Computing ζ(3) from a Rational Start: An Adjoint System for the Apéry Equation
2026-10-19
Earlier posts computed $\zeta(3)$ from a polynomial ODE by an adaptation law. That version started from Taylor values at a rational point. Those values are infinite sums, not rational numbers, so the start was not exact. This post gives a version that starts only from rational constants, by running an adjoint system.
What is wrong with the earlier start
The Apéry sequences $a_n$ and $b_n$ satisfy the same recurrence, and their generating functions $A(z)$, $B(z)$ satisfy the Apéry ODE, with $B$ having a constant forcing $6$ coming from $b_1=6$. A polynomial system that starts at a point $z_0$ inside the disc needs the values $A(z_0), A'(z_0), \dots$ as initial data. These are infinite sums $\sum_n a_n z_0^{n-k}\cdot(\text{falling factorial})$. They are rational only by accident. Truncating the sums gives rational numbers that are approximations, and an approximate start does not prove anything about the limit.
The adjoint idea
A conserved quantity removes the need for the values at $z_0$. If $\lambda$ solves the adjoint equation of a linear system $Y'=-M(t)Y$, then $\lambda(t)\cdot Y(t)$ is constant. Start $\lambda$ at a rational vector $q$. The constant equals $q\cdot Y(z_0)$, and the irrational values $Y(z_0)$ never enter the machine. The flow runs from $z_0$ to an origin where the data are known exactly.
For $\zeta(3)$ the target is a ratio at the singular point $\rho=17-12\sqrt2$, not a linear functional at $z_0$, so one needs a transfer matrix rather than a single adjoint:
- Two moving points. The upper point $b$ solves $b'=bD(b)$, with $D(z)=1-34z+z^2$, so its speed vanishes only at $\rho$ and is approached asymptotically. The lower point $a=re^{-\kappa t}$ runs to $0$.
- The transfer matrix $H(t)$, defined by $Y(b(t))=H(t)Y(a(t))$, starts at the identity $H(0)=I$ because $a=b=r$. No data at $r$ are needed.
- Normalization. Write $G=HS$ with $S=\mathrm{diag}(1,a,a,1)$, then $g=G/\sum G$. The entries of $H$ diverge as $b\to\rho$, but the normalized matrix has a finite limit, and $g_{23}/g_{20}\to\zeta(3)$.
The unbounded linear form and its adjoint
Let $D(z)=1-34z+z^2$, and write $p=z^2D(z)$, $q=z(3-153z+6z^2)$, $r=1-112z+7z^2$, $s=z-5$. Take $Y=(A,A',A'',B,B',B'',1)^T$, where $A'=dA/dz$. The Apéry equation divided by $z$ is $pA'''+qA''+rA'+sA=0$, and the same operator applied to $B$ gives the constant $6$. Run the clock $\dot z=p(z)$. Then $\dot Y=M(z)Y$ with
$$M(z)=\begin{pmatrix}0&p&0&0&0&0&0\\0&0&p&0&0&0&0\\-s&-r&-q&0&0&0&0\\0&0&0&0&p&0&0\\0&0&0&0&0&p&0\\0&0&0&-s&-r&-q&6\\0&0&0&0&0&0&0\end{pmatrix}.$$The mask $\lambda=(\lambda_1,\dots,\lambda_7)$ satisfies $\dot\lambda=-\lambda M$, which is
$$\dot\lambda_1=s\lambda_3,\quad \dot\lambda_2=-p\lambda_1+r\lambda_3,\quad \dot\lambda_3=-p\lambda_2+q\lambda_3,$$$$\dot\lambda_4=s\lambda_6,\quad \dot\lambda_5=-p\lambda_4+r\lambda_6,\quad \dot\lambda_6=-p\lambda_5+q\lambda_6,\quad \dot\lambda_7=-6\lambda_6.$$Then $\lambda\cdot Y$ is constant. With a rational $\lambda(0)$ this gives a rational-weight combination of the values $Y(z_0)$. Those values are infinite sums, so one moving point alone does not give the ratio $B''/A''$ with rational data. The second point below fixes this.
