Goal
Approximate the de Bruijn–Newman constant Λ by studying the heat flow of the Riemann zeta function.
Overview
- Define these equations, describing the heat flow of the Riemann Equation
- such that , the zeroes of are real. This is known as the de Bruijn–Newman constant.
- was proved by Brad Rodgers and Terence Tao.
- The Riemann Hypothesis is equivalent to , hence, given the above, to .
- was proved as part of the polymath project.
- was later proved by Dave Platt and Tim Trudgian
- is the Fourier representation of
- , with the Riemann xi function. So the zeroes of are for the nontrivial zeroes of , and they are all real iff the Riemann Hypothesis holds.
- is the evolution of under the backwards heat equation .
Properties of the Riemann Zeta
- satisfies the functional equation
Plan to bound
- Choose a ( suffices for the time being )
- Take zeroes of
- Computationally estimate how the evolve backwards in time (towards ), getting an estimate of the time at which the trajectories of two neighbouring collide
- Reproduce the above with , obtaining better estimates of .
- Interpretation: a collision at time means that pair leaves the real axis as a complex conjugate pair for , so . Since is already known, a collision at would disprove the Riemann Hypothesis, while finding no collision down to for finitely many zeroes is consistent with it but proves nothing about the remaining zeroes.
Pre-requisites
Finding zeros
What are the best methods known to find zeroes of the . I have no clue how to find zeroes given the lack of knowledge about the properties of and , namely bounds on the change rate of the function.
- Plot the function for and . Visualise these zeroes. I can do this visually, but I have no confidence this gives me a good approximation of zeroes due to continuity concerns.
Modelling the time evolution of the zeroes
Do we have examples ( in code or books ) of approximating path trajectories using heatflow?Andrea suggested Runge-Kutta. Sounds good
Next steps
Both of the points below follow Andrea’s email suggestion
Equation of motion of the zeroes
Let be a simple zero, . Differentiating in and using :
where the last step uses the Hadamard factorisation of ; this is equation (5) of the Rodgers and Tao paper linked above, valid for simple zeroes. The sum is taken symmetrically (, ) since it converges only conditionally. Real zeroes repel as increases, so backwards in time neighbours attract and can collide. Two caveats for numerics: truncating to zeroes drops the conditionally convergent tail, which is not negligible, so the truncated system needs a tail correction or should only be trusted over short times; and near a collision the vector field blows up like .
Runge-Kutta
For an ODE , here and , the classical fourth order Runge-Kutta step of size (negative, to go back in time) is
Since the field is stiff near collisions, use an adaptive embedded pair (Dormand-Prince RK45, as in scipy.integrate.solve_ivp), which estimates the local error from two orders and shrinks as two zeroes approach. Integrate in complex arithmetic so a colliding pair can continue as a conjugate pair, and stop or flag when drops below a tolerance; the collision time is then refined by extrapolating the gap, which closes like , so the squared gap is locally linear in . Since the truncated ODE drifts, re-solve near the predicted zeroes at intermediate times (predictor-corrector) rather than trusting the ODE alone.
Root finding for at fixed
- is real on the real axis and even in , and for (in particular ) all its zeroes are real, so the problem is real 1D root finding on .
- Bracketing: evaluate on a grid much finer than the mean zero spacing, which near height is about (from the density of zeta zeroes at height and ). Each sign change brackets a zero.
- Refinement: Brent’s method (bisection with secant and inverse quadratic interpolation) converges superlinearly and is guaranteed within a bracket; Newton’s method with is faster once close.
- Completeness check: a coarse grid can miss a pair of close zeroes (no sign change). Count the zeroes in independently with the argument principle, around a thin rectangle, and compare with the number of brackets found.
- Evaluation: decays doubly exponentially, so the integral can be truncated at moderate and computed by quadrature in high precision (for example
mpmath). For large the integrand oscillates rapidly and direct quadrature loses precision through cancellation; there the Polymath project used Riemann-Siegel type asymptotic approximations of (see the writeup linked above).