Visualizing an Assassin Puzzle

Over at Math3ma, Tai-Danae Bradley shared the following puzzle, which she also featured in a fantastic (spoiler-free) YouTube video. If you’re seeing this for the first time, watch the video first.

Consider a square in the xy-plane, and let A (an “assassin”) and T (a “target”) be two arbitrary-but-fixed points within the square. Suppose that the square behaves like a billiard table, so that any ray (a.k.a “shot”) from the assassin will bounce off the sides of the square, with the angle of incidence equaling the angle of reflection.

Puzzle: Is it possible to block any possible shot from A to T by placing a finite number of points in the square?

This puzzle found its way to me through Tai-Danae’s video, via category theorist Emily Riehl, via a talk by the recently deceased Fields Medalist Maryam Mirzakhani, who studied the problem in more generality. I’m not familiar with her work, but knowing mathematicians it’s probably set in an arbitrary complex $ n$-manifold.

See Tai-Danae’s post for a proof, which left such an impression on me I had to dig deeper. In this post I’ll discuss a visualization I made—now posted at the end of Tai-Danae’s article—as well as here and below (to avoid spoilers). In the visualization, mouse movement chooses the firing direction for the assassin, and the target is in green. Dragging the target with the mouse updates the position of the guards. The source code is on Github.

Outline

The visualization uses d3 library, which was made for visualizations that dynamically update with data. I use it because it can draw SVGs real nice.

The meat of the visualization is in two geometric functions.

  1. Decompose a ray into a series of line segments—its path as it bounces off the walls—stopping if it intersects any of the points in the plane.
  2. Compute the optimal position of the guards, given the boundary square and the positions of the assassin and target.

Both of these functions, along with all the geometry that supports them, is in geometry.js. The rest of the demo is defined in main.js, in which I oafishly trample over d3 best practices to arrive miraculously at a working product. Critiques welcome 🙂

As with most programming and software problems, the key to implementing these functions while maintaining your sanity is breaking it down into manageable pieces. Incrementalism is your friend.

Vectors, rays, rectangles, and ray splitting

We start at the bottom with a Vector class with helpful methods for adding, scaling, and computing norms and inner products.

function innerProduct(a, b) {
  return a.x * b.x + a.y * b.y;
}

class Vector {
  constructor(x, y) {
    this.x = x;
    this.y = y;
  }

  normalized() { ... }
  norm() { ... }
  add(vector) { ... }
  subtract(vector) { ... }
  scale(length) { ... }
  distance(vector) { ... }
  midpoint(b) { ... }
}

This allows one to compute the distance between two points, e.g., with vector.subtract(otherVector).norm().

Next we define a class for a ray, which is represented by its center (a vector) and a direction (a vector).

class Ray {
  constructor(center, direction, length=100000) {
    this.center = center;
    this.length = length;

    if (direction.x == 0 && direction.y == 0) {
      throw "Can't have zero direction";
    }
    this.direction = direction.normalized();
  }

  endpoint() {
    return this.center.add(this.direction.scale(this.length));
  }

  intersects(point) {
    let shiftedPoint = point.subtract(this.center);
    let signedLength = innerProduct(shiftedPoint, this.direction);
    let projectedVector = this.direction.scale(signedLength);
    let differenceVector = shiftedPoint.subtract(projectedVector);

    if (signedLength > 0
        && this.length > signedLength
        && differenceVector.norm() < intersectionRadius) {
      return projectedVector.add(this.center);
    } else {
      return null;
    }
  }
}

The ray must be finite for us to draw it, but the length we've chosen is so large that, as you can see in the visualization, it's effectively infinite. Feel free to scale it up even longer.

assassin-puzzle

The interesting bit is the intersection function. We want to compute whether a ray intersects a point. To do this, we use the inner product as a decision rule to compute the distance of a point from a line. If that distance is very small, we say they intersect.

In our demo points are not infinitesimal, but rather have a small radius described by intersectionRadius. For the sake of being able to see anything we set this to 3 pixels. If it’s too small the demo will look bad. The ray won’t stop when it should appear to stop, and it can appear to hit the target when it doesn’t.

Next up we have a class for a Rectangle, which is where the magic happens. The boilerplate and helper methods:

class Rectangle {
  constructor(bottomLeft, topRight) {
    this.bottomLeft = bottomLeft;
    this.topRight = topRight;
  }

  topLeft() { ... }
  center() { ... }
  width() { .. }
  height() { ... }
  contains(vector) { ... }

The function rayToPoints that splits a ray into line segments from bouncing depends on three helper functions:

  1. rayIntersection: Compute the intersection point of a ray with the rectangle.
  2. isOnVerticalWall: Determine if a point is on a vertical or horizontal wall of the rectangle, raising an error if neither.
  3. splitRay: Split a ray into a line segment and a shorter ray that’s “bounced” off the wall of the rectangle.

(2) is trivial, computing some x- and y-coordinate distances up to some error tolerance. (1) involves parameterizing the ray and checking one of four inequalities. If the bottom left of the rectangle is $ (x_1, y_1)$ and the top right is $ (x_2, y_2)$ and the ray is written as $ \{ (c_1 + t v_1, c_2 + t v_2) \mid t > 0 \}$, then—with some elbow grease—the following four equations provide all possibilities, with some special cases for vertical or horizontal rays:

$ \displaystyle \begin{aligned} c_2 + t v_2 &= y_2 & \textup{ and } \hspace{2mm} & x_1 \leq c_1 + t v_1 \leq x_2 & \textup{ (intersects top)} \\ c_2 + t v_2 &= y_1 & \textup{ and } \hspace{2mm} & x_1 \leq c_1 + t v_1 \leq x_2 & \textup{ (intersects bottom)} \\ c_1 + t v_1 &= x_1 & \textup{ and } \hspace{2mm} & y_1 \leq c_2 + t v_2 \leq y_2 & \textup{ (intersects left)} \\ c_1 + t v_1 &= x_2 & \textup{ and } \hspace{2mm} & y_1 \leq c_2 + t v_2 \leq y_2 & \textup{ (intersects right)} \\ \end{aligned}$

In code:

  rayIntersection(ray) {
    let c1 = ray.center.x;
    let c2 = ray.center.y;
    let v1 = ray.direction.x;
    let v2 = ray.direction.y;
    let x1 = this.bottomLeft.x;
    let y1 = this.bottomLeft.y;
    let x2 = this.topRight.x;
    let y2 = this.topRight.y;

    // ray is vertically up or down
    if (epsilon > Math.abs(v1)) {
      return new Vector(c1, (v2 > 0 ? y2 : y1));
    }

    // ray is horizontally left or right
    if (epsilon > Math.abs(v2)) {
      return new Vector((v1 > 0 ? x2 : x1), c2);
    }

    let tTop = (y2 - c2) / v2;
    let tBottom = (y1 - c2) / v2;
    let tLeft = (x1 - c1) / v1;
    let tRight = (x2 - c1) / v1;

    // Exactly one t value should be both positive and result in a point
    // within the rectangle

    let tValues = [tTop, tBottom, tLeft, tRight];
    for (let i = 0; i  epsilon && this.contains(intersection)) {
        return intersection;
      }
    } 

    throw "Unexpected error: ray never intersects rectangle!";
  }

Next, splitRay splits a ray into a single line segment and the “remaining” ray, by computing the ray’s intersection with the rectangle, and having the “remaining” ray mirror the direction of approach with a new center that lies on the wall of the rectangle. The new ray length is appropriately shorter. If we run out of ray length, we simply return a segment with a null ray.

