The Traveling Salesman Problem & Christofides Algorithm

---
count: false
class: nord-dark, middle, center

# 🧳 The Traveling Salesman Problem & Christofides Algorithm

### A 3/2-Approximation for Metric TSP

@luk036 👨‍💻 · 2026 📅

---

### 🎯 Traveling Salesman Problem (TSP)

> *"Given a list of cities and the distances between each pair, what is the shortest possible route that visits each city exactly once and returns to the origin city?"*

.font-sm.mb-xs[
```python
# We have N cities with pairwise distances
num_nodes = 20
G = nx.complete_graph(num_nodes)         # fully connected
pos = {i: (random.uniform(0,100), ...)}  # random 2D coordinates
```
]

.mermaid[
<pre>
graph LR
    A((A)) -- 10 --- B((B))
    A -- 15 --- C((C))
    A -- 12 --- D((D))
    B -- 8 --- C((C))
    B -- 9 --- D((D))
    C -- 11 --- D((D))
    style A fill:#e0f2f1,stroke:#00695c,stroke-width:3px
    style B fill:#e0f2f1,stroke:#00695c,stroke-width:3px
    style C fill:#e0f2f1,stroke:#00695c,stroke-width:3px
    style D fill:#e0f2f1,stroke:#00695c,stroke-width:3px
</pre>
]

- **Input**: Complete graph ${\color{salmon}K}_n$ with edge weights ${\color{coral}w}(i, j)$
- **Output**: Hamiltonian cycle of minimum total weight
- **Complexity**: **NP-hard** 🧠 — no polynomial exact algorithm (unless P = NP)

---

### 📐 Why Metric TSP?

#### The Triangle Inequality ✨

Our demos support **two common metrics**:

| Metric | Formula | netlistx.tsp factory |
|:-------|:--------|:-----------------------|
| **L2 (Euclidean)** 🟦 | ${\color{coral}w}(i, j) = \sqrt{\Delta x^2 + \Delta y^2}$ | `make_l2_graph(n, seed)` |
| **L1 (Manhattan)** 🟨 | ${\color{coral}w}(i, j) = \lvert \Delta x \rvert + \lvert \Delta y \rvert$ | `make_l1_graph(n, seed)` |

```python
from netlistx.tsp import make_l2_graph, make_l1_graph

G_l2, pos = make_l2_graph(20, seed=42)  # Euclidean
G_l1, pos = make_l1_graph(20, seed=42)  # Manhattan
```

---

Both satisfy the **triangle inequality**, so the 3/2 guarantee holds for either! ✨

A **metric** satisfies:

1. 🌐 ${\color{coral}w}(i, j) = {\color{coral}w}(j, i)$ (symmetry)
2. 📏 ${\color{coral}w}(i, k) \le {\color{coral}w}(i, j) + {\color{coral}w}(j, k)$ (triangle inequality)
3. ✅ ${\color{coral}w}(i, i) = 0$, ${\color{coral}w}(i, j) > 0$ for $i \ne j$

.mermaid[
<pre>
graph LR
    i((i)) -- w_ij --- j((j))
    i -- w_ik --- k((k))
    j -- w_jk --- k
    style i fill:#ffe0b2,stroke:#e65100,stroke-width:3px
    style j fill:#e0f7fa,stroke:#00838f,stroke-width:3px
    style k fill:#c8e6c9,stroke:#2e7d32,stroke-width:3px
</pre>
]

> **Why it matters**: Metric TSP is still NP-hard, but we can **approximate** it efficiently with a guaranteed bound! 🎉

---

### 🏆 Christofides Algorithm — Overview

#### Nicos Christofides (1976) — A Beautiful Marriage of Graph Theory 💒

.font-sm.mb-xs[
| Step | What | Why |
|------|------|-----|
| **1** 🌲 | Minimum Spanning Tree (MST) | Cheap backbone connecting all nodes |
| **2** 🔢 | Find odd-degree nodes in MST | Eulerian circuits need all-even degrees |
| **3** 💍 | Min-weight perfect matching on odds | Pair them up cheaply |
| **4** 🔗 | Superimpose MST + Matching | Now all degrees are even! |
| **5** 🔄 | Eulerian circuit | Traverse every edge exactly once |
| **6** ✂️ | Shortcut to Hamiltonian cycle | Remove duplicates — triangle inequality saves us |

]

