Strategies to Mitigate Barren Plateaus

Mitigating barren plateaus is the pacing problem of the noisy intermediate-scale quantum era, because every variational algorithm inherits it. In quantum chemistry the standard recipe stacks three levers: a Hartree–Fock warm start feeds a number-conserving ansatz such as unitary coupled cluster (symmetry), kept shallow and adaptively grown (structure) — the ADAPT-VQE recipe. QAOA sidesteps the plateau structurally, its fixed cost-and-mixer layers carrying only $2p$ parameters; in quantum machine learning, local observables plus tree- and MERA-style ansätze are now standard. Two fronts remain open. First, a theoretical unification: the dynamical-Lie-algebra picture suggests expressibility, entangling power, and trainability are one quantity — and that "provably trainable" ansätze may also be classically simulable, hollowing out the promised advantage. Second, the noise-induced plateau, which no ansatz engineering can touch and which ties the field's future to error mitigation and error correction. Whether a variational algorithm can be at once trainable, useful, and beyond classical reach is the defining open question — and why mitigation is where the field's attention now sits.

Imagine searching a desert for its deepest well while blindfolded, judging only the slope beneath your feet. If the sand is flat to your senses everywhere except a tiny oasis, you are lost. A barren plateau is exactly this: away from special regions the gradient falls below the noise floor set by finite shots, so the optimizer is fed pure noise. Every mitigation trick does one of two things: it shrinks the desert so the oasis is a larger fraction of it (symmetry, locality, shallow depth), or it drops you near the oasis (initialization). The villain is concentration of measure — just as almost all of a high-dimensional sphere's area hugs its equator, when $U(\vec\theta)$ explores the full unitary group almost every setting returns the same cost. The most physical lever is locality of the observable, and the reason is a light cone. A parameter's forward cone and a local observable's backward cone overlap in the only region carrying gradient signal — a handful of qubits for a shallow circuit (variance stays polynomial), all $n$ for a deep one (the plateau returns). The first figure makes this explicit: depth, not qubit count alone, decides trainability.

Initialization attacks the same problem differently: gradients are large only while the circuit stays near its start. Begin at (or near) the identity, or a good state such as Hartree–Fock, and the starting gradient is $\mathcal{O}(1)$ — a transient head start, not immunity. Symmetry is the strongest lever when it applies: if every gate commutes with a conserved charge, the dynamics never leave one block, so the averages run over its dimension $d_\lambda$ instead of $2^n$. The second figure shows what decides whether any of this matters — gradient variance versus qubit number, and the cliff between trainable and untrainable.

Above the cliff — local cost, shallow circuit, respected symmetry — a laptop optimizer and a few thousand shots per gradient suffice; below it, no optimizer helps. Mitigation is the art of staying above it.

The phenomenon every strategy must defeat is one line: over the Haar measure the gradient has zero mean and exponentially small variance — the concentration that flattens everything.

where: $C=\mathrm{Tr}[U(\vec\theta)\rho\,U^\dagger(\vec\theta)\,H]$ — the cost; $\partial_{\theta_k}C$ — its derivative in the $k$-th parameter; $n$ — qubit number; $\langle\cdot\rangle_{\mathrm{Haar}}$ — Haar average on $U(2^n)$. The typical gradient $\sqrt{\mathrm{Var}}\sim 2^{-n/2}$ needs $\sim 2^{n}$ shots to resolve. The first lever is the observable. A global projector asks one exponentially fine question about all $n$ qubits at once; a local cost is a sum of single-qubit questions — same minimum, polynomial signal at shallow depth.

