Barren Plateaus in Variational Landscapes

A barren plateau is the reason a variational quantum computer can hold the answer to a problem in its reach and still be structurally unable to find it. The entire near-term programme — variational eigensolvers for chemistry, QAOA for optimization, parameterized circuits for machine learning — rests on one hope: that a classical optimizer, handed the measured gradient of a cost function $C(\boldsymbol\theta)=\langle\psi(\boldsymbol\theta)|H|\psi(\boldsymbol\theta)\rangle$ , can walk a quantum circuit downhill toward a good solution. The barren-plateau result, proved by McClean, Boixo, Smelyanskiy, Babbush and Neven in 2018, shows the hope is fragile for the most natural circuit designs. Make an ansatz expressive enough to reach anywhere in state space and its landscape becomes flat almost everywhere — flat not by a little, but to a precision that shrinks exponentially in the qubit number $n$ . This frontier synthesis takes the cost, gradient, and Haar-measure machinery its prerequisites own and asks what happens when the circuit is random enough to average like the unitary group. Three ideas organize everything. First, the mechanism is not bad luck but concentration of measure: on the exponentially large unitary group almost every circuit sits exponentially close to a trivial average, so both the mean gradient and its variance are computed exactly by Weingarten calculus and both come out crushed as $\mathrm{Var}[\partial_k C]\sim 2^{-n}$ . Second, the wall is information-theoretic, not algorithmic — a landscape flat to $2^{-n}$ cannot be navigated by any optimizer without spending $\sim 2^{n}$ measurement shots. Third, the effect is a unifying obstruction: expressibility, observable locality, entanglement, and hardware noise are four routes to the same exponential flatness, and the escape from each is structural.

Key takeaways

A barren plateau is a region of a variational cost landscape where the gradient has zero mean and a variance that vanishes exponentially in the qubit number: $\mathbb{E}[\partial_k C]=0$ while $\mathrm{Var}[\partial_k C]\sim 2^{-n}$ , so the surface is nearly flat everywhere.
The cause is concentration of measure: once a parameterized circuit is expressive enough to form an approximate unitary 2-design, Haar averaging by Weingarten calculus forces the first and second gradient moments to their trivial, exponentially suppressed values.
On a plateau the cost itself concentrates on the maximally-mixed value $\mathrm{Tr}(H)/2^{n}$ with fluctuations only $\mathcal{O}(2^{-n/2})$ — an optimizer sees a near-constant function buried in shot noise.
Resolving such a gradient above statistical noise costs $M\sim 2^{n}$ measurement shots per step, so the barrier defeats gradient-based, gradient-free, and second-order methods equally.
Plateaus have four routes with one shared signature: expressibility (deep random circuits), global observables, excessive entanglement, and hardware noise, where the variance decays as $(1-p)^{L}$ in depth $L$ rather than $2^{-n}$ in width.
Every mitigation — local cost functions, shallow or structured ansätze, identity-block initialization, warm starts — works by lowering expressibility so the circuit never becomes a 2-design.
Key idea: Deep or unstructured parameterized circuits approximate Haar-random unitaries; Haar averaging concentrates the cost so tightly that the gradient variance — and hence its typical size — falls as $2^{-n}$ , making gradient signals exponentially smaller than the shot noise required to resolve them.

Where this lives today

The barren plateau is the organizing obstacle of the entire variational, NISQ-era programme, and diagnosing it has become a routine first step in serious ansatz design. Before committing to a hardware-efficient or chemistry-inspired circuit, practitioners estimate the gradient variance versus qubit number on small instances and extrapolate; an exponential decay is a red flag that scaling up will fail regardless of the optimizer. That diagnostic now shapes practice across the field. In VQE, global observables are decomposed into local terms and paired with shallow circuits so the lightcone argument keeps the variance polynomial, and adaptive constructions such as ADAPT-VQE grow the ansatz gate by gate to stay in the trainable regime; unitary-coupled-cluster and Hamiltonian-variational ansätze restrict the explored subspace by design. In QAOA and quantum machine learning the noise-induced plateau sets a hard depth budget on real devices, and the entanglement-induced variant (Marrero et al., Patti et al.) warns against ansätze that scramble a bipartition too aggressively. The active research fronts are all responses to the theorem: identity-block and small-angle initializations that start the circuit off the Haar-typical region; classically informed warm starts; symmetry-preserving and equivariant ansätze that break the circuit's randomness; and layerwise or block-local training. A sharper and more sobering thread runs alongside them — recent "provable absence of barren plateaus" results show that many circuits engineered to be trainable are, for the same structural reasons, also classically simulable, raising the live and unresolved question of whether escaping the plateau and retaining a genuine quantum advantage are compatible at all.