  splitRay(ray) {
    let segment = [ray.center, this.rayIntersection(ray)];
    let segmentLength = segment[0].subtract(segment[1]).norm();
    let remainingLength = ray.length - segmentLength;

    if (remainingLength < 10) {
      return {
        segment: [ray.center, ray.endpoint()],
        ray: null
      };
    }

    let vertical = this.isOnVerticalWall(segment[1]);
    let newRayDirection = null;

    if (vertical) {
      newRayDirection = new Vector(-ray.direction.x, ray.direction.y);
    } else {
      newRayDirection = new Vector(ray.direction.x, -ray.direction.y);
    }

    let newRay = new Ray(segment[1], newRayDirection, length=remainingLength);
    return {
      segment: segment,
      ray: newRay
    };
  }

As you have probably guessed, rayToPoints simply calls splitRay over and over again until the ray hits an input “stopping point”—a guard, the target, or the assassin—or else our finite ray length has been exhausted. The output is a list of points, starting from the original ray’s center, for which adjacent pairs are interpreted as line segments to draw.

  rayToPoints(ray, stoppingPoints) {
    let points = [ray.center];
    let remainingRay = ray;

    while (remainingRay) {
      // check if the ray would hit any guards or the target
      if (stoppingPoints) {
        let hardStops = stoppingPoints.map(p => remainingRay.intersects(p))
          .filter(p => p != null);
        if (hardStops.length > 0) {
          // find first intersection and break
          let closestStop = remainingRay.closestToCenter(hardStops);
          points.push(closestStop);
          break;
        }
      }

      let rayPieces = this.splitRay(remainingRay);
      points.push(rayPieces.segment[1]);
      remainingRay = rayPieces.ray;
    } 

    return points;
  }

That’s sufficient to draw the shot emanating from the assassin. This method is called every time the mouse moves.

Optimal guards

The function to compute the optimal position of the guards takes as input the containing rectangle, the assassin, and the target, and produces as output a list of 16 points.

/*
 * Compute the 16 optimal guards to prevent the assassin from hitting the
 * target.
 */
function computeOptimalGuards(square, assassin, target) {
...
}

If you read Tai-Danae’s proof, you’ll know that this construction is to

  1. Compute mirrors of the target across the top, the right, and the top+right of the rectangle. Call this resulting thing the 4-mirrored-targets.
  2. Replicate the 4-mirrored-targets four times, by translating three of the copies left by the entire width of the 4-mirrored-targets shape, down by the entire height, and both left-and-down.
  3. Now you have 16 copies of the target, and one assassin. This gives 16 line segments from assassin-to-target-copy. Place a guard at the midpoint of each of these line segments.
  4. Finally, apply the reverse translation and reverse mirroring to return the guards to the original square.

Due to WordPress being a crappy blogging platform I need to migrate off of, the code snippets below have been magically disappearing. I’ve included links to github lines as well.

Step 1 (after adding simple helper functions on Rectangle to do the mirroring):

  // First compute the target copies in the 4 mirrors
  let target1 = target.copy();
  let target2 = square.mirrorTop(target);
  let target3 = square.mirrorRight(target);
  let target4 = square.mirrorTop(square.mirrorRight(target));
  target1.guardLabel = 1;
  target2.guardLabel = 2;
  target3.guardLabel = 3;
  target4.guardLabel = 4;

Step 2:

  // for each mirrored target, compute the four two-square-length translates
  let mirroredTargets = [target1, target2, target3, target4];
  let horizontalShift = 2 * square.width();
  let verticalShift = 2 * square.height();
  let translateLeft = new Vector(-horizontalShift, 0);
  let translateRight = new Vector(horizontalShift, 0);
  let translateUp = new Vector(0, verticalShift);
  let translateDown = new Vector(0, -verticalShift);

  let translatedTargets = [];
  for (let i = 0; i < mirroredTargets.length; i++) {
    let target = mirroredTargets[i];
    translatedTargets.push([
      target,
      target.add(translateLeft),
      target.add(translateDown),
      target.add(translateLeft).add(translateDown),
    ]);
  }

Step 3, computing the midpoints:

  // compute the midpoints between the assassin and each translate
  let translatedMidpoints = [];
  for (let i = 0; i  t.midpoint(assassin)));
  }

Step 4, returning the guards back to the original square, is harder than it seems, because the midpoint of an assassin-to-target-copy segment might not be in the same copy of the square as the target-copy being fired at. This means you have to detect which square copy the midpoint lands in, and use that to determine which operations are required to invert. This results in the final block of this massive function.

  // determine which of the four possible translates the midpoint is in
  // and reverse the translation. Since midpoints can end up in completely
  // different copies of the square, we have to check each one for all cases.
  function untranslate(point) {
    if (point.x  square.bottomLeft.y) {
      return point.add(translateRight);
    } else if (point.x >= square.bottomLeft.x && point.y <= square.bottomLeft.y) {
      return point.add(translateUp);
    } else if (point.x < square.bottomLeft.x && point.y <= square.bottomLeft.y) {
      return point.add(translateRight).add(translateUp);
    } else {
      return point;
    }
  }

  // undo the translations to get the midpoints back to the original 4-mirrored square.
  let untranslatedMidpoints = [];
  for (let i = 0; i  square.topRight.x && point.y > square.topRight.y) {
      return square.mirrorTop(square.mirrorRight(point));
    } else if (point.x > square.topRight.x && point.y <= square.topRight.y) {
      return square.mirrorRight(point);
    } else if (point.x  square.topRight.y) {
      return square.mirrorTop(point);
    } else {
      return point;
    }
  }

  return untranslatedMidpoints.map(unmirror);

And that’s all there is to it!

Improvements, if I only had the time

There are a few improvements I’d like to make to this puzzle, but haven’t made the time (I’m writing a book, after all!).

  1. Be able to drag the guards around.
  2. Create new guards from an empty set of guards, with a button to “reveal” the solution.
  3. Include a toggle that, when pressed, darkens the entire region of the square that can be hit by the assassin. For example, this would allow you to see if the target is in the only possible safe spot, or if there are multiple safe spots for a given configuration.
  4. Perhaps darken the vulnerable spots by the number of possible paths that hit it, up to some limit.
  5. The most complicated one: generalize to an arbitrary polygon (convex or not!), for which there may be no optional solution. The visualization would allow you to look for a solution using 2-4.

Pull requests are welcome if you attempt any of these improvements.

Until next time!

Formulating the Support Vector Machine Optimization Problem

The hypothesis and the setup

This blog post has an interactive demo (mostly used toward the end of the post). The source for this demo is available in a Github repository.

Last time we saw how the inner product of two vectors gives rise to a decision rule: if $ w$ is the normal to a line (or hyperplane) $ L$, the sign of the inner product $ \langle x, w \rangle$ tells you whether $ x$ is on the same side of $ L$ as $ w$.

Let’s translate this to the parlance of machine-learning. Let $ x \in \mathbb{R}^n$ be a training data point, and $ y \in \{ 1, -1 \}$ is its label (green and red, in the images in this post). Suppose you want to find a hyperplane which separates all the points with -1 labels from those with +1 labels (assume for the moment that this is possible). For this and all examples in this post, we’ll use data in two dimensions, but the math will apply to any dimension.

problem_setup

Some data labeled red and green, which is separable by a hyperplane (line).

The hypothesis we’re proposing to separate these points is a hyperplane, i.e. a linear subspace that splits all of $ \mathbb{R}^n$ into two halves. The data that represents this hyperplane is a single vector $ w$, the normal to the hyperplane, so that the hyperplane is defined by the solutions to the equation $ \langle x, w \rangle = 0$.

