---
title: Topology Efficiency Index (TEI)
url: https://www.emergentmind.com/topics/topology-efficiency-index-tei
type: topic
---

# Topology Efficiency Index (TEI)

The Topology Efficiency Index (TEI) is a rigorously defined metric used to evaluate the efficiency of topological structures in networks and multi-agent systems. TEI quantifies the trade-off between a system’s performance and the resource cost (either in communication acts or hardware complexity) imposed by a specific connectivity topology. It has been introduced and analyzed in resource-centric network evaluation [2601.19008] as well as in the context of communication-efficient multi-agent reinforcement learning [2511.09171], where its interpretation and mathematical formulation vary to match domain requirements.

## 1. Mathematical Formulation

In multi-agent communication settings, TEI is defined as the ratio of the empirical success rate to the total number of communication acts over an epoch:

\[
\Phi_{\mathrm{TEI}_t} = \frac{\mathscr{S}_t}{C_t}
\]

where:
- $\mathscr{S}_t \in [0,1]$ is the task success rate (fraction of successfully completed episodes),
- $C_t$ is the total number of directed communication acts, computed as the sum of all non-self adjacency matrix entries across $L$ communication rounds and $N$ agents:

\[
C_t = \sum_{l=1}^{L} \;\sum_{i=0}^{N-1} \sum_{\substack{j=0\\j\neq i}}^{N-1} G_{i,j}^{\,t(l)}
\]

In hardware network design, TEI quantifies the normalized cost required to maintain a non-blocking, full-throughput topology:

\[
\mathrm{TEI} = \frac{H}{L_{\mathrm{host}}} \cdot [k + \gamma k^2] \cdot \frac{N}{M}
\]

where:
- $H$ is the mean hop count per packet,
- $L_{\mathrm{host}}$ is the connectivity to hosts,
- $k$ is the router radix,
- $\gamma = \beta/\alpha$ is the ratio of crossbar to interface cost,
- $N/M$ is the router-to-host ratio [2601.19008].

Both contexts share the principle that TEI is minimized when high system-level performance is achieved with minimal connectivity cost—either in messages or hardware.

## 2. Theoretical Properties

The properties of TEI are dictated by its mathematical construction:

- **Domain**: In multi-agent systems, $\mathscr{S}_t \in [0,1]$, $C_t \in \{1,2,\dots,N(N-1)L\}$, giving $\Phi_{\mathrm{TEI}_t} \in [0,1]$ or higher if $C_t$ is small and $\mathscr{S}_t$ nonzero.
- **Extremal Behavior**: $\Phi_{\mathrm{TEI}_t}=0$ for zero success rate, while very large TEI values appear if $\mathscr{S}_t>0$ with extremely low communication.
- **Monotonicity**: For fixed $\mathscr{S}_t$, increasing $C_t$ decreases TEI, while for fixed $C_t$, increasing $\mathscr{S}_t$ increases TEI [2511.09171].
- **Hardware context**: TEI increases with hop count $H$ or required router complexity $k$, and is sensitive to the technological ratio $\gamma$ between link and router costs [2601.19008].

## 3. TEI in Multi-Agent Communication Protocols

TEI serves as an explicit post-hoc communication-efficiency metric to evaluate and compare protocols in multi-agent reinforcement learning. Given communication adjacency matrices (indicating message passing) and binary success indicators per episode, TEI can be computed as follows:

```python
def compute_TEI(episodes, L):
    total_success = 0
    total_comms = 0
    E = len(episodes)
    for e in episodes:
        total_success += e.success
        for l in range(L):
            G = e.adjacency_matrices[l]
            for i in range(N):
                for j in range(N):
                    if j != i:
                        total_comms += G[i][j]
    S_t = total_success / E
    C_t = total_comms
    return S_t / C_t if C_t != 0 else 0
```

In empirical studies on the Traffic Junction environment, TEI values allow performance-to-cost comparison across MAGIC, CommNet, TarMAC, GA-Comm, and IC3Net algorithms, with typical TEI in the $10^{-7}$ range [2511.09171]. Increasing communication rounds generally decreases TEI, while efficiency-augmented loss functions significantly improve it.

## 4. TEI in Network Hardware Design

In network architecture, TEI is characterized as the normalized hardware cost per unit host-throughput required for non-blocking operation. The explicit formula,

\[
\mathrm{TEI} = (H/L_\mathrm{host}) \cdot [k + \gamma k^2] \cdot (N/M)
\]

endows TEI with the following comparative properties across common topologies:

| Topology           | Router Radix $k$         | TEI Expression (normalized by $\alpha$)                  |
|--------------------|--------------------------|----------------------------------------------------------|
| 2D Torus           | $5$                      | $(\sqrt{N}/4)[5+\gamma 25]$                              |
| Hypercube          | $\log_2 N + 1$           | $(\log_2 N + 1) + \gamma (\log_2 N+1)^2$                 |
| Flat Butterfly     | $2\sqrt{N}-1$            | $(2\sqrt{N}-1) + \gamma (2\sqrt{N}-1)^2$                 |
| Fat Tree           | const $k$                | $2(1+\gamma k)\cdot[\frac{\ell-1/2}{k}]$, $\ell \sim \log_2 M$ |
| n-plane Star       | $M$                      | $[1+\gamma M]/n$                                         |

For large scales, indirect topologies (Fat Trees) minimize TEI by capping router radix at the technology-optimal value, while direct networks become inefficient as $\beta k^2$ dominates cost [2601.19008].

## 5. Applications and Comparative Insights

In multi-agent learning, TEI is not used as a training objective but rather as a principal metric for post-hoc comparative evaluation. When protocol topologies are fixed (e.g., fully connected), $C_t$ is nearly constant, so TEI is analytic and not amenable as a learning signal. Higher TEI values signal protocols that are more parsimonious with bandwidth and, therefore, better suited for constrained environments.

In network hardware, TEI facilitates selection among candidate topologies by expressing the absolute resource efficiency under strict performance guarantees. Topologies optimizing TEI under cost model parameters ($\alpha$, $\beta$) and problem scale ($M$, $N$) are identified via explicit formulas, enabling engineering design with clear efficiency trade-offs.

## 6. Practical Design Guidance

From the resource-centric analysis, clear regimes emerge:
- For small networks ($N \lesssim 16$), low-radix direct topologies or parallel star planes may win in absolute cost.
- For medium networks ($16 \lesssim N \lesssim 10^3$), hypercube-like logarithmic-diameter networks typically yield minimal TEI.
- For very large networks, indirect Fat Tree topologies with fixed, technology-optimal router radix minimize TEI [2601.19008].

A key design rule is that redundancy is most efficiently implemented by parallel network planes (multi-plane stars) rather than by intrinsic path diversity within a single graph.

## 7. Significance and Limitations

TEI offers a compact, dimensionless score enabling incisive comparisons of efficiency across protocols, algorithms, or hardware designs. Its interpretability at the extremes, analytic tractability, and empirical correlation with deployment cost (whether message or hardware) have made it a central metric in both communication protocol design [2511.09171] and network architecture analysis [2601.19008]. However, TEI does not capture emergent behavioral properties beyond its numerator and denominator and assumes homogeneity in communication or traffic; deployment contexts with highly variable loads or heterogeneous success definitions may require complementary analysis.

A plausible implication is that the widespread adoption of TEI (or resource-centric analogs) could standardize efficiency benchmarking across domains involving structured connectivity and resource-constrained coordination.

Source: https://www.emergentmind.com/topics/topology-efficiency-index-tei