---

.pull-left[
.mermaid[
<pre>
flowchart TD
    A["Complete Graph G"] --> B["🌲 MST"]
    B --> C["🔢 Odd-degree nodes O"]
    C --> D["💍 Min Perfect Matching M on O"]
    B --> E["🔗 Combine MST ∪ M → Multigraph"]
    D --> E
    E --> O((o))
    style A fill:#fff9c4,stroke:#f57f17,stroke-width:3px
</pre>
]
]
.pull-right[
.mermaid[
<pre>
flowchart TD
    O((o)) --> F["🔄 Eulerian Circuit"]
    F --> G["✂️ Shortcut → Hamiltonian Cycle"]
    G --> H["✅ TSP Tour (≤ 1.5× OPT)"]
    style H fill:#c8e6c9,stroke:#2e7d32,stroke-width:3px
</pre>
]
]
---

### 🌲 Step 1 — Minimum Spanning Tree

Build a tree connecting **all** nodes with minimum total weight.

```python
mst = nx.minimum_spanning_tree(G, weight="weight")
# Kruskal's or Prim's — O(m log n)
```

.mermaid[
<pre>
graph TB
    subgraph "MST (n=6)"
        0((0)) --- 1((1))
        0 --- 2((2))
        1 --- 3((3))
        1 --- 4((4))
        4 --- 5((5))
    end
    style 0 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 1 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 2 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 3 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 4 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 5 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
</pre>
]

---

#### Key Property 📌

$${\color{coral}w}(\text{MST}) \le {\color{coral}w}(\text{OPT}_{\text{TSP}})$$

Why? Removing one edge from the optimal TSP tour gives a **spanning tree**, and MST is the **minimum** spanning tree. So the MST is a **lower bound** on the optimal TSP solution! 💡

---

### 🔢 Step 2 — Odd-Degree Nodes

Every tree has an **even** number of odd-degree nodes. 🤯

```python
odd_degree_nodes = [v for v, d in mst.degree() if d % 2 != 0]
# This list is ALWAYS even-length!
```

.font-sm.mb-xs[
| Node | Degree | Odd? |
|:----:|:------:|:----:|
| 0 | 2 | ❌ |
| 1 | 3 | ✅ |
| 2 | 1 | ✅ |
| 3 | 1 | ✅ |
| 4 | 3 | ✅ |
| 5 | 0 (leaf) | — |

]
]
.pull-right[
.mermaid[
<pre>
graph TB
    0((0)) ---|deg=2| 1((1))
    0 ---|deg=2| 2((2))
    1 ---|deg=3| 3((3))
    1 ---|deg=3| 4((4))
    4 ---|deg=3| 5((5))
    style 1 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
    style 2 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
    style 3 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
    style 5 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
</pre>
]
]

---

**Why?** $$\sum_{v \in {\color{salmon}V}} \deg(v) = 2|{\color{lime}E}|$$ is even, and the number of odd-degree vertices must be even (handshaking lemma 🖐️).

---

### 💍 Step 3 — Minimum Weight Perfect Matching

Pair up the odd-degree nodes as **cheaply as possible**.

```python
odd_subgraph = G.subgraph(odd_degree_nodes)
matching = nx.min_weight_matching(odd_subgraph, weight="weight")
# Blossom algorithm — O(n³) 💐
```

.mermaid[
<pre>
graph TB
    subgraph "MST edges (solid)"
        0((0)) --- 1((1))
        0 --- 2((2))
        1 --- 3((3))
        1 --- 4((4))
        4 --- 5((5))
    end
    subgraph "Matching edges (dashed)"
        2 -.- 3
        4 -.- 5
    end
    style 0 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 1 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
    style 2 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
    style 3 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
    style 4 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 5 fill:#ffcdd2,stroke:#c62828,stroke-width:3px
</pre>
]

---

#### Why this works 🔑

$${\color{coral}w}(\text{matching}) \le \frac{1}{2} {\color{coral}w}(\text{OPT}_{\text{TSP}})$$

Taking the optimal TSP tour, alternate edges form 2 perfect matchings on the odd set — one of them ≤ half the tour.