where: $C_G$ — global cost with the full $n$-qubit projector, barren at every depth; $C_L$ — local cost, a sum of single-qubit projectors $|0\rangle\langle 0|_j$ (identity elsewhere), trainable to depth $\mathcal{O}(\log n)$. Both are minimized by $U\rho\,U^\dagger=|\vec 0\rangle\langle\vec 0|$, differing only away from it. The scaling laws come from the calculus of Haar averages. The first-moment identity — the twirl $\int dU\,U A U^\dagger=\mathrm{Tr}(A)\,\mathbb{1}/2^n$ — annihilates any traceless operator (in particular a commutator), forcing the mean gradient to zero, so the variance is what matters. It is fixed by the second moment: for a $2$-design the exact identity, verified by direct Weingarten calculation, is the workhorse behind every plateau theorem.

where: $A,B,C,D$ — the operators inserted by two copies of $\rho$ and $H$; $2$-design — an ensemble whose first two moments equal Haar's. Initialization enters through the starting angles: keeping the total rotation of order one prevents scrambling into a design early.

where: $W_l$ — fixed gates in layer $l$; $V_l$ — Hermitian generator of the $l$-th rotation; $\sigma^2\propto 1/L$ — small-angle variance keeping the accumulated rotation $\mathcal{O}(1)$. The block-identity variant instead sets the two halves to invert each other, $U(\vec\theta_0)=U_1U_2^{-1}=\mathbb{1}$, beginning exactly at the identity with an $\mathcal{O}(1)$ gradient. Symmetry is the sharpest lever: a conserved charge block-diagonalizes the circuit, shrinking the averages to the sector dimension.

where: $G$ — symmetry generator (total spin $S_z$, particle number $\hat N$, a point-group operation); $\mathcal{H}_\lambda$ — the invariant sector the state lives in; $d_\lambda=\dim(\mathcal{H}_\lambda)$. Cured only when $d_\lambda=\mathrm{poly}(n)$; a half-filling sector, $\binom{n}{n/2}\sim 2^n/\sqrt{n}$, is still exponential and does not help alone. The gradient we estimate comes from the parameter-shift rule; its statistical cost is set by the variance.

where: $\hat e_k$ — unit vector along parameter $k$; the $\pi/2$ shift is exact for a generator with eigenvalues $\pm\tfrac12$; $N_s$ — shots to resolve a gradient of size $\sqrt{\mathrm{Var}}$. On a plateau, $N_s\sim 2^n$. Finally, the enemy no structural lever touches: incoherent noise contracts the state toward the maximally mixed state, flattening the gradient exponentially in depth.

where: $p$ — per-layer local noise rate (e.g. depolarizing); $L$ — circuit depth; $G_0$ — the noiseless gradient bound. Set by depth and noise alone, the decay ignores initialization, locality, and symmetry entirely.

How gradient variance scales under the leading regimes (ansatz locally $2$-design within its cone):

1 · The exponential shot wall, and how locality removes it (fully worked)

A deep $2$-design circuit with a single Pauli observable has $\mathrm{Var}[\partial_\theta C]\approx 2^{-n}$. Compare the cost of one gradient component at $n=4,\ 20,\ 50$ against a local cost at shallow depth with $\mathrm{Var}\gtrsim 1/n^2$.

Step 1. Fix the shot rule. Equation (6) gives $N_s\gtrsim 1/\mathrm{Var}$: global/deep needs $N_s\sim 2^{n}$, local/shallow $N_s\sim n^2$.

Step 2. The global wall. $n=4$: $\mathrm{Var}=0.0625$, $N_s\sim\mathbf{16}$. $n=20$: $N_s\sim\mathbf{1.0\times10^{6}}$ — a million per component. $n=50$: $\mathrm{Var}\approx 8.9\times10^{-16}$, $N_s\sim\mathbf{1.1\times10^{15}}$.

Step 3. In wall-clock time: at an optimistic $10^{6}$ shots/s, $1.1\times10^{15}$ shots take $\approx\mathbf{35\ years}$ — for one gradient component of one step. A global cost at $n=50$ is not slow but impossible.