The physical picture

Imagine hunting for one shallow well on a continent-sized plain that is dead flat to within the width of an atom everywhere except a single tiny crater. Random parameter initialization parachutes you somewhere onto the plain; the slope beneath your feet is so faint it is drowned by the trembling of your only instrument — the shot noise of a finite number of measurements. Add a qubit and the plain doubles in area while the residual slope halves in root-mean-square: the featureless expanse grows and flattens at the same exponential rate. This is not a pathological landscape chosen to defeat you. It is the typical one, because on the exponentially large space of circuits almost every instance looks like the average, and the average has no slope at all. The first figure makes the concentration visible: as $n$ grows the cost curve collapses onto its mean and its wiggles shrink like $2^{-n/2}$ .

Figure

As the qubit number n grows the cost landscape C(θ) collapses onto the maximally-mixed value Tr(H)/2ⁿ, its peak-to-peak features shrinking like 2·2⁻ⁿ⁄² — the n = 2 curve still swings widely, the n = 6 curve is muted, and the n = 12 curve hugs the mean line: a barren plateau

The deep reason is symmetry. A highly expressive circuit reaches essentially any unitary with roughly equal probability, so averaging a fixed observable's gradient over the ensemble is nearly averaging over the whole unitary group with the Haar measure — positive and negative contributions cancel almost perfectly. Concentration of measure then guarantees that not just the average but almost every individual instance lies exponentially close to that trivial value: you cannot dodge the plateau by re-rolling the starting angles, because nearly all of them land you on it. Locality is the one genuine lever. A global observable overlaps a typical random state with exponentially small probability, so its landscape is flat even at constant depth; a single-qubit Pauli is only weakly scrambled by a shallow circuit, so its gradient survives up to logarithmic depth. Noise is a second, independent flattener — every layer contracts the state toward the maximally mixed fixed point. The second figure is the quantity a practitioner actually measures, the gradient variance versus qubit number, and it shows exactly where trainability dies.

Figure

Gradient variance Var(∂C) ≈ 2⁻ⁿ collapses exponentially with qubit number: 6.3×10⁻² at n = 4, 9.5×10⁻⁷ at n = 20, 8.9×10⁻¹⁶ at n = 50, crossing the shot-noise floor 1/M = 10⁻⁶ of a million-shot estimate near n = 20 — beyond that the gradient is buried in statistical noise, the trainability cliff

The core equations

Everything starts from the cost function of a variational algorithm: the expectation value of a problem observable in the state a parameterized circuit prepares. It is what the optimizer tries to lower, and the object whose gradient will turn out to vanish.

C(\boldsymbol\theta)=\langle 0|\,U^{\dagger}(\boldsymbol\theta)\,H\,U(\boldsymbol\theta)\,|0\rangle \qquad (1)

where: $U(\boldsymbol\theta)$ — parameterized circuit on $n$ qubits built from $L$ gate layers; $H$ — Hermitian problem observable; $|0\rangle$ — fixed reference input $|0\rangle^{\otimes n}$ ; $\boldsymbol\theta$ — vector of variational parameters. A deeper $U$ or a more global $H$ both push the circuit toward the Haar-random behaviour that creates the plateau. To differentiate with respect to one angle, split the circuit at the gate that carries it. This standard device localizes the derivative onto a single Hermitian generator.

U(\boldsymbol\theta)=U_{-}\,e^{-i\theta_k V_k}\,U_{+},\qquad V_k^2=\mathbb{1} \qquad (2)

where: $V_k$ — Hermitian generator of the $k$ -th gate, with $V_k^2=\mathbb{1}$ as every Pauli rotation has; $U_{\pm}$ — the parts of the circuit before / after the differentiated gate; $\theta_k$ — the single parameter being varied. Only the middle factor depends on $\theta_k$ , so the product rule strikes exactly once. Carrying the derivative through and grouping the pieces, the gradient becomes the expectation of one commutator — a compact form ready to be averaged over the circuit ensemble.

\partial_k C = i\,\mathrm{Tr}\!\big(\rho\,[V_k,\,\tilde H]\big),\qquad \rho=U_{+}|0\rangle\langle 0|U_{+}^{\dagger},\;\;\tilde H=U_{-}^{\dagger}HU_{-} \qquad (3)

where: $\rho$ — state entering the differentiated gate; $\tilde H$ — observable propagated back through $U_{-}$ ; $[\,\cdot,\cdot\,]$ — commutator $[A,B]=AB-BA$ . Read it as: the slope in direction $k$ is how badly the generator $V_k$ fails to commute with the back-propagated observable, sampled in the state $\rho$ . Averaging the commutator over a Haar-random (1-design) sub-circuit sends the observable to a multiple of the identity, and the identity commutes with everything, so the mean gradient is exactly zero.