As we saw last time, $ w$ encodes the following rule for deciding if a new point $ z$ has a positive or negative label.

$ \displaystyle h_w(z) = \textup{sign}(\langle w, x \rangle)$

You’ll notice that this formula only works for the normals $ w$ of hyperplanes that pass through the origin, and generally we want to work with data that can be shifted elsewhere. We can resolve this by either adding a fixed term $ b \in \mathbb{R}$—often called a bias because statisticians came up with it—so that the shifted hyperplane is the set of solutions to $ \langle x, w \rangle + b = 0$. The shifted decision rule is:

$ \displaystyle h_w(z) = \textup{sign}(\langle w, x \rangle + b)$

Now the hypothesis is the pair of vector-and-scalar $ w, b$.

The key intuitive idea behind the formulation of the SVM problem is that there are many possible separating hyperplanes for a given set of labeled training data. For example, here is a gif showing infinitely many choices.

svm_lots_of_choices.gif

The question is: how can we find the separating hyperplane that not only separates the training data, but generalizes as well as possible to new data? The assumption of the SVM is that a hyperplane which separates the points, but is also as far away from any training point as possible, will generalize best.

optimal_example.png

While contrived, it’s easy to see that the separating hyperplane is as far as possible from any training point.

More specifically, fix a labeled dataset of points $ (x_i, y_i)$, or more precisely:

$ \displaystyle D = \{ (x_i, y_i) \mid i = 1, \dots, m, x_i \in \mathbb{R}^{n}, y_i \in \{1, -1\}  \}$

And a hypothesis defined by the normal $ w \in \mathbb{R}^{n}$ and a shift $ b \in \mathbb{R}$. Let’s also suppose that $ (w,b)$ defines a hyperplane that correctly separates all the training data into the two labeled classes, and we just want to measure its quality. That measure of quality is the length of its margin.

Definition: The geometric margin of a hyperplane $ w$ with respect to a dataset $ D$ is the shortest distance from a training point $ x_i$ to the hyperplane defined by $ w$.

The best hyperplane has the largest possible margin.

This margin can even be computed quite easily using our work from last post. The distance from $ x$ to the hyperplane defined by $ w$ is the same as the length of the projection of $ x$ onto $ w$. And this is just computed by an inner product.

decision-rule-3

If the tip of the $ x$ arrow is the point in question, then $ a$ is the dot product, and $ b$ the distance from $ x$ to the hyperplane $ L$ defined by $ w$.

A naive optimization objective

If we wanted to, we could stop now and define an optimization problem that would be very hard to solve. It would look like this:

$ \displaystyle \begin{aligned} & \max_{w} \min_{x_i} \left | \left \langle x_i, \frac{w}{\|w\|} \right \rangle + b \right | & \\ \textup{subject to \ \ } & \textup{sign}(\langle x_i, w \rangle + b) = \textup{sign}(y_i) & \textup{ for every } i = 1, \dots, m \end{aligned}$

The formulation is hard. The reason is it’s horrifyingly nonlinear. In more detail:

  1. The constraints are nonlinear due to the sign comparisons.
  2. There’s a min and a max! A priori, we have to do this because we don’t know which point is going to be the closest to the hyperplane.
  3. The objective is nonlinear in two ways: the absolute value and the projection requires you to take a norm and divide.

The rest of this post (and indeed, a lot of the work in grokking SVMs) is dedicated to converting this optimization problem to one in which the constraints are all linear inequalities and the objective is a single, quadratic polynomial we want to minimize or maximize.

Along the way, we’ll notice some neat features of the SVM.

Trick 1: linearizing the constraints

To solve the first problem, we can use a trick. We want to know whether $ \textup{sign}(\langle x_i, w \rangle + b) = \textup{sign}(y_i)$ for a labeled training point $ (x_i, y_i)$. The trick is to multiply them together. If their signs agree, then their product will be positive, otherwise it will be negative.

So each constraint becomes:

$ \displaystyle (\langle x_i, w \rangle + b) \cdot y_i \geq 0$

This is still linear because $ y_i$ is a constant (input) to the optimization problem. The variables are the coefficients of $ w$.

The left hand side of this inequality is often called the functional margin of a training point, since, as we will see, it still works to classify $ x_i$, even if $ w$ is scaled so that it is no longer a unit vector. Indeed, the sign of the inner product is independent of how $ w$ is scaled.

Trick 1.5: the optimal solution is midway between classes

This small trick is to notice that if $ w$ is the supposed optimal separating hyperplane, i.e. its margin is maximized, then it must necessarily be exactly halfway in between the closest points in the positive and negative classes.

In other words, if $ x_+$ and $ x_-$ are the closest points in the positive and negative classes, respectively, then $ \langle x_{+}, w \rangle + b = -(\langle x_{-}, w \rangle + b)$. If this were not the case, then you could adjust the bias, shifting the decision boundary along $ w$ until it they are exactly equal, and you will have increased the margin. The closest point, say $ x_+$ will have gotten farther away, and the closest point in the opposite class, $ x_-$ will have gotten closer, but will not be closer than $ x_+$.

Trick 2: getting rid of the max + min

Resolving this problem essentially uses the fact that the hypothesis, which comes in the form of the normal vector $ w$, has a degree of freedom in its length. To explain the details of this trick, we’ll set $ b=0$ which simplifies the intuition.

Indeed, in the animation below, I can increase or decrease the length of $ w$ without changing the decision boundary.

svm_w_length.gif

I have to keep my hand very steady (because I was too lazy to program it so that it only increases/decreases in length), but you can see the point. The line is perpendicular to the normal vector, and it doesn’t depend on the length.

Let’s combine this with tricks 1 and 1.5. If we increase the length of $ w$, that means the absolute values of the dot products $ \langle x_i, w \rangle $ used in the constraints will all increase by the same amount (without changing their sign). Indeed, for any vector $ a$ we have $ \langle a, w \rangle = \|w \| \cdot \langle a, w / \| w \| \rangle$.

In this world, the inner product measurement of distance from a point to the hyperplane is no longer faithful. The true distance is $ \langle a, w / \| w \| \rangle$, but the distance measured by $ \langle a, w \rangle$ is measured in units of $ 1 / \| w \|$.

units.png

In this example, the two numbers next to the green dot represent the true distance of the point from the hyperplane, and the dot product of the point with the normal (respectively). The dashed lines are the solutions to <x, w> = 1. The magnitude of w is 2.2, the inverse of that is 0.46, and indeed 2.2 = 4.8 * 0.46 (we’ve rounded the numbers).

Now suppose we had the optimal hyperplane and its normal $ w$. No matter how near (or far) the nearest positively labeled training point $ x$ is, we could scale the length of $ w$ to force $ \langle x, w \rangle = 1$. This is the core of the trick. One consequence is that the actual distance from $ x$ to the hyperplane is $ \frac{1}{\| w \|} = \langle x, w / \| w \| \rangle$.

units2.png

The same as above, but with the roles reversed. We’re forcing the inner product of the point with w to be 1. The true distance is unchanged.

In particular, if we force the closest point to have inner product 1, then all other points will have inner product at least 1. This has two consequences. First, our constraints change to $ \langle x_i, w \rangle \cdot y_i \geq 1$ instead of $ \geq 0$. Second, we no longer need to ask which point is closest to the candidate hyperplane! Because after all, we never cared which point it was, just how far away that closest point was. And now we know that it’s exactly $ 1 / \| w \|$ away. Indeed, if the optimal points weren’t at that distance, then that means the closest point doesn’t exactly meet the constraint, i.e. that $ \langle x, w \rangle > 1$ for every training point $ x$. We could then scale $ w$ shorter until $ \langle x, w \rangle = 1$, hence increasing the margin $ 1 / \| w \|$.