---

### 🔗 Step 4 — Combine: MST ∪ Matching

.font-sm.mb-xs[
```python
multigraph = nx.MultiGraph(mst)                        # start with MST
multigraph.add_edges_from((u, v, G[u][v]) for u, v in matching)  # add matching
```
]

Now **every node has even degree**! ✅

.font-sm.mb-xs[
```
deg(v) before (MST)       →  odd for some
deg(v) after adding matching →  EVEN for EVERY node
```
]

.mermaid[
<pre>
graph TB
    subgraph "Combined Multigraph"
        0((0)) --- 1((1))
        0 --- 2((2))
        1 === 3((3))
        1 --- 4((4))
        2 === 3((3))
        4 === 5((5))
        4 --- 5((5))
    end
    style 0 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 1 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 2 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 3 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 4 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 5 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
</pre>
]

> **All degrees even** ⟹ **Eulerian multigraph** exists! 🚀

---

### 🔄 Step 5 — Eulerian Circuit

Traverse **every edge exactly once** and return to start. Hierholzer's algorithm — $O(|{\color{lime}E}|)$.

```python
eulerian_circuit = list(nx.eulerian_circuit(multigraph))
# e.g., [0→1, 1→3, 3→2, 2→0, 0→...]
```

.mermaid[
<pre>
graph LR
    subgraph "Eulerian Walk"
        direction LR
        0 --> 1 --> 3 --> 2 --> 0 --> 2
    end
    style 0 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 1 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 2 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 3 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
</pre>
]

---

#### Visual Example 👁️

.mermaid[
<pre>
graph TB
    0((①)) === 1((②))
    0 === 2((③))
    0 -.-x|visited twice| 2
    1 === 3((④))
    1 === 4((⑤))
    2 === 3
    4 === 5((⑥))
    4 -.-x 5
    linkStyle 0 stroke:#4c4,stroke-width:2px
    linkStyle 1 stroke:#4c4,stroke-width:2px
    linkStyle 2 stroke:#4c4,stroke-width:2px
    linkStyle 3 stroke:#4c4,stroke-width:2px
    linkStyle 4 stroke:#4c4,stroke-width:2px
    linkStyle 5 stroke:#4c4,stroke-width:2px
    style 0 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 1 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 2 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 3 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 4 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
    style 5 fill:#fff9c4,stroke:#f57f17,stroke-width:3px
</pre>
]

Path: $0 \to 1 \to 3 \to 2 \to 0 \to 2 \to \dots$

> **Problem** 😕: Nodes repeat! (Eulerian ≠ Hamiltonian)

---

### ✂️ Step 6 — Shortcut to Hamiltonian Cycle

**Triangle inequality** to the rescue! 🦸: when we encounter an already-visited node, **skip it**.

```python
path = []
visited = set()
for u, v in eulerian_circuit:
    if u not in visited:
        path.append(u)        # first time → keep
        visited.add(u)
path.append(path[0])  # close the loop
```

.mermaid[
<pre>
flowchart LR
    subgraph "Eulerian: 0→①→③→②→①→④→⑤→④→①→0"
        A
    end
    subgraph "Shortcut: 0→①→③→②→④→⑤→0"
        B
    end
    A -- "✂️ remove duplicates" --> B
    style A fill:#e3f2fd,stroke:#1565c0,stroke-width:3px
    style B fill:#ffccbc,stroke:#bf360c,stroke-width:3px
</pre>
]

---

#### Why it's valid ✅

Triangle inequality guarantees:

$${\color{coral}w}(\text{shortcut edge}) \le {\color{coral}w}(\text{skipped path})$$

So skipping never **increases** total cost! The resulting tour is **no longer** than the Eulerian circuit. 🎯

---

### 📊 Putting It All Together — The Bound

#### The 3/2 Approximation Guarantee 🏅

$${\color{coral}w}(\text{Tour}) = {\color{coral}w}(\text{MST} \cup \text{Matching}) \le {\color{coral}w}(\text{OPT}) + \frac{1}{2}{\color{coral}w}(\text{OPT}) = \frac{3}{2} {\color{coral}w}(\text{OPT})$$

> **Best possible polynomial-time approximation** for Metric TSP (unless P = NP)! 🏆

.pull-left[
.mermaid[
<pre>
xychart-beta
    title "Cost buildup vs OPT"
    x-axis ["MST", "Matching", "Total Tour", "OPT"]
    y-axis "Cost" 0 --> 100
    bar [40, 20, 60, 45]
</pre>
]
]
.pull-right[
.font-sm.mb-xs[
| Component | Bound | Ratio |
|:----------|:------|:-----:|
| 🌲 MST | $\le \text{OPT}$ | 1.0× |
| 💍 Matching | $\le \frac{1}{2}\text{OPT}$ | 0.5× |
| **🔗 Total** | $\le \frac{3}{2}\text{OPT}$ | **1.5×** |
| 🎯 OPT | — | 1.0× |