\mathbb{E}_{U}\big[\partial_k C\big]=0 \qquad (4)

where: $\mathbb{E}_{U}$ — average over the unitary ensemble (Haar or a design) applied to $U_{+}$ or $U_{-}$ . The plateau has no preferred direction: the expected slope is flat everywhere, so the optimizer begins with no bias toward the solution. The mean vanishing is only half the story — a fluctuating gradient with zero mean could still be usable. The second moment settles it: computed with the 2-design (Weingarten) integral, the variance carries an explicit factor $1/(2^{2n}-1)$ .

\mathrm{Var}\big[\partial_k C\big]\;\approx\;\frac{\mathrm{Tr}(H^2)\,\mathrm{Tr}(V_k^2\rho)}{2^{2n}-1}\;\sim\;2^{-n} \qquad (5)

where: $\mathrm{Tr}(H^2)$ — Hilbert–Schmidt norm of the observable, at most $\|H\|^2\,2^{n}$ ; $2^{2n}=d^2$ — square of the Hilbert-space dimension $d=2^{n}$ . Because $\mathrm{Tr}(H^2)\le\|H\|^2 d$ while the denominator is $d^2$ , the ratio scales as $1/d=2^{-n}$ : this is the barren plateau. An exponentially small gradient can be detected only against exponentially small statistical noise, so the number of measurement shots per gradient estimate is set by the reciprocal of the variance.

M\;\gtrsim\;\frac{1}{\mathrm{Var}[\partial_k C]}\;\sim\;2^{\,n} \qquad (6)

where: $M$ — measurement repetitions (shots) needed to resolve the gradient above shot noise. This is the operational face of the plateau: not merely a small number, but an exponentially expensive one to see.

Derivation AVanishing mean and exponentially small variance of the gradient

The goal is the flagship result of the whole subject: to show that when even one sub-circuit is drawn from an approximate unitary 2-design, the parameter-shift gradient of the cost has exactly zero mean and a variance that decays like $2^{-n}$ , so the landscape is provably a barren plateau rather than merely a difficult one. We localize the gradient onto a single commutator, average with the first Weingarten moment to kill the mean, then square and average with the second moment to expose the $1/(d^2-1)$ suppression, and finally bound the surviving trace. Take $C=\langle 0|U^{\dagger}HU|0\rangle$ with $U=U_-\,e^{-i\theta_k V_k}\,U_+$ and $V_k^2=\mathbb{1}$ .

Step 1. Differentiate and group: only the middle factor carries $\theta_k$ , so the product rule gives $\partial_k C = i\,\langle 0|U_+^{\dagger}[V_k,\tilde H]U_+|0\rangle$ with $\tilde H=U_-^{\dagger}HU_-$ , manifestly real because $i[V_k,\tilde H]$ is Hermitian. Why: the derivative of a long circuit collapses to a single commutator sandwiched by the front sub-circuit, which is what makes the ensemble average tractable.

Step 2. Average the observable over $U_-$ using the first Weingarten (1-design) moment $\int\!\mathrm{d}U\,U^{\dagger}HU=\tfrac{\mathrm{Tr}(H)}{d}\mathbb{1}$ , so $\mathbb{E}_{U_-}[\tilde H]=\tfrac{\mathrm{Tr}(H)}{d}\mathbb{1}$ . Why: the Haar average of any conjugated operator is proportional to the identity — the ensemble forgets everything about $H$ except its trace.

Step 3. Take the mean of the gradient: $\mathbb{E}[\partial_k C]\propto \mathrm{Tr}\!\big(\rho\,[V_k,\mathbb{1}]\big)=0$ . Why: a commutator with the identity vanishes identically, so the expected slope is exactly flat and a 1-design already suffices to zero the mean — the variance now carries the whole story.

Step 4. For the second moment, squaring brings two factors of $U$ and two of $U^{\dagger}$ , requiring the 2-design (second Weingarten) moment $\int\!\mathrm{d}U\,(U^\dagger A U)\otimes(U^\dagger B U)=\tfrac{\mathrm{Tr}(A)\mathrm{Tr}(B)\,\mathbb{1}+\mathrm{Tr}(AB)\,\mathbb{S}}{d^2-1}-\tfrac{\cdots}{d(d^2-1)}$ with $\mathbb{S}$ the swap; the surviving contraction yields $\mathbb{E}[(\partial_k C)^2]\approx \tfrac{\mathrm{Tr}(H^2)\,\mathrm{Tr}(V_k^2\rho)}{d^2-1}$ . Why: the first-moment average already sent the mean to zero, so the entire variance lives in this integral, whose prefactor $1/(d^2-1)$ is the algebraic seed of the suppression.

