The Sum Log2 Limit: Why an Unstable System's Poles Mathematically Dictate the Minimum Data Rate for Stability

In 2004, researchers Girish Nair and Robin Evans published a paper in the SIAM Journal on Control and Optimization that proved a hard boundary on physical reality: an unstable system inherently generates information, and a controller must transmit data faster than that generation rate to keep the system from crashing. Published alongside parallel work by Sekhar Tatikonda and Sanjoy Mitter in the IEEE Transactions on Automatic Control, this mathematical proof bridged the gap between Hendrik Bode's control theory and Claude Shannon's information theory.[1][2]

The resulting framework, known as the Data-Rate Theorem, establishes that for any unstable linear system, the minimum data rate required for stability is strictly bounded by a specific formula. That formula requires the data rate to be strictly greater than the sum of the base-2 logarithms of the system's unstable open-loop poles. This is not a limit of current technology or a constraint of poorly written software; it is a mathematical absolute built into the fabric of the physical universe.[1][2]

To understand why this boundary exists, consider a classic inverted pendulum—a stick balanced on a moving cart. Because gravity constantly pulls the stick downward, any microscopic deviation from perfect verticality compounds exponentially over time. The system naturally wants to fall, and as it falls, it generates uncertainty about its exact angle and position.[3]

The speed of this exponential divergence is defined by the system's "poles," which are the eigenvalues of its mathematical state matrix. In discrete time, a pole with an absolute value greater than 1 means the system is inherently unstable. The larger the pole, the faster the stick falls, and the faster the uncertainty about its position grows.[1][3]

To stop the stick from falling, a sensor must measure the angle and send that data across a network to a motor on the cart. But if the network connecting the sensor and motor has limited bandwidth, it can only send a finite number of bits per second. The sensor is forced to quantize the physical reality into a limited digital signal.[2]

The theorem proves that the stick generates uncertainty at a rate of exactly the sum of the base-2 logarithms of its unstable poles. If the network's data rate falls below this threshold, the physical uncertainty grows faster than the digital controller can reduce it. As Tatikonda and Mitter wrote in their 2004 IEEE paper, "The minimum data rate for stabilization is equal to the rate at which the plant generates information."[2]

The math is entirely unforgiving. For a standard 0.5-meter inverted pendulum with an unstable continuous-time pole at 5.42 radians per second, the system generates roughly 7.82 bits of entropy per second. A feedback loop transmitting at 7 bits per second will inevitably fail, no matter how advanced the control algorithm is or how powerful the motor on the cart might be.[3]

This brings us to a central reality of modern engineering: in the era of networked robotics, the limiting factor is often not mechanical engineering or artificial intelligence, but Shannon capacity. The physical dynamics of the object being controlled dictate the absolute minimum bandwidth required from the network.[3]

A common counter-argument from the software industry suggests that a sufficiently advanced artificial intelligence could predict the system's dynamics, thereby reducing the need for constant data transmission. If a neural network knows exactly how the stick falls, the argument goes, the sensor should only need to transmit occasional updates.[3]

The Data-Rate Theorem mathematically forecloses this escape route. The poles dictate the intrinsic rate of entropy generation due to unmeasurable disturbances, thermal fluctuations, and quantum-level noise. Predictive AI can handle the known, deterministic dynamics, but it cannot compress the fundamental, unpredictable uncertainty generated by the unstable poles.[1][2]

This intersection of disciplines is becoming critically relevant today. Drone swarms, remote robotic surgery, and autonomous vehicle platoons all rely on wireless networks to close their feedback loops. These are highly unstable physical systems relying on digital communication channels that are subject to interference, packet loss, and latency.[3]

When a 5G network experiences a momentary latency spike, it effectively drops the available data rate. If that rate dips below the sum of the base-2 logarithms of the physical system's poles, the system mathematically must diverge. The physical world, through its inherent instability, dictates the required bandwidth of the digital one.[3]