In other words, the coup de grâce, provided all the constraints are satisfied, the optimization objective is just to maximize $ 1 / \| w \|$, a.k.a. to minimize $ \| w \|$.

This intuition is clear from the following demonstration, which you can try for yourself. In it I have a bunch of positively and negatively labeled points, and the line in the center is the candidate hyperplane with normal $ w$ that you can drag around. Each training point has two numbers next to it. The first is the true distance from that point to the candidate hyperplane; the second is the inner product with $ w$. The two blue dashed lines are the solutions to $ \langle x, w \rangle = \pm 1$. To solve the SVM by hand, you have to ensure the second number is at least 1 for all green points, at most -1 for all red points, and then you have to make $ w$ as short as possible. As we’ve discussed, shrinking $ w$ moves the blue lines farther away from the separator, but in order to satisfy the constraints the blue lines can’t go further than any training point. Indeed, the optimum will have those blue lines touching a training point on each side.

svm_solve_by_hand

 

I bet you enjoyed watching me struggle to solve it. And while it’s probably not the optimal solution, the idea should be clear.

The final note is that, since we are now minimizing $ \| w \|$, a formula which includes a square root, we may as well minimize its square $ \| w \|^2 = \sum_j w_j^2$. We will also multiply the objective by $ 1/2$, because when we eventually analyze this problem we will take a derivative, and the square in the exponent and the $ 1/2$ will cancel.

The final form of the problem

Our optimization problem is now the following (including the bias again):

$ \displaystyle \begin{aligned} & \min_{w}  \frac{1}{2} \| w \|^2 & \\ \textup{subject to \ \ } & (\langle x_i, w \rangle + b) \cdot y_i \geq 1 & \textup{ for every } i = 1, \dots, m \end{aligned}$

This is much simpler to analyze. The constraints are all linear inequalities (which, because of linear programming, we know are tractable to optimize). The objective to minimize, however, is a convex quadratic function of the input variables—a sum of squares of the inputs.

Such problems are generally called quadratic programming problems (or QPs, for short). There are general methods to find solutions! However, they often suffer from numerical stability issues and have less-than-satisfactory runtime. Luckily, the form in which we’ve expressed the support vector machine problem is specific enough that we can analyze it directly, and find a way to solve it without appealing to general-purpose numerical solvers.

We will tackle this problem in a future post (planned for two posts sequel to this one). Before we close, let’s just make a few more observations about the solution to the optimization problem.

Support Vectors

In Trick 1.5 we saw that the optimal separating hyperplane has to be exactly halfway between the two closest points of opposite classes. Moreover, we noticed that, provided we’ve scaled $ \| w \|$ properly, these closest points (there may be multiple for positive and negative labels) have to be exactly “distance” 1 away from the separating hyperplane.

Another way to phrase this without putting “distance” in scare quotes is to say that, if $ w$ is the normal vector of the optimal separating hyperplane, the closest points lie on the two lines $ \langle x_i, w \rangle + b = \pm 1$.

Now that we have some intuition for the formulation of this problem, it isn’t a stretch to realize the following. While a dataset may include many points from either class on these two lines $ \langle x_i, w \rangle = \pm 1$, the optimal hyperplane itself does not depend on any of the other points except these closest points.

This fact is enough to give these closest points a special name: the support vectors.

We’ll actually prove that support vectors “are all you need” with full rigor and detail next time, when we cast the optimization problem in this post into the “dual” setting. To avoid vague names, the formulation described in this post called the “primal” problem. The dual problem is derived from the primal problem, with special variables and constraints chosen based on the primal variables and constraints. Next time we’ll describe in brief detail what the dual does and why it’s important, but we won’t have nearly enough time to give a full understanding of duality in optimization (such a treatment would fill a book).

When we compute the dual of the SVM problem, we will see explicitly that the hyperplane can be written as a linear combination of the support vectors. As such, once you’ve found the optimal hyperplane, you can compress the training set into just the support vectors, and reproducing the same optimal solution becomes much, much faster. You can also use the support vectors to augment the SVM to incorporate streaming data (throw out all non-support vectors after every retraining).

Eventually, when we get to implementing the SVM from scratch, we’ll see all this in action.

Until then!

The Inner Product as a Decision Rule

The standard inner product of two vectors has some nice geometric properties. Given two vectors $ x, y \in \mathbb{R}^n$, where by $ x_i$ I mean the $ i$-th coordinate of $ x$, the standard inner product (which I will interchangeably call the dot product) is defined by the formula

$ \displaystyle \langle x, y \rangle = x_1 y_1 + \dots + x_n y_n$

This formula, simple as it is, produces a lot of interesting geometry. An important such property, one which is discussed in machine learning circles more than pure math, is that it is a very convenient decision rule.

In particular, say we’re in the Euclidean plane, and we have a line $ L$ passing through the origin, with $ w$ being a unit vector perpendicular to $ L$ (“the normal” to the line).

decision-rule-1

If you take any vector $ x$, then the dot product $ \langle x, w \rangle$ is positive if $ x$ is on the same side of $ L$ as $ w$, and negative otherwise. The dot product is zero if and only if $ x$ is exactly on the line $ L$, including when $ x$ is the zero vector.

decision-rule-2

Left: the dot product of $ w$ and $ x$ is positive, meaning they are on the same side of $ w$. Right: The dot product is negative, and they are on opposite sides.

Here is an interactive demonstration of this property. Click the image below to go to the demo, and you can drag the vector arrowheads and see the decision rule change.

decision-rule

Click above to go to the demo

The code for this demo is available in a github repository.

It’s always curious, at first, that multiplying and summing produces such geometry. Why should this seemingly trivial arithmetic do anything useful at all?

The core fact that makes it work, however, is that the dot product tells you how one vector projects onto another. When I say “projecting” a vector $ x$ onto another vector $ w$, I mean you take only the components of $ x$ that point in the direction of $ w$. The demo shows what the result looks like using the red (or green) vector.

In two dimensions this is easy to see, as you can draw the triangle which has $ x$ as the hypotenuse, with $ w$ spanning one of the two legs of the triangle as follows:

decision-rule-3.png

If we call $ a$ the (vector) leg of the triangle parallel to $ w$, while $ b$ is the dotted line (as a vector, parallel to $ L$), then as vectors $ x = a + b$. The projection of $ x$ onto $ w$ is just $ a$.

Another way to think of this is that the projection is $ x$, modified by removing any part of $ x$ that is perpendicular to $ w$. Using some colorful language: you put your hands on either side of $ x$ and $ w$, and then you squish $ x$ onto $ w$ along the line perpendicular to $ w$ (i.e., along $ b$).

And if $ w$ is a unit vector, then the length of $ a$—that is, the length of the projection of $ x$ onto $ w$—is exactly the inner product product $ \langle x, w \rangle$.

Moreover, if the angle between $ x$ and $ w$ is larger than 90 degrees, the projected vector will point in the opposite direction of $ w$, so it’s really a “signed” length.

decision-rule-4

Left: the projection points in the same direction as $ w$. Right: the projection points in the opposite direction.

And this is precisely why the decision rule works. This 90-degree boundary is the line perpendicular to $ w$.