Step 5. Bound the trace and read the scaling: for bounded $\|H\|$ , $\mathrm{Tr}(H^2)\le\|H\|^2 d$ , so with $d=2^{n}$ we get $\tfrac{d}{d^2-1}\sim\tfrac1d=2^{-n}$ and hence $\mathrm{Var}[\partial_k C]\sim 2^{-n}$ . Why: the observable norm can grow at most like the dimension, but the denominator grows like the dimension squared — the mismatch is exactly the exponential suppression that defines the plateau. *Verified: reproduces McClean et al. (2018), Nat. Commun. 9, 4812; the Weingarten first and second moments are standard (Watrous, The Theory of Quantum Information); $\mathrm{Tr}[V_k,\mathbb{1}]=0$ and $d/(d^2-1)\sim 2^{-n}$ check by inspection, and numerically $2^{-4}=6.25\times10^{-2}$ , $2^{-20}=9.54\times10^{-7}$ .*

Go deeper · Derivation B · 1 minNoise-induced barren plateau from channel contraction

Derivation BNoise-induced barren plateau from channel contraction

The aim is to show that a circuit interleaved with local depolarizing noise has gradients exponentially small in depth $L$ , entirely independent of how the parameters are initialized — a second, mechanistically distinct route to a plateau that no clever starting point or local cost can cure. We model each layer as unitary evolution followed by single-qubit depolarizing noise, track the contraction of the state toward the maximally mixed point in the Pauli basis, and bound the gradient by the shrinking non-identity part. Let $\rho_0$ be the input and $\mathcal{N}_p$ the per-qubit depolarizing channel of rate $p$ .

Step 1. Write the layer channel $\mathcal{N}_p(\sigma)=(1-p)\,\sigma+p\,\tfrac{\mathrm{Tr}(\sigma)}{2}\,\mathbb{1}$ per qubit, whose fixed point is the maximally mixed state $\mathbb{1}/2$ . Why: depolarizing noise mixes in a fraction $p$ of the identity each layer, and the identity component is exactly the structureless state a gradient cannot live on.

Step 2. In the Pauli basis the channel is diagonal, $\mathcal{N}_p:\,P_{\boldsymbol a}\mapsto (1-p)^{|\boldsymbol a|}\,P_{\boldsymbol a}$ , damping every nontrivial Pauli of weight $|\boldsymbol a|$ by $(1-p)$ per active qubit; interleaved unitaries permute Pauli labels but preserve the contraction bound. Why: any observable or state deviation from $\mathbb{1}/2^n$ is a sum of nontrivial Paulis, and each is multiplicatively shrunk by the noise.

Step 3. Compose $L$ layers: $\big\|\rho_L-\tfrac{\mathbb{1}}{2^n}\big\|_2\le (1-p)^{L}\big\|\rho_0-\tfrac{\mathbb{1}}{2^n}\big\|_2$ , so the state converges exponentially to maximal mixedness. Why: repeated contraction multiplies the non-identity part by $(1-p)$ each layer, and $\mathbb{1}/2^n$ is parameter-independent, so only the vanishing deviation can carry any gradient.

Step 4. Bound the gradient: $|\partial_k C|=|\mathrm{Tr}(H\,\partial_k\rho_L)|\le \|H\|_\infty\,\|\partial_k\rho_L\|_1\le 2\|H\|_\infty (1-p)^{L}=\mathcal{O}(e^{-\alpha L})$ with $\alpha=-\ln(1-p)>0$ . Why: the derivative of the contracted state inherits the same $(1-p)^L$ damping, so the gradient decays exponentially in depth regardless of the parameter values. *Verified: matches Wang et al. (2021), Nat. Commun. 12, 6961, whose main theorem bounds the gradient variance above by an exponentially decaying function of $L$ for unital noise; numerically $p=0.01$ gives $\alpha=0.01005$ and $(1-p)^{100}=0.366$ .*

Go deeper · Derivation C · 1 minLocality and shallow depth restore trainability

Derivation CLocality and shallow depth restore trainability

The purpose here is to show why the plateau is escapable rather than a death sentence: if the cost is a sum of geometrically local terms and the circuit is shallow, only a polylogarithmic patch of gates lies in each term's causal lightcone, so the 2-design never forms over the full register, the effective dimension entering the Weingarten estimate is polynomial rather than $2^n$ , and the gradient variance decays only polynomially — while the same argument locates a depth threshold beyond which the plateau returns. Let $H=\sum_j O_j$ be a sum of local operators and $U$ a brickwork circuit of depth $L$ .

