Fleshing out Ofir Gorodetsky's comment: if we define $G(s) = \sum_{n=1}^\infty \mu(n)\phi(n)n^{-s}$, then we have $G(s) = F(s)/\zeta(s-1)$ where $$ F(s) = \prod_p \bigg( 1 - \frac1{p^s-p} \bigg) $$ is absolutely convergent for $\Re s>1$. The rightmost singularities of $G(s)$ are therefore at the points $1+\rho$ where $\rho$ denotes nontrivial zeros of $\zeta(s)$. Assuming the Riemann hypothesis, we thus expect $f(x)/x^{3/2} = x^{-3/2} \sum_{n\le x} \mu(n)\phi(n)$ to have a limiting logarithmic distribution, which will be the same as the distribution of the random variable $$ \sum_{\gamma} \bigg |\frac{F(3/2+i\gamma)}{(3/2+i\gamma)\zeta'(1/2+i\gamma)}\bigg| \Re Z_\gamma, $$ where $\gamma$ runs over the positive imaginary parts of zeros of $\zeta(s)$, and $\{Z_\gamma\}$ is an independent set of random variables each uniformly distributed on the unit circle in $\Bbb C$. This distribution will be roughly bell-shaped but not normal, indeed probably decaying faster than a normal distribution. The function $f(e^y)/e^{3y/2}$ will be an almost periodic function with mean $0$, but will presumably be unbounded above and unbounded below. Numerical computations up to $x=10^6$ support these claims (other than the unboundedness, which will be extremely gradually realized).

This sort of statement, for other functions, goes back to Littlewood's oscillation theorems and continues today in the subfield of comparative prime number theory. See "Limiting distributions of the classical error terms of prime number theory" by Akbary, Ng, and Shahabi for a general framework that addresses problems such as this.

Fleshing out Ofir Gorodetsky's comment: if we define $G(s) = \sum_{n=1}^\infty \mu(n)\phi(n)n^{-s}$, then we have $G(s) = F(s)/\zeta(s-1)$ where $$ F(s) = \prod_p \bigg( 1 - \frac1{p^s-p} \bigg) $$ is absolutely convergent for $\Re s>1$. The rightmost singularities of $G(s)$ are therefore at the points $1+\rho$ where $\rho$ denotes nontrivial zeros of $\zeta(s)$. Assuming the Riemann hypothesis, we thus expect $f(x)/x^{3/2} = x^{-3/2} \sum_{n\le x} \mu(n)\phi(n)$ to have a limiting logarithmic distribution, which will be the same as the distribution of the random variable $$ \sum_{\gamma} \bigg |\frac{F(3/2+i\gamma)}{(3/2+i\gamma)\zeta'(1/2+i\gamma)}\bigg| \Re Z_\gamma, $$ where $\gamma$ runs over the positive imaginary parts of zeros of $\zeta(s)$, and $\{Z_\gamma\}$ is an independent set of random variables each uniformly distributed on the unit circle in $\Bbb C$. (The independence of the $\{Z_\gamma\}$ is based on our belief that the imaginary parts of the nontrivial zeros of $\zeta(s)$ are linearly independent over the rationals.)

This distribution will be roughly bell-shaped but not normal, indeed probably decaying faster than a normal distribution. The function $f(e^y)/e^{3y/2}$ will be an almost periodic function with mean $0$, but will presumably be unbounded above and unbounded below. Numerical computations up to $x=10^6$ support these claims (other than the unboundedness, which will be extremely gradually realized).

This sort of statement, for other functions, goes back to Littlewood's oscillation theorems and continues today in the subfield of comparative prime number theory. See "Limiting distributions of the classical error terms of prime number theory" by Akbary, Ng, and Shahabi for a general framework that addresses problems such as this.

Fleshing out Ofir Gorodetsky's comment: if we define $G(s) = \sum_{n=1}^\infty \mu(n)\phi(n)n^{-s}$, then we have $G(s) = F(s)/\zeta(s-1)$ where $$ F(s) = \prod_p \bigg( 1 - \frac1{p^s-p} \bigg) $$ is absolutely convergent for $\Re s>1$. The rightmost singularities of $G(s)$ are therefore at the points $1+\rho$ where $\rho$ denotes nontrivial zeros of $\zeta(s)$. Assuming the Riemann hypothesis, we thus expect $f(x)/x^{3/2} = x^{-3/2} \sum_{n\le x} \mu(n)\phi(n)$ to have a limiting logarithmic distribution, which will be the same as the distribution of the random variable $$ \sum_{\gamma} \bigg |\frac{F(3/2+i\gamma)}{3/2+i\gamma}\bigg| \Re Z_\gamma, $$ where $\gamma$ runs over the positive imaginary parts of zeros of $\zeta(s)$, and $\{Z_\gamma\}$ is an independent set of random variables each uniformly distributed on the unit circle in $\Bbb C$. This distribution will be roughly bell-shaped but not normal, indeed probably decaying faster than a normal distribution. The function $f(e^y)/e^{3y/2}$ will be an almost periodic function with mean $0$, but will presumably be unbounded above and unbounded below. Numerical computations up to $x=10^6$ support these claims (other than the unboundedness, which will be extremely gradually realized).

This sort of statement, for other functions, goes back to Littlewood's oscillation theorems and continues today in the subfield of comparative prime number theory. See "Limiting distributions of the classical error terms of prime number theory" by Akbary, Ng, and Shahabi for a general framework that addresses problems such as this.

Fleshing out Ofir Gorodetsky's comment: if we define $G(s) = \sum_{n=1}^\infty \mu(n)\phi(n)n^{-s}$, then we have $G(s) = F(s)/\zeta(s-1)$ where $$ F(s) = \prod_p \bigg( 1 - \frac1{p^s-p} \bigg) $$ is absolutely convergent for $\Re s>1$. The rightmost singularities of $G(s)$ are therefore at the points $1+\rho$ where $\rho$ denotes nontrivial zeros of $\zeta(s)$. Assuming the Riemann hypothesis, we thus expect $f(x)/x^{3/2} = x^{-3/2} \sum_{n\le x} \mu(n)\phi(n)$ to have a limiting logarithmic distribution, which will be the same as the distribution of the random variable $$ \sum_{\gamma} \bigg |\frac{F(3/2+i\gamma)}{(3/2+i\gamma)\zeta'(1/2+i\gamma)}\bigg| \Re Z_\gamma, $$ where $\gamma$ runs over the positive imaginary parts of zeros of $\zeta(s)$, and $\{Z_\gamma\}$ is an independent set of random variables each uniformly distributed on the unit circle in $\Bbb C$. This distribution will be roughly bell-shaped but not normal, indeed probably decaying faster than a normal distribution. The function $f(e^y)/e^{3y/2}$ will be an almost periodic function with mean $0$, but will presumably be unbounded above and unbounded below. Numerical computations up to $x=10^6$ support these claims (other than the unboundedness, which will be extremely gradually realized).

This sort of statement, for other functions, goes back to Littlewood's oscillation theorems and continues today in the subfield of comparative prime number theory. See "Limiting distributions of the classical error terms of prime number theory" by Akbary, Ng, and Shahabi for a general framework that addresses problems such as this.