More technically said: Let $ x, y \in \mathbb{R}^n$ be two vectors, and $ \langle x,y \rangle $ their dot product. Define by $ \| y \|$ the length of $ y$, specifically $ \sqrt{\langle y, y \rangle}$. Define by $ \text{proj}_{y}(x)$ by first letting $ y’ = \frac{y}{\| y \|}$, and then let $ \text{proj}_{y}(x) = \langle x,y’ \rangle y’$. In words, you scale $ y$ to a unit vector $ y’$, use the result to compute the inner product, and then scale $ y$ so that it’s length is $ \langle x, y’ \rangle$. Then

Theorem: Geometrically, $ \text{proj}_y(x)$ is the projection of $ x$ onto the line spanned by $ y$.

This theorem is true for any $ n$-dimensional vector space, since if you have two vectors you can simply apply the reasoning for 2-dimensions to the 2-dimensional plane containing $ x$ and $ y$. In that case, the decision boundary for a positive/negative output is the entire $ n-1$ dimensional hyperplane perpendicular to $ y$ (the projected vector).

In fact, the usual formula for the angle between two vectors, i.e. the formula $ \langle x, y \rangle = \|x \| \cdot \| y \| \cos \theta$, is a restatement of the projection theorem in terms of trigonometry. The $ \langle x, y’ \rangle$ part of the projection formula (how much you scale the output) is equal to $ \| x \| \cos \theta$. At the end of this post we have a proof of the cosine-angle formula above.

Part of why this decision rule property is so important is that this is a linear function, and linear functions can be optimized relatively easily. When I say that, I specifically mean that there are many known algorithms for optimizing linear functions, which don’t have obscene runtime or space requirements. This is a big reason why mathematicians and statisticians start the mathematical modeling process with linear functions. They’re inherently simpler.

In fact, there are many techniques in machine learning—a prominent one is the so-called Kernel Trick—that exist solely to take data that is not inherently linear in nature (cannot be fruitfully analyzed by linear methods) and transform it into a dataset that is. Using the Kernel Trick as an example to foreshadow some future posts on Support Vector Machines, the idea is to take data which cannot be separated by a line, and transform it (usually by adding new coordinates) so that it can. Then the decision rule, computed in the larger space, is just a dot product. Irene Papakonstantinou neatly demonstrates this with paper folding and scissors. The tradeoff is that the size of the ambient space increases, and it might increase so much that it makes computation intractable. Luckily, the Kernel Trick avoids this by remembering where the data came from, so that one can take advantage of the smaller space to compute what would be the inner product in the larger space.

Next time we’ll see how this decision rule shows up in an optimization problem: finding the “best” hyperplane that separates an input set of red and blue points into monochromatic regions (provided that is possible). Finding this separator is core subroutine of the Support Vector Machine technique, and therein lie interesting algorithms. After we see the core SVM algorithm, we’ll see how the Kernel Trick fits into the method to allow nonlinear decision boundaries.


Proof of the cosine angle formula

Theorem: The inner product $ \langle v, w \rangle$ is equal to $ \| v \| \| w \| \cos(\theta)$, where $ \theta$ is the angle between the two vectors.

Note that this angle is computed in the 2-dimensional subspace spanned by $ v, w$, viewed as a typical flat plane, and this is a 2-dimensional plane regardless of the dimension of $ v, w$.

Proof. If either $ v$ or $ w$ is zero, then both sides of the equation are zero and the theorem is trivial, so we may assume both are nonzero. Label a triangle with sides $ v,w$ and the third side $ v-w$. Now the length of each side is $ \| v \|, \| w\|,$ and $ \| v-w \|$, respectively. Assume for the moment that $ \theta$ is not 0 or 180 degrees, so that this triangle is not degenerate.

The law of cosines allows us to write

$ \displaystyle \| v – w \|^2 = \| v \|^2 + \| w \|^2 – 2 \| v \| \| w \| \cos(\theta)$

Moreover, The left hand side is the inner product of $ v-w$ with itself, i.e. $ \| v – w \|^2 = \langle v-w , v-w \rangle$. We’ll expand $ \langle v-w, v-w \rangle$ using two facts. The first is trivial from the formula, that inner product is symmetric: $ \langle v,w \rangle = \langle w, v \rangle$. Second is that the inner product is linear in each input. In particular for the first input: $ \langle x + y, z \rangle = \langle x, z \rangle + \langle y, z \rangle$ and $ \langle cx, z \rangle = c \langle x, z \rangle$. The same holds for the second input by symmetry of the two inputs. Hence we can split up $ \langle v-w, v-w \rangle$ as follows.

$ \displaystyle \begin{aligned} \langle v-w, v-w \rangle &= \langle v, v-w \rangle – \langle w, v-w \rangle \\ &= \langle v, v \rangle – \langle v, w \rangle – \langle w, v \rangle +  \langle w, w \rangle \\ &= \| v \|^2 – 2 \langle v, w \rangle + \| w \|^2 \\ \end{aligned}$

Combining our two offset equations, we can subtract $ \| v \|^2 + \| w \|^2$ from each side and get

$ \displaystyle -2 \|v \| \|w \| \cos(\theta) = -2 \langle v, w \rangle, $

Which, after dividing by $ -2$, proves the theorem if $ \theta \not \in \{0, 180 \}$.

Now if $ \theta = 0$ or 180 degrees, the vectors are parallel, so we can write one as a scalar multiple of the other. Say $ w = cv$ for $ c \in \mathbb{R}$. In that case, $ \langle v, cv \rangle = c \| v \| \| v \|$. Now $ \| w \| = | c | \| v \|$, since a norm is a length and is hence non-negative (but $ c$ can be negative). Indeed, if $ v, w$ are parallel but pointing in opposite directions, then $ c < 0$, so $ \cos(\theta) = -1$, and $ c \| v \| = – \| w \|$. Otherwise $ c > 0$ and $ \cos(\theta) = 1$. This allows us to write $ c \| v \| \| v \| = \| w \| \| v \| \cos(\theta)$, and this completes the final case of the theorem.

$ \square$

The Reasonable Effectiveness of the Multiplicative Weights Update Algorithm

papad

Christos Papadimitriou, who studies multiplicative weights in the context of biology.

Hard to believe

Sanjeev Arora and his coauthors consider it “a basic tool [that should be] taught to all algorithms students together with divide-and-conquer, dynamic programming, and random sampling.” Christos Papadimitriou calls it “so hard to believe that it has been discovered five times and forgotten.” It has formed the basis of algorithms in machine learning, optimization, game theory, economics, biology, and more.

What mystical algorithm has such broad applications? Now that computer scientists have studied it in generality, it’s known as the Multiplicative Weights Update Algorithm (MWUA). Procedurally, the algorithm is simple. I can even describe the core idea in six lines of pseudocode. You start with a collection of $ n$ objects, and each object has a weight.

Set all the object weights to be 1.
For some large number of rounds:
   Pick an object at random proportionally to the weights
   Some event happens
   Increase the weight of the chosen object if it does well in the event
   Otherwise decrease the weight

The name “multiplicative weights” comes from how we implement the last step: if the weight of the chosen object at step $ t$ is $ w_t$ before the event, and $ G$ represents how well the object did in the event, then we’ll update the weight according to the rule:

$ \displaystyle w_{t+1} = w_t (1 + G)$

Think of this as increasing the weight by a small multiple of the object’s performance on a given round.

Here is a simple example of how it might be used. You have some money you want to invest, and you have a bunch of financial experts who are telling you what to invest in every day. So each day you pick an expert, and you follow their advice, and you either make a thousand dollars, or you lose a thousand dollars, or something in between. Then you repeat, and your goal is to figure out which expert is the most reliable.

This is how we use multiplicative weights: if we number the experts $ 1, \dots, N$, we give each expert a weight $ w_i$ which starts at 1. Then, each day we pick an expert at random (where experts with larger weights are more likely to be picked) and at the end of the day we have some gain or loss $ G$. Then we update the weight of the chosen expert by multiplying it by $ (1 + G / 1000)$. Sometimes you have enough information to update the weights of experts you didn’t choose, too. The theoretical guarantees of the algorithm say we’ll find the best expert quickly (“quickly” will be concrete later).