Step 1. Bound the causal cone: a local term $O_j$ is affected only by gates in its backward lightcone, of width $\mathcal{O}(2^{L})$ qubits for a brickwork circuit of depth $L$ . Why: information in a finite-depth circuit spreads at a finite effective speed, so a shallow circuit couples each observable to only a small contiguous block.

Step 2. Identify the effective dimension: only the lightcone qubits are scrambled, so the relevant Weingarten dimension is $d_{\mathrm{eff}}=2^{|\mathrm{lightcone}|}$ , which is $\mathrm{poly}(n)$ when $L=\mathcal{O}(\log n)$ . Why: the exponential suppression is governed by however many qubits actually average like Haar, not by the total register size.

Step 3. Replace $d$ by $d_{\mathrm{eff}}$ in equation (5): $\mathrm{Var}[\partial_k C]\sim 1/d_{\mathrm{eff}}\sim 1/\mathrm{poly}(n)=\Omega(1/\mathrm{poly}(n))$ . Why: with a polynomial effective dimension the gradient shrinks only polynomially, so a polynomial number of shots resolves it and the landscape is trainable.

Step 4. Cross the threshold: once $L\gtrsim\mathcal{O}(n)$ the lightcone spans the whole register, $d_{\mathrm{eff}}\to 2^{n}$ , and $\mathrm{Var}[\partial_k C]\sim 2^{-n}$ again. Why: enough depth scrambles the entire system into a 2-design, so even a local cost eventually plateaus — trainability is a statement about depth as much as about locality. *Verified: matches the shallow-circuit theorem of Cerezo et al. (2021), Nat. Commun. 12, 1791, linking local costs and $\mathcal{O}(\log n)$ depth to polynomially-decaying variance; the lightcone reasoning is standard for brickwork circuits.*

Limiting cases

How the gradient scales is decided by three levers — observable locality, circuit depth, and hardware noise — and the same variance formula specializes to qualitatively different regimes:

Case	Result	Meaning
Deep expressive circuit (2-design)	$\mathrm{Var}[\partial_k C]\sim 2^{-n}$	McClean et al.: a global 2-design gives exponential decay in the qubit number $n$ regardless of observable locality — the full barren plateau.
Shallow circuit, local cost	$\mathrm{Var}[\partial_k C]\sim \Omega\!\left(1/\mathrm{poly}(n)\right),\ L=\mathcal{O}(\log n)$	Cerezo et al.: a local observable at logarithmic depth keeps polynomially-large gradients — trainable.
Shallow circuit, global cost	$\mathrm{Var}[\partial_k C]\sim 2^{-n}$	A global observable plateaus even at $L=\mathcal{O}(1)$ depth: locality of $H$ , not just depth, matters.
Noisy circuit, depth $L$	$\big\|\partial_k C\big\|\lesssim (1-p)^{L}\propto e^{-\alpha L}$	Wang et al.: generic local noise of rate $p$ per layer damps gradients exponentially in depth, independent of ansatz structure.

Common trapA barren plateau is a bad local minimum a smarter optimizer could escape

It is neither a local minimum nor a symptom of a weak optimizer. The landscape is genuinely flat — flat to a precision that shrinks like $2^{-n}$ — over an exponentially large region, so there is no slope for any method to follow. The killer is not the vanishing average gradient ( $\mathbb{E}[\partial_k C]=0$ holds trivially by symmetry, even for easily trainable landscapes) but the vanishing variance: when $\mathrm{Var}[\partial_k C]\sim 2^{-n}$ , almost every individual component is exponentially small, so no step size, second-order method (Newton), gradient-free search (Nelder–Mead, SPSA), or quantum natural gradient helps — all must resolve cost differences of size $2^{-n}$ and all drown in the same shot noise. Nor can you re-roll your way out with many random initializations, because concentration of measure puts almost all of them on the plateau. Equally, do not conflate a barren plateau (global flatness) with ordinary local minima or saddles, which are landscape features at fixed resolution.

Worked examples

1 · The trainability cliff, fully worked

A hardware-efficient ansatz on $n$ qubits reaches an approximate unitary 2-design with a bounded observable of order-one norm, so $\mathrm{Var}[\partial_k C]\approx 2^{-n}$ . Tabulate the gradient variance, the typical gradient magnitude, and the shots needed to resolve it for $n=4,10,20,50$ , and locate the cliff for a million-shot budget.

