infsup

by Zinan Huang 🌸

From a Triple Integral to a Polynomial ODE for ζ(3), via Feynman's Trick

2026-07-15


How does one turn the number $\zeta(3) = 1 + 1/8 + 1/27 + 1/64 + \cdots$ into the output of a polynomial dynamical system? This post walks through one route from start to finish: begin with a classical triple integral, apply Feynman’s trick twice to collapse it into a single integral, split that integral to remove a singularity, and finally read off the polynomial ODE whose solution converges to $\zeta(3)$.

The result is an 8-variable polynomial initial-value problem (PIVP) — a system $\dot{y} = p(y)$ with polynomial right-hand side and explicit initial conditions — whose accumulator variable $s_3(t) \to \zeta(3)$ as $t \to \infty$. Every step is elementary, and every step is written out.

An earlier post built a 5-variable bounded PIVP for $\zeta(3)$ starting from the Fermi integral $\int_0^\infty x^2/(1+e^x)\,dx$. The present construction starts from a different integral — the Bose–Einstein form $\int_0^\infty t^2/(e^t - 1)\,dt$ — and handles its singularity at the origin by splitting and substituting rather than by rewriting the integrand. The two routes produce genuinely different ODE systems.

1. The Starting Point: A Triple Integral

The identity

$$\int_0^1 \int_0^1 \int_0^1 \frac{1}{1 - xyz}\,dx\,dy\,dz \;=\; \zeta(3)$$

follows by expanding the integrand as a geometric series. Since $\lvert xyz \rvert < 1$ on the open unit cube,

$$\frac{1}{1 - xyz} = \sum_{n=0}^{\infty} (xyz)^n,$$

and integrating term by term gives

$$\sum_{n=0}^{\infty} \int_0^1 \int_0^1 \int_0^1 (xyz)^n\,dx\,dy\,dz = \sum_{n=0}^{\infty} \frac{1}{(n+1)^3} = \sum_{k=1}^{\infty} \frac{1}{k^3} = \zeta(3).$$

This is a clean formula, but a triple integral is not directly useful for building an ODE. The next step is to collapse it into a single integral.

2. Feynman’s Trick: From $1/k$ to $1/k^3$

Feynman’s trick — differentiation under the integral sign — converts powers of $1/k$ into integrals against exponentials.

Starting point. The Laplace transform of the constant function gives

$$\int_0^{\infty} e^{-kt}\,dt = \frac{1}{k}, \qquad k > 0.$$

First differentiation. Differentiate both sides with respect to the parameter $k$:

$$\int_0^{\infty} \frac{\partial}{\partial k}\bigl(e^{-kt}\bigr)\,dt = \int_0^{\infty} (-t)\,e^{-kt}\,dt = -\frac{1}{k^2}.$$

So $\int_0^{\infty} t\,e^{-kt}\,dt = 1/k^2$.

Second differentiation. Differentiate once more:

$$\int_0^{\infty} (-t^2)\,e^{-kt}\,dt = -\frac{2}{k^3},$$

which gives the integral representation

$$\frac{1}{k^3} = \frac{1}{2}\int_0^{\infty} t^2\,e^{-kt}\,dt.$$

(The same formula follows from the Gamma function identity $\Gamma(s)/k^s = \int_0^\infty t^{s-1} e^{-kt}\,dt$ with $s = 3$ and $\Gamma(3) = 2$. Feynman’s trick is the elementary derivation that avoids invoking the Gamma function.)

3. Summing into a Single Integral

Substitute the integral representation into $\zeta(3) = \sum_{k=1}^\infty 1/k^3$:

$$\zeta(3) = \sum_{k=1}^{\infty} \frac{1}{2}\int_0^{\infty} t^2\,e^{-kt}\,dt = \frac{1}{2}\int_0^{\infty} t^2 \sum_{k=1}^{\infty} e^{-kt}\,dt.$$

The inner sum is a geometric series: $\sum_{k=1}^{\infty} e^{-kt} = e^{-t}/(1 - e^{-t}) = 1/(e^t - 1)$. Therefore

$$2\,\zeta(3) = \int_0^{\infty} \frac{t^2}{e^t - 1}\,dt.$$

This is the integral representation of $\zeta(3)$ via the Bose–Einstein kernel. It is elegant, but it has a singularity at $t = 0$: the integrand behaves like $t^2 / t = t$ near zero (since $e^t - 1 \approx t$), which is integrable, but the function $1/(e^t - 1)$ itself blows up. An ODE system that starts integration at $t = 0$ cannot tolerate a pole in the initial data.

