This effect was reported on `https://habr.com/ru/articles/812405/`

. The exposition contains glaring errors, but the effect is real! Here is the plot:

Here I take normalized differences of zeros of ζ-function $d_n≔(Z_{n+1}-Z_n)\log(|Z|/2π)$ (after this normalization they are ∼1 in magnitude, with the same argument), multiply them by the mod-4-character i$^n$, and plot the running sums. These sums change by about ±1 or ±i, and in the scale of the plot above these differences are small, so the corresponding points “seem to be on a continuous curve” — but *a priori* one does not expect this curve to be “very special”. **Question:** is there any explanation for the observed pattern?

**Update:** I added $||$ in $|Z|$ above, and explained why the resulting points are expected “to go along a curve” (which might have been “Brownian”!). (BTW, $|Z|$ is a shortcut for “something like $|Z_n|$ or $|Z_{n+1}|$” — and I average them in the script below.)

Myself, a couple of years ago I noticed somewhat-similar oscillations of the average value of $|ζ'|$ at zeros when averaged on log-scale — however, the effect discovered by Tzimie seems to be much more pronounced, — and maybe even has a chance to explain what I observed. My

`gp/PARI`

code (using 2M zeros calculated with 35 decimal places):

```
fileno_r = fileopen("zeros6-more-precise.gz","r");
{ global(n=0, BS=4, xds = vector(BS), n_b = 0, x_prev = 0, TOT=vector(3), CLR=vector(3));
while(x = fileread(fileno_r), \\ invariant: first <= last <= first + vals2 - 1.
x = eval(x);
n++; n_b++; if(n_b>BS, n_b=1); xds[n_b] = (x-x_prev)*log((x+x_prev)/4/Pi); x_prev = x;
if(n_b == 1 && n>1, \\ have 4 diffs; "color" the variation: non-const characters: 1 -1 1 -1, 0 1 0 -1, 1 0 -1 0
CLR[1] = xds[2] - xds[3] + xds[4] - xds[1];
CLR[2] = xds[2] - xds[4];
CLR[3] = xds[1] - xds[3];
TOT += CLR;
print(TOT[1], "\t", TOT[2], "\t", TOT[3], "\t", CLR[1], "\t", CLR[2], "\t", CLR[3]);
);
);
}
```

(I did not try to use characters for different moduli — except 2 — calculated by the script above; “with mod 2” the behavior is less anomalous.)

`gp/PARI`

reads files “by evaluation”, I recommend checking that the downloaded file “contains numbers only”. Like in`zcat FILE.gz | perl -wlne "die if /[^\d.]/"`

(Use single quotes on *nix.) $\endgroup$