Step 1. Because the mean is zero, the typical magnitude of the gradient is the square root of its variance: $|\partial_k C|_{\mathrm{typ}}\sim\sqrt{\mathrm{Var}}=\sqrt{2^{-n}}=2^{-n/2}$ . The shots follow from equation (6): $M\sim 1/\mathrm{Var}\sim 2^{n}$ . Both are pure functions of $n$ .

Step 2. Evaluate at $n=4$ : $\mathrm{Var}=2^{-4}=\mathbf{6.25\times10^{-2}}$ , $|\partial_k C|_{\mathrm{typ}}=2^{-2}=0.25$ , $M\sim 2^{4}=\mathbf{16}$ . A four-qubit toy is comfortably trainable — a handful of shots sees the slope.

Step 3. Evaluate at $n=10$ and $n=20$ : at $n=10$ , $\mathrm{Var}=2^{-10}=9.77\times10^{-4}$ , $|\partial_k C|_{\mathrm{typ}}=2^{-5}\approx0.031$ , $M\sim1024$ . At $n=20$ , $\mathrm{Var}=2^{-20}=\mathbf{9.54\times10^{-7}}$ , $|\partial_k C|_{\mathrm{typ}}=2^{-10}\approx9.8\times10^{-4}$ , $M\sim2^{20}\approx\mathbf{1.05\times10^{6}}$ . This is the cliff: a standard $M=10^{6}$ -shot estimate has a shot-noise floor $1/M=10^{-6}$ , and the variance $9.54\times10^{-7}$ has just fallen through it — beyond about 20 qubits the gradient is buried in statistical noise.

Step 4. Evaluate at $n=50$ : $\mathrm{Var}=2^{-50}=\mathbf{8.88\times10^{-16}}$ , $|\partial_k C|_{\mathrm{typ}}=2^{-25}\approx2.98\times10^{-8}$ , $M\sim 2^{50}\approx\mathbf{1.13\times10^{15}}$ . One optimization step would need of order a quadrillion measurements. The plateau turns a "50-qubit trainable model" into a physical impossibility — not because the answer is unrepresentable, but because its gradient is unmeasurable.

2 · A noise-induced plateau and the depth budget

A depth- $L$ circuit on a device with weak per-layer noise contracts the state toward maximally mixed, so the gradient variance decays as $\mathrm{Var}[\partial_k C]\sim (1-p)^{L}$ with an effective per-layer factor $1-p=0.99$ (a one-percent contraction per layer). Find how the gradient shrinks with depth and the deepest circuit that keeps the variance above a usable floor of $10^{-4}$ .

Step 1. The gradient magnitude scales as $\sqrt{\mathrm{Var}}\sim(1-p)^{L/2}$ . At $L=100$ , $(0.99)^{100}=\mathbf{0.366}$ , so the gradient is reduced to $\sqrt{0.366}\approx0.61$ of its noiseless value — a mild but real erosion after a hundred layers.

Step 2. Push deeper. At $L=500$ , $0.99^{500}=6.57\times10^{-3}$ ; at $L=1000$ , $0.99^{1000}=\mathbf{4.32\times10^{-5}}$ : the variance has fallen more than four orders of magnitude purely from noise, with no 2-design assumed. This is a distinct mechanism from the expressibility plateau — driven by depth and error rate, not Haar randomness.

Step 3. Solve for the depth budget. Requiring $(0.99)^{L}>10^{-4}$ gives $L<\dfrac{\ln 10^{-4}}{\ln 0.99}=\dfrac{-9.210}{-0.01005}=\mathbf{916\ \text{layers}}$ . A one-percent-per-layer device can run at most about 900 layers before noise alone flattens the landscape — a hard ceiling on usable depth that tightens as the error rate rises, and the quantitative basis for depth budgets in QAOA and quantum machine learning.

Physics of the portrait

The barren plateau is where a piece of pure high-dimensional geometry — concentration of measure, a nineteenth- and twentieth-century idea about spheres — reappeared in 2018 as the central obstacle of near-term quantum computing.