Step 4. Evaluate the local/shallow curve: $N_s\sim n^2$ gives $n=50\Rightarrow\mathbf{2500}$ shots. At $n=50$, $\sim 10^{15}$ versus $\sim 2.5\times10^{3}$ shots — twelve orders of magnitude, the whole difference between impossible and routine, bought by locality and shallow depth.

2 · Identity-block initialization gives an order-one starting gradient (fully worked)

Show that starting at the identity gives an $\mathcal{O}(1)$ gradient. Take a single qubit, $\rho=|0\rangle\langle 0|$, observable $H=X$, and gate $e^{-i\theta V}$ with $V=Y/2$ so $e^{-i\theta Y/2}=R_y(\theta)$; evaluate at $\theta=0$.

Step 1. Use the commutator gradient from Derivation 1, $\partial_\theta C\big|_{0}=i\,\mathrm{Tr}(\rho[V,H])$, with $[V,H]=[Y/2,X]=-iZ$. The commutator is a single local Pauli — no exponential suppression.

Step 2. Insert the state: $\partial_\theta C\big|_{0}=i\,\mathrm{Tr}(|0\rangle\langle 0|(-iZ))=\langle 0|Z|0\rangle=\mathbf{1}$ — fully $\mathcal{O}(1)$.

Step 3. Cross-check: $R_y(\theta)|0\rangle$ gives $C(\theta)=\langle X\rangle=\sin\theta$, so $\partial_\theta C|_{0}=\cos 0=\mathbf{1}$; parameter-shift reproduces it, $\tfrac12[\sin(\tfrac\pi2)-\sin(-\tfrac\pi2)]=\mathbf{1}$.

Step 4. At the identity the net unitary is trivial and the ensemble is far from a $2$-design, so the gradient is $\mathcal{O}(1)$, not $2^{-n}$. But this is a starting condition, not a guarantee: after enough steps the parameters drift and, if the ansatz can form a design, the variance collapses again — which is why identity-block initialization must be paired with locality or bounded depth.

Barren plateaus were named only in 2018; their mitigation is a recent, unfinished theory whose central result — that trainability depends on the cost function, not just the hardware — belongs to the physicist who proved it and mapped the field.

Texts and papers. McClean et al., "Barren plateaus in quantum neural network training landscapes," Nat. Commun. 9, 4812 (2018) · Cerezo et al., "Cost function dependent barren plateaus in shallow parametrized quantum circuits," Nat. Commun. 12, 1791 (2021) · Grant et al., "An initialization strategy for addressing barren plateaus," Quantum 3, 214 (2019) · Wang et al., "Noise-induced barren plateaus in variational quantum algorithms," Nat. Commun. 12, 6961 (2021) · Larocca et al., "Diagnosing barren plateaus," Quantum 6, 824 (2022) · Cerezo et al., "Variational Quantum Algorithms," Nat. Rev. Phys. 3, 625 (2021) · Watrous, Ch. 7 (Weingarten calculus). The $2$-design identity (3) was cross-checked against direct Haar sampling at $n=3$: the Weingarten formula gives $0.195$ and a Monte Carlo estimate gives $0.192$. The local-cost log-depth threshold uses the brickwork cone size $m(L)=2L+1$, and the half-filling sector dimension is $\binom{n}{n/2}\sim 2^n/\sqrt{n}$.

Prerequisites. Barren Plateaus in Variational Landscapes · Parameterized Quantum Circuits and Ansatz Design · Cost-Function and Measurement Estimation in Variational Algorithms · Parameter-Shift Rules and Quantum Gradients. Leads to. Variational Quantum Eigensolver · Expressibility, Entangling Capability, and Trainability Trade-offs · Hardware-Efficient and Chemistry-Inspired Ansätze · Classical Optimizers for Variational Quantum Algorithms.

Ask the tutor: Ask me to work the variance collapse from the Weingarten identity step by step, to size the shot budget for a plateau at your qubit count, or to design a symmetry- and locality-respecting ansatz for a Hamiltonian.