]
]

---

### 🧪 Demo Walkthrough — L2 vs L1 (n = 20)

Both metrics share the **same code path** — only the distance formula differs:

```python
from netlistx.tsp import make_l2_graph, make_l1_graph, \
        christofides_tsp, solve_christofides_2opt_tsp

G_l2, pos = make_l2_graph(20, seed=42)   # Euclidean
G_l1, pos = make_l1_graph(20, seed=42)   # Manhattan

refined_l2 = solve_christofides_2opt_tsp(G_l2)
refined_l1 = solve_christofides_2opt_tsp(G_l1)
```

#### Results Comparison 📊

.font-sm.mb-xs[
```
L2 (Euclidean) — Initial: 407.12  →  Refined: 389.23  (Δ = 17.89)
L1 (Manhattan) — Initial: 459.99  →  Refined: 444.69  (Δ = 15.30)
```
]

.font-sm.mb-xs[
| Metric | Christofides | +2-Opt | Improvement |
|:-------|:------------:|:------:|:-----------:|
| 🟦 L2 (Euclidean) | 407.12 | 389.23 | 17.89 |
| 🟨 L1 (Manhattan) | 459.99 | 444.69 | 15.30 |

]

---

#### Visual Comparison 🖼️

> 🟦 L2 distances are **always ≤ L1** for the same points (Pythagoras), so absolute values differ — but **both improve** with 2-Opt! 🎯

]
> 🔴 Red = Christofides  ·  🔵 Blue = after 2-Opt

---

### 🎨 Larger Instances — L2 vs L1 (n = 100)

Scaling the same comparison to 100 cities:

]

---

#### Results 📊

.font-sm.mb-xs[
| Metric | Christofides | +2-Opt | Improvement |
|:-------|:------------:|:------:|:-----------:|
| 🟦 L2 (Euclidean) | 875.50 | 809.07 | **66.43 (7.6%)** |
| 🟨 L1 (Manhattan) | 1112.35 | 1044.58 | **67.77 (6.1%)** |

]
#### Observations 👁️
- 🔵 Blue tours are smooth with **no crossing edges** — 2-Opt guarantees this
- Both metrics produce valid, near-optimal TSP tours
- The Christofides + 2-Opt combination is **metric-agnostic** — it works for any distance satisfying triangle inequality

---

### ⏱ Complexity Analysis

| Step | Algorithm | Complexity |
|:-----|:----------|:----------:|
| 🌲 MST | Kruskal / Prim | $O(m \log n)$ |
| 🔢 Odd-degree scan | Degree check | $O(n)$ |
| 💍 Min-weight perfect matching | Blossom algorithm 🌸 | $O(n^3)$ |
| 🔗 Combine | Edge addition | $O(n)$ |
| 🔄 Eulerian circuit | Hierholzer | $O(m)$ |
| ✂️ Shortcut | Linear scan | $O(n)$ |
| **Total** | **Christofides** | **$O(n^3)$** 🐌 |

---

#### Scalability 📈

| $n$ (cities) | Rough time |
|:------------:|:----------:|
| 20 | ~ms ⚡ |
| 100 | ~seconds |
| 1,000 | ~minutes |
| 10,000 | ~hours (matching dominates) |

> **Bottleneck**: Minimum weight perfect matching (Blossom). For huge $n$, use heuristics instead.

---

### 🧮 Why Can't We Do Better?

.mermaid[
<pre>
xychart-beta
    title "Best-known approximation ratios for Metric TSP"
    x-axis ["1976\n Christofides", "2020\n Karlin-Klein-OveisGharan", "2022\n Improved", "Ultimate Limit\n (unless P=NP)"]
    y-axis "Ratio" 1.0 --> 2.0
    line [1.5, 1.5, 1.5, 1.0]
</pre>
]

---

**Wait, is Christofides still state-of-the-art?** 🤔

| Year | Result | Ratio |
|:----:|:-------|:-----:|
| 1976 🏛 | **Christofides** | **1.5** |
| 2020 | Karlin, Klein, Oveis Gharan | $1.5 - \epsilon$ |
| 2022 | Improved analysis | $1.5 - 2\cdot 10^{-6}$ |
| ? | Inapproximability | $> 1.0045$ (under ETH) |