Origins · 1 minbarren plateau and how this was discovered

barren plateau

the exponential flattening of variational quantum landscapes, McClean et al. 2018 · concentration of measure meets variational quantum computing

The barren plateau is not a person but a phenomenon, and its lineage runs through both mathematics and physics. Its mathematical root is concentration of measure: Paul Lévy observed early in the twentieth century, and Vitali Milman sharpened in the 1970s, that a smooth function on a very high-dimensional sphere is almost constant — nearly all of its mass sits exponentially close to the mean. The unitary group is such a high-dimensional space, and Haar integrals over it were made computable by the Weingarten calculus of Don Weingarten and, in its modern combinatorial form, Benoît Collins and Piotr Śniady in 2006. The quantum-computing chapter opened in 2018, when Jarrod McClean, Sergio Boixo, Vadim Smelyanskiy, Ryan Babbush, and Hartmut Neven at Google proved in "Barren plateaus in quantum neural network training landscapes" that a random, sufficiently deep parameterized circuit inherits exactly this concentration: its cost gradient has zero mean and variance $\sim 2^{-n}$ . The result reframed the field. Marco Cerezo, Patrick Coles, and collaborators showed in 2021 that severity depends on cost-function locality, rescuing shallow local-cost circuits; the same year Samson Wang and co-workers established that hardware noise induces plateaus of its own; and Zoë Holmes with Coles tied the whole effect to a single scalar, the ansatz's expressibility. Together they turned a discouraging theorem into a design principle — trainability demands structured, low-expressibility ansätze, not generic universal ones.

History

1971

— Milman sharpens concentration of measure on high-dimensional spheres

2006

— Collins and Śniady formalize the Weingarten calculus for Haar unitary integrals

2018

— McClean, Boixo, Smelyanskiy, Babbush, and Neven prove barren plateaus in random parameterized circuits

2021

— Cerezo and co-workers link plateau severity to cost-function locality; Wang and co-workers identify noise-induced plateaus

2022

— Holmes and co-workers connect ansatz expressibility to gradient magnitude

Going further

Texts and reviews. McClean, Boixo, Smelyanskiy, Babbush, Neven, "Barren plateaus in quantum neural network training landscapes," Nature Communications 9, 4812 (2018) · Cerezo, Sone, Volkoff, Cincio, Coles, "Cost function dependent barren plateaus in shallow parametrized quantum circuits," Nature Communications 12, 1791 (2021) · Wang, Fontana, Cerezo, Sharma, Sone, Cincio, Coles, "Noise-induced barren plateaus in variational quantum algorithms," Nature Communications 12, 6961 (2021) · Holmes, Sharma, Cerezo, Coles, "Connecting ansatz expressibility to gradient magnitudes and barren plateaus," PRX Quantum 3, 010313 (2022) · Cerezo et al., "Variational Quantum Algorithms," Nature Reviews Physics 3, 625 (2021) · Preskill, "Quantum Computing in the NISQ era and beyond," Quantum 2, 79 (2018) · Nielsen and Chuang, Quantum Computation and Quantum Information (2010), ch. 2, 8 (density operators, channels) · Watrous, The Theory of Quantum Information (2018) (Haar measure, Weingarten calculus).

Provenance. The gradient moment formulae are from McClean et al., Nat. Commun. 9, 4812 (2018); the cost-locality scaling from Cerezo et al., Nat. Commun. 12, 1791 (2021); the noise-induced plateau from Wang et al., Nat. Commun. 12, 6961 (2021); the expressibility bound from Holmes et al., PRX Quantum 3, 010313 (2022). A verifier should recompute the first and second Weingarten moments (Watrous), confirm $\mathbb{E}[\partial_k C]=0$ via $\mathrm{Tr}[V_k,\mathbb{1}]=0$ , verify the $d/(d^2-1)\sim 2^{-n}$ scaling of equation (5), and check the shot cost $M\sim 1/\mathrm{Var}$ against standard shot-noise estimation. The worked-example figures are exact: $2^{-4}=6.25\times10^{-2}$ , $2^{-20}=9.54\times10^{-7}$ , $2^{-50}=8.88\times10^{-16}$ ; the noise example uses $0.99^{L}$ with depth budget $\ln(10^{-4})/\ln(0.99)\approx916$ , and $\alpha=-\ln(0.99)=0.01005$ . All standard quantities (Haar moments, depolarizing channel form, Chebyshev-type concentration) match these textbook sources.

Prerequisites. Haar Measure and Unitary t-Designs · Density Operators and Partial Trace · Parameter-Shift Rule · Variational Quantum Eigensolver (VQE) · Quantum Channels and Noise Models. Leads to. Classical Surrogates and Dequantization · Magic State Distillation · Topological Quantum Computation · Quantum Neural Network Expressibility.

Ask the tutor: Ask me to walk the Weingarten derivation of the $2^{-n}$ variance line by line, to work out the shot budget or noise depth-limit for a circuit of your size, or to compare the expressibility, locality, entanglement, and noise routes to a barren plateau.

