPOSSIBLE             Training: FEASIBLE
  (exponential shots to see       (polynomial shots to
   nonzero gradient)               resolve gradient)

MECHANISM:
  PQC generators {iHₖ} ─→ generate ─→ 𝔤 = Lie({iHₖ})
                                         │
                          large 𝔤: circuit forms 2-design
                          → cost landscape is nearly flat
                          → quantum concentration of measure

6. Types of Barren Plateaus

Expressibility-induced BP

From large DLA — generic hardware-efficient circuits. Gradient $\sim 2^{-n}$. Primary focus of DLA theory.

Global cost BP

Observable $O$ acts on all qubits globally. Even structured circuits exhibit BP with global cost. Fix: use local cost functions.

Noise-induced BP

Deep circuits on noisy hardware: depolarising noise exponentially suppresses the signal part of the cost. Gradient noise ratio $\sim e^{-\gamma L}$ for depth $L$.

Entanglement-induced BP

Highly entangled reference states $|\psi_0\rangle$ suppress gradients. Even shallow circuits exhibit BP if the initial state is highly entangled.

7. Mitigation Strategies

1. Problem-Inspired (Structured) Ansätze

Choose $\{H_k\}$ from the problem's symmetry group. Example: for molecular simulation, use Fermionic operators — the DLA is $\mathfrak{su}(n/2)^2$ with $|\mathfrak{g}| = O(n^2)$, polynomial.
$$|\mathfrak{g}| \in O(\text{poly}(n)) \implies \text{Var}[\partial_\mu \mathcal{L}] \in \Omega\!\left(\frac{1}{\text{poly}(n)}\right)$$

2. Layer-wise Pre-training

Train circuit blocks sequentially (depth $L=1$ first, then extend). Each block has small DLA $\implies$ gradient is non-vanishing. Concatenate trained blocks.

3. Local Cost Functions

$$\mathcal{L}_{\text{local}} = \sum_{j=1}^{n} \langle O_j \rangle \quad \text{vs.} \quad \mathcal{L}_{\text{global}} = \langle O_1 \otimes \cdots \otimes O_n \rangle$$

Local cost functions have at most polynomial barren plateaus (gradient $\sim n^{-1}$) compared to exponential for global cost.

4. Quantum Natural Gradient (QNG)

$$\boldsymbol{\theta}_{t+1} = \boldsymbol{\theta}_t - \eta\, F^+(\boldsymbol{\theta}_t)\,\nabla_{\boldsymbol{\theta}} \mathcal{L}$$

Uses the quantum Fisher information matrix $F$ (Fubini-Study metric) as preconditioner. Does not eliminate BP but accelerates convergence when gradients are non-zero.

5. ADAPT-VQE / Greedy Construction

Select the next generator $H_k$ with the largest gradient magnitude. Ensures the active DLA dimension grows slowly, keeping gradients tractable throughout training.

6. Quantum-PEFT

LoRA-style reparameterisation: restrict to a low-rank subspace of parameter space. Equivalent to constraining the DLA to a polynomial-dimension subalgebra.

8. Expressibility–Trainability Tradeoff

\begin{aligned} &\text{Expressibility: } \mathcal{E} \propto |\mathfrak{g}| \\[6pt] &\text{Trainability: } \mathcal{T} \propto \frac{1}{|\mathfrak{g}|} \\[6pt] &\implies \mathcal{E} \cdot \mathcal{T} = O(1) \end{aligned}
Eq. (10) — Fundamental expressibility–trainability tradeoff
Trainability
 ↑
 │  ADAPT-VQE  ★
 │  QAOA (shallow)  ★
 │  Matchgates  ★
 │
 │              ★  Chemistry-inspired
 │
 │                              ★  Hardware-efficient (shallow)
 │
 │                                    ★  Hardware-efficient (deep)
 └────────────────────────────────────────────────→ Expressibility
                                          (full SU(2ⁿ))

Goal: design circuits in the "Goldilocks zone" — enough expressibility
to approximate the target state, small enough DLA to be trainable.

9. Open Questions

Q1: Can we efficiently compute $\dim(\mathfrak{g})$ for a given PQC? (Currently requires exponential-time symbolic computation for generic circuits.)
Q2: What is the tight lower bound on gradient variance? The current bound from DLA is an upper bound — tightness for non-2-design circuits is open.
Q3: Do classical ML techniques (e.g. skip connections, batch norm) have quantum analogues that alleviate BP?
Q4: How does noise-induced BP interact with expressibility-induced BP? Worst case: both active simultaneously.
Research frontier (2025–26): Cerezo et al.'s 2025 work establishes exact formulas for variance in terms of DLA representation theory. Current open problem: extending the DLA analysis to non-unitary quantum channels (noisy circuits) and mid-circuit measurement settings.

Related Papers

Vision
QViT — Quantum Vision Transformers
Avoids BP via orthogonal (RBS) layers
NLP/LLM
Quixer, HyQuT, QTHA
Hybrid transformers navigating the BP landscape