Define & explore C
Rows are powers of z; columns are independent GUE samples shared across coefficients.
Up to 60 seconds per realization; the plot accumulates continuously. The app chooses n automatically; larger budgets allow larger matrices.
Browse the 12 discovered shapes ↗
Click an entry to select it; type a complex value such as 1+2i. Row and column buttons select a whole row or column.
Drag to change C and the limiting density live. Sliders stay where you leave them; changing the selection resets their offsets. Additive moves can activate zero entries.
Paste / copy C
One row per line, entries separated by spaces or commas.
Sampling & interpretation
A bank of 65,536 independent GUE realizations is fixed for the lifetime of this page. Every change to C reuses the same bank. Changing n takes a nested leading submatrix of each Gaussian field, normalized by 1/√n. Diagonal variance is 1/n; real and imaginary off-diagonal variances are 1/(2n).
Small n gives a fast preview while editing. After a pause, measured update times determine how far n can grow within your time budget, with an internal limit dn ≤ 1024. Over-budget jobs are interrupted and retried smaller. Browser scheduling can overshoot the target; initial worker loading is separate. Size changes preserve a populated plot. Once at least 2,048 roots at 75% or more of the current n are available, older roots below that cutoff are retired. Roots from different n values are pooled, and the sampled n range is displayed. Each realization is counted once for the current C. Up to two million roots are retained.
Repeated samples estimate the finite-n mean spectral measure, not the exact large-n limit. More samples reduce noise; increasing n addresses finite-size effects. The histogram uses a square-root stretch of bin counts. Scatter points have sharp edges: 5 pixels below 1,024 visible roots, 3 pixels below 16,384, then 1 pixel. Gamma and exposure map counts to opacity, keeping sparse points visible and dense overlaps distinguishable. No blur is applied. Gaussian fields support indices up to 4000; requested entries use a bounded cache and are reproduced identically if evicted.
Starting sampler…
Limiting spectral measure · n → ∞
Petaloid law · live as you drag C
Low density High densityFixed brightness scale · equal axis scales
Computing limiting measure…
Numerical grid evaluation of the limiting measure. Mass on curves appears as thin bright bands; gold marks an atom at zero. Band width depends on grid resolution. The grid sharpens to 512 × 512, then 1024 × 1024 when you stop dragging.
Formula & interpretation
For fⱼ(z) = ∑ₖ Cₖⱼzᵏ, compute σ² = ∑ⱼ |fⱼ|² and g = ∑ⱼ fⱼ². The display evaluates the logarithmic potential in Theorem 2.13 and its grid Laplacian, with any exact factor zᵐ removed and its atom m/d shown separately. This also displays mass supported on curves. Area density is restricted to A₂D using the ellipse condition in Corollary 2.14. A₁D crossings are retained within one grid step to show curve-supported mass. The raster is not blurred outside those cells. Boundary discretization and common polynomial zeros can affect the approximate visible mass; it is not renormalized.
Shah & Zeng · Spectra of Random Polynomial Matrices: the Petaloid Law
Empirical spectral density
Fixed GUE bank · one coefficient matrix
Low density High density16 × 16 bins · adaptiveEqual axis scales
0 / 1,024 visible roots · 1,024 more to refine
The display refines automatically when you stop editing.