Is a barren plateau the same thing as a local minimum?

No, and conflating them is the most common misconception. A local minimum is a specific point where the gradient is zero but the surrounding landscape still has structure — you are stuck in a basin, and a lucky perturbation or a different optimizer might escape it. A barren plateau is an entire exponentially large region where the gradient is essentially zero everywhere, to a precision that shrinks like $2^{-n}$ . There is no basin to climb out of because there is no structure at all: the cost is nearly constant. That is why the two demand opposite responses — you fight local minima with better search heuristics, but you can only fight barren plateaus by changing the circuit or the cost so the landscape stops being flat in the first place.

What is a unitary 2-design, and why does it matter here?

A unitary $t$ -design is an ensemble of circuits whose first $t$ statistical moments exactly match those of the fully random Haar distribution. A 2-design matches the first two moments — averages of products of up to two circuit copies and two conjugates. That is precisely what the gradient's mean (one copy of $U$ ) and variance (two copies) depend on, so once your ansatz reaches an approximate 2-design those two quantities take their Haar values, which Weingarten calculus shows are exponentially suppressed. Crucially you do not need genuine randomness: generic expressive hardware-efficient circuits approach a 2-design with only polynomial depth, which is why the plateau shows up in ordinary, non-exotic ansätze rather than only in contrived ones.

If almost every initialization is on the plateau, why do small experiments still train?

Because the plateau is an asymptotic, large- $n$ statement, and small circuits are far from that limit. At $n=4$ the gradient variance is $2^{-4}\approx0.06$ — a perfectly healthy slope that a handful of shots resolves, as the worked example shows. The exponential suppression only becomes fatal once $n$ is large enough that $2^{-n}$ falls below your shot-noise floor, which for a million-shot budget happens near $n=20$ . This is exactly what makes barren plateaus so dangerous in practice: a method validated on a few qubits looks like it works, and the failure only appears when you scale up to the sizes where a quantum computer would actually be useful.

Does switching to a gradient-free optimizer avoid the problem?

No. It is tempting to think that because the plateau is defined through the gradient, a method that never computes one — Nelder–Mead, particle swarm, SPSA, Bayesian optimization — might sidestep it. But every one of these ultimately decides where to move by comparing cost values at nearby points, and on a plateau those values differ by only $\mathcal{O}(2^{-n/2})$ , buried in shot noise for exactly the reason the gradient is. The obstruction is information-theoretic: the information needed to descend is exponentially expensive to measure, and no reformulation of the search — gradient-based, gradient-free, or second-order — changes the number of shots that measurement requires.

How is a noise-induced plateau different from the standard one?

They share the exponential signature but arise from different physics. The standard, expressibility-induced plateau is a property of an ideal noiseless circuit random enough to be a 2-design; it depends on how expressive the ansatz is and scales as $2^{-n}$ in the qubit count. A noise-induced plateau needs no randomness at all — it comes from hardware imperfections that, layer by layer, contract the state toward maximally mixed, so the variance decays as $(1-p)^{L}$ in the depth $L$ . The practical distinction matters: you fight the expressibility plateau by making the ansatz less expressive (locality, structure), but you fight the noise plateau by making the circuit shallower or the hardware cleaner, since it is depth and error rate, not expressibility, that drive it — and no initialization removes it.

Can you just rescale the cost or use a huge learning rate to climb the plateau?

No — the flatness is in the measured landscape, not a units convention you can redefine away. Rescaling $H$ by a constant multiplies both the gradient and its shot noise by the same factor, leaving the signal-to-noise ratio, and therefore the shot cost, unchanged. A large learning rate does not help either: you cannot take a confident step in a direction you cannot reliably measure, and on a plateau the measured direction is dominated by statistical noise, so a big step is just a big random jump. The only genuine levers are those that change the landscape itself — a local cost, a shallow or structured ansatz, an off-Haar initialization — because they attack the variance $\mathrm{Var}[\partial_k C]$ rather than its cosmetic scale.

Does escaping the barren plateau guarantee a quantum advantage?

Not necessarily, and this is one of the sharpest open questions in the field. The structural features that make a circuit trainable — locality, shallow depth, limited entanglement, restricted expressibility — are, in several recent results, the very features that also make its outputs efficiently classically simulable. In other words, some ansätze proven to have no barren plateau turn out to compute nothing a laptop could not. This does not mean the two goals are always incompatible, but it does mean that "we made it trainable" and "it beats classical methods" are separate claims that must each be established. Reconciling trainability with genuine quantum advantage — finding the regime that is both learnable and classically hard — is where much of the current effort is aimed.