In fact, let’s play a game where you, dear reader, get to decide the rewards for each expert and each day. I programmed the multiplicative weights algorithm to react according to your choices. Click the image below to go to the demo.

mwua

This core mechanism of updating weights can be interpreted in many ways, and that’s part of the reason it has sprouted up all over mathematics and computer science. Just a few examples of where this has led:

  1. In game theory, weights are the “belief” of a player about the strategy of an opponent. The most famous algorithm to use this is called Fictitious Play, and others include EXP3 for minimizing regret in the so-called “adversarial bandit learning” problem.
  2. In machine learning, weights are the difficulty of a specific training example, so that higher weights mean the learning algorithm has to “try harder” to accommodate that example. The first result I’m aware of for this is the Perceptron (and similar Winnow) algorithm for learning hyperplane separators. The most famous is the AdaBoost algorithm.
  3. Analogously, in optimization, the weights are the difficulty of a specific constraint, and this technique can be used to approximately solve linear and semidefinite programs. The approximation is because MWUA only provides a solution with some error.
  4. In mathematical biology, the weights represent the fitness of individual alleles, and filtering reproductive success based on this and updating weights for successful organisms produces a mechanism very much like evolution. With modifications, it also provides a mechanism through which to understand sex in the context of evolutionary biology.
  5. The TCP protocol, which basically defined the internet, uses additive and multiplicative weight updates (which are very similar in the analysis) to manage congestion.
  6. You can get easy $ \log(n)$-approximation algorithms for many NP-hard problems, such as set cover.

Additional, more technical examples can be found in this survey of Arora et al.

In the rest of this post, we’ll implement a generic Multiplicative Weights Update Algorithm, we’ll prove it’s main theoretical guarantees, and we’ll implement a linear program solver as an example of its applicability. As usual, all of the code used in the making of this post is available in a Github repository.

The generic MWUA algorithm

Let’s start by writing down pseudocode and an implementation for the MWUA algorithm in full generality.

In general we have some set $ X$ of objects and some set $ Y$ of “event outcomes” which can be completely independent. If these sets are finite, we can write down a table $ M$ whose rows are objects, whose columns are outcomes, and whose $ i,j$ entry $ M(i,j)$ is the reward produced by object $ x_i$ when the outcome is $ y_j$. We will also write this as $ M(x, y)$ for object $ x$ and outcome $ y$. The only assumption we’ll make on the rewards is that the values $ M(x, y)$ are bounded by some small constant $ B$ (by small I mean $ B$ should not require exponentially many bits to write down as compared to the size of $ X$). In symbols, $ M(x,y) \in [0,B]$. There are minor modifications you can make to the algorithm if you want negative rewards, but for simplicity we will leave that out. Note the table $ M$ just exists for analysis, and the algorithm does not know its values. Moreover, while the values in $ M$ are static, the choice of outcome $ y$ for a given round may be nondeterministic.

The MWUA algorithm randomly chooses an object $ x \in X$ in every round, observing the outcome $ y \in Y$, and collecting the reward $ M(x,y)$ (or losing it as a penalty). The guarantee of the MWUA theorem is that the expected sum of rewards/penalties of MWUA is not much worse than if one had picked the best object (in hindsight) every single round.

Let’s describe the algorithm in notation first and build up pseudocode as we go. The input to the algorithm is the set of objects, a subroutine that observes an outcome, a black-box reward function, a learning rate parameter, and a number of rounds.

def MWUA(objects, observeOutcome, reward, learningRate, numRounds):
   ...

We define for object $ x$ a nonnegative number $ w_x$ we call a “weight.” The weights will change over time so we’ll also sub-script a weight with a round number $ t$, i.e. $ w_{x,t}$ is the weight of object $ x$ in round $ t$. Initially, all the weights are $ 1$. Then MWUA continues in rounds. We start each round by drawing an example randomly with probability proportional to the weights. Then we observe the outcome for that round and the reward for that round.

# draw: [float] -&gt; int
# pick an index from the given list of floats proportionally
# to the size of the entry (i.e. normalize to a probability
# distribution and draw according to the probabilities).
def draw(weights):
    choice = random.uniform(0, sum(weights))
    choiceIndex = 0

    for weight in weights:
        choice -= weight
        if choice &lt;= 0:
            return choiceIndex

        choiceIndex += 1

# MWUA: the multiplicative weights update algorithm
def MWUA(objects, observeOutcome, reward, learningRate numRounds):
   weights = [1] * len(objects)
   for t in numRounds:
      chosenObjectIndex = draw(weights)
      chosenObject = objects[chosenObjectIndex]

      outcome = observeOutcome(t, weights, chosenObject)
      thisRoundReward = reward(chosenObject, outcome)

      ...

Sampling objects in this way is the same as associating a distribution $ D_t$ to each round, where if $ S_t = \sum_{x \in X} w_{x,t}$ then the probability of drawing $ x$, which we denote $ D_t(x)$, is $ w_{x,t} / S_t$. We don’t need to keep track of this distribution in the actual run of the algorithm, but it will help us with the mathematical analysis.

Next comes the weight update step. Let’s call our learning rate variable parameter $ \varepsilon$. In round $ t$ say we have object $ x_t$ and outcome $ y_t$, then the reward is $ M(x_t, y_t)$. We update the weight of the chosen object $ x_t$ according to the formula:

$ \displaystyle w_{x_t, t} = w_{x_t} (1 + \varepsilon M(x_t, y_t) / B)$

In the more general event that you have rewards for all objects (if not, the reward-producing function can output zero), you would perform this weight update on all objects $ x \in X$. This turns into the following Python snippet, where we hide the division by $ B$ into the choice of learning rate:

# MWUA: the multiplicative weights update algorithm
def MWUA(objects, observeOutcome, reward, learningRate, numRounds):
   weights = [1] * len(objects)
   for t in numRounds:
      chosenObjectIndex = draw(weights)
      chosenObject = objects[chosenObjectIndex]

      outcome = observeOutcome(t, weights, chosenObject)
      thisRoundReward = reward(chosenObject, outcome)

      for i in range(len(weights)):
         weights[i] *= (1 + learningRate * reward(objects[i], outcome))

One of the amazing things about this algorithm is that the outcomes and rewards could be chosen adaptively by an adversary who knows everything about the MWUA algorithm (except which random numbers the algorithm generates to make its choices). This means that the rewards in round $ t$ can depend on the weights in that same round! We will exploit this when we solve linear programs later in this post.

But even in such an oppressive, exploitative environment, MWUA persists and achieves its guarantee. And now we can state that guarantee.

Theorem (from Arora et al): The cumulative reward of the MWUA algorithm is, up to constant multiplicative factors, at least the cumulative reward of the best object minus $ \log(n)$, where $ n$ is the number of objects. (Exact formula at the end of the proof)

The core of the proof, which we’ll state as a lemma, uses one of the most elegant proof techniques in all of mathematics. It’s the idea of constructing a potential function, and tracking the change in that potential function over time. Such a proof usually has the mysterious script:

  1. Define potential function, in our case $ S_t$.
  2. State what seems like trivial facts about the potential function to write $ S_{t+1}$ in terms of $ S_t$, and hence get general information about $ S_T$ for some large $ T$.
  3. Theorem is proved.
  4. Wait, what?

Clearly, coming up with a useful potential function is a difficult and prized skill.