4. Splitting the Integral

To work around the singularity, split at $t = 1$:

$$2\,\zeta(3) = \underbrace{\int_0^1 \frac{t^2}{e^t - 1}\,dt}_{I_1} \;+\; \underbrace{\int_1^{\infty} \frac{t^2}{e^t - 1}\,dt}_{I_2}.$$

Each piece will be transformed so that (a) the domain of integration becomes $[0, \infty)$, and (b) the integrand is finite at the new origin.

Transforming $I_1$

The pole sits at $t = 0$, which is the start of the integration. Two substitutions move and then remove it.

Step 1: Flip. Set $t = 1 - s$, so $dt = -ds$ and $t \in [0,1]$ becomes $s \in [1,0]$:

$$I_1 = \int_0^1 \frac{(1-s)^2}{e^{1-s} - 1}\,ds.$$

Now the singularity is at $s = 1$ (the end of the domain), not the start.

Step 2: Stretch to infinity. Set $s = 1 - e^{-r}$, so $ds = e^{-r}\,dr$ and $s \in [0,1]$ becomes $r \in [0, \infty)$. Then $1 - s = e^{-r}$, and

$$I_1 = \int_0^{\infty} \frac{(e^{-r})^2}{e^{e^{-r}} - 1}\cdot e^{-r}\,dr = \int_0^{\infty} \frac{e^{-3r}}{e^{e^{-r}} - 1}\,dr.$$

At $r = 0$ the integrand equals $1/(e^1 - 1) = 1/(e - 1) \approx 0.582$ — finite, as needed.

Transforming $I_2$

This is simpler. Set $u = t - 1$ to shift $[1, \infty)$ to $[0, \infty)$:

$$I_2 = \int_0^{\infty} \frac{(u+1)^2}{e^{u+1} - 1}\,du.$$

At $u = 0$ the integrand is $(1)^2/(e - 1) = 1/(e-1)$ — also finite.

5. Building the Polynomial ODE

Now comes the key step: express each integrand as a component of a polynomial ODE system.

Subsystem for $I_1$

Define four state variables:

$$\begin{cases} v(t) \;=\; e^{e^{-t}}, \\[4pt] u_n(t) \;=\; \dfrac{e^{-nt}}{v(t) - 1}, \quad n = 1, 2, 3. \end{cases}$$

The integrand of $I_1$ is exactly $u_3(t)$.

Deriving the ODEs. The derivative of $v$ is

$$\dot{v} = e^{e^{-t}} \cdot e^{-t} \cdot (-1) = -v \cdot e^{-t}.$$

The trick is to eliminate $e^{-t}$: since $u_1 = e^{-t}/(v-1)$, we have $e^{-t} = u_1(v-1)$, so

$$\dot{v} = -v\,u_1(v-1).$$

For $u_n$, apply the quotient rule to $e^{-nt}/(v-1)$. After substituting $\dot{v} = -v\,u_1(v-1)$ and simplifying (the $(v-1)^2$ denominators cancel),

$$\dot{u}_n = u_n(v\,u_1 - n).$$

The resulting subsystem is polynomial:

$$\begin{cases} \dot{v} \;=\; -v\,u_1(v - 1), \\[4pt] \dot{u}_1 \;=\; u_1(v\,u_1 - 1), \\[4pt] \dot{u}_2 \;=\; u_2(v\,u_1 - 2), \\[4pt] \dot{u}_3 \;=\; u_3(v\,u_1 - 3). \end{cases}$$

Subsystem for $I_2$

Define three state variables:

$$\begin{cases} r(t) \;=\; \dfrac{1}{1 - e^{-(t+1)}}, \\[6pt] w_n(t) \;=\; \dfrac{(t+1)^n}{e^{t+1} - 1}, \quad n = 1, 2. \end{cases}$$

The integrand of $I_2$ is $w_2(t)$.

