Trainability Lab
Barren Plateau Demo
Explore how gradient variance shrinks as variational circuits grow. This lightweight simulator models the statistical pattern: global costs collapse exponentially with width, while local costs degrade more gently.
Cost Scope
6 qubits, depth 8, global cost, modeled variance 5.98e-3.
Learning mode
When the learning signal disappears
A barren plateau is a volume knob stuck near silence: gradients become too small to guide learning.
Compare global and local costs at the same width, then increase depth.
Definitions on demand
Gradient variance
How broadly gradient values vary across random parameter samples.
Barren plateau
A region where gradients concentrate near zero, making optimization direction difficult to detect.
Log scale
An axis where equal spacing represents multiplication, used here to reveal changes across many orders of magnitude.
Gradient Variance vs. Width
Sampled Gradients