In this proof our potential function is the sum of the weights of the objects in a given round, $ S_t = \sum_{x \in X} w_{x, t}$. Now the lemma.

Lemma: Let $ B$ be the bound on the size of the rewards, and $ 0 < \varepsilon < 1/2$ a learning parameter. Recall that $ D_t(x)$ is the probability that MWUA draws object $ x$ in round $ t$. Write the expected reward for MWUA for round $ t$ as the following (using only the definition of expected value):

$ \displaystyle R_t = \sum_{x \in X} D_t(x) M(x, y_t)$

 Then the claim of the lemma is:

$ \displaystyle S_{t+1} \leq S_t e^{\varepsilon R_t / B}$

Proof. Expand $ S_{t+1} = \sum_{x \in X} w_{x, t+1}$ using the definition of the MWUA update:

$ \displaystyle \sum_{x \in X} w_{x, t+1} = \sum_{x \in X} w_{x, t}(1 + \varepsilon M(x, y_t) / B)$

Now distribute $ w_{x, t}$ and split into two sums:

$ \displaystyle \dots = \sum_{x \in X} w_{x, t} + \frac{\varepsilon}{B} \sum_{x \in X} w_{x,t} M(x, y_t)$

Using the fact that $ D_t(x) = \frac{w_{x,t}}{S_t}$, we can replace $ w_{x,t}$ with $ D_t(x) S_t$, which allows us to get $ R_t$

$ \displaystyle \begin{aligned} \dots &= S_t + \frac{\varepsilon S_t}{B} \sum_{x \in X} D_t(x) M(x, y_t) \\ &= S_t \left ( 1 + \frac{\varepsilon R_t}{B} \right ) \end{aligned}$

And then using the fact that $ (1 + x) \leq e^x$ (Taylor series), we can bound the last expression by $ S_te^{\varepsilon R_t / B}$, as desired.

$ \square$

Now using the lemma, we can get a hold on $ S_T$ for a large $ T$, namely that

$ \displaystyle S_T \leq S_1 e^{\varepsilon \sum_{t=1}^T R_t / B}$

If $ |X| = n$ then $ S_1=n$, simplifying the above. Moreover, the sum of the weights in round $ T$ is certainly greater than any single weight, so that for every fixed object $ x \in X$,

$ \displaystyle S_T \geq w_{x,T} \leq  (1 + \varepsilon)^{\sum_t M(x, y_t) / B}$

Squeezing $ S_t$ between these two inequalities and taking logarithms (to simplify the exponents) gives

$ \displaystyle \left ( \sum_t M(x, y_t) / B \right ) \log(1+\varepsilon) \leq \log n + \frac{\varepsilon}{B} \sum_t R_t$

Multiply through by $ B$, divide by $ \varepsilon$, rearrange, and use the fact that when $ 0 < \varepsilon < 1/2$ we have $ \log(1 + \varepsilon) \geq \varepsilon – \varepsilon^2$ (Taylor series) to get

$ \displaystyle \sum_t R_t \geq \left [ \sum_t M(x, y_t) \right ] (1-\varepsilon) – \frac{B \log n}{\varepsilon}$

The bracketed term is the payoff of object $ x$, and MWUA’s payoff is at least a fraction of that minus the logarithmic term. The bound applies to any object $ x \in X$, and hence to the best one. This proves the theorem.

$ \square$

Briefly discussing the bound itself, we see that the smaller the learning rate is, the closer you eventually get to the best object, but by contrast the more the subtracted quantity $ B \log(n) / \varepsilon$ hurts you. If your target is an absolute error bound against the best performing object on average, you can do more algebra to determine how many rounds you need in terms of a fixed $ \delta$. The answer is roughly: let $ \varepsilon = O(\delta / B)$ and pick $ T = O(B^2 \log(n) / \delta^2)$. See this survey for more.

MWUA for linear programs

Now we’ll approximately solve a linear program using MWUA. Recall that a linear program is an optimization problem whose goal is to minimize (or maximize) a linear function of many variables. The objective to minimize is usually given as a dot product $ c \cdot x$, where $ c$ is a fixed vector and $ x = (x_1, x_2, \dots, x_n)$ is a vector of non-negative variables the algorithm gets to choose. The choices for $ x$ are also constrained by a set of $ m$ linear inequalities, $ A_i \cdot x \geq b_i$, where $ A_i$ is a fixed vector and $ b_i$ is a scalar for $ i = 1, \dots, m$. This is usually summarized by putting all the $ A_i$ in a matrix, $ b_i$ in a vector, as

$ x_{\textup{OPT}} = \textup{argmin}_x \{ c \cdot x \mid Ax \geq b, x \geq 0 \}$

We can further simplify the constraints by assuming we know the optimal value $ Z = c \cdot x_{\textup{OPT}}$ in advance, by doing a binary search (more on this later). So, if we ignore the hard constraint $ Ax \geq b$, the “easy feasible region” of possible $ x$’s includes $ \{ x \mid x \geq 0, c \cdot x = Z \}$.

In order to fit linear programming into the MWUA framework we have to define two things.

  1. The objects: the set of linear inequalities $ A_i \cdot x \geq b_i$.
  2. The rewards: the error of a constraint for a special input vector $ x_t$.

Number 2 is curious (why would we give a reward for error?) but it’s crucial and we’ll discuss it momentarily.

The special input $ x_t$ depends on the weights in round $ t$ (which is allowed, recall). Specifically, if the weights are $ w = (w_1, \dots, w_m)$, we ask for a vector $ x_t$ in our “easy feasible region” which satisfies

$ \displaystyle (A^T w) \cdot x_t \geq w \cdot b$

For this post we call the implementation of procuring such a vector the “oracle,” since it can be seen as the black-box problem of, given a vector $ \alpha$ and a scalar $ \beta$ and a convex region $ R$, finding a vector $ x \in R$ satisfying $ \alpha \cdot x \geq \beta$. This allows one to solve more complex optimization problems with the same technique, swapping in a new oracle as needed. Our choice of inputs, $ \alpha = A^T w, \beta = w \cdot b$, are particular to the linear programming formulation.

Two remarks on this choice of inputs. First, the vector $ A^T w$ is a weighted average of the constraints in $ A$, and $ w \cdot b$ is a weighted average of the thresholds. So this this inequality is a “weighted average” inequality (specifically, a convex combination, since the weights are nonnegative). In particular, if no such $ x$ exists, then the original linear program has no solution. Indeed, given a solution $ x^*$ to the original linear program, each constraint, say $ A_1 x^*_1 \geq b_1$, is unaffected by left-multiplication by $ w_1$.

Second, and more important to the conceptual understanding of this algorithm, the choice of rewards and the multiplicative updates ensure that easier constraints show up less prominently in the inequality by having smaller weights. That is, if we end up overly satisfying a constraint, we penalize that object for future rounds so we don’t waste our effort on it. The byproduct of MWUA—the weights—identify the hardest constraints to satisfy, and so in each round we can put a proportionate amount of effort into solving (one of) the hard constraints. This is why it makes sense to reward error; the error is a signal for where to improve, and by over-representing the hard constraints, we force MWUA’s attention on them.

At the end, our final output is an average of the $ x_t$ produced in each round, i.e. $ x^* = \frac{1}{T}\sum_t x_t$. This vector satisfies all the constraints to a roughly equal degree. We will skip the proof that this vector does what we want, but see these notes for a simple proof. We’ll spend the rest of this post implementing the scheme outlined above.

Implementing the oracle

Fix the convex region $ R = \{ c \cdot x = Z, x \geq 0 \}$ for a known optimal value $ Z$. Define $ \textup{oracle}(\alpha, \beta)$ as the problem of finding an $ x \in R$ such that $ \alpha \cdot x \geq \beta$.