The twenty-state system
The two moving points are $a(t)$ with $a'=-\kappa a$, and $b(t)$ with $b'=bD(b)$, where $\kappa>0$ is rational. The transfer matrix $H$ satisfies $Y(b)=HY(a)$. Set $S=\mathrm{diag}(1,a,a,1)$, $G=HS$, $g=G/\sum_{ij}G_{ij}$, and $R=1/D(a)$. The complete system has 20 coordinates, $(a,\,b,\,R,\,g_{ij}\ (0\le i,j\le3),\,y)$:
$$a'=-\kappa a,\qquad b'=bD(b),\qquad R'=\kappa\,a\,(-34+2a)\,R^2,$$$$F=K(b)\,g+g\,L(a,R),\qquad \sigma=\sum_{i,j}F_{ij},\qquad g'=F-\sigma\,g,\qquad y'=4\,(g_{23}-y\,g_{20}),$$where
$$K(b)=\begin{pmatrix}0&D(b)&0&0\\0&0&D(b)&0\\ b(5-b)&b(27-3b)&b(51-3b)&6b\\0&0&0&0\end{pmatrix},\qquad L(a,R)=\kappa\begin{pmatrix}0&a&0&0\\0&-1&1&0\\ R(5-a)&aR(27-3a)&-1+aR(51-3a)&6R\\0&0&0&0\end{pmatrix}.$$The initial values, all together, are
$$a(0)=b(0)=r,\qquad R(0)=\frac1{D(r)},\qquad g(0)=\frac{\mathrm{diag}(1,r,r,1)}{2+2r},\qquad y(0)=0,$$with $r$ any rational number in $(0,\rho)$, where $\rho=17-12\sqrt2$. For $r=1/100$ this gives $R(0)=10000/6601$ and $g(0)=\mathrm{diag}(50/101,\,1/202,\,1/202,\,50/101)$. Every initial value is rational, and no Taylor value appears.
Integrating this system with $\kappa=1$ and $r=1/1000$ gives $y(30)-\zeta(3)\approx6.7\times10^{-15}$ in floating point. This is a numerical check of exact rational data, not a proof.
Why the system is correct
The matrix $K$ is the first-order form of the Apéry equation, with the constant $1$ in the fourth coordinate carrying the forcing $6$. The matrix $L$ is the lower point’s generator after the scaling $S$, and the identity $G'=KG+GL$ with $G=HS$ follows from $H'=K(b)H+\kappa H M(a)$, $M=K/D$. Normalization gives $g'=F-\sigma g$ exactly, since $\sum g=1$ is preserved. I checked numerically that the normalized transfer matrix obtained from $H'=K(b)H+\kappa HM(a)$ agrees with the $g$ of the system above to $10^{-13}$.
The readout $y$ is a relaxation of the ratio, $y'=4g_{20}\,(g_{23}/g_{20}-y)$, whose equilibrium is the ratio and whose rate $4g_{20}$ is positive. Division is avoided this way.
Numbers
Integrating the system from the rational start above, with $\kappa=1$:
| $r$ | $t$ | $y(t)-\zeta(3)$ |
|---|---|---|
| $1/1000$ | 20 | $1.5\times10^{-11}$ |
| $1/1000$ | 30 | $6.7\times10^{-15}$ |
| $1/100$ | 20 | $1.3\times10^{-10}$ |
| $1/100$ | 30 | $2.9\times10^{-14}$ |
These are floating-point integrations of an exact rational start. They are a check of the claim, not a proof of it.
What is still cited
The selection step is not derived from the system alone. The limit equations have a family of equilibria, and the trajectory picks out $\zeta(3)$. The reason is the generating-function characterization: $\zeta(3)$ is the unique constant $c$ for which $B-cA$ has radius $1/\rho$ rather than $\rho$. That characterization follows from the exact remainder formula $e_n=-a_n\sum_{k>n}\frac{6}{k^3a_ka_{k-1}}$, and the limit $b_n/a_n\to\zeta(3)$ is Apéry’s theorem, which we cite. So the construction is exact in its start and its equations, and it uses Apéry’s theorem for the identification of the limit.
The adaptation law, restated
The earlier adaptation law $\dot\zeta=\varepsilon(\rho-\zeta)$ tracks the ratio $\rho=B''/A''$. The relaxation $y'=4g_{20}(g_{23}/g_{20}-y)$ above is the same kind of tracking, applied to the normalized transfer matrix. The two agree on the limit, $\zeta(3)$. The difference is the start: the earlier one needed irrational values at $z_0$, the new one needs only rational constants.
「知是行之始,行是知之成。」— 王阳明《传习录·下》
Knowing is the beginning of acting; acting is the completion of knowing.