> Despite tiny improvements, **Christofides remains the conceptual foundation** after 50 years! 🎉

---

### 🧪 Variations & Extensions

#### Beyond Metric TSP 🔭

.font-sm.mb-xs[
| Variant | Constraint | Christofides work? |
|:--------|:-----------|:------------------:|
| 🌐 **Metric TSP** | Triangle inequality | ✅ 1.5-approx |
| 🛩 **Asymmetric TSP** | ${\color{coral}w}(i, j) \ne {\color{coral}w}(j, i)$ | ❌ |
| 🗺 **Graphical TSP** | Distances = shortest path | ✅ |
| 📦 **TSP with neighborhoods** | Visit regions, not points | ❌ |
| 🚛 **Vehicle Routing (VRP)** | Multiple vehicles | ❌ (different problem) |
| ⏰ **TSP with time windows** | Deadline constraints | ❌ |

]

---

### 📜 History & Legacy

#### Timeline of TSP 🕰️

.mermaid[
<pre>
timeline
    title A Brief History of TSP
    1800s : "Traveling Salesman" mentioned in German traveling salesmen manuals
    1930 : Karl Menger formalizes the problem
    1972 : Karp proves TSP is NP-complete 🔨
    1976 : Christofides publishes 1.5-approximation 🏆
    1990s : Concorde solver cracks 1000+ city instances
    2004 : 85,900-city instance solved optimally
    2020s : First improvement over Christofides (by 0.0002%)
</pre>
]

#### Why It Matters 🌍

- 🚚 **Logistics** — delivery route optimization
- 🧬 **Genome sequencing** — DNA fragment assembly
- 💻 **VLSI design** — drilling, circuit board routing
- 🏭 **Manufacturing** — CNC machine path planning
- 🛰 **Satellite** — telescope scheduling

---

### 🔬 Hands-on: Try It Yourself

#### Run the Demos 🖥

.font-sm.mb-xs[
```bash
# Euclidean (L2) demos
python experiments/tsp_demo.py         # n=20
python experiments/tsp2opt_demo.py     # n=100

# Manhattan (L1) demos
python experiments/tsp_l1_demo.py      # n=20
python experiments/tsp_l1_n100_demo.py # n=100
```
]

#### API Quick Reference 📖

.font-sm.mb-xs[
```python
from netlistx.tsp import (
    make_l2_graph,          # Factory: Euclidean metric
    make_l1_graph,          # Factory: Manhattan metric
    christofides_tsp,       # 3/2-approximation
    solve_christofides_2opt_tsp,  # Christofides + 2-Opt
    calculate_total_distance,      # Evaluate a tour
)
```
]

---

#### Experiment Ideas 🔧

1. **Vary $n$**: try 10, 50, 100, 500 cities — how does runtime scale?
2. **Compare L2 vs L1**: run both on the same random seed — how do tours differ?
3. **Compare to Nearest Neighbor** heuristic (greedy):
   $${\color{coral}w}(\text{NN}) / {\color{coral}w}(\text{OPT}) \approx O(\log n)$$
4. **Break the triangle inequality** ➡ observe ratio degrade
5. **Visualize each step** — add `plt.pause()` between steps for an animation!

#### Starter Code 🚀

.font-sm.mb-xs[
```python
def nearest_neighbor_tsp(G, start=0):
    """Greedy TSP — O(n²), no guarantee"""
    path, visited = [start], {start}
    curr = start
    while len(path) < len(G):
        nxt = min((n for n in G[curr] if n not in visited),
                  key=lambda x: G[curr][x]['weight'])
        path.append(nxt); visited.add(nxt); curr = nxt
    path.append(start)
    return path
```
]

---

### 🎓 What You Learned Today

| Concept | Key Insight |
|:--------|:------------|
| 🌲 **MST as lower bound** | Any spanning tree ≤ optimal tour |
| 🔢 **Odd-degree parity** | Handshaking lemma → even count |
| 💍 **Perfect matching** | Pairs odds at ≤ half OPT cost |
| 🔗 **Union preserves evenness** | Eulerian condition met |
| 🔄 **Eulerian circuit** | Traverse all edges exactly once |
| ✂️ **Shortcutting** | Triangle inequality protects cost |
| 📐 **3/2 bound** | Best possible poly-time guarantee |

---

#### The Magic Formula 🪄

$$\boxed{{\color{coral}w}(\text{Christofides}) \le \frac{3}{2} {\color{coral}w}(\text{OPT})}$$

#### One Sentence Summary 📝

> *Build a cheap backbone (MST), fix the parity (matching), take an Eulerian walk, and shortcut — triangle inequality ensures you never pay more than 50% above optimal.*

---

### 🧩 Proof Sketch (For the Curious)

#### Theorem 📐

> Christofides algorithm yields a TSP tour of weight at most $\frac{3}{2} \cdot \text{OPT}$.