For the case of this linear region $ R$, we can simply find the index $ i$ which maximizes $ \alpha_i Z / c_i$. If this value exceeds $ \beta$, we can return the vector with that value in the $ i$-th position and zeros elsewhere. Otherwise, the problem has no solution.

To prove the “no solution” part, say $ n=2$ and you have $ x = (x_1, x_2)$ a solution to $ \alpha \cdot x \geq \beta$. Then for whichever index makes $ \alpha_i Z / c_i$ bigger, say $ i=1$, you can increase $ \alpha \cdot x$ without changing $ c \cdot x = Z$ by replacing $ x_1$ with $ x_1 + (c_2/c_1)x_2$ and $ x_2$ with zero. I.e., we’re moving the solution $ x$ along the line $ c \cdot x = Z$ until it reaches a vertex of the region bounded by $ c \cdot x = Z$ and $ x \geq 0$. This must happen when all entries but one are zero. This is the same reason why optimal solutions of (generic) linear programs occur at vertices of their feasible regions.

The code for this becomes quite simple. Note we use the numpy library in the entire codebase to make linear algebra operations fast and simple to read.

def makeOracle(c, optimalValue):
    n = len(c)

    def oracle(weightedVector, weightedThreshold):
        def quantity(i):
            return weightedVector[i] * optimalValue / c[i] if c[i] &gt; 0 else -1

        biggest = max(range(n), key=quantity)
        if quantity(biggest) &lt; weightedThreshold:
            raise InfeasibleException

        return numpy.array([optimalValue / c[i] if i == biggest else 0 for i in range(n)])

    return oracle

Implementing the core solver

The core solver implements the discussion from previously, given the optimal value of the linear program as input. To avoid too many single-letter variable names, we use linearObjective instead of $ c$.

def solveGivenOptimalValue(A, b, linearObjective, optimalValue, learningRate=0.1):
    m, n = A.shape  # m equations, n variables
    oracle = makeOracle(linearObjective, optimalValue)

    def reward(i, specialVector):
        ...

    def observeOutcome(_, weights, __):
        ...

    numRounds = 1000
    weights, cumulativeReward, outcomes = MWUA(
        range(m), observeOutcome, reward, learningRate, numRounds
    )
    averageVector = sum(outcomes) / numRounds

    return averageVector

First we make the oracle, then the reward and outcome-producing functions, then we invoke the MWUA subroutine. Here are those two functions; they are closures because they need access to $ A$ and $ b$. Note that neither $ c$ nor the optimal value show up here.

    def reward(i, specialVector):
        constraint = A[i]
        threshold = b[i]
        return threshold - numpy.dot(constraint, specialVector)

    def observeOutcome(_, weights, __):
        weights = numpy.array(weights)
        weightedVector = A.transpose().dot(weights)
        weightedThreshold = weights.dot(b)
        return oracle(weightedVector, weightedThreshold)

Implementing the binary search, and an example

Finally, the top-level routine. Note that the binary search for the optimal value is sophisticated (though it could be more sophisticated). It takes a max range for the search, and invokes the optimization subroutine, moving the upper bound down if the linear program is feasible and moving the lower bound up otherwise.

def solve(A, b, linearObjective, maxRange=1000):
    optRange = [0, maxRange]

    while optRange[1] - optRange[0] &gt; 1e-8:
        proposedOpt = sum(optRange) / 2
        print(&quot;Attempting to solve with proposedOpt=%G&quot; % proposedOpt)

        # Because the binary search starts so high, it results in extreme
        # reward values that must be tempered by a slow learning rate. Exercise
        # to the reader: determine absolute bounds for the rewards, and set
        # this learning rate in a more principled fashion.
        learningRate = 1 / max(2 * proposedOpt * c for c in linearObjective)
        learningRate = min(learningRate, 0.1)

        try:
            result = solveGivenOptimalValue(A, b, linearObjective, proposedOpt, learningRate)
            optRange[1] = proposedOpt
        except InfeasibleException:
            optRange[0] = proposedOpt

    return result

Finally, a simple example:

A = numpy.array([[1, 2, 3], [0, 4, 2]])
b = numpy.array([5, 6])
c = numpy.array([1, 2, 1])

x = solve(A, b, c)
print(x)
print(c.dot(x))
print(A.dot(x) - b)

The output:

Attempting to solve with proposedOpt=500
Attempting to solve with proposedOpt=250
Attempting to solve with proposedOpt=125
Attempting to solve with proposedOpt=62.5
Attempting to solve with proposedOpt=31.25
Attempting to solve with proposedOpt=15.625
Attempting to solve with proposedOpt=7.8125
Attempting to solve with proposedOpt=3.90625
Attempting to solve with proposedOpt=1.95312
Attempting to solve with proposedOpt=2.92969
Attempting to solve with proposedOpt=3.41797
Attempting to solve with proposedOpt=3.17383
Attempting to solve with proposedOpt=3.05176
Attempting to solve with proposedOpt=2.99072
Attempting to solve with proposedOpt=3.02124
Attempting to solve with proposedOpt=3.00598
Attempting to solve with proposedOpt=2.99835
Attempting to solve with proposedOpt=3.00217
Attempting to solve with proposedOpt=3.00026
Attempting to solve with proposedOpt=2.99931
Attempting to solve with proposedOpt=2.99978
Attempting to solve with proposedOpt=3.00002
Attempting to solve with proposedOpt=2.9999
Attempting to solve with proposedOpt=2.99996
Attempting to solve with proposedOpt=2.99999
Attempting to solve with proposedOpt=3.00001
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3  # note %G rounds the printed values
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
Attempting to solve with proposedOpt=3
[ 0.     0.987  1.026]
3.00000000425
[  5.20000072e-02   8.49831849e-09]

So there we have it. A fiendishly clever use of multiplicative weights for solving linear programs.

Discussion

One of the nice aspects of MWUA is it’s completely transparent. If you want to know why a decision was made, you can simply look at the weights and look at the history of rewards of the objects. There’s also a clear interpretation of what is being optimized, as the potential function used in the proof is a measure of both quality and adaptability to change. The latter is why MWUA succeeds even in adversarial settings, and why it makes sense to think about MWUA in the context of evolutionary biology.

This even makes one imagine new problems that traditional algorithms cannot solve, but which MWUA handles with grace. For example, imagine trying to solve an “online” linear program in which over time a constraint can change. MWUA can adapt to maintain its approximate solution.

The linear programming technique is known in the literature as the Plotkin-Shmoys-Tardos framework for covering and packing problems. The same ideas extend to other convex optimization problems, including semidefinite programming.

If you’ve been reading this entire post screaming “This is just gradient descent!” Then you’re right and wrong. It bears a striking resemblance to gradient descent (see this document for details about how special cases of MWUA are gradient descent by another name), but the adaptivity for the rewards makes MWUA different.

Even though so many people have been advocating for MWUA over the past decade, it’s surprising that it doesn’t show up in the general math/CS discourse on the internet or even in many algorithms courses. The Arora survey I referenced is from 2005 and the linear programming technique I demoed is originally from 1991! I took algorithms classes wherever I could, starting undergraduate in 2007, and I didn’t even hear a whisper of this technique until midway through my PhD in theoretical CS (I did, however, study fictitious play in a game theory class). I don’t have an explanation for why this is the case, except maybe that it takes more than 20 years for techniques to make it to the classroom. At the very least, this is one good reason to go to graduate school. You learn the things (and where to look for the things) which haven’t made it to classrooms yet.

Until next time!