Deriving the ODEs. For $r$: differentiating $(1 - e^{-(t+1)})^{-1}$ and using $e^{-(t+1)} = (r-1)/r$ gives

$$\dot{r} = -r(r-1).$$

For $w_n$: note that $w_0 = 1/(e^{t+1} - 1) = r - 1$, so $w_n = (t+1)^n w_0$. The product rule and $\dot{w}_0 = \dot{r} = -r(r-1)$ give

$$\dot{w}_n = n\,w_{n-1} - r\,w_n.$$

The subsystem:

$$\begin{cases} \dot{r} \;=\; -r(r - 1), \\[4pt] \dot{w}_1 \;=\; (r - 1) - r\,w_1, \\[4pt] \dot{w}_2 \;=\; 2\,w_1 - r\,w_2. \end{cases}$$

6. The Complete 8-Variable PIVP

Add an accumulator $s_3$ whose derivative is $(u_3 + w_2)/2$ (since $2\zeta(3) = I_1 + I_2$). The full system, with state vector $(v, u_1, u_2, u_3, r, w_1, w_2, s_3)$:

$$\begin{cases} \dot{v} \;=\; -v\,u_1(v - 1), \\[4pt] \dot{u}_1 \;=\; u_1(v\,u_1 - 1), \\[4pt] \dot{u}_2 \;=\; u_2(v\,u_1 - 2), \\[4pt] \dot{u}_3 \;=\; u_3(v\,u_1 - 3), \\[4pt] \dot{r} \;=\; -r(r - 1), \\[4pt] \dot{w}_1 \;=\; (r - 1) - r\,w_1, \\[4pt] \dot{w}_2 \;=\; 2\,w_1 - r\,w_2, \\[4pt] \dot{s}_3 \;=\; \dfrac{1}{2}(u_3 + w_2). \end{cases}$$

Initial conditions at $t = 0$:

$$\begin{cases} v(0) \;=\; e, \\[4pt] u_1(0) \;=\; u_2(0) \;=\; u_3(0) \;=\; \dfrac{1}{e - 1}, \\[4pt] r(0) \;=\; \dfrac{1}{1 - e^{-1}}, \\[4pt] w_1(0) \;=\; w_2(0) \;=\; \dfrac{1}{e - 1}, \\[4pt] s_3(0) \;=\; 0. \end{cases}$$

Result:

$$\zeta(3) = \lim_{t \to \infty} s_3(t).$$

The right-hand side of every equation is a polynomial in the eight state variables. The initial conditions involve only $e$ and rational functions of $e$. The system is autonomous — time does not appear explicitly — and all trajectories remain bounded.

7. Why This Works: A Sanity Check

Numerical integration (fourth-order Runge–Kutta, $t$ up to 60) gives

$$s_3(60) \approx 1.2020569031594956,$$

which differs from the known value $\zeta(3) = 1.2020569031595942\ldots$ by less than $10^{-13}$.

The convergence is exponential: the error $\lvert s_3(t) - \zeta(3) \rvert$ decays roughly as $e^{-t}$. This is because both integrands decay exponentially — the $I_1$ integrand is $O(e^{-3t})$ and the $I_2$ integrand is $O(e^{-t})$ — so the accumulated integral saturates exponentially fast.

8. Remarks

Dimension count. The 5-variable system in the earlier post is more economical. The present 8-variable system is less compact but has a different virtue: each subsystem ($I_1$ and $I_2$) is derived by a transparent, mechanical procedure — define the integrand’s building blocks, differentiate, eliminate non-polynomial terms using algebraic relations among the state variables. The method generalizes to other integrals of the same flavor.

The role of splitting. The split at $t = 1$ is not mathematically deep — any point in $(0, \infty)$ would work — but it is necessary. The original integrand $t^2/(e^t - 1)$ has a pole at $t = 0$, and a polynomial ODE cannot start from singular initial data. Splitting and substituting is the simplest way to regularize.

Feynman’s trick as a dimension reducer. The triple integral $\iiint 1/(1-xyz)$ lives in three dimensions. Feynman’s trick (or equivalently, the Gamma function identity) collapses it to a single integral, at the cost of introducing an exponential kernel. The ODE construction then handles the exponential by absorbing it into state variables. The entire pipeline — triple integral $\to$ Feynman trick $\to$ single integral $\to$ split $\to$ polynomial ODE — is a systematic dimension-reduction and algebraization machine.


「知者行之始,行者知之成。」—— 王阳明《传习录》

“Knowing is the beginning of doing; doing is the completion of knowing.” — Wang Yangming, Instructions for Practical Living