**Proof** 🧾:

1. **MST bound**: ${\color{coral}w}(\text{MST}) \le \text{OPT}$ (remove edge from OPT → spanning tree)
2. **Matching bound**: Let ${\color{salmon}O}$ be odd-degree vertices. OPT tour induces 2 perfect matchings on ${\color{salmon}O}$ by taking alternating edges. The cheaper has weight $\le \frac{1}{2} {\color{coral}w}(\text{OPT})$, so:
   ${\color{coral}w}(\text{matching}) \le \frac{1}{2} \text{OPT}$
3. **Combined**: ${\color{coral}w}(\text{MST} \cup \text{matching}) \le \text{OPT} + \frac{1}{2}\text{OPT} = \frac{3}{2}\text{OPT}$
4. **Eulerian + Shortcut**: Triangle inequality ensures shortcutting doesn't increase cost. $\therefore {\color{coral}w}(\text{final tour}) \le \frac{3}{2} \text{OPT} \quad \blacksquare$

---

.mermaid[
<pre>
graph LR
    subgraph "OPT Tour"
        direction LR
        a1((•)) --- a2((•)) --- a3((•)) --- a4((•)) --- a1
    end
    subgraph "Matching 1 (dashed)"
        b1((•)) -.- b2((•))
        b3((•)) -.- b4((•))
    end
    subgraph "Matching 2 (dotted)"
        c1((•)) -.- c2((•))
        c3((•)) -.- c4((•))
    end
    style a1 fill:#fce4ec,stroke:#880e4f,stroke-width:3px
    style a2 fill:#fce4ec,stroke:#880e4f,stroke-width:3px
    style b1 fill:#fce4ec,stroke:#880e4f,stroke-width:3px
    style b2 fill:#fce4ec,stroke:#880e4f,stroke-width:3px
    style b3 fill:#fce4ec,stroke:#880e4f,stroke-width:3px
    style b4 fill:#fce4ec,stroke:#880e4f,stroke-width:3px
</pre>
]

---

### References 📖

- **Christofides, N. (1976)**. *Worst-case analysis of a new heuristic for the travelling salesman problem*. Carnegie Mellon University. [Report 388]
- **Karlin, A.R., Klein, N., Oveis Gharan, S. (2020)**. *A (Slightly) Improved Approximation Algorithm for Metric TSP*. STOC 2021. [doi:10.1145/3406325.3451009]
- **NetworkX Documentation**: [https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.approximation.traveling_salesman.christofides.html](https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.approximation.traveling_salesman.christofides.html)

---

### Code & Project 🗂

.font-sm.mb-xs[
| Resource | Location |
|:---------|:---------|
| 🔬 Demo code | `experiments/tsp_demo.py` (L2, n=20) |
| 🔬 Demo code | `experiments/tsp2opt_demo.py` (L2, n=100) |
| 🔬 Demo code | `experiments/tsp_l1_demo.py` (L1, n=20) |
| 🔬 Demo code | `experiments/tsp_l1_n100_demo.py` (L1, n=100) |
| 🖼 Figure (L2 n=20) | `experiments/tsp_demo_n20.svg` |
| 🖼 Figure (L2 n=100) | `experiments/tsp2opt_demo_n100.svg` |
| 🖼 Figure (L1 n=20) | `experiments/tsp_l1_demo_n20.svg` |
| 🖼 Figure (L1 n=100) | `experiments/tsp_l1_n100_demo.svg` |
| 📦 Python lib | `netlistx.tsp` — `make_l2_graph()`, `make_l1_graph()` |
| ⚡ C++ lib | `netlistx-cpp` — `EuclideanWeight`, `ManhattanWeight` |
| 📊 Slides | `experiments/slides_christofides_tsp.md` |

]

---

## Questions? 🤔

---

> *"The Traveling Salesman Problem: where computer science, mathematics, and logistics meet — and where approximation algorithms shine."* ✨